OPTIMAL CONTROL FOR STERILIZATION OF CANNED FOODS by Manoj Mangesh Nadkarni B.Tech., Indian Institute of Technology, Bombay (1981) Submitted to the Department of Chemical Engineering in Partial Fulfillment of the Requirements of the Degree of MASTER OF SCIENCE IN CHEMICAL ENGINEERING at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY September 1983 © Massachusetts. Institute of Technology 1983 Signature of Author: Deptment . of Chemical Engineering 2 August, 1983 Certified by: . Alan-Hatton, Thesis Supervisor Accepted by: Chairman, Departmental Graduate Committee Archives MASSACHUSETTS INSTITUTE OF TECHNOLOGY OCT 241983 LIBRARIES OPTIMAL CONTROL FOR STERILIZATION OF CANNED FOODS by MANOJ MANGESH NADKARNI Submitted to the Department of Chemical Engineering on August 2, 1983 in partial fulfillment of the requirements for the Degree of Master of Science in Chemical Engineering ABSTRACT A mathematical model has been developed for maximizing the nutrient retention during thermal sterilization of canned foods. Pontryagin's minimum principle as applied to distributed parameter systems, has been used in the analysis. The necessary conditions for optimization, derived by variational methods, lead to bang-bang type of control which requires the sterilization and cooling. to be carried out at maximum rates of heating A numerical procedure is outlined for the solution of the resulting two-point boundary-value problem, which consists of a system of nonlinear, coupled partial differential equations. The unique behaviour of the costate variables makes it possible to integrate only the state equations in the forward direction in the time domain, and to iterate on the final desired reduction of micro organisms. A single point switching shows less nutrient degradation than any combination of multiple switchings. Optimization with constraints on the retort temperature results in a dual point switching where the control is either at a maximum or minimum, or is zero. The model has been applied to the sterilization of pork puree. Comparison with the nutrient retention obtained using other temperature policies, shows that bang-bang control with single. point switching represents the optimal solution. Thesis Supervisor: Dr. T. Alan Hatton Title: Assistant Professor of Chemical Engineering 2 ACKNOWLEDGEMENTS This is a necessary, but not a sufficient expression of my indebtedness to several individuals. Undoubtedly, the person to whom I owe the most, is my thesis advisor, Prof. T. Alan Hatton, whose able guidance and endless efforts will never be forgotten. His continuous encouragement and the willingness to spend long hours of discussions have enabled me to complete this thesis. Among others, I am grateful to Prof. Robert A. Brown for useful discussions on numerical methods, both in the class as well as outside the class. A word of very special thanks goes to Prof. Robert R. Tenney of the department of electrical engineering at MIT, for patiently explaining to me many intricacies of the optimal control theory. His help is greatly appreciated. The financial support during this research work was provided by MIT. There are many friends at MIT and elsewhere, who have helped me in different ways, perhaps most importantly, by giving a moral support. I owe it to all of them. Finally, humble appreciation for what he has done for dedicate this thesis to my father. 3 me, I wish in a to LIST Number OF FIGURES Title Page 2.1 Results of Saguy and Karel (1979) 24 4.1 Modified Control for Constraint on the Retort Temperature 60 5.1 Basis 71 for Obtaining a Single Point Switching 5.2 Condition for Optimal Switching 74 5.3 Condition 75 for Switching for the case of no Constraint on the Final Concentration of Micro Organisms 5.4 5.5 Bang-Bang Control: Single Switching 77 Final Concentration of Micro Organisms 79 as a Function of Switching Time 5.6 Results for Computer Simulation for BangBang Control 80 5.7 Effect of Variation Control u 81 5.8 Variation 5.9 Nutrient Degradation at Different Activation Energies 84 5.10 Optimal Control for Upper Constraint on the Retort Temperature 86 5.11 Optimal Control with Lower Constraint on the Retort Temperature 87 5.12 Bang-Bang Control: Two Switching Points 94 5.13 Results for Two Switching Points 95 5.14 Optimal Concentration Using Two Switching Points 96 in Process 4 in the Limits of the Time 83 5 Title Number Page 5.15 Optimal Control: Three Switching Points 97 5.16 Optimal Concentration with Three Switching Points 99 5.17 Variation 99 in Second Switching Time t 2 ( Three Switching Points ) 6.1 Maximum Nutrient Retention at Constant Uniform Temperature 6.2 Results of Teixeira 6.3 Spatial Domain 107 6.4 Optimal Control for Sterilization of Pork Puree 109 6.5 Temperature Distribution at the CenterPlane 111 6.6 Nutrient Degradation during Cooling Process 113 6.7 Variation in Maximum Rates of Heating and Cooling 115 6.8 Nutrient Retention at Different Inlet Hot-Fill Temperatures 117 6.9 Process Time and Total Sterilization Time at Different Inlet Hot-Fill Temperatures 118 6.10 Variation 119 104 et al.(1975) in Limits of the Control 103 u LIST OF Number TABLES Title Page 4.1 Variables and Parameters for System Equations 44 5.1 System Equations 66 5.2 Initial, Terminal and Boundary Conditions for System Equations 67 6.1 Process Data 102 6.2 Results for Two Dimensional Model 114 6 CONTENTS Abstract 2 Acknowledgements 3 List of Figures 4 List of Tables 6 Chapter Introduction 9 Chapter Principles of Thermal Processing 14 2.1 Thermobacteriological Considerations 14 2.2 Retention of Nutrient Quality 18 2.3 Approaches to Optimization 21 Optimal Control Theory 25 3.1 The Minimum Principle 26 3.2 Bang-Bang Control 32 3.3 Distributed Minimum Principle 34 3.4 Remarks 38 Mathematical Model for Maximizing Nutrient Retention 40 4.1 Model Formulation 40 4.2 Optimality Conditions Using the Variational Approach 47 4.3 Comments on Optimality Conditions 55 4.4 Constraint on the Retort Temperature 57 Chapter Chapter 7 8 Chapter 5 Numerical and Analytical Solutions Strategy 64 64 5.1 Computational 5.2 Analytical Solution 85 5.3 Multiple Switchings: Numerical Solutions 91 Two Dimensional Models 101 6.1 Preliminaries 101 6.2 Results 108 6.3 Discussion and Conclusion 120 6.4 Suggestions for Future Work 123 Chapter 6 Appendix I Nomenclature References 125 130 CHAPTER 1 INTRODUCTION Sterilization of canned foods is one of the major tions in the food processing industries. The opera- sterilization process involves a suitable heat treatment of canned foods that the can contents are virtually free of any isms which could degrade the food. It micro is organ- impractical achieve complete sterility; the cans are commercially ized so that a desired degree of destruction of isms is achieved. tion is that the One of the major thermal problems processing nutrients and food quality. The and organoleptic properties, the to steril- micro organ- in the steriliza- causes destruction deterioration and so of of nutrients destruction of micro organisms follow similar kinetics, and hence higher kill rates for micro organisms also quality. involve greater loss in While it is highly undesirable to incur any degrada- tion of nutrients, it is nevertheless essential to ensure a factor sufficient reduction of micro organisms, usually by of five nutrient to ten log cycles, and hence nutrient loss is unavoidable. Much attention has been paid in recent years to izing the nutrient quality in a sterilization process to a specified reduction of micro 9 organisms. maximsubject Sterilization 10 continuously or can be carried out in a batch retort In both cases, the hydrostatic sterilizers. retort ture and the process time are the two variables mine the degree of sterilization. made to design temperatures, retention. a with suitable a view to been have the maximizing deter- retort for policy in tempera- which Several studies control as nutrient Most of these studies are based on optimizing some either functional form for the temperature-time relationship, analytically or experimentally, and are approximate in nature. However, the recent advances in computer-aided process design, and the development of modern control theory have provided the The potential for a new approach to optimization. originally this approach is the "minimum principle" which was developed for lumped systems by Pontryagin and for basis co-workers in the early sixties (Pontryagin et al., 1962). Subsequent developments have extended the concepts to incorporate distributed parameter systems. The major advantage associated minimum principle is that now with state the distributed the variables control variables can be spatially distributed in of dimensions. The control system then While the distributed information and can more systems than can lumped model accurately models, it can also the number coupled, and provide represent is any involves nonlinear partial differential equations of first order. and much second more the physical more complex. 11 Compl&te analytical However, with the are solutions finite of availability possible. not generally and difference the Galerkin finite element methods for discretizing partial differential equations, numerical solutions can be achieved to a high degree of accuracy. food Application of the distributed minimum principle to problems processing should optimization of nutrient quality sophisticated approaches. better provide We make a no (e.g. regarding the functional form those than high priori assumption temperature versus in temperature Also, the distributed nature of the model allows for a rigorous mathematical formulation of the problem with simplifying studies. short conditions Rather it is the necessary the minimum principle which determine the optimal profile. less on based time, ramp, sinusoidal etc.) of the retort temperature time relationship. for procedures assumptions than those In this thesis, the optimal canned foods is investigated using in embodied thermal the fewer earlier processing distributed of maximum principle as a starting point. In what follows, the salient aspects of food processing operations and the associated kinetics are discussed. In the second chapter, the basics of sterilization processes and the Note that the terms 'minimum principle' and 'maximum principle' are used synonymously. Actually, they differ only in sign. 12 reviewed. for work esearch past nutrient optimizing retention are with the The third chapter is primarily concerned tools for optimal control. time It deals with the continuous minimum principle, analyzes some of the the of limitations lumped model and establishes the need for a distributed model. described also A generalized distributed parameters model is here. In the fourth chapter, the emphasis is on the application of the minimum principle to the problem of optimizing nutrient retention. The are optimality for conditions necessary derived here starting from the basic principles of variational the calculus. The discussion on numerical methods for solving system above-mentioned equations, fifth the in follows chapter. We have considered here one dimensional slab geometry for mathematical nature and of the optimal computational control is also presented in this chapter. chapter, we are mainly concerned it known, derive analytical solutions for system In the the is possible to These are equations. sixth and the exists for puree, comparison final of a with an application dimensional model to sterilization of pork for which significant data Once convenience. a two process purposes. Most of the previous research has failed to account rigorously for sterilization during the cooling cycle, necessitating some new definitions. We elaborate two main aspects of our model here, namely the minimum principle leading to optimal nutrient 13 retention and the distributed nature of the system giving more accurate predictions. Conclusions and some suggestions for future work follow. Though the present work is primarily concerned maximization of the nutrient quality subject to the with an integral constraint on the final concentration of micro organisms, the results would very well apply to any process which is governed by a diffusion-like equation with the surface concentration or temperature as a control variable. be to enhance one kind of reaction other, where both would have The objective product similar and then suppress temperature kinetics with different activation energies. would the dependent CHAPTER 2 PRINCIPLES OF THERMAL PROCESSING It is of prime concern for the food to be able to design a flexible heat sterilizing canned foods. processing treatment engineer schedule The objective is always to that sufficient destruction of contaminating micro is achieved with a minimum loss in quality. the beginning of the twentieth nutrient century, several ensure organisms Since researchers have reported analytical and experimental work in this Bigelow et al. method (1920) gave the first general for area. for calculating thermal process times, and shortly thereafter Ball (1923) defined several parameters of interest in thermal processing and developed mathematical formulas which are still used in the food processing industry. Several others have done pioneering work in this area during the last few decades. There have been numerous textbooks describing the practice of thermal processing, such as those by Beverloo (1975), Loncin and Merson (1979), Charm theory and Leniger and (1978) and Harris and Karmas (1975). 2.1 TERMOBACTERIOLOGICAL CONSIDERATIONS The micro organisms present 14 in food are destroyed by 15 thermally induced changes in the original of the proteins usually assumed This leads to an in the cells. the cells to reproduce. to by a first structure inability of micro The destruction occur chemical of organisms is irreversible order, reaction often described by the rate equation (2.1) dN/dt = - K N exp(-E/RT) t is the time and the term where N is the concentration, K exp(-E/RT) is the temperature degradation of organoleptic quality such as vitamin C and The only difference in kinetics. values rate Chemical reactions occurring in food which lead constant. trients reaction dependent of the constants higher for nutrients and thiamine, these micro for (1982) and Lund (1975) give a detailed of activation energies for different follow is cases K and the activation than of destruction to nu- similar that the energy E are much organisms. listing micro of Thompson the values organisms and nutrients. However, it is usual practice in the food engineering literature to denote the temperature dependence of the reac- tion rate constant by K = Kref exp [ b(T-Tref) so that an equivalent organisms expression and nutrients (2.2) ] is for the destruction of micro 16 (2.3) dN/dt = - Kref N exp [b(T-Tref)] Furthermore, microbiologists and define two more parameters, D and z. to reduce the reactant concentration given temperature, while z is the food technologists D is the time by a factor increase required of 10, in at temperature necessary for reducing the value of D by a factor of 10. the definition of the reaction rate constant K, a one From obtains D=2.303/K, so that the kinetic expression becomes dN/dt = - (2 .303/Dref).N exp [(T-Tref)/(z/2303)] Here the time and the Dre f values are usually minutes, the z and the temperature and celsius (or Fahrenheit). The reference (2.4) expressed values in temperature in degrees Tref is normally taken as 121.1 degree C ( 250 F). Points along the center of a cylindrical the slowest heating region. The mechanism for can heat is assumed to be primarily conduction, though some currents may be set up in Leniger (1975) and Beverloo qualitatively. the case have of constitute transfer convective semi-liquid discussed this foods. aspect The coldest spot may occur below the geometric center of the can if during processing. there is The process no agitation design of containers calculations may be based on the temperature history at these points. Many a time, however, it is desirable to achieve a certain level of average 17 concentration over the entire can rather than tion at a few critical points. Some other the factors concentrawhich can affect the degree of sterilization are: the size and the shape of the cans, movement of cans in the sterilizer, empty space in the sealed can and perhaps the volume of composition of the vapour/gas mixture in this space. Hayakawa (1978) gives a summary of how the empirical analytical formulas have been developed in the past. Olson (1957) have developed tables for the and Ball and variation in process values with respect to heat transfer coefficients and the difference between the temperature of the center point and the surface of the containers. Stumbo (1953) developed mathematical model which based the process time on the a proba- bility of survival of micro organisms in the whole container, and not just the center of the can. Smith Tung (1982) have reviewed various compared their accuracy. Most recently, formula methods Most of the currently used and energy have sterili- zation methods seem to have a high factor of safety and there is considerable scope for saving and hence and improving a lethality the organoleptic quality. Lenz and Lund (1977 a,b) have developed Fourier number method for estimating sterility center of the can. Lund et al. value This is an extension of the work at the done by (1972), which introduces the concept of adiabatic equilibration temperature. The mass average temperature of 18 the food contents is higher than during the heating period. the centerline Containers are temperature heated until the average temperature inside the can is sufficient to induce the desired kill rates. Then they are held adiabatically until the centerpoint temperature rises distribution is achieved. and a uniform temperature This concept was first applied by Lund for quick blanching operations. 2.2 RETENTION OF NUTRIENT QUALITY In recent kinetics years, of nutrient research degradation nutrient quality of thermally has been directed to the and to the prediction of the Harris and processed foods. Karmas (1975) and Lund (1977) have discussed heat processing on nutrients. the effects of It may be ocassionally possible to improve the taste of the canned food by heat treatment, but that is rarely an objective. In most cases, the nutritional and the organoleptic quality of the food reduces as of thermal processing. The kinetics of nutrient are similar to those of the destruction of a result degradation micro organisms, two parameters of interest being the rate of nutrient destruction Kr, at a reference temperature Tr, for a first reaction, and the Arrhenius activation energy Ea . parameters are equivalent to the Dr-value perature Tr, and the z value, discussed at in These reference the order two tem- preceding 19 section. It has been observed by Joslyn and Heid (1963) that at higher temperatures, the destruction rate of bacteria accelerates more rapidly than the degradation rate of nutrients, thus favoring high temperature short time Ammerman (1957) studied the effects of (HTST) heat processes. treatment different processing temperatures and equal microbial ity values, on selected food constituents, by thermal calculations Stearothermophilus. organisms are Stumbo used C. (1973) energy-values for these micro lethal- analyzing effect on colours, flavours and nutrients such as The two main kinds of micro as vitamin a organisms the as This means that for a and B. activation 50-80 given C. in kcal/mole which are significantly higher than those of nutrients 20 kcal/mole). the basis Botulinum gives with (about increase in processing temperatures, the rate of micro organism destruc- tion will increase more rapidly rate nutrient than the degradation. This observation forms the basis of for optimizing rate constants processes to yield maximum nutrient retention. The activation energies and the reaction for different nutrients depend on factors such as medium, the composition of the food medium and reduction potential. There are several e.g. texture, colour and the flavour of pH its quality of the oxidation attributes, food, which exhibit similar degradation due to thermal processing. Their activa- 20 tion energies vary in the range 10 to 25 kcal/mole 1971). The activation energy for thiamine, has been studied extensively, is not a (Timbers, nutrient strongly which dependent on food medium or composition, and hence the mechanism of thermal degradation of thiamine appears to be the same for different media. the reference at the reference and the pH. However, the reaction rate constant at temperature (or equivalently the temperature) is a function D of the value medium For example, the thiamine destruction at pH 6.6 is about 12 faster than at pH 3.2 . Other nutrients include vitamin C and chlorophyll. frequently times studied Their activation energies vary between 10 and 25 kcal/mole. Mulley et al. (1975 a,b,c) have for thiamine destruction in various studied food reaction products observed that, in natural foods thiamine is more rates and heat have resis- tant than thiamine in buffered and aqueous solutions, although a first order reaction mechanism seems to hold for both acidic (pH less than 4.5) and non-acidic Ohlsson (1980 a) and Castillo et al. foods. (1980) More recently, have developed quantitative expressions for retention of nutrients. Thompson (1982) discusses some of the pitfalls the simplified approach of first order, non-cyclic reaction for predicting nutrient other nutrients using irreversible, degradation.- The degradation rate for a particular nutrient may be of the concentration of of and a hence function it may 21 concentrations change in a complex manner as various reactant However, in the absence change during thermal processing. and any detailed more to complex reactions, an approximation al. of reaction is kinetics Hill by given for appli- predicting nutrient quality. A general treatise on the cability be (1971) models mathematical are among others who have developed may order first (1977) and Jen et Downes and Hayakawa justified. for modeling mathematical accurate of and Grieger-Block (1980). 2.3 APPROACHES TO OPTIMIZATION There has been growing concern over the nutritional value of canned foods and this has led to active research programmes Broadly speaking, optimiza- for optimizing thermal processes. tion refers to a alternatives, that process which some preset quantities, Various approaches have selecting for procedure will subject been certain heat treatments of used canned or the design minimize maximize or certain constraints. to for this foods, one of equipment. For purpose. may optimize the temperature-time profile, or the size of containers, various , among wish shape and Teixeira to et (1975) have considered different geometries for containers al. to yield better heat penetration and thereby to optimize thiamine retention. Flat container geometries or cylindrical cans with 22 However, in very high height to diameter ratios are favoured. practice the shape and the size of containers is determined by other considerations such as cost of manufacturing and attractiveness of the product. rotation providing Processing equipment is designed for and agitation of cans during the sterilization process. There this area. designs patented are several variations and Hydrostatic sterilizers, which can process ture policy developed in more used now retort The optimal subsequent the continuously cans are in a pressurized high temperature space, frequently (Fairbrother, 1982). in temperawill chapters, be applicable to both batch-type processes as well as to continuous sterilizers. Teixeira et al. have (1969) shown on dependent retention of nutrients is also energy or the z value of nutrients. that maximum the activation the For a low z (high value E) nutrient, a process with low temperature and high time preferred while for those having high z values, shorter and higher process temperatures are Ball (1971), Teixeira et al. (1975) and and Thijssen (1980) have considered variable retort temperatures to ize the nutrient retention. However, study the influence of other factors as and simplifications inference. made, if any, one should well before times Hayakawa desirable. as is and Kochen maxim- carefully assumptions drawing For example, Teixeira and co-workers have any neglec- 23 ted the sterilization as well as the micro tion during the cooling detail in chapter 6. cycle. This organism will be destruc- discussed in Loncin and Merson (1979) briefly discuss the use of the Euler-Lagrange equation in optimizing nutrient quality. Saguy and Karel (1979) principle as applied to have lumped used Pontryagin's models, for minimum maximizing the nutrient retention during sterilization of canned foods. is possibly the only published application of optimal theory to food processing operations. optimal temperature profiles Figure obtained using 2.1 This control shows this the technique principle. These temperature policies are shown to improve the thiamine retention by more than two percent as compared with other schemes. Their results are encouraging, although some of their assumptions may not be strictly correct. They have averaged the micro organism and nutrient concentrations over the entire can, and have then used first order the averaged quantities. to be the fact that kinetics The reason for this averaging Pontryagin's applied only to the lumped models. minimum principle This is a for seems can severe be limita- tion of the model as presented by Saguy and Karel. There appears to be some scope for the formulation of the optimization problem more rigorously by the spatially distributed nature of taking the into system, account and applying the minimum principle. This is the prime task of then the 24 thesis research that follows in the subsequent chapters. t. a 6- 4 0 WJ g A a I.- 0 0.2 0.4 TIME, 0.6 0.8 1.0 ldimensionless) Optimal retort temperature (B0). mass average temperature (Bin, and central point temperature (B,) during the sterilization process (A 12 can). a I 44 4 w IC TIME, procn (dimensionloss) Optimal retort wnperature profiles for the sterilization of peapurse in #303 can and pork puree in A2K~ can. Fig. 2.1 Results of Saguy and Karel (1979) 3 CHAPTER OPTIMAL CONTROL THEORY Pontryagin's Minimum Principle is one of the major contributions to the development of optimal control theory. It has found a variety of applications in many different types of control in especially problems, electrical wish engineering. In this chapter, we to and review aerospace the basic statement of the minimum principle in one of its commonly used represent forms, using the state space approach throughout to the system. In what follows, we a consider deterministic system with a known relationship between the system states and the input control. The aim is to find the which drives the state X(t) to a desired functions continous and continously differentiable to first noted). The existence It objective. important to assume that all scalar and vector (unless otherwise control particular is are derivatives of a solution to the control problem must also be assumed in deriving the necessary conditions. The latter part of the chapter important aspects related to our subsequent bang-bang control and the distributed pr incipl e. 25 introduces research, minimum (or two namely maximum) 26 3.1 THE MINIMUM PRINCIPLE The minimum principle for lumped models is Lumped models are this section. discussed the by characterized in fact that the state variables are functions of time alone, and that the system dynamics can be described by ordinary equations. To begin, assume that the interest is completely represented by a differential linear of process physical nonlinear or system of coupled or uncoupled differential equations of the form, (3.1) X = f (X,u,t) where the m-dimensional control vector u(t) determines the n-dimensional state vector X(t). A fixed time interval [to,tf] for the process; is considered concepts to the case of it will be easy to variable terminal extend time. the General statements of the initial and final conditions are given by Q [(to),to] (3.2) = 0 and (3-3) R [X(tf),tf] = O respectively. Obviously, the dimensions of Q and less than or equal to n. R will be The following analysis considers the case where initial conditions are specified by equation (3.2). 27 The objective of the optimal control problem is to search for a control policy u(t) which will minimize a cost function, J. The exact form judgement. of the cost function The optimality conditions, namely is a matter the of conditions which will lead to a minimization of J, will depend strongly on the nature to able of J. Hence it is important be quantify the cost or the performance index correctly. Most of the time, the cost function consists of two terms, namely cost associated with the terminal state, and cost over the entire time domain. the to the accumulated In mathematical terms, the cost function J can be written as J = C [X(tf),tf] + g[X(t),u(t),t] dt (3.4) to Costate variables P(t), which are similar to the Lagrange multipliers for the case of static optimization, are to the system equations and incorporated into the adjoined cost func- tion such that J = C rI(tf),tf] + 5 {g[X(t),u(t),t] + to p T (t).rf(X(t),u,t)-X(t)] dt (3.5) A scalar function Hamiltonian is defined as H[X(t),u(t),P(t),t] = g[X(t),u(t),t] + PT(t).fEX(t)u(t)t] (3.6) 28 Incorporating the Hamiltonian into the cost function, one obtains J = C[X(tf),tf] + . H[X(t),u(t),P(t),t] - Pr(t).X I dt (3.7) *t A necessary condition for optimality is that the first variation in J must be zero for independent variations in the state vector X and the control the vector u. canonical equations for the state and (Pontryagin the This yields costate variables et al., 1962) X = JH/aP = f(X,u,t) (3.8) P = - a H/a (3.9) and an additional optimality condition, aH / (3.10) u = 0 The state variables X(t) are free to end on any manifold (i.e. can have any values at the terminal terminal time). A transversality condition of the form P(tf) = DC / x(tf) is obtained for the (3.11) costate variables. Thus, the terminal condition on the costate variables is determined by the nature of C, the cost for having a particular end point state. The control system equations constitute a two point boundary value 29 problem. This is an important characteristic of the minimum principle, where the costate variables always evolve backwards in time. When the terminal time is not fixed, there is tional equation corresponding to the variation in tf H[X(tf),u(tf),P(tf),tf] + C/atf + (R /tf).L an addi- which is = 0 (3.12) where the end point constraint on the state variables, described by an r dimensional vector R [equation (3.3)] is also adjoined to the cost function by an r-fold Lagrange multiplier L. Note that r is less than or equal to n. The transversality conditions are then given by P(tf) = C/@X(tf) + (R Thus there are r T /;X(tf)). L additional (3.13) equations given by (3.3) to account for the r dimensional unknown vector L. In cases where the control variable vector is constrained, the necessary condition given by equation (3.10) can hold only if the domain u completely in the interior of of the optimal the control constraints. speaking, it is no longer possible to take independent tions in u and the optimality condition (3.10) must is Strictly variabe re- placed by the more general form, H[X*(t),u*(t),P*(t),t] < [X(t),u(t),P*(t),t1 (3.14) 30 where the superscript * denotes optimal states. In essence, the necessary condition for optimality is the global minimization of the Hamiltonian H within the constrained domain for u. The variational approach has generally been deriving optimality conditions [Berkowitz(1961), used in Denn(1969)], although it may not always be possible to apply the techniques of variational calculus, for example, in the case or state variable inequality constraints. proof of the minimum principle is that Falb (1966). It gives a new A given interpretation of more control by to rigorous Athans the and costate variables P; namely, the costate equation for P describes motion of a normal of a hyperplane trajectory. along optimal The optimal trajectory represents in cost with change in the state space X cluded as a state variable. optimality, any portion which considers temporal According of optimal trajectory itself. time) the the the where to optimal state variation time the the is in- principle of trajectory is an This forms the basis for the proof variations and spatial perturbations in u that the motion of the hyperplane in . tf In described (free effect, by the terminal it shows costate equations does indeed lead to necessary conditions for minimizing the Hamiltonian. The proof of the minimum principle also yields one of the key properties for the state and the costate variables, namely, 31 < X(t), P(t) > = XT(t).P(t) (3.15) = constant variables i.e. the inner product of the state and the costate is invariant with time. It is not necessary to minimum the formal to use it. able to be principle know variational Ray and (1969) Gould cases. approach is adequate in most The the of proof appli- (1981) have discussed some of the chemical engineering cations of the minimum principle for lumped models (i.e. where the system can by represented be differential ordinary equations). Szepe and Levenspiel (1968) derived an solution for the optimal in the case temperature catalyst of Among kinetics. irreversible deactivation via first order, analytical columns separation processes, the application to distillation operations by Jones (1974). Other uses tower cooling to has been considered by Robinson (1970) and include optimization of polymerization temperature and initial initiator concentration for batchwise Jeng, 1978), two radical step catalytic (Chen polymerization chain in reactions bed reactors in the presence of deactivating of using the applied minimum principle. Pontryagin's minimum Murase in et principle fed batch al. in fixed immobilized enzyme catalyst (Patwardhan and Sadana, 1982). Yamane (1977) maximized the metabolite yield beds packed (Chang and Reilly, 1976), and the optimal operation and et al. culture (1970) optimizing have the 32 in temperature profile for ammonia synthesis a multitubular reactor heat exchanger system. 3.2 BANG-BANG CONTROL A very interesting case arises when the Hamiltonian given linear in the control vector u. The necessary condition by equation (3.10) does not respect to u since extremize the Hamiltonian H/;u is now independent of u. It is with can vector immediately seen that if the domain for the control be u to can be driven negative infinity by choosing infinitely large or infinitely is infinitely large, then the Hamiltonian However, small values of u. most in physical situations, there will be some bound on u, for example, Umin < u < (3.16) max The condition for optimality is given by equation (3.14), which in this case will be u = umax for H / u< O and u = umin for H / u > 0 (3-17) Thus the Hamiltonian is minimized boundaries of the domain of u. The by operating at the occur at the switchings 33 singular points where H/au=O. When the term not just at certain finite points but in a for a problem certain state-space interval, is said to have a singular arc, the contribution aH/au continuous the arc. is solution Along not be further the the singular is zero extremized choice of u. An additional equation is required to manner to of the u term to the Hamiltonian and the Hamiltonian can zero, by any solve for u, which is obtained by considering d/dt ( H/,u ) = 0 If the above equation (3.18) is insufficient, time derivatives of (H/au) successively higher can be set to zero to solve for u in terms of the state and costate variables. Bellman et al. (1956), Johnson (1965) have discussed bang-bang control and the and Sage singular arising in optimal control. One of the early (1968) solutions applications of the variational approach for deriving necessary conditions for a singular control was given by Desoer (1959). Canon (1964) have presented a solution in switchings for the problem of minimum space vehicles, where the control is terms fuel Athans of used bang-bang control for and deriving applied for Gibilaro optimal multiple consumption whenever a given amount of consumed fuel would result most efficient motion. Farhadpour and 'firing', in (1981) inlet in the have reactant concentrations in the case of an unsteady state operation of a 34 continously stirred tank reactor. Most of the examples of bang bang control, including those which are mentioned above, are for lumped systems. problems are, however, distributed in nature of the minimum principle to distributed and all Many application parameter models is discussed in the next section. 3.3 DISTRIBUTED MINIMUM PRINCIPLE The emphasis in this section will be on the development of the minimum principle for systems which are distributed nature. Distributed systems are characterized by partial differential equations in the spatial coordinate vector y time t. and Butkovskii and Lerner (1960) were among the first consider the minimum principle as applied to in to distributed parameter systems. Consider a distributed system described by DX(y,t) / t = f y,t,u(y,t),aX/y, ... , kx/yk, .. ] (319) The initial conditions on the state variables are specified at t=t o, and the boundary conditions on the distributed variables are specified in terms of X(t,y), ;X/ay, etc. at the ies of the spatial domain. The objective boundar- is to find an optimal control u(y,t) which will extremize the cost function 35 J = .C [X(ytf), k .X(y,tf)/y tf] dy + 5g[X(y,t), kX/-ay', y, u(y,t), tdy t As before, (3.20) i the first term in the cost associated with the terminal represents dt the accumulated spatial domain. The state, cost Hamiltonian function while over H the the is is the cost second term entire defined time as and before, namely, H[ X(y,t), aX/y k(y,t), y, t, u(y,t), P(y,t)] g[X(y,t), kX/y_ (y,t), u(y,t), y,t] + P(y,t).f[X(y,t), X/;y (yt), (y,t), y, t] (3.21) The system equations are adjoined via spatially distributed costate cost function Hamiltonian. can then be to variables expressed The process time is assumed though extension to a free the terminal in and the terms of the be is first variation in J, namely 5J, is found by function P(y,t) to time dent variations in the state, costate and cost fixed, al- possible. The taking control indepen- variables, and one then obtains the following necessary conditions: aX/ t = H/P = f (3.22) 36 _P/,t) = - H/ax - (-1 )k k/ yk H/ (3.23) x/[_1 and H[X*(yt), (DaX v,y, t, _ (y,t)] H[X*(y,t), (X/y* Equation (3.24) is , y, t, u(y,t)] the general form (3.24) for the global minimization of the Hamiltonian. If the control vector u is function of time alone, i.e. u(y,t)=u(t), then the a above condition reduces to SH[X*(y,t), ( kX/ayk)*, y, t, u*(t)] dy H[X*(yt), ( akX/y ), y, t, u(t)]dy (3.25) Similarly, for u(y,t)=u(y), this reduces to f H )kk),y,t, u (y)] dt 5 H[X_(yt), (;k/ap*, y, t, u(y)] Note also that when u is unconstrained, equal to zero for optimality, (3.26) t H/au can be either at each point in the case of equation (3.24), or over the entire spatial domain time domain in the cases of conditions (3.25) set and or the (3.26) respectively. (Ray and Szekely, 1973) The distributed components of the state or its spatial derivatives are generally variable specified vector on the 37 boundaries of y, and the initial state,X(y,to ) is transversality The be completely known. to assumed conditions on the applied to the costate variables evolve from the variations state vector x. The end point conditions (t=tf) are given as: P(y,tf) = C(y,tf) The transversality / X(y,tf) (3.27) conditions on the y of boundaries are derived for the distributed costate variables as ak-{ H /k kx/yk] = }1 / Equation (3.28) holds for all O (3.28) values conditions for optimality, namely (3.22) associated transversality conditions t. of to (3.27) derived using the variational approach. The necessary The and the (3.28) are (3.24) and steps detailed involved in using this approach can be best illustrated for particular problem at hand. This is shown in the next a chapter for the problem of maximizing the nutrient retention of canned foods during sterilization processes. There have been a few applications of the distributed maximum principle to chemical engineering processes. Hahn et al. (1971) have computed an optimal start up policy for a plug flow reactor wherein a first order, reversible and reaction takes place. The distributed maximum exothermic principle was also applied to the problem of catalyst deactivation by Ogunye and Ray (1971) and by Gruyaert and Crowe (1974). 38 Nishida et al. (1976) have compared linear and lumped models and distributed models, with their to problems in reaction engineering. They have nonlinear application considered catalytic series reaction in a tubular, nonisothermal where the control catalyst variable. activity This is appears the to radially the only a reactor distributed example of bang-bang control as applied to a distributed system. 3.4 REMARKS The necessary conditions principle, namely equations derived (3.8) to using (3.10) the or maximum (3.12) to (3.14) for lumped systems and equations (3.22) to (3.24) distributed systems, optimality. correspond Actually, the first variation in only J to local yields only the extremum of the cost function. By taking the second local variation .2J, it can be shown that this extremum does correspond local minimum. However, note that the global for to minimization a of the Hamiltonian does not imply global minimization of the cost function over all values of u (Athans and Falb, 1966). Thus more than one solution may be possible for the given system of state and costate equations and the associated transversality conditions. Some other observations are also in order. The ian H is always minimized as long as the optimal Hamiltonu is a 39 continous function. This may not occur, for example, switching points in bang bang control, since at the contribution is zero of u to the Hamiltonian these at the points irrespective of the choice of u. Thus the Hamiltonian can not be extremized at the singular points, which frequently referred to as is bang-bang why suboptimal control. should be noted that the only difference between and the minimum principle is that equation (3.14) or (3.24) must be the control is Further, it the Hamiltonian globally maximized maximum in the rather than minimized; other necessary conditions remain unchanged. CHAPTER 4 MATHEMATICAL MODEL FOR MAXIMIZING NUTRIENT RETENTION The main task in this chapter is to develop a cal model for optimal control of the mathematiprocess. sterilization The system equations are cast in dimensionless form, and based on the variational approach, a detailed proof is presented for shown the derivation of the optimality conditions. It is how policy. the problem formulation leads to a bang-bang control The control must be modified when inequality constraints are incorporated into state variables, such as a constraint on the retort temperature. The resulting policies control less optimal in such cases than those obtained with will be fewer or no constraints. 4.1 MODEL FORMULATION During the sterilization process, heat transfer the can occurs primarily by conduction. This holds or semi-solid foods for which the convection for currents inside solid inside the can are insignificant. Many of the liquid canned foods, on the other hand, have pH less than 4.5, so requirements for these foods are not thermal processing the convective heat 40 that severe. transfer sterilization Also,during coefficients 41 at the surface of a can are very high, thus assume that the temperature at the can enabling surface one equals to the retort temperature. Heat conduction inside the can is governed by the equa- tion, , T(y',t')/ t = V2 T(y',t') (4.1) with appropriate boundary conditions on the time and spatial domains. Thus, T = T i at t = 0 (4.2) and, at the planes of symmetry in the spatial domain, we have 'V T = 0 where (4.3) is the unit outward normal vector. At the external boundaries of the spatial domain, T = TR (t') (4.4) The retort temperature TR(t), which is the variable to be controlled, affects the system behaviour boundary condition (4.4). Denn et al. only (1966) solutions to problems where the control may through have operate the boundaries. Here, however, we find it more apply the minimum principle by transferring the the discussed only at convenient to control from 42 the boundaries to the system equations. For this a purpose, new variable T2 (y',t') is defined as follows: T2(Y',t') = T(y',t') - TR(t') (4.5) The heat conduction equation is now modified to read ~T2(Y',t')/ t' = V 2 T 2 (y',t') - dTR/dt' (4.6) with the initial condition 2 = Ti - TR(t'=O) (4.7) at t'=O At the planes of symmetry, we have _.VT = 0 2 and at the boundaries (4.8) of y', T2 = 0 (4.9) This formulation transfers the problem inhomogeneity from the boundary conditions to the differential itself, equation and this can have some advantages in the eigenfunction sion solution of the problem. expan- It is interesting to note now the rate of change of retort temperature u', or should be considered as the control variable rather that dTR/dt', than the concentrations are retort temperature itself. The nutrient and the micro governed by the equations, organism 43 3C N / t' = - Ko,N.CN exp(-EN/RT) (4.10) = - KO,M.CM exp(-EM/RT) (4.11) and CCM / t respectively. The initial conditions on the nutrient and micro organism concentrations are: (4.12) CN = CNo at t'=O and (4.13) CM = CMo at t'=O respectively. Table 4.1 describes which appear in the the variables formulation of and system the parameters equations. In dimensionless form, the system equations can be represented as dX1 / dt = u X2 / at = DX3 / (4.14) 2X2 - u (4-15) t = - a 1.X 3 exp[-E/(X1 +X 2+1)] X 4 / at = - a 2 .X4 exp[The initial and boundary (4.16) E/(XI 1+X 2 +I)] conditions (4.17) on the variables are given as follows. At t=O, we have dimensionless 44 TABLE 4.1: VARIABLES AND PARAMETERS FOR SYSTEM EQUATIONS. Variables/Parameters. Dimensional Time t' Characteristic Length L Space y' Temperature T(y',t') Retort Temperature TR(t') Initial Temperature Ti 1 y = Initial Nutrient Concentration CNo Micro organisms Concentration C (' Initial Micro org. Concentration CMo Activation Energy Nutrients, Micro org. Ea EN EM X1=(TR - Ti)/ T i 0 t') X 2 = (T - TR)/ T i X 3 = CN/ CNo 1 t') X 4 = CM/ CMo 1 E = Ea/RTi = E /(X 1 +X 2 +l) Exponential Term Reaction Rate Constants '/ L X1+X2 T - TR C (' L2 t =t'/ Temperature Difference Nutrient Concentration Dimensionless K a = KoL2/ Contd. 45 TABLE 4.1 (Continued) Variables/Parameters Rate of Change of Retort Temperature Dimensional u'= dTR/dt' Thermal Conductivity oC Ratio of Activation Energies F Dimensionless u = dXl/dt = E /E M N 46 X1 X2 = X 3 (4.18a) = X10 - (4.18b) X1 0 = 1 (4.18c) = 1 (4.18d) and X If the initial retort temperature hotfill temperature of the is the cans, then same X10 as will the inlet be zero. Boundary conditions are required only for X2. At the planes of symmetry, Q.VX 2 (4.19) = 0 and at the external boundaries of yI, (4.20) X2 = The system is autonomous, i.e. not explicitly dependent on time t. In shorthand notation, it can be represented as ax/ at = f[x, , (4.21) v 2 X2 ] The objective is to maximize the nutrient retention the entire can during sterilization. minimizing the cost function J, where This is equivalent over to 47 J = 9 (- (4.22) X3 ) dy dt or J = 5 t=o y 2 +1)] fa 1 .x 3 exp[-E/(X+X processes, sterilization In most final on the average it is essential micro of concentration to Then level. organism achieve a specified reduction in micro the constraint (4.23) dy dt organisms can be described as 5x (/V). 4 dy < C 1 (4.24) Typical where V is the dimensionless volume of the container. values of C 1 may be in the range from 10- 5 to 10-10 . 4.2 OPTIMALITY CONDITIONS USING THE VARIATIONAL APPROACH optimi- The general form of the necessary conditions for zing a distributed chapter. The aim in system this was presented section specific necessary conditions for will in to be maximization the previous the develop of nutrient using retention by considering the dynamics of the system and the principles of variational calculus. One dimensional geometry is considered first to keep the mathematics However, a logical extension to two and three slab simpler. dimensions can be conveniently made. The spatial variable y has domain [0,1]. 48 The Hamiltonian H = + as (4.25) T f where P(t) is represents and H is defined the distributed costate variable the right hand side of the system is the integrand of the cost vector, equation function f (4.21) defined by equation (4.23). Hence the Hamiltonian can be written as H = alX 3 exp[-E/(X 1+X 2+1)] + P 1 u + P 2( D 2X 2 / P 3 (-alX3 exp[-E/(X1+X 2+1)]) + P4 (-a2 X 4 exp[- y2 _ u ) + 8E/(X1+X 2+1 )]) (4.26) The state equations (4.14) to (4.17) are adjoined to cost function via the costate variable vector P and the the constraint on the final concentration of micro organisms via a Lagrange multiplier . The modified cost function can then be written as J = [ SX4 dy - C1]t=tf + + P1 (u - dX1 /dt) i aX 3 exp[-E/(X1+X2+1)1 + P2 ( a2X 2 / ay2 _ u - + P3 (-a 1 X3 exp[-E/(X 1+X2 +1)] + P4 (-a2 X4 exp[-.BE/(Xl+X 2+1)] - 2/ t) X3 / )t) X4 / at) I dy dt (4.27) 49 which can be expressed in terms of the Hamiltonian as J = U [ S' 4 dy - C 1] t=tf + f §§[H- PT( X/ a2t)] dy dt (4.28) In the following analysis, the process time tf is assumed to be a predetermined factor having a fixed value, so that no variations in tf are considered. Then one can obtain the first variation in J by varying X, Xyys u and P. Thus I S t X 4 dy I I[[ t=tf + o _ ( sp)T( $2X2/;y2 has only one non H/ aX+ (_p)T DH/ aP O T + ( Su) H/Du + ( i y) Here Xy X)T H/Xyy - PTD( X)/ at X/ at)] dy dt zero component, (4.29) namely X2,yy Note that pT ( $x)/Dt = [PT( tX)/ - ( X)T DP/ t (4.30) Also, consider the terms )[( X2) ( H/aX2 ,yy)/?y]/ [ ( DH/ 2,yy)/y] + ( ay = [ 2) ( SX2 )/ ay]. 2 (9H/ X2,yy)/&Y The first term on the right hand side of equation be expressed by using another variational relation (4.-31) (4.31) can 50 [ 11( X2)/ ay] I Z3( D ( L20H/ 2) -')X2,yy) / ..)Y ; ( Sx2)/ sy] /ay .H/ ( 'SX2, yy) Combining equations (4.31) and X2, yy and (4.32) 2, yy (4.32) rearranging, one obtains ( X2,yy) ( X2 ) 2 ( ?H/ aX2,yy DH/ aX2,yy 'a ()H/ ;IX2, y) . a( Cx2)/ y] / s + a[ ( Sx2) . ( aH/ ax2,yy) /y] /) ;Iy / ay (4-33) The expressions given by equations (4.14) and (4.17) incorporated into the variation SJ of are (4.29) equation to obtain tF SJ= > SX 4 dy' t=tf + ( D + (u) H/ u + [( H/ )Tp/ at - aPT( S]/ + ( PT H/ a P $ 2,y) ( (x2)a2( H/ x:I 2, yy V a 2 + ( aH/ ) D( SX2 ) / a;Y]/'" + 1. ) ;)(2a / DX2, yy)/ ar/ t ( pT( DXI at) I aI dydt (4-34) For the first variation in J to be zero, the coefficients of the independent variations, ( P) , ( X) and u must be zero, which leads to the following necessary conditions: D X/ aIt = H/ P = f (4-35) 51 aP/at y2 - 2( H/I a6)/ / - (436) (4.37) H/ au = O and The second term on the right hand namely 2( H/ side equation _Xyy)/ay2 is non zero only X 2. Also, strictly speaking, equation only if the variations in u constrained of domain are for (4.37) allowed to the will be since u is a for u. Secondly, variable valid be within the function of (3.25) will apply. i.e. since the control variable u does not change over of time alone, a necessary condition the type (4.36), the spatial domain of y, the Hamiltonian H need be minimized only over the entire domain of y at any time t, rather than at each point in y. This means that equation (4.37) will modify to { S Hdy}/u'(-H/= (4.38) u) dy = 0 o S or more generally, H[X*(y,t), ( 2 y2)* ,y, (t)] dy< t, S[I(yt),,( ;X2/Qy2) , y, t,u(t)]dy Next, note that the Hamiltonian as defined by equation (4.39) (4.26) is linear in the control variable u, so that BH / u = ( P1 - P2 ) (4.40) 52 The equation (4.38) cannot yield an optimal value of aH/au is independent of policy is required. u, and hence The constraint on a u bang-bang the rate since control of rise of retort temperature, u, will be of the type (4.41) Umin < u < Umax where umax corresponds to the maximum rate of heating and Umin corresponds to the minimum rate of heating (maximum rate of cooling). The desired control policy will then be of the form u = umax if SD /H u dy = and u = Umin if H /u S(P1-P2) dy < 0 dy = §P1-P2) dy > 0 (4.42) Switching(s) in the control will occur at point(s) term I where I =2(P1-P 2 ) dy] is equal to zero. For the existence the of a singular arc, the term I must remain zero continously over a certain time period. This will be discussed in some detail later. It is readily demonstrated that the Hamiltonian minimized only over the entire (4.39)] and not at each individual spatial domain need [equation point in the domain Consider for example, a spatial discretization of be the of y. system equations (4.14) through (4.17). At each time interval a given partial differential equation is converted into a coupled the ordinary differential equations. If number of minimum 53 principle developed in the preceding chapter is applied this new, lumped model, an equivalent optimal control found. The Hamiltonian at each time interval will be of contributions from each ordinary differential Thus the condition for the global minimization of to can the be sum equation. Hamiltonian will indeed yield equation (4.39) as the number of discretizations goes to infinity. The transversality conditions are obtained by setting the remaining terms in equation (4.34) for SJ to zero. Thus one obtains [ 4 dX dYt f f t [ (x) I' X)dy] - e4 + b Cc~H/ ( H/ X 2,). ( 2x )dt] X2 )/ y dt = o (4.43) C Terms inside each of the square brackets must be zero, which leads to P1 (tf) = (4-44a) P 2 (tf) = 0 (4.44b) and P 3 (tf) = 0 (4.44c) Also, P4 (tf) = V (4.45) 54 Since X is specified at t=O, ( X) Note that 'y= O at X2 / is uniformly zero y = O and X 2 = 0 at y = 1 at t=O. so that the transversality conditions on P2 become X2,yy DH/ a ( H/ = P2 = 0 at y = 1 a2, yyP = y / (4.46) = 0 at y = 0 2 (4 47) Note also that the first necessary condition, equation (4.35), gives back the system state equations. The governing equations for each of the costate variables are: aP1 / t = - (1-P 3 ).a1 X 3 exp[-E/(X1 +X 2+1)]. E/(X1+X 2+1) 2 + p4.a2 .X 4 exp[-BE/(X 1+X aP2 / at =- 2 1 )]. BE/(X 1 +X2 +1 )2 (4.48) (1-P3 ).a 1 .X 3 . exp[-E/(X1 +X 2+ )]. E/(X 1+X 2+1) 2 + P 4 .a 2 X 4 exp[- BE/(X 1+X 2+1 )]PE/(X1+X 2+ 1 ) 2 _ 2P2 / Dy2 (4.49) which is equivalent to aP2 / at = P1l/ t - 2P 2/ y2 DP3/ t =- (1-P3 ).a1 exp[-E/(X+X2+1)] (4.49a) (4.50) and laP4/ t = P 4.a 2 exp[- BE/(X1 +X 2+1 )] (4 51 ) 55 4.3 COMMENTS ON OPTIMALITY CONDITIONS The complete system of canonical equations by equations (4.14) through (4.17) for is state described variables equations (4.48) thru (4.51) for costate variables P minimization of the Hamiltonian Boundary conditions and (4.42) initial for the conditions and the control for the X, u. state variables are given as follows. At y = O, X 2 /ay (O,t) = 0 at y = 1, X 2 (1,t) At (4.19a) = 0 (4.20) t = 0, X1 (0) = (4.18a) X 2 (,O) = X2in (4.18b) X = 1 (4.18c) = 1 (4.18d) 3 (y,O) X4 (y,) The 'initial' conditions on the costate variables are time t = tf, and are given by equations while the boundary conditions for P2 are (4.44) given and by (4.46) and (4.47). Note that the Lagrange multiplier at (4.45), equations X is an unknown parameter, for which there is an additional constraint equation of the form (4.24) with the equality constraint as 56 the limiting case. i.e. (4.52) 5x 4 dy = C 1 (1/V) $ Some observations are in order. Integrating the equation (4.16) from zero to t, one gets t X3 (y,t) = exp I a1 exp[-E/(X1 +X2 +1)] dt - (4.53) I D Let P3' = (1-P 3 ), so that the governing equation for P3' becomes ?P3'/t = P3 'a1 exp [-E/(X1 (4.54) +X 2 + 1 )] with P 3 '(tf) and = 1, integrating tP'(t = - the preceding obtains equation exp exp[-E/(X 1 +X P3'(t) = exp - 2 1)] a1 exp[-E/(X1+x2+1)] from t = tf to t, one dt t (4-55) Using the fact that +· g dt = tF g dt - t g dt it is easy to show P3 '(y,t) = X 3 (y,tf) / X 3 (y,t) or P 3 (y,t) = 1 - X 3 (y,tf)/X 3 (y,t) (4.56) (4.-57) 57 and Similarly, it follows that for the state variable X 4 the costate variable P4, we have X4 (y,t).P4 (y,t) = X 4 (Y,tf).P 4 (y,tf) = f(y) (4.58) Equations (4.57) and (4.58) represent a key property for of the product the costate variables, namely that the inner state and costate variables is time invariant. This is similar to the condition described by for (3.15) equation to models. These equations also eliminate the need lumped integrate two of the costate variables backward in time, or they can used to provide a check on the of accuracy be numerical the integration backward in time for the costate variables. 4.4 CONSTRAINT ON THE RETORT TEMPERATURE Most of the sterilizers operate in the food industry in a certain range between a maximum and a minimum temperature. many cases, the minimum temperature of the retort is In anywhere from 70 to 100 degree C, some times equaling the inlet temperfor atures for the cans. The maximum temperatures continuous or batch retorts vary from 125 to 140 degree C, and seldom do they exceed these limits. Clearly, the maximum temperatures at any point inside the can will be Excessively high localized temperatures 'cooking' of the food. well must It within be these limits. to prevent avoided will be essential to 58 incorporate a constraint on the dimensionless state variable X 1, of the form, b1 X1 4 b2 (4.59) Constraints on the state variables can mally using the penalty-function be approach treated [Bryson for- and Ho (1975), Sage (1968)]. For the constraint described by equation (4.59), we define a new variable X5 , whose dynamics are governed by aX5/ t = (b1 - X1 ) .(b 1 - X 1) + (b2 - X1 ).t((X 1 - b2) (4.60) where the Heaviside step function Yj is defined as })(f) = p for f>O and 3-(f) = 0 for f<0 (4.61) X5 It can be immediately seen that is the accumulated penalty for violating the constraint. The quantitative measure of the penalty is given by a positive number p which will have a larger value the more stringent the constraint. constraint is to be always satisfied, we should X 5 (tf) = O. The deviation the penetration into in X 5 (tf) from zero the forbidden domain. If ideally will have represent Adjoining system equation (4.60) to the cost function and the the the Hamilton- 59 ian of the original problem via a new costate variable P5, one obtains a new Hamiltonian H', H' = H + P (bl-X 1) - (bl-X 1 ) + (b 2-X 1 ) f(X-b ) ] (4.62) 2 The costate equation is given as 3P5/ t = 0 (4.63) There are no initial or final conditions on P5, while X 5 (t=O) = 0 (4.64) and ideally, X 5 (t=tf) = 0 (4.65) (4.63), one can see that P5 = constant. From equation value of this constant should be chosen so as to X 5 satisfies equations ensure The that (4.64) and (4.65). Figure 4.1 shows a portion of a control policy where of the switchings occurs at t=ts, for the straint on the retort temperature. The and temperature profiles are consider a case where a shown constraint case of no corresponding as dotted of the imposed on the surface temperature of the can. con- control lines. form one Next, (4.59) We can consider three possibilities, namely u>O, u=O or u<O for but sufficiently close retort temperature X1 to t 1. For the will be first greater case, than b2 is now t>t1 u>O, the and thus (X1 -b2 ) will contribute to the penalty X5, given by equation (4.60). Clearly then X5 (tf) will be greater than zero. By 60 Fcr a: -J 0n- z 0L) LLJ z 0 0 t1 t$ t2 tf T IME FIG. 4.1 Modified Control for Constraint on the Retort Temperature 61 contraposition, in order for X 5 (tf) to be equal to constraint must always be satisfied, cannot be greater the i.e. Xl<b2 . Hence u In fact, this a than zero for t=tl+. without is For analysis. trivial example of the use of penalty function this case, one could easily argue, zero, aid the of the penalty function approach, that in order for the constraint to be always satisfied, u must be less than or equal to zero for on X 1 is satisfied, the t=t1 +. Again as long as the constraint Hamiltonian H' defined by equation (4.62) will be the same as the Hamiltonian H for the unconstrained case. It is implicitly assumed that zero, which must be the case for umin min will be less corresponding than to the maximum rate of cooling. The control variable u must be chosen so as to minimize the Hamiltonian (4.26) over the entire spatial H, domain. defined Note by also equation that the coefficient of u in the expression for 5 H dy, namely the term I [I = ~ (P 1-P 2 ) dy], is negative at time t = t1 (point A in figure 4.1). Since the term I is a continous function, it will continue to be negative for t>t1 but sufficiently close to t1. t=t1 +, its contribution to the Hamiltonian will be a positive number and thus the Hamiltonian cannot be minimized by selecting u to be less than zero. This leads to the conclusion that be Hence if u is chosen to be less than zero for u must zero after t=t1 , which means that the retort temperature remain constant at b 2. This will in turn modify the must profile 62 for the integral term I, so that it may become zero at time t, not necessarily equal to ts. The switching some to u=umin the lower constraint Xl=bI is approached from above. It should be noted will occur thereafter. A similar analysis holds for that the contribution of u to the the case term equal to zero when X1 is on the boundaries This is however different from the usual when .H.dy is of the uniformly constraint. case of control . In singular control, the contribution of singular u to the Hamiltonian is driven to zero because the coefficient of u is zero along the singular of singular arcs the arc. More importantly, Hamiltonian can not in the case be minimized with respect to the choice of u, as opposed to the above case where u=O is an optimality condition for minimizing the Hamiltonian. The possibility of singular arcs must not either in the above case or in the case where be overlooked there is no constraint on X 1. For a singular arc to exist, one must have d/dt[ S(P1-P 2) dy ] O (4.66) over a certain range of time. However this is very unlikely to occur since the system state and costate equations are highly nonlinear with several exponential terms and partial derivatives with respect to spatial coordinates. It will difficult to foresee a situation wherein the integral be very term I 63 is continously kept zero by simultaneously adjusting different state and costate variables (Tenney, 1982). The canonical partial differential equations described in this chapter are coupled, nonlinear and distributed, for which a complete analytical solution is not solutions often involve discretization possible. of the Numerical partial ferential equations and iteration on the resulting two difpoint boundary value problem. This is dealt with in the next chapter along with a discussion on the slab geometry. results for one dimensional 5 CHAPTER NUMERICAL AND ANALYTICAL SOLUTIONS Optimal control of distributed parameter systems involves the solution of a equations. These are system of partial and state canonical costate equations, differential often at coupled and non linear, with boundary conditions specified different ends of the time domain for different variables. most Clearly, numerical solutions are the only choice for these kinds of problems. A procedure for tions is presented in this chapter. This computa- numerical is switchings. tions for single switching are compared with tiple switchings, thereby providing bang-bang control with single a case a of solu- Numerical for mul- evidence that those strong switching by followed closed form solution to the state equations, for the bang-bang control with multiple of is the optimal solution. 5.1 COMPUTATIONAL STRATEGY It was shown in the preceding chapter that the of the Hamiltonian with respect to the control u linearity leads bang-bang control where the control now operates only boundaries of the admissible domain. 64 We first on consider to the a 65 one solution to the problem described in terms of the dimen- sional slab geometry with no constraints on the state variable X1 . It will be possible to extend the ideas with case of cylindrical can geometries to further the temperature surface constraints. The governing state, costate and control equations, which were derived preceding the in collected form in table 5.1 . presented are chapter, in condi- The associated boundary tions on the state and costate variables are outlined in table 5.2 . There X 3 (yt), are X 4 (y,t), nine variables, X 2 (y,t), X(t), namely, P 1 (ySt), P 2 (y,t), P 3 (y,t), P4 (y,t) and u(t); which are to be solved using the nine equations in table 5.1 . 4 are known at t=O, while Note that the state variables X 1 to the costate variables P1 to P3 are costate variable P4 is unspecified at specified at either t=tf. end. It is The equal the integral constraint on the final micro organism concentration equation by equation to an unknown at t=tf, (4.24)]. The optimal which control u evolves is from determined (4.42). Spatial discretization is necessary to convert the system of partial differential equations to a set of given ordinary differential equations. Walsh (1971) and more recently Houstis et al.(1978) have reviewed the finite difference and finite element methods for solving multidimensional partial differential equations. Both finite difference and finite element 66 SYSTEM TABLE 5.1 EQUAT IONS I. FOR STATE VARIABLES (4.14) dX 1 / dt = u ax2/14t = 2 (4-15) u (4.16) ;X3 / t = - a 1 .X3 exp[-E/(X 1+X 2+1)] t = - a 2 .X4 exp[- 'X 4 / i. FOR COSTATE 3P1 / t = - (1-P 3 ).al.X VARIABLES 3 expr-E/(X1 +X2 +1)]1 P4 .a 2 .I 4 .I. exp- expf-E/(I1+X 2+1)]. E/(X1+X2+1)].E/(XI +X 2 2+1) t = P4.a2 E/(X1 +X2+1 )2 + _ 2P2/ ay2 (4.51) FOR THE CONTROL VARIABLE u = Umax if a.nd u ==umin i dy = /u S SE / u dy = (P 1 -P 2 ) dy < P1-P2 (4.49) (4.50) E/(X1+X2+1 )] exp[- + (4.48) =- (1-P3).a1 exp[-E/(X1+X 2+1)] P3/at PD4/ 3. t = - (1-P 3 ).al. E/(Xl+12+1)2 E/(X 1+X2 +)2 P4 .a 2 .X4 expr-, E/(X1+X2+,). aP2/ p (4.17) BE/(Xl+X2+1)1 ) dy > 0 (4.42) 67 TABLE 5.2: INITIAL, TERMINAL AND BOUNDARY CONDITIONS FOR SYSTEM EQUATIONS (A) Initial Conditions: At t = 0, X1 = X10 (4.18a) X 2 = -X10 (4.18b) 1 (4.18c) X4 = 1 4 (4.18d) X 33 = (B) Terminal Conditions: At t = tf P11 = 0 (4.44a) P2 (4.44b) = 0 P3= 0 (4.44c) (4.45) P4 p = 1 V' X4 dy C1 (4.24) (C) Boundary Conditions: At y = 0, ( DX 2 /y) ( = 0 P 2 /Dy) = 0 (4.19a) (4.47) At y = 1, X2 = 0 (4.20) = 0 (4.46) 68 methods were tried in this work. In the finite difference method, the spatial domain was divided into ten equally spaced intervals for the spatially distributed variables X2 (y,t) P2(y,t), while for the rest, five were used. For the finite Galerkin equally spaced element intervals using approach into bilinear basis functions, the spatial domain was divided four intervals of .4, width .2 .3, .1 and respectively. fourth Integration in the time domain was done using a the Runge-Kutta-Gill (RKG) method. Details of and RKG order algorithm are given by Finlayson (1980). Brown (1979) gives a detailed discussion on the Galerkin finite element method with use of kinds of the different basis functions. While this code would be highly accurate two or more spatial dimensions, it offered negligible tage in terms of accuracy for the case of the one It slab geometry. the requires algebraic equations at each time and step considerable amount of computer time. The code, on the other hand, was easier to time for computation and was more thus finite formulate, accurate advan- dimensional a to solution for with of system consumes a difference took a less higher number of meshes than those used for finite elements. Hence in the subsequent analysis, the finite code was developed for difference retained. A fifth order quadrature formula was averaging the nutrient and micro organism concentrations over 69 the entire volume. By this formula, I N f dy = h (5 Af( Here N is equal to six for five equally spaced intervals, each of width h=.2 . The f(Yi) are the values of the function f at each of the nodes. The coefficients A i were evaluated to be: A1 = A6 = .329861 A 2 = A 5 = 1.302083 A3 = A4 .868056 accuracy This quadrature scheme provides significantly higher than the lower order formulas such as Simpson's or rule trapezoidal rule, especially when steep gradients or the exponen- tially decreasing functions are involved. The general strategy for computation is as follows. reduced system of canonical partial differential equations is to a set of 53 ordinary differential equations ( one X1; for six each for X 3,X 4,P 1,P 3 and P4; eleven each for X 2 and P2 We begin by assuming a single switching point at The ) where t=t1 the control u changes from u=umax for t(t1 to u=umin for t>t1 . Equations (4.14) to (4.17) for the state variables X1 X4 are integrated in the forward direction beginning through at t=O. Iterations are performed to correct the value of the switching time t, so that a desired final concentration of the micro 70 This organisms is achieved at t=tf. state equations are always is possible independent of since the the costate equations, while the control u which affects the behaviour the state equations, is in turn governed by the of costate variables, namely equation (4.42). Next, the costate variables P1 through P4 are integrated backwards in time, beginning with t=tf. Note that the costate variable P4 is unspecified at constraint on t=tf, which is compensated for by the integral X4 , namely equation (4.24). Figure S(P1-P2) 5.1 shows the plots of the integral dy) as a function of time, for the unknown , where P4 (tf)= analysis, the time axis is different . (Here and rescaled term in with I (I values the of subsequent respect to the process time tf.) We find that for large positive values of the integral is always negative. As we decrease = , , we observe a singular point (i.e. a point at which the integral I becomes zero) at t=.9 for =5 and at t=.75 for =2.5. For very small positive values, as well as for any negative values of v, the integral is observed to be always positive. One can infer from these results that it is possible to make the term I equal zero at a desired value of t1 by suitably choosing Thus the general scheme for solving the system tions outlined in table 5.1 would be to integrate equations to in the forward direction and switching point t1 to obtain the desired iterate reduction to . of equa- the state for in the micro 71 1iC) LI Z z 0 -10 -20 .0 · .2 .8 .6 .4 1.0 TIME F-ig. 5.1 Basis for Obtaining Switching a Single Point 72 organisms The costate at the end of the process. are equations then integrated backwards in time to iterate for the value for the unknown . which yields the integral I positive equal to zero at t=t1 and negative solution sent a complete to for t<t1 . t>tl, This will repre- equations canonical the of with boundary conditions described in table 5.2 cut The above approach is a short very well partly because of the behaviour unique costate variables described in section 4.3 . A way of solving a two problem point boundary value works which method, of very the general is the shooting technique. Keller (1972) describes this method for system of two ordinary differential two additional (ODE's) equations ODE's need be integrated on condition for the second variable. This unknown the at be few hundred ordinary differential equations to least solved is convergence a at computer each iteration, would require a tremendous amount of the be technique search in the present case. However, it would result in a initial could approach extended to an n-dimensional Newton-Raphson time and storage, and again where in time if one uses Newton-Raphson algorithm to iterate a not always the distributed parameters model with several states, though much theoretical guaranteed. This is a major limitation work has been done in this area. The of computational outlined earlier in this section, appears to be way of solving this problem. an strategy efficient 73 Figure 5.2 shows a plot of the term I for the case fixed process time of 24 minutes, values as noted in the diagram. with The the other switching required reduction the micro organisms. It is interesting to note a parameter is after approximately 5 minutes to achieve a 1/10 of that for case of no constraint on the final concentration of the in the micro organisms, P4 (tf)=O, which makes P4 uniformly zero everywhere. The results for this case have been shown in figure 5.3, which illustrates that the integral I should always be positive t(tf, or u should always be minimum for this case. This indeed make sense, retention with no if one were constraint on to maximize the final for would the nutrient micro organism concentration. Once we are convinced that the single switching represent a possible solution to the given set we need not be concerned further with the of equations, costate equations. This observation is based on the postulate that the on the final concentration of micro does condition organisms, uniquely determines the switching point tl, and that there would always be some value of for which the integral term I would have unique singular point at t=t1 . This reduces the work to solving only the state equations in a computational the forward direction. There are two other possibilities. possible to have multiple switchings Firstly, in it bang-bang may be control. 74 .2 _.J : ( 0. w z -.2 0. .25 -5 -75 1. TIME F i- - 5. Condit ion for Optimal Switching 75 1.5 1.C -J z 0 0. .25 .50 TIME .75 1.0 Fig. 5.3 Condition for Switching for the case of no Constraint on the Final Concentration of Micro Organisms 76 This is discussed in detail in the subsequent sections. Generally second possibility is that of a singular arc. any optimal control problem, it is difficult very The to prove the mathematically the existence of a singular arc, except in case of a few, simple, lumped models (Sage, 1968). However this problem, the presence of singular arcs seems to contain unlikely, since the governing equations linear, exponentially decreasing terms so it difficult to achieve conditions whereby the integral continue to remain zero for a finite time be interval is I in very several that in non very would (Tenney, 1982). Figure 5.4 shows the optimal control policy point switching with no constraint on the retort The switching occurs approximately process, and the maximum surface halfway temperature X 1 =.12 (119 degree C). The dimensionless and for temperature. through the corresponds to the parameters used are: a1 = 3.6 x 1011 , a 2 = 6.956 x 10 2 8 E = 25 , = 2.5 , Tin = 350 K Umax = (dX1 /dt)max = .25 , min = (dX1 /dt)min =-.25 Process time = 30 minutes L = 4 cm , C = .133 cm 2 /min. single dimensional 77 U 0o Li nHL Ur Um 0. -50 .25 .75 1-0 TIME FIG. 5-4 BANG- BANG CONTROL : SINGLE SWITCHING 78 It is interesting to observe from figure 5.5 how increase in switching time from t=O to t=30 minutes the gradually reduces the average micro organism concentration at the end of the process. While a ten log cycle reduction can by switching about halfway, only an be additional achieved log cycle reduction would occur if the heating continues till the end of the process. Clearly, for the fixed process time of 30 minutes in this case the ten log cycle reduction The temperature distribution and the seems best micro organism and log cycle After the switching occurs, the temperature in the outer nutrient concentrations corresponding to ten suited. the reduction are shown in figure 5.6 . portions of the can begins to fall rapidly, while the contents at the center of the can are still at This leads to the attainment of a a higher desired were organism destruction in less time than temperature. degree the of micro can heated uniformly throughout to the retort temperature. Such a control policy is in agreement with the high short temperature time process, followed by a rapid cooling of the products. In our model, the maximum rates of heating and cooling are considered to be independent parameters, and the optimal solution is found once these limits are known. However, the nature of bang-bang control, one should really expect increase in nutrient retention with an increase in the of u. from an limits Figure 5.7 illustrates this behaviour where as much as 79 4 ^ rlu U) z (5 0 0 10 () (D -1 > 10 O0 .25 .50 TIME .75 Fig. 5.5 Final Concent ration of Micro Organisms as a Fun cltion of Switching Time 1.0 80 C .- z m r~E IU L iE z 0LQ? o Li u I O -- LLJ w u n z O ( , c n Do 3l nitv 3d i13i 1_ - Ln Ln LLi W *n Li Ein 1o d rO 0 0n IC lN31 l nN V4SINv90 0831v 81 .16 A I. 75 .14 Li Z .5 (Z LU I z .12 z LO .25 LL .10 0. 0 ·25 .5 .75 1.0 MAX. RATE OF CHANGE OF TEMPE RATURE Fig.5-7 Effect of Variation in the Limits of the Control u 82 50 percent increase in nutrient retention may be one increases the rate of heating deg C/min) to about .75 (10.5 and cooling deg observed from C/min). .225 (3.15 Correspondingly, the switching time is seen to decrease with a rise in u. that the marginal changes in the optimal switching time decrease rapidly with as retention increasing Note and limits the umax and umin' If we increase the process time, there is scope for further reduction in micro organisms, hence the switching time decreases as shown in figure 5.8, other parameters unchanged. For example, a process time of 35 be if chosen the energy cost is one remaining minutes of the should important considerations. Different nutrients in the food have different activation energies. For those with lower activation energies, the reaction rate constant is also low, so that the net effect that of a higher nutrient retention. This is shown 5.9 for the case of ten log cycle in reduction is figure of micro organisms. In many cases, because of the may observe a lower and temperature, in which case an physical upper the constraint, bound control on the policy Umax, umin or zero depending on takes when discrete the retort should modified as outlined in section 4.4 . This leads to of dual switching, where the control one the be case values, constraint is 83 14 w I- 12 z I L 10 U0 8 30 45 PROCESS TIME F g. 5.8 60 min. Variation7 in Process Time 84 .. .1 Li cr .1 Z Z LL C vC 10 20 ACTIVATION 30 50 40 ENERGY E kcal/mol Fig. 5.9 Nutrient Degradation at Different Activation Energies 85 reached. Two typical cases are shown in figures 5.10 and 5.11, where the upper and lower constraints are respectively. Clearly, such a constraint on the perature represents a suboptimal case so that in effect retort lower tem- nutrient concentrations will be predicted by our kinetic models. To summarize, we have presented in this section a general computational scheme to solve canonical equations for the case of single point bang-bang control. There are several indepen- dent parameters affecting the nutrient retention, such as process time tf, the maximum rates of change max and umin, activation energies of the micro organisms and finally temperature. Effects of each the of of these temperature nutrients constraints on and the parameters the retort has studied in our computer simulation. Next, we wish to analytical solutions to the state equations. the been consider Complete closed form solutions are impossible. However, once the nature of the control u is known, the diffusion solved for the case of multiple equation switching points are known or can be assumed and if (4.15) the iterated can be switching upon. This follows in the next section. 5.2 ANALYTICAL SOLUTION In this section a general solution is developed heat conduction equation (4.15), when the control is for the subject 86 Li : Ld LJ w1 H1 J0 z L) 0 Fig. 5.10 .25 .5 TIME .75 1. Optimal Control for Upper Constraint on7 the Retort Temperatu re 87 1 W L1 W J 0 z nT O. .25 .5 TIME .75 1. with Lower Constraint on the Retort Fig. 5.11 Optimal Control Temperature 88 to multiple switchings. The single point switching or the dual point switching (when there is a constraint on the temperature) will be special cases of this solution. We retort begin with the equation, DX 2 / t =v2X2 - u (4.15) having boundary conditions _.7 X 2 = 0 on the planes of symmetry (4.19) and X 2 = 0 on the surface. (4.20) The corresponding eigenvalue problem will be V2V n = n 2V n (5.2) n with boundary conditions _.V Vn = 0 on the planes of symmetry (5.3) and Vn = 0 on the surface. (5.4) Vn(Y) is an orthogonal function representing a solution to the differential equation (5.2) with associated tions (5.3) and (5.4). (See, for example, boundary Myers, condi- 1971.) general solution to equation (4.15) can be written as The 89 00 X 2 (y,t) = Z <X 2 ,Vn>.Vn(y) (5.5) nl- where <a,b> denotes an inner product over the entire spatial domain v, and is defined by <a,b> = §a.b dv (5.6) For the one dimensional Vn(y) = slab geometry, we have cos( ny) (5.7) and Xn = (2n+1) T/2 (5.8) Multiplying equation (4.15) by Vn and integrating over the volume of the container, one obtains ( X2/)t)Vn dv = 72X2Vn dv - uVn dv SX2.V2Vn dv - Vu.Vn dv i V (5.9) V or I(~X 2 /$t)Vn dv = which follows on integration by parts twice. With the (5.10) aid of equation (5.2), equation (5.10) can be reduced to the form d(X2 ,Vn>/dt = _-n 2 < 2 ,Vn> - u< ,Vn> This equation has a solution (5.11) 90 <X 2, Vn > = exp(-\ n 2 t) [<X2'Vn>t=O - (1,Vn> ~ exp(An2t ') u(t') (5.12) dt'] O where, for the control policies considered here, the side is readily evaluated. Thus, on the right-hand the kth switching time Sexp( n2 t') u(t') integral if tk is we have dt' = (l/ n2 ) E - ukl[exp(nftk) 0 exp(n 2 tkl] + uN [exp( 2 t) - exp(An2 tN)]I in where uk_i is the control value interval effect (tk l,tk). Also, to = 0 and N = suitable rearrangement of the terms (5.13) during max{kl in the the tk<tl. time With summation in (5.13), the final solution (5.5) can be written as oo X 2 (Y,t) = 7 - (1/n 2)uN Vn(Y)<1,Vn>[X2 0 exp(-\n2 t) 2t)uo.exp(-An 1 (Uk-Ukl).exp(-1n[t-tk]) (5.14) The dimensionless retort temperature will be x (t) = X10 + L k- i Uk-li(tk - tkl) The nutrient and the expressed as micro + uN.(t - tN) organism concentrations (5.15) can be 91 t X 3 (y,t) = exp Sal a- exp[-E/(X1+X 2+1)] dtl (5.16) and X 4 (Y,t) = exp {- Ia 2 exp[- E/(Xi+X2+1)] dtj (517) respectively. The analytical solution eliminates the need discretization which is required canonical equations numerically. to solve However, essential to iterate on the combination which would organisms. give desired final of for the it spatial system will now switching concentration of be points of micro Substantial numerical integration in time is still required if the analytical solution to the used. In the present analysis, we have heat opted equation for numerical solutions, rather than an analytical plus is complete numerical solution approach. Analytical solutions would be of great help if any instability were to arise in the numerical method due to a very fine discretization in spatial co-ordinates. 5.3 MULTIPLE SWITCHINGS: NUMERICAL SOLUTIONS It was shown in the earlier part of this chapter bang-bang control with a single switching point is a that possible solution to our problem of optimizing nutrient retention since it satisfies the necessary conditions presented in tables and 5.2 . It is worth stressing again that there can be 5.1 more 92 than ne solution to the given set of canonical equations, and this each such case would be a suboptimal control. The aim of section is to explore numerically the possibility of multiple The cases. switchings which could constitute such suboptimal optimal solution would be the one which maximizes the nutrient concentration at the end of the process. One of the possibilities, not discussed earlier, is of a zero point switching. In other words, if is any upper hout the process and if there is Such reached. a once is solution on the this con- mathematically for possible, if one recalls from figure 5.1 that values of V, the integral I is uniformly throug- constraint retort temperature, u will be equal to zero, straint u=umax that very high It seems negative. possible to achieve the optimal nutrient concentration at the end of the process with such a control policy; however this is not physically acceptable, since the containers must be cooled down to the room temperature eventually, and hence least at one switching point is necessary. Next, consider the possibility of two switching points. For this purpose, several combinations of two switching points could be considered. The values of different parameters are the same as those used for the case of single switching point. We begin with the knowledge that for the single switching, t= .48 as shown in figure 5.4 . Obviously, multiple switching, the first switching for point the must case be of less 93 than this value of t, if the process with u=umax. Figure 5.12 illustrates different combinations of two switch- ing points. As the first switching begins point t (independent variable) moves to the left and away from t 1 =.48, switching point is seen to approach closer the to second the first switching point, primarily determined by the condition that desired level of micro organism concentration is the end of the process. The corresponding achieved retort a at temperature profiles for the five combinations are also shown. Figure 5.13 shows the results of the computer simulation for one case, where the two switching points are .468 and .664 respectively. The final nutrient concentration for tions of two switching points is shown different in figure combina5.14. One observes that as the first switching point moves away from the optimal value for single switching, the nutrient concentration continues to decrease. Clearly, the case of single switching (t1 =.48), which is the upper most point in figure 5.14, gives the optimal nutrient retention. Any sterilization process should begin with heating should end with cooling, hence it is desirable to have at the beginning of the process and u=umin at the end process. This will limit us to consideration of only and u=umax of the an odd number of switching points. Consider, for example, the case of three switching points, shown in figure 5.15. Here the second 94 4 3 2 w Ld L 1 D 0- 1 Li p- C U --2 0 ·25 .5 -75 1 TIME Fig. 5.12 Bang-Bang Control Switching Points : Two 95 1 -25 10 Z 05 z4 6500 O 1 05° 0 U 10 1. zL -75) .5 z 0. .5 0. DISTANCE FROM CENTER Fig. 5.13 Results for Two Swi tching (Case 1) 1. Points 96 12 ,12 .11 z z .11 z LL .1C .10 .420 .432 FIRST .444 .45(5 SWITCHING .468 .480 TIME Fig. 5.14 Optimal Concentration Using Two Switching Poi ntss 97 LL c- Li4 Li 5 4 3 0 25 5 75 10 TIME Fig. 5.15 Optimal Control : Three Switching Points 98 switching point t2 is held fixed at .48 and the first ing point t1 is varied from .468 to .42. The third point is determined by the integral constraint micro organism concentration. The resulting concentrations for different cases 5.16. are Again, we find that the nutrient switch- switching on the final presented final nutrient in concentration figure continues to increase as the conditions for optimal single switching are approached. The two plots, obtained with different time step sizes, also illustrate the need for very small time steps for a more accurate integration in time and a correct iteration on switching points. when the second Another interesting behaviour switching earlier, is changed (figure time, which 5.17). This deviation from the optimal conditions and hence the resulting nutrient for was is observed held constant again leads single concentrations to a switching, are Also, the three point switchings seem to give higher lower. nutrient retention than two point switchings in the same range. Thus, there appears to be point switching is the strong optimal evidence control that strategy various combinations of single, double and triple studied. It can be inferred from the principle of single among the switchings optimality (Athans and Falb, 1966) that, for the case of multiple switchings with more than three switching points, trajectory can always be broken down into time the control intervals which two or three switchings occur. One could achieve a in more 99 .12 I I I time step th=.0 0 0 4 z w .120 / Mth .004 z .115 I - I I .444 .432 _ .42 .456 TIME S WI TCHING FIRST Fig.5.16 Optimal Concentration - _ .120 Li z wE with Three - t 2 :A8, z i A4T c)2 .+ Points Switching *125 I .468 J / / .115 / 7- // /l .11C , - *432 - .468 FIRST SWLTCHING TIME Fig.5.17 Variation in Second Switching Timet2 (3 Switching Points ) 100 optimal nutrient concentration if these two or three switching point controls were replaced by an point single equivalent switching which achieves the same final concentration of micro can argument further, one switching will always be conclude better that single the any than this Extending organisms at the end of these time intervals. point combination of multiple switchings. the single switching control, with or without a constraint on the retort Henceforth, we will concentrate only temperature. The next step is to compare the single switching model with the earlier on results published work in food sterilization engineering. This will be the next chapter, using a cylindrical can geometry. of the research done in CHAPTER 6 TWO DIMENSIONAL MODELS 6.1 PRELIMINARIES The optimal control developed policy the in earlier puree chapters is applied to the sterilization of canned-pork a in this chapter. For this purpose, we consider reduction can geometry and a five log cycle organism. zation process is aimed-at maximizing the in pork puree. The first thermal degradation of and nutrients average The Sterili- thiamine retention kinetics irreversible order, the Stearothermophilus concentration of micro organisms. Bacillus is assumed to be the main food spoilage in cylindrical destruction of for micro organisms have been described by dN/dt = -(2.303 N/Dref). exp[(T-Tref)/(z/2.303)] as in the work of Teixeira et (2.3) and al.(1975) Saguy and Karel(1979). The pertinent kinetic data and process parameters listed in table 6.1 . As a starting point, it is worthwhile nutrient retention attainable at a temperature certain inside the can. This is shown in 101 to estimate constant figure the uniform 6.1. It 102 TABLE 6.1: PROCESS DATA 1. Can dimensions: 307x409 (55/16 inches diameter, 73/16 inches length) 2. Can name: A/2 3. Product: Pork puree 4. Thermal conductivity of pork puree: .09226 cm2 /min. 5. Reference 6. D value Tref = 121.1 degree temperature: for thiamine at the reference C (250 temperature: F) D re f 178.6 min. 7. z value 8. D for thiamine: value for the z = 25.56 degree micro organism C at the reference temperature: De min. ref f = 4 m. 9. z value 10. Desired for the micro reduction organism: of micro z = 10 degree C organisms: 5 log cycles. C1=10-5 For can specifications, please refer to Lopez, A. (1975). i.e. 103 1i z 0 Z LJ LU cr 5 z LU zD 0 100 120 TEMPERATURE Fig .6.1 140 °C Maximum Nutrient Retention at Constant Uniform Temperature 104 gives-an upper bound on the nutrient retention if the retort temperature is resistance, the actual known. maximum Obviously, possible due to maximum conduction thiamine retention will be less than that predicted by the above criterion. Teixeira et al. (1975) and Saguy and Karel (1979) have investigated optimal temperature profiles for sterilization of pork puree with the process parameters as outlined 6.1. They have assumed a constant process time of 89 Teixeira and co-workers observed that, perature profiles, the ramp gives among the in minutes. different maximum table tem- nutrient retention. Their results are shown in figure 6.2. 2 2 2 W i 19 4 I- w on ) ThE (MINUTES) Results from investigationof linear ramp functions,showing percentthiamineorenrtion ersusindividual ramnp policies. Fig. 6.2 Results of Teixeira et al.(1975) 105 It is interesting to note that the distributed for principle which requires bang-bang control leads to the same control policy optimization, are there when maximum active no The application of the constraints on the retort temperature. minimum principle to the distributed parameters model explains why one must observe the optimum with the ramp function. While the results of Teixeira et al. and Saguy and are interesting, the exact of values nutrient maximum the Karel retention should be viewed with caution, since the assumptions the made by these researchers concerning sterilization during Teixeira cooling process are very different. For example, al. assumed that the of destruction organisms micro et is complete when the retort temperature drops below 107 degree C, although the nutrient degradation should below this temperature; they did not account degradation in their calculations. to occur this added continue Saguy and for Karel, on the other hand, have aimed at maximizing thiamine retention at the end of the heating cycle during the sterilization (which is assumed to be complete when the retort process temperature drops below the inlet hot-fill temperature, 71 degree C). They have neglected the nutrient degradation during the cycle, but have accounted for the destruction of micro isms till the end of cooling. cooling organ- 106 These discrepancies in the manner in which and nutrient degradation are considered sterilization during the cooling process, lead us to make some more justifying assumptions and to define the terms more precisely, as follows. 1. The destruction of micro dation is assumed to occur organisms by a and the thiamine first order, degra- irreversible kinetics as described by equation (2.3), which will be valid at least in the temperature range from Tmax to Tmin. Tmax can be taken as the maximum retort temperature (130-140 degree C), and Tmin as the inlet hot fill temperature which in this is 71 degree C. It is assumed that there is no further case degra- dation or micro organism destruction below Tmin. 2. The sterilization process time, or briefly, the process time is defined as the time for thermal sterilization, at the end of which the retort temperature, or the temperature at the surface of water temperature, which the more can, precisely, equals in our case is taken the cooling to be 25 degree C. This change to cooling water temperature may be gradual and continous in time, or may be a step change. 3. Once the retort temperature equals the cooling temperature, the additional time required for the water temperature at each point inside the can to drop below Tmin (71 degree C), will be termed the cooling time. Normally, during cooling, the 107 from retort-temperature drops temperature rather rapidly, so to Tmin that the cooling cooling the defined the cooling process to begin when the retort drops below Tmin. However, time we will use zation process time and the cooling time. note that some sterilization as well as tempera- the definition. The total process time will comprise the It is important retort temperature drops from Tmin to cooling water ture , and during the cooling time. '7 &.j (1,1) (0,0) Fig. 6.3 Spatial Domain -l -. (1.0) - to degradation nutrient - __- - - first sterili- will be observed towards the end of the process time when (01 ) as One could have instead defined above will always be positive. ture water V -0 the tempera- 108 For the two dimensional can geometry, the conduction equation (4.15) can be written as X2 /at = 2X / 2 y2 + (1/y)( where a is the ratio of the + a.a 2 X 2/ az 2 2X2 / ;y) diameter to the length cylindrical container. The spatial domain is shown 6.3 . in It is divided into a 10x10 mesh for numerical by the finite difference method. u (6.1) - The resulting 194 of a figure solution ordinary differential equations are integrated in time using the fourth order Runge-Kutta-Gill algorithm as before. 6.2 RESULTS From our definition of the process time, it follows that this parameter will not be an independent quantity, but be determined by the fact that the bang-bang type with u (the rate of change control of is of temperature) discrete values of umax, umin or zero. Figure taking the shows one 6.4 such optimal control policy where uma x and umin are be equal to 3.87 degree C/min. and - 3.87 respectively. The initial retort temperature is to fall after 68 minutes, and is below taken C/min. 105.4 degree C(265 temperature the to degree C(222-F) and it rises to a maximum of 129.6 degree whereupon it is held constant. The retort will inlet ), begins hot fill 109 3 J 0acrZ 0 U - 3. 0 QZ w I- x 0 W m Hi 0J a: I r viVl . I111111 Fig. 6.4 Optimal Control for of Pork Puree Sterilization 110 temperature of 71 degree C at the end of 83 minutes. process time tf by our definition is 95 minutes. The ers need be held in cooling water for another The contain- 32 minutes before the sterilization is complete, i.e. the temperature at each point inside the can is below Tmin Figure 6.5 shows the radial distribution at the center-plane (i.e. z=O) at different of temperature There times. direction negligible variation in temperature in the vertical with steep boundary layers at the upper and surfaces. lower inside Figure 6.6 shows how the nutrient concentration varies the can at the end of the process time tf=95 is minutes and at the end of total sterilization time of 127 minutes. There is a central significant drop in the nutrient concentration in the region of the can during the cooling process. The volume average concentration of nutrient is seen to decrease by about one percent during cooling. The average nutrient retention at the end of total sterilization is 43.6 percent for this case. at the end of the process time The volume average temperature tf is about 76 degree C. One of the limitations of the models is that the sterilization is assumed when the mass average temperature drops below this is not correct, since the central continue to be sterilization. hotter, causing be complete Tmin. However, to portions further lumped of the degradation can and 111 U 0 Ld Q LJ 0- .25 0. RADIAL .50 .75 1. DISTANCE FROM CENTER Fig.6.5 Temperature Distribution at the Center-Plane 112 .70 .65 Z LU D zp- -60 Z 0. RADIAL .6 -4 .2 DISTANCE FROM CENTER .Fig. 6.6 Nutrient Degradation Cool ing Process during 113 It is also not strictly correct to nutrient at the end of a fixed retention the compare average time interval of before heating cycle, since the higher the retort temperature the the completion of the heating cycle, the higher to retention will be. Refer, for example, summarizes the results of dimensional models. the of magnitude 1.93 degree which for simulation computer in the second case, if we As shown umax and umin to be nutrient 6.2, table by reported retention of 45.3 percent as C/min., 1979.) However, during the rest of the the as nutrient highest and Saguy two choose nutrient retention at the end of 89 minutes can be as high 46.21 percent. (Compare this value with the the Karel, sterilization, total the average nutrient concentration reduces by 2.6 percent, so that the final nutrient concentration is 43.6 percent. Figure 6.7 shows the variation in the process time tf and the the total time for sterilization as one increases tude of umax and cooling, we Umin. move With closer to higher rates step changes temperature, as a result of which the considerably. total time decrease average temperature is cooling. This in leads with to increased a rates the other of and retort and the hand, the time process the heating inside the can at the end of higher time tf, On of magni- the process heating and slight increase in cooling time. 114 o -· W Dt: U c C,4 o coN. - r-H- '.,o ,r4o 4.)Luo o 0 C'4 0 zoc, toU") -,4 .. 0 o0 O > H , -.4 4 4)0 + 0 3 a 4- 0 .4 4J ,,i m,-N r- 0 0% r N o LA r- LA r- o o% Lu" L V L N4 % co r- Ln I 0 U 4 I d L E4rC 0 ,I i cO uO n) m w 44 t .I V -, r t C 4 V v 4l + 4 4C I N N m O · Ln -4 -, 4 o rl 0 -4 Zr %4 -4 . .4 ' . · 04-'40 L -IW N4 · ,4 04 0,.) Q 0-i-,- i - ') · .r-H· , 0. - N * ' 0;* * -0 f-. LA e4 co e4 ., - n ,- 0 .4"4J a act HQ). °, 46 H 0 .l o (N , ( 00 LA oo( 4i r .N O~ ~ ~ 0@H-H N O .~ I. I N U E- r)'· U1,.-(N 0 z E N~ En t l 3O4c .r1 rl0 - 4l4i -4) to 0 u r ~ = · to n I . co CO % a 0 Eco . '0 -, Nl (V NN . 0 4. CO H O ,-.o CO O .N N 00 O fo m '0 D Lo 0 4) 4) 00 ' -,O o co C · 0r4 , C r4 0 - *,-, H 0 03 0 0 4) 0 *4 0 0 U Ea 4 6 H L co o 0) PA 0 4 115 U 90< c E 80< LJ w a. 2 70 Li ~- w (9 60 3 1 RATE 7 5 OF TEMPERATURE CHANGE °C/min. Fig. 6.7 Variation in Maximum Heating and Cooling Rates of LU 116 It is interesting to note that the inlet hot-fill perature for the cans is one of the major variables the nutrient retention. Figure 6.8 that shows nutrient retention can increase by up to hot-fill temperature is increased to 50 100 affecting the average percent degree tem- as C. the Another major advantage in using higher inlet temperature for foods is that the sterilization time is reduced significantly, as shown in figure 6.9. This observation supports strongly the idea that the food contents should be as hot as possible just prior to canning, and the should sterilization be carried out immediately after the canning operation. The final optimal nutrient concentration varies in a very narrow range from 43.3 percent to 43.6 percent as one enhances the maximum rates of heating and cooling by a factor of four (table 6.2). This is shown in figure 6.10 for simulations done with two different time step sizes. Note maximum is observed at Umax= 4 degree flattens out as the time step is that C/min, (=.089 min). From the nature of the higher apparent which however, reduced by 1/5. in table 6.2 correspond to the dimensionless higher limits of u, a an time bang-bang nutrient The results step .001 control, with retention should be expected since the Hamiltonian would be maximized by operating at the highest values of illustrated for the one the This fact is well geometry with no con- control. dimensional straints on the retort temperature (figure 5.7). The reason 117 52 5( 0 0 z Li 47- - z 45 42.! 70 80 INLET 90 TEMPERATURE 100 C Fig. 6.8 Nutrient Retention at Dif f erent Inet Hot -FillI Temperatures 118 13 12 10 L- H 9( 7' 6" BC 70 90 80 INLET TEMPE RATURE 100 °C Fig. 6.9 Process Time and Total Sterilization Time at Different Inlet Hot-Fill Temperatures 119 44 0o a zw 43 frH z 42 2 4 6 RATE OF TEMPERATURE CHANGE Fig.610 Variation I °C/min in Limits of the Control U 120 for not observing this phenomenon in the present case, seems to be associated with the numerical scheme for integration time domain. With higher rates of heating and are large changes in temperatures in a very cooling, short which require very small time steps for accurate This explains the flattening of the nutrient in figure 6.10. We could expect a trient retention, if reduced by 1/10 or the marginal minimum 1/100. A step finer there time-span, integration. retention increase size curve in were spatial ter time and discretization prevented However, the excessive amount of storage required us from any further for such reduction nu- further would be another important way to improve the accuracy of numerical algorithm. in the compu- integrations, in the step has size in the time or spatial domain. Some important inferences can be drawn from the results illustrated in this section. These follow next along with the discussion on other related aspects of the thesis. 6.3 DISCUSSION AND CONCLUSION The mathematical developments in the preceding have shown that bang-bang control with single-point represents the optimal solution to the problem the nutrient retention. This is true for a applications where the system dynamics of general are chapters switching maximizing class governed by of a 121 diffusion equation with the control operating boundaries, and the cost function and all other on system tions being independent of the control. The aim of the was to apply the techniques of optimal control maximize the nutrient retention during thermal the equa- thesis theory to processing of canned foods. The results presented for one dimensional slab geometry will be applicable for the sterilization of cylindrical though numbers may vary. For example, it is easier to with a one dimensional model that all solutions with switchings represent only suboptimal cases, cans, show multiple and that the bang-bang control with a single switching point is the optimal solution. This is true when there are no constraints retort temperature, which is rarely found in one can show that constant rise and fall of retort temperature respectively, is the optimal solution heating when the practice. temperature during on Again with maximum and cooling there are con- straints on the retort temperature. This is one of the general practices followed in food process industry. Several researchers in the past have tried various kinds of functional forms and parametric optimization techniques for maximizing nutrient retention. (See for example, Teixeira al. (1969 a,b), Thijssen et al. (1978) and Ohlsson (1980 among others.) Most of this work has been et b,c) based trial-and-error optimization, there being no guarantee that on a 122 partieular type of functional form or the control policy will be the most optimal. Our approach, on the other hand, does not make a priori assumptions regarding the functional form of the retort temperature; it derived from the is rather maximum the principle necessary which conditions lead us to the optimal control. Finally, we have modified the model for a two dimensional cylindrical geometry and have applied it to the sterilization of pork puree, on which significant optimization has been done in the past. Our results compare favourably with values of nutrient retention (1975), consistent with the policy derived by them is reported fact a by that bang-bang the Teixeira the optimal control with highest et al. control single switching. The work of Saguy and Karel (1979) served as point for this research. They have shown how a Pontryagin's minimum principle can be applied to the problem of nutrient retention. However, many of their less rigorous than those made and here, starting maximizing assumptions this work were extends their analysis in a logical manner. One of the major contributions of our model lies in its distributed nature, so that one can compute temperatures and nutrient and micro organism accurately. concentrations More importantly, there are no errors resulting from the lumping of these quantities. This has also helped us to rigorously define 123 the process time and the cooling time, and to account tly for the sterilization during the was not done in the work of others. sterilization of pork puree, we cooling correc- process, With the example have illustrated which of the that our optimal control policy will lead to maximum nutrient retention under given conditions. Last but not the least, it is interesting inlet hot-fill temperature of the canned foods to note is one that of the major variables, affecting the nutrient retention. Hence every effort should be made to keep this temperature as high as possible prior to sterilization. 6.4 SUGGESTIONS FOR FUTURE WORK The last point mentioned in the preceding section gives us some clues to the direction of future research. There is a significant scope for further improving the nutrient retention and at the same time, reducing the process time, by the on-line processes prior to sterilization, to modifying ensure high inlet hot-fill temperatures. This research is likely to foster more in an industrial environment. In this thesis, we have concentrated on conduction heated foods. However, there are several kinds of semi-solid foods as well, wherein the convective currents set up inside the can will enhance the heat transfer. Normally, the industrial 124 sterilizers are designed to provide end-to-end, parallel axis so that or other type of rotation and of agitation convection can occur inside the can. While much work has been done in this area, no one has sound theoretical model for predicting cans, the yet experimental developed convective a heat transfer coefficients. A substantial portion of this thesis has been devoted the mathematical development of the distributed principle. One can find a vast scope for theoretical to minimum as well as applied research in this field. However, from our point of view, the techniques of optimal control are a means to an end, and not an end in themselves. APPENDIX I NOMENCLATURE Definition Symbol coefficients in the quadrature formula (i=1 to 6) C cost associated with the terminal state CM micro organism concentration, no./cm 3 CMo initial concentration of micro organisms, no./cm 3 CN nutrient concentration, g/cm3 CNo initial concentration of the nutrient, g/cm3 C1 final concentration of micro organisms D can diameter, cm D time required to reduce the reactant concentration by a factor of 10, min. Dr Dref value of D at a certain reference temperature, min. E dimensionless activation energy EM activation energy for destruction of micro organisms, kcal/mol EN activation energy for nutrient degradation, kcal/mol E activation energy, kcal/mol a 125 126 Definition Symbol H Hamiltonian I integral term J cost function K,K o reaction rate constant, min - 1 L characteristic length: slab thickness for one dimensional geometry, can height for two dimensional geometry L Lagrange multiplier vector N concentration, mol or no./cm 3 P costate variables vector Pi costate variable associated with the state variable Xi (i=l to 5) (P1- P2) dy Q vector expressing initial conditions R gas constant, kcal/mol.K R vector expressing terminal conditions T temperature, T iTin T T max min TR Tr,Tref T2 °C or K initial temperature of the can, °C or K maximum temperature for the retort, °C minimum temperature below which no further nutrient degradation or micro organisms destruction occurs. C retort temperature, C or K reference temperature, K temperature differential (T-TR), 0C or K 127 Definition Symbol V volume of the can Vn characteristic functions in the eigenfunction expansion X state variables vector X1 dimensionless retort temperature X2 dimensionless X3 dimensionless nutrient concentration X4 dimensionless micro organisms concentration X5 penalty for violating the constraint Xlo initial value of X 1 X2o initial value of X 2 a diameter to length ratio for the can a1 reaction rate constant for nutrient degradation a2 reaction rate constant for destruction of micro organisms b term in the kinetic expression(2.2), equivalent of activation energy bl lower constraint on the retort temperature b2 upper constraint on the retort temperature f right-hand side of the system equations in a vector form f(x) function of x equivalent of T 2 128 Definition Symbol f(xi ) value of the function g time variable cost term h element width in spatial discretization t time tI time, min. t c cooling time tf process time ti ith switching time t initial time 1 o at x. u control variable u control variable vector u' control variable, Uma x ax Umin' min 1 0 C/min. maximum value of the control minimum value of the control spatial co-ordinate Y y z spatial co-ordinate vector dimensional spatial co-ordinate, cm increase in temperature necessary to reduce the D value by a factor of 10 z spatial co-ordinate along the vertical axis of the can 129 Definition Symbol C< thermal conductivity, cm 2/min. t ratio of the activation energies EM/EN e unit outward normal vector v> value (O integrand of the cost function X An of P 4 at t=tf heaviside step function eigenvalue in the solution by eigenfunction expansion REFERENCES Ammerman, G.R. 1957. The effect of equal lethal heat treatments at various times and temperatures upon selected food components. PhD thesis. Purdue Univ., W. Lafayette, Indiana. Athans, M. and Falb, P.L. 1966. "Optimal Control: An Introduction to the 'meory anda its Applications", McGraw Hill Book Co.,New York. Athans, M. and Canon, M.D. 1964. Fuel optimal singular control of a non linear second order system. Preprints 1964 Joint Autom. Control Conf. pp 245-255. Ball, C.O. 1923. Thermal process time for canned Nat. Res. Council 7, part I, # 37. food. Ball, C.O. and Olson, F.C.W. 1957. "Sterilization Technology", McGraw Hill, Inc. New York. Bellman, R., Glicksberg, I., and Gross, 0. 1956. Bull. in Food On the bang-bang control problem. Quart. Appl. Math. 14:11. Berkowitz, L.D. 1961. Variational methods in problems control and programming. J. Math. Anal. & Appl. 3:145. of See also, 1962. On control problems with variables. J. Math. Anal. and Appl. 5:488. bounded state Bigelow, W.P., Bohart, G.S., Richardson, A.C., and Ball, 1920. Heat penetration in processing canned foods. Canners Association Bull. 16L. Brown, R.A. 1979. Finite element methods for the of capillary surfaces. J. Comp. Phys. 33:217. C.O. Nat. calculation Bryson, A.E. Jr. and Ho Y. 1975. "Applied Optimal Control: Optimization, Estimation and Control", Revised Printing, Hemisphere Publishing Co. Washington D.C. Butkovskii, A.G. and Lerner, A.Y. 1960. The optimal control of systems with distributed parameters. Autom. and Remote Control, 21:13 See also, Butkovskii, A.G. 1962. The maximum principle systems with distributed parameters. Autom. Control, 22:1156. and 130 for and optimum Remote 131 1979. Distributed parameter control systems. Avomatica i Telemekhanika, 10:16. Castillo, P.F., Barreiro, J.A., and Salas, G.R. 1980. Predic- tion of nutrient retention in thermally processed heat conduction food packaged in retortable pouches. J. Food Sci. 45:1513. Chang, H. and Reilly P.J. 1976. Optimal catalyst profiles for a two step reaction sequence in a packed bed. Chem. Eng. Sci. 31:413. Charm, S.E. 1978. "Fundamentals of edition. Food Engineering", pp 189-231. AVI Pub. Co., Westport, third Conn. Chen, S. and Jeng, W.F. 1978. Minimum end time policies for a Sci. batchwise radical chain polymerization. Chem. Eng. 33:735. Denn, M.M., Gray, R.D., and Ferron, J.R. 1966. Optimization a class of distributed parameter systems. Ind. Fundamentals, 5:59. & Eng. in Chem. see also, Denn, M.M. 1969. "Optimization by Variational Methods", McGraw Hill Book Co., New York. Desoer, C.A. 1959. The bang-bang servo problem treated variational techniques. Information & Control, 2:333. by Downes, T. and Hayakawa, K.I. 1977. A procedure for estimating the retention of components of thermally conductive processed foods. Lebensm-Wiss.u-Technol. 10:256. Fairbrother, R.G. 1982. American Can Co., Barrington Technical Center, Illinois. Personal Communication. Farhadpour, F.A. and Gibilaro L.G. 1981. On the optimal unsteady state operation of a continous stirred tank reactor. Chem. Eng. Sci. 36:143. See also, 1975. Continous unsteady state operation of a stirred tank reactor. Chem. Eng. Sci. 30:997. in Analysis Finlayson, B.A. 1980. "Nonlinear Engineering', pp 28-32, McGraw Hill, Inc. New York. Gould, L. 1969. "Chemical Pub. Co., Reading, Mass. Gruyaert, F. and Crowe, C.M. Process 1974. Control". On the Chemical Addison-Wesley optimization of 132 distributed parameter systems with boundary control: A counter example for the maximum principle. AIChE J. 20:1124. Hahn, D.R., Fan, L.T., and Hwang, C.L. 1971. Optimal start-up control of a jacketed tubular reactor. AIChE J. 17:1394. Harris, R.S. and Karmas, E. 1975. "Nutritional Evaluation Food Processing", AVI Pub. Co., Westport, Conn. Hayakawa, K.I. and Ball, O.C. temperatures 1971. Theoretical formulas of for various in cans of solid food and for evaluating heat processes. J. Food Sci. 36:306. Hayakawa, K.I. 1978. A critical review of mathematical procedures for determining proper heat sterilization processes. Food Technol. 32(3):59. Hill, C.G.Jr. and Grieger-Block R.A. 1980. Kinetic data: Generation, interpretation and use. Food Technol. 34(2):56. Houstis, E.N., Lynch, R.E., Rice, J.R., and Papatheodorou T.S. 1978. Evaluation of numerical methods for elliptic partial differential equations. J. Comp. Phys. 27:323. Jen, Y., Manson, J.E., Stumbo, C.R., and Zahradnik, J.W. 1971. A procedure for estimating sterilization of and quality factor degradation in thermally processed foods. J. Food Sci. 36:692. Johnson, C.D. 1965. Singular solutions for problems in optimal and control. From "Advances in Control Systems: Theory Applications", edited by C.T. Leondes, vol. II, Academic Press, Inc. New York. Jones, A.G. 1974. Optimal operation crystallizer. Chem. Eng. Sci. 29:1075. of a cooling batch Joslyn, M.A. and Heid, L. 1963. "Food Processing Vol.II" , AVI Pub. Co., Westport, Conn. for Methods "Numerical 1972. Keller, H.B. New York. Boundary-Value Problems", Blaisdell, Operations, Two-Point 1975. "Food Process Leniger, H.A. and Beverloo, W.A. Engineering", D. Reidel Pub. Co., Dordrecht, Holland / Boston, Mass. Lenz, M.K. and Lund, D.B. 1977 a. The Lethality-Fourier number method: Experimental varification of a model for calculating average quality factor retention in conduction heating canned foods. J. Food Sci. 42:997. 133 , 1977 b. The Lethality-Fourier number method: Experimental verification of a model for calculating temperature profiles and lethality in conduction heating canned foods. J. Food Sci. 42:990. See also, , 1980. Experimental procedures for determining destruction kinetics of food components. Food Technol. 34(2):51. Loncin, M. and Merson, R.L. 1979. "Food Engineering: Principles and Selected Applications", pp 329-365. Academic Press, Inc. New York. Lopez, A. 1975. "A Complete Course in Canning", tenth edition. pp 114. The Canning Trade, Inc. Baltimore, Maryland. Lund, D.B., Bruin, S.Jr., and Lazar, M.E. 1972. Internal temperature distribution during individual quick blanching. Food Technol. 37:167. Lund, D.B. 1975. Effects of heat processing on nutrients. From "Nutritional Evaluation of Food Processing", pp 205-239. Edited by R.S. Harris and E. Karmas. AVI Pub. Co., Westport, Conn. , 1977. Design of thermal processes nutrient retention. Food Technol. 31(2):71. for maximizing and also, , 1982. Influence of processing on nutrients in foods. J. Food Protection. 45:367. Mulley, E.A., Stumbo, C.R., and Hunting, W.M. 1975 a. Kinetics of thiamine degradation by heat. A new method for studying reaction rates in model systems and food products at high temperatures. J. Food Sci. 40:985. Kinetics of thiamine degradation by heat. Effects form of the vitamin on its rate of destruction. J. 1975 of pH Food b. and Sci. 40:989. Thiamine: A chemical index of the sterilization thermal processing. J. Food Sci. 40:993. 1975 c. efficacy of Murase, A., Roberts, H.L., and Converse, A.O. 1970. Optimal thermal design of an autothermal ammonia synthesis reactor. Ind. Eng. Chem. Process Des. Develop. 9:503. 134 Myers, G.E. 1971. "Analytical Methods in Conduction Transfer, pp 74-142, McGraw Hill Book Co. New York. Nishida, N., Liu, Y.A., Lapidus L., and Hiratsuka Heat S. 1976. An singular e:ffective computational algorithm for suboptimal and/or bang-bang control. AIChE J. 22:505. Ogunye, A.F. and Ray, W.H. 1971. Optimal control policies tubular reactors experiencing catalytic decay. AIChE for J. 17(1):43. Ohlsson, T. 1980 a. Temperature dependence of sensory quality changes during thermal processing. J. Food Sci. 45:836. , 1980 b. Optimal sterilization flat containers. J. Food Sci. 45:848. temperatures for , 1980 c. Optimal sterilization temperatures for sensory quality in cylindrical containers. J. Food Sci. 45:1517. Patwardhan, V.S. and Sadana, A. 1982. The optimum policy for a deactivating immobilized enzyme reactor. Chem. Eng. Comm. 15:169. Pontryagin, L.S., Boltyanskii, V., Gamkrelidze, temperature fixed bed R., 192. "The Mathematical Theory of Mishchenko,E. Processes", Interscience Publishers, Inc. New York. Ray, W.H. and Szekely, J. 1973. "Process 258-259, Wiley Interscience, New York. Optimization", Ray, W.H. 1981. "Advanced Process Control", pp 84-118, Hill Book Co., New York. Robinson, E.R. 1970. The optimal control of an batch distillation column. Chem. Eng. Sci. 25:921. Sage, A.P. 1968. "Optimum Systems Inc., Englewood Cliffs, N.J. and Optimal Control", pp McGraw industrial Prentice Hall, Saguy, I. and Karel, M. 1979. Optimal retort temperature profile in optimizing thiamine retention in conduction type heating of canned foods. J. Food Sci. 44:1485. Smith, T. and Tung, M.A. 1982. Comparison of formula methods for calculating thermal process lethality. J. Food Sci. 47:626. Stumbo, C.R.193 New procedures for evaluating thermal 135 processes for foods in cylindrical containers. 7:309. Food , 1973. "Thermobacteriology in Food second edition. Academic Press, Inc. New York. Technol. Processing", Optimal temperature Szepe, S. and Levenspiel, O. 1968. deactivation, policies for reactors subject to catalyst I-batch reactor. Chem. Eng. Sci. 23:881. Teixeira, A.A., Dixon, J.R., Zahradnik, J.W., and Zinsmeister, G.E. 1969 b. Computer optimization of nutrient retention in the thermal processing of conduction heating foods. Food Technol. 23:845. 1969 a. Computer determination of spore survival distributions in thermally processed conduction heating foods. Food Technol. 23:352. Teixeira, A.A., Zinsmeister, G.E., and Zahradnik, J.W. 1975. Computer simulation of variable retort control and container geometry as a possible means of improving thiamine retention in thermally processed foods. J. Food Sci. 40:656. Tenney, R.R. 1982. Department of Electrical Cambridge, Mass. Personal Communication. Eng., MIT, Thijssen, H.A.C., Kerkhof, P.J.A.M., and Liefkens, A.A.A. 1978. Short-cut method for the calculation of sterilization conditions yielding optimal quality retention for conduction type heating of packed foods. J. Food Sci. 43:1096. Thijssen, H.A.C. and Kochen, L. 1980. Calculation of optimum sterilization conditions for packed conduction type foods. J. Food Sci. 45:1267. Thompson, D.R. 1982. The challange in predicting nutrient changes during food proceesing. Food Technol. 36(2):97. Timbers, G.E. 1971. Some aspects of quality degradation during the processing and storage of canned foods. Ph.D. thesis, Rutgers, The State Univ., New Brunswick, N.J. Walsh, J. 1971. Finite difference and finite element of approximation. Proceedings of Royal Society of Series A.323:155. Yamane, T., Kume, T., Sada, E., and Takamatsu, simple optimization technique for fed-batch methods London. T. 1977. culture. A J. 136 Ferment. Technol. 55(6):587.