0045-7949/8X s3.00 + 0.00 8 1988 Per~tttoo Press plc compurers & stnicwas Vol. 29, No. 2, pp. 34>3JO, 1988 Printed ia Great Britain. THE ADJOINT METHOD FOR DETERMINING INFLUENCE LINES A. D. BELEGUNDU Mechanical Engineering Department, The Pennsylvania State University, University Park, PA 16802, U.S.A. (Received 25 August 1986) Abstract-A method for computing influence lines is presented. Influence lines are widely used in structural design to determine worst-case locations of live loads. The method is a result of giving physical interpretation to certain adjoint variables used in the field of design sensitivity analysis. The method is general, and can be easily implemented in finite element programs. Computational aspects and example problems are discussed. I. INTRODUCIION Recently, certain design-specifically adjoint variables where p] is an (n x n) structural stiffness matrix and F is an (n x 1) vector of applied loads. From eqn (I), we have z as a function of the load vector F. That is, we have used in optimal in design sensitivity analysiswere given a physical interpretation [I]. The adjoint variables were shown to be the influence coefficients associated with a given function. In the present paper, the above interpretation is shown to result in a new and general approach for determining influence lines, which can be easily integrated into existing finite element codes. Influence lines are used by structural analysts to determine worst-case locations of live loads [2-4]. Influence Iines show the influence on a certain function of a unit load as it travels across the structure. The function can be anything that varies as the load moves across the span, such as moment or shear at a given section in the girder or beam, or bar force in a particular truss member, or deflection of a given point on a structure. The ordinate of the influence line, at a given point, is referred to as an influence coefficient, and represents the value of the function due to a unit load acting at that point. The adjoint variable method for determining influence lines is derived in the next section. Example problems are given in Sec. 3. In the first of the example problems, the adjoint method is compared to the Muller-Breslau approach existing in the literature. A summary is given in the last section. z=[i(l-‘F, (2) from which dz/dF=[I(]-‘. (3) We can write g s g(W)). Now dg/dF = dg(z)/dz dz/dF. Using eqn (3) in (4), and noting symmetric matrix, we have fK1dgJdFT = dg(z)/dzT. that (4) [K] is a (5) Letting 1= dg/dFT, which is an (n x 1) vector, we can write eqn (5) as 2. DERIVATiON OF THE ADJOINT METHOD p]lz = dg(z)/dzT. Consider a function g(z) where g can be the shear force or moment at a beam section, or a bar force in a particular truss member, or a deflection of a given point on a structure. The (n x 1) vector z is a vector of nodal ~spla~ments obtained from the finite element equations Wlz= F, (6) Now consider the case when g is a linear function of z. This is true in linear structural analysis when g is a moment or shear, bar force, or a deflection. Since z is a linear function of F, it also follows that g is a linear function of F. C’onsequently, we have the equation (1) 345 g(F+u)=g(F)+dg/dF.u, (7) A. 346 D. BELEGUNDU where g =g(z(F)) =g(F), and u is a change in F. If the structure is initially unstressed, we have g = 0 when F=O. Letting F=O, u=(O ,..., l,..., O)T represent a unit load along the ith degree of freedom of the structure, and noting that li = dg/dF,, eqn (7) yields g(u) = li. (8) That is, li is the value of the function g due to a unit load acting along the ith degree of freedom. By definition [2], 1 is the influence coefficient vector associated with the function g. Therefore, it can be seen that solution of eqn (6) yields the influence coefficients vector 1. Noting the resemblance between eqn (6) and the equilibrium equations in eqn (1), we can think of Iz as a displacement vector associated with the load vector dg/dz. This important observation leads to the following conclusion. In the finite element method, element shape functions are used to obtain analytical expressions for displacements at interior points in the structure from the nodal displacement vector z. Likewise, element shape functions can be used to obtain analytical expressions for the influence coefficients within the elements from the nodal values of A. These expressions then can be used to generate the influence lines (or curves). It should be noted that when g is a linear function of z, the vector of dg/dz which forms the right-handside of eqn (6) is a constant vector, which does not depend on z. The analytical expression of g(z) is readily provided by the finite element method. Consequently, calculation of dg/dz is routinely handled in computer programs. Details related to element shape functions and calculation of dg/dz are given the next section. with the adjoint variable vector I determined from the adjoint equations in eqn (6). This method of computing derivatives has its origins in optimal control theory [S]. In [1], a physical interpretation was given to the adjoint variable vector rZ.The algorithm presented above and example problems in the next section demonstrate that the adjoint method based on eqn (6) is a powerful method for determining influence lines. 3. EXAMPLE Four example problems are presented. the first is a four-bar truss, where calculations are done by hand to illustrate the adjoint method. The problem is also solved using the Muller-Breslau method. The remaining three problems are solved using a computer program. Four -bar truss Consider the four-member statically indeterminate truss shown in Fig. 1. It is necessary to obtain the influence coefficients for the vertical reaction at the roller support at node 2. In this problem, EA = 1 for each member, where E is Young’s modulus and A is cross-sectional area. Using the stiffness method of analysis, the equilibrium equations Kz = F in (1) can be written as where Zi, Fi, i = 1, 2, 3, are indicated in Fig. 1. Using the finite element method, the reaction at node 2 can be expressed as R = -A Algorithm The steps involved in the adjoint method determining influence lines are as follows. PROBLEMS (0.96 z, + 2.50 z~). (10) for Step 1. Compute dg(z)/dz. Step 2. Solve for 1 from eqn (6). Step 3. Obtain expressions for influence coefficients within the elements using element shape functions. Then construct the influence lines. The present author uses the name ‘adjoint method for the new method presented because eqn (6) is used in optimal design to evaluate sensitivity coefficients or derivatives of functions with respect to design parameters. More specifically, let b = (k x 1) be a vector of design variables, and g(z, b) be a function with z implicity dependent on b through the equilibrium equations [K(b)]z = F(b). Then, the derivative dg/db is calculated from The influence coefficients for this problem are simply the values of R due to unit loads along each degree of freedom. The influence coefficients will be derived using the adjoint equations of this paper, and by another method taken from the literature. F3rZ3 T 4 f 40 1 t-------i Fig. 1. Four-member truss. 347 The adjoint method for determining influence lines Adjoint method. From eqn (IO), we have dR - = -& dz which, when substituted (0.96,0,2.50), ment method, for any general function example, consider (11) into eqn (6), yields resulting in the influence coefficients being given by A = -{0.237,0, l.O)r. -2, (14) where R, and R, are the reactions at two different supports. In the adjoint method, dg/dz = dR, fdz + dR,/dz is evaluated in the usual manner, and rZ is obtained from eqn (6). Use of the Muller-Breslau principle for eqn (14) is more involved and requires special logic in the computer program. Alternatively, use of superposition requires solution of two righthand-side vectors, as opposed to only one in eqn (6). (12) It should be noted, of course, that calculating 1 from eqn (6) for large problems is routinely handled in computer programs. Muller-Breslau principle. According to this principle [2,6], the roller support is removed and replaced by a load P (Fig. 2). The influence coefficients, say K,, are then given by IC;= g(z)= R,(4 + 4 (~1, g(z). For (13) Pratt truss Consider the Pratt truss in Fig. 3(a), which is a problem obtained from [2]. It is necessary to obtain the influence coefficients associated with the bar force in member 3-9. This and subsequent problems are solved using a computer program. The calculation of dg/dz for this problem can be illustrated by considering a typical two-dimensional truss element as shown in Fig. 4. Using the finite element approach, element forces in the local reference frame are given by where Si are the displacements along z, , z,, zj due to load P, and A is the displa~ment along P Fig. 2). Straightforward calculations result in A=95P 6, = 22.5 P 6, = 0 p’ = F’]z’, (15) where k’ is the element stiffness matrix and z’ is the element nodal displacements in the local (x’ - y’) reference. Equation (15) can be written in expanded form as r 1 0 -1 01 f-q 6, = 95 P, which, in view of eqn (13) yields IC = - f0.237,0,1 .O)r, which is the same as that obtained earlier. From this example, one can see that implementing the Muller-Breslau principle requires special modifications of the computer program, since removal of the roller support requires modifications of the boundary conditions. Furthermore, the principle is not as efficient and as easily applicable as the adjoint method within the context of the finite ele- P Fig. 2. Muller-Breslau method. Evidently, p ; = pi = 0 and p ; = -p ; . The bar force is simply the component pi in the element force vector, which follows the convention that positive means tension. We thus have P; = k;UW. (17) where k; is the third row of [k’]. The vector ze = (z,, 22, Z3r zJT is the element displacement in global reference as shown in Fig. 4, and [fl] is a transformation matrix, with z’ = [/?]z’. From eqn (17), we obtain where [asjaz] is a (4 x n) matrix consisting of 0s and 1s. The vector ap;/az obtained from eqn (18) now forms the right-hand-side of eqn (6), which yields the influence coefficient vector 1. The vector I is given in Table 1. It should be noted that the calculations in eqn (18) are routinely handled in computer programs. A. D. BELEGUNJXJ 348 (b) 0.239 Fig. 3. (a) Pratt truss. (b) Influence line for bar force in member 3-9 in Pratt truss. (c) Influence line for shear in Pratt frame--values to be multiplied by lo-‘. In Fig. 3(b), an influence line is constructured by considering only the influence coefficients associated with the y degrees of freedom at nodes 1-7 along the bottom chord. The ordinates of Fig. 3(b) represent values of the force in member 3-9 due to downward unit loads at nodes 1-7. example is converted into a frame by treating all joints as rigid. All members have a rectangular cross-section with width = 10 in. and depth = 100 in. Young’s modulus E = 30 x 106psi. The function g , for which influence coefficients are required is defined as Pratt frame g =f,lF,, The application of the adjoint method to frames is illustrated here. The Pratt truss of the previous (19) where f, is the maximum flexural shear stress in the cross-section of element 3-9 and F3 = 14,400 psi is the Fig. 4. A typical two~mensional truss element. The adjoint method for determining influence lines 349 Table 1. Influence coefficient vectors for example problems Pratt frame Pratt truss Node/ direction Value Node/ direction lx 2x 0.0 0.0 2x ITx 4y 0.167 0.0 0.333 0.0 -0.5 0.0 -0.333 0.0 -0.167 -0.222 0.167 -0.222 - 0.667 -0.222 -0.5 - 0.222 -0.333 -0.222 -0.167 IX 3yx & Syx 1:: 1iX lT9 20 30 48 58 68 7% 88 98 100 118 128 allowable shear stress. The function g in eqn (19) represents the ratio of actual to allowable shear stress in element 3-9. Since f, = 1.5 V/A, with V = shear force and A = cross-sectional area, we have dg/dz = 1.S(d V/dz)/(A . F,). (20) In eqn (20), dV/dz is computed in a manner similar to that described in the Pratt truss example earlier. Substitution of dgjdz into eqn (6) yields the influence coefficient vector I, which is given in Table 1. Also, the influence line constructed from the influence coefficients corresponding to the y degree of freedom of nodes l-7 is shown in Fig. 3(c). The influence curves are constructed using element shape functions, as explained in the next example. Beam Consider the beam shown in Fig. 5(a). It is necessary to determine the influence line associated with the moment at support b of this beam. A finite element model of this beam is shown in Fig. 5(b). It may be noted that a lesser number of elements will Beam Value ( x 10m5) 0.136 0.129 0.239 0.107 -0.084 0.002 -0.477 -0.004 -0.287 -0.0 -0.147 -0.055 0.152 -0.160 -0.223 -0.185 -0.565 -0.195 -0.297 -0.194 -0.146 0.181 E-03 1.167 E-03 -7.661 E-03 1.363 E-03 0.310 E-03 0.413 E-03 0.401 E-03 1.637 E-03 - 7.689 E-03 1.205 E-03 0.377 E-03 0.425 E-03 Node/ direction Value 2x 0.0 -0.550 -0.165 0.0 -1.110 0.0 0.0 0.495 -0.002 - 1,015 -0.044 -0.002 -0.683 0.168 -0.002 0.132 -0.002 0.332 0.012 -0.002 0.226 -0.044 -0.002 -0.036 -0.002 -0.120 -0.006 -0.002 -0.096 0.012 -0.002 0.018 ; 3x i; 4x e 5x ; 6x ; 7x e 8x Y e 9x ; 10x 0 11x Y 9 12x ; 13x 8 also yield the same solution, since exact shape functions are used in the finite element analysis of beams (as per elementary beam theory). Solution of eqn (6) yields the 1 vector, as given in Table 1. In order to draw the irdluence lines, however, it is necessary to know the values of 1 within the elements. This is obtained by using element shape functions as described below. A typical two-dimensional beam element has six degrees of freedom as indicated in Fig. 6. The value of 1 at an inte~ediate point x (Fig. 6), is given by 4x) = i N,(x)& (21) i-l where Ni are Hermite cubic shape functions as given in, for example, [7]. For element 1-2, we have 1, = 2, = L, = A4= 0 and, from Table 1, A5= -0.550, 1, = -0.165. In eqn (20), it should be noted that 1 (x) represents the value of the moment at support b due to a vertical load at point x. Consequently, for element l-2, we have n(~)=(3<~-2<~)1,+f-<*+,~)L&, (22) A. D. BELEGLJNDU b I=900 12 3 4567 8 &‘ d I= 1600 ~I=1200 9 IO II 12 13 tbl 7% ‘z&h’ (c) -0.12 -1.1 I -0.096 -1.015 Fig. 5. (a} Beam problem. (b) Finite element model of beam. (c) Influence line for moment at support b of beam. 4. SUMMARY Fig. 6. A typical two-dimensional beam element. where < =x/L, written as and L = 5’. Equation ;3&-LA,)+ f (22) can be ;-2L,+L&), (23) 0 which, upon substituting &, yields n(x)= -- REFERENCES numericai values for 1, and 2 A new method for computing influence lines based on eqn (6) is presented. Element shape functions can then be used to construct influence lines. The method is applicable to statically determinate or indeterminate structures alike, and is applicable to any problem which can be modeled using finite elements. In addition to structural engineering problems, mechanical design problems (such as determining influence lines for bearing reactions in cams modeled by finite elements) can also be handled. The adjoint method, as it is referred to herein, is especially suitable for implementation into finite element programs. The method can be implemented in a general manner, and its logic is not problem dependent. x3 30.86 + 480.76 ’ (24) where 0 G x G 5’. In general, a(x) for any element can be obtained from eqn (21), and is routinely handled in finite eIement codes. The complete influence line for the beam associated with the moment at support b is constructed in Fig. 5(c). It should be noted that use of eqn (21) can include axial and shear effects, depending on the choice of Ni. Inclusion of such effects by the moment ~st~bution method used in [2, p. 4961 is very difficult. 1. A. D. Belegundu, Interpreting adjoint equations in structural optimization. AXE, J. Struct. Engng 112, 1971-1976 (1986). 2. C. H. Norris and J. B. Wilbur, Elementary Structural Analysis. McGraw-Hill, New York (1960). 3. W. J. Lamach, Inruence Lines: for Statically Indeterminate Plane Structures. Macmillan, London (1964). 4. J. F. Fleming and A. P. Shah, Truss influence lines on a time sharing system. Comput. Struct. 4, 933-950 (1974). 5. A. D. Belegundu, Lagrangian approach to design sensitivity analysis. ASCE, J. Engng Mech. 111, 680-695 (1980). 6. T. M. Carlton, Principle of virtual work in relation to Muller-Breslau’s principle. Inc. J. Mech. Sci. 22, 523-525 (1980). 7. J. S. Przemieniecki, Theory of Matrix Structural Analysis. McGraw-Hill, New York (1968).