Journal of Journal Manufacturing of Manufacturing Systems Systems Vol. 25/No. Vol.225/No. 2 2006 2006 Identification of Dimensional Variation Patterns on Compliant Assemblies Jaime A. Camelio, Dept. of Mechanical Engineering and Engineering Mechanics, Michigan Technological University, Houghton, Michigan, USA Hyunjune Yim, Dept. of Mechanical Engineering, Hongik University, Seoul, South Korea Abstract Dimensional imperfection of the final body assembly may be traced back to any combination of the three types of faults: non-nominal shape or size of parts, non-nominal location or deteriorated condition of fixtures, and welding distortion faults. Because of its importance, the dimensional quality of automotive body has been pursued in the assembly plants of leading manufacturers, mainly by applying statistical process control (SPC) methods to the assembly measurement data. In general, assembly products are measured at the end of the assembly line because it is not economically viable to measure parts at each station and it is difficult to measure tooling misalignment. When the SPC gives a precursor for dimensional quality deterioration, the root cause of the situation needs to be identified and, if possible, eliminated. Most advancement in the root cause identification technology has been oriented toward fixture faults (Ceglarek and Shi 1995, 1996; Hu and Wu 1992; Liu and Hu 2005; Apley and Shi 1998; Chang and Gossard 1998; Camelio, Hu, and Yim 2005; Wang and Nagarkar 1999; Ding, Ceglarek, and Shi 2002). This is mainly because fixture faults have been found to be the major cause for defective automotive body assembly (Ceglarek and Shi 1995). Significant research has been conducted on fixture fault diagnosis, including the use of principal component analysis (PCA) and designated component analysis (DCA). PCA is a systematic methodology to diagnose fixture faults by extracting the principal components from the measured data on the subassembly (Hu and Wu 1992). Because of its entire dependence on the measurement data, the root cause of principal components extracted may not This paper presents a methodology to diagnose sources of dimensional variation for compliant parts from the measurement data of final assemblies. The method is developed for a single-station assembly process. The proposed diagnosis tool is based on applying a predictive variation propagation model to determine the part-to-part and tooling interaction in the assembly system. The variation propagation model allows identifying the impact of different faulty component patterns in the final assembly product. Using the predictive assembly fault patterns and the designated component analysis, the contribution of each fault in the total system variation may be identified. The methodology incorporates an optimal sensor placement algorithm to determine the key measurement points in the assembly. Two case studies were conducted to illustrate that the methodology is capable of identifying part variation patterns from assembly measurement data, even under significant levels of noise. Although the methodology is presented for a single assembly station, it can be extended to a multiple-station assembly scenario using a multistation variation propagation model. Keywords: Assembly System Diagnosis, Compliant Assembly, Variation Reduction 1. Introduction Quality improvement in automotive body assembly is one of the major tasks undertaken by the leading automobile manufacturers. Among various quality measures, the dimensional quality of the final body assembly, also called body in white, has been regarded as the most important characteristic because its deterioration affects many quality aspects of the vehicle—for example, noise and vibration characteristics such as wind noise, convenience of use such as the amount of effort required to open and close the door, and the aesthetic appearance. 65 Journal of Manufacturing Systems Vol. 25/No. 2 2006 lems on the assembly. The objective of the present work is to develop a method to detect part variation patterns in an assembly station from the measurements on the station subassembly output and not from the incoming components. The significance of this work lies in that an assembly may exhibit dimensional variations in the absence of faults in the fixture and welding process if the incoming parts have dimensional variation or non-nominal shapes, which may not be detected or may be misinterpreted by the methods of fixture fault diagnosis previously developed. It is understood that decoupling the effects of parts variation from those of fixture faults or those of welding faults is not trivial and could be a challenging research topic. In the present study, however, all fixtures and the welding processes are assumed to be perfect, leaving the decoupling problem as a future research topic. As in the case of fixture faults, detection of part dimensional variation may be considered as an inverse problem of the forward problem—the generation of assembly dimensional variation caused by the component variation fed into the assembly process. There exist a great number of models and analysis methods for this forward problem, in the categories called ‘variation simulation’ and ‘tolerance analysis.’ Most of the methods, available in the form of commercial software, assume rigid behavior of parts. In such methods, the assembly variation is determined by “adding” the variation of parts according to its geometric and kinematic relations (Greenwood and Chase 1988), regardless of the “adding” method used–worst-case analysis, root sum squares analysis, or Monte Carlo simulation. Noting the importance of the compliant behavior of the assembly components, Liu and Hu (1997) developed a mechanistic variation simulation method by integrating engineering structural models with statistical analysis. In this model, the deformation of parts during the clamping process and the subsequent springback of the assembly are analyzed by FEA. As a result, the assembly variation is expressed as linear function of part variation through the use of sensitivity matrices. The present work may be considered as an attempt to solve the inverse problem of the one modeled in Liu and Hu (1997). The solution technique for this inverse problem will be the least-squares method, facilitated by the optimal placement of sensors (Camelio, Hu, and Yim 2005). always have a physical meaning. To address this problem, the method of designated component analysis (DCA) was developed by Liu and Hu (2005). This approach first defines mutually orthogonal dimensional variation patterns (called diagnostic vectors) for a part, each of which is associated with one or more fixture faults. Then, actual measurement data are decomposed into the components of the diagnostic vectors, disclosing the major components and thus the associated fixture faults. DCA is a powerful method for simple problems, but it is not always obvious how to define physically meaningful, mutually orthogonal patterns for complex part geometries. More physics-based diagnostic vectors, which are only constrained to be linearly independent with each other, were derived by Ceglarek and Shi (1996). The diagnostic vectors are defined by simple geometric analyses of rigid body movement for each fixture fault. The same diagnostic vectors were used for diagnosis of multiple fixture faults using the least squares method by Apley and Shi (1998). This diagnosis approach was adopted by Camelio, Hu, and Yim (2005) for the fixture fault diagnosis in the assembly of complaint parts, where the diagnostic vectors corresponding to fixture faults were obtained by finite element analyses (FEA). Furthermore, Camelio, Hu, and Yim (2005) also presented a methodology to determine the optimal locations of sensors on the parts by applying the effective independence (EfI) method so that different diagnostic vectors can readily be distinguished from each other. In addition, Ding, Ceglarek, and Shi (2002) developed a method to identify fixture faults in multistage manufacturing processes by applying a mapping procedure that combines PCA and the pattern recognition approach to a state-space multistage model. In Ding, Ceglarek, and Shi’s approach, the parts were all assumed to be rigid and perfect, and fixture faults were the only cause of non-nominal product dimensions. Camelio, Hu, and Ceglarek (2003) incorporated the effects of part variation and weld gun variation into the state-space model of a multistation assembly system, yet they did not address the diagnosis problem. As mentioned earlier, fixture faults are not the only cause leading to the deteriorated dimensional quality of a product. Variations in the parts and in the welding process can also result in dimensional prob66 Journal of Manufacturing Systems Vol. 25/No. 2 2006 2. Methodology vector, often assumed to have a known covariance matrix Σ w = σ2w I , where I is the identity matrix. The problem studied in this paper is how to find the part variation vector, u, from the measurement vector, y. The following sections provide detailed descriptions of the steps taken by the diagnosis methodology. 2.1 Problem Statement An automotive body is an assembly consisting of hundreds of sheet metal parts. The product key dimensions, or key product characteristics, are typically measured at several intermediate assembly stations as well as at the end of the body assembly line. Each measurement system consists of a set of sensors that measures displacements at some points of the assembly or subassembly. The deviations of these dimensions with respect to their nominal values are caused by various sources of variation: part geometry errors, faults on fixtures, and/or errors in the joining process. As indicated in the introduction, the objective of this work is to develop a methodology to identify the variation patterns of the parts based on the data measured on the assemblies. It is assumed that the fixtures and the assembly process are both perfect, thus leaving the parts variation as the only source of variation. In an initial effort to investigate this issue, the present study considers a single assembly station where two or more compliant parts are assembled, as shown in Figure 1. The proposed methodology will ultimately be extended to a series of multiple assembly stations. In the single assembly station shown in Figure 1, the parts variation, denoted by u, yields the assembly variation, denoted by v. These variation vectors, u and v, represent deviations of specific points on the parts/components or assembly from their nominal positions in a determined direction. The assembly variation, v, is in general measured using a coordinate measurement machine (CMM) or an optical CMM. Noise from the measurement process or from the unmodeled characteristics is also considered in the model. The noise vector, w, is a random 2.2 Modeling of the Problem Based on the concept of geometric covariance (Camelio, Hu, and Marin 2004), it can be assumed that the deformation vector of the assembly components can be decomposed in the contribution of different deformation patterns. Therefore, the part variation vector u can be modeled as a linear combination of q patterns, b1, b2, …, bq, each of which may represent a potential variation pattern of the parts, as will be explained in section 2.3. Then, the deformation vector for the parts can be written as u = x1 ⋅ b1 + x2 ⋅ b2 + … + xq ⋅ bq = B ⋅ x where xi (i = 1, 2, …, q) are the coefficients of the linear combination and represent the weight of the each pattern in u. The dimensions of the vectors, bi and u, are both assumed to be p × 1; x vector dimensions are q × 1; and B dimensions are p × q. p corresponds to the number of input sources or key control characteristics (KCC). As a result of the part variation given by Eq. (1), the assembly dimensions will vary as shown in Figure 1. Several variation analysis models have been presented to describe the relationship between the parts variation and assembly variation. For the spotwelding process of compliant parts, as in the automotive body assembly, it is possible to predict the assembly variation from the components/parts variation (Liu and Hu 1997). This prediction is made us- Noise, w Components Variation u (1) Assembly Variation Assembly Process v Measurements Measurement System Figure 1 Propagation of Variations in an Assembly Process 67 y Journal of Manufacturing Systems Vol. 25/No. 2 2006 ing the sensitivity matrix defined by Liu and Hu’s model. The sensitivity matrix can be obtained for each assembly process by conducting finite element analyses and using the method of influence coefficients. These finite element analyses require information on the nominal geometry of the parts, the fixturing scheme of each part, and the locations of the welding spots. This information must be obtained from the existing assembly line or from the design drawings. The method of Liu and Hu (1997) may be expressed mathematically as ⎧ v1 ⎫ ⎡ s11 ⎪v ⎪ ⎢s ⎪ ⎪ ⎢ 21 v=⎨ 2⎬=⎢ ⎪#⎪ ⎢ # ⎪⎩vm ⎪⎭ ⎢ s ⎣ m1 s12 s22 # sm 2 " s1 p ⎤ ⎧ u1 ⎫ ⎥ " s2 p ⎥ ⎪⎪u2 ⎪⎪ ⎨ ⎬ = S⋅u % # ⎥⎪ # ⎪ ⎥ " smp ⎥⎦ ⎪⎩u p ⎪⎭ (4) that the measurement vector, y, has the same dimension, m, as the assembly variation vector, v. These m points of measurement must be carefully selected, as will be discussed in detail in section 2.4, and thus the sensitivity matrix, S, must be defined such that S ⭈ u yields v at those m selected points. The columns of the matrix D, often called the diagnostic vectors, may be interpreted as the effect of deformation pattern bi represented by each diagnostic vector, di = S ⭈ bi , on the assembly variation measurement. The components of the vector x (Eq. 3) may accordingly be interpreted as the weights of those components existing in the measurement y. From Eqs. (3) and (4), the problem of the present study may be interpreted as finding the weights of diagnostic vectors x from the assembly measurement data y given the diagnostic matrix D. In other words, the diagnostic problem is to estimate the amount of variation in the measured data that can be explained by each designated pattern or fault. A fault is defined as a source of variation instead of a shift in the mean. Therefore, if the variance of a pattern (xi) is significant, a fault will be identified. The covariance matrix of the designated components, xi, can be calculated as follows: (2) where the element of the sensitivity matrix, sij, measures the sensitivity of the assembly variation at the i-th point to the incoming part variation at the j-th point. This model is based on several assumptions: (1) deformation on parts in the assembly process is within the linear elastic range; (2) the material is homogeneous and isotropic; (3) fixtures and welding guns are rigid; (4) there is no thermal deformation due to welding; and (5) the stiffness matrix for non-nominal part shapes remains the same as that for nominal shapes. The validity of these assumptions is presented in Liu and Hu (1997). Note that the assembly variation vector, v, in Eq. (2) has dimensions of m × 1, generally different from those of the parts variation vector, u. The value of m here corresponds to the number of measurement points on the assembly, which will be discussed shortly. The measurement vector, y, in Figure 1 is simply modeled as the sum of the assembly variation and noise. That is, using Eqs. (1) and (2), y = v + w = S⋅u + w = S⋅B⋅x + w Σ Y = D ⋅ Σ X ⋅ DT + Σ w From Eq. (5), the covariance matrix of the measurements points, ⌺Y, can be seen as a linear stackup of the variation effect of the designated patterns, D ⭈ ⌺X ⭈ DT, and the nonmodeled effects variance, ⌺w. 2.3 Definition of Part Variation Patterns The part variation vectors, b1, b2, …, bq, in Eq. (1) are defined such that they correspond to patterns typically observed in the specific component fabrication process. For example, a deformation pattern for an assembly part/component can be a bending effect from the stamping process or a twisting effect from the material handling equipment. These patterns can be determined by experts or measured from available past data from suppliers. Figure 2 shows some examples of deformation patterns for a rectangular plate part that has been clamped along its left-hand edge. For complex assembly geometries, such patterns will be constructed using finite element simulations or obtained from previous components measurement data (Camelio, Hu, and Marin 2004). (3) Given that the matrices S and B are constant and known, Eq. (3) can be simplified as follows: D = S⋅B y = D⋅x + w (5) (4) where D, often called the diagnostic matrix, is defined as an m × q matrix. The matrix D represents the effect of the different deformation patterns of the components on the assembly. It should be noted from Eq. 68 Journal of Manufacturing Systems Vol. 25/No. 2 2006 (a) Bending about x-axis, b1 (b) Twisting about y-axis, b2 z y x (c) Bending about y-axis, b3 Figure 2 Examples of Part Variation Patterns (Camelio, Hu, and Marin 2004) the locations of the measurement points are very important. To facilitate the fault identification, m measurement points must be selected such that the diagnostic vectors, that is, the column vectors of D in Eq. (4), may easily be differentiated from each other. Camelio, Hu, and Yim (2005) proposed a methodology for optimal placement of sensors. This method is essentially an iterative numerical method that finds the most effective sensor locations by starting with a great number of candidate locations. The candidate locations are typically at the nodes of the finite element mesh of the assembly components. The method eliminates the least-effective sensor in each iteration step, based on the effective independence (EfI) criterion. This iteration proceeds until the desired number of sensors are left. During this iterative process, the number of rows of the sensitivity matrix, S, in Eq. (2) and thus that of the diagnostic matrix, D, in Eq. (3), are reduced from the number of all finite element nodes to the desired number of measurement points. The effectiveness of the sensor locations obtained by the procedure assures that they provide the greatest linear independency among the diagnostic vectors, each associated with a single fault of the parts. In this paper, the optimal sensor placement problem is solved using a genetic algorithm (GA) (Goldberg 1989). The objective function in this method is to maximize the determinant of the Fisher information Considering the patterns shown in Figure 2 and defining the sources of variation as three points along the edge in the x-direction (points along the right-hand edge of the plate, one at the midpoint and two at the ends), the matrix B will have the following form: B = [ b1 b2 ⎡1 1 1 ⎤ b3 ] = ⎢⎢1 0 0 ⎥⎥ ⎢⎣1 −1 1 ⎥⎦ (6) Each pattern vector, bi, is defined as the deviations from nominal at the sources of variation, which are often located at the welding points. Thus, the iT th pattern vector is written as bi = {bi1, bi2, …, bip} , where p is the total number of welding points over the parts, and bij is the deviation from the nominal at the j-th welding point for the i-th pattern. As will be discussed in section 3.1, the pattern vectors, bi, must be linearly independent of each other. It must be noted that the vectors bi may be also defined for deviations occurring in the in-plane directions, however, for simplification, in this example they are only defined as deviation normal to the surface plane on a reduced set of key control characteristic points. 2.4 Placement of Measurement Points Because the problem is to identify the part fault patterns from the measurement data on the assembly, 69 Journal of Manufacturing Systems Vol. 25/No. 2 2006 T matrix, Q = D D. Among the subsets consisting of m sensor locations, constructed from all candidate sensor locations (in this case, corresponding to the finite element nodes of the assembly), the one associated with the maximum value of determinant of Q is sought using the GA, and it is selected as the optimal set of sensor locations. The GA method has been found to be significantly faster than the iterative elimination method in Camelio, Hu, and Yim (2005). Off-Line Preparation Define Parts Variation Patterns, H Locate Sensors and Calculate Diagnostic Matrix, D On-Line Diagnosis 2.5 Diagnosis Methodology The entire diagnosis methodology for the problem defined and modeled above consists of two steps, as shown in Figure 3, the off-line preparation procedure, and the online diagnosis procedure. The off-line procedure mostly has been explained in the previous sections. That is, the part variation pattern matrix, B, is first established by assuming typical patterns of parts (see section 2.3). Then, the sensitivity matrix, S, is obtained by conducting finite element analyses and using the method of influence coefficients. In this step, the number of rows of S is the total number of the finite element nodes of the assembly. The final step in the off-line procedure is to construct the diagnostic matrix, D, as S ⭈ B and to reduce its number of rows on D to the desired number of measurements, m, by the optimal sensor placement method described in section 2.4. The diagnosis procedure is conducted online as shown in Figure 3. The measurement data are obtained online from the m measurement sensors, which provide y(j), j = 1, 2, …, N, for the j-th assembly. For each measured part, a measurement (j) vector, y , is obtained. The corresponding weight (j) vector, x , is estimated using the least-squares method, as ( xˆ ( j ) = DT D ) −1 Construct Sensitivity Matrix, S DT y ( j ) Calculate Assembly Patterns Significance (LSM) Measurement Data, y Calculate Patterns Correlation Identification of Multiple Part Faults Figure 3 Part Variation Diagnosis Methodology After the weight vector, x, is estimated using Eq. (7) for N assemblies, the variance of the contribution of each diagnostic vector (corresponding to each of the q part/component variation patterns) is calculated as ( 1 N ( j) ∑ xˆi − xˆ i N − 1 j =1 i = 1,2," , q σi2 = ), 2 (8) where xˆ i = ∑ j =1 xˆ i( j ) N . The matrix form of the covariance can be expressed as N ( ∑ xˆ = DT D ) −1 ( DT ∑ y D DDT ) −1 It must be noted that the covariance of the measurement data, ⌺y, includes the effect of the deformation patterns variation, ⌺x, and the effect of the noise variation, ⌺w. Therefore, to estimate the contribution of each deformation pattern, the noise effect should be removed from the estimation ∑ xˆ . The effect of the noise in the estimation of x is ∑ xˆ ( w) and corresponds to the effect on the measurement data if only noise is present. (7) (j) where xˆ ( j ) is the least-squares estimator of x . The estimator xˆ ( j ) differs from the real contribution because is has the noise effect included. This approach T requires the matrix D D to be invertible. In other T words, D D must be of full rank. This restriction will be further discussed in section 3.1. 70 Journal of Manufacturing Systems Vol. 25/No. 2 2006 ( ∑ xˆ ( ) = DT D w ) −1 ( DT ∑ w D DDT ) 3. Special Considerations −1 3.1 Diagnosability Diagnosability is defined here as the capability of the proposed diagnostic procedure to detect and isolate a specific pattern or multiple patterns of part variation. The diagnosability analysis investigates under which conditions the proposed methodology will be able to identify the faults. From Eq. (7), the estimate x̂ of the part deformation pattern contributions to the assembly variation T T cannot be obtained if D D is not invertible. D D will be invertible if the columns of D are linearly independent or, equivalently, if D is full rank (Strang 1988). Recalling that D = S ⭈ B [Eq. (4)] and given that the part patterns (columns of the matrix, B) are forced to be linearly independent as discussed in T section 2.3, D D will be invertible if the columns of S are linearly independent. Then, the diagnosability condition is given by The percent contribution of each pattern, Ci, to the total variation can be calculated individually as a proportion of the measurement data variation explained by each pattern. The total variation of the system is defined as the sum of the variance of all patterns. The effect of each pattern is calculated as the effect of the estimated variance of each significant pattern in the total variation of the significant patterns. Therefore, the contribution of each pattern is obtained as follows: ( C = trace ( Σ ) ×100%, −Σ ( )) diag Σ xˆ − Σ xˆ ( i w) w xˆ xˆ (9) i = 1,2," , q where Σ xˆ − Σ xˆ ( w ) denotes a matrix that includes only the significant patterns. The significant patterns may be identified using a hypothesis testing presented in section 3.2. The function diag is defined as the diagonal of a matrix. The trace function is defined as the sum of the variances of all significant components. If the i-th diagnostic vector (or equivalently the i-th assembly pattern) is found to be significant, the i-th part variation pattern, bi, is regarded as a significant source for the assembly variation. In addition, because multiple faults may result in significant correlations among the assembly variation patterns, the correlations between patterns may be calculated in Eq. (8). rank ( S ) = p where p is the number of columns in S and corresponds to the number of sources of variation. The rank function counts the number of independent columns in a matrix. It should be clear that this independence requirement means the linear independence of vectors, not their statistical independence. In fact, designated components are often correlated with each other as measurements are made. The correlation matrix is used in the diagnosis methodology. Furthermore, if D is not of full rank, the sensor placement procedure explained in section 2.4 also fails because the determinant of the Fisher information matrix becomes zero. Even in the cases where D is not of full rank, partial diagnosis can be done. A full-rank submatrix, Dind, may be constructed by selecting a linearly independent set of columns of D. The other columns of D constitute the complement submatrix, Dcind , then D = ⎡⎣ D ind Dcind ⎤⎦ . By definition, the columns (or assembly patterns) in the submatrix, Dcind , are linear combinations of the patterns in matrix Dind. The leastsquares estimation can then be performed only for the linearly independent patterns in Dind as N ηij = r r ∑ xˆ i( ) xˆ (j ) r =1 N ( ) ( ) r ∑ xˆ i( ) r =1 2 N r ∑ xˆ (j ) , 2 (11) (10) r =1 i, j = 1,2," , q The correlation among the patterns can be useful to identify the root cause in the presence of multiple faults occurring simultaneously. It should be noted that it is only meaningful to analyze the correlation among the significant patterns. However, statistical independence among the patterns is usually expected. ( x̂ ind = DTind ⋅ D ind ) −1 DTind y (12) where x̂ ind is a subvector of x consisting only of those components of x corresponding to Dind. The 71 Journal of Manufacturing Systems Vol. 25/No. 2 2006 significance of these independent assembly patterns is evaluated by Eqs. (8) and (9). In the latter case of limited diagnosability, the dependency among assembly patterns means that the diagnosis method cannot identify and isolate the difference among the dependent patterns. Patterns dependent on each other can be grouped into ⍀j, j = 1, 2, …. Patterns in a group, ⍀j, are dependent on each other but independent of patterns in other subgroups. If one or more of the patterns in a group is identified as significant, it means that any pattern in the group can be the source of variation. pothesis testing for the variance of a normal population, the test in Eq. (15) may be conducted to identify if the contribution of a designated pattern is statistically significant. The interpretation of the test is that for each deformation pattern i the variance of the pattern must be statistically greater that the variance of the noise w to diagnose a pattern. ( ( ) −1 = σ2w DT D ) ( DT ⎤ Σ y ⎡ DT D ⎦⎥ ⎣⎢ −1 ) −1 DT ⎤ ⎦⎥ w) w xi 2 w xi xi 2 2 i = 1,… , q (15) To evaluate the hypothesis, we will use the test statistic: χ20 = ( m − 1) Sx2 (σ ) (w) i 2 (16) xˆi 2 where Sx is the estimate of the variance of the deformation pattern i, and m is the number of observations in the data y. The null hypothesis H0 will be rejected if i χ20 > χ2α ,m −1 (17) If the null hypothesis is rejected for a specific deformation pattern j, then the pattern is a significant contributor of the system variation. 4. Case Studies The proposed methodology is illustrated with two examples. The first example considers the assembly of two rectangular sheet metal parts. This simple example is used to clarify each step of the methodology without including difficulties coming from complex geometries. A second example considers a realistic automotive example, that is, the assembly of a fender and its reinforcement. T (13) 4.1 Assembly of Two Metal Sheets The proposed diagnosis methodology is first exemplified with the assembly of two sheet metal parts, as shown in Figure 4. Each part has dimensions of 400 mm x 400 mm x 1 mm. The material is mild steel with Young’s modulus E = 207 GPa and Poisson’s ratio, = 0.3. Both parts are constrained along one edge parallel to the y-axis. The parts are joined together at three weld points (designated as W1, W2, and W3 in Figure 4) located along the mating edge. where Σ y = Σ w = σ2w I has been used. From Eq. (4), the covariance matrix for x̂ can be calculated as Σ xˆ = Σ x + Σ(xˆ 2 xi H1 : 3.2 Statistical Significance Under Noise Using Eq. (9), the contribution of each part variation pattern to the total variation of the assembly can be determined. However, it is necessary to examine the statistical significance of each fault identified. For example, even for the in-control case (i.e., when no faults are present), the noise in the measured data may generate large values for Ci. Therefore, a large value of Ci does not necessarily imply that the i-th pattern is a significant source of variation. The statistical significance of each fault may be determined using statistical hypothesis testing. The proposed approach assumes that the covariance matrix of w, ⌺w, is known and xi, i = 1, 2, …, q, (deformation patterns deviation) exhibit a normal distribution. Based on these assumptions, the covariance matrix for x, ∑ xˆ , can be estimated for in-control conditions (⌺Y = ⌺w). The noise contribution of the measuring system and the locating fixture is included in the term ⌺w. The covariance matrix of x̂ , w ∑(x̂ ) , solely due to noise can be estimated from Eq. (7) as follows: ∑ xˆ ( w ) = ⎡ DT D ⎣⎢ ( σ ) = ( σ( ) ) ( σ ) > ( σ( ) ) H0 : (14) Therefore, if there is a fault present the estimation of the contribution of the faults will be larger than the estimation of x solely due to noise. Using hy- 72 Journal of Manufacturing Systems Vol. 25/No. 2 2006 part variation patterns. The matrix, D, originally has dimen400 mm sions of 242 × 6. Using the W1 (12) 400 mm optimal sensor placement apW2 (17) proach described in section 2.4, W3 (2) Part 2 three optimal measurement points were selected as the Part 1 Z welding points. This means that X at these welding points the asY sembly variation patterns may be most easily differentiated Figure 4 Assembly Example of Two Rectangular Plates from each other. The measurement points selected are shown The procedure in Figure 3 is followed. Six part as Nodes 12, 17, and 2 in Figure 4. The resulting variation patterns, bi, are considered in this example. diagnostic matrix, D, of reduced dimensions 3 × 6 is Each part is assumed to exhibit three patterns (simiD = [ d1 d 2 d3 d 4 d 5 d6 ] lar to the patterns shown in Figure 2): bending along the y-axis, twisting along the x-axis, and bending 0.281 0.265 0.112 0.281 0.265 ⎤ ⎡ 0.112 ⎢ along the x-axis. Therefore, the part variation pat0 0 −0.019 ⎥⎥ = ⎢ 0.103 −0.019 0.103 terns for the two parts may be written as the out-of⎢⎣ 0.112 −0.281 0.265 0.112 −0.281 0.265 ⎥⎦ plane deviations of the parts from the nominal at the welding points. A similar approach can be considFrom matrix D, it can be seen that d1 = d4, d2 = d5, ered for in-plane rigid motion. The resulting part and d3 = d6. Recall that there is one-to-one correvariation pattern matrix, B, with dimensions of 6 × 6 spondence between the assembly patterns, di’s, and is found to be the part variation patterns, bi’s. Then, the dependence among di’s observed above means that the diagnosis methodology cannot distinguish between the B = [ b1 b2 b3 b 4 b5 b6 ] same type of part variation patterns of the two parts. ⎡0.574 0.707 0.707 ⎤ The reason for this is, of course, the symmetry of ⎢0.574 ⎥ ∅3×3 0 0 ⎢ ⎥ the problem, that is, identical part geometries, and ⎢0.574 −0.707 0.707 ⎥ identical part variation patterns for the two parts. In =⎢ ⎥ 0.574 0.707 0.707 general, if there were lack in complete symmetry, ⎢ ⎥ ⎢ ⎥ there would have been no diagnosability problem. ∅3×3 0.574 0 0 ⎢ ⎥ Following the method in section 3.1, the follow0.574 −0.707 0.707 ⎦ ⎣ ing groups may be defined for the dependent patterns: ⍀1 = {d1, d4}; ⍀2 = {d2, d5}; ⍀3 = {d3, d6}. where b1, b2, b3 represent the variation patterns for Selecting the first members of the three groups to one part, b4, b5, b6 for the other part, and ∅3×3 deform a submatrix of the diagnostic matrix gives notes a zero matrix having dimensions of n × m. Then, using the finite element method and the D ind = [ d1 d 2 d3 ] method of influence coefficients, the sensitivity ma⎡0.112 0.281 0.265 ⎤ trix, S, is obtained for this problem. The finite ele0 = ⎢⎢ 0.103 −0.019 ⎥⎥ ment mesh of each part has 121 nodes, and thus the sensitivity matrix has dimensions of 242 × 6. As ⎢⎣0.112 −0.281 0.265 ⎥⎦ mentioned previously, sensitivity matrix represents Allowing only one part to have variation, a simulathe response of the assembly to a unit deviation of tion study was performed. Cases of the three part variaparts at each welding point. tion patterns, b1, b2, b3, are exclusively simulated. For Next, using Eq. (4), the diagnostic matrix, D, is each case, the number of assemblies simulated was N computed. Recall that the columns of D represent = 100, and the faulty pattern was assumed to have a the assembly variation patterns corresponding to the 400 mm 73 Journal of Manufacturing Systems Vol. 25/No. 2 2006 M5 Fault Faults d1 d2 d3 M1 M2 M3 M4 No Fault 0% 20% 40% 60% 80% Noise to Signal Ratio Figure 5 Simulation Results for Rectangular Plate Assembly variance of 1 mm2. In addition, different levels of noise have been simulated, with a variance of a given percentage of the pattern variance. Figure 5 shows the results of the study. The presence of a part fault will be identified as 1, and a zero value means that the fault is not significant based in the hypothesis testing. Figure 5 shows the capability of the methodology to detect a part fault with respect to the level of noise. It can be seen that under the noise level over 7 percent, Pattern 1, d1, will have a missed alarm. The probability of a missed alarm is lower for Pattern 2, d2; that is, Pattern 2 is diagnosable for noise level up to 49 percent. This seems to imply that the second pattern produces stronger signal than the others, thus making it more discernible even under high level of noise. Similar results show that the probability of false alarm is low for same level of noise as the simulated in Figure 5. Figure 6 Finite Element Mesh of Fender-Reinforcement Assembly the y-direction) at Nodes 2769 (L3), 4598 (L4), and 472 (L5). For the reinforcement component, the locating scheme consists of a hole-pin (constraining displacements in the x-, y- and z-directions) at Node 10815 (L1), a slot-pin (constraining displacement in the y- and z-direction) at Node 11988 (L2), and a locating pad (constraining displacements in the ydirection) at Node 10805 (L3). The parts are joined together at four welding points that constrain Nodes 1001/11103 (W 1), 4530/12163(W 2), 1395/10178 (W3), and 2699/11628 (W4). Fixtures and welding locations are shown in Figure 7 for the fender and in Figure 8 for the reinforcement. In this study, only two component deformation patterns of the reinforcement are considered: bending and twisting. The two reinforcement variation patterns are defined, in terms of the out-of-plane (ydirection) deviations from the nominal at the welding locations, as 4.2 Assembly of Automotive Fender A more realistic case study was conducted by simulating a fender-reinforcement assembly. Figure 6 shows the finite element mesh for the assembly. The assembly consists of two parts: (1) a fender component, and (2) a reinforcement that is assembled to the top of the fender along its longitudinal direction. Each part was assumed to be located following the 3-2-1 locating scheme (Cai, Hu, and Yuan 1996). For the fender, the locating scheme consists of a hole (constraining displacements in the x- and z-directions) at Node 3098 (L1), a slot (constraining displacements in the z-direction) at Node 855 (L2), and three locating pads (constraining displacements in ⎡ ⎢ ⎢ B = [ b1 b2 ] = ⎢ ⎢ ⎢ ⎢⎣ ∅ 4×2 0.5 0.5 0.5 0.5 ⎤ 0 ⎥⎥ 0.707 ⎥ ⎥ −0.707 ⎥ 0 ⎥⎦ The matrix B includes the key control characteristics (sources of variation) for both parts, the fender 74 Journal of Manufacturing Systems Vol. 25/No. 2 2006 L1 W2 4530 W3 1395 3098 W4 2699 W1 1001 W1 1001 L5 L2 L1 472 855 10815 L4 L2 L3 4598 855 2769 W2 12163 W3 1395 W4 11628 L3 Figure 8 Fixture and Welding Locations for Reinforcement 2769 points reduces the dimensions of D to 5 × 2. The reduced matrix, D, is ⎡ ⎢ ⎢ D = [ d1 d 2 ] = ⎢ ⎢ ⎢ ⎢⎣ Figure 7 Fixture and Welding Locations for Fender and the reinforcement. However, this case study only includes failure modes on the reinforcement. Given that the reinforcement is more rigid than the fender itself, the effect of the fender can be neglected. Therefore, it should be noted that the patterns have zeros for the welding points in the fender component. Using the method of influence coefficients, the sensitivity matrix, S, is obtained. The input variables are the sources of variation defined at the welding points, Wi, i = 1, …, 4. The output variables are at the 7317 finite element nodes in the assembly. The sensitivity matrix has dimensions of 7313 × 8. Using Eq. (4), the diagnostic matrix, D, may be computed, which has dimensions of 7317 × 2. Among the potential 7317 locations for measurements, only five measurement points were selected using the optimal sensor placement methodology based on genetic algorithms described in section 2.4. The measurement points selected are Nodes 1612 (M1), 3013 (M2), 4962 (M3), 4966 (M4), and 5707 (M5), as shown in Figure 6. It must be noted that based on the genetic algorithm, the five selected sensors may not be an optimal solution and only be a good set of sensors. In addition, to identify two faults only two sensors are needed. Therefore, some of the measurement points are close together in areas were the deformation or the signal or the sensor will be larger. Accordingly, this selection of five measurement 0.5558 0.5633 0.8187 0.8147 0.3026 −0.2544 ⎤ −0.2602 ⎥⎥ 0.0764 ⎥ ⎥ 0.0799 ⎥ 0.3887 ⎥⎦ This reduced matrix, D, is of full rank. This means that the two assembly variation patterns, resulting from the two part variation patterns, are independent of each other. Both patterns are thus fully diagnosable with the five measurement points. The diagnosis methodology was evaluated similarly to the previous example. The methodology was capable of successfully diagnosing the two part variation patterns—bending and twisting with noise levels under 30% of the system variation. 5. Conclusions A methodology has been developed for the diagnosis of the variation contribution of assembly components using measurement data from the final assembled products. The method assumes that typically occurring part variation patterns are known from the characteristics of the specific manufacturing process of the parts. Using the previously developed method of influence coefficients for compliant parts assembly, the sensitivity of components geometrical variation on the assembly product is determined. Finally, by using the sensitivity matrix and selecting a set of optimal sensor locations, the contribution of 75 Journal of Manufacturing Systems Vol. 25/No. 2 2006 assembly variation patterns (which constitute a diagnostic matrix) as measured by the sensors is computed. Because these assembly variation patterns correspond to the part variation patterns, it is possible to diagnose particular part variation patterns by identifying the significant assembly variation patterns. Diagnosability of part variation patterns using the proposed methodology has also been studied in relation to the full-rankness of the diagnostic matrix. Furthermore, the effects of the measurement noise, which may also represent the unmodeled behavior on the diagnosis, were discussed. To illustrate the capability of the proposed methodology, two case studies were performed: an assembly of two identical rectangular plates, and a fender-reinforcement assembly. In both cases, the methodology developed in this study was capable of identifying the part variation patterns even under some levels of noise. In the first simple case, some diagnosability issues occurred because the diagnostic matrix turned out to be less than full rank due to the complete symmetry of the problem. Given that, in general, real problems do not possess such complete symmetry, as in the second case of fender-reinforcement assembly, such a problem is not expected to occur in real cases. Having established a methodology for the diagnosis of part variation in a single-station assembly, the next natural problem to study will be the extension of the method to multiple-station assembly process. No major obstacle is expected in doing so, except the increased complexity of calculations. Camelio, J.; Hu, S.J.; and Marin, S. (2004). “Compliant assembly variation analysis using components geometric covariance.” ASME Journal of Mfg. Science and Engg. (v126, n2), pp355-360. Camelio, J.; Hu, S.J.; and Yim, H. (2005). “Sensor placement for effective diagnosis of multiple faults in fixturing of compliant parts.” ASME Journal of Mfg. Science and Engg. (v127, n1), pp68-74. Ceglarek, D. and Shi, J. (1995). “Dimensional variation reduction for automotive body assembly.” Mfg. Review (v8, n2), pp139-154. Ceglarek, D. and Shi, J. (1996). “Fixture failure diagnosis for autobody assembly using pattern recognition.” ASME Journal of Engg. for Industry (v118), pp55-66. Chang, M. and Gossard, D.C. (1998). “Computational method for diagnosis of variation-related assembly problems.” Int’l Journal of Production Research (v36, n11), pp2985-2995. Ding, Y.; Ceglarek, D.; and Shi, J. (2002). “Fault diagnosis of multistage manufacturing processes by using state space approach.” ASME Journal of Mfg. Science and Engg. (v124), pp313-322. Goldberg, D.E. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Reading, MA: Addison-Wesley. Greenwood, W.H. and Chase, K.W. (1988). “Worst case tolerance analysis with nonlinear problems.” Journal of Engg. for Industry (v110), pp232-235. Hu, S.J. and Wu, S.M. (1992). “Identifying root causes of variation in automobile body assembly using Principal Component Analysis.” Transactions of NAMRI/SME (v20), pp311-316. Liu, S.C. and Hu, S.J. (1997). “Variation simulation for deformable sheet metal assemblies using finite element methods.” ASME Journal of Mfg. Science and Engg. (v119), pp368-374. Liu, Y.G. and Hu, S.J. (2005). “Assembly fixture fault diagnosis using Designated Component Analysis.” ASME Journal of Mfg. Science and Engg. (v127, n2), pp358-368. Strang, G. (1988). Linear Algebra and Its Applications, 3rd ed. Harcourt College Publishers. Wang Y. and Nagarkar, S.R. (1999). “Locator and sensor placement for automated coordinate checking fixtures.” ASME Trans., Journal of Mfg. Science and Engg. (v121), pp709-719. Authors’ Biographies Dr. Jaime Camelio is currently an assistant professor in the Mechanical Engineering-Engineering Mechanics Dept. at Michigan Technological University. Previously he was a consultant in the Automotive/ Operations Practice at A.T. Kearney Inc. Dr. Camelio obtained his BS and MS in mechanical engineering from the Catholic University of Chile in 1994 and 1995, respectively. In 2002, he received his PhD from the University of Michigan. After graduation, he held a postdoctoral research position and later a research scientist position in the Dept. of Mechanical Engineering at the University of Michigan, Ann Arbor. His research interests are in assembly systems modeling, uncertainty management, systems diagnosis/prognosis, and remanufacturing. Dr. Camelio received the 2007 SME Kuo K. Wang Outstanding Young Manufacturing Engineer award. Acknowledgments The authors acknowledge the case study data provided by Mr. Eric Chen from General Motors Corporation. References Apley, D. and Shi, J. (1998). “Diagnosis of multiple fixture faults in panel assembly.” Journal of Mfg. Science and Engg. (v120), pp793-801. Cai, W.; Hu, S.J.; and Yuan, J.X. (1996). “Deformable sheet metal fixturing: principles, algorithms, and simulations.” Journal of Mfg. Science and Engg. (v118), pp318-324. Camelio, J.; Hu, S.J.; and Ceglarek, D. (2003). “Modeling variation propagation of multi-station assembly systems with compliant parts.” ASME Journal of Mechanical Design (v125, n4), pp673-681. Hyunjune Yim is director of the PACE Center and a professor in the Mechanical and System Design Engineering Dept. at Hongik University (Seoul, Korea). He received his bachelor’s (1984) and master’s (1986) degrees in mechanical engineering from Seoul National University and PhD (1993) in mechanical engineering from the Massachusetts Institute of Technology. Dr. Yim’s current research interests include tolerance analysis and design, digital manufacturing, and digital product development. 76