A method for detecting abnormality in automobile resistance welding based on multi-channel feature integration

Abstract

The present invention discloses a method for detecting anomalies in automobile resistance welding based on multi-channel feature integration. The method collects process data of automobile resistance welding that meets welding quality standards and organizes it into a real number matrix; performs multi-channel feature expansion on the real number matrix based on mean value, standard deviation, skewness, kurtosis and negative entropy; performs standardization processing on the extended feature vectors by calculating the mean value and standard deviation of each row vector in an extended feature matrix formed by merging all extended feature vectors; performs singular value decomposition on the standardized extended feature matrix, and then obtains a feature integration matrix and an upper limit of a normal variation range of automobile resistance welding; for the latest process data of automobile resistance welding, similarly obtains the standardized extended feature vector, combines the feature integration matrix to obtain a score vector, and then obtains an anomaly detection index, and determines whether welding quality anomalies occur by comparing with the upper limit; the advantage is that the entire process does not involve complex calculations and the required computing load is very small.

Description

Technical Field

[0001] The present invention relates to a resistance spot welding anomaly detection technology, and in particular to an automobile resistance welding anomaly detection method based on multi-channel feature integration. Background Art

[0002] Resistance spot welding involves assembling the welded parts into overlapping joints and pressing them between two electrodes. Resistance heat is then used to melt the base metal, forming the weld. Compared to traditional riveting or other welding methods, resistance spot welding offers advantages such as high joint quality, fewer auxiliary steps, high production efficiency, no need for additional welding materials, and ease of mechanization and automation. It has been widely used in industrial manufacturing fields such as automotive, aerospace, and electronics.

[0003] Resistance welding accounts for over 95% of all welding in automobiles. However, without effective and reliable welding quality inspection technology, to ensure vehicle body strength, the number of welds typically increases by 20% to 30%. While this approach maintains overall weld quality, it results in significant waste of manpower, material, and capital resources. Because automotive resistance welding is a very short process, weld quality is often determined in a single pass. Immediate assessment of weld quality abnormalities is required after completion, providing guidance for re-welding.

[0004] Although automotive resistance welding is a complex and time-consuming process, the factors that influence welding quality are primarily five signals: welding voltage, welding current, dynamic resistance, electrode pressure, and electrode displacement. While the entire process from initiation to completion is intermittent, electrical and mechanical sensors can capture signal data from the entire process. This provides a foundation for data-driven anomaly detection in automotive resistance welding. Furthermore, with the current momentum of intelligent manufacturing, developing multi-parameter integrated monitoring and utilizing multi-information fusion technologies are effective approaches to improving automotive resistance welding quality monitoring. The signal data corresponding to each welding process can be used to identify any anomalies in automotive resistance welding. Therefore, efficiently processing and analyzing the changing characteristics of the signal data corresponding to each welding process is crucial.

[0005] Given the complex nature of the automotive resistance welding process, the electrical and mechanical signals captured during each weld, while traceable, exhibit varying patterns over time. Furthermore, the amount of data provided varies from weld to weld, inevitably leading to uneven batch lengths. Therefore, implementing data-driven anomaly detection for automotive resistance welding requires not only analyzing and extracting the potential variations in the electrical and mechanical signals during the welding process from a channel perspective, but also employing technologies that can effectively integrate multi-channel features to instantly detect anomalies in the process. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for detecting anomalies in automobile resistance welding based on multi-channel feature integration, which expands the multi-channel features of the process data of automobile resistance welding, and uses a new feature analysis technology to integrate the expanded multi-channel features, and then uses the integrated features to instantly detect whether the welding quality is abnormal. The entire feature expansion and feature integration process does not involve complex calculations, and the required computing load is very small, which can better adapt to the short process characteristics of automobile resistance welding.

[0007] The technical solution adopted by the present invention to solve the above technical problems is: a method for detecting automobile resistance welding anomalies based on multi-channel feature integration, which is characterized by specifically comprising the following steps:

[0008] Step 1: Preprocessing the process data of automotive resistance welding that meets welding quality standards, specifically including steps 1.1 to 1.4 shown below;

[0009] Step 1.1: Collect B times of automobile resistance welding process data that meet the welding quality standards, and organize the process data of each automobile resistance welding from the beginning to the end into a real number matrix consisting of 5 columns of data, thereby obtaining B real number matrices, which are recorded as X1, X2, ..., X B ; Among them, the b-th real matrix is ​​X b , X b The process data of the bth automobile resistance welding from the beginning to the end is sorted into, X b The data in the first to fifth columns correspond to the welding voltage, welding current, dynamic resistance, electrode pressure and electrode displacement in K b The data collected at the sampling time node, X b The dimension is K b ×5, number b=1,2,…,B;

[0010] Step 1.2: Calculate the mean, standard deviation, skewness, kurtosis and negative entropy of the process data of automobile resistance welding for X1, X2, ..., X1, respectively. B Implement multi-channel feature expansion to obtain X1,X2,…,X B The corresponding extended eigenvectors are denoted as y1, y2, ..., y B ;

[0011] Step 1.3: Replace y1,y2,…,y B Merge into an extended feature matrix with a dimension of 25×B, denoted as Y, Y=[y1,y2,…,y B]; then calculate the average values ​​of the first row vector, the second row vector, to the 25th row vector in Y, and record them as m1, m2, ..., m 25 ; and calculate the standard deviations of the first row vector, the second row vector, to the 25th row vector in Y, which are denoted as z1, z2, ..., z 25 ;

[0012] Step 1.4: Set m1, m2, …, m 25 Form an average value vector with a dimension of 25×1, denoted as M, M=[m1,m2,…,m 25 ] T ; and z1,z2,…,z 25 Form a diagonal matrix of dimension 25×25, denoted as Z, where the data on the diagonal of Z are equal to z1, z2,…, z 25 , the data on the non-diagonal lines in Z are all equal to 0; then according to the formula For y1,y2,…,y B Implement standardization and obtain the expanded feature vectors after standardization, which are recorded as Where, number b = 1, 2, ..., B, the superscript "T" represents the transpose of the matrix or vector, y b Represents X b The corresponding extended eigenvector, y b represents y b The corresponding normalized extended feature vector;

[0013] Step 2: Perform feature integration on the standardized extended feature vector and set the upper limit of the normal variation range of automotive resistance welding, which specifically includes steps 2.1 to 2.3 shown below;

[0014] Step 2.1: Merge into a standardized extended feature matrix with a dimension of 25×B, recorded as Again Implement singular value decomposition, which is: Where U and V represent two unitary matrices of singular value decomposition, and S represents a diagonal matrix consisting of non-zero singular values;

[0015] Step 2.2: According to the formula R = VS -1 Calculate the feature integration matrix R; and according to the formula Θ=UU T Calculate the matrix Θ;

[0016] Step 2.3: Arrange the 10 largest data on the diagonal of Θ in descending order and record them as φ1, φ2, φ3, φ4, φ5, φ6, φ7, φ8, φ9, φ10 ; Then set the upper limit of the normal variation range of automobile resistance welding, recorded as

[0017] Step 3: Implement online abnormality detection on the welding quality of automobile resistance welding, specifically including steps 3.1 to 3.3 shown below;

[0018] Step 3.1: Collect the latest automotive resistance welding process data and organize the process data from the start to the end of the automotive resistance welding into a real number matrix consisting of 5 columns of data, recorded as X new Then, we can analyze the five channels of X from the mean, standard deviation, skewness, kurtosis and negative entropy of the process data of automobile resistance welding. new Implement multi-channel feature expansion to obtain X new The corresponding extended eigenvector is denoted as y new ;

[0019] Step 3.2: According to the formula y new Perform standardization to obtain the expanded feature vector after standardization, which is recorded as ; Then according to the formula right Implement feature integration to obtain the score vector after feature integration, denoted as υ;

[0020] Step 3.3: According to the formula Calculating anomaly detection metrics Re-judge Is it greater than If so, it is considered that the welding quality of this automobile resistance welding is abnormal, and the welding abnormality record is retained for repeated welding; if not, it is considered that the welding quality of this automobile resistance welding meets the welding quality standards.

[0021] The specific implementation process of step 1.2 includes steps A to F shown below:

[0022] Step A: Set b = 1 and set the real matrix X = X b , set j = 1, 2, ..., 5, and at the same time follow the average value equal to μ j , standard deviation equals δ j A N×1 dimensional data vector is randomly generated using the normal distribution, thereby obtaining 5 N×1 dimensional data vectors, which are denoted as u1, u2, …, u5 respectively;

[0023] Step B: Let the first, second, and fifth column vectors in X be denoted as x1, x2, …, x5 respectively; then, according to the formula Calculate the average value corresponding to each column vector in X, and record them as μ1, μ2,…, μ5; where μ j represents the j-th column vector x in X j The corresponding average value, K represents the number of row vectors in X, row number k = 1, 2, ..., K, column number j = 1, 2, ..., 5, x j (k) represents the j-th column vector x in X j The kth data;

[0024] Step C: According to the formula Calculate the standard deviation of each column vector in X, which is denoted as δ1, δ2, ..., δ5; and according to the formula Calculate the skewness corresponding to each column vector in X, and record them as γ1, γ2, ..., γ5 in sequence; according to the formula Calculate the kurtosis of each column vector in X, which are denoted as η1, η2, ..., η5; where δ j represents the j-th column vector x in X j The corresponding standard deviation, γ j represents the j-th column vector x in X j The corresponding skewness, η j represents the j-th column vector x in X j The corresponding kurtosis;

[0025] Step D: According to the formula Calculate the negative entropy corresponding to each column vector in X, and record them as θ1, θ2, ..., θ5 in sequence; where θ j represents the j-th column vector x in X j The corresponding negative entropy, exp() represents an exponential function with a natural constant as the base, u j (n) represents u j The nth element in, n=1,2,…,N, when j is equal to 1,2,…,5 respectively, u j Corresponding representations are u1,u2,…,u5;

[0026] Step E: Combine μ1, μ2, …, μ5, δ1, δ2, …, δ5, γ1, γ2, …, γ5, η1, η2, …, η5, and θ1, θ2, …, θ5 into a 25×1 dimensional extended feature vector, denoted as y; wherein the first to fifth data in y are equal to μ1, μ2, …, μ5 respectively, the sixth to tenth data are equal to δ1, δ2, …, δ5 respectively, the eleventh to fifteenth data are equal to γ1, γ2, …, γ5 respectively, the sixteenth to twentieth data are equal to η1, η2, …, η5 respectively, and the twenty-first to twenty-fifth data are equal to θ1, θ2, …, θ5 respectively;

[0027] Step F: Set up Xb The corresponding extended eigenvector y b =y; then determine whether b is less than B. If so, set b = b + 1, and then set X = X b And return to step B to continue execution; if not, then get X1, X2, ..., X B The corresponding extended eigenvectors y1,y2,…,y B .

[0028] Compared with the prior art, the advantages of the present invention are:

[0029] First, the method of the present invention can expand and integrate features from multiple channels of automotive resistance welding process data, enabling multi-channel and multi-angle monitoring of each weld quality for abnormalities. Second, the entire feature expansion and integration process of the present method does not involve complex calculations, requiring a very low computational load, making it well suited to the short processing time of automotive resistance welding. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flow chart for the implementation of the method of the present invention;

[0031] Figure 2 Detailed diagram of abnormality detection of automobile resistance welding using the method of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

[0033] The present invention proposes a method for detecting abnormalities in automobile resistance welding based on multi-channel feature integration, such as Figure 1 As shown, it specifically includes the following steps:

[0034] Step 1: Preprocess the process data of automotive resistance welding that meets welding quality standards, specifically including steps 1.1 to 1.4 shown below.

[0035] Step 1.1: Collect B times of automobile resistance welding process data that meet the welding quality standards, and organize the process data of each automobile resistance welding from the beginning to the end into a real number matrix consisting of 5 columns of data, thereby obtaining B real number matrices, which are recorded as X1, X2, ..., X B ; Among them, X1 represents the real number matrix formed by the process data of the first automobile resistance welding from the beginning to the end, that is, the first real number matrix, X2 represents the real number matrix formed by the process data of the second automobile resistance welding from the beginning to the end, that is, the second real number matrix, X B The real number matrix representing the process data of the Bth automobile resistance welding from the beginning to the end is the Bth real number matrix, and the bth real number matrix is ​​X b , X bThe process data of the bth automobile resistance welding from the beginning to the end is sorted into, X b The data in the first to fifth columns correspond to the welding voltage, welding current, dynamic resistance, electrode pressure and electrode displacement in K b The data collected at the sampling time node, X b The dimension is K b ×5, number b=1,2,…,B.

[0036] Step 1.2: Calculate the mean, standard deviation, skewness, kurtosis and negative entropy of the process data of automobile resistance welding for X1, X2, ..., X1, respectively. B Implement multi-channel feature expansion to obtain X1,X2,…,X B The corresponding extended eigenvectors are denoted as y1, y2, ..., y B ; Among them, X b The corresponding extended eigenvector y b The dimension is 25×1.

[0037] Here, the specific implementation process of step 1.2 includes steps A to F shown below:

[0038] Step A: Set b = 1 and set the real matrix X = X b , set j = 1, 2, ..., 5, and at the same time follow the average value equal to μ j , standard deviation equals δ j A N×1 dimensional data vector is randomly generated using the normal distribution, thereby obtaining 5 N×1 dimensional data vectors, which are denoted as u1, u2,…, u5 respectively.

[0039] Step B: Let the first, second, and fifth column vectors in X be denoted as x1, x2, …, x5 respectively; then, according to the formula Calculate the average value corresponding to each column vector in X, and record them as μ1, μ2,…, μ5; where μ j represents the j-th column vector x in X j The corresponding average value, μ1 represents the average value corresponding to the first column vector x1 in X, μ2 represents the average value corresponding to the second column vector x2 in X, μ5 represents the average value corresponding to the fifth column vector x5 in X, K represents the number of row vectors in X, row number k = 1, 2, ..., K, column number j = 1, 2, ..., 5, x j (k) represents the j-th column vector x in X j The kth data.

[0040] Step C: According to the formula Calculate the standard deviation of each column vector in X, which is denoted as δ1, δ2, ..., δ5; and according to the formula Calculate the skewness corresponding to each column vector in X, and record them as γ1, γ2, ..., γ5 in sequence; according to the formula Calculate the kurtosis of each column vector in X, which are denoted as η1, η2, ..., η5; where δ j represents the j-th column vector x in X j The corresponding standard deviation, δ1 represents the standard deviation of the first column vector x1 in X, δ2 represents the standard deviation of the second column vector x2 in X, δ5 represents the standard deviation of the fifth column vector x5 in X, γ j represents the j-th column vector x in X j The corresponding skewness, γ1 represents the skewness of the first column vector x1 in X, γ2 represents the skewness of the second column vector x2 in X, γ5 represents the skewness of the fifth column vector x5 in X, η j represents the j-th column vector x in X j The corresponding kurtosis, η1 represents the kurtosis corresponding to the first column vector x1 in X, η2 represents the kurtosis corresponding to the second column vector x2 in X, and η5 represents the kurtosis corresponding to the fifth column vector x5 in X.

[0041] Step D: According to the formula Calculate the negative entropy corresponding to each column vector in X, and record them as θ1, θ2, ..., θ5 in sequence; where θ j represents the j-th column vector x in X j The corresponding negative entropy, θ1 represents the negative entropy corresponding to the first column vector x1 in X, θ2 represents the negative entropy corresponding to the second column vector x2 in X, θ5 represents the negative entropy corresponding to the fifth column vector x5 in X, exp() represents the exponential function with a natural constant as the base, u j (n) represents u j The nth element in, n=1,2,…,N, when j is equal to 1,2,…,5 respectively, u j Corresponding representations are u1,u2,…,u5.

[0042] Step E: Combine μ1, μ2,…, μ5, δ1, δ2,…, δ5, γ1, γ2,…, γ5, η1, η2,…, η5, and θ1, θ2,…, θ5 into a 25×1 dimensional extended feature vector, denoted as y; wherein the first to fifth data in y are equal to μ1, μ2,…, μ5 respectively, the sixth to tenth data are equal to δ1, δ2,…, δ5 respectively, the eleventh to fifteenth data are equal to γ1, γ2,…, γ5 respectively, the sixteenth to twentieth data are equal to η1, η2,…, η5 respectively, and the twenty-first to twenty-fifth data are equal to θ1, θ2,…, θ5 respectively.

[0043] Step F: Set up X bThe corresponding extended eigenvector y b =y; then determine whether b is less than B. If so, set b = b + 1, and then set X = X b And return to step B to continue execution; if not, then get X1, X2, ..., X B The corresponding extended eigenvectors y1,y2,…,y B .

[0044] Step 1.3: Replace y1,y2,…,y B Merge into an extended feature matrix with a dimension of 25×B, denoted as Y, Y=[y1,y2,…,y B ]; then calculate the average values ​​of the first row vector, the second row vector, to the 25th row vector in Y, and record them as m1, m2, ..., m 25 ; and calculate the standard deviations of the first row vector, the second row vector, to the 25th row vector in Y, which are denoted as z1, z2, ..., z 25 ; Among them, the symbol "[]" is a vector or matrix representation symbol, m1 and z1 respectively represent the mean and standard deviation corresponding to the first row vector in Y, m2 and z2 respectively represent the mean and standard deviation corresponding to the second row vector in Y, m 25 and z 25 They represent the mean and standard deviation corresponding to the 25th row vector in Y respectively.

[0045] Step 1.4: Set m1, m2, …, m 25 Form an average value vector with a dimension of 25×1, denoted as M, M=[m1,m2,…,m 25 ] T ; and z1,z2,…,z 25 Form a diagonal matrix of dimension 25×25, denoted as Z, where the data on the diagonal of Z are equal to z1, z2,…, z 25 , the data on the non-diagonal lines in Z are all equal to 0; then according to the formula For y1,y2,…,y B Implement standardization and obtain the expanded feature vectors after standardization, which are recorded as Where, number b = 1, 2, ..., B, the superscript "T" represents the transpose of the matrix or vector, y b Represents X b The corresponding extended eigenvector, represents y b The corresponding normalized extended feature vector, The dimension is 25×1, represents the standardized extended feature vector corresponding to y1, represents the standardized extended feature vector corresponding to y2, represents y B The corresponding normalized expanded feature vector.

[0046] Step 2: Perform feature integration on the standardized extended feature vector and set the upper limit of the normal variation range of automotive resistance welding, which specifically includes steps 2.1 to 2.3 shown below.

[0047] Step 2.1: Merge into a standardized extended feature matrix with a dimension of 25×B, recorded as Again Implement singular value decomposition, which is: Where U and V represent two unitary matrices of singular value decomposition, and S represents a diagonal matrix consisting of non-zero singular values.

[0048] Step 2.2: According to the formula R = VS -1 Calculate the feature integration matrix R; and according to the formula Θ=UU T Calculate the matrix Θ.

[0049] Step 2.3: Arrange the 10 largest data on the diagonal of Θ in descending order and record them as φ1, φ2, φ3, φ4, φ5, φ6, φ7, φ8, φ9, φ 10 ; Then set the upper limit of the normal variation range of automobile resistance welding, recorded as

[0050] Step 3: Implement online abnormality detection on the welding quality of automobile resistance welding, specifically including steps 3.1 to 3.3 shown below.

[0051] Step 3.1: Collect the latest automotive resistance welding process data and organize the process data from the start to the end of the automotive resistance welding into a real number matrix consisting of 5 columns of data, recorded as X new Then, we can analyze the five channels of X from the mean, standard deviation, skewness, kurtosis and negative entropy of the process data of automobile resistance welding. new Implement multi-channel feature expansion to obtain X new The corresponding extended eigenvector is denoted as y new ; Among them, X new The data in the first to fifth columns correspond to the welding voltage, welding current, dynamic resistance, electrode pressure and electrode displacement in K new The data collected at the sampling time node, X new The dimension is K new ×5,y newThe dimension of is 25 × 1. Here, follow the process from step B to step E to obtain y in the same way. new , that is: calculate X according to step B new The average value corresponding to each column vector in X; According to step C, calculate the standard deviation, skewness, and kurtosis corresponding to each column vector in X; According to step D, calculate the negative entropy corresponding to each column vector in X; According to step E, convert X new The mean, standard deviation, skewness, kurtosis, and negative entropy corresponding to all column vectors in are sequentially merged into a 25×1-dimensional extended feature vector, which is y new .

[0052] Step 3.2: According to the formula y new Perform standardization to obtain the expanded feature vector after standardization, which is recorded as ; Then according to the formula right Implement feature integration to obtain the score vector after feature integration, denoted as υ; where, The dimension is 25×1.

[0053] Step 3.3: According to the formula Calculating anomaly detection metrics Re-judge Is it greater than If so, it is considered that the welding quality of this automobile resistance welding is abnormal, and the welding abnormality record is retained for repeated welding; if not, it is considered that the welding quality of this automobile resistance welding meets the welding quality standards.

[0054] Through the above steps 3.1 to 3.3, anomaly detection is performed using the process data of 100 automobile resistance welding operations, and the following is obtained: Figure 2 The anomaly detection image shown, Figure 2 The horizontal axis represents the numbers of automobile resistance welding from 1 to 100, the vertical axis represents the abnormal detection index corresponding to each automobile resistance welding, and the dotted line represents the upper limit of the normal variation range of automobile resistance welding.

[0055] pass Figure 2 It can be found that the method of the present invention can immediately provide abnormal detection results for each automobile resistance welding. The welding points corresponding to the points beyond the dotted line are abnormal, and the corresponding welding records need to be retained for repeated welding.

Claims

1. A method for detecting abnormalities in automobile resistance welding based on multi-channel feature integration, characterized in that The specific steps include: Step 1: Preprocessing the process data of automotive resistance welding that meets welding quality standards, specifically including steps 1.1 to 1.4 shown below; Step 1.1: Collect B times of automobile resistance welding process data that meet the welding quality standards, and organize the process data of each automobile resistance welding from the beginning to the end into a real number matrix consisting of 5 columns of data, thereby obtaining B real number matrices, which are recorded as X1, X2, ..., X B ; Among them, the b-th real matrix is ​​X b , X b The process data of the bth automobile resistance welding from the beginning to the end is sorted into, X b The data in the first to fifth columns correspond to the welding voltage, welding current, dynamic resistance, electrode pressure and electrode displacement in K b The data collected at the sampling time node, X b The dimension is K b ×5, number b=1,2,…,B; Step 1.2: Calculate the mean, standard deviation, skewness, kurtosis and negative entropy of the process data of automobile resistance welding for X1, X2, ..., X1, respectively. B Implement multi-channel feature expansion to obtain X1,X2,…,X B The corresponding extended eigenvectors are denoted as y1, y2, ..., y B ; The specific implementation process of step 1.2 includes steps A to F shown below: Step A: Set b = 1 and set the real matrix X = X b , set j = 1, 2, ..., 5, and at the same time follow the average value equal to μ j , standard deviation equals δ j A N×1 dimensional data vector is randomly generated using the normal distribution, thereby obtaining 5 N×1 dimensional data vectors, which are denoted as u1, u2, …, u5 respectively; Step B: Let the first, second, and fifth column vectors in X be denoted as x1, x2, …, x5 respectively; then, according to the formula Calculate the average value corresponding to each column vector in X, and record them as μ1, μ2,…, μ5; where μ j represents the j-th column vector x in X j The corresponding average value, K represents the number of row vectors in X, row number k = 1, 2, ..., K, column number j = 1, 2, ..., 5, x j (k) represents the j-th column vector x in X j The kth data; Step C: According to the formula Calculate the standard deviation of each column vector in X, which is denoted as δ1, δ2, ..., δ5; and according to the formula Calculate the skewness corresponding to each column vector in X, and record them as γ1, γ2, ..., γ5 in sequence; according to the formula Calculate the kurtosis of each column vector in X, which are denoted as η1, η2, ..., η5; where δ j represents the j-th column vector x in X j The corresponding standard deviation, γ j represents the j-th column vector x in X j The corresponding skewness, η j represents the j-th column vector x in X j The corresponding kurtosis; Step D: According to the formula Calculate the negative entropy corresponding to each column vector in X, and record them as θ1, θ2, ..., θ5 in sequence; where θ j represents the j-th column vector x in X j The corresponding negative entropy, exp() represents an exponential function with a natural constant as the base, u j (n) represents u j The nth element in, n=1,2,…,N, when j is equal to 1,2,…,5 respectively, u j Corresponding representations are u1,u2,…,u5; Step E: Combine μ1, μ2, …, μ5, δ1, δ2, …, δ5, γ1, γ2, …, γ5, η1, η2, …, η5, and θ1, θ2, …, θ5 into a 25×1 dimensional extended feature vector, denoted as y; wherein the first to fifth data in y are equal to μ1, μ2, …, μ5 respectively, the sixth to tenth data are equal to δ1, δ2, …, δ5 respectively, the eleventh to fifteenth data are equal to γ1, γ2, …, γ5 respectively, the sixteenth to twentieth data are equal to η1, η2, …, η5 respectively, and the twenty-first to twenty-fifth data are equal to θ1, θ2, …, θ5 respectively; Step F: Set up X b The corresponding extended eigenvector y b =y; then determine whether b is less than B. If so, set b = b + 1, and then set X = X b And return to step B to continue execution; if not, then get X1, X2, ..., X B The corresponding extended eigenvectors y1,y2,…,y B ; Step 1.3: Replace y1,y2,…,y B Merge into an extended feature matrix with a dimension of 25×B, denoted as Y, Y=[y1,y2,…,y B ]; then calculate the average values ​​of the first row vector, the second row vector, to the 25th row vector in Y, and record them as m1, m2, ..., m 25 ; and calculate the standard deviations of the first row vector, the second row vector, to the 25th row vector in Y, which are denoted as z1, z2, ..., z 25 ; Step 1.4: Set m1, m2, …, m 25 Form an average value vector with a dimension of 25×1, denoted as M, M=[m1,m2,…,m 25 ] T ; and z1,z2,…,z 25 Form a diagonal matrix of dimension 25×25, denoted as Z, where the data on the diagonal of Z are equal to z1, z2,…, z 25 , the data on the non-diagonal lines in Z are all equal to 0; then according to the formula For y1,y2,…,y B Implement standardization and obtain the expanded feature vectors after standardization, which are recorded as Where, number b = 1, 2, ..., B, the superscript "T" represents the transpose of the matrix or vector, y b Represents X b The corresponding extended eigenvector, represents y b The corresponding normalized extended feature vector; Step 2: Perform feature integration on the standardized extended feature vector and set the upper limit of the normal variation range of automotive resistance welding, which specifically includes steps 2.1 to 2.3 shown below; Step 2.1: Merge into a standardized extended feature matrix with a dimension of 25×B, recorded as Again Implement singular value decomposition, which is: Where U and V represent two unitary matrices of singular value decomposition, and S represents a diagonal matrix consisting of non-zero singular values; Step 2.2: According to the formula R = VS -1 Calculate the feature integration matrix R; and according to the formula Θ=UU T Calculate the matrix Θ; Step 2.3: Arrange the 10 largest data on the diagonal of Θ in descending order and record them as φ1, φ2, φ3, φ4, φ5, φ6, φ7, φ8, φ9, φ 10 ; Then set the upper limit of the normal variation range of automobile resistance welding, recorded as Step 3: Implement online abnormality detection on the welding quality of automobile resistance welding, specifically including steps 3.1 to 3.3 shown below; Step 3.1: Collect the latest automotive resistance welding process data and organize the process data from the start to the end of the automotive resistance welding into a real number matrix consisting of 5 columns of data, recorded as X new Then, we can analyze the five channels of X from the mean, standard deviation, skewness, kurtosis and negative entropy of the process data of automobile resistance welding. new Implement multi-channel feature expansion to obtain X new The corresponding extended eigenvector is denoted as y new ; Step 3.2: According to the formula y new Perform standardization to obtain the expanded feature vector after standardization, which is recorded as Then according to the formula right Implement feature integration to obtain the score vector after feature integration, denoted as υ; Step 3.3: According to the formula Calculating anomaly detection metrics Re-judge Is it greater than If so, it is considered that the welding quality of this automobile resistance welding is abnormal, and the welding abnormality record is retained for repeated welding; if not, it is considered that the welding quality of this automobile resistance welding meets the welding quality standards.

Images