Welding gun health feature selection method based on physical consistency weighting

Abstract

The application discloses a welding tongs health feature selection method based on physical consistency weighting and belongs to the technical field of industrial equipment fault diagnosis. Welding cycle data under multiple welding procedures of a servo resistance spot welding tongs is collected, a robust statistical baseline is established by using normal state samples of a source program, and a physical mechanism group is constructed according to the main physical representation of candidate measurement features. Physical consistency indexes and sample weights are calculated according to the deviation intensity of the samples relative to the normal baseline of the program and the coordination degree among the mechanism groups. The health contribution and working condition sensitivity of the features are evaluated under the action of the weights, four types of division and comprehensive sorting are completed, high health contribution and high working condition sensitive features in the selected candidate subset are corrected by using the robust baseline, the number of features is determined through internal one-leave-verification of the source program, and a health feature subset used for health state recognition of a target program is obtained. The application can reduce the interference of program differences on feature evaluation and retain the health information in the working condition sensitive features.

Description

Technical Field

[0001] This invention belongs to the field of welding equipment health monitoring technology, specifically involving a method for selecting welding clamp health characteristics based on physical consistency weighting. Background Technology

[0002] Servo resistance spot welding guns are key equipment in automotive body-in-white welding production lines. Their operation involves coupled physical processes such as mechanical actuation, electrode-workpiece contact, electrical energy transmission and Joule heating, heat transfer, and cooling. Industrial sites typically collect multi-source measurement data, including welding current, welding voltage, dynamic resistance, power, energy, transformer temperature, cooling water flow rate, electrode pressure, and electrode regrinding frequency. This data is then used for welding gun condition monitoring and fault diagnosis.

[0003] The same welding clamp typically performs multiple welding procedures during production. Different welding procedures differ in target current, welding time, electrode pressure, plate assembly, and production cycle time, causing the statistical distribution of candidate measurement variables to change systematically with each welding procedure. Simultaneously, changes in health status, such as electrode wear, abnormal cooling, contact deterioration, and abnormal conductive circuits, can also cause changes in the same measurement variables. Therefore, candidate measurement features often contain both health status information and welding procedure information, easily leading to confusion between procedure differences and health differences.

[0004] Some existing feature selection methods primarily evaluate the candidate measurement features based on statistical correlation, mutual information, model importance, or feature redundancy between the candidate measurement features and health status labels. Under conditions of a single welding procedure or where the distributions of training and test data are relatively similar, these methods can retain features with strong discriminative power. However, under conditions of multiple welding procedures, changes in measurement variables caused by differences in welding procedures are easily misjudged as differences in health status, reducing the stability of the selected features in the absence of welding procedures.

[0005] Some cross-condition diagnostic methods reduce the differences in data distribution between different conditions through methods such as distribution alignment, domain adaptation, or latent space representation learning. These methods typically process the internal representation of the model, making it difficult to directly explain the health contribution and condition sensitivity of the original industrial measurement variables, and also hindering the formation of health feature subsets with clear physical meaning at the level of the original measurement variables.

[0006] Some physical information methods utilize equipment operating mechanisms to construct new features, set model constraints, or interpret diagnostic results. However, existing methods still lack a way to transform the correlation of multiple physical processes in welding into sample-level evaluation information that can be directly used for screening original candidate measurement features. Furthermore, for features that simultaneously possess high health discrimination capabilities and high operational condition sensitivity, direct deletion may result in the loss of effective health information; conversely, direct retention may allow differences in welding procedure benchmarks to continue affecting subsequent identification results.

[0007] Therefore, there is a need for a welding clamp health feature selection method that can utilize the welding clamp operation mechanism to reduce the interference of welding procedure differences on the evaluation of candidate measurement features and improve the stability of health features under multiple welding procedures and no welding procedure conditions. Summary of the Invention

[0008] To address the problems of existing feature selection methods that easily misjudge program differences as health differences under multiple welding procedures, have difficulty distinguishing between health information and working condition information in candidate measurement features, and easily cause the loss of effective health information by directly deleting working condition sensitive features, this invention proposes a welding clamp health feature selection method based on physical consistency weighting.

[0009] This method first maps candidate measurement features to corresponding physical mechanism groups based on the physical correlation between the electrical-energy, pressure actuation, thermal state, cooling state, and electrode process of the servo resistance spot welding clamp, and establishes a robust statistical baseline using samples of the normal state of the source welding program. Then, based on the deviation intensity of each sample from the statistical baseline of the normal state of its respective welding program and the degree of coordination between different physical mechanism groups, a physical consistency index is calculated and sample weights are generated. Under the action of the sample weights, the health contribution and operating condition sensitivity of the candidate measurement features are evaluated respectively, and the candidate measurement features are classified into different types. Subsequently, the candidate measurement features are comprehensively ranked according to feature type, health contribution, and operating condition stability. For candidate feature subsets corresponding to different feature numbers, robust position-scale correction is performed on the selected high health contribution and high operating condition sensitivity features, and the number of features is determined by leave-one-out verification within the source welding program, thereby obtaining a subset of welding clamp health features suitable for cross-program health state identification.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The method for selecting welding clamp health characteristics based on physical consistency weighting includes the following steps:

[0012] Step 1: Data Acquisition and Sample Construction

[0013] Step 1.1: Welding Data Acquisition and Sample Construction

[0014] Welding data from a servo resistance spot welding clamp under multiple welding programs were collected, and each welding data point was used to construct a sample. Each sample included candidate measurement features, health status labels, and the source welding program.

[0015] Step 1.2: Construction of the Source Welding Program Sample Set

[0016] The source welding program set is represented as: The source welding program sample set is represented as: The candidate measurement feature set is represented as: .

[0017] in, A collection of source welding procedures; This is a sample set of source welding programs; The total number of source samples; For the first Candidate measurement feature vectors for each sample For the first One candidate measurement feature, For the first Each sample in candidate measurement features The measured values ​​on The number of candidate measurement features. ; For the first Health status labels for each sample; For the first The welding procedure identifier to which each source sample belongs.

[0018] Step 1.3: Construction of the target welding program sample set

[0019] The unlabeled sample set of the target welding program to be identified is represented as follows: .

[0020] in, For the first Candidate measurement feature vectors for each target program sample; Identify the target welding procedure; The number of samples for the target welding procedure.

[0021] The health status label of the target welding process is not involved in the construction of physical consistency sample weights, feature evaluation, feature ranking, determination of feature quantity, and training of the health status recognition model.

[0022] Step 2: Construction of robust statistical baseline and physical mechanism set under normal conditions

[0023] Step 2.1: Construction of a robust statistical baseline under normal conditions

[0024] From the source welding procedure sample set Extract normal state samples. Let the label of normal state be... .

[0025] For any source welding procedure Its normal state sample index set is represented as: .

[0026] in, For source welding procedures The normal state sample index set, Indicates the first The samples are in a normal state. Indicates the first The sample belongs to the source welding process. .

[0027] For candidate measurement feature set The first in Candidate measurement features Its source welding process The normal state position parameters and scale parameters are as follows: , .

[0028] in, Candidate measurement features In the source welding process Normal state position parameters; These are the corresponding normal state scale parameters; This represents the median; This represents the interquartile range; the coefficient 1.349 is used to convert the interquartile range under normal distribution conditions to a scale similar to the standard deviation. To avoid extremely small positive numbers with a denominator of zero.

[0029] Step 2.2: Construction of the physical mechanism group

[0030] Based on the operating mechanism of the servo resistance spot welding clamp and the main physical characterization function of the candidate measurement features, the candidate measurement features are set. Mapped to A group of physical mechanisms. The mapping function from candidate measurement features to the physical mechanism group is defined as:

[0031] .

[0032] in, Indicates the first Candidate measurement features Mapped to the A group of physical mechanisms, .

[0033] No. The set of candidate measurement features corresponding to each physical mechanism group is represented as follows:

[0034] .

[0035] in, The total number of physical mechanism groups; Number the physical mechanism group; For the first A set of candidate measurement features contained in a physical mechanism group.

[0036] Step 3: Calculation of Physical Consistency Index and Sample Weights

[0037] Step 3.1: Calculation of the degree of deviation of candidate measurement features and the intensity of deviation of mechanism group

[0038] For the One sample, whose source welding procedure is identified as... ,in For the first Candidate measurement features Calculate the candidate measurement features of the sample. The standardized deviation from the normal robust statistical baseline of the source welding procedure: .

[0039] in, For the first The first source sample The degree of standardized deviation of each candidate measurement feature relative to the baseline of the normal state of its source welding process; The larger the value, the better the candidate measurement feature. The more significant the deviation from the baseline of the normal state of the welding procedure, the more pronounced the deviation.

[0040] To avoid a few extreme outliers having an excessive impact on sample weights, the standardization deviation is truncated: .

[0041] in, The degree of standardized deviation after truncation. The upper limit is truncated to indicate the degree of deviation.

[0042] No. The set of candidate measurement features corresponding to each physical mechanism group Calculate the first The average deviation of each sample from this physical mechanism group: .

[0043] in, Indicates the first The number of candidate measurement features included in each physical mechanism group. Indicates sample In the Average deviation intensity on each physical mechanism group; Indicate candidate measurement features Belongs to the A group of physical mechanisms.

[0044] Step 3.2: Calculation of Physical Consistency Index

[0045] Taking into account the overall deviation strength of the samples across each physical mechanism group and the degree of coordination between groups, the first... The physical consistency index for each sample is: .

[0046] in, For the first The sample at the th The average deviation intensity on each physical mechanism group The total number of physical mechanism groups; The discrete penalty coefficients for the mechanism group. This represents the standard deviation of the deviation strength across each physical mechanism group. The first term reflects the average deviation strength of the sample across all physical mechanism groups, while the second term is used to suppress isolated deviations caused by anomalous fluctuations in a single mechanism group. When the sample exhibits significant and relatively coordinated deviations across multiple related physical mechanism groups, Larger.

[0047] The average physical consistency index for all source welding process samples is: .

[0048] in, This represents the total number of source samples.

[0049] Step 3.3: Generation of Physically Consistent Sample Weights

[0050] The first The physical consistency index of each sample is mapped to the initial sample weights. .

[0051] in, For the first Initial physical consistency sample weights for each sample; The sample weight change intensity coefficient. and Let be the lower and upper bounds of the initial sample weights, respectively, and satisfy . ; This indicates that the calculation result is restricted to an interval. Inside.

[0052] The initial sample weights are normalized to obtain the final physically consistent sample weights: .

[0053] in, The sample index used for summation. For the first The initial physical consistency sample weights of each sample; the normalized sample weights satisfy: .

[0054] Step 4: Physical Consistency Weighted Health Contribution and Operating Condition Sensitivity Assessment

[0055] For the Candidate measurement features Using the normalized physical consistency sample weights obtained in step 3 Calculate their health contribution scores respectively. and operating condition sensitivity score .

[0056] Step 4.1: Physical Consistency Weighted Health Contribution Evaluation

[0057] Calculate candidate measurement features respectively The mutual information between health status labels, multi-class health status effect size, normal and abnormal binary effect size, and feature importance in the health status random forest model were analyzed, and each sub-indicator was normalized to... Interval.

[0058] Among them, the physical consistency sample weight It is used for multi-class health status effect size, normal and abnormal binary effect size, and random forest model training; mutual information is calculated directly from the training samples using the source welding program.

[0059] The physically consistent weighted health contribution score is represented as follows: .

[0060] in, Candidate measurement features Normalized mutual information between health status; The normalized multi-class health status effect size after introducing physical consistency sample weights; The normalized normal-abnormal binary effect size after introducing the weights of physically consistent samples; Normalized feature importance obtained after training a healthy random forest model by introducing physically consistent sample weights. , , and The weight coefficients corresponding to mutual information, multi-class health status effect size, normal-abnormal binary effect size, and random forest feature importance, respectively, satisfy: , , The number of the evaluation item for contributions to health.

[0061] The larger the value, the better the candidate measurement feature. The stronger the contribution to the judgment of the welding clamp's health status or fault type.

[0062] Step 4.2: Physical Consistency Weighted Operating Condition Sensitivity Evaluation

[0063] To reduce the impact of the coupling between fault category distribution and welding procedure distribution on the evaluation results, the condition sensitivity of candidate measurement features is calculated only using normal state samples from the source welding procedure. For candidate measurement features... The center drift, inter-program distribution differences, and predictability of the physically consistent weighted program were calculated separately, and each sub-indicator was normalized to... Interval.

[0064] The physical consistency weighted operating condition sensitivity score is expressed as follows: .

[0065] in, For the normalized physical consistency weighted program center drift, To represent the differences in Jensen–Shannon distributions between normalized programs, This is a normalized, physically consistent weighted program predictability index. The physical consistency sample weights are used to calculate program center drift and program predictability, while the distribution differences between programs are calculated based on the empirical distribution of normal state samples. , and These are the weighting coefficients corresponding to program center drift, inter-program distribution differences, and program predictability, respectively, and they satisfy the following: , ,in, This is the number for the sub-item of the operating condition sensitivity evaluation. The larger the value, the better the candidate measurement feature. The more sensitive it is to the welding procedure or operating conditions.

[0066] Step 5: Feature Type Determination

[0067] Let the health contribution score threshold be... The threshold for operating condition sensitivity scoring is ,in, , Based on candidate measurement features Health Contribution Score and operating condition sensitivity score Candidate measurement features are divided into the following four categories:

[0068] Category A: and This contributes to high health by low working condition sensitivity characteristics;

[0069] Category B: and This contributes to high health and high working condition sensitivity characteristics;

[0070] Class C: and This contributes to low-health, low-condition-sensitive characteristics;

[0071] Category D: and This contributes to the high-condition sensitivity characteristics of low health.

[0072] Features of type A have a high ranking priority; although features of type B have a high sensitivity to working conditions, they still contain strong health discrimination information, so they are retained and given the second highest ranking priority. In the evaluation of candidate feature subsets and the health status identification stage, robust baseline correction is performed on the selected features of type B; features of type C and D are not directly deleted, but are given a lower type priority, and whether they enter the final health feature subset is determined by the subsequent comprehensive ranking results and the number of features.

[0073] Step 6: Robust baseline correction for Class B features.

[0074] For the Class B features identified in step 5, their health contribution score, operating condition sensitivity score, and overall ranking position are retained. During the candidate feature subset evaluation and health status identification stage, robust baseline correction is performed only on Class B features within the selected subset.

[0075] For the The first sample There are 1 Class B feature, and the corrected feature values ​​are: .

[0076] in, Indicates sample The welding procedure identifier. and These represent the robust position and scale parameters for the corresponding normal state samples of the program. For the internal verification program and the target welding program, the corresponding parameters are estimated using the median and interquartile range of the unlabeled samples of the program.

[0077] The correction only applies to the selected B-class features; the A, C, and D-class features retain their original values. The correction results are used for subsequent internal validation and training and testing of the health status recognition model, and do not participate in feature scoring and overall ranking again.

[0078] Step 7: Comprehensive sorting, determination of feature quantity, and health status identification.

[0079] Step 7.1: Comprehensive ranking of candidate measurement features

[0080] The candidate measurement features of categories A, B, C, and D determined in step 5 are collectively used to form a feature set to be ranked. For the first... Candidate measurement features A comprehensive ranking score is calculated based on the health contribution score, working condition sensitivity score, and feature type priority: .

[0081] Prioritize feature types; and These are the weighting coefficients for the health contribution item and the working condition stability item, respectively.

[0082] Feature type priority is represented as follows: ,

[0083] in, Candidate measurement features The characteristic type; , , and These are the type priority coefficients for features of types A, B, C, and D, respectively, and they satisfy the following: .

[0084] Step 7.2: Determining the number of features

[0085] Based on comprehensive ranking score Arrange the candidate measurement features from highest to lowest to obtain a sequence. Set the set of candidate feature counts. For any Select before sorting Each candidate measurement feature constitutes a subset of candidate features. .

[0086] Within the source welding program, a leave-one-out verification process is performed. Each time, one source welding program is used as the internal verification program, and the remaining source welding programs are used as internal training programs. The feature scores, types, and rankings obtained in steps 4, 5, and 7.1 are kept unchanged, and the candidate feature subsets are processed according to step 6. Robust baseline correction was performed on the B-type features in the model, and then the health status recognition model was trained and its recognition performance was evaluated.

[0087] Candidate feature subset The internal validation composite index is expressed as:

[0088] in, For the source welding program set, Indicates the source welding procedure. Indicates the set of source welding procedures The cardinality, i.e., the number of internal verification procedures. , , and These represent the source welding process. As an internal verification procedure, it employs a subset of candidate features. The macro-average F1 score, Matthews correlation coefficient, area under the precision-recall curve, and area under the receiver operating characteristic curve were obtained at that time. , , and These are the weighting coefficients for the macro-average F1 score, Matthews correlation coefficient, area under the precision-recall curve, and area under the receiver operating characteristic curve, respectively, and they satisfy the following: And satisfy: , This is the number for the internal validation evaluation indicators.

[0089] The optimal number of features is determined based on the comprehensive internal validation metrics:

[0090] Select before sorting These features constitute a subset of the final health characteristics: .

[0091] Step 7.3: Health Status Identification of the Target Welding Program

[0092] For the subset of final health characteristics The B-type features in the model are subjected to robust baseline correction in step 6; the health status recognition model is trained using all source welding program samples, and the health status of the target welding program samples is then recognized.

[0093] Compared with the prior art, the present invention has the following beneficial effects:

[0094] (1) Transform the physical correlations of electrical conduction, energy input, heat transfer, cooling state and contact wear of welding clamp into sample-level physical consistency weights, so that samples that show obvious and coordinated deviations in multiple physical mechanism groups can obtain higher weights in feature evaluation, thereby enhancing the connection between feature evaluation and welding clamp operation mechanism.

[0095] (2) Evaluating health contribution and working condition sensitivity at the level of industrial measurement variables, and evaluating program response using only normal state samples, helps to reduce the situation where welding process differences are misjudged as health differences.

[0096] (3) Retaining the B-type features with high health contribution and high working condition sensitivity to participate in the comprehensive ranking, and implementing robust baseline correction on the selected B-type features instead of deleting them directly, is conducive to weakening the program baseline difference and retaining health deviation information;

[0097] (4) The number of features is determined by using the one-to-one verification within the source welding program. The health status label of the target welding program does not participate in feature selection and model training, which is conducive to improving the stability of the healthy feature subset under the condition of no welding program. Attached Figure Description

[0098] Figure 1 This is a flowchart of the welding clamp health feature selection method based on physical consistency weighting according to the present invention.

[0099] Figure 2 A schematic diagram showing the classification of candidate measurement features into categories A, B, C, and D.

[0100] Figure 3 The graph shows the performance test results of the health status identification method of the present invention under different target welding procedures. Detailed Implementation

[0101] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0102] Example 1: A Welding Claw Health Feature Selection Method Based on Physical Consistency Weighting

[0103] This embodiment selects weld point data collected from the automotive body-in-white welding production line to verify the weld clamp health feature selection method based on physical consistency weighting proposed in this invention.

[0104] like Figure 1 As shown, the welding clamp health feature selection method based on physical consistency weighting includes seven steps: data acquisition and sample construction, construction of robust statistical baselines and physical mechanism groups under normal conditions, calculation of physical consistency indices and sample weights, evaluation of health contribution and working condition sensitivity based on physical consistency weighting, feature type discrimination, robust baseline correction for type B features, comprehensive ranking, determination of feature quantity, and health status identification. The detailed steps are as follows:

[0105] Step 1: Data Acquisition and Sample Construction

[0106] Step 1.1: Welding Data Acquisition and Sample Construction

[0107] Welding data from a servo resistance spot welding clamp under multiple welding programs were collected, and each welding data point was used to construct a sample. Each sample included candidate measurement features, health status labels, and the welding program.

[0108] The candidate measurement features in this embodiment include:

[0109] Electrode regrinding count (dress_count); Inlet water flow rate (flow_in); First return water branch flow rate (flow_out1); Second return water branch flow rate (flow_out2); Total cooling water flow rate (flow_total); Transformer temperature (T_trans); Average welding current (I_avg); Average welding voltage (U_avg); Average electrode pressure (P_avg); Welding energy per weld point (Energy); Dynamic resistance (R_dyn); Weld point Q value (xal_q).

[0110] Step 1.2: Construction of the Source Welding Program Sample Set

[0111] The source welding program set is represented as: The source welding program sample set is represented as: The candidate measurement feature set is represented as: .

[0112] in, A collection of source welding procedures; This is a sample set of source welding programs; The total number of source samples; For the first Candidate measurement feature vectors for each sample For the first One candidate measurement feature, For the first Each sample in candidate measurement features The measured values ​​on The number of candidate measurement features. ; For the first Health status labels for each sample; For the first The welding procedure identifier to which each source sample belongs.

[0113] Step 1.3: Construction of the target welding program sample set

[0114] The unlabeled sample set of the target welding program to be identified is represented as follows: .

[0115] in, For the first Candidate measurement feature vectors for each target program sample; Welding procedure for target welding process; The number of samples for the target welding procedure.

[0116] The health status label of the target welding process is not involved in the construction of physical consistency sample weights, feature evaluation, feature ranking, determination of feature quantity, and training of the health status recognition model.

[0117] Step 2: Construction of robust statistical baseline and physical mechanism set under normal conditions

[0118] Step 2.1: Construction of a robust statistical baseline under normal conditions

[0119] From the source welding procedure sample set Extract normal state samples. Let the label of normal state be... .

[0120] For any source welding procedure Its normal state sample index set is represented as: .

[0121] in, For source welding procedures The normal state sample index set, Indicates the first The samples are in a normal state. Indicates the first The sample belongs to the source welding process. .

[0122] For candidate measurement feature set The first in Candidate measurement features Its source welding process The normal state position parameters and scale parameters are as follows: , .

[0123] in, Candidate measurement features In the source welding process Normal state position parameters; These are the corresponding normal state scale parameters; This represents the median; This represents the interquartile range; the coefficient 1.349 is used to convert the interquartile range under normal distribution conditions to a scale similar to the standard deviation. To avoid extremely small positive numbers with a denominator of zero.

[0124] Step 2.2: Construction of the physical mechanism group

[0125] Based on the operating mechanism of the servo resistance spot welding clamp and the main physical characterization function of the candidate measurement features, the candidate measurement features are set. Mapped to A group of physical mechanisms. The mapping function from candidate measurement features to the physical mechanism group is defined as:

[0126] .

[0127] in, Indicates the first Candidate measurement features Mapped to the A group of physical mechanisms, .

[0128] No. The set of candidate measurement features corresponding to each physical mechanism group is represented as follows:

[0129] .

[0130] in, The total number of physical mechanism groups; Number the physical mechanism group; For the first A set of candidate measurement features contained in a physical mechanism group.

[0131] Step 3: Calculation of physical consistency index and sample weight.

[0132] Step 3.1: Calculation of the degree of deviation of candidate measurement features and the intensity of deviation of mechanism group

[0133] For the One sample, whose source welding procedure is identified as... ,in For the first Candidate measurement features Calculate the candidate measurement features of the sample. The standardized deviation from the normal robust statistical baseline of the source welding procedure: .

[0134] in, For the first The first source sample The degree of standardized deviation of each candidate measurement feature relative to the baseline of the normal state of its source welding process; The larger the value, the better the candidate measurement feature. The more significant the deviation from the baseline of the normal state of the welding procedure, the more pronounced the deviation.

[0135] To avoid a few extreme outliers having an excessive impact on sample weights, the standardization deviation is truncated: .

[0136] in, The degree of standardized deviation after truncation. The upper limit is truncated to indicate the degree of deviation.

[0137] No. The set of candidate measurement features corresponding to each physical mechanism group Calculate the first The average deviation of each sample from this physical mechanism group: .

[0138] in, Indicates the first The number of candidate measurement features included in each physical mechanism group. Indicates sample In the Average deviation intensity on each physical mechanism group; Indicate candidate measurement features Belongs to the A group of physical mechanisms.

[0139] Step 3.2: Calculation of Physical Consistency Index

[0140] Taking into account the overall deviation strength of the samples across each physical mechanism group and the degree of coordination between groups, the first... The physical consistency index for each sample is: .

[0141] in, For the first The sample at the th The average deviation intensity on each physical mechanism group The total number of physical mechanism groups; The discrete penalty coefficient for the mechanism group is taken as follows in this embodiment: ; This represents the standard deviation of the deviation strength across each physical mechanism group. The first term reflects the average deviation strength of the sample across all physical mechanism groups, while the second term is used to suppress isolated deviations caused by anomalous fluctuations in a single mechanism group. When the sample exhibits significant and relatively coordinated deviations across multiple related physical mechanism groups, Larger.

[0142] The average physical consistency index for all source welding process samples is: .

[0143] in, This represents the total number of source samples.

[0144] Step 3.3: Generation of Physically Consistent Sample Weights

[0145] The first The physical consistency index of each sample is mapped to the initial sample weights:

[0146] .

[0147] in, For the first Initial physical consistency sample weights for each sample; The sample weight change intensity coefficient. and Let be the lower and upper bounds of the initial sample weights, respectively, and satisfy . In this embodiment, the following values ​​are taken: , , ; This indicates that the calculation result is restricted to an interval. Inside.

[0148] The initial sample weights are normalized to obtain the final physically consistent sample weights: .

[0149] in, The sample index used for summation. For the first The initial physical consistency sample weights of each sample; the normalized sample weights satisfy: .

[0150] Step 4: Evaluation of health contribution and operating condition sensitivity based on physical consistency weighting.

[0151] For the Candidate measurement features Using the normalized physical consistency sample weights obtained in step 3 Calculate their health contribution scores respectively. and operating condition sensitivity score .

[0152] Step 4.1: Physical Consistency Weighted Health Contribution Evaluation

[0153] Calculate candidate measurement features respectively The mutual information between health status labels, multi-class health status effect size, normal and abnormal binary effect size, and feature importance in the health status random forest model were analyzed, and each sub-indicator was normalized to... Interval.

[0154] Among them, the physical consistency sample weight It is used for multi-class health status effect size, normal and abnormal binary effect size, and random forest model training; mutual information is calculated directly from the training samples using the source welding program.

[0155] The physically consistent weighted health contribution score is represented as follows:

[0156] .

[0157] in, Candidate measurement features Normalized mutual information between health status; The normalized multi-class health status effect size after introducing physical consistency sample weights; The normalized normal-abnormal binary effect size after introducing the weights of physically consistent samples; Normalized feature importance obtained after training a healthy random forest model by introducing physically consistent sample weights. , , and The weight coefficients corresponding to mutual information, multi-class health status effect size, normal-abnormal binary effect size, and random forest feature importance, respectively, satisfy: , , The numbering of the health contribution evaluation items is taken as follows in this embodiment: , , , . The larger the value, the better the candidate measurement feature. The stronger the contribution to the judgment of the welding clamp's health status or fault type.

[0158] Step 4.2: Physical Consistency Weighted Operating Condition Sensitivity Evaluation

[0159] To reduce the impact of the coupling between fault category distribution and welding procedure distribution on the evaluation results, the condition sensitivity of candidate measurement features is calculated only using normal state samples from the source welding procedure. For candidate measurement features... The center drift, inter-program distribution differences, and predictability of the physically consistent weighted program were calculated separately, and each sub-indicator was normalized to... Interval.

[0160] The physical consistency weighted operating condition sensitivity score is expressed as follows: .

[0161] in, For the normalized physical consistency weighted program center drift, To represent the differences in Jensen–Shannon distributions between normalized programs, This is a normalized, physically consistent weighted program predictability index. The physical consistency sample weights are used to calculate program center drift and program predictability, while the distribution differences between programs are calculated based on the empirical distribution of normal state samples. , and These are the weighting coefficients corresponding to program center drift, inter-program distribution differences, and program predictability, respectively, and they satisfy the following: , ,in, The numbering for the sub-items of the operating condition sensitivity evaluation is taken as follows in this embodiment: , , . The larger the value, the better the candidate measurement feature. The more sensitive it is to the welding procedure or operating conditions.

[0162] Step 5: Feature Type Determination

[0163] Let the health contribution score threshold be... The threshold for operating condition sensitivity scoring is ,in, , Based on candidate measurement features Health Contribution Score and operating condition sensitivity score Candidate measurement features are divided into the following four categories:

[0164] Category A: and This contributes to high health by low working condition sensitivity characteristics;

[0165] Category B: and This contributes to high health and high working condition sensitivity characteristics;

[0166] Class C: and This contributes to low-health, low-condition-sensitive characteristics;

[0167] Category D: and This contributes to the high-condition sensitivity characteristics of low health.

[0168] Features of type A have a high ranking priority; although features of type B have a high sensitivity to working conditions, they still contain strong health discrimination information, so they are retained and given the second highest ranking priority. In the evaluation of candidate feature subsets and the health status identification stage, robust baseline correction is performed on the selected features of type B; features of type C and D are not directly deleted, but are given a lower type priority, and whether they enter the final health feature subset is determined by the subsequent comprehensive ranking results and the number of features.

[0169] Step 6: Robust baseline correction for Class B features.

[0170] For the Class B features identified in step 5, their health contribution score, operating condition sensitivity score, and overall ranking position are retained. During the candidate feature subset evaluation and health status identification stage, robust baseline correction is performed only on Class B features within the selected subset.

[0171] For the The first sample There are 1 Class B feature, and the corrected feature values ​​are: .

[0172] in, Indicates sample The welding procedure identifier. and These represent the robust position and scale parameters for the corresponding normal state samples of the program. For the internal verification program and the target welding program, the corresponding parameters are estimated using the median and interquartile range of the unlabeled samples of the program.

[0173] The correction only applies to the selected B-class features; the A, C, and D-class features retain their original values. The correction results are used for subsequent internal validation and training and testing of the health status recognition model, and do not participate in feature scoring and overall ranking again.

[0174] Step 7: Comprehensive sorting, determination of feature quantity, and health status identification.

[0175] Step 7.1: Comprehensive ranking of candidate measurement features

[0176] The candidate measurement features of categories A, B, C, and D determined in step 5 are collectively used to form a feature set to be ranked. For the first... Candidate measurement features A comprehensive ranking score is calculated based on the health contribution score, working condition sensitivity score, and feature type priority:

[0177] .

[0178] Prioritize feature types; and These are the weighting coefficients for the health contribution item and the working condition stability item, respectively.

[0179] Feature type priority is represented as follows: ,

[0180] in, Candidate measurement features The characteristic type; , , and These are the type priority coefficients for features of types A, B, C, and D, respectively, and they satisfy the following: In this embodiment, the following values ​​are taken: , , , , , .

[0181] Step 7.2: Determining the number of features

[0182] Based on comprehensive ranking score Arrange the candidate measurement features from highest to lowest to obtain a sequence. Set the set of candidate feature counts. In this embodiment, the following values ​​are taken: For any Select before sorting Each candidate measurement feature constitutes a subset of candidate features. .

[0183] Within the source welding program, a leave-one-out verification process is performed. Each time, one source welding program is used as the internal verification program, and the remaining source welding programs are used as internal training programs. The feature scores, types, and rankings obtained in steps 4, 5, and 7.1 are kept unchanged, and the candidate feature subsets are processed according to step 6. Robust baseline correction was performed on the B-type features in the model, and then the health status recognition model was trained and its recognition performance was evaluated.

[0184] Candidate feature subset The internal validation composite index is expressed as:

[0185] in, For the source welding program set, Indicates the source welding procedure. Indicates the set of source welding procedures The cardinality, i.e., the number of internal verification procedures. , , and These represent the source welding process. As an internal verification procedure, it employs a subset of candidate features. The macro-average F1 score, Matthews correlation coefficient, area under the precision-recall curve, and area under the receiver operating characteristic curve were obtained at that time. , , and These are the weighting coefficients for the macro-average F1 score, Matthews correlation coefficient, area under the precision-recall curve, and area under the receiver operating characteristic curve, respectively, and they satisfy the following: And satisfy: , The internal validation evaluation index is numbered; in this embodiment, it is taken as: , , , .

[0186] The optimal number of features is determined based on the comprehensive internal validation metrics:

[0187] Select before sorting These features constitute a subset of the final health characteristics: .

[0188] Step 7.3: Health Status Identification of the Target Welding Program

[0189] For the subset of final health characteristics The B-type features in the model are subjected to robust baseline correction in step 6; the health status recognition model is trained using all source welding program samples, and the health status of the target welding program samples is then recognized.

[0190] To verify the applicability of the method of this invention in the absence of known welding programs, welding programs 7, 31, 32, and 33 were used as target welding programs, while the remaining programs were used as source welding programs. Feature evaluation, type identification, comprehensive ranking, determination of the number of features, and model training were all completed using only the source program samples, and the target program labels were only used for the final test result evaluation.

[0191] Figure 2 The distribution of candidate measurement features in the two-dimensional space of health contribution-condition sensitivity is shown under different target welding procedures. Figure 2 (a) Figure 2 (b) Figure 2 (c) and Figure 2 (d) Corresponding to target welding procedures 7, 31, 32, and 33 respectively; the horizontal axis represents the working condition sensitivity score, and the vertical axis represents the health contribution score. Based on the health contribution score threshold and the working condition sensitivity score threshold, the candidate measurement features are divided into categories A, B, C, and D, and the outer circle indicates the features selected into the final health feature subset by the method of this invention.

[0192] After determining the final subset of health features, health status identification tests were performed on target welding procedures 7, 31, 32, and 33, respectively. The results are as follows: Figure 3 As shown. Among them, Figure 3 (a) Figure 3 (b) Figure 3 (c) and Figure 3 (d) correspond to target welding procedures 7, 31, 32 and 33 respectively. Figure 3 In the table, indicator 1 represents precision, indicator 2 represents the harmonic mean of macro-precision and recall (Macro-F1), indicator 3 represents the Matthews correlation coefficient (MCC), indicator 4 represents the area under the receiver operating characteristic (ROC-AUC), and indicator 5 represents the area under the precision-recall curve (PR-AUC).

[0193] Figure 3 The results show that the method of the present invention has an accuracy of 0.92 to 0.99 under four target welding procedures, a harmonic mean of macro-average precision and recall of 0.87 to 0.97, a Matthews correlation coefficient of 0.78 to 0.95, an area under the receiver operating characteristic curve of 0.98 to 1.00, and an area under the precision-recall curve of 0.94 to 0.99.

[0194] The above results demonstrate that the present invention can reduce the impact of welding procedure differences on feature evaluation and health status identification while retaining effective health information, and maintain good identification performance even when the target welding procedure is not observed.

Claims

1. A method for selecting welding clamp health characteristics based on physical consistency weighting, characterized in that, Includes the following steps: S1, acquire source samples with health status tags under multiple source welding programs using the servo resistance spot welding clamp, and untagged target samples under the target welding program to be identified; denote the set of source welding programs as... The total number of source code samples is denoted as ; S2, based on the normal state samples of each source welding process, establish a normal state robust statistical baseline for candidate measurement features, and map the candidate measurement features to the corresponding physical mechanism group according to the physical characterization effect of the candidate measurement features. S3, based on the standardized deviation of the source sample from the normal state robust statistical baseline of its respective source welding procedure, the deviation strength of each physical mechanism group, and the dispersion between different physical mechanism groups, determine the physical consistency sample weight of the source sample; wherein, step S3 specifically includes: For the One sample, whose source welding procedure is identified as... ,in For the first Candidate measurement features Calculate the candidate measurement features of the sample. The standardized deviation from the normal robust statistical baseline of the source welding procedure: ;in, For the first The source samples at the th Candidate measurement features The measured value on; and Candidate measurement features Normal state position parameters and normal state scale parameters under the source welding procedure; The larger the value, the better the candidate measurement feature. The more significant the deviation from the baseline of the normal state of the welding procedure, the more pronounced the deviation. To avoid a few extreme outliers having an excessive impact on the sample weights, the standardized deviation is truncated. ; in, The degree of standardized deviation after truncation. The upper limit is truncated to indicate the degree of deviation; the first The set of candidate measurement features corresponding to each physical mechanism group Calculate the first The average deviation of each sample from this physical mechanism group: ;in, Indicates the first The number of candidate measurement features included in each physical mechanism group. Indicates sample In the Average deviation intensity on each physical mechanism group; Indicate candidate measurement features Belongs to the A physical mechanism group; considering the overall deviation strength of the sample across each physical mechanism group and the degree of coordination between groups, the first physical mechanism group is defined. The physical consistency index for each sample is: ;in, For the first The sample at the th The average deviation intensity on each physical mechanism group The total number of physical mechanism groups; The discrete penalty coefficients for the mechanism group. The standard deviation of the deviation intensity for each physical mechanism group; the mean of the physical consistency index for all source samples is: ,in, The total number of source samples; the first... The physical consistency index of each sample is mapped to the initial sample weights. ;in, For the first Initial physical consistency sample weights for each sample; The sample weight change intensity coefficient. and Let be the lower and upper bounds of the initial sample weights, respectively, and satisfy . ; This indicates that the calculation result is restricted to an interval. Internally, the initial sample weights are normalized to obtain the final physically consistent sample weights. ;in, The sample index used for summation. For the first The initial physical consistency sample weights of each sample; the normalized sample weights satisfy: ; S4. Using the physical consistency sample weights, calculate the health contribution score and working condition sensitivity score for each candidate measurement feature respectively. S5. Based on the health contribution scoring threshold and the working condition sensitivity scoring threshold, each candidate measurement feature is divided into category A, category B, category C, and category D features. S6, calculate a comprehensive ranking score based on the feature type, health contribution score, and working condition sensitivity score of the candidate measurement features, and arrange the candidate measurement features from high to low according to the comprehensive ranking score; set the set of candidate feature quantities. ,for Select before sorting Each candidate measurement feature constitutes a subset of candidate features. ; S7. For each candidate feature subset, robust baseline correction is applied only to the selected Class B features within the subset, while Class A, C, and D features retain their original values. The corrected candidate feature subset is used to train the health status recognition model, and its recognition performance on the internal validation procedure is evaluated. The optimal number of features is determined based on the comprehensive internal validation index of each candidate feature subset. Before comprehensive sorting The candidate measurement features were determined as a subset of the final health features; S8. Robust baseline correction is performed on the B-type features in the final health feature subset. The health status recognition model is trained using all source welding program samples. The trained health status recognition model is then used to identify the health status of the target welding program sample.

2. The method for selecting welding clamp health characteristics based on physical consistency weighting according to claim 1, characterized in that, In step S2, the source welding program sample set is represented as: The candidate measurement feature set is represented as: ;in, A collection of source welding procedures; This is a sample set of source welding programs; The total number of source samples; For the first Candidate measurement feature vectors for each sample For the first One candidate measurement feature, For the first Each sample in candidate measurement features The measured value on The number of candidate measurement features, ; For the first Health status labels for each sample; For the first The welding procedure identifier to which each source sample belongs; for the source welding procedure Its normal state sample index set is represented as: ;in, For source welding procedures The normal state sample index set, Indicates the first All samples are in a normal state; Indicates the first The sample belongs to the source welding process. For candidate measurement feature sets The first in Candidate measurement features Its source welding process The normal state position parameters and scale parameters are as follows: , ;in, Candidate measurement features In the source welding process Normal state position parameters; These are the corresponding normal state scale parameters; This represents the median; This represents the interquartile range; the coefficient 1.349 is used to convert the interquartile range under normal distribution conditions to a scale similar to the standard deviation. To avoid extremely small positive numbers with a denominator of zero, the mapping function from candidate measurement features to the physical mechanism set is defined as follows: ;in, Indicates the first Candidate measurement features Classified to the A group of physical mechanisms, ;No. The set of candidate measurement features corresponding to each physical mechanism group is represented as follows: ;in, The total number of physical mechanism groups; Number the physical mechanism group; For the first A set of candidate measurement features contained in a physical mechanism group.

3. The method for selecting welding clamp health characteristics based on physical consistency weighting according to claim 1, characterized in that, In step S4, the health contribution score is obtained by weighting the mutual information between candidate measurement features and health status labels, the multi-class health status effect size, the normal and abnormal binary effect size, and the feature importance in the health status identification model. ,in, Candidate measurement features Normalized mutual information between health status; The normalized multi-class health status effect size after introducing physical consistency sample weights; The normalized normal-abnormal binary effect size after introducing the weights of physically consistent samples; Normalized feature importance obtained after training a healthy random forest model by introducing physically consistent sample weights; , , and The weight coefficients corresponding to mutual information, multi-class health status effect size, normal-abnormal binary effect size, and random forest feature importance, respectively, satisfy: , , The health contribution evaluation item is numbered; the operating condition sensitivity score is obtained by weighting the physical consistency weighted program center drift, inter-program distribution differences, and physical consistency weighted program predictability. ,in, For the normalized physical consistency weighted program center drift, To represent the differences in Jensen–Shannon distributions between normalized programs, The normalized physical consistency weighted program predictability index; the physical consistency sample weight is used to calculate program center drift and program predictability, and the distribution difference between programs is calculated based on the empirical distribution of normal state samples. , and These are the weighting coefficients corresponding to program center drift, inter-program distribution differences, and program predictability, respectively, and they satisfy the following: , ,in, This is the number for the sub-item of the operating condition sensitivity evaluation.

4. The method for selecting welding clamp health characteristics based on physical consistency weighting according to claim 3, characterized in that, In step S5, the health contribution score threshold and the working condition sensitivity score threshold are respectively: , Candidate measurement features are divided into the following four categories: Category A: and Contributing low-condition sensitivity characteristics to high health; Category B: and Contributing high-condition sensitivity characteristics to high health; Category C: and Category D: Contributes low-condition sensitivity characteristics to low health; and This contributes to the high-condition sensitivity characteristics of low health.

5. The method for selecting welding clamp health characteristics based on physical consistency weighting according to claim 4, characterized in that, In step S6, the comprehensive ranking score of the candidate measurement features is: , Priority based on feature type; and These are the weight coefficients for the health contribution item and the operating condition stability item, respectively. The feature type priority is represented as follows: ,in, Candidate measurement features The characteristic type; , , and These are the type priority coefficients corresponding to features of types A, B, C, and D, respectively, and they satisfy the following: .

6. The method for selecting welding clamp health characteristics based on physical consistency weighting according to claim 2, characterized in that, In steps S7 and S8, for the first The sample corresponds to the selected candidate feature subset, and the final health feature subset is the th sample. There are 1 Class B feature, and the corrected feature values ​​are: ,in, Indicates sample The welding procedure identifier; and These are the robust position and scale parameters for the normal state samples of the corresponding program, respectively; for the internal verification program and the target welding program, the corresponding parameters are estimated using the median and interquartile range of the unlabeled samples of the program.

7. The method for selecting welding clamp health characteristics based on physical consistency weighting according to claim 1, characterized in that, In step S7, during the leave-one-out-of-program verification within the source welding program, one source welding program is used as the internal verification program each time, while the remaining source welding programs are used as internal training programs. The health contribution score, operating condition sensitivity score, feature type, and overall ranking remain unchanged. (Candidate feature subset) The internal validation composite index is expressed as: in, For the source welding program set, Indicates the source welding procedure. Represents a set The base number, i.e., the number of internal verification procedures; , , and These represent the source welding process. As an internal verification procedure, it employs a subset of candidate features. The macro-average F1 score, Matthews correlation coefficient, area under the precision-recall curve, and area under the receiver operating characteristic curve were obtained at that time. , , and These are the weighting coefficients for the macro-average F1 score, Matthews correlation coefficient, area under the precision-recall curve, and area under the receiver operating characteristic curve, respectively, and they satisfy the following: And satisfy: , The internal validation evaluation index number; optimal number of features: Select before sorting These features constitute a subset of the final health characteristics: .

Images