A welding gun abnormality detection method based on multi-physical chain consistency modeling

Abstract

The present application relates to a kind of welding tongs abnormality detection method based on multi-physical chain consistency modeling, belong to industrial welding quality monitoring and equipment health management technical field.The method is by collecting welding current, voltage, electrode pressure, cooling water flow, transformer temperature, process program number, electrode dressing frequency and other multi-source process signals, respectively construct the regression baseline model of heating chain, cooling chain and contact chain three physical links in only using normal sample, obtain the expected value of energy input, heat dissipation capacity and interface contact state;Further calculate the residual of each link, and group z-score standardization and symmetric truncation according to working condition, obtain three chain standardization residual vector;On this basis, the covariance matrix of residual vector is estimated, the Mahalanobis distance is used as multi-chain physical consistency deviation degree, the determination of abnormal welding point is realized, and the physical explanation of abnormal source is given through chain level radar chart.

Description

Technical Field

[0001] This invention belongs to the field of welding equipment health monitoring and intelligent manufacturing technology, specifically involving a welding clamp anomaly detection method based on multi-physics chain consistency modeling. Background Technology

[0002] Resistance spot welding is a critical joining process in the mass production of high-end equipment such as automotive body-in-white and aerospace structural components. On these automated production lines, the welding gun, as the core execution unit, simultaneously undertakes multiple functions including high-current conduction, high-pressure loading, and heat management. The health of the equipment itself, such as electrode wear, cooling pipe blockage, contact degradation of conductive components, and overheating of rectifier elements, directly determines the stability of welding quality, process consistency, and overall equipment utilization. With the continuous increase in production line cycle time and the extension of continuous operating cycles, the aforementioned equipment-level performance degradation often exhibits latent and gradual characteristics, gradually becoming a key risk source for quality defects such as electrode adhesion, incomplete welds, and spatter, as well as unplanned downtime.

[0003] Currently, industrial on-site monitoring of the welding process mainly focuses on determining the weld quality. Common techniques include empirical threshold alarms for welding current / voltage waveforms, statistical monitoring of welding energy, and statistical process control or simple machine learning models based on single signals. However, these result-oriented monitoring methods have significant limitations when dealing with abnormalities in the welding clamp's condition:

[0004] First, the fault differentiation is insufficient. Abnormalities in a single current, voltage, or energy parameter make it difficult to effectively distinguish whether the root cause of the fault is a decrease in cooling system efficiency, deterioration of electrode-workpiece interface contact, or power system fluctuations, resulting in ambiguous alarm information and poor maintenance guidance.

[0005] Secondly, the mechanistic coupling is neglected. The welding process is essentially a process of tight coupling of multiple physical fields, including electricity, heat, and mechanics. Most existing methods have not established a quantitative model of this inherent coupling relationship, and therefore cannot characterize the propagation and manifestation of anomalies in the energy input, transfer, and dissipation links, thus monitoring remains superficial.

[0006] Finally, health assessments are limited to a single dimension. Traditional methods typically set thresholds for a single indicator, lacking a multi-dimensional and synergistic assessment of the overall health of the welding clamp, and thus failing to capture early and comprehensive performance degradation caused by the interplay of multiple factors.

[0007] To overcome the above limitations and realize the transformation from quality result monitoring to equipment anomaly detection, a monitoring method is needed that can integrate the working mechanism of welding clamps, adapt to changing working conditions, and accurately distinguish and locate multiple types of equipment-level anomalies. Summary of the Invention

[0008] The application aims to provide a welding tong abnormality detection method based on multi-physical chain consistency modeling, so as to solve the problems of inaccurate device-level abnormality recognition, difficult fault root differentiation and lack of physical interpretability caused by relying on a single signal and ignoring the electric-thermal-mechanical coupling mechanism in current welding monitoring.

[0009] To achieve the above-mentioned purpose, the core of the technical solution adopted by the application is to construct three physical chain link reference models of the resistance spot welding tong heating chain, the cooling chain and the contact chain, to form a multi-level abnormality recognition and quantification framework from local physical mechanism to system overall coordination through intra-chain consistency residual analysis and inter-chain Mahalanobis Distance measurement.

[0010] The method can realize multi-level abnormality recognition from local mechanism to system coordination, has clear physical explanation, has high adaptability and robustness, and significantly improves the accuracy of resistance spot welding tong device-level health monitoring and welding quality control.

[0011] The application adopts the following technical solutions:

[0012] A welding tong abnormality detection method based on multi-physical chain consistency modeling, comprising data acquisition and preprocessing, three-chain physical link modeling, intra-chain residual calculation and working condition adaptive standardization, inter-chain covariance estimation and multi-chain consistency scoring, abnormality judgment and performance evaluation. The specific contents are as follows.

[0013] Step 1: Data acquisition and preprocessing

[0014] Acquire multi-source sensor data of the welding process, including welding current, voltage, electrode pressure, cooling water flow, transformer temperature, process program number and electrode dressing frequency; clean, align and extract features from the data, calculate resistance derived features to form a sample data set.

[0015] Step 2: Three-chain physical link modeling, the welding process is divided into heating chain model, cooling chain model and contact chain model.

[0016] Step 2.1: Heating chain modeling, energy input is regarded as a nonlinear function of multi-source process variables, and its steady-state mapping relationship is learned through data-driven method, so as to obtain stable energy input reference under complex working conditions: wherein, is the average welding current, is the average welding voltage, related to instantaneous electric power; is the electrode pressure, affecting the contact area and contact resistance of the welding spot; is the cooling water flow, affecting the local temperature field of the electrode and the conductive arm; is the transformer temperature, used to represent the thermal state of the welding tong; electrode grinding times, representing the electrode wear degree; process program number, used to distinguish the system dynamics under different welding modes.

[0017] Step 2.2: Cooling chain modeling, modeling the transformer temperature as a function of energy input and cooling conditions: where, single-point welding energy input, cooling water flow, electrode grinding times, representing the electrode wear degree; process program number, used to distinguish the system dynamics under different welding modes.

[0018] Step 2.3: Contact chain modeling, modeling the resistance as a nonlinear mapping relationship of multiple source process variables: where, welding average current, welding average voltage, related to instantaneous electric power; electrode pressure, affecting the contact area and contact resistance of the welding spot; single-point welding energy input, cooling water flow, affecting the local temperature field of the electrode and the conductive arm; transformer temperature, used to represent the thermal state of the welding gun; electrode grinding times, representing the electrode wear degree; process program number, used to distinguish the system dynamics under different welding modes.

[0019] Step 2.4: Three-chain physical link model construction, for each physical chain (heat chain / cooling chain / contact chain), define the mapping function within the chain: where, is the original input feature vector of the chain, is the model parameter. The training of the model only uses the normal welding spot sample data set to ensure that the learned is the healthy baseline mapping, so that the residual error of the observation-prediction value can reflect the abnormal deviation. The training target adopts the mean square error (MSE) form: , the base learner selects the random forest regressor, and combines the pre-processing and the regressor into a unified pipeline: .

[0020] Step 3: In-chain residual calculation and working condition adaptive standardization

[0021] To quantify the intra-chain bias and make it adaptive to the condition variation (e.g. different program_id), the invention first defines the intra-chain residual as the difference between the observation and the intra-chain baseline prediction, then normalizes and truncates the residual under the condition, and finally gets the normalized residual matrix for inter-chain covariance estimation.

[0022] For sample , and any physical chain , let the observation of the chain be , and the intra-chain baseline prediction be , then the intra-chain residual is defined as: The residuals of the three chains form the residual vector: Considering that different conditions / programs (denoted as corresponding to program_id) will cause systematic bias, the intra-chain residual mean and standard deviation are estimated separately for each set of normal samples : , To avoid numerical instability or division by zero, , .

[0023] For sample , perform z-score normalization under its belonging condition : To suppress the impact of extreme values on subsequent covariance estimation, symmetric truncation is used: where , and the normalized residual vector of each sample is obtained by combining: .

[0024] Step 4: Inter-chain covariance estimation and multi-chain consistency score

[0025] Stack all the normal training samples' residual vectors by column or by row to form the matrix This matrix is the input for inter-chain covariance estimation: Based on , estimate the covariance of the three-chain residuals To improve robustness in small sample or noisy scenarios, the invention provides two strategies: shrinkage estimation: use Ledoit-Wolf shrinkage estimator to get , denote its output as the covariance matrix and precision matrix (inverse covariance): , , .

[0026] If shrinkage estimation fails or returns a non-finite value, fall back to sample covariance.

[0027] Sample Covariance and Regularization: Compute sample covariance with diagonal regularization to ensure invertibility: where, is a small positive number, is the identity matrix.

[0028] Given samples of standardized residual vectors and the estimated covariance inverse (precision matrix) , define its Mahalanobis Distance (multi-chain consistency measure) as: .

[0029] The Mahalanobis Distance takes into account both the magnitude of the residual of each chain and the correlation between chains (off-diagonal elements of the covariance matrix). When two chains deviate simultaneously but are highly correlated with each other, the Mahalanobis Distance corrects the information redundancy of their joint anomaly; when the deviation occurs in a direction unrelated to other chains, the distance can more sensitively reflect the anomaly of the individual chain.

[0030] Step 5: Anomaly Judgment and Performance Evaluation

[0031] Using the area under the receiver operating characteristic curve (ROC-AUC), the area under the precision-recall curve (PR-AUC), false alarm rate (FAR), and missed alarm rate (MDR) as indicators, the abnormal welding points are judged, and the radar chart is used for anomaly explainable analysis.

[0032] Compared with the prior art, the beneficial effects are: the welding tongs anomaly detection method based on multi-physical chain consistency modeling provided by the present application has the following remarkable advantages: the heating chain, cooling chain and contact chain models constructed by the present application have clear physical meaning. The anomaly not only reflects the excessive final score, but also can be traced back to the chain link, providing direct and clear diagnostic direction for on-site maintenance, overcoming the problem of unknown alarm reason of black box model. The working condition grouping standardization strategy is introduced to eliminate the system deviation between different welding programs, and adapt to the multi-line and multi-variety production scene. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is the block diagram of the method of the present application;

[0034] Figure 2 is the anomaly point detection result (the first 10 anomaly points) of the method of the present application;

[0035] Figure 3 is the three-chain anomaly contribution analysis result. DETAILED DESCRIPTION

[0036] The specific embodiments of the present application are further described below with reference to the accompanying drawings. It should be understood that the following examples are only used to illustrate the principles of the present application and do not limit the protection scope of the present application.

[0037] Example 1: A welding gun abnormality detection method based on multi-physical chain consistency modeling

[0038] This embodiment selects the welding point data collected in the automobile body-in-white welding production line to verify the welding gun abnormality detection method based on multi-physical chain consistency modeling proposed by the present application. As shown in the figure, the welding gun abnormality detection method based on multi-physical chain consistency modeling includes five steps of data acquisition and preprocessing, three-chain physical link modeling, intra-chain residual calculation and working condition adaptive standardization, inter-chain covariance estimation and multi-chain consistency scoring, abnormality judgment and performance evaluation, and the specific detailed steps are as follows: Figure 1

[0039] Step 1: Data acquisition and preprocessing

[0040] Collect multi-source sensor data of the welding process, including welding current, voltage, electrode pressure, cooling water flow, transformer temperature, process program number and electrode grinding frequency; clean, align and extract features from the data, calculate resistance derived features to form a sample data set.

[0041] Step 2: Three-chain physical link modeling, which divides the welding process into a heating chain model, a cooling chain model and a contact chain model.

[0042] Step 2.1: Heating chain modeling, energy input is regarded as a nonlinear function of multi-source process variables, and its steady-state mapping relationship is learned through data-driven method, so as to obtain stable energy input reference under complex working conditions: wherein, is the welding average current, welding average voltage, related to instantaneous electric power; electrode pressure, affecting the contact area and contact resistance of the welding point; cooling water flow, affecting the local temperature field of the electrode and the conducting arm; transformer temperature, used to represent the thermal state of the welding gun; electrode grinding frequency, representing the degree of electrode wear; process program number, used to distinguish the system dynamics under different welding modes.

[0043] Step 2.2: Cooling chain modeling, modeling the transformer temperature as a function of energy input and cooling conditions: wherein, is the single-point welding energy input, cooling water flow, ​Electrode grinding times, representing the degree of electrode wear; Process program number, used to distinguish system dynamics under different welding modes.

[0044] Step 2.3: Contact chain modeling, modeling the resistance as a nonlinear mapping relationship of multiple source process variables: Wherein, is the average welding current, is the average welding voltage, related to the instantaneous electric power; is the electrode pressure, affecting the contact area and contact resistance of the welding spot; is the single-point welding energy input, is the cooling water flow, affecting the local temperature field of the electrode and the conductive arm; is the transformer temperature, used to represent the thermal state of the welding gun; Electrode grinding times, representing the degree of electrode wear; Process program number, used to distinguish system dynamics under different welding modes.

[0045] Step 2.4: Three-chain physical link model construction, for each physical link (heating chain / cooling chain / contact chain), define the mapping function within the chain: Wherein, is the original input feature vector of the chain, is the model parameter. The training of the model only uses the normal welding spot sample data set to ensure that the learned is the healthy baseline mapping, so that the residual error of the observed value-predicted value can reflect the abnormal deviation. The training target adopts the mean square error (MSE) form: , the base learner selects the random forest regressor, and combines the pretreatment with the regressor into a unified pipeline (pipeline): .

[0046] Step 3: In-chain residual calculation and working condition adaptive standardization

[0047] In order to quantify the in-chain deviation and make it adaptive to working condition changes (such as different program_id), the invention first defines the in-chain residual as the difference between the observed value and the in-chain baseline predicted value, then standardizes and truncates the residual under the working condition, and finally obtains the standardized residual matrix used for inter-chain covariance estimation.

[0048] For a sample , and any physical chain , let the observed quantity of the chain be , and the in-chain baseline prediction be , then the in-chain residual is defined as: The residuals of the three chains form a residual vector: , considering different working conditions / programs (denoted as , which will cause systematic bias, estimate the mean and standard deviation of the intra-chain residuals for each set of normal samples separately: , , to avoid numerical instability or division by zero, , .

[0049] The samples are z-score standardized under their respective working conditions : , to suppress the impact of extreme values on subsequent covariance estimation, symmetric truncation is used: where , the standardized residual vector of each sample is obtained by combining: .

[0050] Step 4: Inter-chain covariance estimation and multi-chain consistency score

[0051] Stack all the normal training samples' standardized residuals by column or by row to form a matrix : This matrix is the input for inter-chain covariance estimation: , based on estimate the covariance of the three-chain residuals , to improve robustness in small sample or noisy scenarios, the invention provides two strategies: shrinkage estimation: use Ledoit-Wolf shrinkage estimator to obtain , denote its output as the covariance matrix and precision matrix (inverse covariance): , .

[0052] If shrinkage estimation fails or returns a non-finite value, fall back to sample covariance.

[0053] Sample covariance and regularization: calculate the sample covariance and add diagonal regularization to ensure invertibility: where is a small positive number, is the identity matrix.

[0054] Given the standardized residual vector of sample and the estimated covariance inverse (precision matrix) , define its Mahalanobis distance (multi-chain consistency measure) as: .

[0055] The Mahalanobis distance considers both the amplitude of the residual of each chain and the correlation between chains (non-diagonal elements of the covariance matrix). When two chains deviate simultaneously but are highly correlated with each other, the Mahalanobis distance corrects the information redundancy of their joint anomaly; when the deviation occurs in a direction unrelated to other chains, the distance can more sensitively reflect the anomaly of the individual chain.

[0056] Step 5: Abnormality judgment and performance evaluation

[0057] The area under the receiver operating characteristic curve (ROC-AUC), the area under the precision-recall curve (PR-AUC), the false alarm rate (FAR) and the missed alarm rate (MDR) are used to realize the judgment of abnormal welding points, and the radar chart is used for abnormal explainable analysis.

[0058] In order to verify the advantages of the method, the autoencoder (AE), long short-term memory network (LSTM), convolutional neural network (CNN), isolation forest (IF), support vector machine (SVM), local outlier factor (LOF), K-means clustering (K-Means), and graph neural network (GNN) methods are compared, and the results are shown in Table 1.

[0059] Table 1 Comparison results

[0060] Method name ROC-AUC PR-AUC FAR MDR AE 0.8773 0.0808 0.0674 0.5385 LSTM 0.8919 0.0874 0.1393 0.1538 CNN 0.8648 0.4806 0.1720 0.1759 IF 0.7527 0.3080 0.2652 0.3518 SVM 0.6453 0.2564 0.3781 0.3981 LOF 0.8486 0.4749 0.1702 0.2962 K-Means 0.6909 0.2403 0.5752 0.0648 GNN 0.6944 0.2430 0.1426 0.5238 The method 0.9904 0.6825 0.0122 0.0769

[0061] As can be seen from Table 1, the method of the present application is superior to the comparative algorithm in various indicators. The ROC-AUC of the method is 0.9904, which is significantly higher than all deep learning baselines (such as 0.8773 of AE, 0.8919 of LSTM, 0.8648 of CNN) and traditional methods (IF, SVM, LOF, K-Means, etc.), indicating that the method has the strongest ability to distinguish normal and abnormal samples under different thresholds. In the PR-AU index which can better reflect the performance of class imbalance, the method reaches 0.6825, which is also significantly improved compared with the better-performing CNN (0.4806) and LOF (0.4749), indicating that the method can still maintain stable detection ability in the scene where the actual abnormality ratio is low. In the FAR index reflecting the false alarm rate, the method is 0.0122. At the same time, the MDR is controlled at 0.0769, which is much lower than most baseline methods (such as 0.5385 of AE, 0.3518 of IF, 0.2962 of LOF, etc.), reflecting the more balanced and acceptable control ability of false alarms and omissions in the engineering field.

[0062] Figure 2 The first 10 abnormal points of the abnormal point detection result of the method of the present application are shown in the radar chart of FIG. 6, and the chain-level explainability analysis of the first 10 detected abnormal welding points is shown in FIG. 7. Figure 3 As shown in the figure, the dashed line represents the contribution of the heat chain, the solid line represents the contribution of the cooling chain, and the dotted line represents the contribution of the contact chain. The values are all normalized to the interval [0, 1], and the larger the value, the more significant the contribution of the chain to the anomaly. From the figure, it can be observed that the cooling chain (solid line) has the highest overall value and the smallest fluctuation, indicating that most of the abnormal welding points are related to the decline in heat dissipation capacity and insufficient cooling, which is consistent with the engineering phenomenon that the fluctuation of cooling water flow and the blockage of the cooling circuit can easily lead to abnormal temperature rise of the welding point and even spatter failure in actual production. The heat chain (dashed line) and the cooling chain have basically the same trend, indicating that when the cooling capacity decreases, the welding current and heat accumulation jointly act on the cooling chain, causing the imbalance of the cooling chain, thereby forming a heat-cooling collaborative abnormal behavior, verifying the rationality of the physical chain coupling mechanism established by the present application. The contact chain (dotted line) has a relatively low overall contribution and a large fluctuation, and only a slight rise occurs in a small number of welding points (such as the first and fourth abnormal points). This indicates that most of the abnormalities are not caused by insufficient electrode pressure or sudden change of contact resistance, but are dominated by the imbalance of the heat chain and the cooling chain, further proving that the proposed abnormality detection method can distinguish between heat / cooling instability and contact degradation of two different failure mechanisms.

[0063] In summary, the radar chart verifies that the three-chain model of the present application not only can realize abnormality detection, but also can realize physical-level explanation of the source of the abnormality, providing clear mechanism support and decision basis for welding process monitoring.

Claims

1. A welding torch abnormality detection method based on multi-physical chain consistency modeling, characterized by, Comprising the following steps: S1: based on a normal welding spot sample data set , respectively, a heating chain model, a cooling chain model and a contact chain model of the resistance spot welding clamp are constructed and trained, the three physical link models are respectively used to represent the benchmark mapping relationship of energy input, heat dissipation ability and interface contact state in the normal condition during the welding process; S2: for , respectively, input into the three physical link models of the heating chain model, the cooling chain model and the contact chain model, obtain the predicted value of each link, and calculate the residual between the observed value and the predicted value of each link; S3: based on the welding process program, the residuals of are counted, and the residuals of the are standardized and truncated based on the statistical quantity of the process program to obtain a standardized residual vector; S4: based on the standardized residual vector of , estimate the covariance matrix and its inverse matrix representing the cooperative relationship between the residuals of the three physical link models of the heating chain model, the cooling chain model and the contact chain model; S5: calculate the Mahalanobis distance of the standardized residual vector of the relative to the covariance matrix as a multi-chain consistency score; S6: according to the multi-chain consistency score result, determine whether the welding clamp state corresponding to the is abnormal.

2. The welding torch anomaly detection method based on multi-physical chain consistency modeling according to claim 1, wherein, In step S1, the input features of the heat generation chain model include welding average current, welding average voltage, electrode pressure, cooling water flow, transformer temperature, electrode dressing frequency, and process program number: ; Input characteristics of the cooling chain model include welding energy input, cooling water flow rate, electrode dressing frequency, and process program number: ; The input features of the contact chain model include welding average current, welding average voltage, electrode pressure, welding energy input, transformer temperature, cooling water flow, electrode dressing frequency, and process program number: .

3. The welding torch anomaly detection method based on multi-physical chain consistency modeling according to claim 2, characterized in that, In step S1, for each physical chain: wherein, is the heating chain, is the cooling chain, is the contact chain; define the mapping function within the chain: wherein, is the original input feature vector of the chain, is the model parameter, the training of the model only uses the normal solder sample data set to ensure that the learned is a healthy baseline mapping, so that the residual error of the observation-prediction value can reflect the abnormal deviation; the training target adopts the mean square error (MSE) form: ; the base learner selects a random forest regressor, and the preprocessing and the regressor are combined into a unified pipeline: .

4. The welding torch anomaly detection method based on multi-physical chain consistency modeling of claim 1, wherein, In step S2, the predicted value of each link is obtained, and the residual between the observed value and the predicted value of each link is calculated; the sample , and the physical chain: , let the observed value of the chain be , and the baseline prediction of the chain be ; then the intra-chain residual is defined as: The residuals of the heating chain model, the cooling chain model and the contact chain model constitute a residual vector, .

5. The method of claim 1, wherein the method is based on a multi-physical consistency modeling of the welding torch. In step S3, the standardization is z-score standardization, i.e. using the mean and standard deviation of the residuals of the process group to which the process belongs for the standardization calculation, taking into account different process conditions , which can cause systematic deviations, the mean and standard deviation of the in-chain residuals are estimated separately for each set of normal samples , , , in order to avoid numerical instability or division by zero, , , the samples are z-score standardized under their own process condition , .

6. The method of claim 1, wherein the method is based on a multi-physical consistency modeling of the welding torch. In step S3, the truncation processing is to limit the standardized residual in a preset symmetric interval, and a symmetric truncation is adopted to suppress the influence of extreme values on the subsequent covariance estimation: wherein, , and the standardized residual vector of each sample is obtained by combination .

7. The method of claim 1, wherein the method is based on a multi-physical consistency modeling of the welding torch. In step S4, the covariance matrix is estimated using Ledoit-Wolf shrinkage estimation or regularized sample covariance methods to improve the robustness of the estimation, shrinkage estimation: using the Ledoit-Wolf shrinkage estimator to obtain , denoting its output as the covariance matrix and precision matrix, , , if shrinkage estimation fails or returns a non-finite value, fall back to sample covariance, sample covariance and regularization: compute the sample covariance and add diagonal regularization to ensure invertibility, where, is a small positive number, is the identity matrix.

8. The method of claim 1, wherein the method is based on a multi-physical consistency modeling of the welding torch. In step S5, the Mahalanobis distance of the standardized residual vector of the given sample to the estimated covariance inverse is defined as the multi-chain consistency score. The Mahalanobis distance takes into account both the magnitude of the residual of each chain and the correlation between chains. When two chains deviate simultaneously but are highly correlated with each other, the Mahalanobis distance corrects the information redundancy of their joint anomaly. When the deviation occurs in a direction unrelated to other chains, the distance can more sensitively reflect the anomaly of the individual chain.​​ 9. The method of claim 1, wherein the method is based on a multi-physical consistency modeling of the welding torch. In step S6, the evaluation indexes are area under the receiver operating characteristic curve, area under the precision-recall curve, false positive rate and false negative rate, etc., to realize the decision of the abnormal welding points, and the radar chart is used for abnormal explainable analysis.

Images