On-line prediction and evaluation system for laser welding quality of sheet metal

Abstract

The present application provides a kind of sheet laser welding quality online prediction and evaluation system, including multi-source data acquisition and monitoring module, for real-time acquisition of the multi-source heterogeneous process signal of sheet laser welding process;Quality prediction module is configured to include the cascade prediction architecture of macro reference prediction stage and micro residual compensation stage, to predict the prior prediction field of residual stress and deformation based on macro physical characteristics in macro reference prediction stage, in micro residual compensation stage, the space-time condition vector containing the micro residual feature extracted from prior prediction field is input to predict residual compensation term and prediction confidence interval, and the prediction result of residual stress and deformation is obtained by superimposing prior prediction field and residual compensation term.The present application provides a kind of data efficient, quantifiable online monitoring scheme for sheet laser welding process.

Description

Technical Field

[0001] This invention relates to the field of laser welding monitoring technology, and in particular to an online prediction and evaluation system for the quality of thin plate laser welding. Background Technology

[0002] Although laser welding has advantages such as high energy density, low heat input and controllable deformation, its process quality is extremely sensitive to fluctuations in parameters such as power and speed. It also involves a complex physical process involving multiple couplings of heat, force and metallurgy, which leads to the inherent bottlenecks of traditional "trial and error method" having long cycle, high cost and difficulty in obtaining the global optimal solution.

[0003] While numerical simulation can analyze the process mechanism, it is computationally expensive and time-consuming, making it difficult to meet the needs of rapid production line iteration. On the other hand, pure data-driven machine learning methods, although responsive, are limited by the "black box" nature of the system, resulting in insufficient prediction reliability and reliance on a large amount of labeled data.

[0004] Existing monitoring systems based on multi-source sensing also face problems such as a lack of effective correlation between sensing data and quality indicators, the inability of simulation models to be directly applied online, and insufficient generalization performance of machine learning models in areas with scarce samples. In thin plate welding, an application scenario with stringent requirements for deformation and residual stress control, the limitations of existing technical approaches are particularly prominent. Summary of the Invention

[0005] This invention provides an online prediction and evaluation system for the quality of laser welding of thin plates, which solves the problem that residual stress and out-of-plane deformation are difficult to predict quickly in the prior art, and realizes an online evaluation and prediction system for the quality of laser thin plate welding based on the fusion of physical mechanism and data-driven approach.

[0006] This invention provides an online prediction and evaluation system for the quality of laser welding of thin plates, comprising: The multi-source data acquisition and monitoring module is used to acquire multi-source heterogeneous process signals in the thin plate laser welding process in real time. The feature engineering module is used to extract macroscopic physical features from the preprocessed process signal, wherein the macroscopic physical features include at least a thermal input index calculated based on the process signal and a stiffness index reflecting the structure's resistance to deformation. The quality prediction module is configured as a cascaded prediction architecture including a macroscopic baseline prediction stage and a microscopic residual compensation stage. The macroscopic baseline prediction stage is used to predict the prior prediction field of residual stress and deformation based at least on the macroscopic physical characteristics. The microscopic residual compensation stage performs probabilistic reconstruction of the residual distribution based at least on the microscopic residual characteristics, and outputs the residual compensation term and prediction confidence interval. The microscopic residual characteristics are obtained based on the result of spatial differentiation processing of the prior prediction field. The fusion output module is used to superimpose the prior prediction field with the residual compensation term to generate and output the prediction results of residual stress and deformation.

[0007] According to the present invention, an online prediction and evaluation system for thin plate laser welding quality is provided. The macroscopic benchmark prediction stage adopts a ridge regression model, which takes the macroscopic physical characteristics as input and outputs the prior prediction field.

[0008] According to the present invention, an online prediction and evaluation system for the quality of laser welding of thin plates is provided, wherein the microscopic residual compensation stage includes a cascaded random forest regressor and a conditional diffusion probability model, wherein: The random forest regression branch is configured to take a spatiotemporal condition vector including the micro residual features, the prior prediction field and the macro physical features as input, and output a mean residual correction term. The conditional diffusion probability model branch is configured to use the spatiotemporal conditional vector and the mean residual correction term as control conditions to sample and generate the random residual term and the prediction confidence interval from the learned residual conditional probability distribution. The mean residual correction term and the random residual term together constitute the residual compensation term.

[0009] The present invention provides an online prediction and evaluation system for the quality of laser welding of thin plates. The ridge regression model is trained on a macroscopic dataset constructed through finite element simulation with the training objective of minimizing the mean squared error loss function with L2 regularization, and is used to output the prior prediction field. The conditional diffusion probability model is trained with the goal of minimizing the noise prediction error loss function. The training label of the branch of the conditional diffusion probability model is the residual field. The residual field is obtained by calculating the original residual field between the prior prediction field and the actual measurement value, and then subtracting the mean residual output of the random forest regressor from the original residual field.

[0010] The online prediction and evaluation system for thin plate laser welding quality provided by the present invention further includes a cloud map generation module, which is used to perform spatial interpolation and continuous surface reconstruction on the residual stress and deformation prediction results generated by the fusion output module to generate a full-field residual stress cloud map and a full-field deformation cloud map.

[0011] The online prediction and evaluation system for thin plate laser welding quality provided by the present invention further includes a cloud map comprehensive evaluation module, which is used to perform offline quantitative analysis on the full-field residual stress cloud map and the full-field deformation cloud map, automatically identify the residual stress peak area and the out-of-plane deformation abnormal area, and compare the identification results with experimental measurement data and historical process data to output a comprehensive quality evaluation index.

[0012] The thin plate laser welding quality online prediction and evaluation system provided by the present invention further includes a real-time prediction and evaluation module, which is used to comprehensively score and determine the risk level of the residual stress and deformation prediction results generated by the fusion output module, output online quality early warning signals, and provide a visual interface for real-time operation guidance.

[0013] The online prediction and evaluation system for thin plate laser welding quality provided by the present invention further includes a correlation analysis module, which is used to perform feature contribution analysis on the output of the quality prediction module based on the SHAP value, quantify the influence weight of each process parameter on residual stress and deformation, and generate a process sensitivity analysis report.

[0014] The online prediction and evaluation system for thin plate laser welding quality provided by the present invention further includes a preprocessing module for performing preprocessing on the multi-source heterogeneous process signal to generate the preprocessed process signal, wherein the preprocessing includes one or more of resampling, timing alignment, outlier filtering and dimensionless normalization.

[0015] According to the present invention, a thin plate laser welding quality online prediction and evaluation system is provided, wherein the multi-source data acquisition and monitoring module is further used to divide the welding process into multiple analysis units according to a preset time interval, and to map the process signals acquired in each analysis unit to the corresponding physical space interval of the weld seam through timestamps. Automatically identify abnormal process parameter segments within each analysis unit and record the abnormality type, abnormal parameter value, and timestamp.

[0016] The thin plate laser welding quality online prediction and evaluation system provided by this invention completes model training with a very small number of samples by constructing a physics-driven cascaded prediction architecture. Based on multi-source heterogeneous data in the welding process, it realizes real-time quantitative prediction of residual stress and deformation, and simultaneously generates confidence intervals and uncertainty evaluations for online quality monitoring and anomaly early warning throughout the welding process. Ultimately, it provides a data-efficient and quantifiable online monitoring solution for the thin plate laser welding process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the schematic diagrams of the architecture of the online prediction and evaluation system for thin plate laser welding quality provided by the present invention; Figure 2 This is a flowchart illustrating the quality prediction module in the online prediction and evaluation system for thin-plate laser welding quality provided by this invention. Figure 3 This is the second schematic diagram of the architecture of the online prediction and evaluation system for thin plate laser welding quality provided by the present invention; Figure 4 This is a flowchart illustrating the online prediction and evaluation system for thin-plate laser welding quality provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] The following is combined Figures 1 to 4 This invention introduces an online prediction and evaluation system for the quality of thin-plate laser welding, such as... Figure 1 As shown, it includes: The multi-source data acquisition and monitoring module 101 is used to acquire multi-source heterogeneous process signals in the thin plate laser welding process in real time. The feature engineering module 102 is used to extract macroscopic physical features from the preprocessed process signal, wherein the macroscopic physical features include at least a thermal input index calculated based on the process signal and a stiffness index reflecting the deformation resistance of the structure. The quality prediction module 103 is configured as a cascaded prediction architecture including a macroscopic baseline prediction stage and a microscopic residual compensation stage. The macroscopic baseline prediction stage is used to predict the prior prediction field of residual stress and deformation based at least on the macroscopic physical characteristics. The microscopic residual compensation stage performs probabilistic reconstruction of the residual distribution based on the microscopic residual characteristics and outputs the residual compensation term and prediction confidence interval. The microscopic residual characteristics are obtained based on the result of spatial differentiation processing of the prior prediction field. The fusion output module 104 is used to superimpose the prior prediction field with the residual compensation term to generate and output the prediction results of residual stress and deformation.

[0021] Specifically, the multi-source data acquisition and monitoring module is used to collect key dynamic signals and process parameters in real time during the laser welding process to construct multi-source heterogeneous process signals.

[0022] Optionally, key dynamic signals include laser power, weld deviation, welding speed, and temperature.

[0023] Process parameters include pre-set defocusing amount, protective gas pressure, and spot diameter; they also include plate thickness, heat input, surface energy density, and bending stiffness index.

[0024] Based on this, the aforementioned key dynamic signals and process parameters are preprocessed to perform time-series alignment and dimensionless standardization, and then input into the feature engineering module for extracting macroscopic physical features.

[0025] Among them, the macroscopic physical characteristics include thermal input indicators calculated based on key dynamic signals acquired in real time and stiffness indicators calculated based on process parameters, to reflect the overall thermodynamic and structural response characteristics and to be used for preliminary prediction of residual stress and deformation.

[0026] The quality prediction module receives macroscopic physical features output by the feature engineering module to predict residual stress and deformation through a cascaded prediction architecture.

[0027] The cascaded prediction architecture includes a macro-benchmark prediction stage (Model-I) and a micro-residual compensation stage (Model-II). Model-I can be configured as a linear prediction model, such as Lasso regression or elastic net, to capture the slowly changing, large-span parts of the stress / deformation field. This part is defined by the main process signals and structural stiffness and has strong regularity. Model-II can be configured as a nonlinear model to predict details that are difficult for Model-I to predict, such as those with drastic changes and local concentrations.

[0028] Based on this, unlike the conventional decoupling method according to the welding physical process in the prior art, the present invention decomposes the prediction task of welding stress and deformation field into two coupled sub-tasks: global trend prediction and local residual compensation, thereby achieving hierarchical decoupling of the prediction task.

[0029] Furthermore, in the macroscopic baseline prediction stage, by employing a linear prediction model with strong inductive bias, using macroscopic physical characteristics such as thermal input indicators and structural stiffness indicators as input, it can capture the smooth overall trend dominated and controlled by physical mechanisms with extremely low model complexity, and output a macroscopic prediction field that constitutes the framework of stress and deformation distribution across the entire field as the prior prediction field. Since this stage only needs to learn the mapping relationship between macroscopic physical characteristics and macroscopic trends, rather than fitting the entire complex physical field, the sample size required for training is effectively reduced. At the same time, the linear model itself has small parameter size, high numerical stability, and is not prone to overfitting under small sample conditions, making it suitable for prediction in small sample scenarios.

[0030] It should be noted that the prior prediction field in the first stage is not only the preliminary prediction result, but also a key input source for the micro residual compensation in the second stage. Specifically, the feature engineering module receives the prior prediction field output from the first stage and performs spatial differentiation on it to extract micro residual features such as spatial gradients and higher-order Laplacian operators at each node. This characterizes the spatial variation pattern of the prior prediction field in the local region, thus providing rich structural prior information for the second stage prediction. This allows Model-II to focus on modeling high-frequency residuals directly on the known smooth skeleton without implicitly inferring local nonlinear behavior from the original monitoring data / process parameters, thereby significantly reducing the learning difficulty of Model-II.

[0031] In the micro residual compensation stage, Model-II uses micro residual features as conditional sampling to generate residual compensation terms. Since the micro residual features have already encoded the local geometry of the prior prediction field, they provide Model-II with a high-quality initialization space. Based on this, Model-II can be configured as a conditional generative model to model the residuals. In this case, the model only needs a very small number of samples to learn the probability distribution of the residual field.

[0032] This invention decomposes the prediction of welding stress and deformation into two stages: macroscopic skeleton construction and microscopic detail compensation, through precision-level decoupling. This allows the model in each stage to run with minimal data requirements on its preferred task. Through the two-stage collaboration, the system can complete the learning process with only a very small number of finite element simulation samples, thus achieving the prediction of welding stress and deformation.

[0033] Furthermore, the fusion output module integrates the prior prediction field predicted in the macroscopic benchmark prediction stage and the residual compensation term in the microscopic residual compensation stage as the prediction results of residual stress and deformation. Finally, based on the multi-source heterogeneous process signals obtained by real-time monitoring, it realizes real-time monitoring of the thin plate laser welding process.

[0034] This invention constructs a physics-driven cascaded prediction architecture, completing model training with a very small number of samples. Based on multi-source heterogeneous data in the welding process, it achieves real-time quantitative prediction of residual stress and deformation, and simultaneously generates confidence intervals and uncertainty assessments for online quality monitoring and anomaly early warning throughout the welding process. Ultimately, it provides a data-efficient and quantifiable online monitoring solution for thin plate laser welding.

[0035] In a preferred embodiment, the multi-source data acquisition and monitoring module is further used to divide the welding process into multiple analysis units according to a preset time interval, and to map the process signals acquired in each analysis unit to the corresponding physical space interval of the weld seam through timestamps. Automatically identify abnormal process parameter segments within each analysis unit and record the abnormality type, abnormal parameter value, and timestamp.

[0036] Specifically, the system collects welding dynamic signals in real time through multi-source sensing devices, including laser power, weld deviation, welding speed and temperature, while recording process parameters such as defocusing amount, shielding gas pressure and spot diameter and storing them in the multi-source data acquisition and monitoring module.

[0037] Meanwhile, the system divides the welding process into multiple analysis units (T-elements) according to a preset time interval Δt. By using the acquisition timestamp of the process signal, the process signal of each T-element is mapped to the actual physical space interval corresponding to the T-element, so as to achieve a precise correspondence between process data and weld space location. All acquired data and anomaly records are stored synchronously with the time series for the quality traceability and anomaly location of the entire welding process, thus providing a reliable foundation for anomaly identification and subsequent analysis.

[0038] Based on this, the multi-source data acquisition and monitoring module can automatically identify abnormal process parameter segments within each T-element, including power drops, temperature anomalies, weld deviation drift, and speed anomalies. Optionally, the multi-source data acquisition and monitoring module also records the anomaly type, anomaly parameter value, timestamp, and potential impact on welding quality for each anomaly T-element, and provides anomaly handling or process optimization suggestions.

[0039] The online prediction and evaluation system for thin plate laser welding quality of the present invention also includes a preprocessing module for performing preprocessing on the multi-source heterogeneous process signals to generate the preprocessed process signals. The preprocessing includes one or more of the following: resampling, timing alignment, outlier filtering, and dimensionless normalization.

[0040] The preprocessing module receives the raw monitoring data and process parameters sent by the multi-source data acquisition and monitoring module as multi-source heterogeneous process signals, and performs resampling, time alignment, outlier filtering (robust detection based on standard deviation), linear interpolation completion, and dimensionless standardization on them.

[0041] Optionally, the preprocessing module also performs data cleaning on the discrete deformation point data acquired by point laser scanning, including the removal of outliers.

[0042] This ultimately forms a data benchmark for predicting residual stress and deformation.

[0043] In a preferred embodiment, the system further includes a feature engineering module for processing the preprocessed process signal into a feature vector for model prediction.

[0044] Specifically, the feature engineering module receives the preprocessed process signal sent by the preprocessing module, extracts macroscopic physical features including thermal input and stiffness indices, and uses them as input to Model-I: ; In the formula, Characterizing macroscopic physical parameters, P, V, and H represent laser power, welding speed, and plate thickness, respectively; Characterizing thermophysical eigenvectors, Q, , , , D and These represent heat input per unit length, area energy density, bending stiffness index, heat flow mode index, Peckley number surrogate, deformation sensitivity index, and volumetric thermal strain, respectively, which together characterize the influence of plate thickness, heat input, and structural bending resistance on residual stress and deformation.

[0045] Among them, the heat input index Q is calculated based on laser power and welding speed; the bending stiffness index is calculated based on plate thickness and material properties, characterizing the structural ability of the thin plate to resist welding deformation; the heat flow mode index characterizes the conduction mode of welding heat flow in the thin plate; the Peckle number surrogate reflects the relative relationship between the moving speed of the welding heat source and the heat diffusion speed; the deformation sensitivity index characterizes the sensitivity to welding deformation; and the volumetric thermal strain is the estimated value of thermal strain calculated from the welding heat input and the thermal expansion properties of the material.

[0046] Furthermore, for Model-II, the feature engineering index is also based on the prior prediction field of Model-I's real-time predictions. Construct a 24-dimensional spatiotemporal conditional vector c as the input vector for Model-II: ; Among these methods, statistical analysis is performed on the temperature sensing signals collected in real time during the welding process to extract the average temperature. Maximum temperature and temperature standard deviation Together with the protective gas flow rate F, they constitute the dynamic temperature characteristics. .

[0047] For each node in the weld and heat-affected zone, a normalized spatial coordinate system is constructed. and its quadratic term This provides positional encoding for subsequent spatial interpolation and full-field mapping, and utilizes the output of Model-I. Interactive features are generated with spatial coordinates, ultimately forming spatial features. .

[0048] For Model-I output Perform spatial differentiation to extract neighboring node values. and Spatial gradient and Laplace operator This constitutes the micro residual characteristics. .

[0049] In the online prediction and evaluation system for thin plate laser welding quality of the present invention, the macroscopic benchmark prediction stage adopts a ridge regression model, which takes the macroscopic physical characteristics as input and outputs the prior prediction field.

[0050] As a preferred implementation, the macroscopic baseline prediction stage uses an L2 regularized ridge regression model to map process parameters, physical derivatives, and thermal characteristics to the macroscopic prior prediction field, obtaining preliminary prediction results for residual stress and deformation. The expression for this model is: ; In the formula, Let W be the prior prediction field output by Model-I, and W be the learnable weight matrix. This is the mapping function for the ridge regression model.

[0051] In the online prediction and evaluation system for thin-plate laser welding quality of this invention, the microscopic residual compensation stage includes a cascaded random forest regressor and a conditional diffusion probability model, wherein: The random forest regression branch is configured to take a spatiotemporal condition vector including the micro residual features, the prior prediction field and the macro physical features as input, and output a mean residual correction term. The conditional diffusion probability model branch is configured to use the spatiotemporal conditional vector and the mean residual correction term as control conditions to sample and generate the random residual term and the prediction confidence interval from the learned residual conditional probability distribution. The mean residual correction term and the random residual term together constitute the residual compensation term.

[0052] Furthermore, based on the macro-benchmark prediction results, this implementation method uses a cascaded random forest regressor and a conditional probability diffusion model as Model-II to predict the residual compensation term.

[0053] Optionally, the conditional probability model is a one-dimensional conditional probability model.

[0054] The Random Forest Regressor (RF) is used to correct the mean residual and output the mean residual. The conditional probability diffusion model, based on mean residual correction, reconstructs the local nonlinear residual distribution and outputs random residuals. and its standard deviation .

[0055] Specifically, during the training phase, the random forest regressor uses... Using the input vector as input, and minimizing the residual between the prior prediction field output by Model-I and the actual measured values ​​obtained from experiments as the training objective, multiple decision trees are constructed and the prediction results of each tree are averaged to learn from the input vector. The deterministic mapping to the residual mean can be understood as capturing the systematic bias portion of the residuals that can be explained by structured features.

[0056] The one-dimensional conditional probability model is trained after the random forest regressor is trained, with the training objective of minimizing the noise prediction error. During training, noise is gradually added to the true residual obtained after compensation by the random forest regressor until it becomes pure Gaussian noise. Then, the network learns to gradually remove noise from the noise and recover the inverse conditional probability distribution of the original true residual.

[0057] Based on this, during the inference phase, i.e., the monitoring of the welding process, the random forest regressor first receives the spatiotemporal condition vector, including microscopic residual features, extracted from the prior prediction field, and outputs a deterministic mean residual correction term. This characterizes the expected compensation for deviations caused by predictable factors that the system should apply under the current operating conditions.

[0058] Then, the cascaded one-dimensional conditional probability diffusion model, using the spatiotemporal conditional vector and the mean residual correction term as control conditions, samples from the learned residual distribution to generate a random residual term that conforms to the characteristics of the residual residual distribution under the current operating condition. Furthermore, the standard deviation of each sampling result was calculated through multiple random samplings. This allows us to output the prediction confidence interval.

[0059] Based on this, the mean residual correction term and the random residual term can be superimposed to form the residual compensation term.

[0060] In this embodiment, a 95% confidence interval is constructed based on the sampling results of DDPM: In the formula, The residual stress and deformation prediction results are obtained by superimposing the prior prediction field and the residual compensation term.

[0061] Understandably, since the random forest regressor has already pre-extracted the predictable mean portion of the residuals, the conditional probability model only needs to reconstruct the probability of the remaining residual distribution, which has smaller amplitudes and a more concentrated structure. This allows it to achieve a complete prediction of the residual compensation term and the output of its confidence interval without having to access the actual measured values.

[0062] In summary, this invention addresses the problems of high cost, long experimental cycle, and limited sample size in obtaining actual welding test samples. It constructs a cascaded prediction architecture suitable for small sample conditions. Unlike directly using large-scale deep neural networks for end-to-end fitting of process parameters and welding quality, this invention introduces finite element prior results, thermophysical characteristics, spatial gradient characteristics, and residual compensation mechanisms. This decomposes the complex welding quality prediction problem into two relatively low-complexity sub-problems: "physical prior benchmark prediction" and "local residual correction." Therefore, effective training can be completed with a limited number of measured samples.

[0063] The cascaded prediction framework comprised of "physical prior constraints, low-complexity benchmark models, residual decomposition learning, and probabilistic uncertainty modeling" enables the model to minimize the degrees of freedom in the learning function space, increase the effective information contained in a single sample, and suppress overfitting problems common under small sample conditions. This allows for high-precision, fast, and reliable welding quality prediction results even with limited experimental data. Compared to traditional finite element methods and conventional end-to-end data-driven models, this framework offers lower sample requirements, higher computational efficiency, stronger prediction reliability, and better engineering applicability.

[0064] Furthermore, in the task of local residual correction, this invention combines residual compensation with probabilistic modeling to provide both the confidence interval or uncertainty range of the prediction results while outputting deterministic prediction results. Specifically, based on the macroscopic baseline prediction results and the random forest mean residual correction results, a conditional diffusion probability model is introduced to reconstruct the local nonlinear residual distribution and output the residual standard deviation or confidence interval. Thus, the prediction results not only provide numerical estimates of residual stress, deformation, or other welding quality indicators, but also reflect the reliability of the model under different working conditions, spatial locations, or local areas, providing a basis for quality assessment, risk judgment, and process decision-making in engineering applications.

[0065] Based on this, the process of the quality prediction module is as follows: Figure 2 As shown.

[0066] In the thin plate laser welding quality online prediction and evaluation system of the present invention, the ridge regression model is trained on a macroscopic dataset constructed by finite element simulation with the training objective of minimizing the mean square error loss function with L2 regularization, and is used to output the prior prediction field. The conditional diffusion probability model is trained with the goal of minimizing the noise prediction error loss function. The training label of the branch of the conditional diffusion probability model is the residual field. The residual field is obtained by calculating the original residual field between the prior prediction field and the actual measurement value, and then subtracting the mean residual output of the random forest regressor from the original residual field.

[0067] The micro residual compensation branch (Model-II) employs a cascaded training strategy, wherein: The cascaded first-layer random forest regressor is trained with the original residual field between the prior prediction field and the actual measurement value as the label, and the spatiotemporal condition vector obtained by integrating the micro residual features is used as the input to establish a deterministic mapping from the spatiotemporal condition vector to the mean residual. The conditional diffusion probability model branch of the cascaded second layer is trained with the residual field after compensation by the random forest regressor as the label, in order to learn the conditional probability distribution of the local nonlinear residual under the spatiotemporal conditional vector, and output the random residual and its confidence interval.

[0068] In this implementation, the core architecture adopts a two-stage cascaded correction strategy based on physical prior guidance. The specific training and mapping process is as follows: 1. Macroscopic Prior Mapping Stage (Model-I Training): First, the welding process is simulated using the traditional sequential coupled thermo-finite element method to obtain the residual stress field and deformation field of the entire field. This finite element simulation result is directly used as the training label. Model-I (Ridge Regression Model) optimizes the weight matrix W by minimizing the mean squared error loss function with L2 regularization as the training objective. This establishes a rapid mapping relationship between welding process parameters, welding process characteristics, and finite element simulation data, enabling rapid prediction of the macroscopic prior field of the entire field.

[0069] 2. Microscopic Residual Correction and Uncertainty Quantification Stage (Model-II Training): After completing Model-I training, the predicted output values ​​of the ridge regression model are input into Model-II as physical prior benchmarks. Simultaneously, actual welding experiments are conducted under the same process parameters as the simulation. A small amount of measured residual stress data is obtained at key locations using standardized measurement methods, and this measured data is used as the baseline true value. By calculating the difference between the prior predicted field output by Model-I and the baseline true value at each corresponding measurement point, the original residual field is constructed, thereby explicitly capturing the systematic bias caused by the simplification of the simulation model.

[0070] Corresponding to the cascaded architecture, Model-II also employs a cascaded training strategy for deterministic residual correction and uncertainty interval estimation: Deterministic residual learning: The macroscopic physical features of each working condition sample are input into the trained Model-I to obtain the prior prediction field. Spatial differentiation is performed on this field to extract microscopic residual features, which are then integrated into a spatiotemporal conditional vector. A random forest regressor is trained with the goal of minimizing the original residuals, establishing a deterministic mapping from the microscopic residual features to the residual mean.

[0071] Uncertainty interval estimation: After the random forest regressor is trained, the residual is calculated after mean residual correction compensation. The one-dimensional conditional diffusion probability model (CDPM) is trained with this residual as the label, so that it learns the conditional probability distribution of the residual residual under a given conditional vector, thereby completing the quantitative output of the whole field residual stress residual field and uncertainty interval.

[0072] This invention utilizes a small amount of sparse, measured residual stress or deformation data to perform cascaded residual correction on the full-field finite element simulation results, thereby reconstructing a high-fidelity full-field stress and deformation field that more closely approximates the actual welding physical state. This method explicitly compensates for systematic biases in the finite element model caused by factors such as mesh generation, simplified boundary conditions, uncertainties in high-temperature material parameters, errors in heat source model parameters, and nonlinear iterative solutions. This not only significantly improves the physical realism and reliability of the full-field mapping data but also effectively enhances the generalization ability of subsequent prediction models under small-sample measured constraints, avoiding predictive bias shifts caused by simply fitting low-fidelity simulation data.

[0073] Through the above method, in specific experiments, the present invention can complete model training based on only 16 sets of welding test samples and obtain high prediction accuracy. The determination coefficient of the prediction results can reach about 0.99, indicating that the method can significantly reduce the model's dependence on large-scale experimental data.

[0074] The thin plate laser welding quality online prediction and evaluation system of the present invention also includes a cloud map generation module, which is used to perform spatial interpolation and continuous surface reconstruction on the residual stress and deformation prediction results generated by the fusion output module to generate a full-field residual stress cloud map and a full-field deformation cloud map.

[0075] In one feasible implementation, a high-resolution full-field deformation cloud map is generated from the discrete point data cleaned by the preprocessing module using spatial interpolation and continuous surface reconstruction methods.

[0076] Then, the welding process parameters are bound to the collected deformation data to establish a one-to-one correspondence, forming an index mapping table of process parameters and deformation field grid nodes. This facilitates quick searching by process parameter name and provides convenient data access for feature extraction, quality prediction, and anomaly analysis.

[0077] Residual stress data and a small number of measured residual stress points are extracted from the simulation cloud map based on finite element method. The simulation data and measured data are then fused to generate a calibration reference cloud map, which is combined with the corresponding process parameters to form input data. The fused data is then mapped to the weld and thin plate surface through spatial interpolation and continuous reconstruction to achieve visualization of the stress field from the node level to the whole field. The generated residual stress cloud map is associated with process parameters and anomaly information to achieve the correspondence between time series and spatial location, providing data support for prediction verification, quality backtracking, and anomaly area location.

[0078] The online prediction and evaluation system for thin plate laser welding quality of the present invention also includes a cloud map comprehensive evaluation module, which is used to perform offline quantitative analysis on the full-field residual stress cloud map and the full-field deformation cloud map, automatically identify the residual stress peak area and the out-of-plane deformation abnormal area, and compare the identification results with experimental measurement data and historical process data to output a comprehensive quality evaluation index.

[0079] In one feasible implementation, the cloud map comprehensive evaluation module acquires the full-field deformation cloud map and residual stress cloud map after welding is completed. It reads, preprocesses, and checks the continuity of the cloud map data. Combined with welding process parameters and T-level anomaly information, the cloud map is converted into quantifiable evaluation data to ensure that the deformation amplitude and stress distribution of each region can accurately reflect the actual welding situation and provide high-quality input for subsequent analysis.

[0080] The processed cloud map data is comprehensively analyzed and anomaly identified. By comparing the local and overall trends of the cloud map, the residual stress peak area, abnormal deformation area and weld segment with potential high risk are automatically marked, and a comprehensive quality evaluation index is generated. At the same time, the abnormal patterns and trend changes are analyzed to provide an actionable decision basis for welding process improvement, welding quality retrospective and process optimization, and to ensure that potential risks in the welding process can be detected and identified in a timely manner.

[0081] By correlating and comparing the comprehensive evaluation results of cloud maps with experimental measurement data and historical process data, the accuracy and reliability of the prediction model are verified. At the same time, the variation law of welding quality under different process parameters is evaluated, providing scientific basis and data support for welding process optimization, residual stress and deformation control, and process design of future production batches, thus realizing a closed loop of offline high-precision quality assessment and process improvement.

[0082] The thin plate laser welding quality online prediction and evaluation system of the present invention also includes a real-time prediction and evaluation module, which is used to comprehensively score and determine the risk level of the residual stress and deformation prediction results generated by the fusion output module, output online quality early warning signals, and provide a visual interface for real-time operation guidance.

[0083] In one feasible implementation, the real-time prediction module performs a comprehensive scoring and risk level determination on the real-time residual stress and deformation results output by the quality prediction module, thereby achieving online quality early warning.

[0084] The real-time prediction and evaluation module is also used to correlate real-time prediction results with T-element anomaly detection, process parameter characteristics, and historical quality data to realize anomaly location and dynamic process optimization decision-making in the welding process.

[0085] Optionally, the real-time prediction and evaluation module also provides an online visualization interface for real-time quality situation awareness and operational guidance.

[0086] The online prediction and evaluation system for thin plate laser welding quality of the present invention also includes a correlation analysis module, which is used to perform feature contribution analysis on the output of the quality prediction module based on the SHAP value, quantify the influence weight of each process parameter on residual stress and deformation, and generate a process sensitivity analysis report.

[0087] In one feasible implementation, the correlation analysis module performs global and local correlation analysis based on the collected multi-source process signals and real-time predicted residual stress and deformation data, including nonlinear correlation analysis, univariate and multivariate relationship evaluation, to quantify the influence of each process parameter on residual stress and deformation, thereby providing data support for welding process window design, key parameter identification and process optimization.

[0088] Optionally, it can also be used to perform feature contribution analysis on the output of the prediction model using model interpretation and visualization methods, including using SHAP values ​​or similar interpretation techniques to reveal the contribution weight of each process parameter in the prediction of residual stress and deformation. At the same time, it combines scatter analysis and process parameter-target response mapping to achieve dynamic visualization and process sensitivity analysis, and generate a process sensitivity analysis report to provide a reference for process decision-making and risk assessment.

[0089] Based on this, a complete online prediction and evaluation system for thin plate laser welding quality, such as Figure 3 As shown, a complete online prediction and evaluation process for the quality of thin-plate laser welding is as follows: Figure 4 As shown.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An online prediction and evaluation system for the quality of laser welding of thin plates, characterized in that, include: The multi-source data acquisition and monitoring module is used to acquire multi-source heterogeneous process signals in the thin plate laser welding process in real time. The feature engineering module is used to extract macroscopic physical features from the preprocessed process signal, wherein the macroscopic physical features include at least a thermal input index calculated based on the process signal and a stiffness index reflecting the structure's resistance to deformation. The quality prediction module is configured as a cascaded prediction architecture including a macroscopic baseline prediction stage and a microscopic residual compensation stage. The macroscopic baseline prediction stage is used to predict the prior prediction field of residual stress and deformation based at least on the macroscopic physical characteristics. The microscopic residual compensation stage performs probabilistic reconstruction of the residual distribution based at least on the microscopic residual characteristics, and outputs the residual compensation term and prediction confidence interval. The microscopic residual characteristics are obtained based on the result of spatial differentiation processing of the prior prediction field. The fusion output module is used to superimpose the prior prediction field with the residual compensation term to generate and output the prediction results of residual stress and deformation.

2. The online prediction and evaluation system for thin plate laser welding quality according to claim 1, characterized in that, The macroscopic baseline prediction stage employs a ridge regression model, taking the macroscopic physical characteristics as input and outputting the prior prediction field.

3. The online prediction and evaluation system for thin plate laser welding quality according to claim 2, characterized in that, The micro residual compensation stage includes a cascaded random forest regressor and a conditional diffusion probability model, wherein: The random forest regression branch is configured to take a spatiotemporal condition vector including the micro residual features, the prior prediction field and the macro physical features as input, and output a mean residual correction term. The conditional diffusion probability model branch is configured to use the spatiotemporal conditional vector and the mean residual correction term as control conditions to sample and generate the random residual term and the prediction confidence interval from the learned residual conditional probability distribution. The mean residual correction term and the random residual term together constitute the residual compensation term.

4. The online prediction and evaluation system for thin plate laser welding quality according to claim 3, characterized in that, The ridge regression model is trained on a macroscopic dataset constructed through finite element simulation with the training objective of minimizing the mean squared error loss function with L2 regularization, and is used to output the prior prediction field. The conditional diffusion probability model is trained with the goal of minimizing the noise prediction error loss function. The training label of the branch of the conditional diffusion probability model is the residual field. The residual field is obtained by calculating the original residual field between the prior prediction field and the actual measurement value, and then subtracting the mean residual output of the random forest regressor from the original residual field.

5. The online prediction and evaluation system for thin plate laser welding quality according to any one of claims 1-3, characterized in that, It also includes a cloud map generation module, which performs spatial interpolation and continuous surface reconstruction on the residual stress and deformation prediction results generated by the fusion output module to generate full-field residual stress cloud maps and full-field deformation cloud maps.

6. The online prediction and evaluation system for thin plate laser welding quality according to claim 5, characterized in that, It also includes a cloud map comprehensive evaluation module, which is used to perform offline quantitative analysis on the full-field residual stress cloud map and the full-field deformation cloud map, automatically identify the residual stress peak area and the out-of-plane deformation abnormal area, and compare the identification results with experimental measurement data and historical process data to output a comprehensive quality evaluation index.

7. The online prediction and evaluation system for thin plate laser welding quality according to any one of claims 1-3, characterized in that, It also includes a real-time prediction and evaluation module, which is used to comprehensively score and determine the risk level of the residual stress and deformation prediction results generated by the fusion output module, output online quality early warning signals, and provide a visual interface for real-time operation guidance.

8. The online prediction and evaluation system for thin plate laser welding quality according to any one of claims 1-3, characterized in that, It also includes a correlation analysis module, which is used to perform feature contribution analysis on the output of the quality prediction module based on the SHAP value, quantify the influence weight of each process parameter on residual stress and deformation, and generate a process sensitivity analysis report.

9. The online prediction and evaluation system for thin plate laser welding quality according to any one of claims 1-3, characterized in that, It also includes a preprocessing module for performing preprocessing on the multi-source heterogeneous process signals to generate the preprocessed process signals, wherein the preprocessing includes one or more of resampling, timing alignment, outlier filtering and dimensionless normalization.

10. The online prediction and evaluation system for thin plate laser welding quality according to any one of claims 1-3, characterized in that, The multi-source data acquisition and monitoring module is also used to divide the welding process into multiple analysis units according to a preset time interval, and to map the process signals acquired in each analysis unit to the corresponding physical space interval of the weld through timestamps. Automatically identify abnormal process parameter segments within each analysis unit and record the abnormality type, abnormal parameter value, and timestamp.

Images