A resistance welding quality evaluation model incremental learning method suitable for complex industrial environments

Abstract

The application discloses a resistance welding quality evaluation model incremental learning method suitable for complex industrial environments, comprising the following steps: collecting multi-source time series data in a welding process, and constructing standard curves of each working condition based on the multi-source time series data; based on the standard curves, combining dynamic time warping and principal component analysis to evaluate the time sequence difference degree between the current working condition and the initial working condition, and determining an incremental learning sequence; constructing an incremental learning model, wherein the incremental learning model is used to extract time sequence features and weighted key time point information, and freeze the backbone network parameters of a meta model trained based on initial working condition data; according to the incremental learning sequence, sequentially fine-tuning the adapter of the incremental learning model to obtain an incremental learning model suitable for the current working condition; and using the incremental model suitable for the current working condition to evaluate the resistance welding quality. The application can effectively solve the model failure problem caused by data distribution drift under different working conditions.

Description

Technical Field

[0001] This invention belongs to the field of quality monitoring technology, and in particular relates to an incremental learning method for resistance welding quality evaluation models suitable for complex industrial environments. Background Technology

[0002] Resistance welding methods, such as resistance projection welding and resistance spot welding, are crucial joining processes in automobile manufacturing, aerospace, and high-speed train manufacturing. The quality of these connections directly impacts product safety and reliability. However, the resistance welding process is affected by various factors, including electrode wear, surface contact conditions, and assembly specifications. The time-varying nature of the process and the complexity of the environment make it difficult to guarantee the stability of resistance welding quality. Therefore, real-time monitoring of weld quality to ensure its stability and reliability is a key aspect of guaranteeing the quality of high-end equipment manufacturing.

[0003] In industrial settings, electrode wear, material replacement, and assembly gap fluctuations can cause data distribution drift, leading to the rapid failure of offline models and rendering them unsuitable for new operating conditions, necessitating retraining. Retraining requires a large amount of sample data, but labeling all samples through destructive testing is extremely expensive. Furthermore, even with significant investment in labeling all samples, due to process optimization in production sites, the number of qualified resistance welding samples often far exceeds the number of abnormal samples, potentially resulting in an imbalanced dataset and a low defect detection rate for the new model. Therefore, there is an urgent need for an adaptive model suitable for complex industrial environments to meet the demands of resistance welding quality monitoring under diverse and challenging conditions. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an incremental learning method for resistance welding quality evaluation models applicable to complex industrial environments, enabling adaptive evolution of the model under multiple working conditions and evaluation of weld quality.

[0005] To achieve the above objectives, this invention provides an incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments, comprising: Collect multi-source time series data during the welding process, and construct standard curves for each working condition based on the multi-source time series data; Based on the standard curve, the temporal difference between the current operating condition and the initial operating condition is evaluated by combining dynamic time warping and principal component analysis, and the incremental learning order is determined. Construct an incremental learning model, wherein the incremental learning model is used to extract temporal features and weighted key time point information, and freeze the backbone network parameters of the meta-model trained based on the initial working condition data; According to the incremental learning sequence, the adapters of the incremental learning models are extracted sequentially and fine-tuned for each working condition to obtain an incremental learning model suitable for the current working condition.

[0006] Optionally, the multi-source time series data includes: voltage, resistance, upper electrode displacement, lower electrode displacement, pressure, and current information during the welding process.

[0007] Optionally, a standard curve for each working condition is constructed based on the multi-source time series data, and the arithmetic mean of qualified samples at each time step is calculated. The qualified sample pair is the training set in each working condition that has undergone a tensile test, and those that pass the test are considered qualified samples. Based on the arithmetic mean, obtain the standard curve for the corresponding working condition.

[0008] Optionally, based on the standard curve, and combining dynamic time warping and principal component analysis to evaluate the temporal difference between the current operating condition and the initial operating condition, the incremental learning order is determined by including: The difference between the standard curve under the current operating condition and the standard curve under the initial operating condition is quantitatively calculated using dynamic time warping to obtain the initial difference vector; Principal component analysis is used to reduce the dimensionality of the initial difference vector and extract the main information to obtain a low-dimensional difference vector. Calculate the Euclidean distance between the low-dimensional difference vectors of different working conditions and the initial working condition to determine the incremental learning order.

[0009] Optionally, the incremental learning model includes: an LSTM temporal feature extraction module, an Attention mechanism module, a lightweight adapter, and a Softmax classifier; The LSTM time series feature extraction module is used to capture the cross-stage time series dependencies of multi-source time series data, extract the time series features at each time point, and input the time features at each time point into the Attention mechanism module. The Attention mechanism module weights the time features at each time point, emphasizes the features of key time periods, and obtains weighted global features. The lightweight adapter is used to process the weighted global features and store the processed features using an adapter suitable for the current operating condition. The Softmax classifier is used to normalize the features after linear transformation and output the final prediction result.

[0010] Optionally, capturing cross-stage temporal dependencies in multi-source time series data and extracting the temporal features at each time point includes: via input gate i t Forgotten Gate f t Cell state c t and output gate o tSynergistic effect updates hidden state: Based on the hidden states, obtain the hidden state sequence; Based on the hidden state sequence, extract the temporal features at each time point.

[0011] Optionally, via input gate i t Forgotten Gate f t Cell state c t and output gate o t The collaborative updates to hidden states include: ; Where σ is the Sigmoid activation function, ⊙ represents element-wise multiplication, and W, U, and b are learnable parameters. i W is the input weight matrix of the current input gate. f Let W be the input weight matrix of the current forget gate. c W is the input weight matrix for the current state. o U is the input weight matrix of the current output gate. i U is the cyclic weight matrix of the current input gate. f U is the current cyclic weight matrix of the forget gate. c U is the cyclic weight matrix for the candidate states. o Let b be the cyclic weight matrix of the current output gate. i Let b be the bias vector of the input gate. f Let b be the bias vector of the forget gate. c Let b be the bias vector of the candidate state. o Let x be the bias vector of the output gate. t h is the input feature vector at the current time t. t-1 Let c be the hidden state vector from the previous time step t-1. t-1 Let i be the cell state vector at the previous time step t-1. t is the input gate activation vector.

[0012] Optionally, the weighted global feature processing includes: The weighted global features are normalized layer by layer, then nonlinearly transformed by the first fully connected layer and the GELU activation function, then processed by the second fully connected layer, and finally the input and output are added together through residual connection.

[0013] Compared with the prior art, the present invention has the following advantages and technical effects: 1. Enhance model generalization and adaptability: By quantifying the differences in working conditions and optimizing the learning path through a progressive evolution strategy, combined with backbone network parameter freezing and adapter inheritance mechanisms, the model can adaptively evolve under multiple working conditions, effectively solving the model failure problem caused by data distribution drift under different working conditions.

[0014] 2. Balancing historical and new knowledge learning: Employing an incremental learning model, it achieves efficient learning of new knowledge without retraining, while fully preserving the learning results of historical operating conditions to avoid knowledge forgetting. The feature extraction capability of the meta-model can be stably reused under different operating conditions.

[0015] 3. Reduced industrial application costs: Dynamic learning rate and regularization strategies effectively solve the problem of overfitting with small samples in current working conditions. Only 30 labeled samples are needed to achieve the model performance target, which greatly reduces the data labeling cost in industrial scenarios. Compared with the existing technology that "requires 1000+ labeled samples to train a stable model", the data labeling cost is reduced by 97%. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is an architecture diagram of a multi-sensor signal classification model based on LSTM-Attention and an adapter, according to an embodiment of the present invention. Figure 2 This is a flowchart of the model training process according to an embodiment of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0019] This embodiment proposes an incremental learning method for resistance welding quality evaluation models suitable for complex industrial environments, which specifically includes the following steps: Collect multi-source time series data during the welding process, and construct standard curves for each working condition based on the multi-source time series data; Based on the standard curve, the temporal difference between the current operating condition and the initial operating condition is evaluated by combining dynamic time warping and principal component analysis, and the incremental learning order is determined. Construct an incremental learning model, wherein the incremental learning model is used to extract temporal features and weighted key time point information, and freeze the backbone network parameters of the meta-model trained based on the initial working condition data; According to the incremental learning sequence, the adapters of the incremental learning models are extracted sequentially and fine-tuned for each working condition to obtain an incremental learning model suitable for the current working condition.

[0020] Specifically, step 1: data preprocessing, by installing displacement and current sensors in the projection welding machine, six kinds of raw time series data are collected during the welding process, namely voltage, resistance, upper electrode displacement, lower electrode displacement, pressure and current information, and standard curves for each working condition are constructed. Step 2: Multi-condition time-series difference assessment and incremental learning path planning. Based on Dynamic Time Warping (DTW) and Principal Component Analysis (PCA), a progressive evolution mechanism is constructed to quantify the difference between the current condition and the initial condition, and to determine the incremental learning order of the model. The specific process is as follows: Step 2.1: Quantitatively calculate the difference between the standard curves of the six characteristics under each working condition and the initial working condition standard curve using DTW to obtain the initial difference vector. .

[0021] Step 2.2: Use PCA to reduce the dimensionality of the difference vector and extract the main information to obtain the final low-dimensional difference vector. (m is the number of principal components whose cumulative variance contribution rate exceeds 85%.) Step 3: Construct an assemblyable LSTM-Attention incremental learning model. LSTM layers are used to extract time-series features, and the Attention layer focuses on different time steps to concentrate on information at key time points. Subsequently, the parameters of the LSTM-Attention backbone network of the meta-model trained based on the initial working condition data are frozen to preserve historical knowledge.

[0022] Step 3.1: LSTM through the input gate i t 、 Forgotten Gate f t 、 Cell state c t and output gate o t Synergistic effect updates hidden state: ; Where σ is the sigmoid activation function, ⊙ represents element-wise multiplication, and W, U, and b are learnable parameters. i W is the input weight matrix of the current input gate. fLet W be the input weight matrix of the current forget gate. c W is the input weight matrix for the current state. o U is the input weight matrix of the current output gate. i U is the cyclic weight matrix of the current input gate. f U is the current cyclic weight matrix of the forget gate. c U is the cyclic weight matrix for the candidate states. o Let b be the cyclic weight matrix of the current output gate. i Let b be the bias vector of the input gate. f Let b be the bias vector of the forget gate. c Let b be the bias vector of the candidate state. o Let x be the bias vector of the output gate. t h is the input feature vector at the current time t. t-1 Let c be the hidden state vector from the previous time step t-1. t-1 Let i be the cell state vector at the previous time step t-1. t is the input gate activation vector.

[0023] Step 3.2: The Attention layer calculates the attention score using the learnable query vector q and normalizes it to obtain the weights. Different weights represent the importance of the state at different time points. By integrating the features from each time period, a global feature representation is obtained. : ; ; Step 4: Extract the adapter of the meta-model trained based on the initial working condition data, and fine-tune it for each working condition in sequence according to the planned incremental learning order.

[0024] Step 4.1: For each adapter, first adjust the attention output h. out The layers are normalized, then pass through a fully connected layer (FC1), GELU activation, and a nonlinear transformation of the fully connected layer (FC2), and the original features are preserved through residual connections. The function is a Gaussian cumulative distribution function, and W1b1, W2, b2 are adapter-specific parameters. The features after linear transformation are normalized using the Softmax function, and the final prediction result is output.

[0025] ; ; ; .

[0026] Step 4.2: Use a dynamic learning rate based on sample size to match the fine-tuning intensity with the sample size. In the formula, nsamples is the current sample size for fine-tuning, nmin is the minimum number of samples (set to 1), and nmax is the number of samples in the incremental learning training set under the current conditions (set to 30). To avoid the denominator being 0. The decayratio ranges from [0,1]. The larger the sample size, the smaller the ratio, and the lower the fine-tuning intensity.

[0027] ; .

[0028] Furthermore, the multi-source time series data includes: voltage, resistance, upper electrode displacement, lower electrode displacement, pressure, and current information during the welding process.

[0029] Furthermore, standard curves for each working condition are constructed based on the multi-source time series data, and the arithmetic mean of qualified samples at each time step is calculated. The qualified sample pair is the training set in each working condition that has undergone a tensile test, and those that pass the test are considered qualified samples. Based on the arithmetic mean, obtain the standard curve for the corresponding working condition.

[0030] Specifically, in step 1, the standard curve is obtained by calculating the arithmetic mean of all qualified samples at each time step under the same operating condition. For an operating condition containing N qualified samples, let the time series of the i-th sample be: ; Where T is the sequence length. Then the standard curve for this operating condition... Defined as: ; in, .

[0031] Furthermore, based on the standard curve, and combining dynamic time warping and principal component analysis to evaluate the temporal difference between the current operating condition and the initial operating condition, the incremental learning sequence is determined, including: The difference between the standard curve under the current operating condition and the standard curve under the initial operating condition is quantitatively calculated using dynamic time warping to obtain the initial difference vector; Principal component analysis is used to reduce the dimensionality of the initial difference vector and extract the main information to obtain a low-dimensional difference vector. Calculate the Euclidean distance between the low-dimensional difference vectors of different working conditions and the initial working condition to determine the incremental learning order.

[0032] Specifically, the implementation of DTW in step 2.1 includes: For the standard time-series curves of six types of welding sensor signals (voltage, current, pressure, resistance, upper displacement, and lower displacement) under different operating conditions, the difference between the standard curves of different operating conditions and the initial operating condition is calculated dimension by dimension using DTW. The calculation method is as follows: ; in, and These are the standard curves for the new and old operating conditions (with lengths of m and n, respectively). Let be the Euclidean distance of the k-th regular point pair. This represents the path weight.

[0033] The specific implementation of PCA in step 2.2 includes: For the six-dimensional DTW difference vector Principal Component Analysis (PCA) is used for dimensionality reduction and compression. Through orthogonal linear transformation, PCA maps high-dimensional difference vectors to a low-dimensional space composed of principal components, where each principal component is a linear combination of the original vectors, and their variance contributions decrease sequentially. Its optimization objective is: ; in, The weight vector of the principal components. This represents variance. By retaining the principal components that contribute the most to the variance, the six-dimensional difference vector can be compressed into a low-dimensional comprehensive difference degree. By calculating the covariance matrix and performing eigenvalue decomposition, the top m principal components with a cumulative variance contribution rate exceeding 85% are retained to obtain the final low-dimensional difference vector. Then, the Euclidean distances of the final low-dimensional difference features between different working conditions and the initial working condition are calculated and sorted to obtain the incremental learning order of the model.

[0034] The incremental learning model includes: an LSTM temporal feature extraction module, an Attention mechanism module, a lightweight adapter, and a Softmax classifier; The LSTM time series feature extraction module is used to capture the cross-stage time series dependencies of multi-source time series data, extract the time series features at each time point, and input the time features at each time point into the Attention mechanism module. The Attention mechanism module weights the time features at each time point, emphasizes the features of key time periods, and obtains weighted global features. The lightweight adapter is used to process the weighted global features and store the processed features using an adapter suitable for the current operating condition. The Softmax classifier is used to normalize the features after linear transformation and output the final prediction result.

[0035] Furthermore, capturing cross-stage temporal dependencies in multi-source time series data and extracting the temporal features at each time point includes: via input gate i t Forgotten Gate f t Cell state c t and output gate o t Synergistic effect updates hidden state: Based on the hidden states, obtain the hidden state sequence; Based on the hidden state sequence, extract the temporal features at each time point.

[0036] Furthermore, the weighted global feature processing includes: The weighted global features are normalized layer by layer, then nonlinearly transformed by the first fully connected layer and the GELU activation function, then processed by the second fully connected layer, and finally the input and output are added together through residual connection.

[0037] Specifically, in the meta-model pre-training phase, a basic meta-model is first trained based on a working condition with the largest sample size and most balanced categories. An end-to-end model is then constructed, incorporating LSTM temporal feature extraction, an attention mechanism, a lightweight adapter, and a Softmax classifier. This model takes preprocessed six classes of multi-source sensor signals as input and outputs the predicted weld quality category. The LSTM temporal feature extraction module captures the cross-stage temporal dependencies of the welding signal and extracts temporal features through the synergistic effect of the input gate, forget gate, cell state, and output gate. The attention mechanism module calculates the attention weights at each time step using a learnable query vector, enhancing the feature contributions of key time nodes such as resistance mutations and generating weighted global features. The lightweight adapter module performs LayerNorm normalization, GELU activation, and two fully connected transformations on the attention output features, preserving the original features through residual connections. Each working condition corresponds to a dedicated adapter to store knowledge related to that condition. After completing the meta-model pre-training phase, the parameters of the LSTM-Attention are frozen, and the meta-model adapter parameters are saved.

[0038] The adapter inheritance phase comprises three stages: Dynamic Time Warping (DTW), Principal Component Analysis (PCA), and Euclidean distance calculation. First, the DTW algorithm is used to calculate the minimum alignment distance between the standard curves of each incremental load case and the initial load case, resulting in a six-dimensional difference vector. Then, Principal Component Analysis (PCA) is used to reduce the dimensionality of the difference vector, retaining the top m principal components with a cumulative variance contribution rate exceeding 85%, thus eliminating information redundancy. Finally, the Euclidean distance between each incremental load case and the initial load case in the PCA-reduced space is calculated. The model evolution order is determined based on the distance magnitude, prioritizing the learning of load case data with smaller differences from the initial load case to reduce the difficulty of incremental learning.

[0039] The differentiated evolutionary fine-tuning stage includes two phases: extracting historical adapters and dynamically adjusting the adapters using a dynamic learning rate. First, based on the model's evolutionary order, historical adapters from the previous scenario are extracted to effectively transfer historical knowledge. In the dynamic learning rate fine-tuning phase, the decay ratio is calculated based on the current training sample size; the larger the sample size, the lower the fine-tuning intensity. The minimum learning rate is limited to 1 / 5 of the base learning rate to avoid convergence stagnation. The training epochs are set to 15, and L2 regularization is added to suppress overfitting with small sample sizes.

[0040] The following is a detailed description of this embodiment with reference to the accompanying drawings: The method of this invention collects welding parameters from actual welding production processes and divides them into seven working conditions based on differences in welding current and electrode pressure. The working condition with the largest sample size and the most balanced category distribution is selected as the initial working condition, and the remaining six are incremental working conditions. A standard curve is constructed as described in step one, and a meta-model is trained using the welding data of the initial working condition. The parameters of the meta-model's LSTM-Attention backbone network are then frozen. When faced with a new working condition, the model's evolution path is determined based on the difference evaluation results of the progressive evolution strategy, and a new adapter is trained using a dynamic learning rate fine-tuning strategy, ultimately generating a dedicated model suitable for the specific new working condition. Under the initial working condition, the meta-model achieves a classification accuracy of 93.2%.

[0041] like Figure 1 As shown, the model architecture of this invention consists of five parts, and the processing flow of the model can be summarized as: data preprocessing, LSTM feature extraction, Attention focusing on key time steps, adapter for specific working conditions, and classification output.

[0042] The input layer receives real-time data from six sensors during the resistance welding process: current, voltage, force, upper electrode displacement, lower electrode displacement, and dynamic resistance.

[0043] In the LSTM layer, six sensor signals are input to six parallel LSTM units. This design allows for the specialized learning of the unique temporal patterns and long-term dependencies of each signal. After processing the entire time series, each LSTM unit outputs a series of hidden states. The LSTM passes through input gate i... t Forgotten Gate t Cell state c t and output gate o t Synergistic effect updates hidden state: ; Where σ is the Sigmoid activation function, ⊙ represents element-wise multiplication, and W, U, b are learnable parameters. The resulting hidden state sequence It contains all the feature information extracted by the LSTM network.

[0044] The attention layer addresses the issue of varying importance of different time points in the final quality evaluation. The attention mechanism receives all LSTM hidden states. An attention weight α is calculated for each hidden state. This weight represents the importance of information at the corresponding time point. The sum of all weights is 1. The model automatically learns which key moments in the welding process are most decisive for judging the weld quality. All hidden states are then weighted and summed according to their attention weights to generate a global feature vector h. out .

[0045] The adapter is the core component of this invention that enables incremental learning. It consists of four parts: LayerNorm, fully connected layer 1 with GELU activation function, fully connected layer 2, and residual connection. There are N corresponding adapters for N operating conditions.

[0046] LayerNorm: Normalizes the input features.

[0047] Fully connected layer 1+GELU activation function: Performs feature transformation to reduce the dimensionality of the data.

[0048] Fully connected layer 2: Transforms the feature dimensions back to their original size.

[0049] Residual connection: The input and output of the module are added together to ensure that the basic knowledge learned in the initial working condition is not destroyed during incremental learning, while the model can be fine-tuned to adapt to the new working condition.

[0050] During incremental learning, the parameters of the LSTM-Attention backbone network are frozen, and only a small number of parameters in the adapter are updated. This allows for rapid adaptation to new operating conditions while significantly saving computational resources.

[0051] The output of the final residual connection is passed through a Softmax classifier to generate the final classification result.

[0052] When process adjustments or welding material changes are made on-site, it is necessary to learn from multiple operating conditions. Therefore, how to perform incremental model learning and design the learning path are the primary issues to be addressed. To enable the model to adapt to new operating conditions as quickly as possible and maintain high accuracy in the initial stages of entering new conditions, this embodiment proposes a model incremental evolution strategy based on Dynamic Time Warping (DTW) and Principal Component Analysis (PCA) to address the differences in standard curves under different operating conditions.

[0053] The standard curve for a specific operating condition is constructed from its set of qualified samples. Specifically, for an operating condition containing N qualified samples, each sample is a time series of length T. Let the time series of the i-th sample be: ; Then the standard curve for this working condition Defined by the following formula: ; The standard curves for current, pressure, voltage, dynamic resistance, upper displacement, and lower displacement under different operating conditions after various processing steps are compared. For six new operating conditions, the difference between their standard curves and the initial standard curves is calculated. This calculation is based on the DTW algorithm and is performed dimension-by-dimensionally (i.e., for each sensor signal). For each dimension, the minimum alignment path and the corresponding cumulative distance between two standard curves are obtained using the DTW algorithm. From this, an initial difference vector can be obtained. .

[0054] Wherein, the j-th dimension component d j The DTW distance between the old and new standard curves represents the distance of the j-th type of sensor signal, where d is the DTW distance in a single dimension. j It is given by the following formula: ; in, and These are the standard curves for the new and old operating conditions (with lengths of m and n, respectively). Let be the Euclidean distance of the k-th regular point pair. This represents the path weight.

[0055] Because the six types of parameters are strongly coupled (such as the positive correlation between current and voltage), PCA is used to reduce the dimension of the difference vector: after standardization, the covariance matrix is ​​calculated and eigenvalue decomposition is performed to extract the top n principal components with a cumulative variance contribution rate of more than 85%; then the Euclidean distance between each working condition and the initial working condition in the 3D PCA space is calculated and sorted according to the distance between the working conditions, and finally the incremental learning path is determined.

[0056] In industrial practice, the sample size of new operating conditions often accumulates gradually as production progresses. Using fixed fine-tuning parameters can easily lead to underfitting when samples are scarce and overfitting when samples are abundant. To address this issue, this embodiment designs a dynamic learning rate decay mechanism based on sample size to match the fine-tuning intensity with the data scale. The formula for calculating the decay coefficient (decayratio) is as follows: ; ; Where, n samples n is the number of samples currently used for fine-tuning. min and n max These are the preset minimum and maximum sample size reference values ​​for a single working condition (1 and 30 respectively in this embodiment), and ϵ is a very small constant used to avoid the denominator being zero. This coefficient is constrained within the interval [0,1]. The smaller the sample size, the larger its value, and the stronger the fine-tuning; conversely, the larger the sample size, the smaller its value, and the more conservative the fine-tuning, thereby achieving adaptive control of the learning process.

[0057] This embodiment performs quality assessment and monitoring on the weld points of resistance projection welding of nuts.

[0058] Model training process as follows Figure 2 As shown, the voltage, resistance, upper electrode displacement, lower electrode displacement, pressure, and current information during the initial welding process are input into the model to construct a meta-model. The parameters of the LSTM-Attention backbone network are then frozen, and further adjustments are made based on the parameters of the attached... Figure 2 The standard curves for spot welding under different working conditions are used to determine the incremental learning order of the model through a progressive evolution strategy, generate specific adapters, and form the corresponding model for the new working conditions.

[0059] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An incremental learning method for resistance welding quality evaluation models suitable for complex industrial environments, characterized in that, include: Collect multi-source time series data during the welding process, and construct standard curves for each working condition based on the multi-source time series data; Based on the standard curve, the temporal difference between the current operating condition and the initial operating condition is evaluated by combining dynamic time warping and principal component analysis, and the incremental learning order is determined. Construct an incremental learning model, wherein the incremental learning model is used to extract temporal features and weighted key time point information, and freeze the backbone network parameters of the meta-model trained based on the initial working condition data; According to the incremental learning order, the adapters of the incremental learning models are extracted sequentially and fine-tuned for each working condition to obtain an incremental learning model suitable for the current working condition. The quality of resistance welding is evaluated using an incremental model applicable to the current operating conditions.

2. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 1, characterized in that, The multi-source time series data includes: voltage, resistance, upper electrode displacement, lower electrode displacement, pressure, and current information during the welding process.

3. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 2, characterized in that, Based on the multi-source time series data, standard curves for each working condition are constructed, and the arithmetic mean of qualified samples at each time step is calculated. The qualified sample pair is the training set in each working condition that has undergone a tensile test, and those that pass the test are considered qualified samples. Based on the arithmetic mean, obtain the standard curve for the corresponding working condition.

4. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 2, characterized in that, Based on the standard curve, and combining dynamic time warping and principal component analysis to evaluate the temporal difference between the current operating condition and the initial operating condition, the incremental learning sequence is determined as follows: The difference between the standard curve under the current operating condition and the standard curve under the initial operating condition is quantitatively calculated using dynamic time warping to obtain the initial difference vector; Principal component analysis is used to reduce the dimensionality of the initial difference vector and extract the main information to obtain a low-dimensional difference vector. Calculate the Euclidean distance between the low-dimensional difference vectors of different working conditions and the initial working condition to determine the incremental learning order.

5. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 2, characterized in that, The incremental learning model includes: an LSTM temporal feature extraction module, an Attention mechanism module, a lightweight adapter, and a Softmax classifier; The LSTM time series feature extraction module is used to capture the cross-stage time series dependencies of multi-source time series data, extract the time series features at each time point, and input the time features at each time point into the Attention mechanism module. The Attention mechanism module weights the time features at each time point, emphasizes the features of key time periods, and obtains weighted global features. The lightweight adapter is used to process the weighted global features and store the processed features using an adapter suitable for the current operating condition. The Softmax classifier is used to normalize the features after linear transformation and output the final prediction result.

6. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 5, characterized in that, Capturing cross-stage temporal dependencies in multi-source time series data and extracting the temporal features at each time point includes: via input gate i t Forgotten Gate f t Cell state c t and output gate o t Synergistic effect updates hidden state: Based on the hidden states, obtain the hidden state sequence; Based on the hidden state sequence, extract the temporal features at each time point.

7. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 6, characterized in that, via input gate i t Forgotten Gate f t Cell state c t and output gate o t The collaborative updates to hidden states include: ； Where σ is the Sigmoid activation function, ⊙ represents element-wise multiplication, and W, U, and b are learnable parameters. i W is the input weight matrix of the current input gate. f Let W be the input weight matrix of the current forget gate. c W is the input weight matrix for the current state. o U is the input weight matrix of the current output gate. i U is the cyclic weight matrix of the current input gate. f U is the current cyclic weight matrix of the forget gate. c U is the cyclic weight matrix for the candidate states. o Let b be the cyclic weight matrix of the current output gate. i Let b be the bias vector of the input gate. f Let b be the bias vector of the forget gate. c Let b be the bias vector of the candidate state. o Let x be the bias vector of the output gate. t h is the input feature vector at the current time t. t-1 Let c be the hidden state vector from the previous time step t-1. t-1 Let i be the cell state vector at the previous time step t-1. t is the input gate activation vector.

8. The incremental learning method for a resistance welding quality evaluation model suitable for complex industrial environments according to claim 5, characterized in that, The weighted global feature processing includes: The weighted global features are normalized layer by layer, then nonlinearly transformed by the first fully connected layer and the GELU activation function, then processed by the second fully connected layer, and finally the input and output are added together through residual connection.

Images