Injection molding process fault diagnosis method based on domain generalization time sequence feature fusion

Abstract

The application discloses an injection molding process fault diagnosis method based on domain generalization time sequence feature fusion, relates to the technical field of industrial intelligent manufacturing and process monitoring, and comprises the following steps: obtaining high-resolution time sequence data and scalar process parameters of each injection cycle and performing pretreatment; inputting the processed time sequence data into a time sequence feature encoder to generate a time sequence feature vector, and inputting the pretreated scalar process parameters into a scalar feature encoder to generate a scalar feature vector; performing feature fusion on the time sequence feature vector and the scalar feature vector through a cross-modal attention gate fusion module to output a joint feature vector; inputting the joint feature vector into a domain generalization module, extracting domain invariant features through a complementary mechanism of causal feature disentanglement, multi-Gaussian distribution alignment and invariant feature mask adversarial training; and inputting the extracted domain invariant features into a fault detection and classification module to output a fault diagnosis result. The application improves the accuracy of cross-condition fault detection.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent manufacturing and process monitoring technology, specifically to a fault diagnosis method for injection molding process based on domain generalization temporal feature fusion. Background Technology

[0002] Injection molding is currently the most widely used method for processing thermoplastics. Globally, over 110,000 new injection molding machines are added annually, with an average service life of over 10 years. The demand for injection molded parts in the manufacturing industry is extremely strong. However, the injection molding process involves multiple coupled physical fields such as temperature, pressure, flow rate, holding pressure, and cooling. The process parameters are complex and susceptible to disturbances such as raw material batch fluctuations, mold wear, and changes in environmental temperature and humidity, making it highly prone to product defects or process failures. This leads to increased scrap rates, downtime losses, and energy waste.

[0003] Traditional injection molding quality monitoring methods mainly rely on manual inspections by operators and post-process sampling, making it difficult to achieve online, real-time process fault early warning. In recent years, intelligent process monitoring methods based on machine learning and deep learning have made significant progress, but existing methods still have many shortcomings. First, most existing methods only utilize scalar process parameters output by the machine control system, such as maximum injection pressure, switching injection pressure, and metering time, neglecting the rich dynamic process information contained in high-resolution time-series signals such as injection pressure curves and injection flow rate curves. This results in insufficient feature representation capabilities and low sensitivity to fine-grained quality anomalies. Second, injection molding process conditions are highly variable; parameters such as barrel temperature, mold temperature, and injection speed vary significantly between different production batches and different molds, leading to significant shifts in the distribution of process data under different production conditions. Models trained under one condition experience a sharp performance drop when transferred to a new condition, exhibiting severely insufficient model generalization ability. Furthermore, existing deep learning fault diagnosis models typically require a large number of labeled fault samples for supervised training, while in actual industrial scenarios, fault samples are scarce and labeling costs are high. Models are prone to overfitting under small sample conditions, limiting industrial usability. In addition, existing methods rarely take into account both multimodal data fusion and domain generalization capabilities, and cannot fully utilize information from multi-source heterogeneous sensors under different operating conditions.

[0004] To address the aforementioned issues, existing research in the field of fault diagnosis has proposed several representative domain generalization methods, including: Single-Source Domain Generalization Network (ICDDG) based on improved causal disentanglement, which uses triplet loss and prototype learning to separate causal and non-causal features, improving the model's generalization ability in unseen target domains; Single-Source Domain Generalization Method based on Multi-Gaussian Attention (MGA-SDG), which projects multi-scale fault features onto a Gaussian feature space and applies covariance loss to achieve robust cross-domain feature representation; Domain Generalization Method based on Invariant Feature Cleaning (IFPDG), which uses a two-stage game mechanism of global training and adversarial training to extract purely invariant features; Domain Generalization Method based on Meta-Learning Gradient Alignment and Data Augmentation (MGADA), which introduces Mixup augmentation and gradient consistency constraints within the meta-learning framework; and Source Free Domain Adaptation Method (SFDA), which achieves cross-domain knowledge transfer through pseudo-labels and entropy minimization when source domain data is inaccessible.

[0005] However, injection molding process data exhibits a typical periodic temporal structure, with pressure and flow curves showing significant distributional drift under different process conditions. Furthermore, product quality labels such as weight and geometric dimensions are continuous rather than discrete categories, posing additional challenges to domain generalization modeling. Therefore, there is an urgent need for an intelligent fault diagnosis method specifically designed for injection molding processes, possessing both multimodal temporal feature fusion and cross-condition generalization capabilities. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: This application provides a fault diagnosis method for injection molding process based on domain generalization temporal feature fusion, including: Acquire high-resolution timing data and scalar process parameters for each injection molding cycle and perform preprocessing; The preprocessed high-resolution time series data is input into a pre-built time series feature encoder to generate a time series feature vector. At the same time, the preprocessed scalar process parameters are input into a pre-built scalar feature encoder to generate a scalar feature vector. The temporal feature vector and the scalar feature vector are fused across modally using a cross-modal attention-gated fusion module to output a joint feature vector; The joint feature vector is input into the domain generalization module, and the domain invariant features are extracted through the complementary mechanisms of causal feature deentanglement, multi-Gaussian distribution alignment, and invariant feature mask adversarial training. The extracted domain-invariant features are input into the fault detection and classification module, which simultaneously performs binary anomaly detection, multi-class fault classification, and product quality index regression prediction, and outputs fault diagnosis results.

[0007] In one possible implementation, the temporal feature encoder includes a multi-scale convolutional block, the output of which is connected to the input of a channel attention module. The channel attention module adopts a Squeeze-and-Excitation structure, and its output is connected to the input of a Transformer encoder. The output of the Transformer encoder is connected to the input of a global average pooling layer. The scalar feature encoder includes a multilayer perceptron, which includes a first fully connected layer, a first batch normalization layer, a first ReLU activation function layer, a first Dropout regularization layer, a second fully connected layer, a second batch normalization layer, a second ReLU activation function layer, a second Dropout regularization layer, a third fully connected layer, a third batch normalization layer, a third ReLU activation function layer, and a third Dropout regularization layer connected in sequence. The output of the global average pooling layer and the output of the third Dropout regularization layer are both connected to the input of the cross-modal attention gating fusion module. The cross-modal attention gating fusion module includes a query mapping layer, a key mapping layer, and a value mapping layer. The input of the query mapping layer is connected to the output of the third Dropout regularization layer. The inputs of the key mapping layer and the value mapping layer are connected to the output of the global average pooling layer. The outputs of the query mapping layer, the key mapping layer, and the value mapping layer are all connected to the input of the cross-modal attention calculation unit. The outputs of the cross-modal attention calculation unit and the third Dropout regularization layer are both connected to the input of the gating fusion unit. The output of the gating fusion unit is connected to the input of the dimensionality reduction projection layer. The output of the dimensionality reduction projection layer is connected to the input of the domain generalization module. The domain generalization module includes a causal feature deentanglement unit, the input of which is connected to the output of the dimensionality reduction projection layer. The output of the causal feature deentanglement unit is connected to the input of a multi-Gaussian distribution alignment unit and an invariant feature mask adversarial training unit, respectively. The invariant feature mask adversarial training unit includes a gradient inversion layer and a domain discriminator. The input of the domain invariant feature output unit is connected to the output of the causal feature deentanglement unit, the multi-Gaussian distribution alignment unit, and the invariant feature mask adversarial training unit, respectively. The output of the domain invariant feature output unit is connected to the fault detection and classification module.

[0008] In one possible implementation, the preprocessed high-resolution time-series data is input into a pre-built time-series feature encoder to generate a time-series feature vector, including: The preprocessed high-resolution time series data is input into the multi-scale convolutional block. The short-range abrupt change features, mid-range trend features and long-range morphological features of the time series signal are extracted separately and then concatenated by the channel dimension to obtain the multi-scale feature map. The multi-scale feature map is input into the channel attention module, and the statistical information of each channel is obtained through global average pooling. Then, the attention weights of each channel are generated through a fully connected layer with a preset compression ratio and a Sigmoid activation function. The attention weights are multiplied channel-by-channel by the multi-scale feature map to obtain a weighted feature map; The weighted feature map is input into the Transformer encoder, which captures the long-distance dependencies between different time points in the temporal data through a multi-head self-attention mechanism and positional encoding, and outputs the temporal feature map after the encoder. The temporal feature map is input into a global average pooling layer, and a global average pooling operation is performed along the time dimension to compress the variable-length feature map into a fixed-length temporal feature vector.

[0009] In one possible implementation, the preprocessed scalar process parameters are input into a pre-built scalar feature encoder to generate a scalar feature vector, including: Obtain multiple scalar process parameters corresponding to each injection cycle, and construct a scalar parameter vector according to the preset parameter order; Normalize each scalar process parameter in the scalar parameter vector to obtain a normalized scalar parameter vector; After the normalized scalar parameter vector is input into the first fully connected layer of the scalar feature encoder, the normalized scalar parameter vector is mapped to the first preset dimension, and then processed sequentially through the first batch of normalization layers, the first ReLU activation layer and the first Dropout layer to obtain the first hidden feature vector. After the first hidden feature vector is input into the second fully connected layer, the first hidden feature vector is mapped from the first preset dimension to the second preset dimension, and then processed sequentially through the second batch normalization layer, the second ReLU activation layer and the second Dropout layer to obtain the second hidden feature vector; After the second hidden feature vector is input into the third fully connected layer, the second hidden feature vector is mapped from the second preset dimension to the first preset dimension, and then processed sequentially through the third batch normalization layer, the third ReLU activation layer and the third Dropout layer to obtain the scalar feature vector.

[0010] In one possible implementation, the temporal feature vector and the scalar feature vector are fused across modally using a cross-modal attention-gated fusion module to output a joint feature vector, including: The time-series feature vectors are linearly mapped to obtain key vectors and value vectors, respectively; the scalar feature vectors are linearly mapped to obtain query vectors. The cross-modal attention weights are calculated based on the query vector and the key vector, and then the value vector is weighted and summed using these cross-modal attention weights to obtain the cross-modal attention output vector. The calculation formula is as follows: in, This is the cross-modal attention output vector. Let be the dimension of the key vector. Scaling factor The query matrix is ​​obtained by linear projection of scalar features. The key matrix is ​​obtained by linear projection of the temporal features. The value matrix is ​​obtained by linear projection of the time series features. The Softmax activation function is used. express Transpose of; The cross-modal attention output vector is concatenated with the scalar feature vector to obtain a fusion vector; The gating weights are calculated based on the fusion vector, and then the fusion vector is weighted element-wise using the gating weights to obtain the gating fusion features. The calculation formula is as follows: in, As a feature of gating fusion, For the fusion vector, For scalar eigenvectors, For element-wise multiplication, The learnable weight matrix of the gated fusion unit. For the bias term of the gated fusion unit; The gated fusion features are used to obtain a joint feature vector through linear projection.

[0011] In one possible implementation, the joint feature vector is input to a domain generalization module, and domain-invariant features are extracted through a complementary mechanism of causal feature deentanglement, multi-Gaussian distribution alignment, and invariant feature mask adversarial training, including: The joint feature vector is input into a causal feature decoupling unit, which uses a variational autoencoder structure to decouple the joint feature vector, forming a causal feature vector and a non-causal feature vector. The calculation formulas are as follows: in, For causal feature vectors, Given input hour The posterior distribution of; The mean is The covariance matrix is ​​a diagonal matrix. Gaussian distribution; These are non-causal feature vectors. Given input hour The posterior distribution, The mean of the non-causal features output by the encoder. The variance of the non-causal features output by the encoder. The mean is The covariance matrix is ​​a diagonal matrix. Gaussian distribution; The distribution of the causal feature vectors is constrained by KL divergence loss, so that the overall distribution of the causal feature vectors is aligned with the standard Gaussian prior distribution. Meanwhile, by improving the triplet loss, the inter-class separation and intra-class aggregation of the causal feature vectors are constrained, so that the feature distance between the causal feature vectors corresponding to different working conditions in the same category is less than the intra-class distance threshold, while the feature distance between the causal feature vectors corresponding to samples of different categories is greater than the inter-class distance threshold. The causal feature vectors are then input into a multi-Gaussian distribution alignment unit. Class-conditional Gaussian distributions are estimated for each fault category's causal feature vector in both the source and target domains. The difference in class-conditional covariance matrices for the same fault category under different operating conditions is calculated. By minimizing the cross-domain covariance matrix difference, the feature distribution of the same fault category remains consistent under different process conditions. The calculation formula is as follows: ; in, For covariance alignment loss, For fault category index, For the first in the source domain Class-conditional covariance matrix of causal characteristics of class-type faults. For the first in the target domain Class-conditional covariance matrix of causal characteristics of class-type faults. Denotes the square of the Frobenius norm; By inverting the gradient during backpropagation through the gradient reversal layer, a domain discriminator is trained to distinguish between the source and target domains. At the same time, a feature extractor is trained to make the domain discriminator unable to distinguish between them, forcing the feature extractor to learn domain-invariant features. By combining the causal features output by the causal feature deentanglement unit, the distribution alignment results output by the multi-Gaussian distribution alignment unit, and the adversarial training results, domain-invariant features are obtained.

[0012] In one possible implementation, a gradient inversion layer negates the gradient during backpropagation, training a domain discriminator to distinguish between the source and target domains. Simultaneously, a feature extractor is trained to render the discriminator indistinguishable, forcing the feature extractor to learn the formula for calculating domain-invariant features: in, To counteract mask loss, For sample index, For the first Domain labels for each sample, For the first The domain-invariant features output by the domain generalization module for each sample The output probability of the domain discriminator. For the discriminator to the first The target domain prediction probability is given by the invariant features of each sample.

[0013] In one possible implementation, the total loss function of the domain generalization module is: in, Let the total training loss function be the domain generalization module. Comprehensive loss due to fault classification For causal features, KL divergence loss, To counteract mask loss, For covariance alignment loss, For causal KL divergence loss The balancing hyperparameters, Covariance alignment loss The balancing hyperparameters, To combat mask loss The balancing hyperparameters, Let KL divergence be the KL divergence. For the posterior distribution, It is the standard prior distribution; For binary classification, cross-entropy loss, For multi-class cross-entropy loss, To recoup Huber's losses, , The weights are used to balance the classification and regression losses.

[0014] In one possible implementation, the extracted domain-invariant features are input into the fault detection and classification module, which simultaneously performs binary anomaly detection, multi-class fault classification, and product quality index regression prediction, outputting fault diagnosis results, including: The domain-invariant features are respectively input into a binary classification detection head, a multi-class fault classification head, and a regression prediction head that are set in parallel. The binary classification detection head is used to predict the anomaly probability of the domain-invariant features and output the failure probability of the current injection molding cycle. When the failure probability is greater than the preset anomaly detection threshold, it is determined that there is a quality anomaly in the current injection molding cycle; The multi-class fault classification head is used to predict the fault category probability of the domain invariant features, output the probability corresponding to each fault category, and determine the fault category to which the current injection molding cycle belongs based on the Softmax classifier. The regression prediction head performs quality index regression prediction on the domain-invariant features, and outputs the product weight deviation and geometric dimension deviation. Based on the failure probability, the failure category, and the deviation of the product quality index, a failure diagnosis result for the current injection molding cycle is generated.

[0015] In one possible implementation, the binary classification detection head predicts the anomaly probability of the domain-invariant features, and the formula for calculating the failure probability of the current injection molding cycle is as follows: in, This represents the probability of failure. Cross-entropy loss for binary classification The corresponding standard activation function, This is the learnable weight matrix of the binary classification detection head. For the cross-domain invariant time series features output by the domain generalization module, This represents the learnable bias term of the binary classification head, with the subscript b corresponding to the binary classification branch.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: This application systematically applies a domain generalization deep learning framework to fault diagnosis in the injection molding process, filling a gap in existing technologies. Simultaneously, it innovatively fuses high-resolution time-series signals from 2048 injection pressure / flow curves with scalar process parameters in a multimodal manner. Compared to using scalar parameters alone, this improves the anomaly detection recall rate by approximately 18.6% and reduces the false alarm rate by approximately 31.2%.

[0017] The multi-scale temporal convolution and Transformer hybrid encoder proposed in this application can effectively capture three types of fault modes during the injection process: short-range mutation, medium-range trend and long-range morphology. The inference latency on 2048 steps of temporal data is less than 15ms, which meets the real-time monitoring requirements of injection molding machine in a single cycle.

[0018] The domain generalization module in this application combines three complementary regularization strategies: causal deentanglement, multi-Gaussian alignment, and adversarial masking. Under a wide range of process fluctuations (±10% barrel temperature, ±20% injection speed), the model performance degradation does not exceed 4.3%, demonstrating excellent robustness across different operating conditions. The online adaptive update mechanism enables the model to adapt to new production conditions at extremely low cost, with rapid fine-tuning time less than 2 minutes, eliminating the need to recollect large amounts of labeled data and significantly reducing industrial deployment costs.

[0019] This method has good scalability and can be applied to injection molding machines of different specifications, injection molding materials and mold structures, and has broad industrial application potential. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a fault diagnosis method for injection molding process based on domain generalization temporal feature fusion, provided for an embodiment of this application; Figure 2 This is a structural diagram of a temporal feature encoder provided in an embodiment of this application; Figure 3 This is a structural diagram of the cross-modal attention-gated fusion module provided in an embodiment of this application; Figure 4 This is a structural diagram of the domain generalization module provided in an embodiment of this application; Figure 5 The following is a visualization of timing signals during the injection molding process provided in the embodiments of this application, wherein (a) is the normal cycle injection pressure curve, (b) is the abnormal cycle of pressure peak, (c) is the normal cycle injection flow curve, and (d) is the abnormal cycle of flow drop. Figure 6 The following are comparison charts showing the distribution of multiple key process parameters provided in the embodiments of this application: (a) is a comparison chart of the distribution of injection pressure process parameter, (b) is a comparison chart of the distribution of injection speed process parameter, (c) is a comparison chart of the distribution of barrel temperature process parameter, (d) is a comparison chart of the distribution of holding pressure process parameter, (e) is a comparison chart of the distribution of mold temperature process parameter, and (f) is a comparison chart of the distribution of cooling time process parameter. Figure 7 The following is a four-category fault diagnosis confusion matrix diagram provided in the embodiments of this application, wherein (a) is the DS1 confusion matrix and (b) is the DS3 confusion matrix. Detailed Implementation

[0021] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0022] Figure 1 A flowchart illustrating a fault diagnosis method for injection molding process based on domain generalization temporal feature fusion, provided in an embodiment of this application, is shown below. Figure 1 This embodiment of a fault diagnosis method for injection molding process based on domain generalization temporal feature fusion includes: S101: Acquire high-resolution timing data and scalar process parameters for each injection molding cycle and perform preprocessing.

[0023] In this embodiment, high-resolution timing data and scalar process parameters for each injection cycle are acquired from the injection molding machine control system. The timing data includes injection pressure timing signals and injection flow rate timing signals. Each cycle has a timing length of N steps, where N is greater than or equal to 1024. In this embodiment, 2048 time steps of injection pressure and injection flow rate timing signals are acquired per injection cycle, with a time resolution of approximately 0.025 s / step and a sampling rate of 200 Hz. The scalar process parameters include 19 scalar features such as maximum injection pressure, switching injection pressure, injection time, material pad volume, metering time, temperature of each heating zone in the barrel, mold temperature, and back pressure.

[0024] The collected data underwent normalization preprocessing. Time-series data were normalized using z-score based on batch statistics, i.e., the following was performed for each channel: in, For the normalized data, These are the original values. This is the mean of the data in this batch. This represents the standard deviation of the data in this batch. The smoothing constant takes a value of This is used to prevent division by zero errors.

[0025] Scalar parameters are normalized using min-max to eliminate dimensional differences; the calculation formula is as follows: in, The minimum value of this feature across all samples. This is the maximum value of the feature across all samples.

[0026] At the same time, based on the weight of the product Deviation from the mean of geometric dimensions Exceed Double standard deviation The criteria automatically generate quality anomaly labels: when The time was marked as abnormal. The value is 2 or 3, and the anomaly rate in normal production data is approximately 4.5%.

[0027] S102, the preprocessed high-resolution time series data is input into the pre-built time series feature encoder to generate a time series feature vector, and the preprocessed scalar process parameters are input into the pre-built scalar feature encoder to generate a scalar feature vector.

[0028] See Figure 2 In this embodiment, the temporal feature encoder includes a multi-scale convolutional block. The output of the multi-scale convolutional block is connected to the input of the channel attention module, the output of the channel attention module is connected to the input of the Transformer encoder, and the output of the Transformer encoder is connected to the input of the global average pooling layer. The multi-scale convolutional block uses three one-dimensional convolutional layers with kernel sizes of 7, 15, and 31 in parallel to capture the short-range abrupt changes in pressure spikes during the filling phase, the mid-range trend of pressure decay during the holding phase, and the long-range periodic features of the overall injection curve shape, respectively. The output feature map at each scale has a dimension of 64. The three outputs are concatenated by the channel dimension to obtain a multi-scale feature map with a dimension of 192.

[0029] The channel attention module adopts a Squeeze-and-Excitation structure. It obtains channel statistics through global average pooling, and then generates channel attention weights through two fully connected layers with a compression ratio of 16 and a sigmoid activation function. It adaptively weights the feature channels of convolutional branches at different scales to emphasize fault-related frequency components.

[0030] The Transformer encoder contains at least four self-attention layers, each containing eight attention heads with a head dimension of 32 and a feedforward network dimension of 512. It captures long-distance dependencies between different moments in the temporal data through a multi-head self-attention mechanism and positional encoding, and models the temporal context information of the entire injection cycle. The positional encoding adopts learnable positional embedding.

[0031] The global average pooling layer performs global average pooling along the time dimension, compressing the variable-length feature map into a fixed-length temporal feature vector.

[0032] The scalar feature encoder comprises a multilayer perceptron, which includes a first fully connected layer, a first batch of normalization layers, a first ReLU activation function layer, a first Dropout regularization layer, a second fully connected layer, a second batch of normalization layers, a second ReLU activation function layer, a second Dropout regularization layer, a third fully connected layer, a third batch of normalization layers, a third ReLU activation function layer, and a third Dropout regularization layer, all connected sequentially. The outputs of the global average pooling layer and the third Dropout regularization layer are both connected to the input of the cross-modal attention gating fusion module. The first fully connected layer has a dimension of 64, the second fully connected layer has a dimension of 128, and the third fully connected layer has a dimension of 64. Through layer-by-layer nonlinear mapping, the original scalar parameters are encoded into high-dimensional feature representations.

[0033] In the process of generating temporal feature vectors, preprocessed high-resolution temporal data is input into a multi-scale convolutional block. Short-range abrupt change features, mid-range trend features, and long-range morphological features of the temporal signal are extracted separately and then concatenated along the channel dimension to obtain a multi-scale feature map. This multi-scale feature map is then input into a channel attention module, where global average pooling is used to obtain statistical information for each channel. Attention weights for each channel are then generated using a fully connected layer with a preset compression ratio and a sigmoid activation function. These attention weights are multiplied channel-by-channel with the multi-scale feature map to obtain a weighted feature map. The weighted feature map is then input into a Transformer encoder, where multi-head self-attention and positional encoding are used to capture long-distance dependencies between different time points in the temporal data. The encoder-generated temporal feature map is then output and input into a global average pooling layer. Global average pooling is performed along the time dimension to compress the variable-length feature map into a fixed-length temporal feature vector.

[0034] In the process of generating scalar feature vectors, multiple scalar process parameters corresponding to each injection molding cycle are obtained, and scalar parameter vectors are constructed according to a preset parameter order. Each scalar process parameter in the scalar parameter vector is normalized to obtain a normalized scalar parameter vector. The normalized scalar parameter vector is input into the first fully connected layer of the scalar feature encoder and mapped to a first preset dimension. It is then processed sequentially through the first batch of normalization layers, the first ReLU activation layer, and the first Dropout layer to obtain a first hidden feature vector. The first hidden feature vector is input into the second fully connected layer and mapped from the first preset dimension to the second preset dimension. It is then processed sequentially through the second batch of normalization layers, the second ReLU activation layer, and the second Dropout layer to obtain a second hidden feature vector. The second hidden feature vector is input into the third fully connected layer and mapped from the second preset dimension to the first preset dimension. It is then processed sequentially through the third batch of normalization layers, the third ReLU activation layer, and the third Dropout layer to obtain a scalar feature vector.

[0035] S103 performs cross-modal feature fusion on temporal feature vectors and scalar feature vectors through a cross-modal attention-gated fusion module, and outputs a joint feature vector.

[0036] See Figure 3 In this embodiment, the cross-modal attention gating fusion module includes a query mapping layer, a key mapping layer, and a value mapping layer. The input of the query mapping layer is connected to the output of the third Dropout regularization layer. The inputs of the key mapping layer and the value mapping layer are connected to the output of the global average pooling layer. The outputs of the query mapping layer, the key mapping layer, and the value mapping layer are all connected to the input of the cross-modal attention calculation unit. The outputs of the cross-modal attention calculation unit and the third Dropout regularization layer are both connected to the input of the gating fusion unit. The output of the gating fusion unit is connected to the input of the dimensionality reduction projection layer. The output of the dimensionality reduction projection layer is connected to the input of the domain generalization module.

[0037] The temporal feature vector is linearly mapped to obtain the key vector and value vector, respectively. The scalar feature vector is linearly mapped to obtain the query vector. Cross-modal attention weights are calculated based on the query vector and the key vector. The cross-modal attention weights are then used to perform a weighted summation of the value vector to obtain the cross-modal attention output vector. The calculation formula is as follows: in, This is the cross-modal attention output vector. Let be the dimension of the key vector. Scaling factor The query matrix is ​​obtained by linear projection of scalar features. The key matrix is ​​obtained by linear projection of the temporal features. The value matrix is ​​obtained by linear projection of the time series features. The Softmax activation function is used. express The transpose of .

[0038] The cross-modal attention output vector is concatenated with the scalar feature vector to obtain a fusion vector. Gating weights are then calculated based on the fusion vector, and the fusion vector is weighted element-wise using these gating weights to obtain the gated fusion features. The calculation formula is as follows: in, As a feature of gating fusion, For the fusion vector, For scalar eigenvectors, For element-wise multiplication, The learnable weight matrix of the gated fusion unit. As the bias term of the gated fusion unit, the gated fusion features are linearly projected to obtain the joint feature vector. S104 inputs the joint feature vector to the domain generalization module, and extracts the domain invariant features through the complementary mechanisms of causal feature deentanglement, multi-Gaussian distribution alignment and invariant feature mask adversarial training.

[0039] See Figure 4 In this embodiment, the domain generalization module includes a causal feature deentanglement unit. The input of the causal feature deentanglement unit is connected to the output of the dimension reduction projection layer. The output of the causal feature deentanglement unit is connected to the input of the multi-Gaussian distribution alignment unit and the input of the invariant feature mask adversarial training unit, respectively. The invariant feature mask adversarial training unit includes a gradient inversion layer and a domain discriminator. The input of the domain invariant feature output unit is connected to the output of the causal feature deentanglement unit, the output of the multi-Gaussian distribution alignment unit, and the output of the invariant feature mask adversarial training unit, respectively. The output of the domain invariant feature output unit is connected to the fault detection and classification module.

[0040] This embodiment addresses the domain drift problem caused by variations in barrel temperature (±5%~±10%) and injection speed (±10%~±20%) by designing three complementary domain generalization mechanisms.

[0041] The joint feature vector is input into the causal feature decoupling unit, which uses a variational autoencoder structure to decouple the joint feature vector, forming causal and non-causal feature vectors. The calculation formulas are as follows: in, For causal feature vectors, Given input hour The posterior distribution of; The mean is The covariance matrix is ​​a diagonal matrix. Gaussian distribution; These are non-causal feature vectors. Given input hour The posterior distribution, The mean of the non-causal features output by the encoder. The variance of the non-causal features output by the encoder. The mean is The covariance matrix is ​​a diagonal matrix. The Gaussian distribution.

[0042] The distribution of causal feature vectors is constrained by KL divergence loss, aligning the overall distribution of causal feature vectors with the standard Gaussian prior distribution. At the same time, the inter-class separation and intra-class aggregation of causal feature vectors are constrained by improved triplet loss, so that the feature distance between causal feature vectors corresponding to different working conditions in the same class is less than the intra-class distance threshold, while the feature distance between causal feature vectors corresponding to samples of different classes is greater than the inter-class distance threshold.

[0043] Subsequently, the causal feature vectors are input into a multi-Gaussian distribution alignment unit. The class-conditional Gaussian distributions of the causal feature vectors for each fault category are estimated in both the source and target domains. The differences in the class-conditional covariance matrices for the same fault category under different operating conditions are calculated. By minimizing the cross-domain covariance matrix differences, the feature distributions of the same fault category are kept consistent under different process conditions. The calculation formula is as follows: ; in, For covariance alignment loss, For fault category index, For the first in the source domain Class-conditional covariance matrix of causal characteristics of class-type faults. For the first in the target domain Class-conditional covariance matrix of causal characteristics of class-type faults. This represents the square of the Frobenius norm.

[0044] By inverting the gradient during backpropagation using a gradient reversal layer, a domain discriminator is trained to distinguish between the source and target domains. Simultaneously, a feature extractor is trained to render the discriminator indistinguishable, forcing the feature extractor to learn domain-invariant features. The calculation formula is as follows: in, To counteract mask loss, For sample index, For the first Domain labels for each sample, For the first The domain-invariant features output by the domain generalization module for each sample The output probability of the domain discriminator. For the discriminator to the first The target domain prediction probability is given by the sample-invariant features. The domain-invariant features are obtained by combining the causal features output by the causal feature de-entanglement unit, the distribution alignment results output by the multi-Gaussian distribution alignment unit, and the adversarial training results.

[0045] The total loss function of the domain generalization module is: in, Let the total training loss function be the domain generalization module. Comprehensive loss due to fault classification For causal features, KL divergence loss, To counteract mask loss, For covariance alignment loss, For causal KL divergence loss The balancing hyperparameters, Covariance alignment loss The balancing hyperparameters, To combat mask loss The balancing hyperparameters, Let KL divergence be the KL divergence. For the posterior distribution, It is the standard prior distribution; For binary classification, cross-entropy loss, For multi-class cross-entropy loss, To recoup Huber's losses, , The weights are used to balance the classification and regression losses.

[0046] S105 inputs the extracted domain-invariant features into the fault detection and classification module, and simultaneously performs binary anomaly detection, multi-class fault classification and product quality index regression prediction, and outputs fault diagnosis results.

[0047] In this embodiment, the domain-invariant features are input to a binary classification detection head, a multi-class fault classification head, and a regression prediction head that are set in parallel. The binary classification detection head predicts the anomaly probability of the domain-invariant features and outputs the fault probability of the current injection molding cycle. The calculation formula is as follows: in, This represents the probability of failure. Cross-entropy loss for binary classification The corresponding standard activation function, This is the learnable weight matrix of the binary classification detection head. For the cross-domain invariant time series features output by the domain generalization module, This represents the learnable bias term of the binary classification detection head, with the subscript b corresponding to the binary classification branch. When the failure probability is greater than the preset abnormality detection threshold, it is determined that there is a quality abnormality in the current injection molding cycle.

[0048] Fault category probability prediction is performed on domain-invariant features using a multi-class fault classification head, and the probability corresponding to each fault category is output. The calculation formula is as follows: in, For multi-type fault classification, the fault probability is given. The Softmax activation function is used. The learnable weight matrix for multi-class fault classification heads. is the learnable bias vector for the multi-class fault classification head, where the subscript c corresponds to the multi-class branch.

[0049] The system uses a Softmax classifier to determine the fault category of the current injection molding cycle. The fault categories include: normal, abnormal pressure, abnormal flow, and temperature drift.

[0050] The regression prediction head performs quality index regression prediction on the domain-invariant features, outputs product weight deviation and geometric dimension deviation, and generates fault diagnosis results for the current injection molding cycle based on fault probability, fault type and product quality index deviation.

[0051] In addition, it includes an automatic abnormal label generation step and an online adaptive update step: In the automatic label generation step, the data in the machine startup warm-up stage is labeled as a transitional state and does not participate in fault classification training. Normal and abnormal samples are automatically balanced, and when the abnormality rate is less than 5%, an oversampling strategy is used to compensate for class imbalance.

[0052] In the online adaptive update step, when the model inference confidence is detected to be lower than the threshold for N consecutive cycles, the online fine-tuning mechanism is triggered. Data from at least 50 injection molding cycles are collected from the new working conditions, all weights of the temporal feature encoder and scalar feature encoder are frozen, and only the top-level parameters of the classification head and domain generalization module are fine-tuned with a few samples. After the fine-tuning is completed, the new model parameters are pushed through OTA. The entire update process takes no more than 2 minutes and does not interrupt normal production.

[0053] Secondly, this application also provides an injection molding process anomaly detection and fault diagnosis system based on domain generalization temporal feature fusion, including: a data acquisition module, a temporal feature extraction module, a scalar feature encoding module, a cross-modal fusion module, a domain generalization module, a diagnostic output module, and an edge computing unit, each module performing the corresponding function in the above method steps.

[0054] Subsequently, experimental verification was conducted using a publicly available injection molding dataset published by the University of Kassel, Germany. The dataset contains two classes of injection molded parts: DS1 (Housing Part, approximately 58.9 g in weight, approximately 663 bar injection pressure) and DS3 (Stacking Box II, approximately 113.5 g in weight, approximately 1001 bar injection pressure), representing two domains under different process conditions. DS1 contains 1167 injection cycles, and DS3 contains 1332 injection cycles.

[0055] Each injection molding cycle synchronously collects 2048 injection pressure and injection flow rate timing signals, as well as over 30 scalar process parameters including injection pressure, holding pressure, metering time, material pad volume, back pressure, injection time, and barrel temperature in each zone. Quality anomaly labels are based on product weight. The criteria are automatically generated, and the DS1 anomaly rate is approximately 4.5%.

[0056] The temporal feature encoder is implemented as follows: the multi-scale convolutional block adopts three parallel one-dimensional convolutions with kernel sizes of 7, 15, and 31, a stride of 1, and 64 output channels. After BN and ReLU activation, the channels are concatenated to form a 192-dimensional feature map. The SE channel attention compression ratio is 16. The Transformer encoder contains 4 layers, each with 8 self-attention heads and a head dimension of 32. The feedforward network width is 512, and the Dropout is 0.1. Layer normalization is applied before the attention and feedforward networks. Finally, the temporal feature vector is output after global average pooling.

[0057] Cross-modal attention gating fusion module (CMAG): Calculates cross-modal attention weights based on query vector and key vector, and uses the cross-modal attention weights to perform a weighted summation on the value vector to obtain the cross-modal attention output vector. After concatenating the cross-modal attention output vector with the scalar feature vector, it is mapped through a linear layer to obtain the gating weights. After gating, it is linearly projected to obtain the joint feature vector.

[0058] Domain Generalization Module (DGM): In the causal unentanglement unit, the encoder output mean and logarithmic variance belong to R64; in the multi-Gaussian alignment unit, the class conditional covariance matrix is ​​calculated from intra-batch samples of the same class; in the adversarial mask, the domain discriminator is a 2-layer fully connected network with a gradient inversion coefficient of 1.0, and is trained alternately by the Adam optimizer.

[0059] See Figure 5 Representative normal and abnormal periods from the DS1 and DS3 datasets were used for time-series signal visualization. (a) shows the injection pressure curve during the normal period, (b) shows the abnormal pressure peak period, (c) shows the injection flow rate curve during the normal period, and (d) shows the abnormal flow rate decline period. The figures show that the injection pressure curve exhibits a typical rapid rise-slow decay waveform during the normal period; the abnormal pressure peak period shows an abnormal peak value during the pressure holding phase, exceeding the normal value by approximately 30%; and the abnormal flow rate decline period shows a lower flow rate during the injection phase, dropping to approximately 55% of the normal level. This verifies that the time-series signals contain rich fault discrimination information.

[0060] See Figure 6 The distribution of DS1 and DS3 across six key process parameters—injection pressure, injection speed, barrel temperature, holding pressure, mold temperature, and cooling time—is compared. (a) shows the distribution of injection pressure, (b) shows the distribution of injection speed, (c) shows the distribution of barrel temperature, (d) shows the distribution of holding pressure, (e) shows the distribution of mold temperature, and (f) shows the distribution of cooling time. The figures show significant differences in injection pressure, barrel temperature, and holding pressure, verifying the existence of cross-domain distribution drift and highlighting the necessity of domain generalization methods.

[0061] See Figure 7Figure (a) shows the DS1 confusion matrix, and Figure (b) shows the DS3 confusion matrix. In Figure (a), the recall rate for the normal class is 98.1%, the recall rate for the pressure spike class is 90.2%, the recall rate for the flow decrease class is 87.8%, the recall rate for the temperature drift class is 85.4%, and the overall F1 score is 91.7%. In Figure (b), the overall F1 score for DS3 is 90.4%, and the recall rates for each class remain above 86%, verifying the robustness of the model in cross-domain scenarios.

[0062] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0063] The above description is merely a specific embodiment of this application. 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 protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A fault diagnosis method for injection molding process based on domain generalization temporal feature fusion, characterized in that, include: Acquire high-resolution timing data and scalar process parameters for each injection molding cycle and perform preprocessing; The preprocessed high-resolution time series data is input into a pre-built time series feature encoder to generate a time series feature vector. At the same time, the preprocessed scalar process parameters are input into a pre-built scalar feature encoder to generate a scalar feature vector. The temporal feature vector and the scalar feature vector are fused across modally using a cross-modal attention-gated fusion module to output a joint feature vector; The joint feature vector is input into the domain generalization module, and the domain invariant features are extracted through the complementary mechanisms of causal feature deentanglement, multi-Gaussian distribution alignment, and invariant feature mask adversarial training. The extracted domain-invariant features are input into the fault detection and classification module, which simultaneously performs binary anomaly detection, multi-class fault classification, and product quality index regression prediction, and outputs fault diagnosis results. The temporal feature encoder includes a multi-scale convolutional block, the output of which is connected to the input of a channel attention module. The channel attention module adopts a Squeeze-and-Excitation structure, and its output is connected to the input of a Transformer encoder. The output of the Transformer encoder is connected to the input of a global average pooling layer. The scalar feature encoder includes a multilayer perceptron, which includes a first fully connected layer, a first batch normalization layer, a first ReLU activation function layer, a first Dropout regularization layer, a second fully connected layer, a second batch normalization layer, a second ReLU activation function layer, a second Dropout regularization layer, a third fully connected layer, a third batch normalization layer, a third ReLU activation function layer, and a third Dropout regularization layer connected in sequence. The output of the global average pooling layer and the output of the third Dropout regularization layer are both connected to the input of the cross-modal attention gating fusion module. The cross-modal attention gating fusion module includes a query mapping layer, a key mapping layer, and a value mapping layer. The input of the query mapping layer is connected to the output of the third Dropout regularization layer. The inputs of the key mapping layer and the value mapping layer are connected to the output of the global average pooling layer. The outputs of the query mapping layer, the key mapping layer, and the value mapping layer are all connected to the input of the cross-modal attention calculation unit. The outputs of the cross-modal attention calculation unit and the third Dropout regularization layer are both connected to the input of the gating fusion unit. The output of the gating fusion unit is connected to the input of the dimensionality reduction projection layer. The output of the dimensionality reduction projection layer is connected to the input of the domain generalization module. The domain generalization module includes a causal feature deentanglement unit, the input of which is connected to the output of the dimensionality reduction projection layer. The output of the causal feature deentanglement unit is connected to the input of a multi-Gaussian distribution alignment unit and an invariant feature mask adversarial training unit, respectively. The invariant feature mask adversarial training unit includes a gradient inversion layer and a domain discriminator. The input of the domain invariant feature output unit is connected to the output of the causal feature deentanglement unit, the multi-Gaussian distribution alignment unit, and the invariant feature mask adversarial training unit, respectively. The output of the domain invariant feature output unit is connected to the fault detection and classification module.

2. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 1, characterized in that, The preprocessed high-resolution time-series data is input into a pre-constructed time-series feature encoder to generate time-series feature vectors, including: The preprocessed high-resolution time series data is input into the multi-scale convolutional block. The short-range mutation features, mid-range trend features, and long-range morphological features of the time series signal are extracted separately and then concatenated by the channel dimension to obtain a multi-scale feature map. The multi-scale feature map is input into the channel attention module, and the statistical information of each channel is obtained through global average pooling. Then, the attention weights of each channel are generated through a fully connected layer with a preset compression ratio and a Sigmoid activation function. The attention weights are multiplied channel-by-channel by the multi-scale feature map to obtain a weighted feature map; The weighted feature map is input into the Transformer encoder, which captures the long-distance dependencies between different time points in the temporal data through a multi-head self-attention mechanism and positional encoding, and outputs the temporal feature map after the encoder. The temporal feature map is input into a global average pooling layer, and a global average pooling operation is performed along the time dimension to compress the variable-length feature map into a fixed-length temporal feature vector.

3. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 1, characterized in that, The preprocessed scalar process parameters are input into a pre-constructed scalar feature encoder to generate a scalar feature vector, including: Obtain multiple scalar process parameters corresponding to each injection cycle, and construct a scalar parameter vector according to the preset parameter order; Normalize each scalar process parameter in the scalar parameter vector to obtain a normalized scalar parameter vector; After the normalized scalar parameter vector is input into the first fully connected layer of the scalar feature encoder, the normalized scalar parameter vector is mapped to the first preset dimension, and then processed sequentially through the first batch of normalization layers, the first ReLU activation layer and the first Dropout layer to obtain the first hidden feature vector. After the first hidden feature vector is input into the second fully connected layer, the first hidden feature vector is mapped from the first preset dimension to the second preset dimension, and then processed sequentially through the second batch normalization layer, the second ReLU activation layer and the second Dropout layer to obtain the second hidden feature vector; After the second hidden feature vector is input into the third fully connected layer, the second hidden feature vector is mapped from the second preset dimension to the first preset dimension, and then processed sequentially through the third batch normalization layer, the third ReLU activation layer and the third Dropout layer to obtain the scalar feature vector.

4. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 1, characterized in that, The temporal feature vector and the scalar feature vector are fused across modally using a cross-modal attention-gated fusion module to output a joint feature vector, including: The time-series feature vectors are linearly mapped to obtain key vectors and value vectors, respectively; the scalar feature vectors are linearly mapped to obtain query vectors. The cross-modal attention weights are calculated based on the query vector and the key vector, and then the value vector is weighted and summed using these cross-modal attention weights to obtain the cross-modal attention output vector. The calculation formula is as follows: in, This is the cross-modal attention output vector. Let be the dimension of the key vector. This is the scaling factor. The query matrix is ​​obtained by linear projection of scalar features. The key matrix is ​​obtained by linear projection of the temporal features. The value matrix is ​​obtained by linear projection of the time series features. The Softmax activation function is used. express Transpose of; The cross-modal attention output vector is concatenated with the scalar feature vector to obtain a fusion vector; The gating weights are calculated based on the fusion vector, and then the fusion vector is weighted element-wise using the gating weights to obtain the gating fusion features. The calculation formula is as follows: in, As a feature of gating fusion, For the fusion vector, For scalar eigenvectors, For element-wise multiplication, The learnable weight matrix of the gated fusion unit. For the bias term of the gated fusion unit; The gated fusion features are used to obtain a joint feature vector through linear projection.

5. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 1, characterized in that, The joint feature vector is input into the domain generalization module, and domain-invariant features are extracted through a complementary mechanism of causal feature deentanglement, multi-Gaussian distribution alignment, and invariant feature mask adversarial training, including: The joint feature vector is input into a causal feature decoupling unit, which uses a variational autoencoder structure to decouple the joint feature vector, forming a causal feature vector and a non-causal feature vector. The calculation formulas are as follows: in, For causal feature vectors, Given input hour The posterior distribution of; The mean is The covariance matrix is ​​a diagonal matrix. Gaussian distribution; These are non-causal feature vectors. Given input hour The posterior distribution, The mean of the non-causal features output by the encoder. The variance of the non-causal features output by the encoder. The mean is The covariance matrix is ​​a diagonal matrix. Gaussian distribution; The distribution of the causal feature vectors is constrained by KL divergence loss, so that the overall distribution of the causal feature vectors is aligned with the standard Gaussian prior distribution. Meanwhile, by improving the triplet loss, the inter-class separation and intra-class aggregation of the causal feature vectors are constrained, so that the feature distance between the causal feature vectors corresponding to different working conditions in the same category is less than the intra-class distance threshold, while the feature distance between the causal feature vectors corresponding to samples of different categories is greater than the inter-class distance threshold. The causal feature vectors are then input into a multi-Gaussian distribution alignment unit. Class-conditional Gaussian distributions are estimated for each fault category's causal feature vector in both the source and target domains. The difference in class-conditional covariance matrices for the same fault category under different operating conditions is calculated. By minimizing the cross-domain covariance matrix difference, the feature distribution of the same fault category remains consistent under different process conditions. The calculation formula is as follows: ； in, For covariance alignment loss, For fault category index, For the first in the source domain Class-conditional covariance matrix of causal characteristics of class-type faults. For the first in the target domain Class-conditional covariance matrix of causal characteristics of class-type faults. Denotes the square of the Frobenius norm; By inverting the gradient during backpropagation through the gradient reversal layer, a domain discriminator is trained to distinguish between the source and target domains. At the same time, a feature extractor is trained to make the domain discriminator unable to distinguish between them, forcing the feature extractor to learn domain-invariant features. By combining the causal features output by the causal feature deentanglement unit, the distribution alignment results output by the multi-Gaussian distribution alignment unit, and the adversarial training results, domain-invariant features are obtained.

6. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 5, characterized in that, By inverting the gradient during backpropagation using a gradient reversal layer, a domain discriminator is trained to distinguish between the source and target domains. Simultaneously, a feature extractor is trained to render the discriminator indistinguishable, forcing the feature extractor to learn the formula for calculating domain-invariant features: in, To counteract mask loss, For sample index, For the first Domain labels for each sample, For the first The domain-invariant features output by the domain generalization module for each sample The output probability of the domain discriminator. For the discriminator to the first The target domain prediction probability is given by the invariant features of each sample.

7. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 1, characterized in that, The total loss function of the domain generalization module is: in, Let the total training loss function be the domain generalization module. Comprehensive loss due to fault classification For causal features, KL divergence loss, To counteract mask loss, For covariance alignment loss, For causal KL divergence loss The balancing hyperparameters, Covariance alignment loss The balancing hyperparameters, To combat mask loss The balancing hyperparameters, Let KL divergence be the KL divergence. For the posterior distribution, It is the standard prior distribution; For binary classification, cross-entropy loss, For multi-class cross-entropy loss, To recoup Huber's losses, , The weights are used to balance the classification and regression losses.

8. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 1, characterized in that, The extracted domain-invariant features are input into the fault detection and classification module, which simultaneously performs binary anomaly detection, multi-class fault classification, and product quality index regression prediction, outputting fault diagnosis results, including: The domain-invariant features are respectively input into a binary classification detection head, a multi-class fault classification head, and a regression prediction head that are set in parallel. The binary classification detection head is used to predict the anomaly probability of the domain-invariant features and output the failure probability of the current injection molding cycle. When the failure probability is greater than the preset anomaly detection threshold, it is determined that there is a quality anomaly in the current injection molding cycle; The multi-class fault classification head is used to predict the fault category probability of the domain invariant features, output the probability corresponding to each fault category, and determine the fault category to which the current injection molding cycle belongs based on the Softmax classifier. The regression prediction head performs quality index regression prediction on the domain-invariant features, and outputs the product weight deviation and geometric dimension deviation. Based on the failure probability, the failure category, and the deviation of the product quality index, a failure diagnosis result for the current injection molding cycle is generated.

9. The method for fault diagnosis of injection molding process based on domain generalization temporal feature fusion according to claim 8, characterized in that, The binary classification detection head predicts the anomaly probability of the domain-invariant features, and the formula for calculating the failure probability of the current injection molding cycle is as follows: in, This represents the probability of failure. Cross-entropy loss for binary classification The corresponding standard activation function, This is the learnable weight matrix of the binary classification detection head. For the cross-domain invariant time series features output by the domain generalization module, This represents the learnable bias term of the binary classification head, with the subscript b corresponding to the binary classification branch.

Images