Power quality disturbance positioning and identification method based on phase perception and multi-modal fusion

Abstract

This invention discloses a power quality disturbance location and identification method based on phase sensing and multimodal fusion, relating to the field of power system monitoring and fault diagnosis technology. This power quality disturbance location and identification method based on phase sensing and multimodal fusion includes the following steps: acquiring the raw voltage signal of a three-phase power system; extracting the sample embedding vector of each phase based on the raw voltage signal and comparing it with the normal sample vector, and determining the disturbed phase of the power quality disturbance by combining an adaptive threshold; extracting time-series feature vectors and bispectral image features from the disturbed phase, and aligning and co-fusing them to obtain fused features; identifying the power quality disturbance type based on the fused features, achieving phase-level location of the three-phase PQD: through an innovative phase-sensing embedding and prototype comparison mechanism, it achieves accurate and rapid location of abnormal phases in a three-phase system for the first time without relying on the disturbed phase label, solving the core pain points of traditional methods.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring and fault diagnosis technology, specifically to a power quality disturbance location and identification method based on phase sensing and multimodal fusion. Background Technology

[0002] Power quality disturbances are a significant factor affecting the safe and stable operation of power systems. With the large-scale grid connection of renewable energy and the widespread application of complex industrial loads, power quality disturbance problems are becoming increasingly complex. Traditional disturbance identification methods are mostly designed for single-phase systems, extracting features through signal processing techniques (such as Fourier transform and wavelet transform) and combining them with machine learning algorithms for classification.

[0003] However, practical power transmission networks generally adopt a three-phase structure, and phase asymmetry is a common problem. Traditional single-phase identification models analyze the signals of each phase in isolation, which has the fundamental defect of missing phase coupling information, making it impossible to accurately locate the disturbed phase. When facing real three-phase systems, the identification accuracy and generalization ability are significantly reduced.

[0004] In recent years, deep learning has provided new insights for PQD (Problem-to-Device) recognition. Two main technical approaches have emerged: one is an end-to-end method based on one-dimensional convolutional neural networks, which, while computationally efficient, lacks sufficient representation of global information such as the perturbation's time-frequency structure and nonlinear dynamic characteristics; the other is a two-stage method based on feature encoding and image classification. This approach converts the original signal into an image (such as a Gram angle field, recursive graph, or time-frequency graph), and then uses an image classification model for recognition. However, the performance of this type of method is highly dependent on the choice of encoding method; inappropriate conversion can lead to redundant or distorted image information. For example, common time-frequency analysis methods (such as short-time Fourier transform) cannot preserve the phase information between frequency components and require long measurement times to obtain high resolution.

[0005] Bispectral analysis, as a type of higher-order spectral analysis, can detect phase coupling information between frequency components in a signal. It has unique advantages in revealing the characteristics of disturbances caused by system nonlinearity and performs excellently in noisy signal processing and short-time data analysis. Although bispectral analysis has been applied in rotating machinery fault diagnosis, its potential in three-phase power quality disturbances, especially in achieving coordinated diagnosis of disturbance phase location and fine-grained identification, has not yet been fully explored.

[0006] Therefore, there is an urgent need for a three-phase power quality disturbance intelligent diagnosis solution that can overcome the above-mentioned defects and has high positioning accuracy, high recognition rate, strong robustness and high computational efficiency. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a power quality disturbance location and identification method based on phase sensing and multimodal fusion. This method aims to achieve: 1. Unlike the traditional method of splitting the three phases and calculating them separately, this method can simultaneously identify three-phase disturbances and accurately locate the phase level of disturbances in a three-phase system.

[0008] 2. Perform fine-grained perturbation type identification on the located perturbation phase.

[0009] 3. While ensuring high accuracy, the two-stage inference mechanism significantly improves computational efficiency.

[0010] 4. Enhance the robustness of the model in complex noisy environments by utilizing bispectral images and multimodal fusion.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a power quality disturbance location and identification method based on phase sensing and multimodal fusion, comprising the following steps: Acquire the raw voltage signal of a three-phase power system; The sample embedding vectors of each phase are extracted based on the original voltage signal and compared with the normal sample vectors. The perturbation phase that has caused power quality disturbance is determined by combining the adaptive threshold. Temporal feature vector extraction and bispectral image feature extraction are performed on the perturbation phase, and alignment and co-fusion are performed to obtain fused features; Identification of power quality disturbance types based on fused features.

[0012] Furthermore, the process for obtaining a normal sample vector is as follows: Obtain the normalized, denoised, and fixed-length truncated three-phase normal voltage signal X: X={x A x B x c}, where x A x B x c These are the instantaneous voltage values ​​of phases A, B, and C at a certain moment; A three-layer one-dimensional convolutional network is used as the encoder f θ Single-phase independent feature extraction is performed on the three-phase normal voltage signal X, and a high-dimensional feature vector h is output. p Then, a two-layer fully connected network is used as the projection head g. φ , the feature vector h p Mapping to a low-dimensional embedding space and normalizing it, the output is a 256-dimensional normalized embedding vector e. p : ; Where p is the phase index; Comparison of learning objectives using normalized temperature-scaled cross-entropy loss: ; Among them (e) p , i ,e p,j ) represents a positive sample pair, sim(e p , i ,e p,j ) represents cosine similarity, ε is the temperature scaling factor, i represents the current sample index, j represents the positive sample index, and L p contrast Represents the comparative loss, when L p contrast When minimizing, two normal samples e of the same phase are... p,i and e p,j Aggregate to obtain aggregate group D normal ; For aggregation group D normal The mean of all normal samples is calculated to obtain the normal sample vector: ; Where N p It is aggregation group D normal The number of normal samples, c p This is the normal sample vector for phase p.

[0013] Furthermore, the process of extracting the sample embedding vectors of each phase based on the original voltage signal, comparing them with the normal sample vectors, and determining the disturbed phase of the power quality disturbance by combining an adaptive threshold is as follows: The original three-phase voltage signal X1 is processed by a one-dimensional convolution encoder f. θ Perform single-phase independent feature extraction, and then pass it through the projection head g. φ By performing mapping, the sample embedding vector e of each phase is obtained. p ; The extracted phase samples are embedded into vector e. p With the normal prototype vector c p Distance quantization is performed between them, and the Euclidean distance d is calculated. p : ; Generate a binary phase mask m by combining an adaptive threshold τ. p : ; Phase mask m p When the value is 1, it indicates that the phase disturbance is abnormal, that is, the disturbing phase that causes power quality disturbance, and the phase mask m p A value of 0 indicates that the disturbance in that phase is normal.

[0014] Furthermore, the adaptive threshold acquisition process is as follows: If the normal sample distance set D p If the distribution approximately follows a Gaussian distribution, indicating that the three-phase power system is in steady-state operation, then the formula for calculating the adaptive threshold is: ; Where, μd p It is the distance mean, σd p is the standard deviation, and k is a constant for controlling sensitivity; If the normal sample distance set D p If the distribution is non-Gaussian, then the highest anomaly samples need to be excluded using statistical boundaries. Therefore, the formula for calculating the adaptive threshold is: ; Where a is the threshold coefficient, Percentile 95 It is the 95th percentile boundary; If the normal sample distance set D p If measurement noise and outliers exist, the robustness of the threshold calculation needs to be improved. Therefore, the formula for calculating the adaptive threshold is: ; Where b is the threshold coefficient, median(D) p ) represents the distance from the median, MAD(D) p ) represents the median absolute difference, and m is a constant used to adjust the MAD to be equivalent to the standard deviation under a Gaussian distribution.

[0015] Furthermore, the bispectral image acquisition process is as follows: Bispectral analysis was performed on the voltage signal of the disturbed phase to calculate its bispectral matrix. : ; Where f1 and f2 are two independent frequency components, and P(f1) and P(f2) are the Fourier transform coefficients of the voltage signal at frequencies f1 and f2, respectively. * (f 1+ f2) is the signal at frequency (f 1+ The Fourier transform coefficients P(f2) at f2) 1+ The complex conjugate of f2); The bispectral matrix is ​​visualized as a bispectral image in the form of a contour plot.

[0016] Furthermore, the process of extracting the temporal feature vector of the perturbation phase is as follows: Let the three-phase original voltage timing signal containing the disturbance phase be denoted as X∈R. 3×L Its length is fixed at L=1024 sampling points; Let X∈R 3×L The input is fed into a lightweight hierarchical structure constructed from three layers of depthwise separable convolutions, batch normalization, and activation functions to obtain temporal feature vectors: ; ; Where F l Let L be the output feature of layer l, l be the current feature extraction layer number, BN(.) be the batch normalization function, ReLU(.) be the activation function, and DSConv1D be the output feature of layer l. 5×1 (.) represents a one-dimensional depthwise separable convolution with a kernel size of 5×1, F l (:,i) represents the characteristics of all channels of the output of the l-th layer at time step i, f t This is the final temporal feature vector.

[0017] Furthermore, the process of feature extraction from bispectral images is as follows: For bispectral image X BSP The input is fed into a convolutional embedding layer to extract local spatial patterns, and the output of the convolutional layer is a cross-scale feature F. conv : ; Among them, K d and K p These are depthwise convolution kernels and pointwise convolution kernels, respectively; σ(·) represents the ReLU nonlinear activation function; and BN is batch normalization. By capturing responses at different spatial scales through parallel average pooling and max pooling, and after linear mapping, the cross-scale features F conv They are unified into the same embedding space, forming multi-scale features F ms : ; Where Avg() represents the average pooling response, Max() represents the max pooling response, Concat(,) represents the concatenation operation, and W s and b s Represents the linear mapping parameters; Multiscale features F ms Input a Transformer structure, output a sequence of tokens: ; Where T is the token sequence, W e It is a linear embedding matrix; The token sequence is processed using a multi-head attention mechanism, and the attention output is A(Q,K,V): ; Where Q, K, and V are the query, key, and value vectors, respectively, and d k Let QK be the dimension of the key vector, Softmax(.) be the normalization function, and QK be the dimension of the key vector. T Attention score; The multi-head attention mechanism splits the Q, K, and V matrices into H independent heads, performs A(Q,K,V) computation in parallel, concatenates the outputs of all heads, and then projects them through a linear layer W. out Get the output h MHA : ; h MHA This is the output of the multi-head attention mechanism sublayer, which needs to be residually connected and normalized with the input T of this layer to obtain the processed feature sequence T': ; The processed feature sequence T' is then processed by a feedforward network and global pooling to obtain the final image features f. v : ; In this function, GlobalPool(.) is the global pooling function, LayerNorm(.) is the layer normalization function, and FFN(.) is the feedforward network operation.

[0018] Furthermore, the process of aligning and co-fusing the temporal feature vector and bispectral image features to obtain the fused features is as follows: The time series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C After weighted processing, the merged features are spliced ​​together to form a preliminary aligned fusion feature f. aligned ; Using linear layer pairs f aligned Dimensionality reduction is performed to obtain the reduced feature h0: ; Among them, W aligned b0 and b0 are the weights and biases of the linear dimensionality reduction layer; Inputting h0 into a multilayer perceptron, which becomes the core of collaborative fusion: ; ; Where W1, b1, W2, and b2 are the weights and biases of the multilayer perceptron, ReLU(.) is the activation function, DroPout() is the regularization function, h1 is the output of the intermediate hidden layer of the multilayer perceptron, and h2 is the final output of the multilayer perceptron. Finally, a residual connection is introduced to add the dimensionality-reduced feature h0 to the output h2 of the MLP, and then the fused output is obtained through layer normalization: ; Where LayerNorm() is the layer normalization function, f fused It is the final fusion feature.

[0019] Furthermore, the step of transferring the time-series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C After weighted processing, the merged features are spliced ​​together to form a preliminary aligned fusion feature f. aligned The process is as follows: The time series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C Projected onto the common dimension D through a shared linear layer u And an importance weight for each feature is generated through a lightweight gating network: ; Where f i z represents the i-th input feature vector. i It is the projected feature, W p (i) ∈R Du*Di and b p (i) It is the projection matrix and bias of the feature type, W g and b g These are shared gated network parameters, σ(·) is the Sigmoid function, which compresses the weights to the (0,1) interval, and α i It is f i Adaptive importance weights; The projected features are weighted according to their importance and then concatenated to form a preliminarily aligned fused feature f. aligned : ; Where α t α v、 α A、 α B and α C These are the feature vectors z after projection through the gating network. t z v、 z A、 z B and zC The adaptive importance weights are given, and [;] represents the vector concatenation operation.

[0020] Furthermore, the process of identifying power quality disturbance types based on fused features is as follows: After the perturbation phase localization stage, a binary anomaly mask is generated, based on the mask label M. i,j Selective activation of the classification head: ; Where I is the indicator function, with an output of 0 for normal and 1 for abnormal (abnormal is activated by the classification head), d(,) is the distance function, and e i,j It is the embedding vector of the i-th sample and the j-th phase, c j It is the prototype center of the j-th phase, τ j It is the threshold for determining the j-th phase; The final category label C is selected through a gating mechanism. j : ; Among them, W j It is the weight matrix of the j-th class head, b j C1 is the corresponding bias vector, argmax() is the aggregation function that yields the predicted class label, and C1 is the normal class label.

[0021] The present invention has the following beneficial effects: 1. Phase-level localization of three-phase PQD was achieved: Through an innovative phase-sensing embedding and prototype comparison mechanism, the accurate and rapid localization of abnormal phases in a three-phase system was achieved for the first time without relying on the perturbation phase label, solving the core pain point of traditional methods.

[0022] 2. Improved feature representation completeness and robustness: The model extracts temporal local features and global image features in parallel through a dual-branch structure, and uses bispectral analysis to mine the phase coupling information of the signal. Then, information complementarity is achieved through a heterogeneous fusion module, which makes the model have a strong representation ability and robustness against complex disturbances and noise.

[0023] 3. The computational efficiency and practicality of the model have been optimized: The two-stage inference mechanism ensures that complex multimodal classification calculations only apply to the abnormal phases, effectively skipping the calculations of a large number of normal phases, significantly reducing the average inference time of the model, and making it more suitable for industrial scenarios with high real-time requirements.

[0024] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0025] Figure 1This is a flowchart of the power quality disturbance location and identification method based on phase sensing and multimodal fusion of the present invention.

[0026] Figure 2 This is a flowchart illustrating the process of determining the disturbance phase in the power quality disturbance location and identification method based on phase sensing and multimodal fusion according to the present invention.

[0027] Figure 3 This is a flowchart of the time-series feature vector extraction and bispectral image feature extraction in the power quality disturbance location and identification method based on phase sensing and multimodal fusion of the present invention.

[0028] Figure 4 This is a flowchart of the fused features obtained in the power quality disturbance location and identification method based on phase sensing and multimodal fusion of the present invention.

[0029] Figure 5 This is the bispectral image converted in the power quality disturbance localization and identification method based on phase sensing and multimodal fusion of the present invention.

[0030] Figure 6 This is a schematic diagram of the overall average training accuracy and loss curve of an embodiment of the power quality disturbance location and identification method based on phase sensing and multimodal fusion of the present invention.

[0031] Figure 7 This is a diagram of the three-phase overall average confusion matrix in an embodiment of the power quality disturbance location and identification method based on phase sensing and multimodal fusion of the present invention.

[0032] Figure 8 This is a schematic diagram showing the waveform representation of some PQD signals in an embodiment of the power quality disturbance location and identification method based on phase sensing and multimodal fusion of the present invention. Detailed Implementation

[0033] Please see Figure 1 The present invention provides a technical solution: a power quality disturbance location and identification method based on phase sensing and multimodal fusion, comprising the following steps: Acquire the raw voltage signal of the three-phase power system.

[0034] It should be noted that the raw voltage signal of the three-phase power system is acquired using a voltage sensor, with a sampling frequency set to 20kHz and a fixed signal length of 1024 sampling points. Normalization preprocessing is then performed on the acquired signal.

[0035] The sample embedding vectors of each phase are extracted based on the original voltage signal and compared with the normal sample vectors. The perturbed phases that have experienced power quality disturbances are determined by combining the adaptive threshold.

[0036] The normal sample vector acquisition process is as follows: Obtain the normalized, denoised, and fixed-length truncated three-phase normal voltage signal X: X={x A x B x c}, where x A x B x c These are the instantaneous voltage values ​​of phases A, B, and C at a certain moment; A three-layer one-dimensional convolutional network is used as the encoder f θ Single-phase independent feature extraction is performed on the three-phase normal voltage signal X, and the encoder f θ The convolution kernel sizes are 5, 3, and 3, and the number of channels are 64, 128, and 256, respectively, outputting a high-dimensional feature vector h. p Then, a two-layer fully connected network is used as the projection head g. φ , the feature vector h p Mapping to a low-dimensional embedding space and normalizing it, the output is a 256-dimensional normalized embedding vector e. p : ; Where p is the phase index; Comparison of learning objectives using normalized temperature-scaled cross-entropy loss: ; Among them (e) p , i ,e p,j ) represents a positive sample pair, sim(e p , i ,e p,j ) represents cosine similarity, ε is the temperature scaling factor, i represents the current sample index, j represents the positive sample index, and L p contrast Represents the comparative loss, when L p contrast When minimizing, two normal samples e of the same phase are... p,i and e p,j Aggregate to obtain aggregate group D normal ; For aggregation group D normal The mean of all normal samples is calculated to obtain the normal sample vector: ; Where N p It is aggregation group D normal The number of normal samples, c p This is the normal sample vector for phase p.

[0037] The process for acquiring normal sample vectors was clearly defined. By combining a three-layer one-dimensional convolutional encoder with a two-layer fully connected projector, efficient feature extraction and low-dimensional embedding of normal voltage signals were achieved. Then, through contrastive learning using normalized temperature-scaled cross-entropy loss, the embedding vectors of normal samples in the same phase were effectively aggregated. Finally, stable and representative normal sample vectors were obtained through averaging. This design provides a precise and unified comparison benchmark for subsequent perturbation phase localization, solving the problems of susceptibility to interference and poor consistency of normal sample benchmarks in traditional methods. This ensures the reliability of subsequent distance quantization and perturbation determination, laying a solid foundation for phase-level localization.

[0038] like Figure 2 As shown, the process of extracting the sample embedding vector of each phase based on the original voltage signal, comparing it with the normal sample vector, and determining the disturbed phase of power quality disturbance by combining it with an adaptive threshold is as follows: The original three-phase voltage signal X1 is processed by a one-dimensional convolution encoder f. θ Perform single-phase independent feature extraction, and then pass it through the projection head g. φ By performing mapping, the sample embedding vector e of each phase is obtained. p ; The extracted phase samples are embedded into vector e. p With the normal prototype vector c p Distance quantization is performed between them, and the Euclidean distance d is calculated. p : ; Generate a binary phase mask m by combining an adaptive threshold τ. p : ; Phase mask m p When the value is 1, it indicates that the phase disturbance is abnormal, that is, the disturbing phase that causes power quality disturbance, and the phase mask m p A value of 0 indicates that the disturbance in that phase is normal.

[0039] The process for determining the perturbation phase has been refined. First, the sample embedding vector of the signal to be detected is extracted using an encoder and projector head that are consistent with the normal sample vector, ensuring the consistency of feature representation. Then, the difference between the sample embedding vector and the normal sample vector is quantified using Euclidean distance, and the normal phase and perturbation phase are intuitively distinguished by combining binary phase mask. The entire process does not rely on perturbation phase labels, achieving unsupervised and accurate perturbation phase localization. This overcomes the core deficiency of traditional single-phase analysis, which isolates each phase signal and ignores inter-phase coupling information, leading to inaccurate localization. Moreover, the judgment logic is simple and efficient, providing a clear target for subsequent targeted feature extraction and recognition.

[0040] The adaptive threshold acquisition process is as follows: If the normal sample distance set Dp If the distribution approximately follows a Gaussian distribution, indicating that the three-phase power system is in steady-state operation, then the formula for calculating the adaptive threshold is: ; Where, μd p It is the distance mean, σd p is the standard deviation, and k is a constant for controlling sensitivity; If the normal sample distance set D p If the distribution is non-Gaussian, then the highest anomaly samples need to be excluded using statistical boundaries. Therefore, the formula for calculating the adaptive threshold is: ; Where a is the threshold coefficient, Percentile 95 It is the 95th percentile boundary; If the normal sample distance set D p If measurement noise and outliers exist, the robustness of the threshold calculation needs to be improved. Therefore, the formula for calculating the adaptive threshold is: ; Where b is the threshold coefficient, median(D) p ) represents the distance from the median, MAD(D) p ) represents the median absolute difference, and m is a constant used to adjust the MAD to be equivalent to the standard deviation under a Gaussian distribution.

[0041] To address the distribution characteristics of the normal sample distance set under different operating conditions, three adaptive threshold calculation methods were designed: a Gaussian distribution adaptation formula for steady-state conditions, a statistical boundary formula for non-Gaussian distributions, and a robust formula for noisy scenarios. This dynamic threshold adjustment mechanism ensures the sensitivity of disturbance detection during steady-state operation while effectively avoiding missed or false detections under unsteady or complex operating conditions containing measurement noise and outliers. It significantly improves the adaptability and robustness of threshold calculation, solving the problem of poor applicability of traditional fixed thresholds in complex power system operating conditions, and making disturbance phase determination more consistent with actual operating scenarios.

[0042] like Figures 3-5 As shown, temporal feature vector extraction and bispectral image feature extraction are performed on the perturbation phase, and alignment and co-fusion are performed to obtain fused features.

[0043] The process of acquiring bispectral images is as follows: Bispectral analysis was performed on the voltage signal of the disturbed phase to calculate its bispectral matrix. : ; Where f1 and f2 are two independent frequency components, and P(f1) and P(f2) are the Fourier transform coefficients of the voltage signal at frequencies f1 and f2, respectively. * (f 1+ f2) is the signal at frequency (f 1+ The Fourier transform coefficients P(f2) at f2) 1+ The complex conjugate of f2); The bispectral matrix is ​​visualized as a bispectral image in the form of a contour plot.

[0044] Bispectral analysis is used to calculate the bispectral matrix of voltage signals, capturing the phase coupling information between frequency components that traditional time-frequency analysis methods cannot retain. This is particularly suitable for revealing the essential characteristics of nonlinear disturbances. Furthermore, bispectral analysis performs exceptionally well in noisy signals and short-time data analysis, aligning well with the characteristics of power quality disturbances. Visualizing the bispectral matrix as a contour plot transforms the abstract phase coupling relationship into an intuitive spatial pattern, providing a carrier rich in core disturbance information for subsequent image feature extraction. This overcomes the limitations of traditional methods that rely on single signal features and are insufficient for representing complex disturbances, enhancing the model's ability to identify complex, short-time disturbances.

[0045] The process of extracting the time-series feature vector of the perturbation phase is as follows: Let the three-phase original voltage timing signal containing the disturbance phase be denoted as X∈R. 3×L Its length is fixed at L=1024 sampling points; Let X∈R 3×L The input is fed into a lightweight hierarchical structure constructed from three layers of depthwise separable convolutions, batch normalization, and activation functions to obtain temporal feature vectors: ; ; Where F l Let L be the output feature of layer l, l be the current feature extraction layer number, BN(.) be the batch normalization function, ReLU(.) be the activation function, and DSConv1D be the output feature of layer l. 5×1 (.) represents a one-dimensional depthwise separable convolution with a kernel size of 5×1, F l (:,i) represents the characteristics of all channels of the output of the l-th layer at time step i, f t This is the final temporal feature vector.

[0046] Specifically: Let X∈R 3×L The input is fed into a lightweight hierarchical structure constructed by the first layer of depthwise separable convolution, batch normalization, and activation functions, and the output is the basic temporal feature F1 of 64 channels; The F1 value is fed into a lightweight hierarchical structure constructed by the second layer of depthwise separable convolution, batch normalization, and activation function, expanding the number of channels to 96, and outputting feature F2. The F2 input is fed into a lightweight hierarchical structure constructed by the third layer of depthwise separable convolution, batch normalization, and activation functions to obtain the high-level temporal feature F3 with 128 channels.

[0047] A lightweight hierarchical structure is constructed using three layers of depthwise separable convolution, batch normalization, and ReLU activation function. When extracting time-series features of disturbed phase voltage signals, this structure ensures accurate capture of local time-series details while significantly reducing computational complexity through depthwise separable convolution, thus balancing feature extraction effectiveness and computational efficiency. The fixed signal length of 1024 sampling points ensures the consistency and comparability of time-series features from different samples, avoiding feature distortion caused by differences in signal length. The lightweight structure also meets the computational efficiency requirements of real-time power system monitoring, solving the problems of traditional one-dimensional convolutional models that either fail to capture sufficient local features or suffer from excessive computational complexity.

[0048] For bispectral image X BSP The input is fed into a convolutional embedding layer to extract local spatial patterns, and the output of the convolutional layer is a cross-scale feature F. conv : ; Among them, K d and K p These are depthwise convolution kernels and pointwise convolution kernels, respectively; σ(·) represents the ReLU nonlinear activation function; and BN is batch normalization. By capturing responses at different spatial scales through parallel average pooling and max pooling, and after linear mapping, the cross-scale features F conv They are unified into the same embedding space, forming multi-scale features F ms : ; Where Avg() represents the average pooling response, Max() represents the max pooling response, Concat(,) represents the concatenation operation, and W s and b s Represents the linear mapping parameters; Multiscale features F ms Input a Transformer structure, output a sequence of tokens: ; Where T is the token sequence, W e It is a linear embedding matrix; The token sequence is processed using a multi-head attention mechanism, and the attention output is A(Q,K,V): ; Where Q, K, and V are the query, key, and value vectors, respectively, and d k Let QK be the dimension of the key vector, Softmax(.) be the normalization function, and QK be the dimension of the key vector. T Attention score; The multi-head attention mechanism splits the Q, K, and V matrices into H independent heads, performs A(Q,K,V) computation in parallel, concatenates the outputs of all heads, and then projects them through a linear layer W. out Get the output h MHA : ; h MHA This is the output of the multi-head attention mechanism sublayer, which needs to be residually connected and normalized with the input T of this layer to obtain the processed feature sequence T': ; The processed feature sequence T' is then processed by a feedforward network and global pooling to obtain the final image features f. v : ; In this function, GlobalPool(.) is the global pooling function, LayerNorm(.) is the layer normalization function, and FFN(.) is the feedforward network operation.

[0049] Local spatial patterns of bispectral images are extracted through convolutional embedding layers. Multi-scale spatial responses are captured in parallel using average pooling and max pooling. Then, the global dependencies between features are mined using the Transformer's multi-head attention mechanism, achieving deep feature extraction layer by layer. This process preserves local details while integrating global structural information, resulting in image features rich in perturbation-based global representations that complement temporal local features, effectively compensating for the incomplete representation of perturbation information by single features. Simultaneously, the design of a lightweight feedforward network, residual layer normalization, and global pooling improves computational stability while ensuring feature extraction quality, avoiding image feature redundancy or distortion.

[0050] The process of aligning and co-fusing temporal feature vectors and bispectral image features to obtain fused features is as follows: The time series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C After weighted processing, the merged features are spliced ​​together to form a preliminary aligned fusion feature f. aligned ; Using linear layer pairs f alignedDimensionality reduction is performed to obtain the reduced feature h0: ; Among them, W aligned b0 and b0 are the weights and biases of the linear dimensionality reduction layer; Inputting h0 into a multilayer perceptron, which becomes the core of collaborative fusion: ; ; Where W1, b1, W2, and b2 are the weights and biases of the multilayer perceptron, ReLU(.) is the activation function, DroPout() is the regularization function, h1 is the output of the intermediate hidden layer of the multilayer perceptron, and h2 is the final output of the multilayer perceptron. Finally, a residual connection is introduced to add the dimensionality-reduced feature h0 to the output h2 of the MLP, and then the fused output is obtained through layer normalization: ; Where LayerNorm() is the layer normalization function, f fused It is the final fusion feature.

[0051] First, weighted processing is used to achieve initial alignment of temporal features, image features, and sample embedding vectors. Then, linear dimensionality reduction is applied to reduce feature dimensions and computational burden. The MLP structure enables deep collaborative fusion of heterogeneous features. Finally, residual connections are used to preserve the original feature information, avoiding the loss of effective information during the fusion process. This multi-stage, multi-level fusion mechanism fully leverages the complementary advantages of different modal features, significantly improving the completeness and discriminative power of the fused features. It solves the problems of information misalignment and weight imbalance leading to poor representation results in traditional multimodal fusion, providing high-quality feature support for subsequent accurate identification of perturbation types.

[0052] The time-series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C After weighted processing, the merged features are spliced ​​together to form a preliminary aligned fusion feature f. aligned The process is as follows: The time series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C Projected onto the common dimension D through a shared linear layer u And an importance weight for each feature is generated through a lightweight gating network: ; Where f i z represents the i-th input feature vector. i It is the projected feature, W p (i) ∈R Du*Di and b p (i) It is the projection matrix and bias of the feature type, W g and b g These are shared gated network parameters, σ(·) is the Sigmoid function, which compresses the weights to the (0,1) interval, and α i It is f i Adaptive importance weights; The projected features are weighted according to their importance and then concatenated to form a preliminarily aligned fused feature f. aligned : ; Where α t α v、 α A、 α B and α C These are the feature vectors z after projection through the gating network. t z v、 z A、 z B and z C The adaptive importance weights are given, and [;] represents the vector concatenation operation.

[0053] The specific implementation method of feature weighted concatenation was clarified. By using a shared linear layer, different types of features, such as temporal features, image features, and sample embedding vectors, are uniformly projected onto a common dimension, thus removing dimensionality barriers for feature alignment and subsequent fusion. A lightweight gated network dynamically generates adaptive importance weights for each feature through the sigmoid function, automatically distinguishing the contribution of different features based on the type of perturbation and signal characteristics, avoiding the subjectivity and limitations of manually setting weights. This adaptive weighted concatenation method allows the initial fused features to focus on information more critical to perturbation identification, effectively improving the quality of the initial fused features and providing more targeted inputs for subsequent linear dimensionality reduction, deep fusion, and accurate identification, further optimizing the effect of multimodal fusion.

[0054] Identification of power quality disturbance types based on fused features. Figure 7 This is a schematic diagram of the identification and classification results.

[0055] The process of identifying power quality disturbance types based on fused features is as follows: After the perturbation phase localization stage, a binary anomaly mask is generated, based on the mask label M.i,j Selective activation of the classification head: ; Where I is the indicator function, with an output of 0 for normal and 1 for abnormal (abnormal is activated by the classification head), d(,) is the distance function, and e i,j It is the embedding vector of the i-th sample and the j-th phase, c j It is the prototype center of the j-th phase, τ j It is the threshold for determining the j-th phase; The final category label C is selected through a gating mechanism. j : ; Among them, W j It is the weight matrix of the j-th class head, b j C1 is the corresponding bias vector, argmax() is the aggregation function that yields the predicted class label, and C1 is the normal class label.

[0056] Based on the anomaly mask generated by perturbation phase localization, the classification head of the corresponding perturbation phase is selectively activated, performing fine-grained type identification only on the perturbation phase. This skips redundant classification calculations for normal phases, significantly reducing the model's average inference time and meeting the engineering requirements of real-time power system monitoring. The gating mechanism, through aggregation functions combined with the weights and biases of the classification heads, accurately outputs perturbation type labels, ensuring the accuracy of the classification results. This collaborative design of localization and selective classification achieves efficient integration of perturbation phase localization and type identification, solving the problems of poor generalization ability and low computational efficiency of traditional methods for perturbation type identification in three-phase systems. Furthermore, the explicit setting of normal category labels makes the identification results clearer and more practical.

[0057] This embodiment adopts a power quality disturbance localization model based on phase sensing and multimodal fusion. The network consists of a phase sensing anomaly detection module and a dual-branch heterogeneous feature fusion network.

[0058] Training objective: Model training is a joint optimization process, involving two main loss functions: Contrastive learning loss: applied to the embedding space of the phase sensing anomaly detection module, used to narrow the distance between normal samples and their phase prototypes, ensuring the stability and representativeness of the prototype center.

[0059] Weighted cross-entropy loss: applied to the classification output of the dual-branch heterogeneous feature fusion module, used to accurately identify 25 types of perturbations, and uses a weighted approach to handle class imbalance.

[0060] The total loss function is a weighted sum of the two and guides the updating of model parameters.

[0061] Specifically, in this embodiment, the collected raw three-phase voltage time-series data is first used as the raw dataset. Data augmentation operations are then performed on the training data to increase data diversity and improve the model's generalization ability. To comprehensively verify the model's performance and robustness, the experiment covers 25 types of power quality disturbances. This type set includes 10 single disturbances (C1 normal, C2 sag, C3 sag, C4 spike, C5 flicker, C6 gap, C7 interruption, C8 harmonic, C9 impulse transient, C10 oscillatory transient); and 15 complex composite disturbances (C11 sag + harmonic, C12 flicker + harmonic, C13 gap + harmonic, C14 sag + harmonic, C15 interruption + flicker, C16 sag + flicker, C17 sag + flicker, C18 oscillatory transient + flicker, C19 flicker + interruption + harmonic, C20 gap + sag + harmonic, C21 oscillatory transient + sag + harmonic, C22 impulse transient + sag + flicker, C23 spike + sag + harmonic, C24 flicker + oscillatory transient + spike, C25 oscillatory transient + sag + harmonic + gap). By covering these 25 types (corresponding to labels C1-C25), especially the identification of complex perturbations, the model's ability to make judgments and its engineering applicability in complex power system environments are ensured.

[0062] Dataset partitioning: The original data is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for parameter learning of the network model; the validation set is used to evaluate the model's performance during training and to tune hyperparameters; and the test set is used to evaluate the model's final performance.

[0063] Data augmentation: Performing data augmentation operations on the training set to increase data diversity and effectively improve the model's generalization ability.

[0064] Hyperparameter settings and initialization: All learnable parameters of the model are initialized. Key hyperparameter settings during training are as follows: AdamW optimizer is used, with an initial learning rate of 0.001 and a weight decay coefficient of 0.001; batch size is 32; and the total number of training epochs is 100. A cosine annealing restart strategy is used for learning rate scheduling to dynamically adjust the learning rate and improve model convergence performance.

[0065] In each iteration, the model performs precise forward propagation calculations, where the signal first passes through a single-phase encoder and a projector to generate a phase embedding vector e. p The system identifies the disturbed phase and sends the abnormal phase to the next stage for fine-grained classification. The three-phase voltage signals and bispectral image data from the data-augmented training set are input into a dual-branch encoder. Features are extracted from the signals and images using the dual-branch encoder, and these features are then compared with e... p Together, alignment, weighting, and residual fusion are performed in the heterogeneous feature alignment and collaborative fusion module to obtain the final fused feature f.fused Subsequently, the probability distributions of 25 perturbation types were calculated using a three-phase independent fully connected classification layer and a softmax output layer. This was then applied based on the network model's output and the corresponding ground truth labels.

[0066] Gradients are calculated using the backpropagation algorithm. Following the chain rule, backpropagation calculates gradients layer by layer, starting from the output layer and working backwards to the input layer. During gradient calculation, the gradient of each layer's parameters is calculated, multiplied by the learning rate, and used as the new parameters for the next iteration. During training, the network model is periodically evaluated using a validation set. Data from the validation set is input into the trained model (with gradient calculation disabled), and performance metrics such as loss and accuracy are calculated on the validation set to assess model convergence. The best-performing model weights on the validation set are saved to prevent overfitting.

[0067] Once the model converges, the trained network model is applied to actual power quality monitoring equipment to identify and classify the real-time collected three-phase voltage data, thus promptly detecting power quality problems. The preferred embodiment of this invention achieves a final experimental accuracy of 99.82% across 25 disturbance types. (See training curves). Figure 6 For details, please refer to the following: Figure 7 The confusion matrix demonstrates that the proposed method has practical engineering value for deployment on edge computing platforms for online monitoring and disturbance identification.

[0068] Table 1 shows the results of comparing the trained network described in this embodiment with other popular neural networks on 20dB, 30dB, 40dB, and noise-free test datasets. As can be seen from Table 1, the power quality disturbance localization and identification method based on phase sensing and multimodal fusion of this invention has higher accuracy and stronger noise resistance compared to other popular methods under different noise environments. This method maintains a smaller number of parameters while achieving higher verification accuracy than popular models. Compared with other widely used models, it has better noise resistance and faster single-sample inference time.

[0069] Table 1: Average accuracy of this model and other models ; Since the model has been experimentally validated, the examples provided are the identification results of this model under the following 25 types of power quality disturbances, as shown in Table 2: Table 2: Identification of 25 Types of Power Quality Disturbances ; like Figure 8 The table shows the waveform representation of some PQD signals.

[0070] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0074] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0075] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A power quality disturbance location and identification method based on phase sensing and multimodal fusion, characterized in that, Includes the following steps: Acquire the raw voltage signal of a three-phase power system; The sample embedding vectors of each phase are extracted based on the original voltage signal and compared with the normal sample vectors. The perturbation phase that has caused power quality disturbance is determined by combining the adaptive threshold. Temporal feature vector extraction and bispectral image feature extraction are performed on the perturbation phase, followed by alignment and co-fusion to obtain fused features. The bispectral image feature extraction process is as follows: The bispectral image X BSP The input is fed into a convolutional embedding layer to extract local spatial patterns, and the output of the convolutional layer is a cross-scale feature F. conv : ； Among them, K d and K p These are depthwise convolution kernels and pointwise convolution kernels, respectively. σ1(·) represents the ReLU nonlinear activation function, and BN(·) is the batch normalization function. By capturing responses at different spatial scales through parallel average pooling and max pooling, and after linear mapping, the cross-scale features F conv They are unified into the same embedding space, forming multi-scale features F ms : ； Where Avg() represents average pooling response, Max() represents max pooling response, Concat(,) represents concatenation operation, and W s and b s Represents the linear mapping parameters; Multiscale features F ms Perform flattening and linear projection to output the token sequence: ； Where T is the token sequence, W e It is a linear embedding matrix; The token sequence is input into the Transformer encoder, and the token sequence is processed using a multi-head attention mechanism. The attention output is A(Q,K,V). ； Where Q, K, and V are the query, key, and value vectors, respectively, and d k Let QK be the dimension of the key vector, Softmax(·) be the normalization function, and QK be the dimension of the key vector. T Attention score; The multi-head attention mechanism splits the Q, K, and V matrices into H independent heads, performs A(Q,K,V) computation in parallel, concatenates the outputs of all heads, and then projects them through a linear layer W. out Get the output h MHA : ； h MHA This is the output of the multi-head attention mechanism sublayer, which needs to be residually connected and normalized with the input T of this layer to obtain the processed feature sequence T': ； The processed feature sequence T' is then processed by a feedforward network and global pooling to obtain the final image features f. v : ； Where GlobalPool(·) is global pooling, LayerNorm(·) is layer normalization function, and FFN(·) is feedforward network operation; Identification of power quality disturbance types based on fused features.

2. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 1, characterized in that, The normal sample vector acquisition process is as follows: Obtain the normalized, denoised, and fixed-length truncated three-phase normal voltage signal X: X={x A x B x c }, where x A x B x c These are the instantaneous voltage values ​​of phases A, B, and C at a certain moment; A three-layer one-dimensional convolutional network is used as the encoder f θ Single-phase independent feature extraction is performed on the three-phase normal voltage signal X, and a high-dimensional feature vector h is output. p Then, a two-layer fully connected network is used as the projection head. , the feature vector h p Mapping to a low-dimensional embedding space and normalizing it, the output is a 256-dimensional normalized embedding vector e. p : ； Where p is the phase index; Comparison of learning objectives using normalized temperature-scaled cross-entropy loss: ； in Indicates a positive sample pair. Let L represent the cosine similarity, ε be the temperature scaling factor, i represent the current sample index, j represent the positive sample index, and L be the cosine similarity. p contrast Let L represent the contrast loss, N represent the total number of current samples, and M represent the total number of positive samples. p contrast When minimizing, two normal samples e of the same phase are... p,i and e p,j Aggregate to obtain aggregate group D normal ; For aggregation group D normal The mean of all normal samples is calculated to obtain the normal sample vector: ； Where N p It is aggregation group D normal The number of normal samples, c p Let m be the normal sample vector of phase p, and m be the normal sample index.

3. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 2, characterized in that, The process of extracting the sample embedding vector of each phase based on the original voltage signal, comparing it with the normal sample vector, and determining the disturbed phase of power quality disturbance by combining it with an adaptive threshold is as follows: The original three-phase voltage signal X1 is processed by a one-dimensional convolution encoder f. θ Perform single-phase independent feature extraction, and then pass it through the projection head g. φ Mapping is performed to obtain the sample embedding vectors for each phase. ; The extracted samples of each phase are embedded into a vector. Compared with the normal sample vector c p Distance quantization is performed between them, and the Euclidean distance d is calculated. p : ； Generate a binary phase mask m by combining an adaptive threshold τ. p : ； Phase mask m p When the value is 1, it indicates that the phase disturbance is abnormal, that is, the disturbing phase that causes power quality disturbance, and the phase mask m p A value of 0 indicates that the phase is normal and no disturbance has occurred.

4. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 3, characterized in that, The adaptive threshold acquisition process is as follows: If the normal sample distance set D p If the distribution of the three-phase power system approximately follows a Gaussian distribution, indicating that the three-phase power system is operating in steady state, then the formula for calculating the adaptive threshold is: ； Where, μD p It is the distance mean, σD p is the standard deviation, and k is a constant for controlling sensitivity; If the normal sample distance set D p If the distribution is non-Gaussian, then the highest anomaly samples need to be excluded using statistical boundaries. Therefore, the formula for calculating the adaptive threshold is: ； Where a is the threshold coefficient, Percentile 95 It is the 95th percentile boundary; If the normal sample distance set D p If measurement noise and outliers exist, the robustness of the threshold calculation needs to be improved. Therefore, the formula for calculating the adaptive threshold is: ； Where b is the threshold coefficient, median(D) p ) represents the distance from the median, MAD(D) p ) represents the median absolute difference, and m1 is a constant used to adjust the MAD to be equivalent to the standard deviation under a Gaussian distribution.

5. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 1, characterized in that, The process of acquiring bispectral images is as follows: Bispectral analysis was performed on the voltage signal of the disturbed phase to calculate its bispectral matrix s. 3,x (f1,f2): ； Where f1 and f2 are two independent frequency components, and P(f1) and P(f2) are the Fourier transform coefficients of the voltage signal at frequencies f1 and f2, respectively. * (f1+f2) is the complex conjugate of the Fourier transform coefficients P(f1+f2) of the signal at frequency (f1+f2); The bispectral matrix is ​​visualized as a bispectral image in the form of a contour plot.

6. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 1, characterized in that, The process of extracting the time-series feature vector of the perturbation phase is as follows: Let the three-phase original voltage timing signal containing the disturbance phase be denoted as X∈R. 3×L Its length is fixed at L=1024 sampling points; Let X∈R 3×L The input is fed into a lightweight hierarchical structure constructed from three layers of depthwise separable convolutions, batch normalization, and activation functions to obtain temporal feature vectors: ； ； Where F l Let L be the output feature of layer l, l be the current feature extraction layer number, BN(·) be the batch normalization function, ReLU(·) be the activation function, and DSConv1D be the output feature of layer l. 5×1 (·) represents a one-dimensional depthwise separable convolution with a kernel size of 5×1, F l (:,t1) represents the characteristics of all channels of the output of the l-th layer at time step t1, f t For the final time-series feature vector, C l This represents the number of channels in the l-th layer.

7. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 1, characterized in that, The process of aligning and co-fusing temporal feature vectors and bispectral image features to obtain fused features is as follows: The time series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C After weighted processing, the merged features are spliced ​​together to form a preliminary aligned fusion feature f. aligned ; Using linear layer pairs f aligned Dimensionality reduction is performed to obtain the reduced feature h0: ； Among them, W aligned b0 and b0 are the weights and biases of the linear dimensionality reduction layer; Inputting h0 into a multilayer perceptron, which becomes the core of collaborative fusion: ； ； Where W1, b1, W2, and b2 are the weights and biases of the multilayer perceptron, ReLU(·) is the activation function, Dropout() is the regularization function, h1 is the output of the intermediate hidden layer of the multilayer perceptron, and h2 is the final output of the multilayer perceptron. Finally, a residual connection is introduced to add the dimensionality-reduced feature h0 to the output h2 of the MLP, and then the fused output is obtained through layer normalization: ； Where LayerNorm() is the layer normalization function, f fused It is the final fusion feature.

8. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 7, characterized in that, The time-series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C After weighted processing, the merged features are spliced ​​together to form a preliminary aligned fusion feature f. aligned The process is as follows: The time series feature vector f t Image features f v The sample embedding vector e of each phase A e B and e C Projected onto common dimension D through a shared linear layer u And an importance weight for each feature is generated through a lightweight gating network: ； Among them B h B represents the h-th input feature vector, where h is t. h f t When h is v, B h f v When h is A, B h For e A When h is B, B h For e B When h is C, B h For e C , z h It is the projected feature, W p (h) ∈R Du*Dh and b p (h) These are the projection matrix and bias of the feature type, where D h W represents the original dimension of the h-th input feature vector. g and b g These are shared gated network parameters, σ²(·) is the Sigmoid function, which compresses the weights to the (0,1) interval, and α h It is z h Adaptive importance weights; The projected features are weighted according to their importance and then concatenated to form a preliminarily aligned fused feature f. aligned : ； Where α t α v、 α A、 α B and α C These are the feature vectors z after projection through the gating network. t z v、 z A、 z B and z C The adaptive importance weights are given, and [;] represents the vector concatenation operation.

9. The power quality disturbance location and identification method based on phase sensing and multimodal fusion according to claim 1, characterized in that, The process of identifying power quality disturbance types based on fused features is as follows: After the perturbation phase localization stage, a binary anomaly mask is generated, based on the mask label M. i,p Selective activation of the classification head: ； Where I is the indicator function, with an output of 0 for normal and 1 for abnormal (activating the classification head when an abnormality occurs), d(,) is the distance function, and e i,p It is the embedding vector of the i-th sample in the p-th phase, c p It is the normal sample vector of phase p. It is the threshold for determining the p-th phase; The final category label C is selected through a gating mechanism. p : ； Among them, W p It is the weight matrix of the classifier head corresponding to phase p, b p It is the corresponding bias vector, f fused The final fused feature is argmax(), which is the aggregation function that yields the predicted class label, and C1 is the normal class label.

Images