A power quality disturbance detection method of a fusion cross-distillation feature adversarial network

Abstract

This invention belongs to the field of power quality detection technology for smart distribution networks. It discloses a power quality disturbance detection method integrating a cross-distillation feature adversarial network (DCAF-WGAN). The method includes signal upscaling, a cross-distillation fusion method, and a DCAF-WGAN model. After upscaling the power signal to a two-dimensional image using a recursive graph and Markov transition field, a cross-distillation fusion method is introduced. This method utilizes a knowledge transfer strategy, a distillation loss function, and nonlinear attention weighting to extract complementary features and achieve modal enhancement. Finally, a DCAF-WGAN model is constructed, using a dilated convolution module, channel attention mechanism, and spatial attention mechanism integrated in the discriminator to perform disturbance feature recognition. This invention can effectively address complex power quality disturbance problems in power systems, ensuring the stable operation of the power grid and industrial equipment.

Description

Technical Field

[0001] This invention belongs to the field of power quality detection technology for smart distribution networks, specifically relating to a power quality disturbance detection method that integrates cross-distillation feature adversarial networks. Background Technology

[0002] With the increasing intelligence of new energy sources and power electronic equipment, the proportion of nonlinear loads in power systems is rising. The complex operating characteristics of power grids and significant power quality disturbances are impacting the stable operation of power grids and industrial equipment. Therefore, high-precision and robust power quality disturbance detection methods are crucial for power system safety and reducing economic losses.

[0003] Traditional power quality disturbance (PQD) detection methods rely on electrical quantity feature extraction and signal processing techniques. Common feature extraction methods include Fourier transform, empirical mode decomposition, wavelet transform, and S-transform. The obtained features are used as input to a classifier to identify disturbances. Common classifiers include support vector machines, decision trees, random forests, and Naive Bayes. In recent years, with the development of artificial intelligence technology, data mining methods have been widely used in PQD detection. Compared to traditional classifiers, neural network models can utilize multi-layer mapping structures to mine disturbance features and have good noise resistance.

[0004] Despite some progress in existing technologies, significant shortcomings remain in methods for extracting complex perturbation features and improving model stability. Traditional feature extraction and analysis methods are based on one-dimensional time-series windows, while PQD signals have nonlinear characteristics, and one-dimensional signals may miss time-frequency correlation information of complex perturbations. Some perturbation recognition models based on two-dimensional information, although improving feature dimensionality, are prone to problems such as network parameter redundancy, training collapse, and gradient instability. WGAN based on Wasserstein distance and its improved methods have achieved stable improvements in adversarial training through gradient penalty, but these models lack the ability to collaboratively model multi-channel features, making it difficult to accurately extract time-domain and frequency-domain details of perturbation signals.

[0005] Therefore, for the complex disturbance scenarios of smart distribution networks, it is particularly important to construct a power quality detection method with stable training process, high prediction accuracy and strong classification ability for the normal and safe operation of the power grid. Summary of the Invention

[0006] To address the above issues, this invention proposes a power quality disturbance detection method that integrates cross-distillation feature adversarial networks. This method extracts data features and utilizes the DCAF-WGAN model to achieve accurate classification and detection, maintaining stable performance under complex operating conditions.

[0007] The power quality disturbance detection method based on the fusion of inter-distillation characteristic adversarial networks described in this invention comprises the following steps: Step 1: Convert the one-dimensional power quality disturbance signal into a two-dimensional image using recursive graph (RP) and Markov transition field (MTF): RP generates a two-dimensional image by reconstructing the local trajectory of the signal in phase space; MTF generates a two-dimensional image by discretizing the signal into finite states and calculating the state transition probabilities. Step 2: Construct a mutual distillation fusion module. By introducing a feature distillation loss function between the RP branch and the MTF branch, knowledge transfer and peer learning are carried out. The contribution of the two types of images is adaptively adjusted by using a nonlinear attention weighting mechanism to generate an information-enhanced feature map after mutual distillation of the temporal recursive features of the RP image and the state transition probability features of the MTF image. Step 3: Introduce Wasserstein distance and gradient penalty terms into GAN, i.e. WGAN-GP. Based on this, introduce dilated convolution and multi-level attention mechanisms (channel attention module and spatial attention module) into the discriminator of GAN to form the DCAF-WGAN model. Step 4: Input the enhanced feature map obtained from the inter-distillation fusion in Step 2 into the DCAF-WGAN model to classify various types of power quality disturbances.

[0008] Furthermore, in step 1, Recursive Graph (RP) is a two-dimensional encoding method based on phase plane reconstruction. RP is used for two-dimensional image transformation, including: One-dimensional time series in the phase plane The matrix is ​​obtained by reconstructing its phase space. The formula is as follows: , In the formula, d represents the embedding dimension; Indicates the delay time; Represents the trajectory related to the embedding dimension and latency; This indicates that the time series is at time 1. The value; n represents the total number of samples, and m represents the time index to the m-th step in the phase space; This represents the last data sample in the sequence; This represents the m-th reconstructed vector; The similarity between vectors is determined based on Euclidean distance, and the recurrence value matrix is ​​calculated using the following formula: , In the formula, It is the Euclidean paradigm. The set distance threshold, Represents a binary function. It is the independent variable in the function, and in this formula it equals .

[0009] Furthermore, in step 1, Markov Transition Field (MTF) is a time series dimensionality upscaling method. Using MTF for two-dimensional image transformation includes: Assuming time series , where s n This represents the nth sample in an MTF time series. The value range is divided into Q parts, such that each value in the time series is mapped to a corresponding... , representing the i-th state interval obtained after discretizing the value range of the time-series signal, where i=1,2,…Q; , calculated based on the transition of the one-dimensional power quality disturbance signal between different states. The transition probabilities between these probabilities form the Markov transition matrix W, as shown in the following formula: , In the formula, Representing data The elements in are The probability of following, i.e. s i This represents the i-th sampling point of the time series signal; Rearrange the probabilities in chronological order, expand the Markov transition matrix W, and generate the Markov transition field M, as shown in the following formula: , In the formula, Indicates the interval from Transferred to The probability of that, i.e. .

[0010] Further, in step 2, the RP and MTF images are input into convolutional neural networks with different parameters but the same structure. Each layer of the convolutional neural network includes convolution operations, batch normalization, and activation functions, and deep semantic features are extracted by downsampling layer by layer; let the RP image be... MTF image is Their characteristics are expressed as follows: , , In the formula, and Let represent the convolutional feature extraction mapping functions of RP and MTF, respectively; During training, a symmetric feature distillation loss function is used for interactive learning between the RP and MTF branches; feature spacing is extracted using KL divergence, and the stopping gradient is used to guide one branch as a teacher model to fit the target features to the other branch. The loss function is defined as follows: , In the formula, For similarity measurement functions, , The contribution of the two modes can be changed by setting different values ​​to achieve the distillation intensity balance coefficient in different directions. The features after distillation are further modeled using an attention-weighted fusion method, specifically as follows: , In the formula, Represents the sigmoid activation function. Represents element-wise multiplication. , These represent the channel weights learned by RP and MTF in the fusion method, respectively; To enhance the adaptability of the fusion representation, the attention weighting mechanism is extended to a non-linear form: , In nonlinear expressions, and Indicates the predicted spatial channel weights. To prevent division by zero of extremely small constants; Finally, the output is normalized, and the result is mapped to the grayscale space so that the pixel values ​​of the output image fall within the grayscale range. Within the specified range, enhance image contrast; expressed as follows: .

[0011] Furthermore, in step 3, the objective function of WGAN-GP is: , In the formula, This is the negative of the generator's objective function, used to make the discriminator output as large as possible to fool the discriminator; This represents the discriminator's score for the real sample; a higher value indicates that the sample is more real. This represents the score given by the discriminator to the generated sample; a lower value indicates a better sample recognition performance. This is a gradient penalty term; For real data With generated data Random interpolation points between points are used to calculate the gradient penalty term; The L2 norm of the interpolation point. The strength of the gradient penalty.

[0012] Furthermore, in step 3, the DCAF-WGAN recognition model includes a discriminator and a generator; The discriminator takes the fused and enhanced feature map as input. First, it performs preliminary downsampling feature extraction through a regular convolutional layer, and then further extracts deep features through another regular convolutional layer. Subsequently, it uses a dilated convolutional layer to capture a wider range of perturbation information. Then, it connects the channel attention module and the spatial attention module to adaptively weight and fuse the importance of different channels and the salience of spatial regions. Finally, it outputs the sample authenticity score and the classification probability of the power quality perturbation type through a flattening layer and a fully connected layer. The generator receives a random noise vector, reshapes it through a fully connected layer, and then gradually upsamples it through a multi-layer deconvolution structure to generate a fake sample with the same size as the real fused feature map. The discriminator and generator are trained adversarially to construct the DCAF-WGAN recognition model.

[0013] During the training phase, a DCAF-WGAN recognition model is constructed by using a two-dimensional image generated by the fusion of cross-distillation features as input and alternating adversarial training between the generator and the discriminator. During the application phase, the perturbation signal is input into the model to obtain the power quality perturbation classification results of each perturbation signal.

[0014] The beneficial effects of this invention are as follows: This invention innovatively introduces a mutual distillation feature fusion mechanism, enabling the recursive structural features of the RP (Recursive Progressive Map) and the state transition probability features of the MTF (Medium-Terminal Array) image to complement each other through bidirectional knowledge transfer and nonlinear attention weighting, thus reducing the feature loss of a single encoding method. It also constructs a DCAF (Diverterless Convolutional Algorithm) structure, cascading dilated convolution with a channel-space dual attention mechanism, expanding the receptive field of the convolutional layer and enhancing its sensitivity to weak perturbations and noise. Compared with existing methods based on simple image overlay or traditional WGAN, this invention achieves richer time-frequency joint representation and stronger feature discrimination capability. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method. Figure 2 This is a diagram illustrating the dimensionality upscaling process of the PQD signal. Figure 3 The diagram shows the inter-distillation fusion method based on RP-MTF. Figure 4 This is a diagram of the DCAF-WGAN structure. Figure 5 The waveform diagram of the PQD signal; Figure 6 Figure 1 shows the ablation experiment of the model. Figure 7 For the simulation confusion matrix diagram; Figure 8 To compare the training accuracy of the models; Figure 9 The confusion matrix diagram for the experiment. Detailed Implementation

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0017] like Figure 1 As shown, the power quality disturbance detection method of the fused cross-distillation feature adversarial network described in this invention includes the following steps: Step 1: Convert the one-dimensional power quality disturbance signal into a two-dimensional image using recursive graph (RP) and Markov transition field (MTF): RP generates a two-dimensional image by reconstructing the local trajectory of the signal in phase space; MTF generates a two-dimensional image by discretizing the signal into finite states and calculating the state transition probabilities. Step 2: Construct a mutual distillation fusion module. By introducing a feature distillation loss function between the RP branch and the MTF branch, knowledge transfer and peer learning are carried out. The contribution of the two types of images is adaptively adjusted by using a nonlinear attention weighting mechanism to generate an information-enhanced feature map after mutual distillation of the temporal recursive features of the RP image and the state transition probability features of the MTF image. Step 3: Introduce Wasserstein distance and gradient penalty terms into GAN, i.e. WGAN-GP. Based on this, introduce dilated convolution and multi-level attention mechanism into the discriminator of GAN to form DCAF-WGAN model. Step 4: Input the enhanced feature map obtained from the inter-distillation fusion in Step 2 into the DCAF-WGAN model to classify various types of power quality disturbances.

[0018] like Figure 2 As shown, the conversion process from a one-dimensional time-series signal to two-dimensional time-frequency images of RP and MTF is intuitively demonstrated.

[0019] Step 1 involves using RP to perform two-dimensional image conversion, including: One-dimensional time series in the phase plane The matrix is ​​obtained by reconstructing its phase space. The formula is as follows: , In the formula, d represents the embedding dimension; Indicates the delay time; Represents the trajectory related to the embedding dimension and latency; This indicates that the time series is at time 1. The value; n represents the total number of samples, and m represents the time index to the m-th step in the phase space; This represents the last data sample in the sequence; This represents the m-th reconstructed vector; The similarity between vectors is determined based on Euclidean distance, and the recurrence value matrix is ​​calculated using the following formula: , In the formula, It is the Euclidean paradigm. The set distance threshold, Represents a binary function. It is the independent variable in the function, and in this formula it equals .

[0020] Two-dimensional image conversion using MTF includes: Assuming time series , where s n This represents the nth sample in an MTF time series. The value range is divided into Q parts, such that each value in the time series is mapped to a corresponding... , representing the i-th state interval obtained after discretizing the value range of the time-series signal, where i=1,2,…Q; , calculated based on the transition of the one-dimensional power quality disturbance signal between different states. The transition probabilities between these probabilities form the Markov transition matrix W, as shown in the following formula: , In the formula, Representing data The elements in are The probability of following, i.e. s i This represents the i-th sampling point of the time series signal; Rearrange the probabilities in chronological order, expand the Markov transition matrix W, and generate the Markov transition field M, as shown in the following formula: , In the formula, Indicates the interval from Transferred to The probability of that, i.e. .

[0021] like Figure 3 As shown, in step 2, the RP and MTF images are respectively input into convolutional neural networks with different parameters but the same structure to extract deep semantic features; let the RP image be... MTF image is Their characteristics are expressed as follows: , , In the formula, and Let represent the convolutional feature extraction mapping functions of RP and MTF, respectively. During training, a symmetric feature distillation loss function is used for interactive learning between the RP and MTF branches. Feature spacing is extracted using KL divergence, and the stopping gradient is used to guide one branch as the teacher model to fit the target features to the other branch. The loss function is defined as: , In the formula, For similarity measurement functions, , The distillation intensity balance coefficients in different directions are used to vary the contribution of the two modes by setting different values. The features after distillation are further modeled using an attention-weighted fusion method, specifically represented as follows: , In the formula, Represents the sigmoid activation function. Represents element-wise multiplication. , These represent the channel weights learned by RP and MTF in the fusion method, respectively; Extending the attention-weighted mechanism to a non-linear form: , In the formula, and Indicates the predicted spatial channel weights. To prevent division by zero by extremely small constants, the output is finally normalized and mapped to grayscale space so that the pixel values ​​of the image output fall within the [0,1] interval, thereby improving image contrast. This is expressed as follows: .

[0022] like Figure 4 As shown, in step 3, WGAN introduces Wasserstein distance instead of JS divergence. Wasserstein distance represents the minimum cost of moving from one distribution to another, i.e., the minimum moving average distance. This distance allows WGAN to generate more stable and smooth gradient signals during training. Furthermore, to maintain Lipschitz continuity and further avoid gradient vanishing and mode collapse problems, WGAN-GP replaces the weight clipping in the original WGAN with a gradient penalty term (GP). Gradient penalty refers to constraining the L2 norm of the original input gradient to around 1, preventing abnormal gradient updates. The objective function of WGAN-GP is: , In the formula, This is the negative of the generator's objective function, used to make the discriminator output as large as possible to fool the discriminator. This represents the discriminator's score for the real sample; a higher value indicates that the sample is more real. This represents the score given by the discriminator to the generated sample; a lower value indicates a better sample recognition performance. This is a gradient penalty term; For real data With generated data Random interpolation points between points are used to calculate the gradient penalty term; The L2 norm of the interpolation point. The strength of the gradient penalty.

[0023] The original WGAN-GP discriminator convolutional layer uses a downsampling convolutional structure. Each feature extraction layer consists of multiple convolutional blocks, mostly using Conv2d two-dimensional convolution operations, with convolutional kernels of... Dilated convolution introduces a dilation rate. Expand the convolution to: , in, Indicates the weights of the convolution kernel. Represents the spatial coordinates of the feature map. The input feature map is used. Therefore, dilated convolution expands the receptive field to: , Here, RF represents the receptive field function. Based on the above principle, keeping the number of channels and non-activation functions unchanged, only the middle and later convolutional layers of the discriminator are replaced with dilated convolutions, preserving the perturbation information. The feature map spatial weighted sum after the dilated convolution operation is expressed as: , CAF assigns different weights to each feature map and establishes channel relationships using global pooling and MLP (Multi-Level Processing). The MLP typically consists of two fully connected layers. The feature map after dilated convolution... After performing average pooling and max pooling operations, the result is: , , Using MLP fusion as follows: , in, and For the two-parameter matrix of the MLP, the total output is: , SAF performs pooling operations on top of CAF: , , Then, by splicing them together, we get: , enter Convolutional kernels are fused using MLP: , Output feature map: ; After adding CAF, the feature maps are reweighted in the channels, as shown in the following formula: , in, Indicates the first Channel importance weights. SAF, on the other hand, performs attention modulation in space, as shown in the following formula: ; In the hidden layers, complex features are constructed using ReLU or Leaky ReLU. The generator's loss function is: , in, The generated samples represent the generator. This represents the rating given by the evaluator to the generated sample.

[0024] During the training phase, a DCAF-WGAN recognition model is constructed by using a two-dimensional image generated by the fusion of cross-distillation features as input and alternating adversarial training between the generator and the discriminator. During the application phase, the perturbation signal is input into the model to obtain the power quality perturbation classification results of each perturbation signal.

[0025] Simulation tests were conducted under 14 different labels. The confusion matrices from the 20th, 43rd, 82nd, and 100th iterations were selected as comparison results. With increasing iteration count, the model's classification accuracy continuously improved, and the model became increasingly stable in both time and frequency domain feature discrimination. At 100 iterations, the confusion matrix was essentially stable, exhibiting a diagonal distribution, and the accuracy for each label category reached over 96.8%. Simulation results are as follows... Figure 7 As shown.

[0026] In one example, the model in step 4 is built using the Python and TensorFlow frameworks. The specific hardware configuration and runtime environment versions are shown in the table below. The training batch size is 64, the total number of training iterations is 100, the optimizer is Adam, and the gradient penalty coefficient is set to 10.

[0027] In this example, the specific parameter settings are shown in Table 1.

[0028] Table 1 Hardware and software configuration

[0029] The discriminator's convolutional layers are divided into two parts. The kernel size *k* of the first two convolutional operations remains 5x5, but the number of channels is increased to 64 and 128 respectively, preserving the ability to process image details. The LeakyReLU activation function avoids the gradient vanishing problem while enhancing the model's feature resolution at low signal-to-noise ratios. A dilated convolutional layer is then introduced, with a kernel size of 3x3 and a dilation rate *d* set to 2, expanding the receptive field without increasing parameters compared to ordinary convolutional operations. Furthermore, the discriminator introduces CAF and SAF to achieve local and global attention fusion of perturbation information features. Finally, after flattening by a fully connected layer, Softmax activates the samples and outputs the results.

[0030] In this example, the specific parameter settings are shown in Table 2.

[0031] Table 2 Model Parameters

[0032] To verify the effectiveness of the proposed method, simulation and experimental studies were conducted. Figure 5 The waveform diagram of a typical PQD signal shows the time-domain characteristics of various disturbance signals. Figure 6 The figure shows the ablation experiment curves. As can be seen, the complete DCAF-WGAN model (model0) outperforms other ablation models in both the rate of loss reduction and the magnitude of accuracy improvement. The confusion matrices at iterations 20, 43, 82, and 100 are used for comparison. With increasing iteration count, the model's classification accuracy continuously improves, and the model becomes increasingly stable in both time and frequency domain feature discrimination. At 100 iterations, the confusion matrix is ​​essentially stable, exhibiting a diagonal distribution, and the accuracy for each label category reaches over 96.8%. Figure 7 As shown, Figure 7 (a) is the confusion matrix diagram for the 20th iteration. Figure 7 (b) is the confusion matrix diagram for the 43rd iteration. Figure 7 (c) is the confusion matrix diagram for the 82nd iteration. Figure 7 (d) is the confusion matrix diagram for the 100th iteration. Figure 8 A graph comparing the training accuracy of different models. Figure 8 (a) represents the training set accuracy iteration curves of five models: Transformer, MobilenetV2, Resnet18, DCGAN, and DCAF-WGAN. Figure 8(b) represents the validation set accuracy iteration curves of the five models. As can be seen from the figure, the training set and validation set accuracy curves of the DCAF-WGAN model of this invention converge faster, have higher final accuracy, and exhibit less fluctuation. Figure 9 To generate a confusion matrix for the experimental data, the method was validated on an electromagnetic confusion simulation system platform. The specific structure included: a main control box, a temperature control box, a liquid storage tank, a disturbance electronic meter, and a substation topology. The results show that the proposed method maintains high classification accuracy and low misclassification rate under real-world operating conditions.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A method for detecting power quality disturbances by integrating cross-distillation feature adversarial networks, characterized in that, Includes the following steps: Step 1: Convert the one-dimensional power quality disturbance signal into a two-dimensional image using recursive graph (RP) and Markov transition field (MTF): RP generates a two-dimensional image by reconstructing the local trajectory of the signal in phase space; MTF generates a two-dimensional image by discretizing the signal into finite states and calculating the state transition probabilities. Step 2: Construct a mutual distillation fusion module. By introducing a feature distillation loss function between the RP branch and the MTF branch, knowledge transfer and peer learning are carried out. The contribution of the two types of images is adaptively adjusted by using a nonlinear attention weighting mechanism to generate an information-enhanced feature map after mutual distillation of the temporal recursive features of the RP image and the state transition probability features of the MTF image. Step 3: Introduce Wasserstein distance and gradient penalty terms into GAN, i.e. WGAN-GP. Based on this, introduce dilated convolution and multi-level attention mechanism into the discriminator of GAN to form DCAF-WGAN model. Step 4: Input the enhanced feature map obtained from the inter-distillation fusion in Step 2 into the DCAF-WGAN model to classify various types of power quality disturbances.

2. The power quality disturbance detection method based on a fusion of inter-distillation characteristic adversarial networks according to claim 1, characterized in that, Step 1 involves using RP to perform two-dimensional image conversion, including: One-dimensional time series in the phase plane The matrix is ​​obtained by reconstructing its phase space. The formula is as follows: ， In the formula, d represents the embedding dimension; Indicates the delay time; Represents the trajectory related to the embedding dimension and latency; This indicates that the time series is at time 1. The value; n represents the total number of samples, and m represents the time index to the m-th step in the phase space; This represents the last data sample in the sequence; This represents the m-th reconstructed vector; The similarity between vectors is determined based on Euclidean distance, and the recurrence value matrix is ​​calculated using the following formula: ， In the formula, It is the Euclidean paradigm. The set distance threshold, Represents a binary function. It is the independent variable in the function, and in this formula it equals .

3. The power quality disturbance detection method based on a fusion inter-distillation characteristic adversarial network according to claim 2, characterized in that, Step 1 involves using MTF to perform two-dimensional image conversion, including: Assuming time series , where s n This represents the nth sample in an MTF time series. The value range is divided into Q parts, such that each value in the time series is mapped to a corresponding... , representing the i-th state interval obtained after discretizing the value range of the time-series signal, where i=1,2,…Q; , calculated based on the transition of the one-dimensional power quality disturbance signal between different states. The transition probabilities between these probabilities form the Markov transition matrix W, as shown in the following formula: ， In the formula, Representing data The elements in are The probability of following, i.e. s i This represents the i-th sampling point of the time series signal; Rearrange the probabilities in chronological order, expand the Markov transition matrix W, and generate the Markov transition field M, as shown in the following formula: ， In the formula, Indicates the interval from Transferred to The probability, i.e. .

4. The power quality disturbance detection method based on a fusion of inter-distillation characteristic adversarial networks according to claim 1, characterized in that, In step 2, the RP and MTF images are input into convolutional neural networks with different parameters but the same structure. Each layer of the convolutional neural network includes a convolution operation, batch normalization, and an activation function, and deep semantic features are extracted by downsampling layer by layer. Let the RP image be... MTF image is , Their characteristics are expressed as follows: ， ， In the formula, and Let represent the convolutional feature extraction mapping functions of RP and MTF, respectively; During training, a symmetric feature distillation loss function is used to perform interactive learning between the RP and MTF branches; Feature spacing is extracted using KL divergence, and stopping gradients are used to guide one branch as a teacher model to fit the target features to the other branch. The loss function is defined as follows: ， In the formula, For similarity measurement functions, , The equilibrium coefficients for distillation intensity in different directions; The features after distillation are further modeled using an attention-weighted fusion method, specifically as follows: ， In the formula, Represents the sigmoid activation function. Represents element-wise multiplication. , These represent the channel weights learned by RP and MTF in the fusion method, respectively; Extending the attention-weighted mechanism to a non-linear form: ， In the formula, and Indicates the predicted spatial channel weights. To prevent division by zero of extremely small constants; Finally, the output is normalized, and the result is mapped to the grayscale space so that the pixel values ​​of the output image fall within the grayscale range. Within the interval, it can be expressed as follows: 。 5. The power quality disturbance detection method based on a fusion inter-distillation feature adversarial network according to claim 1, characterized in that, In step 3, the objective function of WGAN-GP is: ， In the formula, It is the negative of the generator's objective function; This represents the discriminator's score for the real sample; This represents the score given by the discriminator to the generated sample; This is a gradient penalty term; For real data With generated data Random interpolation points between; The L2 norm of the interpolation point. The strength of the gradient penalty.

6. The power quality disturbance detection method based on a fusion of inter-distillation characteristic adversarial networks according to claim 5, characterized in that, In step 3, the DCAF-WGAN recognition model includes a discriminator and a generator; The discriminator takes the fused and enhanced feature map as input. First, it performs preliminary downsampling feature extraction through a regular convolutional layer, and then further extracts deep features through another regular convolutional layer. Subsequently, it uses a dilated convolutional layer to capture a wider range of perturbation information. Then, it connects the channel attention module and the spatial attention module to adaptively weight and fuse the importance of different channels and the salience of spatial regions. Finally, it outputs the sample authenticity score and the classification probability of the power quality perturbation type through a flattening layer and a fully connected layer. The generator receives a random noise vector, reshapes it through a fully connected layer, and then gradually upsamples it through a multi-layer deconvolution structure to generate a fake sample with the same size as the real fused feature map. The discriminator and generator are trained adversarially to construct the DCAF-WGAN recognition model.

Images