Power quality disturbance detection method and device based on SWD and joint convolution

Abstract

The invention provides a power quality disturbance detection method based on SWD and joint convolution. The method comprises the following steps: step 1, collecting electric energy quality disturbance data and marking a label, and carrying out overlapping sampling on each category in the electric energy quality disturbance data to generate an electric energy quality disturbance signal sample; 2, decomposing the electric energy quality disturbance signal sample through an SWD algorithm, and reconstructing decomposed sample components; 3, transforming the samples reconstructed in the step 2 to obtain a time-frequency graph; 4, constructing a joint neural convolutional network, and training the joint convolutional neural network by using the time-frequency graph to obtain a power quality disturbance recognition model; and 5, inputting the time-frequency diagram corresponding to the to-be-detected power quality disturbance data into the power quality disturbance identification model to obtain a classification result of power quality disturbance identification. According to the invention, disturbance signal processing and disturbance type identification can be realized, and the accuracy of power quality disturbance detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power quality disturbance detection, and in particular to a power quality disturbance detection method and device based on SWD and joint convolution. Background Art

[0002] Under the background of "dual carbon", the penetration rate of photovoltaic power sources in distribution networks is increasing, and the fault characteristics of distribution network lines have undergone fundamental changes. The fault characteristics of photovoltaic power sources mainly depend on their physical structure, grid connection mode and control strategy. The fault current shows strong nonlinearity, phase control and high harmonic content, which are significantly different from traditional synchronous power sources. The complex fault characteristics caused by a high proportion of photovoltaic access will bring serious challenges to the power quality in the system. In order to meet users' high requirements for power quality, the management of power quality is particularly important.

[0003] The classification and identification of power quality disturbances is the premise of power quality management. Early power quality disturbance identification technology mainly relied on manual identification of disturbance features, and identification was performed based on low frequency analysis recording (LOFAR) and envelope modulation (DEMON) spectra. Related work under the traditional technical framework revolves around feature extraction and classifier design, but this method relies on manual experience for judgment, has limited accuracy, and has low recognition efficiency when faced with a large amount of disturbance signal data. With the breakthrough progress made in the theory of deep learning algorithms, identification technology has ushered in new method innovations. Summary of the invention

[0004] In view of the problem that the existing power quality disturbance detection technology mainly relies on manual experience to identify power disturbances, and has low recognition efficiency and recognition accuracy, the present invention proposes a power quality disturbance detection method based on SWD and joint convolution, which utilizes the SWD group decomposition algorithm and the joint convolution module to simultaneously realize disturbance signal processing and disturbance type recognition, thereby improving the recognition efficiency and recognition accuracy of power quality disturbance detection.

[0005] In a first aspect, the present invention provides a power quality disturbance detection method based on SWD and joint convolution, comprising:

[0006] Step 1: Collect power quality disturbance data and mark labels, and perform overlapping sampling on each category in the power quality disturbance data to generate power quality disturbance signal samples;

[0007] Step 2: Decompose the power quality disturbance signal sample by SWD algorithm, and reconstruct the decomposed sample components;

[0008] Step 3: Transform the samples reconstructed in step 2 to obtain a time-frequency diagram;

[0009] Step 4: construct a joint neural convolutional network, and use the time-frequency graph to train the joint convolutional neural network to obtain a power quality disturbance recognition model;

[0010] Step 5: Input the time-frequency diagram corresponding to the power quality disturbance data to be measured into the power quality disturbance identification model to obtain the classification result of the power quality disturbance identification.

[0011] Furthermore, the joint neural convolution network includes a feature fusion module, an SGE module, a frequency dynamic convolution module and a classifier;

[0012] Among them, the feature fusion module is used to extract multi-scale features in the input time-frequency diagram and fuse the multi-scale features to obtain a fused feature diagram; the SGE module further enhances the features in the fused feature diagram based on the output of the feature fusion module to obtain a feature enhancement diagram; the frequency dynamic convolution module further extracts the frequency domain and time domain information of the feature enhancement diagram through a dynamic convolution kernel to obtain an intermediate feature diagram; the classifier obtains the power quality disturbance identification result based on the output of the frequency dynamic convolution module.

[0013] Furthermore, the overlapping sampling in step 1 is as follows:

[0014]

[0015] Where N represents the number of samples of power quality disturbance data collected, L t represents the length of the sampled power quality disturbance data, L represents the length of a single sample signal, and s represents the moving step size of each overlapping sampling.

[0016] Furthermore, the decomposition of the power quality disturbance signal sample by the SWD algorithm in step 2 is to decompose the power quality disturbance signal sample into multiple signal components by iteratively applying the SWF filter, specifically including:

[0017] Step 201: setting relevant parameters of the SWF filter according to the frequency of the power quality disturbance signal sample;

[0018] Step 202: Filter the power quality disturbance signal sample using a SWF filter to obtain an output signal;

[0019] Step 203: updating the power quality disturbance sample signal input to the SWF filter according to the output signal;

[0020] Step 204: Repeat steps 201 to 203 until the power spectrum of the power quality disturbance signal input to the SWF filter is less than a threshold value, thereby obtaining a plurality of signal components.

[0021] Furthermore, in step 2, the decomposed sample components are reconstructed, specifically including:

[0022] Calculate the GINI index for each sample component and determine the average based on the GINI index of all sample components;

[0023] Select the sample component that is greater than the average value as the optimal component;

[0024] The selected optimal components are combined to obtain the reconstructed sample signal; the GINI index formula is as follows:

[0025]

[0026] Among them, x (n) It represents the nth component in the arrangement result obtained by arranging all sample components in ascending order, N represents the total number of sample components, and ||x1|| represents the L1 norm.

[0027] Furthermore, in step 3, the samples reconstructed in step 2 are subjected to continuous wavelet transform to obtain a time-frequency diagram; the continuous wavelet transform formula is as follows:

[0028]

[0029] Among them, f(t) represents the reconstructed sample signal, ψ α,τ (t) represents the wavelet basis function, which is obtained by scaling and translating the mother wavelet:

[0030]

[0031] Among them, α represents the scale factor and τ represents the translation factor.

[0032] Furthermore, the feature fusion module is used to extract multi-scale features from the input time-frequency graph and fuse the multi-scale features to obtain a fused feature graph. The specific process is as follows:

[0033] The time-frequency graph is subjected to a first convolutional layer, a BN layer and a ReLu activation function to obtain a first feature graph;

[0034] The first feature map is simultaneously passed through the second convolution layer and the third convolution layer to obtain the second feature map and the third feature map respectively;

[0035] The second feature map and the third feature map are summed and the fused feature map is obtained through the SiLu activation function.

[0036] Furthermore, the SGE module further enhances the features in the fused feature map based on the output of the feature fusion module to obtain a feature enhancement map. The specific process is as follows:

[0037] Dividing the fused feature map into local feature maps along the channel dimension;

[0038] Approximately calculating the feature vector of the local feature map by global average pooling;

[0039] Performing a dot product operation on the feature vector and the local feature map to generate corresponding coefficients;

[0040] Normalize the corresponding coefficients in the spatial dimension and obtain the attention weight through the sigmoid function;

[0041] Performing a dot product operation on the attention weight and the corresponding local feature map to obtain an enhanced local feature vector;

[0042] All enhanced local feature vectors constitute the feature enhancement map.

[0043] Furthermore, the frequency dynamic convolution module FDC includes: an attention mechanism module, multiple basic convolution kernels and a dynamic perceptron; the corresponding feature processing process includes:

[0044] The feature enhancement graph is passed through an attention mechanism module to extract frequency-adaptive attention weights, wherein the attention mechanism module includes an average pooling layer, a convolution layer, a BN layer, a ReLu activation function, a convolution layer, and a SoftMax layer;

[0045] Using the frequency adaptive attention weights, weighted summation is performed on the weight parameters and bias parameters of the multiple basic convolution kernels to obtain dynamic convolution kernel weights and biases;

[0046] The dynamic convolution kernel weight and bias are used as parameters of the dynamic perceptron, and the parameterized dynamic perceptron is used to process the feature enhancement map to obtain an intermediate feature map.

[0047] In a second aspect, the present invention provides a power quality disturbance detection device based on SWD and joint convolution, comprising:

[0048] A signal acquisition module, used for collecting power quality disturbance data and marking labels, and performing overlapping sampling on each category in the power quality disturbance data to generate power quality disturbance signal samples;

[0049] A signal conversion module, used for performing SWD decomposition on the power quality disturbance signal sample, reconstructing the decomposed sample components, and transforming the reconstructed signal into a time-frequency diagram;

[0050] A model building module, used for building a joint neural convolutional network, and using the time-frequency graph to train the joint convolutional neural network to obtain a power quality disturbance recognition model;

[0051] The classification detection module is used to input the power quality disturbance data to be measured into the power quality disturbance identification model to obtain the classification result of the power quality disturbance identification.

[0052] Beneficial effects of the present invention:

[0053] The present invention proposes a power quality disturbance detection method and device based on SWD and joint convolution, which expands the signal samples by overlapping sampling the collected power quality disturbance data; decomposes the signal samples by SWD algorithm and selects the optimal components according to the GINI index to reconstruct the signal, thereby filtering out the interference components and providing a more accurate data basis for subsequent detection steps; and constructs a joint convolutional neural network to automatically extract features in the time-frequency diagram to complete the detection without relying on manual experience, thereby improving the accuracy and efficiency of power quality disturbance detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A schematic diagram of a flow chart of a power quality disturbance detection method based on SWD and joint convolution provided in an embodiment of the present invention;

[0056] Figure 2 A partial module framework diagram of a joint neural convolutional network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] like Figure 1 As shown, an embodiment of the present invention provides a power quality disturbance detection method based on SWD and joint convolution, including:

[0059] Step 1: Collect power quality disturbance data and mark labels, and perform overlapping sampling on each category in the power quality disturbance data to generate power quality disturbance signal samples;

[0060] Specifically, the power quality disturbance data in the photovoltaic grid-connected system is collected and labeled, and overlapping sampling is performed on each type of power quality disturbance data. The overlapping sampling is shown in the following formula:

[0061]

[0062] Among them, N represents the number of samples collected, L t Represents the length of the sampled power disturbance data, L represents the length of a single sample signal, and s represents the moving step size of each overlapping sampling.

[0063] Step 2: Decompose the power quality disturbance signal sample by SWD algorithm, and reconstruct the decomposed sample components;

[0064] Specifically, the power quality disturbance signal samples are decomposed by the SWD algorithm, and the power quality disturbance signal samples are decomposed into multiple signal components by iteratively applying the SWF filter, including:

[0065] Step 201: setting relevant parameters of the SWF filter according to the frequency of the power quality disturbance signal sample;

[0066] Specifically, according to the frequency of the power quality disturbance signal sample Set the relevant parameters M and δ of the SWF filter:

[0067]

[0068] in, represents the qth identification of frequency ω in the SWD decomposition process, S it (ω) represents the power spectrum of the power quality disturbance signal, q represents the number of times ω corresponds to the center frequency during the operation of the algorithm, and p th is the threshold, M and δ represent the relevant parameters of the SWF group filter.

[0069] Step 202: Filter the power quality disturbance signal sample using the SWF filter to obtain an output signal;

[0070] Specifically, the SWF filter set in step 201 is used to filter the power quality disturbance signal sample to obtain an output signal. In order to obtain a single oscillation component as simple as possible, multiple filtering processes are required until the variance threshold of the results of two consecutive filtering processes is less than D th The SWF filtering iterative processing is completed, and the variance D of the two output signals is as follows:

[0071]

[0072] Among them, x i [n] represents the output signal of the filter after the i-th filtering, x i-1 [n] represents the output signal of the filter after the i-1th filtering.

[0073] Step 203: updating the power quality disturbance sample signal input to the SWF filter according to the output signal;

[0074] Specifically, when the SWF filtering iterative processing is terminated, the obtained output signal is subtracted from the power quality disturbance signal, and the remaining signal is used as a new power quality disturbance signal sample for the next filtering process. The process is as follows:

[0075]

[0076]

[0077] Where x[n] represents the power quality disturbance signal sample, x0[n] represents the new power quality disturbance signal sample, and x i [n] represents the output signal after the SWF filter. represents the extracted sub-oscillation component, τ represents the delay of the algorithm, represents the cross-correlation function, and argmax represents the argmax function.

[0078] Step 204: Repeat steps 201 to 203 until the power spectrum of the power quality disturbance signal input to the SWF filter is less than a threshold value, thereby obtaining a plurality of signal components.

[0079] Specifically, steps 201 to 203 are repeated until the power spectrum of the power quality disturbance signal input to the SWF filter is less than the threshold value, that is, Multiple signal components are obtained.

[0080] Specifically, the decomposed sample components are reconstructed, including:

[0081] Calculate the GINI index for each sample component and determine the average based on the GINI index of all sample components;

[0082] Select the component greater than the average value as the optimal component;

[0083] The selected optimal components are combined to obtain the reconstructed sample signal; the GINI index formula is as follows:

[0084]

[0085] Among them, x (n) It represents the nth component in the result of arranging all sample components in ascending order, for example |x (1) |≤|x (2) |≤…≤|x (N) |, N represents the total number of signal components, and ||x1|| represents the L1 norm.

[0086] Step 3: Transform the samples reconstructed in step 2 to obtain a time-frequency diagram;

[0087] Specifically, the reconstructed sample is a one-dimensional signal, and a two-dimensional time-frequency diagram can be obtained through continuous wavelet transform. The characteristic spectrum of the power quality disturbance signal is represented in the form of a time-frequency diagram, which can reflect the characteristics of the signal more carefully and comprehensively. The continuous wavelet transform is shown as follows:

[0088]

[0089] Among them, f(t) represents the reconstructed sample signal, ψ α,υ (t) represents the wavelet basis function, which is obtained by scaling and translating the mother wavelet:

[0090]

[0091] Among them, α represents the scale factor and υ represents the translation factor.

[0092] Step 4: Construct a joint neural convolutional network and use the time-frequency graph to train the joint convolutional neural network to obtain a power quality disturbance recognition model;

[0093] like Figure 2 As shown, an embodiment of the present invention provides a partial module framework diagram of a joint neural convolutional network; wherein the joint neural convolutional network includes a feature fusion module, an SGE module, a frequency dynamic convolution module FDC and a classifier; wherein the feature fusion module is used to extract multi-scale features in the input time-frequency graph and fuse the multi-scale features to obtain a fused feature graph; the SGE module further enhances the features in the fused feature graph based on the output of the feature fusion module to obtain a feature enhancement graph; the frequency dynamic convolution module further extracts the frequency domain and time domain information of the feature enhancement graph through a dynamic convolution kernel to obtain an intermediate feature graph; the classifier obtains a power quality disturbance recognition result based on the output of the frequency dynamic convolution module.

[0094] Specifically, the feature fusion module is used to extract multi-scale features from the input time-frequency graph and fuse the multi-scale features to obtain a fused feature graph. The specific process is as follows:

[0095] The time-frequency graph passes through the first convolution layer, BN layer and ReLu activation function to obtain the first feature graph;

[0096] Specifically, the time-frequency graph is convolved with a kernel size of conv 7×7 The first convolutional layer extracts the overall features, and the first feature map y1 is obtained through the BN layer and ReLu activation function;

[0097] The first feature map passes through the second convolution layer and the third convolution layer at the same time to obtain the second feature map and the third feature map respectively;

[0098] Specifically, the first feature map y1 is simultaneously convolved by kernels of sizes conv 1×1 and conv 3×3 The second convolutional layer and the third convolutional layer are used to extract subtle features, and two second feature maps y2 and third feature maps y3 carrying different information are obtained respectively;

[0099] The second feature map and the third feature map are summed and the fused feature map is obtained through the SiLu activation function.

[0100] Specifically, the second feature map y2 and the third feature map y3 carrying different information are fused through element-wise summation, and the fused feature map Y is obtained through the SiLu activation function.

[0101] The above process can be expressed as follows:

[0102]

[0103] Among them, Input represents the input time-frequency diagram.

[0104] It can be understood that the information of the feature map is extracted and fused through the convolution layers with convolution kernels of different sizes, thereby enriching the feature extraction information, further enhancing the learning and expression capabilities of the network, and reducing the increase in the depth of the network. And considering that a non-deep network may not have enough nonlinearity, thus limiting its representation ability, the SiLu activation function is used in the feature fusion module.

[0105] Specifically, the SGE module further enhances the features in the fused feature map based on the output of the feature fusion module to obtain a feature enhancement map. The specific process is as follows:

[0106] The fused feature map Y output by the feature fusion module is divided into local feature maps Y along the channel dimension i ;

[0107] Specifically, the input fused feature map Among them, H and W represent the height and width of the feature map respectively, and C represents the channel dimension. In this embodiment, the fused feature map Y is divided into G groups along the channel dimension. Taking one group as an example, each position of the group in the space has a vector representation, that is,

[0108] Approximate the feature vector of the local feature map by global average pooling;

[0109] Specifically, the local feature map Y is approximated by global average pooling. i The characteristic vector g of is as follows:

[0110]

[0111] Perform dot product operation on the feature vector and the local feature map to generate corresponding coefficients;

[0112] Specifically, the feature vector g and the local feature Y i Perform a dot product operation to generate the corresponding coefficient c for each spatial position i , to measure the similarity between them, the corresponding coefficient c i The calculation formula is as follows:

[0113] c i = g·Y i

[0114] Normalize the corresponding coefficients in the spatial dimension and obtain the attention weight through the sigmoid function;

[0115] Specifically, the corresponding coefficient c is then i Perform normalization operation, and then use the sigmoid function to get the attention weight a i , the formula is as follows:

[0116] a i =sigmoid[f BN (c i )]

[0117] Among them, f BN (·) represents the normalization operation, and sigmoid represents the sigmoid function.

[0118] The attention weight a i And the corresponding local feature map Y i Perform dot product operation to obtain the enhanced local feature vector Y i ';

[0119] All enhanced local feature vectors Y i 'Constitute the feature enhancement graph Y'.

[0120] Specifically, the frequency dynamic convolution module FDC includes: an attention mechanism module, multiple basic convolution kernels and a dynamic perceptron; the corresponding feature processing process includes:

[0121] The feature enhancement graph Y' is passed through the attention mechanism module to extract the frequency adaptive attention weight, where the attention mechanism module includes average pooling layer, convolution layer, BN layer, ReLu activation function, convolution layer and SoftMax layer;

[0122] The frequency-adaptive attention weights are used to perform weighted summation on the weight parameters and bias parameters of multiple basic convolution kernels to obtain dynamic convolution kernel weights and biases.

[0123] It can be understood that the basic convolution kernel is related to the module input of FDC, which can further integrate the information after passing through the attention mechanism module, so that the model can better adapt to the changes in input data.

[0124] The dynamic convolution kernel weights and biases are used as parameters of the dynamic perceptron, and the parameterized dynamic perceptron is used to process the feature enhancement map Y' to obtain the intermediate feature map y4.

[0125] The above process can be expressed as follows:

[0126]

[0127] Among them, W(Y') represents the weight of the dynamic convolution kernel, b(Y') represents the bias of the dynamic convolution kernel, and f'(·) represents the dynamic perceptron.

[0128] Specifically, the classifier obtains the power quality disturbance identification result based on the output of the frequency dynamic convolution module.

[0129] Specifically, the classifier obtains the power quality disturbance recognition result based on the intermediate feature map output by the frequency dynamic convolution module, and uses SoftMax as the classifier to classify the power quality disturbance category. The specific formula is as follows:

[0130] y output =SoftMax(Fc(Avgpool 7×7 (y4)))

[0131] Among them, y4 represents the intermediate feature map, FC represents the fully connected layer, Avgpool 7×7 represents the tie pooling operation, y output represents the power quality disturbance detection result, and SoftMax represents the SoftMax function.

[0132] Specifically, the joint convolutional neural network is trained using the time-frequency graph to obtain a power quality disturbance recognition model;

[0133] Specifically, the time-frequency graph is divided into a training set, a validation set, and a test set; the training set and the validation set are input into the photovoltaic grid-connected system power quality disturbance identification model for training. In this embodiment, the power quality disturbance identification model uses the Adam optimizer, sets the learning rate to 0.00001, and the BatchSize to 32. The above parameters are conducive to the convergence of the model. The training set is input into the model for 100 rounds of training. After each round of training, the validation set is tested, and the parameters of the round with the highest accuracy of the validation set during the training process are saved, which is the moment with the best effect. After the training is completed, the parameters are loaded and the test set is identified and predicted.

[0134] Step 5: Input the time-frequency diagram corresponding to the power quality disturbance data to be measured into the power quality disturbance identification model to obtain the classification result of the power quality disturbance identification.

[0135] The embodiment of the present invention further provides a power quality disturbance detection device based on SWD and joint convolution, comprising:

[0136] A signal acquisition module is used to collect power quality disturbance data and mark labels, and perform overlapping sampling on each category in the power quality disturbance data to generate power quality disturbance signal samples;

[0137] A signal conversion module is used to perform SWD decomposition on the power quality disturbance signal sample, reconstruct the decomposed sample components, and transform the reconstructed signal into a time-frequency diagram;

[0138] The model building module is used to build a joint neural convolutional network and train the joint convolutional neural network using the time-frequency graph to obtain a power quality disturbance recognition model;

[0139] The classification detection module is used to input the power quality disturbance data to be tested into the power quality disturbance identification model to obtain the classification result of the power quality disturbance identification.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power quality disturbance detection method based on SWD and joint convolution, characterized in that: include: Step 1: Collect power quality disturbance data and mark labels, and perform overlapping sampling on each category in the power quality disturbance data to generate power quality disturbance signal samples; Step 2: Decompose the power quality disturbance signal sample by SWD algorithm, and reconstruct the decomposed sample components; Step 3: Transform the samples reconstructed in step 2 to obtain a time-frequency diagram; Step 4: construct a joint neural convolutional network, and use the time-frequency graph to train the joint convolutional neural network to obtain a power quality disturbance recognition model; Step 5: Input the time-frequency diagram corresponding to the power quality disturbance data to be measured into the power quality disturbance identification model to obtain the classification result of the power quality disturbance identification.

2. The power quality disturbance detection method based on SWD and joint convolution according to claim 1 is characterized in that: The joint neural convolution network includes a feature fusion module, an SGE module, a frequency dynamic convolution module and a classifier; Among them, the feature fusion module is used to extract multi-scale features in the input time-frequency diagram and fuse the multi-scale features to obtain a fused feature diagram; the SGE module further enhances the features in the fused feature diagram based on the output of the feature fusion module to obtain a feature enhancement diagram; the frequency dynamic convolution module further extracts the frequency domain and time domain information of the feature enhancement diagram through a dynamic convolution kernel to obtain an intermediate feature diagram; the classifier obtains the power quality disturbance identification result based on the output of the frequency dynamic convolution module.

3. The power quality disturbance detection method based on SWD and joint convolution according to claim 1 is characterized in that: The overlapping sampling in step 1 is as follows: Where N represents the number of samples of power quality disturbance data collected, L t represents the length of the sampled power quality disturbance data, L represents the length of a single sample signal, and s represents the moving step size of each overlapping sampling.

4. The power quality disturbance detection method based on SWD and joint convolution according to claim 1 is characterized in that: Decomposing the power quality disturbance signal sample by the SWD algorithm in step 2 is to decompose the power quality disturbance signal sample into multiple signal components by iteratively applying the SWF filter, specifically including: Step 201: setting relevant parameters of the SWF filter according to the frequency of the power quality disturbance signal sample; Step 202: Filter the power quality disturbance signal sample using a SWF filter to obtain an output signal; Step 203: updating the power quality disturbance sample signal input to the SWF filter according to the output signal; Step 204: Repeat steps 201 to 203 until the power spectrum of the power quality disturbance signal input to the SWF filter is less than a threshold value, thereby obtaining a plurality of signal components.

5. The power quality disturbance detection method based on SWD and joint convolution according to claim 1 is characterized in that: In step 2, the decomposed sample components are reconstructed, which specifically includes: Calculate the GINI index for each sample component and determine the average based on the GINI index of all sample components; Select the sample component that is greater than the average value as the optimal component; The selected optimal components are combined to obtain the reconstructed sample signal; wherein, the GINI index formula is as follows: Among them, x (n) It represents the nth component in the arrangement result obtained by arranging all sample components in ascending order, N represents the total number of sample components, and ||x1|| represents the L1 norm.

6. The power quality disturbance detection method based on SWD and joint convolution according to claim 1 is characterized in that: In step 3, the samples reconstructed in step 2 are subjected to continuous wavelet transform to obtain a time-frequency diagram; the continuous wavelet transform formula is as follows: Among them, f(t) represents the reconstructed sample signal, ψ α,τ (t) represents the wavelet basis function, which is obtained by scaling and translating the mother wavelet: Among them, α represents the scale factor and τ represents the translation factor.

7. The power quality disturbance detection method based on SWD and joint convolution according to claim 2 is characterized in that: The feature fusion module is used to extract multi-scale features from the input time-frequency graph and fuse the multi-scale features to obtain a fused feature graph. The specific process is as follows: The time-frequency graph is subjected to a first convolutional layer, a BN layer and a ReLu activation function to obtain a first feature graph; The first feature map is simultaneously passed through the second convolution layer and the third convolution layer to obtain the second feature map and the third feature map respectively; The second feature map and the third feature map are summed and the fused feature map is obtained through the SiLu activation function.

8. The power quality disturbance detection method based on SWD and joint convolution according to claim 2 is characterized in that: The SGE module further enhances the features in the fused feature map based on the output of the feature fusion module to obtain a feature enhancement map. The specific process is as follows: Dividing the fused feature map into local feature maps along the channel dimension; Approximately calculating the feature vector of the local feature map by global average pooling; Performing a dot product operation on the feature vector and the local feature map to generate corresponding coefficients; Normalize the corresponding coefficients in the spatial dimension and obtain the attention weight through the sigmoid function; Performing a dot product operation on the attention weight and the corresponding local feature map to obtain an enhanced local feature vector; All enhanced local feature vectors constitute the feature enhancement map.

9. The power quality disturbance detection method based on SWD and joint convolution according to claim 2 is characterized in that: The frequency dynamic convolution module FDC includes: an attention mechanism module, multiple basic convolution kernels and a dynamic perceptron; the corresponding feature processing process includes: The feature enhancement graph is passed through an attention mechanism module to extract frequency-adaptive attention weights, wherein the attention mechanism module includes an average pooling layer, a convolution layer, a BN layer, a ReLu activation function, a convolution layer, and a SoftMax layer; Using the frequency adaptive attention weights, weighted summation is performed on the weight parameters and bias parameters of the multiple basic convolution kernels to obtain dynamic convolution kernel weights and biases; The dynamic convolution kernel weight and bias are used as parameters of the dynamic perceptron, and the parameterized dynamic perceptron is used to process the feature enhancement map to obtain an intermediate feature map.

10. A power quality disturbance detection device based on SWD and joint convolution, characterized in that: include: A signal acquisition module, used for collecting power quality disturbance data and marking labels, and performing overlapping sampling on each category in the power quality disturbance data to generate power quality disturbance signal samples; A signal conversion module, used for performing SWD decomposition on the power quality disturbance signal sample, reconstructing the decomposed sample components, and transforming the reconstructed signal into a time-frequency diagram; A model building module, used for building a joint neural convolutional network, and using the time-frequency graph to train the joint convolutional neural network to obtain a power quality disturbance recognition model; The classification detection module is used to input the power quality disturbance data to be measured into the power quality disturbance identification model to obtain the classification result of the power quality disturbance identification.

Images