A method, system, device and storage medium for diagnosing power grid faults

Abstract

The present invention relates to the technical field of power grid fault diagnosis, and specifically to a power grid fault diagnosis method, system, device and storage medium. After preprocessing and synchronous compression wavelet transform processing are performed on collected original fault recording data and fault recording simulation data, fault features are obtained. The fault features are successively subjected to full convolution mask processing and decoding processing and input into a power grid fault diagnosis model for model training to obtain a trained power grid fault diagnosis model. Shallow feature extraction is performed on newly input fault recording data, and then multiple convolution processing is performed to obtain deep features. Then, non-parameter attention processing is performed, and weighted deep features are obtained after fusion processing with the deep features. After downsampling processing, the final weighted deep features obtained are fully connected, and the corresponding fault type labels are output, which can improve the accuracy and efficiency of power grid fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid fault diagnosis, and in particular to a power grid fault diagnosis method, system, device and storage medium. Background Art

[0002] The stability and reliability of the power system are crucial to maintaining the smooth operation of modern society. With the advent of the large-scale power grid era characterized by high voltage, AC / DC, and new energy grid connection, the scale of the power system is increasing, and the operation mode is becoming more and more complex, which puts higher requirements on the safety and stability of the system.

[0003] Grid fault diagnosis based on grid fault recording information is an important method to ensure the normal operation and daily maintenance of the power system. When a grid fault occurs, the monitoring system collects a large amount of fault recording data and sends it to the dispatching center from the local automatic device. Grid fault diagnosis technology needs to be able to quickly analyze fault-related data from the massive amount of fault data and find the cause of the fault.

[0004] At present, the more mature power grid fault diagnosis methods can be divided into two categories: one is to use the action information of protection and circuit breaker, that is, the switch quantity information, as the fault information source, and to make judgments and analyses on it through various intelligent technologies; the other is to use the fault recording data and PMU data as the fault information source, and to diagnose the fault by calculating and analyzing the digital information and integrating other fault information sources. Among them, the method of judging by processing the fault recorder recording data has important practical value for the auxiliary judgment of the staff in actual application.

[0005] Traditional fault recording processing methods usually directly input the signal into the fault diagnosis model without preprocessing the recording signal, which results in a lot of useless information in the data, causing the problem of feature information fuzziness, making the trained model less resistant to noise. In addition, existing research work has difficulty processing multi-dimensional data when performing high-dimensional data discrimination tasks, which easily causes feature loss and leads to low accuracy of power grid fault diagnosis. Summary of the invention

[0006] The present invention provides a power grid fault diagnosis method, system, device and storage medium.

[0007] The technical solution of the present invention is as follows:

[0008] A method for diagnosing power grid faults comprises the following steps:

[0009] S1, collecting original fault recording data and fault recording simulation data to obtain a mixed data set, and preprocessing the mixed data set to obtain preprocessed fault recording data;

[0010] S2, perform synchronous compression wavelet transform on the pre-processed fault recording data, compress and rearrange its time-frequency coefficients in the frequency domain on the time-frequency plane, obtain the processed time-frequency distribution data, extract fault feature information based on the time-frequency distribution data, and obtain the fault feature,

[0011] Classify power grid faults based on fault characteristics and obtain power grid fault type labels;

[0012] S3, performing full convolution mask processing on the fault feature to obtain deep information of the fault feature, and decoding the fault feature based on the deep information to obtain a decoded fault feature,

[0013] Input the grid fault type label and the decoded fault characteristics into the grid fault diagnosis model, perform model training, learn the fault type, and obtain a trained grid fault diagnosis model;

[0014] S4: Based on the trained power grid fault diagnosis model, perform fault diagnosis on the newly input fault recording data and output the corresponding fault type label.

[0015] The specific processing of the power grid fault diagnosis model is as follows:

[0016] After the newly input fault recording data is processed by synchronous compression wavelet transform, convolution processing is performed to obtain shallow features.

[0017] The shallow features are input into the depth separable convolution module, and the following processing is performed in the depth separable convolution module: a first depth feature is obtained by a first convolution process, a second depth feature is obtained by a second convolution process after regularization processing on the first depth feature, a third depth feature is obtained by a third convolution process after nonlinear activation processing and global response normalization processing on the second depth feature, a third depth feature is obtained by a parameter-free attention process on the third depth feature to obtain a triple attention weight, and a weighted depth feature is obtained by fusing the triple attention weight and the third depth feature;

[0018] The weighted depth feature is downsampled and then input into the depth separable convolution module for processing again. The process of downsampling and then inputting into the depth separable convolution module for processing again is repeated until the number of processing times of the depth separable convolution module reaches the set value, and the final weighted depth feature is obtained.

[0019] The final weighted deep features are input into the fully connected layer to obtain the probability of each fault type, and the fault type label with the highest probability is output.

[0020] Specifically, the loss function applied in the fully connected layer in S4 is a loss function obtained by weighted summation of a cross entropy loss function and a center loss function, and the formula is as follows:

[0021] ,

[0022] in, L CE is the cross entropy loss function, m is the batch size set, i For the batch i samples, x i For the i The feature map of samples, is the true distribution probability of samples of different categories, is the predicted distribution probability of samples of different categories;

[0023] ,

[0024] in, L C is the center loss function, m is the batch size set, i For the batch i samples, x i For the i The feature map of samples, C yi is the feature center of the category corresponding to the sample;

[0025] ,

[0026] in, is the loss function of the application, is the weight coefficient, L CE is the cross entropy loss function, L C is the center loss function.

[0027] In the S4, the first convolution processing uses a convolution kernel with a convolution kernel size of 7*7, a stride of 1, and a padding pixel size of 3 to extract the first depth feature, the second convolution processing uses a convolution kernel with a convolution kernel size of 1*1 and a stride of 1 to extract the second depth feature, and the third convolution processing uses a convolution kernel with a convolution kernel size of 1*1 and a stride of 1 to extract the third depth feature.

[0028] The global response normalization processing in S4 is specifically to perform global feature aggregation on the second depth feature based on the L2 norm to obtain an aggregated feature vector, perform standard normalization processing on the aggregated feature vector, perform feature calibration to obtain a calibrated second depth feature, and perform a third convolution processing on the calibrated second depth feature.

[0029] The nonlinear activation processing in S4 is specifically to perform nonlinear activation function processing using a Gaussian error linear unit activation function.

[0030] The fault recording simulation data in S1 is specifically generated by random simulation using electronic circuit simulation design software to obtain fault recording simulation data under different fault types.

[0031] The fault recording raw data includes three-phase current and three-phase voltage data.

[0032] The present invention also provides a power grid fault diagnosis system, comprising:

[0033] Data acquisition module: used to collect original fault recording data and fault recording simulation data to obtain a mixed data set, and preprocess the mixed data set to obtain preprocessed fault recording data;

[0034] Data processing module: used to perform synchronous compression wavelet transform processing on the pre-processed fault recording data, compress and rearrange its time-frequency coefficients in the frequency domain on the time-frequency plane, obtain the processed time-frequency distribution data, extract fault feature information based on the time-frequency distribution data, obtain fault features, classify power grid faults based on the fault features, and obtain power grid fault type labels;

[0035] Model training module: used to obtain deep information of fault features after full convolution mask processing of fault features, decode fault features based on deep information to obtain decoded fault features, input power grid fault type labels and decoded fault features into power grid fault diagnosis model, perform model training, learn fault types, and obtain a trained power grid fault diagnosis model;

[0036] Fault classification diagnosis module: It is used to perform fault diagnosis on the newly input fault recording data based on the trained power grid fault diagnosis model and output the corresponding fault type label.

[0037] The specific processing of the power grid fault diagnosis model is as follows:

[0038] The newly input fault recording data is preprocessed and input into the shallow feature extraction layer of the trained power grid fault diagnosis model, and the shallow features are obtained after convolution processing.

[0039] The shallow features are input into the depth separable convolution module, and the following processing is performed in the depth separable convolution module: a first depth feature is obtained by a first convolution process, a second depth feature is obtained by a second convolution process after regularization processing on the first depth feature, a third depth feature is obtained by a third convolution process after nonlinear activation processing and global response normalization processing on the second depth feature, a third depth feature is obtained by a parameter-free attention process on the third depth feature to obtain a triple attention weight, and a weighted depth feature is obtained by fusing the triple attention weight and the third depth feature;

[0040] The weighted depth feature is downsampled and then input into the depth separable convolution module for processing again. The process of downsampling and then inputting into the depth separable convolution module for processing again is repeated until the number of processing times of the depth separable convolution module reaches the set value, and the final weighted depth feature is obtained.

[0041] The final weighted deep features are input into the fully connected layer to obtain the probability of each fault type, and the fault type label with the highest probability is output.

[0042] In addition, the present invention provides a power grid fault diagnosis device, including a processor and a memory, wherein the processor implements the power grid fault diagnosis method as described above when executing a computer program stored in the memory.

[0043] The present invention also provides a computer-readable storage medium storing a computer program, which can implement the steps of the power grid fault diagnosis method as described above when executed by a processor.

[0044] The beneficial effects of the present invention are:

[0045] 1. After preprocessing and synchronous compression wavelet transform processing on the collected original fault recording data and fault recording simulation data, the present invention obtains fault characteristics containing fault feature information, and then inputs them into the fault diagnosis model to diagnose and classify power grid faults, which can effectively solve the problem of fuzzy feature information and improve the anti-noise ability of the fault diagnosis model trained subsequently.

[0046] 2. The present invention is based on a trained power grid fault diagnosis model. It performs shallow feature extraction on the newly input fault recording data and then performs multiple convolution processing to obtain deep features. It then performs parameter-free attention processing and fuses the data with the deep features to obtain weighted deep features. By fusing fault features of different scales, it is not easy to cause feature loss problems when performing high-dimensional data discrimination tasks. By adding parameter-free attention processing, it can not only enhance the model's ability to mine important features, but also enable the entire model to have the ability to eliminate noise and redundant features, thereby improving the accuracy of power grid fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In the attached picture:

[0048] Figure 1 Schematic diagram of the power grid fault diagnosis method in the embodiment. DETAILED DESCRIPTION

[0049] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0050] Example

[0051] This embodiment provides a method for diagnosing power grid faults. Figure 1 , including the following steps:

[0052] S1. Collect original fault recording data and fault recording simulation data to obtain a mixed data set, and preprocess the mixed data set to obtain preprocessed fault recording data.

[0053] Specifically, the present invention uses a fault recorder to collect raw fault recording data; preprocessing includes cleaning and normalizing the raw fault recording data in sequence. The raw fault recording data includes three-phase current and three-phase voltage data. In this embodiment, the fault recording simulation data is specifically generated by random simulation using MULTSIUM electronic circuit simulation design software and using a python interface to drive a digsilent programming interface to obtain fault recording simulation data under different fault types.

[0054] S2, perform synchronous compression wavelet transform on the pre-processed fault recording data, compress and rearrange its time-frequency coefficients in the frequency domain on the time-frequency plane, obtain the processed time-frequency distribution data, extract fault feature information based on the time-frequency distribution data, and obtain the fault feature,

[0055] The power grid faults are classified based on the fault characteristics to obtain the power grid fault type labels.

[0056] Specifically, the processing of synchronous compression wavelet transform is expressed as follows:

[0057] = ,

[0058] in, In order to improve the calculation results of short-time Fourier transform, is the input signal, For signal The Fourier transform of For function exist The conjugate at is the time variable, is the current time, To describe the frequency of the recorded signal, is the Fourier transform of the filter, which represents the frequency response of the system, is a virtual variable, which represents the change of each frequency component in the frequency domain in Fourier transform. Expressed as the output of the Fourier transform, To refer to this calculation operation, Is an imaginary unit.

[0059] Synchro-compression wavelet transform is to complete the compression and rearrangement of MSTFT coefficients on the frequency axis based on the estimated value of the instantaneous frequency operator (IFO). Therefore, the calculation of IFO is a crucial step in the synchronous compression wavelet transform. The first-order expression of IFO is It can be written as:

[0060] ,

[0061] in, is the IFO value, R is the real part of the complex number, plural The phase represents the angle information of the signal changing with time. is the phase change rate converted to frequency, where j Can help convert phase information, is the window function, Represents the original signal input.

[0062] Furthermore, the inverse transform formula of the improved short-time Fourier transform is:

[0063] ,

[0064] in, The final signal is is the window function ( ), The input signal is in the time domain and frequency domain The plural form below indicates, For time.

[0065] The inverse transform of the improved short-time Fourier transform is only integrated in the frequency direction. Therefore, if the improved short-time Fourier transform coefficients are only rearranged in the frequency direction, it will not affect the reconstruction of the signal. The synchronous compression wavelet transform rearrangement formula can be written as:

[0066] ,

[0067] in, is a two-variable function, which represents the fault recording signal pair The response of the filter; For the filter The conjugate value at .

[0068] The original power grid fault signal can be obtained by frequency integrating the coefficients gathered near the ridge line after movement. The reconstruction formula of synchronous compression wavelet transform is expressed as:

[0069] ,

[0070] in, This is the final signal obtained by applying the function In frequency This function is obtained by integrating over time. This function usually represents the reconstruction or recovery of a certain time signal.

[0071] After synchronous compression wavelet transform, more easily processable fault feature information is obtained.

[0072] S3, performing full convolution mask processing on the fault feature to obtain deep information of the fault feature, and decoding the fault feature based on the deep information to obtain a decoded fault feature,

[0073] The power grid fault type label and the decoded fault characteristics are input into the power grid fault diagnosis model, the model is trained, the fault type is learned, and a trained power grid fault diagnosis model is obtained.

[0074] S4: Based on the trained power grid fault diagnosis model, perform fault diagnosis on the newly input fault recording data and output the corresponding fault type label.

[0075] The specific processing of the power grid fault diagnosis model is as follows:

[0076] After the newly input fault recording data is processed by synchronous compression wavelet transform, convolution processing is performed to obtain shallow features.

[0077] The shallow features are input into the depth separable convolution module, and the following processing is performed in the depth separable convolution module: a first depth feature is obtained by a first convolution process, a second depth feature is obtained by a second convolution process after regularization processing on the first depth feature, a third depth feature is obtained by a third convolution process after nonlinear activation processing and global response normalization processing on the second depth feature, a third depth feature is obtained by a parameter-free attention process on the third depth feature to obtain a triple attention weight, and a weighted depth feature is obtained by fusing the triple attention weight and the third depth feature;

[0078] The weighted depth feature is downsampled and then input into the depth separable convolution module for processing again. The process of downsampling and then inputting into the depth separable convolution module for processing again is repeated until the number of processing times of the depth separable convolution module reaches the set value, and the final weighted depth feature is obtained.

[0079] The final weighted deep features are input into the fully connected layer to obtain the probability of each fault type, and the fault type label with the highest probability is output.

[0080] The loss function is applied in the fully connected layer. Specifically, the loss function is obtained by weighted summation of the cross entropy loss function and the center loss function. The formula is as follows:

[0081] ,

[0082] in, L CE is the cross entropy loss function, m is the batch size set, i For the batch i samples, x i For the i The feature map of samples, is the true distribution probability of samples of different categories, is the predicted distribution probability of samples of different categories;

[0083] ,

[0084] in, L C is the center loss function, m is the batch size set, i For the batch i samples, x i For the i The feature map of samples, C yi is the feature center of the corresponding category of the sample.

[0085] Fault recording images have the characteristics of small inter-class differences and large intra-class differences. The intersection-entropy loss function is difficult to optimize the model to an ideal state. Therefore, the center loss function is combined to strengthen the model's optimization of intra-class distances and improve its performance in classification tasks.

[0086] The main goal of the center loss function is to reduce intra-class differences, and it is impossible to predict image classification. Therefore, according to the characteristics of fault recording images, the cross entropy and center loss functions are combined to form the loss function of this paper. The formula is as follows:

[0087] ,

[0088] in, is the loss function of the application, is the weight coefficient, L CEis the cross entropy loss function, L C is the center loss function.

[0089] The first convolution processing uses a convolution kernel with a convolution kernel size of 7*7, a stride of 1, and a padding pixel size of 3 to extract the first depth feature. The second convolution processing uses a convolution kernel with a convolution kernel size of 1*1 and a stride of 1 to extract the second depth feature. The third convolution processing uses a convolution kernel with a convolution kernel size of 1*1 and a stride of 1 to extract the third depth feature.

[0090] The global response normalization process is specifically to perform global feature aggregation on the second depth feature based on the L2 norm to obtain the aggregated feature vector, perform standard normalization on the aggregated feature vector, perform feature calibration to obtain the calibrated second depth feature, and perform the third convolution process on the calibrated second depth feature. The formula is as follows:

[0091] Mapping a spatial feature Xi Aggregate into a global function Vector :

[0092] ,

[0093] in is the image height, is the width, is the total number of channels, is the spatial feature mapping value, and after this operation, we get the vector , is the output three-dimensional vector, is a set of real-valued values.

[0094] Apply a response normalization function to the aggregate values. Specifically, use a standard divergence normalization as follows:

[0095] ,

[0096] Where N is the response normalization function, is a real value, It is The L2 norm of the channel, C is the total number of channels, j is traversing all channels, X j is the feature map of the jth channel.

[0097] Finally, the original input response after feature normalization and calibration is calculated ,

[0098] .

[0099] The nonlinear activation processing is specifically to perform nonlinear activation function processing using a Gaussian error linear unit activation function.

[0100] The present invention also provides a power grid fault diagnosis system, comprising:

[0101] Data acquisition module: used to collect original fault recording data and fault recording simulation data to obtain a mixed data set, and preprocess the mixed data set to obtain preprocessed fault recording data;

[0102] Data processing module: used to perform synchronous compression wavelet transform processing on the pre-processed fault recording data, compress and rearrange its time-frequency coefficients in the frequency domain on the time-frequency plane, obtain the processed time-frequency distribution data, extract fault feature information based on the time-frequency distribution data, obtain fault features, classify power grid faults based on the fault features, and obtain power grid fault type labels;

[0103] Model training module: used to obtain deep information of fault features after full convolution mask processing of fault features, decode fault features based on deep information to obtain decoded fault features, input power grid fault type labels and decoded fault features into power grid fault diagnosis model, perform model training, learn fault types, and obtain a trained power grid fault diagnosis model;

[0104] Fault classification diagnosis module: It is used to perform fault diagnosis on the newly input fault recording data based on the trained power grid fault diagnosis model and output the corresponding fault type label.

[0105] The specific processing of the power grid fault diagnosis model is as follows:

[0106] The newly input fault recording data is preprocessed and input into the shallow feature extraction layer of the trained power grid fault diagnosis model, and the shallow features are obtained after convolution processing.

[0107] The shallow features are input into the depth separable convolution module, and the following processing is performed in the depth separable convolution module: a first depth feature is obtained by a first convolution process, a second depth feature is obtained by a second convolution process after regularization processing on the first depth feature, a third depth feature is obtained by a third convolution process after nonlinear activation processing and global response normalization processing on the second depth feature, a third depth feature is obtained by a parameter-free attention process on the third depth feature to obtain a triple attention weight, and a weighted depth feature is obtained by fusing the triple attention weight and the third depth feature;

[0108] The weighted depth feature is downsampled and then input into the depth separable convolution module for processing again. The process of downsampling and then inputting into the depth separable convolution module for processing again is repeated until the number of processing times of the depth separable convolution module reaches the set value, and the final weighted depth feature is obtained.

[0109] The final weighted deep features are input into the fully connected layer to obtain the probability of each fault type, and the fault type label with the highest probability is output.

[0110] In addition, the present invention also provides a power grid fault diagnosis device, including a processor and a memory, wherein the power grid fault diagnosis method described above is implemented when the processor executes a computer program stored in the memory.

[0111] Finally, the present invention provides a computer-readable storage medium storing a computer program, which can implement the power grid fault diagnosis method as described above when executed by a processor.

[0112] Experimental Section

[0113] The present invention uses MULTSIUM electronic circuit simulation design software and utilizes the Python interface to drive the digsilent programming interface to randomly simulate and generate batch fault recording simulation data of different types. The simulation results show that four types of faults include A, B, and C three-phase grounding. The simulated data are combined with the collected original fault recording data to obtain a mixed data set.

[0114] The numerical characteristics under different faults are as follows:

[0115] (1) Single-phase grounding fault: the single-phase current increases and the single-phase voltage decreases; zero-sequence current and zero-sequence voltage appear; the current increase and voltage decrease are in the same group; the zero-sequence current phase is in the same direction as the fault phase current, and the zero-sequence voltage is in the opposite direction to the fault phase voltage;

[0116] (2) Two-phase short-circuit fault: The two-phase current increases and the two-phase voltage decreases; zero-sequence current and zero-sequence voltage appear; the current increase and voltage decrease are two identical groups; the zero-sequence current vector is located between the two-phase currents of the fault; the fault phase-to-phase voltage leads the fault phase-to-phase current by about 80 degrees; the zero-sequence current leads the zero-sequence voltage by about 110 degrees.

[0117] (3) Two-phase ground short circuit fault: the two-phase current increases and the two-phase voltage decreases; zero-sequence current and zero-sequence voltage appear; the current increase and voltage decrease are the same two groups; the zero-sequence current vector is located between the two-phase currents of the fault; the fault phase-to-phase voltage leads the fault phase-to-phase current by about 80 degrees; the zero-sequence current leads the zero-sequence voltage by about 110 degrees.

[0118] (4) Three-phase short-circuit fault: The three-phase current increases and the three-phase voltage decreases; there is no zero-sequence current or zero-sequence voltage; the fault phase voltage leads the fault phase current by about 80 degrees, and the fault phase-to-phase voltage leads the fault phase-to-phase current by about 80 degrees.

[0119] 1.1 Classification Task Evaluation Indicators

[0120] In order to evaluate the performance of the fault diagnosis model, this paper uses precision, recall and F1 value as evaluation indicators. The calculation method is shown in the following formula. Since the fault diagnosis models designed in this paper are all classification tasks, the evaluation indicators used are the same and will not be repeated in the following text.

[0121] ,

[0122] ,

[0123] ,

[0124] Among them, TP is true positive, FN is false negative, FP is false positive, and TN is true negative.

[0125] Taking the binary classification task as an example, The actual value and the predicted value represent two categories in the binary classification task respectively. The relationship between the four is shown in Table 1.

[0126] Table 1 Relationship among TP, FN, FP and TN

[0127]

[0128] 1.2 Comparative Experiment

[0129] The present invention compares the power grid fault diagnosis method used with the diagnosis methods based on long short-term memory network (LSTM), convolutional neural network (CNN), residual neural network (ResNet), and Inceplon-ResNel model, and experiments are conducted on the two-classification discrimination of fault presence or absence, and the four-classification discrimination of three-phase short circuit, two-phase short circuit, two-phase ground short circuit and single-phase short circuit faults. The hidden layer scale of the model used in each diagnosis method is roughly the same, the number of training rounds is 50 rounds, and the accuracy, precision, recall rate, F1 value and other evaluation indicators are used for comprehensive evaluation. The test set evaluation results are shown in Table 2. Among them, the convolutional neural network represented by CNN shows strong advantages in F1 value and training time within 50 rounds. In the convolutional neural network series, the F1 value generated by the CNN and ResNet model tests is only at a medium level. This is because the deep features extracted by CNN and ResNet during the training process are limited and the feature maps are not rich enough. However, the power grid fault diagnosis model used in the present invention is obviously more advantageous in terms of training time, can effectively integrate the deep features between channels, and has the characteristics of compressing the feature space to improve computing efficiency. Under the same number of rounds and computing scale, it can compress the computing time and have better performance.

[0130] Table 2 Two categories (fault or not)

[0131]

[0132] Table 3 Four categories (three-phase short circuit, two-phase short circuit, two-phase ground short circuit, single-phase short circuit)

[0133]

[0134] The learning rate is set to 0.0075. In order to select the best iteration number epochs value, the loss rate and accuracy of each iteration are recorded during training, and the parameters are optimized using the Adam optimizer. The fault location accuracy and fault recognition accuracy are shown in Table 3. It can be seen from Table 3 that the fault recognition accuracy of the present invention is close to 100%, and the loss value gradually decreases with the increase of epochs, which verifies the feasibility of the model.

[0135] The present invention proposes a power grid fault diagnosis method, system, device and storage medium, which realizes a more effective complex power grid fault diagnosis method by combining feature extraction under data processing with convolutional neural network. The present invention uses a feature extraction method based on synchronous compression wavelet transform to effectively reduce the influence of noise or redundant information in the network, extracts shallow features from the fault recording image, performs deep separable convolution and parameter-free attention processing, and optimizes the loss function, thereby solving the problem of decreased generalization ability of the original model and improving stability. The power grid fault diagnosis model used in the present invention has high accuracy, strong adaptability and good applicability in distribution network fault identification.

[0136] By fusing fault features of different scales, the present invention not only realizes the advantages of the amplitude change characteristics and time-frequency characteristics of the fault recording signal in fault type identification, but also the fused features can more comprehensively characterize the different state information of the fault. Compared with the Inception-ResNet and ResNet models, the fault accuracy is improved by 1.45% and 3.45% on average, and the fault recognition accuracy is improved by 3.47% and 5.50% on average. Compared with CNN and other traditional convolutional models, the present invention not only enhances the ability to mine important features by adding parameter-free attention processing, but also enables the entire model to have the ability to remove noise and redundant features. Even if the fault recording signal is interfered by 5% noise, the fault location and recognition accuracy can still be maintained above 96.75%, indicating that the model has high fault location accuracy. The fault recognition accuracy of the present invention is greatly improved under different types of interference noise. Even in the case of distributed generation (DG) power fluctuations, the diagnosis accuracy is always not less than 97.17%, which fully demonstrates that the model has strong adaptability and robustness in fault location and recognition.

Claims

1. A power grid fault diagnosis method, characterized in that: The following steps are involved: S1, collecting original fault recording data and fault recording simulation data to obtain a mixed data set, and preprocessing the mixed data set to obtain preprocessed fault recording data; S2, perform synchronous compression wavelet transform on the pre-processed fault recording data, compress and rearrange its time-frequency coefficients in the frequency domain on the time-frequency plane, obtain the processed time-frequency distribution data, extract fault feature information based on the time-frequency distribution data, and obtain the fault feature, Classify power grid faults based on fault characteristics and obtain power grid fault type labels; S3, performing full convolution mask processing on the fault feature to obtain deep information of the fault feature, and decoding the fault feature based on the deep information to obtain a decoded fault feature, Input the grid fault type label and the decoded fault characteristics into the grid fault diagnosis model, perform model training, learn the fault type, and obtain a trained grid fault diagnosis model; S4: Based on the trained power grid fault diagnosis model, perform fault diagnosis on the newly input fault recording data and output the corresponding fault type label. The specific processing of the power grid fault diagnosis model is as follows: After the newly input fault recording data is processed by synchronous compression wavelet transform, convolution processing is performed to obtain shallow features. The shallow features are input into the depth separable convolution module, and the following processing is performed in the depth separable convolution module: a first depth feature is obtained by a first convolution process, a second depth feature is obtained by a second convolution process after regularization processing on the first depth feature, a third depth feature is obtained by a third convolution process after nonlinear activation processing and global response normalization processing on the second depth feature, a third depth feature is obtained by a parameter-free attention process on the third depth feature to obtain a triple attention weight, and a weighted depth feature is obtained by fusing the triple attention weight and the third depth feature; The weighted depth feature is downsampled and then input into the depth separable convolution module for processing again. The process of downsampling and then inputting into the depth separable convolution module for processing again is repeated until the number of processing times of the depth separable convolution module reaches the set value, and the final weighted depth feature is obtained. The final weighted deep features are input into the fully connected layer to obtain the probability of each fault type, and the fault type label with the highest probability is output.

2. The power grid fault diagnosis method according to claim 1, characterized in that: The loss function used in the fully connected layer in S4 is specifically a loss function obtained by weighted summation of the cross entropy loss function and the center loss function, and the formula is as follows: ， in, L CE is the cross entropy loss function, m is the batch size set, i For the batch i samples, x i For the i The feature map of samples, is the true distribution probability of samples of different categories, is the predicted distribution probability of samples of different categories; ， in, L C is the center loss function, m is the batch size set, i For the batch i samples, x i For the i The feature map of samples, C yi is the feature center of the category corresponding to the sample; ， in, is the loss function of the application, is the weight coefficient, L CE is the cross entropy loss function, L C is the center loss function.

3. The power grid fault diagnosis method according to claim 1, characterized in that: In the S4, the first convolution processing uses a convolution kernel with a convolution kernel size of 7*7, a stride of 1, and a padding pixel size of 3 to extract the first depth feature, the second convolution processing uses a convolution kernel with a convolution kernel size of 1*1 and a stride of 1 to extract the second depth feature, and the third convolution processing uses a convolution kernel with a convolution kernel size of 1*1 and a stride of 1 to extract the third depth feature.

4. The power grid fault diagnosis method according to claim 1, characterized in that: The global response normalization processing in S4 is specifically to perform global feature aggregation on the second depth feature based on the L2 norm to obtain an aggregated feature vector, perform standard normalization processing on the aggregated feature vector, perform feature calibration to obtain a calibrated second depth feature, and perform a third convolution processing on the calibrated second depth feature.

5. The power grid fault diagnosis method according to claim 1, characterized in that: The nonlinear activation processing in S4 is specifically to perform nonlinear activation function processing using a Gaussian error linear unit activation function.

6. The power grid fault diagnosis method according to claim 1, characterized in that: The fault recording simulation data in S1 is specifically generated by random simulation using electronic circuit simulation design software to obtain fault recording simulation data under different fault types.

7. The power grid fault diagnosis method according to claim 1, characterized in that: The fault recording raw data includes three-phase current and three-phase voltage data.

8. A power grid fault diagnosis system, characterized in that: include: Data acquisition module: used to collect original fault recording data and fault recording simulation data to obtain a mixed data set, and preprocess the mixed data set to obtain preprocessed fault recording data; Data processing module: used to perform synchronous compression wavelet transform processing on the pre-processed fault recording data, compress and rearrange its time-frequency coefficients in the frequency domain on the time-frequency plane, obtain the processed time-frequency distribution data, extract fault feature information based on the time-frequency distribution data, obtain fault features, classify power grid faults based on the fault features, and obtain power grid fault type labels; Model training module: used to obtain deep information of fault features after full convolution mask processing of fault features, decode fault features based on deep information to obtain decoded fault features, input power grid fault type labels and decoded fault features into power grid fault diagnosis model, perform model training, learn fault types, and obtain a trained power grid fault diagnosis model; Fault classification diagnosis module: It is used to perform fault diagnosis on the newly input fault recording data based on the trained power grid fault diagnosis model and output the corresponding fault type label. The specific processing of the power grid fault diagnosis model is as follows: The newly input fault recording data is preprocessed and input into the shallow feature extraction layer of the trained power grid fault diagnosis model, and the shallow features are obtained after convolution processing. The shallow features are input into the depth separable convolution module, and the following processing is performed in the depth separable convolution module: a first depth feature is obtained by a first convolution process, a second depth feature is obtained by a second convolution process after regularization processing on the first depth feature, a third depth feature is obtained by a third convolution process after nonlinear activation processing and global response normalization processing on the second depth feature, a third depth feature is obtained by a parameter-free attention process on the third depth feature to obtain a triple attention weight, and a weighted depth feature is obtained by fusing the triple attention weight and the third depth feature; The weighted depth feature is downsampled and then input into the depth separable convolution module for processing again. The process of downsampling and then inputting into the depth separable convolution module for processing again is repeated until the number of processing times of the depth separable convolution module reaches the set value, and the final weighted depth feature is obtained. The final weighted deep features are input into the fully connected layer to obtain the probability of each fault type, and the fault type label with the highest probability is output.

9. A power grid fault diagnosis device, characterized in that: The method comprises a processor and a memory, wherein the processor implements the power grid fault diagnosis method according to any one of claims 1 to 7 when executing the computer program stored in the memory.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the power grid fault diagnosis method according to any one of claims 1 to 7 is implemented.

Images