Fault identification method and device, equipment and medium

Abstract

The invention relates to a fault identification method and device, equipment and a medium, and relates to the technical field of power distribution networks. The method comprises the following steps: acquiring a first signal corresponding to a target power distribution network; preprocessing the first signal to obtain a second signal corresponding to the target power distribution network; wherein the preprocessing comprises slicing, compression and sampling; inputting the second signal into a fault recognition model to obtain a fault recognition result corresponding to the target power distribution network; wherein the fault identification model is a model obtained through improvement based on a Swin Transform and an auxiliary classification generative adversarial network. By adopting the method, accurate fault identification can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution networks, and particularly relates to a fault identification method and device, equipment and a medium. BACKGROUND

[0002] In recent years, with the rapid growth of electric loads such as new energy vehicle charging piles, smart home devices, and industrial precision instruments, the load types of low-voltage power distribution networks are increasingly diversified, and the operating conditions present the characteristics of high dynamics, strong disturbances, and multiple couplings. Under this background, the risk of series arc faults caused by problems such as aging, loose joints, and insulation damage of electrical lines has increased.

[0003] In the field of series arc fault monitoring of current low-voltage power distribution networks, multiple technical solutions have been formed to meet the demand for fault identification, such as feature detection methods based on hardware, compressed sensing technology, and machine learning algorithms.

[0004] However, the existing solutions all have problems such as being difficult to adapt to fault monitoring under complex conditions, ultimately leading to inaccurate fault identification results. SUMMARY

[0005] The present application provides a fault identification method, device, equipment and medium, which can realize accurate fault identification.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a fault identification method, comprising: obtaining a first signal corresponding to a target power distribution network; preprocessing the first signal to obtain a second signal corresponding to the target power distribution network; wherein the preprocessing includes slicing, compression, and sampling; inputting the second signal into a fault identification model to obtain a fault identification result corresponding to the target power distribution network; wherein the fault identification model is a model improved based on Swin Transformer and auxiliary classification generative adversarial network.

[0007] In one embodiment, preprocessing the first signal to obtain the second signal corresponding to the target power distribution network comprises: based on a sliding time window, slicing the first signal to obtain a first observation group; based on a preset compression ratio, performing compression calculation on the first observation group to obtain a second observation group; based on random sampling, sampling the second observation group to obtain the second signal corresponding to the target power distribution network.

[0008] In one embodiment, based on a preset compression ratio, performing compression calculation on the first observation group to obtain a second observation group comprises: The first observation group is compressed based on a preset compression ratio to obtain a compressed observation group; The first observation group is normalized to obtain a normalized observation group; The compressed observation group and the normalized observation group are compressed based on matrix multiplication to obtain a second observation group.

[0009] In one embodiment, the second observation group is sampled based on random sampling to obtain a second signal corresponding to the target power distribution network, including: The second observation group is zero-padded by filling placeholders to obtain a complete observation group; The placeholders in the complete observation group are marked to obtain a mask sequence corresponding to the complete observation group; The second signal corresponding to the target power distribution network is obtained based on the complete observation group and the mask sequence.

[0010] In one embodiment, the training process of the fault identification model includes: Based on the training signal and the corresponding state label, a training data set is constructed; The training data set is input into an initial model to perform feature extraction, signal reconstruction and state discrimination training to obtain a fault identification model; The initial model is a model improved based on Swin Transformer and auxiliary classification generative adversarial network.

[0011] In one embodiment, the second signal is input into the fault identification model to obtain a fault identification result corresponding to the target power distribution network, including: The second signal is input into the fault identification model to perform signal reconstruction, fault category identification and feature verification to obtain a fault identification result corresponding to the target power distribution network.

[0012] In one embodiment, the fault identification result includes at least one of risk level, fault state, fault time, fault location, fault load category, fault signal feature, fault signal anomaly level and treatment measures.

[0013] In a second aspect, the present application provides a fault identification device, including: An acquisition module for acquiring a first signal corresponding to a target power distribution network; A processing module for pre-processing the first signal to obtain a second signal corresponding to the target power distribution network; wherein the pre-processing includes slicing, compression and sampling; An identification module for inputting the second signal into a fault identification model to obtain a fault identification result corresponding to the target power distribution network; wherein the fault identification model is a model improved based on Swin Transformer and auxiliary classification generative adversarial network.

[0014] In a third aspect, the present application provides a computing device, comprising a memory and a processor; wherein one or more computer programs are stored in the memory, the one or more computer programs comprising instructions; when the instructions are executed by the processor, the computing device performs the method according to any one of the first aspect.

[0015] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, the computer program being used to perform the method according to any one of the first aspect.

[0016] In a fifth aspect, the present application provides a computer program product, comprising one or more computer instructions, when the computer instructions are executed by a computer, the computer performs the method according to any one of the first aspect.

[0017] From the above technical solutions, the present application has at least the following beneficial effects: In the present application, the first signal corresponding to the target power distribution network is obtained, which provides a data basis for subsequent fault analysis; further, the first signal is preprocessed by slicing, compressing and sampling to obtain the second signal corresponding to the target power distribution network, which lays a foundation for obtaining accurate fault identification results; then, the second signal can be input into the fault identification model improved based on SwinTransformer and auxiliary classification generative adversarial network to obtain the fault identification result corresponding to the target power distribution network; the present application provides data for accurate fault analysis by introducing the second signal; moreover, the fault identification model improved based on SwinTransformer and auxiliary classification generative adversarial network is introduced to provide a way for accurate signal fault analysis, and finally accurate fault identification is realized.

[0018] It should be understood that the description of technical features, technical solutions, beneficial effects or similar language in the present application does not imply that all features and advantages can be realized in any single embodiment. On the contrary, it can be understood that the description of a feature or a beneficial effect means that the specific technical feature, technical solution or beneficial effect is included in at least one embodiment. Therefore, the description of technical features, technical solutions or beneficial effects in the specification does not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions and beneficial effects described in the present embodiment can be combined in any appropriate manner. Those skilled in the art will understand that the embodiments can be implemented without one or more specific technical features, technical solutions or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects can be identified in specific embodiments that do not embody all embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 An application environment diagram of a fault identification method provided in an embodiment of the present application; Figure 2 A flowchart of a fault identification method provided in an embodiment of the present application; Figure 3 A flowchart of obtaining a second signal provided in an embodiment of the present application; Figure 4 A structural block diagram of a fault identification device provided in an embodiment of the present application; Figure 5 An internal structure diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The terms "first", "second", and "third" and the like in the specification and the drawings of the present application are used to distinguish different objects, and are not used to limit a specific order.

[0021] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration, and not necessarily to imply any preference or advantage as compared with other embodiments or designs. In the embodiments of the present application, the words "exemplary" or "for example" are used to present the relevant concept in a specific manner.

[0022] In order to make the following embodiments clear and simple, first, a brief introduction of related technologies is given: In recent years, with the rapid growth of new energy automobile charging piles, smart home devices, industrial precision instruments and other power consumption loads, the load types of low-voltage distribution networks are becoming increasingly diversified, and the operating conditions are showing the characteristics of high dynamics, strong disturbances and multi-coupling. Under this background, the risk of series arc caused by aging, loose joints, insulation damage and other problems in electrical lines has increased.

[0023] In the field of series arc fault monitoring of current low-voltage distribution networks, multiple technical solutions have been formed to meet the demand for fault identification, such as hardware-based feature detection methods, compressed sensing technology and machine learning algorithms.

[0024] However, the existing solutions all have problems in adapting to fault monitoring under complex conditions, ultimately leading to inaccurate fault identification results.

[0025] In order to make the technical solutions of the present application clearer and easier to understand, the application scenarios of the technical solutions of the present application will be introduced below in conjunction with the drawings. As shown in Figure 1 The diagram is a schematic diagram of an application scenario provided in an embodiment of the present application.

[0026] In this application scenario, high-precision sensors 104 such as current sensors and voltage sensors are deployed in the power distribution network, and the sensors 104 start to work together to continuously collect the first signals corresponding to the target power distribution network, and transmit these first signals to the server 103 in real time and low delay through wireless communication technologies such as 4G / 5G; then, the server 103 analyzes and calculates after receiving the first signals to obtain the fault identification result corresponding to the target power distribution network. Finally, the fault identification result is pushed to the intelligent terminal 102 of the field operator or the operation personnel in real time through a secure API (Application Programming Interface, application programming interface), and the special application program or graphical interface on the terminal will clearly and intuitively display the target position and action guidance in various forms such as visual graphics, highlighted numbers and text instructions, to assist the operator to make decisions or automatically executed by the system, so as to complete the closed loop from physical world perception to digital instruction generation.

[0027] In order to make the technical solutions of the present application clearer and easier to understand, a fault identification method provided by the embodiments of the present application is introduced below in combination with the above application scenario. As shown in the figure, it is a flow chart of a fault identification method provided by an embodiment of the present application. Figure 2 As shown in the figure, it is a flow chart of a fault identification method provided by an embodiment of the present application.

[0028] S201, obtaining the first signal corresponding to the target power distribution network.

[0029] Among them, the target power distribution network refers to a specific power distribution network that needs to be fault-identified, which has a clear geographical range, topological structure and operating parameters, etc., such as a certain city 10kV urban power distribution network, a certain industrial park 35kV enterprise power distribution network, etc., and its characteristics can include but are not limited to line length, load type, access equipment, etc.; the first signal can refer to the original arc signal collected by the sensor in the target power distribution network, including multi-source current signals in fault state and non-fault state, and the signal format can be digital signal after analog-to-digital conversion of analog signal, which meets the sampling frequency requirement of real-time monitoring of power distribution network, such as not less than the collection requirement corresponding to arc high-frequency disturbance.

[0030] Optionally, the sensors in the target power distribution network can include high-dynamic-range current sensors and analog-to-digital converters, etc.

[0031] Exemplarily, the power grid operation state can be monitored in real time by the power distribution network central control module, and the sensor acquisition module is triggered to work when it is detected that the current fluctuation exceeds the normal threshold, such as the current mutation amplitude > 5%, or a timing acquisition instruction is received; then, the current signals of each line of the target power distribution network, such as the current signals of the outlet end of the distribution box and the low-voltage side of the transformer, can be acquired by using a high dynamic range current sensor, and high-frequency interference signals can also be filtered synchronously to avoid frequency aliasing in the sampling process, such as the high-frequency interference signal filtering can be realized by an anti-aliasing filter module.

[0032] Further, the analog current signal can also be converted into a digital signal to form an initial original signal, such as the signal format conversion can be realized by an analog-to-digital converter; the collected digital signal can also be preliminarily checked to eliminate abnormal signals caused by sensor failure, such as signals with a signal amplitude of 0 or exceeding the sensor range, and valid original signals are retained as the first signal; the signal acquisition time, the acquisition line number and other related information can also be recorded at the same time.

[0033] S202, pre-processing the first signal to obtain a second signal corresponding to the target power distribution network.

[0034] Among them, the pre-processing includes slicing, compression and sampling; the second signal can refer to the standardized data of the first signal after slicing, compression and sampling pre-processing, which can have the characteristics of uniform data length, complete effective feature reservation, adaptive model input format, etc.

[0035] It should be noted that the purpose of preprocessing is to eliminate noise, redundancy and inconsistency in the original data, extract and retain key features, shape the data into a model-friendly format, and thus improve the training efficiency, generalization ability and final analysis accuracy of the model; for example, slicing can cut the first signal of a long time sequence into multiple shorter and continuous signal segments according to certain rules or time windows; compression can reduce the data volume or dimension of the data while retaining key features of the signal, effectively reducing the data volume, and the compression method can be wavelet transform (using wavelet basis functions to decompose the signal, retaining important low-frequency approximation coefficients and high-frequency detail coefficients, discarding unimportant coefficients, and achieving compression and denoising), principal component analysis (projecting high-dimensional signals onto a few main principal component directions, and using the linear combination of these principal components to approximately represent the original signal, achieving dimension reduction), piecewise aggregation approximation (segmenting the signal and using the average value of each segment to represent the data in that segment), Huffman coding and feature extraction; sampling is to unify the sampling rate of all data to a standard value to ensure the consistency of the model input, which can include downsampling and upsampling, downsampling is to extract data points from high sampling rate signals at certain intervals to form low sampling rate signals. Usually, anti-aliasing filtering is performed before downsampling to prevent frequency aliasing; upsampling is to insert new data points between the original data points through interpolation algorithms (such as linear interpolation and spline interpolation) to increase the sampling rate of the signal.

[0036] For example, the first signal of the power distribution network can be obtained first, and the signal change frequency, such as the load fluctuation period, and the key feature distribution, such as the fault transient signal peak position, can be combined to determine the slice length, the starting position and the slice interval; among them, the slice length needs to ensure that the complete feature period is covered, the starting position needs to avoid signal edge noise, and the slice interval needs to avoid feature repetition or omission; then, the slicing operation can be performed on the first signal to decompose the long continuous signal into several fixed-length sub-signal segments, reducing the data complexity for subsequent processing; Further, the data redundancy type, such as time domain redundancy and frequency domain redundancy, can be analyzed for the sliced sub-signal segments, and the appropriate compression algorithm can be selected, such as if it is a smooth load signal, then predictive coding is selected, that is, the current value is predicted through historical data to reduce the error data volume; if it contains transient high-frequency features, then transform coding is selected, such as wavelet transform, which converts the signal to the frequency domain and compresses the low-frequency redundancy, removing redundant data while retaining effective features, completing sub-signal compression; Further, according to the sampling theorem, the sampling frequency is not less than 2 times the highest frequency of the signal, and the sampling frequency is set in combination with the highest frequency of the power distribution network signal, such as a harmonic signal of 2000 Hz, and the distribution of sampling points is determined synchronously, such as encrypting sampling at feature mutations and appropriately sparsely sampling in smooth sections. Discrete sampling is performed on the compressed sub-signals, sampling errors can be processed through hardware filtering (such as an anti-aliasing filter) and software calibration (such as linear interpolation to correct deviations), and finally a second signal with uniform data length, complete effective features, and an input format adapted to the model is generated.

[0037] S203, inputting the second signal into the fault identification model to obtain a fault identification result corresponding to the target power distribution network.

[0038] The fault identification model is a model improved based on a Swin Transformer (sliding window transformer) and an auxiliary classification generative adversarial network; and the fault identification result includes at least one of a fault type, a risk level, a fault state, a fault time, a fault position, a fault load category, a fault signal feature, a fault signal anomaly level, and a treatment measure.

[0039] Optionally, the fault identification result includes rich and key information, which is crucial for fault handling and maintenance of the power distribution network; for example, according to the severity of the fault, the influence range, the possible loss and other factors, the fault can be divided into different levels, such as low risk, medium risk, high risk and extremely high risk. The low risk fault may only be a short signal fluctuation, and has little influence on the overall operation of the power distribution network. However, the high risk or even extremely high risk fault may cause large-scale power failure, serious damage to equipment and other serious consequences, and needs to be immediately handled by taking emergency measures. Through the clear risk level, the operation and maintenance personnel can quickly judge the emergency degree of the fault and reasonably allocate repair resources. The fault state can indicate what kind of fault condition the power distribution network is currently in, such as short circuit fault, open circuit fault, ground fault and the like. Different fault states have different electrical characteristics and processing methods, and accurate judgment of the fault state is the basis for taking correct handling measures subsequently. The fault time can accurately record the time when the fault occurs, which is of great significance for analyzing the cause of the fault, tracing the development process of the fault and evaluating the influence time of the fault on the operation of the power grid. For example, by comparing the sequence of fault times of different devices, it is possible to find the correlation between faults and find the root cause of the fault. The fault location can determine the specific line, node or device in the power distribution network where the fault occurs, which is the key information for quickly repairing the fault. After the fault location is clear, the operation and maintenance personnel can quickly go to the scene for maintenance, reduce the fault outage time and improve the power supply reliability. The fault load category refers to the load type affected by the fault, such as residential power load, industrial power load, commercial power load and the like. Different load categories have different requirements for power supply reliability, and understanding the fault load category helps to evaluate the influence of the fault on different user groups and reasonably arrange the sequence of restoring power supply. The fault signal characteristics can include the characteristics of the second signal in the time domain, frequency domain and the like when the fault occurs, such as the amplitude variation law, frequency component distribution, phase shift and the like. These characteristics are important basis for the fault identification model to judge the fault type and state, and are also key data for studying the fault mechanism and improving the fault identification method. The fault signal anomaly level can reflect the degree of deviation of the fault signal from the normal state, and is classified from slight anomaly to severe anomaly. The higher the anomaly level, the more serious the fault, and the greater the difference between the signal characteristics and the normal state. The handling measures are specific fault handling suggestions according to the above identification results, such as immediately cutting off the fault line, repairing the equipment, adjusting the power grid operation mode and the like. These handling measures provide specific operation guidance for the operation and maintenance personnel, helping them to quickly and effectively solve the fault and restore the normal operation of the power distribution network.

[0040] It should be noted that the Swin Transformer regards the input second signal as a special signal image, and uses its hierarchical structure to extract different levels of features from the original signal step by step, from small local signal mutation details to large overall signal trend changes, which can be accurately captured; the auxiliary classification generative adversarial network adds strong learning and discrimination ability to the fault identification model, which is composed of a generator and a discriminator, the generator is responsible for generating pseudo signals similar to the characteristics of real fault signals, and the discriminator tries to distinguish between real signal characteristics and pseudo signals generated by the generator, in the process of continuous confrontation game, the signal generated by the generator becomes more and more realistic, and the discrimination ability of the discriminator also becomes stronger, so that the whole model understands the fault signal characteristics more deeply, and can accurately identify the fault type from the complex and changeable second signal. This fault identification model combining Swin Transformer and auxiliary classification generative adversarial network has strong feature extraction ability and accurate classification and discrimination ability, and provides strong support for power distribution network fault identification.

[0041] One implementation manner is to input the second signal into the fault identification model to perform signal reconstruction, fault category identification and feature verification, and obtain the fault identification result corresponding to the target power distribution network.

[0042] Among them, signal reconstruction is an operation of fine data repair and feature enhancement on the signal, which can solve the problems of local data loss, signal distortion and feature ambiguity in the signal, and restore the true electrical state. For example, for the missing of 1-2 sampling points of voltage signal due to communication delay, the data is completed by interpolation algorithm (such as linear interpolation, cubic spline interpolation); for the peak distortion of fault instantaneous current signal, the noise and effective features are separated by wavelet transform method, and the smooth current waveform reflecting the essence of fault is reconstructed; Fault category identification is a process of extracting fault features (such as current surge of short-circuit fault, voltage drop of single-phase grounding fault, and zero-sequence current increase of single-phase grounding fault) based on the reconstructed signal, and matching with the pre-set fault feature library, and finally determining the fault type of power distribution network. Common fault categories can include but are not limited to three-phase short-circuit, two-phase short-circuit, single-phase grounding fault, broken line fault, insulation damage fault, etc., and the identification result can specify the specific type of fault and the line section to which it belongs; Feature verification is a process of ensuring the reliability of the result through multi-dimensional verification mechanism, that is, a process of verifying the rationality of matching fault features and categories in reverse. The verification dimensions can include, but are not limited to, electrical quantity conservation verification (such as compliance with Kirchhoff's law of power and current before and after the fault), feature consistency verification (such as single-phase grounding fault should meet zero sequence voltage rise + zero sequence current increase at the same time, if only a single feature is met, a verification warning is triggered), historical data comparison verification (calculate the similarity of the current fault feature and the historical fault feature of the same type on the same line, and if the similarity is lower than the threshold, re-identify), etc.

[0043] It should be noted that the fault identification model includes a Swin Transformer-based super-resolution fault arc generator module, a discriminator based on auxiliary classification, an auxiliary feature extractor, and a generator network loss function. The Swin Transformer-based super-resolution fault arc generator module captures the correlation between windows through a local attention mechanism, realizes cross-window feature extraction, and considers local and global multi-scale features to better restore the detailed texture features of one-dimensional fault arcs. First, the input x in The shallow features are obtained through the convolution layer, and the recovery process is mapped from the signal space to the high-dimensional space. The shallow feature output result is: (3) wherein C1() is a one-dimensional convolution operation; and θc1 is a one-dimensional convolution operation parameter.

[0044] The Swin Transformer layer is introduced to extract the shallow features in depth, mainly including window multi-head self-attention and sliding window multi-head self-attention. The non-overlapping window division and the translation window division result are: (4) wherein w is the window size; s is the moving step, usually set as w / 2; x fix (t) is the feature segment corresponding to the tth non-overlapping window; x shifted (t) is the feature segment corresponding to the tth translation window, and t is the time window index, indicating the window number; x c [] is the feature sequence subjected to window division, and the interval in the bracket indicates the slice interval; w is the window size, that is, the number of sampling points contained in each window; is the segment taken from c to in the sequence x ; s is the moving step; is the window that is translated to the right by After sampling one point, take another segment of the same length.

[0045] The self-attention mechanism of different local windows is calculated through a multi-head attention mechanism. Relative positions are encoded to focus on the relative positions between elements. Assuming that each local window has h heads that are computed in parallel, the operation process is as follows: (5) (6) (7) in, For the first A local window, For the first The first window The query matrix of each attention head is a learnable parameter tensor; For the first The first window The key matrix of each attention head is a learnable parameter tensor; It is the first The first window The value matrix of each attention head is a learnable parameter tensor; Softmax(⋅) is the normalized exponential function; No. The first window The output features of each attention head; Indicates the first The output results of the multi-head attention in each window; Concat(⋅) is the concatenation operation; No. The final window output features after multi-head attention and linear mapping of each window; Represents all windows (total) The set of linear mapping output results (number of elements); Window_reverse(⋅) is the window reverse mapping operation; This represents the feature output after window inversion; Dense(·) is a linear transformation operation, θ d Let B be the parameter tensor of the linear transformation operation, D be the scaling factor, and B be the parameter tensor of the linear transformation operation. (t) Let x be the relative position offset parameter tensor. st This is for splicing and outputting windows.

[0046] Furthermore, a multi-level skip connection method is adopted between different layers to enhance the network's generalization and anti-interference ability while ensuring network depth. Finally, upsampling is performed through subpixel convolution operations, and the final output arc signal is obtained through convolution operations to improve the accuracy of the generator in recovering fault arc signals. (8) where x st is the feature output result of window reflection, that is, the splicing output of the window; C1() represents a convolutional transformation operation, and subscript i represents the layer order number of the convolutional layer, different layers having different parameter sets ; PixelShuffle(⋅) is a pixel rearrangement operation; x final is the signal reconstruction result.

[0047] Based on the discriminator with auxiliary classification, the high-precision fault arc signal restored by the generator and the original category are taken as the input of the discrimination module. The arc fault is divided into four categories: the resistive load category presents a sine wave, and the overload is easy to ignite combustible materials; the gas discharge lamp category produces high-frequency noise due to gas ionization; the motor category presents a triangular wave current and a starting surge; and the switching power supply category has a wide frequency harmonic characteristic. The different operating condition load categories are One-Hot coded, and the output fault state marker is 01, and the non-fault marker is 00. The discriminator module can not only distinguish the true and false samples, but also distinguish the categories of the samples, so as to realize the reconstruction and identification of the arc fault monitoring; therefore, the objective function of the ACGAN (Auxiliary Classifier Generative Adversarial Network) discriminator is divided into two parts, which are the log-likelihood estimation of the true and false samples and the log-likelihood of correct category identification; the deep convolutional network is used to extract the deep features of the generator output super-resolution fault arc signal, and the true / false binary discrimination is completed after the full connection layer and the Sigmoid function activation, and the category identification of the sample is completed through the Softmax activation function; the reconstruction quality and the identification accuracy are improved through the game between the discrimination module and the generation module; the discrimination module of the auxiliary classification module shares the network weight.

[0048] The traditional generative adversarial network may have problems such as mode collapse and unstable training, resulting in unbalanced training and low quality of the output signal result, while the recovery signal based on the random singular value decomposition method in the present application and the auxiliary feature extraction network of the original signal fully exert the consistency characteristics of the potential features of the generator recovered super-resolution signal data and the original data features, so that the generator output signal restores the original signal information as much as possible; the method uses a low-rank orthogonal matrix to approximate the range of the input signal, and then performs SVD (Singular Value Decomposition) decomposition, which saves the network calculation cost compared with the traditional SVD algorithm; the approximate SVD matrix is taken as the potential feature extracted by the auxiliary feature network, and the potential feature and the distribution feature of the original signal are gradually approximated, and the calculation formula is as follows: (9) where H is a low-rank approximation matrix; is the left matrix after the traditional SVD approximate decomposition. is the singular value matrix after the approximate decomposition of the traditional SVD; is the transpose of the right matrix after the approximate decomposition of the traditional SVD;x final is the signal reconstruction result; P(·) is the processing operation of the feature extraction module; ApproxSVD(·) is the approximate singular value decomposition operator.

[0049] The generator network loss function is composed of weighted losses. (10) wherein, L cont is the generated content loss; L ad is the adversarial network loss; L P is the auxiliary network loss; ε, τ and μ are network loss weight coefficients; θ G is the generator optimization parameter; is the generator network loss; is the minimization of the generator network loss.

[0050] The root mean square error is used to calculate the content loss between the recovered signal and the original signal, and the maximum similarity between the recovered data and the original data can be realized by minimizing the generated content loss, which can be specifically represented as: (11) wherein, is the generated content loss; is the output result after the processing operation of the generator; x is the original arc signal; is the two-norm operation; L is the length of the observed signal; x in is the input low-resolution signal, i.e., the second signal.

[0051] The auxiliary network loss function is used as a regularization term of the generator loss function to enhance the local-spatial invariance of the G-Network, avoid pattern collapse, and optimize the learning training process. The latent feature loss function is described by using the l1 norm, which can be specifically represented as: (12) wherein, represents the feature vector output by the feature extraction module after processing the original arc signal.

[0052] To avoid the problem of over-smoothness of the recovered signal and lack of high-frequency fluctuation information, the quality of the recovered signal can be further improved through the adversarial loss between the discriminator and the generator. Unlike the traditional GAN, the D-Network is not only responsible for distinguishing the true and false samples, but also identifies the category of the samples; the discriminator loss function can be represented as: (13) wherein, represents the discriminator network loss, represents the expectation of the real data distribution p H , C(·) is the classification output of the discriminator, the classifier adopts cross-entropy loss as the loss function; c is the class of the sample; D(·) is the output of the discriminator network; p H is the real data distribution; is a low-resolution sample input by the generator; p h is the probability distribution of the low-resolution input sample; p c is the class distribution; represents the output of the generator under the input of the low-resolution sample and the class c.

[0053] Therefore, the super-resolution generator network adversarial loss is: (14) wherein, represents the super-resolution generator network adversarial loss.

[0054] Exemplarily, from the power grid data acquisition system, according to the current monitoring demand of the target power grid, such as monitoring the 10kV East Ring line, the second signal that has completed preprocessing is screened out, and other line signals unrelated to the target power grid are excluded to avoid data interference; further, according to the input data specification of the fault identification model, such as data sampling frequency 1kHz, conversion, and feature dimension time step x signal type, the second signal after screening is format converted, the originally discrete voltage and current signals can be integrated into structured data of time sequence + feature label structure to ensure that the model can be read; further, through an encryption communication protocol, the formatted second signal is transmitted from the power grid edge terminal to the server (local server or cloud platform) where the fault identification model is deployed, and the data integrity needs to be checked in real time during the transmission process, such as through CRC (Cyclic Redundancy Check), to prevent signal loss or tampering; further, the fault identification model receives the transmitted second signal, loads the signal into the model through the input layer interface, and at the same time triggers the model initialization program, such as parameter reset, feature extractor preheating, etc., to prepare for the subsequent signal reconstruction link; ​Further, the fault identification model performs signal reconstruction through detection, repair, enhancement, and output reconstruction processes; for example, the fault identification model calls a built-in signal quality detection module to perform segment-by-segment scanning on the input second signal, identifies defect types such as data loss (blank sampling points caused by temporary sensor failure), signal distortion (current signal appears as a sharp pulse due to electromagnetic interference at the moment of failure), and feature ambiguity (voltage characteristics are not clear due to superimposed load fluctuations and fault signals), and records defect locations and severity; further, targeted repair operations can be performed for different defect types; for data loss, interpolation repair methods can be used, such as using cubic spline interpolation based on the previous and next 10 normal sampling points to complete the missing data; for signal distortion, wavelet threshold denoising methods can be used to separate noise components, such as filtering interference signals with a frequency higher than 500 Hz and retaining fault characteristic signals; for feature ambiguity, signal decomposition methods such as EMD (Empirical Mode Decomposition) can be used to decompose the mixed signal into multiple intrinsic mode functions, and extract the modal components related to the fault; Further, fault features in the reconstructed signal can be enhanced through normalization processing, difference operation, and other feature amplification algorithms, such as amplifying the voltage difference before and after the fault from 0.2 kV to a standardized 0.8, so that the subsequent fault category recognition stage of the model can more clearly capture the fault features; after signal reconstruction is complete, the model transmits the reconstructed signal to the internal feature cache area, and generates a signal reconstruction report at the same time, recording the number of repaired defects, the repair algorithm used, and the quality score of the reconstructed signal, such as 95 points, representing the similarity between the reconstructed signal and the real fault signal, providing a reference for subsequent feature verification.

[0055] Further, the fault identification model performs fault category identification through processes such as feature extraction, pattern matching, preliminary determination, and output of candidate categories. For example, based on the convolutional layer of the CNN (Convolutional Neural Network) and the traditional model, multi-dimensional fault features are extracted from the reconstructed signal, such as time domain features including current peak value, voltage effective value, signal mutation time, frequency domain features including main frequency component and harmonic content of the fault signal, and transient features including voltage sag amplitude within 0.02s after the fault occurs, to finally form a feature vector matrix, such as 100 features x 1 sample. The extracted feature vector matrix is input into the classification decision module of the fault identification model, and is matched with the fault pattern library constructed in the model training stage. The fault pattern library stores standard feature templates of various faults, such as single-phase ground fault standard features of zero sequence voltage ≥ 0.8 times rated voltage and zero sequence current ≥ 5A. Similarity calculation algorithms such as cosine similarity and Euclidean distance are used in the matching process to calculate the similarity between the current feature vector and each fault template. Further, according to the similarity ranking result, the category corresponding to the fault template with the highest similarity can be selected as the preliminary fault category. For example, the similarity between the current feature vector and the single-phase ground fault template is 92%, and the similarity between the current feature vector and the two-phase short circuit template is 45%, so the preliminary determination is single-phase ground fault, and the similarity value is recorded as the basis for subsequent verification. The preliminary fault category and the corresponding similarity and feature matching details (such as zero sequence voltage feature matching degree 95% and zero sequence current feature matching degree 88%) are transmitted to the feature verification module as the candidate fault category. Further, the fault identification model performs feature verification through processes such as calling verification rules, verifying feature consistency, comparing with historical data, and correcting / confirming the category. By loading the built-in feature verification rule library, which is constructed based on the electrical laws of distribution network faults (such as Kirchhoff's law and fault timing logic) and operation and maintenance experience, including multiple verification rules, such as single-phase ground fault must meet both zero sequence voltage rise and zero sequence current increase, and short circuit fault must be accompanied by sudden current increase and voltage drop. The features of the candidate fault category are compared with the verification rules to determine whether they conform to the rules. If the candidate category is single-phase ground fault, but only zero sequence voltage rise is detected and zero sequence current increase is not detected, a feature inconsistency warning is triggered, and the model automatically returns to the feature extraction link to extract key features again. The historical fault database of the power distribution network can also be called to retrieve the same type of fault records of the target power distribution network in the past three years, the similarity of the features of the current candidate category and the historical fault features is calculated, if the similarity is lower than a preset threshold (such as 70%), a secondary matching process is started, and the other categories in the fault mode library are matched again, if the feature consistency verification is passed and the historical data comparison similarity meets the standard (such as ≥80%), the candidate fault category is confirmed as the final category, if the verification fails, the problem can also be located through a reverse tracing algorithm, such as incomplete feature extraction, outdated mode library template, etc., and the fault category identification is re-executed after correction until the fault category that meets the verification rules is output.

[0056] Finally, the fault identification model generates and outputs the fault identification result of the target power distribution network through the processes of integrating information, formatting results, associating operation and maintenance suggestions, and pushing to the terminal, confirms the fault category, fault location (which can be calculated by impedance method combined with line parameters), fault severity (which can be determined based on fault duration and current voltage deviation amplitude), confidence, etc. Information is integrated into a structured data set, and the integrated information is formatted according to the standard format of the power distribution network operation and maintenance system to ensure that the result includes fault basic information-feature basis-confidence, etc.

[0057] Optionally, the fault identification model can also match the corresponding operation and maintenance suggestions according to the fault type and severity, such as suggesting immediately arranging operation and maintenance personnel to carry a grounding fault detector to 3.5km for investigation, giving priority to checking whether the line insulator is damaged, disconnecting the first branch switch of the east ring line during the disposal process to avoid the impact of load, and can push the formatted fault identification result and operation and maintenance suggestions to the terminal equipment of the operation and maintenance personnel through the communication interface of the power distribution network operation and maintenance platform, and trigger an alarm prompt such as a sound alarm, a pop-up prompt, etc. to ensure that the operation and maintenance personnel obtain the fault information and start the disposal process in the first time.

[0058] The training process of the fault identification model includes: Based on the training signal and the corresponding state label, a training data set is constructed, the training data set is input into an initial model, and the training of feature extraction, signal reconstruction and state discrimination is performed to obtain the fault identification model.

[0059] The state label can refer to the label information used for model training, which can include two types: one is the fault state label, such as the fault state label is 01 and the non-fault state label is 00; the other is the load category label, such as it can be divided into four categories of resistive load, gas discharge lamp, motor and switching power supply; the initial model is a model improved and built based on Swin Transformer and auxiliary classification generative adversarial network.

[0060] Exemplarily, the operation signals in the past 3-5 years can be extracted from the historical database of the power distribution network SCADA (Supervisory Control and Data Acquisition System), FTU (Feeder Terminal Unit), and DTU (Distribution Terminal Unit), covering normal operation, faults (three-phase short circuit, two-phase short circuit, single-phase grounding, and broken line), load fluctuations (such as resident electricity peak and industrial motor start-stop), and the like. The signal types include three-phase voltage, three-phase current, zero-sequence voltage, zero-sequence current, and power factor, and the sampling frequency is set to 1 kHz. A power distribution network simulation model can also be built by PSCAD / EMTDC and MATLAB / Simulink to simulate extreme faults (such as high-resistance grounding and intermittent grounding) and complex load combinations (such as resistive load + motor load mixed operation) that are difficult to collect in real scenarios. For example, a gradual change scenario is set for the grounding resistance from 10Ω to 1000Ω to generate zero-sequence current signals under different grounding resistances. The mixed signal of the impact current when the motor starts and the harmonic superposition of the switching power supply is simulated to supplement the scene gap of the real data.

[0061] Further, the collected original signals can be preprocessed to remove invalid data and retain key features. For example, abnormal signals caused by sensor failure (such as extreme data with current suddenly becoming 0 or far exceeding the rated value) can be identified and deleted by the 3σ principle (mean ± 3 times standard deviation), and outliers can be removed. Wavelet threshold denoising method can be used to filter electromagnetic interference (such as high-frequency noise) and measurement errors in the signal, and retain fault transient characteristics such as current spikes during short-circuit fault and zero-sequence voltage mutations during grounding fault for denoising. Fault signals can also be intercepted, such as intercepting the time period of 0.02s before fault to 0.1s after fault to cover the fault transient process. For normal operation signals, a stable segment of 1s in length is intercepted to ensure that the time dimension of each signal is uniform (such as fixed at 120 sampling points, corresponding to 0.12s under 1kHz sampling frequency), facilitating subsequent model input format uniformity.

[0062] Further, in order to facilitate model classification calculation, binary coding can be used to clearly distinguish between "fault" and "non-fault", and to refine fault types. The basic binary label can be used for distinction, such as non-fault state (normal operation, simple load fluctuation) being uniformly labeled as "00"; fault state being labeled as "01", which only distinguishes the basic scenario of fault / non-fault; and the fault type subdivision label can be used for distinction, that is, if the model needs to identify specific fault types, it can be extended to multi-ary coding, such as "000-non-fault, 001-three-phase short circuit, 010-two-phase short circuit, 011-single-phase ground fault, 100-wire breakage" (the number of fault types can be adjusted according to actual needs), and the labeling needs to be combined with the fault recording report of the distribution network to confirm the actual fault type corresponding to each fault signal.

[0063] Further, based on the electrical characteristic differences of the loads of the distribution network, the loads are divided into four categories, and decimal or binary coding is used for labeling, such as resistive load category (such as incandescent lamp, electric heater): current and voltage are in phase, power factor ≈1, labeled as "1"; gas discharge lamp load category (such as fluorescent lamp, high-pressure sodium lamp): contains ballast, current lags behind voltage, and there are 3rd and 5th harmonics, labeled as "2"; motor load category (such as asynchronous motor, synchronous motor): current surge is large during starting (about 5-7 times the rated current), and there is current fluctuation caused by slip during running, labeled as "3"; switching power supply load category (such as charging pile, electronic equipment): current waveform is not sinusoidal, contains high-frequency harmonics (such as more than 20 times), power factor is low, labeled as "4".

[0064] The labeling needs to be combined with the load account (such as the load type connected to a certain line being "residential area + industrial park, containing motor load and switching power supply load") and signal characteristics (such as current surge characteristics during motor starting) to ensure that the load category label of each signal is consistent with the actual situation.

[0065] Optionally, a double mechanism of artificial review combined with algorithm verification can also be used, wherein the verification process of artificial review can be: extracting 30% of the labeled data, and checking whether the label is accurate by a distribution network operation and maintenance expert (familiar with fault and load characteristics) (such as whether the motor starting surge signal is mislabeled as a fault); the process of algorithm verification can be: verifying by feature matching degree, for example, calculating the similarity between a signal labeled as a single-phase ground fault and a standard single-phase ground fault signal template (such as zero sequence voltage rise and zero sequence current increase), and if the similarity is less than 80%, re-labeling is triggered to ensure that the labeling accuracy is ≥98%.

[0066] Further, the labeled data can be divided into a training set, a validation set, and a test set in a ratio of 7:2:1, with clear division of functions. Specifically, the training set is used for iterative updating of model parameters to learn the mapping relationship between signal features and state labels; the validation set is used to adjust model hyperparameters (such as learning rate and batch size) and monitor whether the model is overfitting (such as increasing the training set accuracy but decreasing the validation set accuracy, which requires adjusting the regularization parameter); and the test set is used to simulate the actual application scenario of the model and evaluate the generalization ability of the final model (the test set data does not participate in model training and hyperparameter adjustment to ensure the objectivity of the evaluation results).

[0067] Hierarchical sampling can be used for division to ensure that the proportion of fault types and load categories in the training set, validation set, and test set is consistent with that of the original data (for example, if single-phase ground faults account for 60% of the total number of faults in the original data, then each subset must maintain this proportion to avoid missing data for a certain state in the training set.

[0068] Further, "new labeled samples" can also be generated through data transformation to expand the training set size and improve the model's adaptability to signal variations. Data transformation can include time domain enhancement, such as time shifting of the signal (e.g., shifting the transient peak position of the fault signal by 1-2 sampling points), signal stretching / compression (e.g., stretching a 120-sample-point signal to 130 sample points to simulate the difference between signals of different sampling frequencies); amplitude domain enhancement, such as amplitude fine-tuning of voltage and current signals (e.g., randomly adjusting the amplitude within ±5% to simulate voltage fluctuations in the distribution network), and adding noise disturbance (adding 20dB Gaussian white noise to simulate electromagnetic interference in actual operation); and feature domain enhancement, such as adjusting the frequency domain features of the signal (e.g., ±10% perturbation on the 3rd harmonic amplitude to simulate the harmonic difference of different loads). The enhanced samples should retain the original state labels (e.g., after amplitude fine-tuning of the signal of a single-phase ground fault + motor load, the label is still 011+3) to ensure that the mapping relationship between the signal and the label remains unchanged.

[0069] It should be noted that the initial model is not simply a combination of the two algorithms, but is improved for the time series and feature complexity of the distribution network signal, divided into two modules: a generator (Generator) and a discriminator (Discriminator) (based on the AC-GAN framework), and the discriminator is embedded with SwinTransformer for feature extraction. The specific structure is as follows: The function of the generator is to learn the feature distribution of normal / fault signals and generate realistic signal samples to assist the discriminator in improving feature discrimination. The input is random noise vector + state label (fault label + load label, concatenated into a vector), and the output is a generated signal consistent with the format of the real signal. The network structure uses fully connected layers + deconvolution layers + normalization layers. After the random noise and state label are concatenated, they are mapped to a high-dimensional feature vector through 3 fully connected layers (with 256, 512, and 1024 neurons, respectively). The high-dimensional vector is converted into a time series signal through 4 deconvolution layers (with a kernel size of 3x1 and a stride of 2) (from 1024 dimensions to 120 dimensions). After each deconvolution layer, BatchNorm2d normalization and LeakyReLU activation function (with a slope of 0.2) are added. Finally, the output signal amplitude is normalized to the range [-1, 1] through the Tanh activation function (consistent with the amplitude range of the preprocessed real signal).

[0070] The function of the discriminator is to distinguish whether the input signal is a real sample (signal in the training set) or a generated sample (signal generated by the generator), while completing fault state discrimination and load category discrimination (auxiliary classification function of AC-GAN), and embedding Swin Transformer to accurately extract time series features. The network structure is divided into three parts: feature extraction sub-module (improved Swin Transformer), real / generated discrimination sub-module, and state classification sub-module. Specifically, the power distribution network signal in the feature extraction sub-module (improved Swin Transformer) is time series data, which can convert a one-dimensional signal of 120 sampling points into a two-dimensional feature map of "1x120x1" (adapted to the input format of Swin Transformer). The improved Swin Transformer uses window attention (Window Attention) instead of traditional global attention, divides the 120 sampling points into 6 windows (20 sampling points per window), and reduces the computational complexity. Time series position encoding (instead of image spatial position encoding) is added, which generates an encoding vector related to the position of the sampling point through a sine function and concatenates it to the original signal features to help the model capture the time series dependence of the signal (such as the sequence of signal changes before and after the fault occurs). After passing through 4 Swin Transformer blocks (each containing window attention, cross-window attention, and Feed-Forward Network), the output is a deep time series feature map of "1x120x256", which is then compressed into a global feature vector of 256 dimensions through global average pooling (Global Average Pooling). The real / generation discriminator submodule can input a 256-dimensional global feature vector into a 2-layer fully connected layer (128 neurons, 1), and finally output a probability value of "0-1" through a Sigmoid activation function, where 0 represents a generated sample and 1 represents a real sample; the state classification submodule can be parallelly provided with two classification branches to process fault states and load categories, respectively. For example, the fault state branch includes a 256-dimensional feature, a fully connected layer (64 neurons), a Softmax activation function, and the like, and outputs a probability distribution of fault types (for example, 5 fault types are outputted, corresponding to "non-fault, three-phase short circuit, two-phase short circuit, single-phase ground fault, and broken line"); the load category branch includes a 256-dimensional feature, a fully connected layer (32 neurons), a Softmax activation function, and the like, and outputs a probability distribution of 4 load categories, and finally the category with the maximum probability is taken as the model prediction result.

[0071] Further, the model training parameters are initialized; before inputting the training data set, the initialization of the model parameters, the optimizer, and the loss function needs to be completed to ensure stable start of the training; specifically, the parameter initialization includes the weight parameters of the fully connected layer, the deconvolution layer, and the Swin Transformer block of the generator and the discriminator, which can adopt He normal initialization (suitable for ReLU / LeakyReLU activation function); and the bias parameter is initialized to 0; the moving average coefficient (momentum) of the BatchNorm layer is set to 0.9, and the weight decay is set to 1e-4 to prevent overfitting of parameters; the optimizer selects the Adam optimizer (which takes into account the convergence speed and stability); the learning rate of the generator is set to 2e-4, and the learning rate of the discriminator is set to 1e-4; the training speed of the discriminator needs to be slightly faster than that of the generator to avoid generating too poor samples by the generator; the momentum parameters β1 and β2 are set to 0.5 and 0.999, respectively; the batch size (BatchSize) can be set according to the size of the GPU memory, such as 128 (memory ≥ 16 GB) or 64 (memory 8 GB), to ensure that enough diverse samples are inputted for each training.

[0072] Further, the multi-objective loss function can be designed by combining the adversarial loss and the classification loss of the AC-GAN to ensure that the model optimizes the real / generation discrimination and the state classification ability at the same time. For example, the adversarial loss can be calculated by using the binary cross-entropy loss to calculate the prediction error of the discriminator for real samples / generated samples; the classification loss can be calculated by using the cross-entropy loss to calculate the prediction error of the discriminator for fault states and load categories; the load category classification loss is the cross-entropy between the load label of the real sample and the probability distribution outputted by the load branch of the discriminator; and the total classification loss of the discriminator can be determined according to the adversarial loss, the classification loss, and the load category classification loss.

[0073] The discriminator training can improve the ability of the discriminator to distinguish between real / generated samples and identify state labels; the generator training can improve the ability of the generator to generate realistic signals and match state labels; the implicit training of signal reconstruction capability includes learning to generate signals that match the state characteristics according to the state label during the training process of the generator. The essence is to realize signal reconstruction. For example, when the state label is single-phase ground fault + motor load, the generator needs to generate a signal with "zero sequence voltage rise, zero sequence current increase and motor current fluctuation characteristics". This ability is constrained by the classification loss: if the generated signal does not match the state label, the generator will adjust the parameters until the characteristics of the generated signal match the label, thereby indirectly achieving the training goal of "signal reconstruction".

[0074] The feature extraction capability is strengthened in iterations; in the iteration process, the Swin Transformer module in the discriminator needs to extract key features that distinguish between real / generated and match states from real samples and generated samples in each discriminator training. For example, extract fault features such as transient current spikes and zero sequence component mutations from fault samples, and extract load features such as harmonic distribution and power factor characteristics from load samples. With training iterations, the window attention mechanism of Swin Transformer will gradually optimize and more accurately capture the time-dependent and local features of the signal (such as signal mutations at the moment of fault occurrence), achieving continuous strengthening of feature extraction capability.

[0075] The state discrimination capability is gradually improved in iterations; in the iteration process, the state classification branch of the discriminator is iteratively optimized by the classification loss: for example, during initial training, the model may misjudge the motor start impact current as a short circuit fault, at which time the classification loss will increase, driving the classification branch to adjust the fully connected layer weights and gradually learn the difference between the short duration (0.1s) of the motor impact current and the long duration (≥0.5s) of the short circuit fault current. Eventually, the model can accurately distinguish between fault states and load categories.

[0076] It should be noted that during the training process of the model, the loss curve and the accuracy of the validation set need to be monitored in real time, and reasonable training termination conditions need to be set to avoid overfitting or insufficient training of the model. For example, real-time plotting of the total loss curves of the discriminator and the generator, if the curve fluctuation amplitude is ≤5% in continuous 20 iterations, it means that the loss function converges; or, every 10 training, input the validation set into the model, calculate the fault state discrimination accuracy and load category discrimination accuracy, such as fault state accuracy = (number of samples in the validation set whose fault state prediction is correct / total number of samples in the validation set) x 100%; load category accuracy = (number of samples in the validation set whose load category prediction is correct / total number of samples in the validation set) x 100%.

[0077] Further, the training can be terminated when any of the following conditions is met: The fault state accuracy of the verification set is greater than or equal to 95% and the load category accuracy is greater than or equal to 92% (the threshold is adjusted according to actual needs), and the accuracy does not obviously increase (fluctuation is less than or equal to 1%) for 10 consecutive rounds; The number of training rounds reaches a preset maximum value, such as 200 rounds, and the loss function has converged, avoiding infinite training; Overfitting signs appear, such as the training set accuracy continuously increases, but the verification set accuracy begins to decrease, at which time the training is terminated and the best model parameter before overfitting is loaded.

[0078] After the training is terminated, the best model parameter (the parameter file with the highest verification set accuracy) is loaded to obtain the final fault identification model, and the generalization ability of the model is evaluated through the test set. The parameters, network structure configuration, and preprocessing rules of the generator and the discriminator are saved to ensure that the model structure and parameters can be completely reproduced in subsequent deployment.

[0079] The above fault identification method provides a data basis for subsequent fault analysis by obtaining the first signal corresponding to the target power distribution network; further, the first signal is preprocessed by slicing, compressing, and sampling to obtain the second signal corresponding to the target power distribution network, which lays a foundation for obtaining an accurate fault identification result; then, the second signal can be input into the fault identification model improved based on the Swin Transformer and the auxiliary classification generative adversarial network to obtain the fault identification result corresponding to the target power distribution network; the present scheme provides data for accurate fault analysis by introducing the second signal; moreover, the fault identification model improved based on the Swin Transformer and the auxiliary classification generative adversarial network is introduced to provide a way for accurate signal fault analysis, and finally accurate fault identification is achieved.

[0080] On the basis of the above embodiment, the present embodiment explains and describes S202 in detail. Specifically, the process of obtaining the second signal in the present embodiment includes the following steps, as shown in Figure 3 S301, based on a sliding time window, slicing the first signal to obtain a first observation group.

[0081] The sliding time window is a technique of dividing data according to a fixed time length and making the window slide with time, which is used for segmenting and analyzing data streams; the slicing operation is a segmenting operation of the first signal using the sliding time window, which aims to split the continuous signal into multiple independent observation segments; and the first observation group is a set of multiple signal segments obtained by slicing the first signal using the sliding time window.

[0082] ​It should be noted that the concept of sliding time window can be compared to moving a photo frame with a fixed width along a continuously extending photo strip (data stream), observing only the photos (data) within the photo frame each time, thereby realizing real-time or batch analysis of data; the three important features of sliding time window are: Fixed window size: the time length of the window is pre-set, such as 1 minute, 5 minutes, which determines the data coverage range of each analysis; Continuous sliding advance: the window will move forward according to the set step (sliding interval), such as a window size of 1 minute and a step of 30 seconds, which means a new window will be generated every 30 seconds, and there will be data overlap between windows; Dynamic data range: as the window slides, the data contained in the window will be updated, the old data outside the window time range is discarded, and the new data entering the window time range is included.

[0083] For example, according to the sampling frequency of the first signal and the cycle of the arc fault feature, the length of the sliding time window and the overlap width can be set, for example, in the case of a sampling frequency of 500 kHz, the length of the sliding time window can be set to 200 sampling points, corresponding to a time length of 0.4 ms, and the overlap width can be set to half of the length of the sliding time window , i.e. 100 sampling points; further, the starting analysis position of the first signal is L p , and the ending analysis position is L q , which can be started from the starting analysis position L p , and the window is moved along the time axis, with a sliding step of , i.e. 100 sampling points, to divide the continuous first signal into N non-overlapping signal slices, each with a length of , all signal slices together constitute the first observation group; the number of non-overlapping signal slices N can be calculated as follows:

[0084] Further, the signal energy of each signal slice can be calculated, and if the energy value of a certain slice is less than 50% of the normal signal energy threshold, it is determined to be an invalid slice, which is removed and the window position is adjusted to complete, ensuring that all slices in the first observation group contain valid signal features.

[0085] S302, based on the preset compression ratio, performing compression calculation on the first observation group to obtain the second observation group.

[0086] The compression ratio can refer to the ratio of the length of the compressed signal to the length of the original slice signal (i.e., the first signal), and the value range can be 0.2-0.6, which can balance the data compression efficiency and signal feature retention rate, such as requiring the energy retention rate of the compressed signal to be not less than 80%; the second observation group is the final compressed data obtained by compressing and calculating the first observation group.

[0087] In an implementation manner, based on a preset compression ratio, the first observation group is compressed and calculated to obtain a compressed observation group; the first observation group is normalized to obtain a normalized observation group; and based on matrix multiplication operation, the compressed observation group and the normalized observation group are compressed and calculated to obtain the second observation group.

[0088] The compressed observation group is intermediate compressed data obtained by directly compressing and calculating the first observation group, and has not yet been combined with the normalization processing; the normalization processing is an operation of standardizing data of the first observation group, and usually maps the data to a fixed range to eliminate dimensional differences; the normalized observation group is a standardized data group obtained by normalizing the first observation group, and is used for subsequent matrix operation with the compressed observation group; the matrix multiplication operation can refer to a calculation manner of multiplying the compressed observation group and the normalized observation group by a matrix.

[0089] For example, the normalization processing can be performed on each signal slice in the first observation group to map all signal amplitudes in the slice to the interval [0, 1], and the calculation formula is: normalized amplitude=(original amplitude-minimum amplitude in the slice) / (maximum amplitude in the slice-minimum amplitude in the slice), that is, the normalized observation group is obtained to eliminate the influence of signal amplitude differences between different slices; further, a random sampling matrix with a dimension of m x Lw (m=δ x Lw) can be generated according to the preset compression ratio δ, and the matrix elements adopt Bernoulli distribution (value 0 or 1) or Gaussian distribution (mean 0, variance 1 / m) to ensure that the matrix sparsity is greater than or equal to 70% (the number of non-zero elements in each row is less than or equal to 0.3Lw), thereby reducing the subsequent calculation complexity.

[0090] Further, a block matrix multiplication manner can be used to perform matrix multiplication operation between the random sampling matrix and each signal slice in the normalized observation group to obtain a compressed signal (length m) corresponding to each slice, and all compressed signals constitute the compressed observation group; in the calculation process, each block matrix product is processed in parallel to adapt to the algorithmic limit of the edge device, and the compression process can be represented as:

[0091] wherein, y R m×1 is the signal value after compression; R m×Lw is the random sampling matrix; x R Lw For the processed first observation group, m < L w , and δ is a preset compression ratio, δ = m / L w .

[0092] The signal energy retention rate = (compressed signal energy / original slice signal energy) * 100% can also be calculated for each compressed signal in the compressed observation group. If the retention rate < 80%, a random sampling matrix is regenerated and the compression calculation is repeated until all compressed signals meet the energy retention requirement, and finally a second observation group is obtained.

[0093] In S303, based on random sampling, a second signal corresponding to the target power distribution network is obtained by sampling the second observation group.

[0094] It should be noted that since the length of the compressed signal is smaller, and the length of the compressed signal changes with different compression ratios, in order to ensure the consistency of the length of the input data, a placeholder is added to the unsampled position of the compressed signal to complete the length, and the placeholder can be set to 0.

[0095] In one implementation, the second observation group is zero-padded by filling the placeholder to obtain a complete observation group. The placeholders in the complete observation group are marked to obtain a mask sequence corresponding to the complete observation group. Based on the complete observation group and the mask sequence, a second signal corresponding to the target power distribution network is obtained.

[0096] In order to avoid the influence of the placeholder on the network model training, the placeholder of the compressed signal can be marked based on the mask sequence. For valid sampling points, the elements in the mask sequence are marked as 1, and for the placeholder, the elements in the mask sequence are marked as 0 to obtain a complete observation group. The mask sequence and the complete observation group together constitute the second signal.

[0097] Wherein, the placeholder refers to virtual data used to fill the data vacancy position of the second observation group, which only plays the role of completing the data length and unifying the data structure, and has no actual physical or business significance, which can be distinguished by the mask later; the zero padding processing is an operation for the length defect of the second observation group, which fills it with 0 (i.e. the placeholder) to make it a fixed length data that meets the subsequent processing requirements; the complete observation group is the complete data group obtained after the second observation group is zero-padded, including two parts: one is the original effective compressed data of the second observation group, and the other is the filled placeholder (0), and the data length meets the preset standard; the mask sequence is a marking sequence corresponding to the complete observation group, which is used to distinguish between valid data and placeholders, for example, the position corresponding to the original valid data in the complete observation group is marked as 1, and the position corresponding to the placeholder (0) is marked as 0, which is used to identify valid information and ignore meaningless placeholders later.

[0098] The workflow of the fault identification model includes super-resolution generator signal reconstruction, discriminator identification with auxiliary classification, and auxiliary feature extractor verification. The super-resolution generator signal reconstruction includes input reception, shallow feature extraction, and deep feature mining. The input reception receives the "completed signal + mask sequence" in the second signal, extracts the effective features in the completed signal through the effective sampling point position marked by the mask sequence, the shallow feature extraction uses convolution layers to perform convolution operations on the effective features, maps the signal from the original signal space to the high-dimensional feature space, and obtains the shallow features, the deep feature mining introduces Swin Transformer layers, divides the shallow features through non-overlapping windows (window size w) and translation windows (moving step w / 2), captures the feature correlation between windows using window multi-head self-attention (W-MSA) and sliding window multi-head self-attention (SW-MSA), combines multi-level jump connections to enhance feature transmission, and mines local and global multi-scale features), and signal reconstruction output (i.e., through sub-pixel convolution operation to complete upsampling, map the high-dimensional features back to the signal space, and then process through the convolution layer to output the reconstructed high-precision arc signal.

[0099] The discriminator identification with auxiliary classification includes input reception, feature extraction, true / false discrimination, and class identification. The input reception receives the reconstructed signal output by the generator, the original arc signal class information, the feature extraction extracts the deep features of the reconstructed signal through the deep convolution network, filters the redundant information, the true / false discrimination inputs the deep features into the full connection layer, activates through the Sigmoid function, and outputs the signal authenticity judgment result, and the class identification simultaneously inputs the deep features into the classification unit, activates through the Softmax function, and outputs the corresponding load fault class of the signal. The auxiliary feature extractor verification includes feature extraction and feature verification. The feature extraction uses low-rank matrix approximation to extract the latent features of the reconstructed signal and the original signal respectively, reduces the calculation cost through random singular value decomposition (SVD), and obtains the approximate SVD matrix as the latent feature; the feature verification calculates the similarity of the latent features of the reconstructed signal and the original signal, if the similarity is less than 90%, feedback to the generator to adjust the parameters, and ensure that the reconstructed signal restores the original signal information.

[0100] The fault identification result can include basic information integration, fault load class determination, fault signal feature and abnormal level analysis, risk level evaluation, and processing measure matching.

[0101] The basic information integration is that the line number, the signal collection time, the fault position (such as the L1 line of the No. 3 distribution box) and the fault time (accurate to the millisecond level) are determined in combination with the sensor collection; the fault / non-fault result is determined based on the fault state output by the discriminator; the fault load category determination is that the fault load subtype (such as the resistive load-heater overload fault) is refined according to the load category output by the discriminator and in combination with the fault characteristics of the load category (such as the sine wave of the resistive load fault and the triangular wave of the motor fault); the fault signal feature and abnormality level analysis is that the key parameters (such as the distortion rate, the harmonic peak frequency and the current ripple coefficient) of the reconstructed signal are extracted as the fault signal features; the abnormality level (mild: deviation 5%-10%; moderate: deviation 10%-20%; severe: deviation >20%) is divided according to the deviation of the features from the normal threshold; the risk level evaluation is that the risk level (low risk: no ignition / equipment damage risk; medium risk: slight equipment damage risk; high risk: flammable material ignition or equipment burning risk) is divided in combination with the abnormality level, the on-site environment (such as whether there are flammable materials) and the equipment damage risk; the treatment measure matching is that the preset treatment scheme (such as low risk: reduce the load power + continuous monitoring; high risk: emergency power-off + replace the faulty equipment) is matched based on the risk level and the fault type, and the complete fault identification result is formed. In the embodiment of the application, the first signal is preprocessed by introducing slicing, compression and random sampling, thereby ensuring the high quality of the second signal and laying a foundation for obtaining an accurate fault identification result.

[0102] The above Figures 1 to 3 The fault identification method provided in the embodiments of the application is described in detail, and the device and the equipment provided in the embodiments of the application will be described below with reference to the drawings.

[0103] As Figure 4 shown, the figure is a schematic diagram of a fault identification device provided in the embodiments of the application. The fault identification device 600 includes an acquisition module 601, a processing module 602 and an identification module 603, wherein: The acquisition module 601 is configured to acquire a first signal corresponding to a target power distribution network. The processing module 602 is configured to preprocess the first signal to obtain a second signal corresponding to the target power distribution network. The preprocessing includes slicing, compression and sampling. The identification module 603 is configured to input the second signal into a fault identification model to obtain a fault identification result corresponding to the target power distribution network. The fault identification model is a model improved based on a Swin Transformer and an auxiliary classification generative adversarial network.

[0104] In one embodiment, the processing module 602 is specifically configured to: Slice the first signal based on a sliding time window to obtain a first observation group; Perform compression calculation on the first observation group based on a preset compression ratio to obtain a second observation group; Sample the second observation group based on random sampling to obtain a second signal corresponding to the target power distribution network.

[0105] In an embodiment, the processing module 602 is specifically configured to: Perform compression calculation on the first observation group based on a preset compression ratio to obtain a compressed observation group; Perform normalization processing on the first observation group to obtain a normalized observation group; Perform compression calculation on the compressed observation group and the normalized observation group based on matrix multiplication operation to obtain the second observation group.

[0106] In an embodiment, the processing module 602 is specifically configured to: Perform zero padding processing on the second observation group by filling placeholders to obtain a complete observation group; Mark the placeholders in the complete observation group to obtain a mask sequence corresponding to the complete observation group; Obtain the second signal corresponding to the target power distribution network based on the complete observation group and the mask sequence.

[0107] In an embodiment, the fault identification apparatus 600 further includes: A training module configured to construct a training data set based on a training signal and a corresponding state label; Input the training data set into an initial model to perform training of feature extraction, signal reconstruction, and state discrimination, and obtain a fault identification model; wherein the initial model is a model improved and built based on SwinTransformer and auxiliary classification generative adversarial network.

[0108] In an embodiment, the identification module 603 is specifically configured to: Input the second signal into the fault identification model to perform signal reconstruction, fault category identification, and feature verification, and obtain a fault identification result corresponding to the target power distribution network.

[0109] In an embodiment, the fault identification result includes at least one of a risk level, a fault state, a fault time, a fault location, a fault load category, a fault signal feature, a fault signal anomaly level, and a processing measure.

[0110] The fault identification apparatus 600 according to the embodiments of the present application can correspond to performing the methods described in the embodiments of the present application, and the above-mentioned other operations and / or functions of each module / unit of the fault identification apparatus 600 are respectively implemented to achieve Figure 2 、 Figure 3The corresponding flow of each method in the illustrated embodiment will not be described here in detail for the sake of brevity.

[0111] The embodiments of the present application further provide a computing device. The computing device can be a local computing device or an application server.

[0112] As Figure 5 The figure is a schematic diagram of a computing device provided by the embodiments of the present application. The computing device 700 includes a bus 701, a processor 702, a communication interface 703 and a memory 704. The processor 702, the memory 704 and the communication interface 703 communicate through the bus 701.

[0113] The bus 701 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience, Figure 5 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0114] The processor 702 can be any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0115] The communication interface 703 is used for external communication. For example, the communication interface 703 can be used for communication with the terminal 102. The communication interface 703 is used to send the fault identification result to the terminal 102, so that the terminal 102 displays the fault identification result.

[0116] The memory 704 can include volatile memory, such as random access memory (RAM). The memory 704 can also include non-volatile memory, such as read-only memory (ROM), flash memory, a hard disk drive (HDD), or a solid state drive (SSD).

[0117] The executable code is stored in the memory 704, and the processor 702 executes the executable code to perform the foregoing fault identification method.

[0118] Specifically, in the case of the embodiment shown in the figure, and Figure 4 the foregoing fault identification method. Figure 4 In the case of implementing the modules or units of the fault identification apparatus described in the embodiment by software, the software or program code required for implementing the functions of the modules / units in the foregoing embodiment can be stored in the memory 704 partially or entirely. The processor 702 executes the program code corresponding to each unit stored in the memory 704 to perform the foregoing fault identification method. Figure 4 In the case of implementing the modules or units of the fault identification apparatus described in the embodiment by software, the software or program code required for implementing the functions of the modules / units in the foregoing embodiment can be stored in the memory 704 partially or entirely. The processor 702 executes the program code corresponding to each unit stored in the memory 704 to perform the foregoing fault identification method.

[0119] The embodiment of the present application further provides a computer readable storage medium. The computer readable storage medium can be any available medium or data storage device that can be accessed by a computing device, such as a data center. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk), etc. The computer readable storage medium includes instructions indicating the computing device to perform the foregoing fault identification method.

[0120] The embodiment of the present application further provides a computer program product including one or more computer instructions. When the computer instructions are loaded and executed on a computing device, the foregoing processes or functions according to the embodiment of the present application are generated entirely or partially.

[0121] The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer or data center to another website, computer or data center through a wired (such as a coaxial cable, an optical fiber, a digital subscriber line) or wireless (such as infrared, wireless, microwave, etc.) manner.

[0122] The computer program product is executed by a computer, and the computer executes any of the foregoing fault identification methods. The computer program product can be a software installation package, and when any of the foregoing fault identification methods needs to be used, the computer program product can be downloaded and executed on the computer.

[0123] The foregoing description of the processes or structures corresponding to each figure has its own emphasis, and the parts not described in detail in a certain process or structure can be referred to the related description of other processes or structures.

[0124] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this, any change or replacement within the technical scope disclosed by the present application should be covered in the protection scope of the present application.

Claims

1. A fault identification method, characterized in that, The method includes: Obtain the first signal corresponding to the target distribution network; The first signal is preprocessed to obtain the second signal corresponding to the target distribution network; wherein, the preprocessing includes slicing, compression and sampling; The second signal is input into the fault identification model to obtain the fault identification result corresponding to the target distribution network; wherein, the fault identification model is an improved model based on Swin Transformer and auxiliary classification generative adversarial network.

2. The method according to claim 1, wherein preprocessing the first signal to obtain the second signal corresponding to the target distribution network comprises: Based on a sliding time window, the first signal is sliced ​​to obtain the first observation group; Based on a preset compression ratio, the first observation group is compressed to obtain the second observation group; Based on random sampling, the second observation group is sampled to obtain the second signal corresponding to the target distribution network.

3. The method according to claim 2, wherein the step of performing compression calculations on the first observation group based on a preset compression ratio to obtain a second observation group includes: Based on a preset compression ratio, the first observation group is compressed to obtain a compressed observation group; The first observation group is normalized to obtain the normalized observation group; Based on matrix multiplication, the compressed observation group and the normalized observation group are compressed to obtain the second observation group.

4. The method according to claim 2, wherein sampling the second observation group based on random sampling to obtain the second signal corresponding to the target distribution network includes: By filling in placeholders, the second observation group is padded with zeros to obtain the complete observation group; The placeholders in the completed observation group are marked to obtain the mask sequence corresponding to the completed observation group; Based on the completed observation group and mask sequence, the second signal corresponding to the target distribution network is obtained.

5. The method according to claim 1, wherein the training process of the fault identification model includes: A training dataset is constructed based on the training signals and their corresponding state labels; The training dataset is input into the initial model for feature extraction, signal reconstruction, and state discrimination training to obtain the fault identification model; wherein, the initial model is a model built based on the improved SwinTransformer and auxiliary classification generative adversarial network.

6. The method according to claim 1, wherein inputting the second signal to the fault identification model to obtain the fault identification result corresponding to the target distribution network includes: The second signal is input into the fault identification model for signal reconstruction, fault category identification, and feature verification to obtain the fault identification result corresponding to the target distribution network.

7. The method according to claim 1, wherein the fault identification result includes at least one of risk level, fault state, fault time, fault location, fault load category, fault signal characteristics, fault signal anomaly level, and handling measures.

8. A fault identification device, characterized in that, The device includes: The acquisition module is used to acquire the first signal corresponding to the target distribution network; A processing module is used to preprocess the first signal to obtain a second signal corresponding to the target distribution network; wherein the preprocessing includes slicing, compression and sampling; The identification module is used to input the second signal into the fault identification model to obtain the fault identification result corresponding to the target distribution network; wherein, the fault identification model is a model improved based on Swing Transformer and auxiliary classification generative adversarial network.

9. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 7.

Images