Power system fault classification method, device, equipment and program product

Abstract

The invention is suitable for the technical field of power systems, and provides a power system fault classification method, device and equipment and a program product, and the method comprises the steps: obtaining electrical quantity time sequence data collected by power system monitoring equipment; performing multi-level signal decomposition on the electrical quantity time sequence data to obtain a plurality of signal components; extracting statistical features of the plurality of signal components to construct feature vectors to be classified; and inputting the feature vectors to be classified into a breadth feature mapping classification model trained in advance, and outputting a fault classification result, so that complex fault types can be effectively distinguished, and the classification accuracy and fineness are improved.

Description

Technical Field

[0001] This application belongs to the field of power system technology, and in particular relates to power system fault classification methods, devices, equipment and program products. Background Technology

[0002] In power systems, fault diagnosis is a crucial link in ensuring the safe and stable operation of the power grid. Accurate fault diagnosis relies on the acquisition and analysis of high-precision, high-frequency waveform data of electrical quantities such as voltage and current when a power grid fault occurs. Existing fault diagnosis methods typically involve acquiring electrical quantity data, decomposing the signal into multiple components using signal processing methods, extracting feature parameters from each component, and finally using a classifier for fault identification. However, most of these methods rely on setting fixed thresholds for single indicators such as current surges and voltage dips, or on manually set logical rules for judgment. When the classifier structure is relatively simple or the rules are not well-defined, the fixed thresholds are difficult to adjust when facing complex operating conditions caused by the integration of new energy sources and the increase in power electronic equipment, especially for high-resistance grounding faults or when the signal is subjected to strong noise interference, easily leading to misjudgments or missed judgments. Furthermore, these methods usually only perform simple fault judgments, have poor classification capabilities for specific fault types, weak generalization ability, and are difficult to adapt to the ever-changing power grid operating modes. Therefore, how to quickly, accurately, and automatically diagnose and classify power grid fault states is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0003] This application provides a power system fault classification method, apparatus, equipment, and program product, aiming to solve the technical problems in the prior art that, due to reliance on fixed thresholds and manual rules, fault classification accuracy is low, generalization ability is weak, and it is easy to misclassify or miss faults when facing complex operating conditions.

[0004] In a first aspect, embodiments of this application provide a power system fault classification method, including: Acquire time-series electrical quantity data collected by power system monitoring equipment; The electrical quantity time-series data is subjected to multi-level signal decomposition to obtain multiple signal components; Statistical features of the multiple signal components are extracted to construct a feature vector to be classified; The feature vector to be classified is input into a pre-trained breadth feature mapping classification model, and the fault classification result is output.

[0005] In one possible implementation of the first aspect, the multi-level signal decomposition of the electrical quantity time-series data is further included before: The acquired electrical quantity time-series data are subjected to data dimensionality reduction processing.

[0006] In one possible implementation of the first aspect, the multi-level signal decomposition of the electrical quantity time-series data to obtain multiple signal components includes: The electrical quantity time series data is iteratively decomposed using a decomposition kernel function, generating a compressed feature variable and a redundant information component in each decomposition level. The compressed feature variables of the previous layer are used as the input of the next layer decomposition. After multiple iterations, the compressed feature variables obtained in the last iteration and the redundant information components generated in each iteration are used as the multiple signal components.

[0007] In one possible implementation of the first aspect, the statistical features include at least two of energy, variance, kurtosis, and information entropy.

[0008] In one possible implementation of the first aspect, the broad feature mapping classification model maps the feature vector to be classified into high-dimensional combined features through a broad feature mapping layer, and the decision layer performs classification based on the high-dimensional combined features.

[0009] In one possible implementation of the first aspect, the training process of the breadth feature mapping classification model is as follows: The sample feature vectors labeled with fault type tags are used as the training set and input into the breadth feature mapping layer to obtain the feature mapping output Zr and the breadth enhancement output Hr; The feature mapping output Zr is concatenated with the breadth enhancement output Hr to obtain the combined feature matrix Ar; The combined feature matrix Ar is input into the decision layer to obtain the decision output Dr; Based on the decision output Dr and the fault type label Yr corresponding to the training set, the output weight Wo is calculated to complete the training of the breadth feature mapping classification model.

[0010] In one possible implementation of the first aspect, the step of inputting the feature vector to be classified into a pre-trained wide-range feature mapping classification model and outputting a fault classification result includes: The feature vector to be classified is input into the breadth feature mapping layer of the breadth feature mapping classification model to obtain the combined feature matrix Ae; The fault classification result is output based on the output weight Wo and the combined feature matrix Ae.

[0011] Secondly, embodiments of this application provide a power system fault classification device, comprising: The acquisition module is used to acquire time-series electrical quantity data collected by power system monitoring equipment; The decomposition module is used to perform multi-level signal decomposition on the electrical quantity time-series data to obtain multiple signal components. The feature extraction module is used to extract statistical features of the multiple signal components to construct a feature vector to be classified. The fault classification module is used to input the feature vector to be classified into a pre-trained breadth feature mapping classification model and output the fault classification result.

[0012] Thirdly, embodiments of this application provide a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any of the first aspects.

[0013] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method as described in any of the first aspects.

[0014] Fifthly, embodiments of this application provide a computer program product that, when run on a computer device, causes the computer device to perform the method described in any of the first aspects above.

[0015] The beneficial effects of the embodiments in this application compared with the prior art are: By performing multi-level signal decomposition and extracting statistical features from electrical quantity time-series data, transient information in fault waveforms can be deeply mined. Combined with the powerful nonlinear learning capability of the extensive feature mapping classification model, it can effectively distinguish complex and similar fault types such as high-resistance grounding and short circuits in different phases, significantly improving the accuracy and precision of fault classification and overcoming the shortcomings of traditional methods that are prone to misjudgment and omission under complex operating conditions. Furthermore, the extensive feature mapping model used in this application has the advantages of fast training speed and low prediction overhead, enabling the entire diagnostic process to quickly complete feature extraction and model calculation after a fault occurs, achieving rapid and automated fault classification and greatly shortening fault handling time.

[0016] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a power system fault classification method based on extensive feature mapping, provided for an embodiment of this application; Figure 2 A schematic diagram of the structure of a power system fault classification system based on extensive feature mapping provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0023] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0025] Example 1 This application provides a method for classifying power system faults. Please refer to [link / reference]. Figure 1 This document illustrates the overall flow of a power system fault classification method according to an embodiment of this application. The method aims to achieve accurate, rapid, and automated classification of fault types through in-depth analysis of electrical signals in the power system, thereby shortening fault handling time and improving the reliability of power grid operation.

[0026] Specifically, the method may include the following steps: S101, acquire time-series electrical quantity data collected by power system monitoring equipment.

[0027] As the data input for the entire fault classification process, this step aims to capture raw data from the power system that accurately reflects the dynamic behavior of the system under fault disturbances.

[0028] In one specific implementation of this application, the data originates from power system monitoring equipment installed at critical locations in substations, switching stations, or transmission lines. This power system monitoring equipment may include, but is not limited to, digital fault recorders (DFRs), protection and control devices, phasor measurement units (PMUs), or other online monitoring units with high-frequency data acquisition capabilities.

[0029] The acquired electrical quantity time-series data represents a series of digitally sampled values ​​arranged in chronological order, characterizing the dynamic behavior of the power system before and after a fault. These electrical quantities may include, but are not limited to, the instantaneous values ​​of three-phase voltages (such as Ua, Ub, Uc) and three-phase currents (such as Ia, Ib, Ic). In some application scenarios, they may also include zero-sequence voltage or zero-sequence current.

[0030] To ensure complete capture of the transient process at the time of a fault, especially critical information such as traveling waves and high-frequency oscillations in the initial stage, the data acquisition frequency is typically set high. For example, sampling frequencies of 10kHz, 20kHz, or even higher can be used, far exceeding the sampling rates of conventional measurements. Correspondingly, the length of the acquired data window should be reasonably designed to cover a short period before the fault (e.g., 1-2 power frequency cycles) and a longer period after the fault (e.g., 3-5 power frequency cycles) to form a complete fault waveform record. For example, a data window with a total duration of 40 milliseconds can be set, including 10 milliseconds before the fault (serving as a steady-state reference baseline for normal system operation) and 30 milliseconds after the fault (containing the complete transient process and initial steady-state information after the fault). At a 10kHz sampling rate, a 40-millisecond data window would contain 400 sampling points. For three-phase voltage and three-phase current, a total of 6 channels of data need to be processed, with 400 points per channel. This data is typically stored and transmitted in standard file formats such as the Transient Data Exchange Common Format to ensure interoperability between equipment from different manufacturers.

[0031] Optionally, after acquiring the time-series electrical quantity data collected by power system monitoring equipment, data preprocessing can be performed. Although the raw high-frequency sampling data is complete, its massive volume would lead to huge computational and storage burdens if directly used for subsequent complex signal processing, thus affecting the real-time requirements of fault diagnosis. Therefore, before formal signal decomposition, an optional data preprocessing step can be performed, which aims to reduce or normalize the data while preserving key waveform features as much as possible.

[0032] As an optional implementation, a key point sampling algorithm is used to perform data dimensionality reduction processing on the acquired electrical quantity time series data, so as to reduce the data length of the electrical quantity time series data while retaining the key waveform morphology features of the electrical quantity time series data.

[0033] This keypoint sampling algorithm is not a simple, evenly spaced sampling (i.e., downsampling), as the latter may miss crucial instantaneous abrupt changes in the waveform. Instead, it intelligently selects sampling points by analyzing the local morphological features of the waveform. Specifically, it prioritizes retaining local maxima (peaks), local minima (valleys), zero-crossing points, and points with significant curvature changes (inflection points), as these points often contain the most important information about the signal amplitude, frequency, and phase. In this way, a data sequence with, for example, 1000 sampling points can be reduced to a sequence of length 256. This reduces the computational load in subsequent steps and avoids the loss of crucial impact and oscillation information during fault transients due to reduced data volume, thus achieving a good balance between computational efficiency and information fidelity.

[0034] S102, perform multi-level signal decomposition on the electrical quantity time series data to obtain multiple signal components.

[0035] The original fault signal is a complex, non-stationary hybrid signal, containing not only the power frequency fundamental component but also various high-frequency transient components, DC offset components, and aperiodic components caused by the fault. Direct analysis of such a hybrid signal is quite difficult. Therefore, this step aims to decompose the complex original signal into a series of relatively pure signal components at different scales or frequency bands.

[0036] In one specific implementation of this application, a multi-level signal compression technique is employed. This technique iteratively peels away the detailed information of the signal layer by layer. Specifically, it includes: using a decomposition kernel function to iteratively decompose the electrical quantity time-series data, generating a compressed feature variable and a redundant information component in each decomposition layer; using the compressed feature variable of the previous layer as the input of the next layer, and after multiple iterations, using the compressed feature variable obtained in the last iteration and the redundant information components generated in each iteration as the multiple signal components.

[0037] In the embodiments of this application, the multi-level signal compression technique uses a pair of orthogonal decomposition kernel functions. One kernel function (called the compressed feature variable extraction kernel ωL) acts like a low-pass filter to extract the main components or low-frequency basis of the signal; while the other kernel function (called the redundancy information extraction kernel ωH) acts like a high-pass filter to extract the detailed information or high-frequency redundancy of the signal.

[0038] The multi-level signal compression technique is achieved through iterative decomposition, and its core calculation process is as follows: The data sampled from the keypoints is used as the input signal sequence x. In each decomposition layer, two types of convolution calculations are performed simultaneously: 1) Calculate the compressed feature variable Bbase: Perform convolution and downsampling operations on the input signal x using the low-pass decomposition kernel function ωL. This process is defined by the following formula: Bbase[n]=∑ k (x[k]·ωL[2n-k]) This step aims to extract the main contours and low-frequency approximation information of the signal.

[0039] 2) Calculate the redundant information component Bdetail: Perform convolution and downsampling operations on the same input signal x using the high-pass decomposition kernel function ωH. This process is defined by the following formula: Bdetail[n]=∑ k (x[k]·ωH[2n-k]) This step aims to capture detailed changes in the signal and high-frequency transient information.

[0040] In the above formula, the parameters are defined as follows: Bbase[n] and Bdetail[n] represent the nth compressed feature variable and redundant information component obtained after the decomposition of this layer, respectively; x[k] represents the kth sampling point value of the input signal of the current layer; ωL and ωH are a pair of complementary (usually orthogonal) decomposition kernel functions. n is the sample point index of the decomposed sequence. k is the index of the convolution summation operation.

[0041] The decomposition process is iterative. The compressed feature variable Bbase calculated in this layer is used as the input signal for the next layer, and the above two convolution calculations are repeated. After a set number of iterations (e.g., 7 times), the compressed feature variable of the final layer and the redundant information components generated by all layers are used as the multiple signal components to prepare for subsequent feature extraction.

[0042] To facilitate understanding, an example is provided here. In the first iteration, a preprocessed time-series data x of length 256 is used as input and processed by the two kernel functions mentioned above, resulting in a compressed feature variable Bbase1 and a redundant information component Bdetail1. Bdetail1 represents the highest frequency component in the signal, typically corresponding to the steep impact or traveling wave front at the moment of the fault. Subsequently, in the second iteration, the compressed feature variable Bbase1 obtained in the first iteration is used as a new input, and the same decomposition kernel function is used again to decompose it, resulting in a new compressed feature variable Bbase2 and a new redundant information component Bdetail2. This Bdetail2 represents the second highest frequency component in the signal, possibly corresponding to high-frequency oscillations after the fault.

[0043] This process can be repeated. For example, if we set it to perform 7 iterations of decomposition, in each iteration, the compressed feature variable obtained from the previous layer is used as the input of the current layer, and a new redundant information component is generated. After 7 iterations, we will obtain 7 redundant information components (denoted as Bdetail1, Bdetail2, ..., Bdetail7) and the final compressed feature variable Bbase7 obtained in the last iteration (the 7th iteration).

[0044] Understandably, these eight signal components (Bdetail1 to Bdetail7, and Bbase7) together constitute a multi-scale representation of the original signal. Physically, Bdetail1, Bdetail2, ..., Bdetail7 approximately correspond to the transient disturbance information of the signal across various frequency bands from high to low frequency, while Bbase7 represents the most dominant low-frequency component, close to the power frequency steady-state component. Through this decomposition, the rich transient information contained in the original fault signal is effectively separated into different signal components, laying a solid foundation for subsequent accurate feature extraction.

[0045] S103, extract the statistical features of the multiple signal components to construct the feature vector to be classified.

[0046] In this embodiment, although eight signal components capable of representing information in different frequency bands are obtained after step S102, these components are still time-series data with varying dimensions, making them unsuitable for direct use as input to a classifier. Therefore, this step aims to extract numerical indicators (i.e., statistical features) that can quantify the characteristics of each signal component and combine these features into a fixed-dimensional feature vector as a compact and effective digital fingerprint of the original fault waveform. The statistical features include at least two of energy, variance, kurtosis, and information entropy.

[0047] In one embodiment of this application, energy and variance are used as examples. For each of the eight signal components mentioned above, their energy and variance are calculated as statistical characteristics.

[0048] The formula for calculating energy characteristics is: ,in, is the value of the i-th sampling point in the signal component, and N is the length of the signal component. In the context of power system fault analysis, the energy of a signal component directly reflects the intensity or severity of the disturbance in the corresponding frequency band of the original signal. For example, a sudden increase in the energy of a high-frequency component (such as Bdetail1) usually means that a severe high-frequency transient impact occurred when the fault occurred, which is particularly crucial for distinguishing between metallic short circuits and high-resistance grounding faults.

[0049] The formula for calculating variance characteristics is: ,in, This is the mean of the signal component. Variance measures the dispersion of signal data points relative to their mean, and in fault analysis, it reflects the severity or instability of signal fluctuations in the corresponding frequency band. A high variance value may indicate persistent oscillations or irregular fluctuations in that frequency band.

[0050] For each phase electrical quantity (e.g., analyzing only current) of phases A, B, and C, the above steps S102 and S103 are performed. Taking the analysis of only phase A current as an example, after decomposition, eight signal components are obtained. The energy and variance characteristics are calculated for each component, resulting in 8 × 2 = 16 characteristic values. These 16 characteristic values ​​are arranged in a predetermined order (e.g., [E1, V1, E2, V2, ..., E8, V8], where Ei and Vi represent the energy and variance of the i-th component, respectively), forming a 16-dimensional characteristic vector F. This characteristic vector F is the final numerical description of the current phase A current fault waveform. If all three phase currents are analyzed simultaneously, the characteristic vectors of the three phases can be concatenated to form a 48-dimensional characteristic vector. This high-dimensional vector can more comprehensively reflect the phase-to-phase relationship of the fault. Before inputting the data into the classifier, the feature vectors can be normalized (e.g., max-min normalization or Z-score standardization) to eliminate the dimensional differences between different features and improve the stability and convergence speed of model training.

[0051] S104, input the feature vector to be classified into the pre-trained breadth feature mapping classification model, and output the fault classification result.

[0052] Optionally, the breadth feature mapping classification model maps the feature vector to be classified into high-dimensional combined features through a breadth feature mapping layer, and the decision layer performs classification based on the high-dimensional combined features.

[0053] In this embodiment, the breadth feature mapping classification model employed is a shallow neural network structure, which enhances its learning ability through horizontal expansion rather than vertical deepening. Unlike traditional deep neural networks that require layer-by-layer abstraction and time-consuming backpropagation for iterative optimization, this model boasts advantages such as extremely fast training speed and good generalization performance. The model consists of two core stages: an offline training stage and an online prediction stage.

[0054] First, the breadth-based feature mapping classification model needs to be trained offline. The offline training process is as follows: Step a1, Data Preparation: Collect a large amount of historical fault data. This data should be generated through field waveform recording or power system simulation software and include various known fault types, such as normal operation, single-phase grounding (AG, BG, CG), phase-to-phase short circuit (AB, BC, CA), two-phase grounding short circuit (e.g., ABG), three-phase short circuit (ABC), and high-resistance grounding faults under different transition resistances. For each historical fault waveform data, process it according to the aforementioned S101 to S103 to extract its feature vector. At the same time, according to the fault analysis report or simulation settings, assign a corresponding fault type label to each feature vector (e.g., using one-hot encoding, [1,0,0...] represents AG, [0,1,0...] represents BG, etc.), thus forming a training set containing a large number of feature vectors (training input Xr) and corresponding labels (training labels Yr). A rich and balanced training set is a guarantee of model performance.

[0055] Step a2, Model Construction and Solution: Input all feature vectors Xr from the training set into the model. Internally, the model first uses a broad feature mapping layer to non-linearly transform the input low-dimensional feature vector Xr (e.g., through a sigmoid or ReLU activation function) into a set of wider-dimensional feature mapping nodes using randomly generated weights and biases. The output of this layer is denoted as the feature mapping output Zr. Then, in the broad feature mapping layer, a non-linear transformation is performed on the feature mapping output Zr to obtain a set of augmentation nodes. The output of this layer is denoted as the broad augmentation output Hr.

[0056] The feature map output Zr and the breadth augmentation output Hr are calculated using the following formulas: Zr = φ(Xr·Wz + βz); Hr = ξ(Zr·Wh + βh). Where φ() is the feature mapping function, ξ() is the breadth augmentation function, Wz and βz are the weights and biases of the feature mapping, respectively, and Wh and βh are the weights and biases of the breadth augmentation, respectively.

[0057] Step a3: Subsequently, the feature map output Zr and the breadth-enhanced output Hr are concatenated horizontally to form a wider combined feature matrix Ar, i.e., Ar = [Zr|Hr]. This combined feature matrix Ar can be regarded as a rich representation of the original input features in a higher-dimensional random subspace, which makes the nonlinear relationships hidden in the original features explicit.

[0058] Among them, concatenating Zr and Hr yields the combined feature matrix Ar: Ar=[Zr|Hr].

[0059] Step a4: The ultimate goal of model training is to find an output weight matrix Wo such that the result of multiplying the combined feature matrix Ar by this weight matrix approximates the true training label Yr as closely as possible, i.e., Ar*Wo≈Yr. A characteristic of this model is that the output weights Wo can be directly obtained through a single analytical calculation, without the need for repeated iterative optimization.

[0060] Ar is input into the decision layer, and the decision output Dr is obtained through the decision function g(): Dr = g(Ar·Wd + βd), where Wd and βd are the weights and biases of the decision layer, respectively.

[0061] Based on the decision output Dr and the training set labels Yr, the output weight Wo is calculated using the following formula: Wo=(Dr T ·Dr+λI) -1 Dr T ·Yr; where Wd and βd are the weights and biases of the decision layer, respectively, and λ is a regularization parameter used to prevent overfitting and ensure numerical stability; its value is typically in the range of 10. -8 Up to 10 - ²; I is the identity matrix; Yr is the fault type label of the training set.

[0062] Among them, Wz, βz, Wh, βh, Wd, βd and Wo are all fixed parameters stored during the training phase.

[0063] Step a5: After training is complete, what needs to be saved is not the entire training dataset, but only the core parameters of the model, including the fixed random weights and biases used to generate Zr and Hr, and the calculated output weights Wo. Since the training process does not involve iteration, its training speed far exceeds that of deep learning models for large-scale datasets.

[0064] Once the model is trained, it can be deployed online for real-time fault classification. The online prediction phase is as follows: Step b1: When a new unknown fault occurs, execute S101 to S103 in real time to extract the feature vector of the fault waveform as the input Xe to be classified.

[0065] Step b2: The input Xe to be classified is fed into the trained model. The model uses the previously saved random weights and biases to perform the same feature mapping and enhancement transformation on Xe as in the training phase, obtaining the feature mapping output Ze and the breadth enhancement output He corresponding to the sample.

[0066] Specifically, the feature mapping output is calculated as: Ze = φ(Xe·Wz + βz); the breadth augmentation output is calculated as: He = ξ(Ze·Wh + βh); where Xe is the input feature vector of the test set; Ze and He are the feature mapping output and breadth augmentation output, respectively.

[0067] Step b3: Concatenate Ze and He to obtain the combined feature matrix Ae=[Ze|He] of the sample.

[0068] Step b4: Using the output weights Wo obtained during the training phase, perform matrix multiplication forward computation. Specifically, the combined feature matrix Ae is input into the decision layer to obtain the decision output De, De = g(Ae·Wd + βd); then, the decision output De is multiplied by the output weights Wo to obtain the fault classification prediction result Ypred: Ypred = De·Wo. Ypred is the final fault classification prediction result.

[0069] Optionally, the calculated result Ypred is a vector, where each element corresponds to the model's prediction confidence for each fault type. Typically, the index corresponding to the element with the largest value (or converted to a probability distribution using the Softmax function) is selected as the final classification decision.

[0070] After obtaining Ypred and determining the category index with the highest confidence, the index is converted into a human-readable fault type name, such as A-phase high-resistance ground fault or BC-phase metallic short circuit. This result can then be recorded in the event log, displayed on the human-machine interface of the monitoring system, or sent to the upper-level dispatch center via the communication network, thereby completing the entire automated fault classification process.

[0071] Through the above steps, the method provided in this embodiment can deeply mine the multi-scale transient features of fault signals and utilize the fast and efficient nonlinear classification capability of the extensive feature mapping model to achieve high-precision identification of faults in complex power systems. It is particularly suitable for online monitoring scenarios with high real-time requirements.

[0072] Optionally, each fault diagnosis event (including fault time, line name, file path of raw waveform data, extracted feature vector, fault classification result, etc.) can be persistently stored for subsequent fault tracing, statistical analysis, and model iteration optimization. It can also include an alarm service that immediately generates a clear and easily understood alarm message when a fault is detected. This message can be sent in various ways, such as by popping up an alarm window with audio-visual prompts on the local monitoring master station's human-machine interface, sending an alarm trap to the superior power grid dispatch automation system (EMS) via Simple Network Management Protocol (SNMP), or pushing the alarm message to the mobile terminal (such as a mobile app) of maintenance personnel via Internet of Things (IoT) communication protocols.

[0073] Example 2 This application aims to illustrate that the technical solutions for implementing multi-level signal decomposition are not unique. The core idea of ​​this application—the framework of multi-scale decomposition + feature extraction + broad-map classification—is compatible with various signal processing techniques. This embodiment provides a variant, the main difference from Embodiment 1 being the specific implementation of S102.

[0074] In this embodiment, in S102, a mature and widely used technique in the field of signal processing—Discrete Wavelet Transform (DWT)—is used to replace the multi-level signal compression technique in Embodiment 1. Due to its excellent time-frequency localization characteristics, Discrete Wavelet Transform is particularly adept at analyzing non-stationary signals containing transients.

[0075] In practical implementation, the first step is to select a suitable wavelet basis function. The choice of wavelet basis function has a certain impact on the decomposition effect. Different basis functions have different characteristics (such as support length, symmetry, vanishing moment, etc.) and are suitable for analyzing signals with different characteristics. In this embodiment, Daubechies series wavelets (such as db4, db6) or Symlets wavelets with tight support and certain regularity can be selected, as they perform well in detecting signal abrupt changes.

[0076] After selecting the wavelet basis, the preprocessed time series data is decomposed into multiple levels. Taking a 7-level decomposition as an example, the process is as follows: The first level of decomposition involves simultaneously passing the input signal through a high-pass filter composed of selected wavelet bases and a low-pass filter. The output of the high-pass filter is a set of high-frequency detail coefficients, denoted as cD1, which captures the most rapidly changing components of the signal (such as the traveling wave front). The output of the low-pass filter is a set of low-frequency approximation coefficients, denoted as cA1, which represents the smooth profile of the signal.

[0077] Second-level decomposition: The low-frequency approximation coefficient cA1 obtained in the first level is used as a new input signal and passed through the same high-pass and low-pass filters again to obtain the high-frequency detail coefficient cD2 and the low-frequency approximation coefficient cA2 of the second level.

[0078] Iterative process: Repeat the above process, decomposing the approximation coefficients of the previous layer at each layer. After 7 layers of decomposition, 7 sets of detail coefficients (cD1, cD2, ..., cD7) and the approximation coefficient cA7 of the last layer (the 7th layer) will be obtained.

[0079] Understandably, these eight sets of coefficients (cD1 to cD7, and cA7) constitute a multi-resolution representation of the original signal, functionally equivalent to the eight signal components obtained through multi-level signal compression technology in Example 1. cD1 corresponds to the highest frequency band, cD2 to the second highest frequency band, and so on, while cA7 represents the lowest frequency component in the signal, primarily containing power frequency components. Each set of coefficients reflects the characteristics of the original signal within a specific time-frequency window.

[0080] In the subsequent S103 (extracting statistical features), the processing method is the same as in Example 1. That is, these 8 sets of wavelet coefficients (cD1,...,cD7,cA7) are regarded as 8 independent signal components, and the energy and variance of each component are calculated separately, thus constructing a 16-dimensional feature vector.

[0081] The subsequent S104 (model classification) is the same as in Example 1. That is, the 16-dimensional feature vector obtained based on wavelet decomposition is input into the pre-trained broad feature map classification model to perform classification decisions and output the results. It is important to emphasize that the training set of the broad feature map model used in this example should also be constructed using discrete wavelet transform techniques to extract features, ensuring consistency in feature extraction methods between the training and prediction stages.

[0082] This embodiment, by employing the mature alternative technique of discrete wavelet transform, also achieves multi-scale decomposition and feature extraction of fault signals, demonstrating the universality and flexibility of the technical solution of this application.

[0083] Example 3 This embodiment aims to illustrate that the statistical features used to describe the characteristics of signal components are not limited to energy and variance. By extracting richer and more diverse feature dimensions, the intrinsic properties of fault signals can be characterized more comprehensively and profoundly, which may further improve the performance of classification models, especially for difficult fault types with weak feature differences. This embodiment extends S103 based on Embodiment 1.

[0084] In this embodiment, S101 (data acquisition) and S102 (multi-level signal decomposition) are the same as in Embodiment 1. That is, using the same multi-level signal compression technique, eight signal components (Bdetail1 to Bdetail7, and Bbase7) are obtained.

[0085] The core difference in this embodiment lies in S103, namely the step of extracting statistical features. In Embodiment 1, only energy and variance were calculated for each signal component. In this embodiment, four different statistical features are calculated for each signal component: energy, variance, kurtosis, and information entropy, to construct a more informative feature space.

[0086] The definitions and physical meanings of energy and variance have been explained in Example 1 and will not be repeated here. The two newly added features are described in detail below: Kurtosis: Kurtosis is a statistical measure of the sharpness or thickness of the tail of a data distribution. Its calculation formula can be expressed as: Where E[·] represents expectation, These are the sampled values ​​in the signal components. 'Var' is the mean, and 'Var' is the variance. Kurtosis reflects the distribution of extreme values ​​in a signal waveform. For a standard normal distribution, the kurtosis value is 3. If the kurtosis value is much greater than 3, it indicates that the signal waveform is sharper than a normal distribution, with more extreme values ​​or heavy tails, which is often associated with impulsive events in signal processing. Therefore, kurtosis characteristics are sensitive to detecting transient impulses (such as spikes generated by arc discharge) during fault occurrence.

[0087] Information Entropy: Taking Shannon entropy as an example, information entropy is a measure of the uncertainty or complexity of a signal source. Its calculation formula is as follows: ,in, This represents the probability that the signal amplitude occurs in the i-th pre-divided quantization interval. The calculation first involves quantizing the amplitude range of the signal components (e.g., dividing it into 256 equally wide intervals), and then counting the frequency of sampling points within each interval to obtain the probability distribution. The higher the information entropy value, the higher the complexity and uncertainty of the signal, and the more chaotic the waveform; conversely, the lower the information entropy, the stronger the regularity of the signal, and the more orderly the waveform. When a fault occurs, the appearance of transient components significantly increases the complexity of the signal, causing its information entropy to change. This characteristic helps to distinguish transient processes of different levels of complexity.

[0088] In S103 of this embodiment, for the eight signal components obtained through multi-level signal compression technology, all four features (energy, variance, kurtosis, and information entropy) mentioned above are calculated for each component. Thus, each signal component is described by a 4-dimensional feature vector. The feature vectors of all eight components are concatenated to obtain a feature vector F with a dimension of 8 × 4 = 32.

[0089] This 32-dimensional feature vector contains richer information than the 16-dimensional vector in Example 1. It not only describes the energy intensity and fluctuation of the signal in different frequency bands, but also its impact characteristics (through kurtosis) and intrinsic complexity (through information entropy).

[0090] In the subsequent S104, this 32-dimensional feature vector will be input into the breadth feature map classification model. Accordingly, the input layer dimension of this model should also be 32, and its training process must also be completed using the 32-dimensional feature vector extracted by the method of this embodiment.

[0091] By introducing two additional features, kurtosis and information entropy, the feature vector constructed in this embodiment helps to capture subtle differences between different fault types more precisely. For example, some high-resistance grounding faults may not have obvious energy characteristics, but their transient processes may have unique impact or complexity, which can be effectively captured by kurtosis or information entropy features. This results in higher classification accuracy and robustness when dealing with difficult-to-distinguish faults with similar characteristics.

[0092] Example 4 This application provides an adaptive wide-area feature mapping classification method based on the foregoing embodiments. Conventional machine learning models, including the wide-area feature mapping models used in embodiments 1 to 3, have a fixed structure (e.g., network width, number of nodes) after training. This means that the model uses the same computational resources and complexity to process both simple faults with very obvious features and complex faults with weak features. This embodiment can dynamically adjust its internal structure according to the characteristics of the sample to be classified, thereby optimizing computational efficiency while ensuring accuracy.

[0093] The method flow of this embodiment is based on that of embodiment 1, with the addition of a new step between S103 and S104.

[0094] S101 to S103 are exactly the same as in Example 1. That is, by acquiring data, performing multi-level signal decomposition and feature extraction, statistical characteristics such as the energy and variance of the eight signal components are obtained.

[0095] Next, this embodiment adds a feature saliency evaluation step. This step aims to perform a rapid quantitative evaluation of the saliency of the transient characteristics of the current fault signal before classification. Intuitively, metallic short-circuit faults usually induce violent high-frequency oscillations, and their transient characteristics are relatively significant; while the transient characteristics of high-resistance grounding faults or remote faults may be very weak and easily drowned out by background noise, resulting in lower feature saliency.

[0096] In this embodiment, the significance score S of the current fault is calculated based on the following scoring formula, which comprehensively utilizes the energy characteristics of each high-frequency component extracted in S103: .

[0097] in, It is the first The energy of each high-frequency detail component (e.g., corresponding to Bdetail1, Bdetail2, etc. in Example 1). The energy of each component can be selected from the previous components. The calculation is performed using the highest frequency component, as transient information is primarily concentrated in the high-frequency range. For example, it can be set... . It is the sum of the energy of all 8 signal components (Bdetail1…7,Bbase7). and Directly using energy values ​​may result in significant differences of orders of magnitude (e.g., from...). arrive This leads to computational instability. Using a logarithmic transformation can effectively compress the dynamic range of the energy value, making its changes smoother and thus enhancing the robustness of the score. Adding 1 is to avoid the situation where taking the logarithm is meaningless when the energy value is 0. It is a set of preset weight coefficients, and satisfies This set of weights reflects a priori knowledge that higher-frequency components (such as Bdetail1) typically contribute more to the judgment of transient events, and their weights should be higher accordingly. For example, it can be set... Or other decreasing sequences.

[0098] The entire formula quantifies the prominence or contribution of high-frequency transient energy to the total energy composition by calculating the weighted logarithm of the high-frequency energy and its ratio to the logarithm of the total energy. A higher score S indicates that the energy of the high-frequency component occupies an important position in the overall energy, suggesting that this is a significant transient event.

[0099] After calculating the significance score S, proceed to step c2: adaptively adjust the model structure based on the significance score. The core of this step is to directly use the score S obtained in the previous step to control the hyperparameters of the breadth feature map classification model; specifically, to dynamically determine the number of augmentation nodes used in this classification. .

[0100] In this embodiment, the number of enhancement nodes This can be determined using the following piecewise function:

[0101] in, These are the preset minimum, baseline, and maximum number of enhanced nodes. For example, they can be set... These values ​​are used as hyperparameters of the model and are determined during the offline validation phase. These are the low and high thresholds for significance scoring, which divide the score S into three intervals. These thresholds can be determined through statistical analysis of significance scores on a large amount of historical data. For example, T1 can be set as the 20th percentile of the score distribution, and T2 as the 80th percentile. It is a scaling factor used to control the rate at which the number of nodes increases with the score S within the middle range. It is a floor function to ensure that the number of nodes is an integer.

[0102] The logic of this function lies in implementing an intelligent allocation strategy for computing resources: if the significance score S is lower than the threshold T1, it means that this may be a very weak event or just noise. In this case, a very simple model is used ( This allows for rapid processing, avoiding excessive computational resources on samples lacking clear features and reducing the risk of being misled by noise. If the significance score S is higher than the threshold T2, it indicates a typical fault with strong features. In this case, the most complex model is activated (…). The model uses its strongest nonlinear expressive power to ensure the highest classification accuracy. If the significance score S is between T1 and T2, it usually corresponds to those difficult faults that require the most detailed analysis (such as high-resistance grounding, initial arcing faults, etc.). In this case, the complexity of the model is proportional to the significance of the features; the larger S is, the more significant the feature is. The larger the value, the more adaptive and on-demand the allocation of computing resources can be achieved.

[0103] Finally, in S104 (model classification), adaptive changes are also required. During the prediction phase, when the model receives a sample to be classified, it no longer uses a fixed, pre-trained broad feature map layer. Instead, it first uses the dynamic number of nodes calculated in step c2. Select a corresponding number of weights and biases from a pre-generated and stored random weight pool to temporarily construct a pool containing... The algorithm performs augmentation on each node and calculates the breadth-enhanced output He for that sample. Then, it concatenates He with the feature map output Ze to form a combined feature matrix Ae. Finally, classification is performed using uniform output weights Wo. Importantly, the output weights Wo are calculated during training based on a sufficiently large number of augmentation nodes (e.g., ...). The calculation is derived by selectively using enhanced nodes during prediction, without recalculating Wo.

[0104] The introduction of this adaptive mechanism allows the model's computational complexity and nonlinear expressive power to be customized for each sample to be classified. For simple faults with obvious features, the model automatically simplifies, resulting in faster computation; while for difficult faults with complex or subtle features, the model automatically enhances its nonlinear learning capabilities for deeper analysis. Ultimately, this embodiment effectively improves the classification efficiency and resource utilization of the entire system while ensuring high accuracy for all types of faults (especially difficult faults).

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0106] Corresponding to the power system fault classification method described in the above embodiments, Figure 2 A structural block diagram of a power system fault classification device provided in an embodiment of this application is shown. For ease of explanation, only the parts related to the embodiment of this application are shown.

[0107] Reference Figure 2 A power system fault classification device, comprising: The acquisition module is used to acquire time-series electrical quantity data collected by power system monitoring equipment; The decomposition module is used to perform multi-level signal decomposition on the electrical quantity time-series data to obtain multiple signal components. The feature extraction module is used to extract statistical features of the multiple signal components to construct a feature vector to be classified. The fault classification module is used to input the feature vector to be classified into a pre-trained breadth feature mapping classification model and output the fault classification result.

[0108] Optionally, the power system fault classification device further includes a preprocessing module for performing data dimensionality reduction processing on the acquired electrical quantity time-series data.

[0109] Optionally, the decomposition module is used for: The electrical quantity time series data is iteratively decomposed using a decomposition kernel function, generating a compressed feature variable and a redundant information component in each decomposition level. The compressed feature variables of the previous layer are used as the input of the next layer decomposition. After multiple iterations, the compressed feature variables obtained in the last iteration and the redundant information components generated in each iteration are used as the multiple signal components.

[0110] Optionally, the statistical features include at least two of energy, variance, kurtosis, and information entropy.

[0111] Optionally, the breadth feature mapping classification model maps the feature vector to be classified into high-dimensional combined features through a breadth feature mapping layer, and the decision layer performs classification based on the high-dimensional combined features.

[0112] Optionally, the power system fault classification device also includes a training module for: The sample feature vectors labeled with fault type tags are used as the training set and input into the breadth feature mapping layer to obtain the feature mapping output Zr and the breadth enhancement output Hr; The feature mapping output Zr is concatenated with the breadth enhancement output Hr to obtain the combined feature matrix Ar; The combined feature matrix Ar is input into the decision layer to obtain the decision output Dr; Based on the decision output Dr and the fault type label Yr corresponding to the training set, the output weight Wo is calculated to complete the training of the breadth feature mapping classification model.

[0113] Optionally, the fault classification module is used for: The feature vector to be classified is input into the breadth feature mapping layer of the breadth feature mapping classification model to obtain the combined feature matrix Ae; The fault classification result is output based on the output weight Wo and the combined feature matrix Ae.

[0114] It should be noted that the information interaction and execution process between the above-mentioned devices / units / modules are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0116] This application also provides a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor executes the computer program to implement the steps in any of the above method embodiments.

[0117] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0118] This application provides a computer program product that, when run on a computer device, enables the computer device to perform the steps described in the above-described method embodiments.

[0119] Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 3 As shown, the computer device of this embodiment includes: at least one processor 40 ( Figure 3 (Only one is shown in the diagram), memory 41, and computer program 42 stored in the memory 41 and executable on the at least one processor 40, wherein the processor 40 executes the computer program 42 to implement the steps in any of the above-described visual programming method embodiments.

[0120] The computer device may include, but is not limited to, a processor 40 and a memory 41. Those skilled in the art will understand that... Figure 3The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.

[0121] The processor 40 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0122] In some embodiments, the memory 41 may be an internal storage unit of the computer device, such as a hard disk or memory. In other embodiments, the memory 41 may be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 41 may include both internal and external storage units of the computer device. The memory 41 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory 41 can also be used to temporarily store data that has been output or will be output.

[0123] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a device / computer equipment, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. Computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0124] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0125] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0126] In the embodiments provided in this application, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0127] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for classifying faults in a power system, characterized in that, include: Acquire time-series electrical quantity data collected by power system monitoring equipment; The electrical quantity time-series data is subjected to multi-level signal decomposition to obtain multiple signal components; Statistical features of the multiple signal components are extracted to construct a feature vector to be classified; The feature vector to be classified is input into a pre-trained breadth feature mapping classification model, and the fault classification result is output.

2. The method according to claim 1, characterized in that, Prior to performing multi-level signal decomposition on the electrical quantity time-series data, the process also includes: The acquired electrical quantity time-series data are subjected to data dimensionality reduction processing.

3. The method according to claim 1, characterized in that, The process of performing multi-level signal decomposition on the electrical quantity time-series data yields multiple signal components, including: The electrical quantity time series data is iteratively decomposed using a decomposition kernel function, generating a compressed feature variable and a redundant information component in each decomposition level. The compressed feature variables of the previous layer are used as the input of the next layer decomposition. After multiple iterations, the compressed feature variables obtained in the last iteration and the redundant information components generated in each iteration are used as the multiple signal components.

4. The method according to claim 1, characterized in that, The statistical features include at least two of energy, variance, kurtosis, and information entropy.

5. The method according to claim 1, characterized in that, The breadth feature mapping classification model maps the feature vector to be classified into high-dimensional combined features through a breadth feature mapping layer, and the decision layer performs classification based on the high-dimensional combined features.

6. The method according to claim 5, characterized in that, The training process of the breadth feature mapping classification model is as follows: The sample feature vectors labeled with fault type tags are used as the training set and input into the breadth feature mapping layer to obtain the feature mapping output Zr and the breadth enhancement output Hr; The feature mapping output Zr is concatenated with the breadth enhancement output Hr to obtain the combined feature matrix Ar; The combined feature matrix Ar is input into the decision layer to obtain the decision output Dr; Based on the decision output Dr and the fault type label Yr corresponding to the training set, the output weight Wo is calculated to complete the training of the breadth feature mapping classification model.

7. The method according to claim 6, characterized in that, The step of inputting the feature vector to be classified into a pre-trained breadth feature mapping classification model and outputting the fault classification result includes: The feature vector to be classified is input into the breadth feature mapping layer of the breadth feature mapping classification model to obtain the combined feature matrix Ae; The fault classification result is output based on the output weight Wo and the combined feature matrix Ae.

8. A power system fault classification device, characterized in that, include: The acquisition module is used to acquire time-series electrical quantity data collected by power system monitoring equipment; The decomposition module is used to perform multi-level signal decomposition on the electrical quantity time-series data to obtain multiple signal components. The feature extraction module is used to extract statistical features of the multiple signal components to construct a feature vector to be classified. The fault classification module is used to input the feature vector to be classified into a pre-trained breadth feature mapping classification model and output the fault classification result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, When the computer program product is run on a computer device, it causes the computer device to perform the method as described in any one of claims 1 to 7.

Images