A power grid fault diagnosis method and system based on power data analysis

Abstract

The application discloses a kind of power grid fault diagnosis method and system based on electric power data analysis, it is related to electric power system fault diagnosis related field, including: the multiple electric power data sources of constructing electric power network, extract multiple electric power data;Multiple electric power data are carried out semi-supervised separation according to fault type using blind signal separation algorithm, obtain multiple separation signals;The fault feature similarity between the pairs of multiple electric power data sources is calculated;According to the cross-source end clustering of fault feature similarity calculation result, the multiple separation signal clusters output are respectively input power grid fault diagnosis model analysis, obtain multiple fault types, send power grid fault diagnosis result after joint identification.Solve the technical problem that existing power grid fault diagnosis exists in mixed fault scene and is easy to produce misjudgment due to signal coupling, leading to insufficient diagnostic accuracy and reliability, achieve the technical effect of accurate separation and identification through complex mixed fault, improve fault diagnosis accuracy and reliability.

Description

Technical Field

[0001] This application relates to the field of power system fault diagnosis, and in particular to a power grid fault diagnosis method and system based on power data analysis. Background Technology

[0002] Accurate fault diagnosis is crucial for ensuring stable power system operation, reducing fault losses, and improving power supply service quality. Especially in complex and ever-changing mixed fault scenarios, rapid and accurate diagnosis is a core prerequisite for preventing fault escalation. Currently, the main approach to this problem is to directly combine mixed fault features collected from multiple device data sources for overall fault diagnosis, without differentiating and processing the data signals from different devices. Because fault signals from different devices couple and interfere with each other in mixed fault scenarios, directly combining all fault features for diagnosis leads to fault feature confusion, loss of key information, and a high risk of misjudgment, making it impossible to accurately identify complex mixed fault types.

[0003] At present, the relevant technologies for power grid fault diagnosis have technical problems such as the susceptibility to misjudgment due to signal coupling in mixed fault scenarios, resulting in insufficient accuracy and reliability of diagnosis. Summary of the Invention

[0004] This application provides a power grid fault diagnosis method and system based on power data analysis. It employs techniques such as constructing a power network data source and extracting multi-source power data, performing semi-supervised separation using a blind signal separation algorithm to obtain separated signals from each source, calculating the similarity of fault features across source ends and performing cluster analysis to form different types of signal clusters, inputting these signal clusters into a power grid fault diagnosis model, and finally outputting comprehensive fault diagnosis results through joint identification and diagnostic analysis. These techniques address the technical problem of insufficient accuracy and reliability in existing power grid fault diagnosis methods, which are prone to misjudgment due to signal coupling in mixed fault scenarios. The method achieves the technical effect of improving the accuracy and reliability of fault diagnosis through precise separation and identification of complex mixed faults.

[0005] This application provides a power grid fault diagnosis method based on power data analysis, comprising: constructing multiple power data sources of a power network; extracting multiple power data corresponding to the multiple power data sources; using a blind signal separation algorithm to perform semi-supervised separation of the multiple power data according to fault type, obtaining multiple separated signals corresponding to each power data source; calculating the fault feature similarity of the separated signal pairs among the multiple power data sources, obtaining the fault feature similarity calculation result; performing cross-source clustering according to the fault feature similarity calculation result, until the multiple power data sources are traversed and multiple separated signal clusters are output; inputting the output multiple separated signal clusters into a power grid fault diagnosis model for analysis, obtaining multiple fault types corresponding one-to-one with the multiple separated signal clusters; jointly identifying the multiple fault types and sending the power grid fault diagnosis result.

[0006] In a possible implementation, a blind signal separation algorithm is used to separate the multiple power data according to the fault type, and the following processing is also performed: a preset power signal conversion template is used to resample the multiple power data to obtain multiple resampled power signals; the multiple resampled power signals are mapped to a preset numerical range in the power signal conversion template to obtain multiple standardized power signals; a blind signal separation algorithm is used to perform semi-supervised separation of the multiple standardized power signals according to the fault type to obtain multiple separated signals corresponding to each power data source.

[0007] In a possible implementation, a blind signal separation algorithm is used to perform semi-supervised separation of the multiple standardized power signals according to fault type, and the following processing is performed: obtaining a preset fault type set corresponding to any power data source, wherein each fault type in the preset fault type set includes a fault signal feature template; training a fault type filter bank according to the fault type set and the corresponding fault signal feature template set, wherein the fault type filter bank includes multiple fault type filters corresponding to the multiple power data sources, each fault type filter is a multi-channel filter, and each channel corresponds to a fault type; calling the fault type filter to perform frequency band filtering on the standardized power signal corresponding to each power data source to obtain multiple separated signals corresponding to each power data source.

[0008] In a possible implementation, the following processing is performed: each fault type in the preset fault type set includes a fault signal feature template, which includes frequency range, main energy distribution, harmonic ratio characteristics, instantaneous jump intensity, and reference waveform morphology.

[0009] In a possible implementation, a fault type filter is invoked to perform frequency band filtering on the standardized power signal corresponding to each power data source, and the following processing is performed: obtain the fault type filter matched for each power data source; match and invoke the fault type filter to perform frequency band convolution processing on the standardized power signal corresponding to each power data source to obtain the frequency band response components corresponding to each fault type; obtain multiple separate signals corresponding to each power data source according to the frequency band response components corresponding to each fault type.

[0010] In a possible implementation, the following processing is performed: calculating the fault feature similarity of the separated signal pairs among the plurality of power data source terminals, wherein the fault features include time-frequency features, waveform distance features, and fault mode embedding features, wherein the time-frequency features are two-dimensional time-frequency spectra of the separated signal pairs based on time-frequency distribution, the waveform distance features are the DTW waveform distance of the separated signal pairs, and the fault mode embedding features are high-level semantic fault features representing the separated signal pairs; and obtaining the fault feature similarity of the separated signal pairs by weighting the time-frequency features, waveform distance features, and fault mode embedding features.

[0011] In a possible implementation, the method for outputting the fault feature similarity calculation result performs the following processing: sequentially calculate the m fault feature similarities between the first separated signal in the first power data source and all m separated signals in the remaining power data source; and so on, to obtain k×m fault feature similarities between k separated signals in the first power data source, and output the k×m fault feature similarities as the fault feature similarity calculation result, where k and m are both positive integers greater than or equal to 0.

[0012] In a possible implementation, cross-source clustering is performed based on the fault feature similarity calculation results until multiple power data source ends are traversed and multiple separate signal clusters are output. The following processing is performed: a fault similarity threshold is set, and the fault feature similarity of the separate signal pairs of the multiple power data source ends is traversed sequentially according to the fault similarity threshold; the separate signal pairs with a similarity greater than the fault similarity threshold are divided into the same cluster, and multiple separate signal clusters are output.

[0013] In a possible implementation, the following processing is performed: multiple output separate signal clusters are input into a power grid fault diagnosis model for analysis, wherein the power grid fault diagnosis model includes a Softmax single-label classification layer; the fault type distribution probability of each separate signal cluster is output by the Softmax single-label classification layer, and the first fault type of each separate signal cluster is determined based on the fault type distribution probability, wherein the first fault type is the class label with the highest probability.

[0014] This application also provides a power grid fault diagnosis system based on power data analysis, comprising: a power data extraction module for constructing multiple power data source terminals of a power network and extracting multiple power data corresponding to the multiple power data source terminals; a semi-supervised separation module for performing semi-supervised separation of the multiple power data according to fault type using a blind signal separation algorithm to obtain multiple separated signals corresponding to each power data source terminal; a fault feature similarity calculation module for calculating the fault feature similarity of separated signal pairs among the multiple power data source terminals to obtain fault feature similarity calculation results; and a power grid fault diagnosis module for performing cross-source clustering according to the fault feature similarity calculation results until multiple power data source terminals are traversed and multiple separated signal clusters are output, and the output multiple separated signal clusters are respectively input into a power grid fault diagnosis model for analysis to obtain multiple fault types corresponding one-to-one with the multiple separated signal clusters, and after jointly identifying the multiple fault types, a power grid fault diagnosis result is sent.

[0015] This application proposes a power grid fault diagnosis method and system based on power data analysis. First, multiple power data sources are constructed from the power network. Multiple power data points corresponding to these sources are extracted. Then, a blind signal separation algorithm is used to semi-supervisedly separate the power data according to fault type, obtaining multiple separated signals corresponding to each power data source. Next, the fault feature similarity of the separated signal pairs among the multiple power data sources is calculated, yielding the fault feature similarity calculation results. Finally, cross-source clustering is performed based on the fault feature similarity calculation results until all multiple power data sources are traversed, outputting multiple separated signal clusters. These output separated signal clusters are then input into a power grid fault diagnosis model for analysis, obtaining multiple fault types corresponding one-to-one with each separated signal cluster. After jointly identifying these multiple fault types, the power grid fault diagnosis results are sent. This achieves the technical effect of improving the accuracy and reliability of fault diagnosis through precise separation and identification of complex mixed faults. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a power grid fault diagnosis method based on power data analysis, provided as an embodiment of this application.

[0018] Figure 2This is a schematic diagram of the separation result in a semi-supervised separation process in a power grid fault diagnosis method based on power data analysis provided in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of the structure of a power grid fault diagnosis system based on power data analysis, provided in an embodiment of this application.

[0020] Figure labeling: Power data extraction module 10, semi-supervised separation module 20, fault feature similarity calculation module 30, power grid fault diagnosis module 40. Detailed Implementation

[0021] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0024] This application provides a power grid fault diagnosis method based on power data analysis, such as... Figure 1 As shown, the method includes:

[0025] Step S100: Construct multiple power data source terminals for the power network and extract multiple power data corresponding to the multiple power data source terminals.

[0026] Specifically, the power data source refers to the source in the power network that can generate and provide power data, including various sensors, monitoring devices, and other devices installed at key locations on power equipment or power lines. These devices collect various parameter information during the power operation process in real time, such as current, voltage, power, and frequency, which are the basic data sources for data processing and fault diagnosis.

[0027] Specifically, key monitoring points in the power network are identified, such as the output terminals of power plants, the incoming and outgoing lines of substations, and the access terminals of important power users. Various sensors, such as current transformers, voltage transformers, and frequency sensors, are deployed at these key nodes to collect power data in real time. These sensors are connected to different power equipment or power line segments, forming multiple power data sources. The collected data is transmitted to a data processing center via data acquisition cards and communication modules (such as 4G / 5G communication modules and fiber optic communication modules). The data processing center uses server clusters or cloud computing platforms to store and process large amounts of data.

[0028] For example, taking a medium-sized city power grid as an example, sensors are installed at substations, transmission lines, distribution lines, and some important users in various areas of the city. Current transformers and voltage transformers are installed on the high and low voltage sides of the main transformers in substations to monitor the operating current and voltage of the main transformers; insulator leakage current sensors and conductor temperature sensors are installed on the towers of transmission lines to obtain information on the insulation status and conductor thermal status of the transmission lines; fault indicators are installed at the branches of distribution lines to detect short circuits and grounding faults. These sensors convert the acquired analog signals into digital signals via data acquisition cards, and then transmit the data to the power grid company's data processing center via a 4G communication module. The data processing center uses a distributed storage system capable of storing massive amounts of power data and performing preliminary data cleaning and preprocessing.

[0029] Step S200: A blind signal separation algorithm is used to perform semi-supervised separation of the multiple power data according to the fault type, so as to obtain multiple separated signals corresponding to each power data source.

[0030] Specifically, the blind signal separation algorithm is a class of algorithms that recovers the original independent signal sources using only observed mixed signals without knowing the statistical characteristics of the signal sources or the transmission channel parameters. Its basic principle is based on the statistical independence and non-Gaussianity of signals, using mathematical transformations to decompose the mixed signal into multiple uncorrelated signal components, thereby achieving signal separation. The semi-supervised separation method combines supervised and unsupervised learning. It uses a small amount of labeled data (e.g., known fault types) to supervise and guide the separation process, while simultaneously performing unsupervised separation processing on a large amount of unlabeled data, making full use of limited labeled data resources and improving the accuracy and reliability of the separation results.

[0031] Specifically, the extracted power data undergoes preprocessing, including data standardization and noise removal, to improve the performance of the blind signal separation algorithm. A blind signal separation algorithm is selected, and its parameter settings are determined based on the data characteristics and fault type requirements. The preprocessed data is then input into the blind signal separation algorithm for signal separation, yielding a preliminary separated signal. A semi-supervised learning method is used to associate a small amount of labeled data with known fault types with the separated signal, adjusting and optimizing the separation results to make the separated signal more consistent with fault characteristics.

[0032] For example, in a fault diagnosis system for a regional power grid, suppose data such as current and voltage from multiple transmission lines are collected. This data includes normal operation data and a mixture of data from various fault types (such as single-phase grounding faults and phase-to-phase short-circuit faults). First, the collected data is standardized by adjusting the mean to 0 and the standard deviation to 1 to eliminate the influence of different dimensions on the separation results. Then, a fast independent component analysis (ICA) algorithm is used for signal separation, with kurtosis as the non-Gaussianity metric and a convergence accuracy of 1e-6. After obtaining multiple signal components, a small amount of labeled data with known fault types is compared and analyzed with the separated signals. For example, if the current signal characteristics corresponding to a single-phase grounding fault are known, the separated signal components are matched with these characteristics for similarity. Signal components with high matching degrees are labeled as the separated signals for that fault type. Simultaneously, other unlabeled separated signals are inferentially separated based on the characteristics of the labeled signals. Finally, multiple separated signals are obtained for each power data source, and each separated signal may correspond to different fault characteristics or normal operation characteristics.

[0033] In one possible implementation, a blind signal separation algorithm is used to separate the multiple power data according to the fault type. Step S200 further includes step S210, which involves setting a preset power signal conversion template and resampling the multiple power data using the template to obtain multiple resampled power signals. Specifically, the preset power signal conversion template is a standard framework that defines the desired signal sampling frequency and other related parameters. Because power data collected from different power data sources may have different sampling frequencies and other attributes—for example, some devices collect data 100 times per second, while others collect data 1000 times per second—the preset power signal conversion template specifies a unified target sampling frequency, such as 500 samples per second. The original power data is resampled according to the sampling frequency of the power signal conversion template. For example, for signals with an original sampling frequency lower than the target frequency, interpolation or other methods are used to increase the number of data points to reach the target frequency; for signals with an original sampling frequency higher than the target frequency, decimation or other methods are used to reduce the number of data points, ultimately resulting in multiple resampled power signals that conform to the sampling frequency of the power signal conversion template. By performing signal resampling, the messy and inconsistent raw data is organized into relatively neat and uniform data.

[0034] Step S220 involves mapping the multiple resampled power signals to a preset numerical range in the power signal conversion template to obtain multiple standardized power signals. Specifically, the power signal conversion template also specifies a numerical range, for example, mapping all signal values ​​to the range of [0,1] or [-1,1]. This is because the original numerical ranges of different power data can vary greatly; for example, voltage may vary from hundreds of volts to tens of thousands of volts, and current may range from a few milliamps to thousands of amps. Through the mapping operation, the value of each resampled power signal is proportionally converted to this preset numerical range, thereby obtaining standardized power signals.

[0035] Step S230 involves using a blind signal separation algorithm to semi-supervisedly separate the multiple standardized power signals according to fault type, obtaining multiple separated signals corresponding to each power data source. Specifically, after standardizing the signals, these standardized signals are input into the blind signal separation algorithm. Based on the statistical characteristics of the signals, the algorithm mines independent signal features corresponding to different fault types, thereby decomposing the mixed signals. A small amount of labeled fault data is used to guide the separation process, helping the algorithm better identify the signal components corresponding to different fault features, improving the accuracy of the separation, and thus obtaining multiple separated signals corresponding to each power data source. These separated signals correspond to a single, clear fault feature or normal operation feature.

[0036] This implementation avoids problems such as time series misalignment and difficulty in data comparison caused by different sampling frequencies by performing signal resampling. Standardization helps improve the numerical stability of the algorithm, making the final separated signal more representative and accurate.

[0037] In one possible implementation, a blind signal separation algorithm is used to perform semi-supervised separation of the multiple standardized power signals according to fault type. Step S230 further includes step S231, obtaining a preset fault type set corresponding to any power data source. Each fault type in the preset fault type set includes a fault signal feature template, which includes frequency range, main energy distribution, harmonic ratio characteristics, instantaneous jump intensity, and reference waveform morphology. Specifically, a large amount of historical fault data is collected from power system operation and maintenance records, fault reports, and fault cases provided by power equipment manufacturers. This data includes power signal characteristics when different fault types (such as short-circuit faults, ground faults, overload faults, etc.) occur. The collected historical fault data is analyzed to extract features for each fault type. Based on the analysis results, a preset fault type set containing multiple common fault types is constructed; that is, the preset fault type set is a standard library of fault types, containing various common power fault types, such as short-circuit faults, open-circuit faults, insulation aging faults, etc. Each fault type has a corresponding fault signal feature template, including frequency range (e.g., short circuit faults have obvious characteristics in a certain high frequency range), main energy distribution (energy is concentrated in a certain frequency band), harmonic ratio characteristics (different faults generate a specific proportion of harmonics), instantaneous jump intensity (e.g., instantaneous voltage or current jump amplitude), and reference waveform shape (e.g., whether the waveform is peaked or smooth).

[0038] Step S232: Train a fault type filter bank according to the fault type set and the corresponding fault signal feature template set. The fault type filter bank includes multiple fault type filters corresponding to the multiple power data source ends. Each fault type filter is a multi-channel filter, and each channel corresponds to one fault type. Specifically, determine the initial parameters of the filter, such as center frequency and bandwidth, based on the frequency range of the power signal and the frequency information in the fault signal feature template. For example, if the characteristic frequency range of a certain fault type is 100-200Hz, then the center frequency of the corresponding filter channel can be set at around 150Hz, and the bandwidth can cover 100-200Hz. Select a filter design method, such as a Finite Impulse Response (FIR) filter or an Infinite Impulse Response (IIR) filter. For example, FIR filters have good linear phase characteristics and are suitable for precise frequency band filtering of signals. Design a multi-channel filter bank according to the number of faults in the fault type set, with each channel corresponding to one fault type. The filter is trained using data from a set of fault signal feature templates. The signals from these templates are used as input, and the filter parameters (such as filter coefficients) are adjusted to match the filter's output with the expected fault characteristics. For example, for a short-circuit fault filter channel, the input signal is a signal from a short-circuit fault signal feature template. The filter coefficients are adjusted so that the output signal has a higher amplitude within the short-circuit fault characteristic frequency range and a lower amplitude in other frequency ranges. Optimization algorithms such as least squares can be used to minimize the error between the filter output and the expected output, thus obtaining the optimal filter parameters.

[0039] Step S233 involves calling a fault type filter to perform frequency band filtering on the standardized power signal corresponding to each power data source, resulting in multiple separate signals for each power data source. Specifically, the standardized power signal corresponding to each power data source is used as the input to the fault type filter group, and the fault type filter performs frequency band filtering on the input signal. Each channel's filter filters the input signal according to its designed frequency band range, extracting the characteristic signal of the corresponding fault type. For example, for a fault type filter containing short-circuit faults, ground faults, and overload faults, the filter for the short-circuit fault channel extracts the signal components within the short-circuit fault characteristic frequency range of the input signal, the filter for the ground fault channel extracts the signal components within the ground fault characteristic frequency range, and the filter for the overload fault channel extracts the signal components within the overload fault characteristic frequency range. After filtering, each channel outputs a separate signal, and these separate signals correspond to different fault types. These separate signals represent different fault characteristic components that may exist in the original power signal.

[0040] like Figure 2 The figure shows the separation results of a power data source signal after processing by a 4-channel filter. Each sub-figure corresponds to the separated signal of a different fault type. From top to bottom, they are the separated signal for short-circuit faults, third harmonic faults, overvoltage faults, and ground faults. The horizontal axis represents time, and the vertical axis represents signal amplitude. The figure shows that the filter successfully extracted the characteristic signals of short-circuit faults and third harmonic faults.

[0041] This implementation provides a clear objective for filter design and training by constructing a detailed set of fault types and fault signal feature templates, thereby improving the accuracy and relevance of fault feature extraction. Multi-channel fault type filters can simultaneously filter features from multiple fault types. Each channel independently performs band filtering on the input signal, enabling rapid separation of signals for different fault types. This reduces the time spent checking fault types one by one and improves the efficiency of fault diagnosis.

[0042] In one possible implementation, a fault type filter is invoked to perform frequency band filtering on the standardized power signal corresponding to each power data source. Step S233 further includes step S2331, obtaining the fault type filter matched for each power data source. Specifically, the matching rules between each power data source and the fault type filter are determined based on factors such as the equipment type, installation location, and historical fault conditions of the power data source. Within the fault type filter group, the corresponding fault type filter is found for each power data source by searching the filter group's index or label according to the pre-set matching rules. Each filter has an associated label indicating which type of power data source it is applicable to.

[0043] Step S2332 involves matching and applying fault type filters to perform frequency band convolution processing on the standardized power signals corresponding to each power data source, obtaining the frequency band response components corresponding to each fault type. Specifically, the obtained fault type filters are applied to the corresponding standardized power signals, and frequency band convolution processing is performed. That is, the frequency response characteristics of the filter are convolved with the spectrum of the signal to extract the frequency band response components in the signal that match the frequency characteristics of the filter. More specifically, the frequency domain representations of the signal and the filter can be multiplied, and then the time domain frequency band response components are obtained through inverse Fourier transform. Through frequency band convolution processing, the frequency band response components corresponding to each fault type are obtained. These frequency band response components represent the intensity and characteristics of the frequency components in the original signal related to a specific fault type.

[0044] Step S2333: Obtain multiple separate signals corresponding to each power data source based on the frequency band response components corresponding to each fault type. Specifically, based on the frequency band response components corresponding to each fault type, these components are combined and processed to obtain multiple separate signals corresponding to each power data source. For example, each frequency band response component can be extracted separately and subjected to appropriate amplification, filtering, or other operations to highlight the characteristics of a specific fault type. Finally, multiple separate signals corresponding to each power data source are output, each corresponding to a different fault type, for fault diagnosis and analysis.

[0045] Step S300: Calculate the fault feature similarity of the separated signal pairs among the multiple power data source terminals to obtain the fault feature similarity calculation result.

[0046] Specifically, fault feature similarity is an indicator used to measure the degree of similarity between different signals in terms of fault features. By extracting and quantifying the features of the signals, such as the amplitude, frequency, phase, and waveform shape, similarity measurement methods are used to calculate the degree of matching between two signals in these features. The higher the similarity, the more similar the two signals are in terms of fault features, and they may belong to the same type of fault or have similar fault causes and propagation paths.

[0047] From the obtained multiple separated signals, pairwise signal pairs are selected as the calculation objects. Feature extraction is performed on each signal pair. The extracted features can include signal amplitude features, frequency features, time-domain features (such as mean, variance, peak value, etc.), and frequency-domain features (such as power spectral density, harmonic content, etc.), forming feature vectors. Based on the selected similarity measurement method, the similarity between the feature vectors of each signal pair is calculated, resulting in a fault feature similarity matrix. The similarity measurement method can be Euclidean distance, cosine similarity, correlation coefficient, etc. These methods can measure the similarity between two signals from different perspectives. Euclidean distance measures the absolute distance between signals, cosine similarity measures the directional similarity between signals, and correlation coefficient measures the linear correlation between signals.

[0048] For example, when calculating the fault feature similarity of isolated signals from multiple substations in a power grid, assume that isolated signals from two substations are selected, with three isolated signals from each substation, corresponding to three possible fault features. First, time-domain and frequency-domain features are extracted for each isolated signal. In the time domain, features such as the signal's mean, variance, and peak value (the ratio of peak value to mean) are calculated. In the frequency domain, the power spectral density of the signal is calculated using Fast Fourier Transform (FFT), and the main frequency components and their corresponding amplitudes are extracted. Then, cosine similarity is chosen as the similarity metric, and the cosine similarity between each pair of isolated signals from the two substations is calculated. For example, to calculate the cosine similarity between the first isolated signal from substation A and the first isolated signal from substation B, their feature vectors are multiplied by a dot product, and then divided by the product of the magnitudes of the two vectors to obtain the similarity value. The similarity of all signal pairs is calculated sequentially, forming a 3×3 similarity matrix.

[0049] In one possible implementation, step S300 further includes step S310, calculating the fault feature similarity of the separated signal pairs between the multiple power data source ends. The fault features include time-frequency features, waveform distance features, and fault mode embedding features. The time-frequency features are two-dimensional time-frequency maps of the separated signal pairs based on their time-frequency distribution. The waveform distance features are the DTW waveform distance of the separated signal pairs. The fault mode embedding features are high-level semantic fault features representing the separated signal pairs. Specifically, the separated signal pairs are subjected to time-frequency transformation (such as Short-Time Fourier Transform (STFT) or wavelet transform) to obtain a two-dimensional time-frequency map based on the time-frequency distribution. For example, for a separated signal, an STFT transformation can yield a distribution map on the time-frequency plane, showing the frequency components and energy levels of the signal at different time points. The waveform distance of the separated signal pairs is calculated using the Dynamic Time Warping (DTW) algorithm. The DTW algorithm can effectively measure the similarity of two time series in waveform shape, even if they have some scaling or distortion on the time axis. For example, for two separate signal waveform sequences, the DTW algorithm can find the optimal nonlinear time mapping relationship between them, thereby calculating their waveform distance. Deep learning models are used to extract features from the separate signals, obtaining high-level semantic fault features representing the signal pair. For instance, a convolutional neural network model can be trained, taking the time-frequency spectrum or waveform sequence of the separate signals as input, and extracting fault mode embedding features through convolutional layers, pooling layers, and other operations. These features can reflect the high-level semantic information of the signals, such as fault type and fault severity.

[0050] For each fault feature (time-frequency feature, waveform distance feature, and fault mode embedding feature) of the separated signal pair, their similarity is calculated separately. For example, for time-frequency features, cosine similarity can be used to measure the similarity between two time-frequency spectra; for waveform distance features, the DTW distance can be normalized and used as a measure of similarity (the smaller the distance, the higher the similarity); for fault mode embedding features, cosine similarity can also be used to calculate their similarity.

[0051] Step S320: The fault feature similarity of the separated signal pairs is obtained by weighting the time-frequency features, waveform distance features, and fault mode embedding features. Specifically, according to actual needs and the importance of each feature, corresponding weights are assigned to the time-frequency features, waveform distance features, and fault mode embedding features. For example, experimental analysis shows that time-frequency features are the most critical for fault diagnosis, followed by waveform distance features, while fault mode embedding features are relatively less important. Therefore, they can be assigned weights of 0.5, 0.3, and 0.2, respectively. The similarity of each feature is multiplied by its corresponding weight, and then the weighted similarities are summed to obtain the final fault feature similarity of the separated signal pairs.

[0052] This implementation comprehensively characterizes the fault features of separated signal pairs from multiple perspectives by integrating time-frequency features, waveform distance features, and fault mode embedding features. This multi-dimensional feature fusion approach can more accurately measure the similarity of fault features between separated signal pairs, avoiding the one-sidedness that may occur with single feature measurement.

[0053] In one possible implementation, the fault feature similarity calculation result is output. Step S300 further includes step S330, which sequentially calculates the fault feature similarity between the first separated signal in the first power data source and all m separated signals in the remaining power data sources. Specifically, the first separated signal is selected from a certain power data source as a reference signal. This reference signal is compared with the m separated signals in all the remaining power data sources. Here, m represents the total number of separated signals in all the remaining power data sources, which is determined based on the actual data acquisition and separation results. For each separated signal in the remaining power data sources, the fault feature similarity with the reference signal is calculated using a fault feature similarity calculation method (such as weighted calculation of time-frequency features, waveform distance features, and fault mode embedding features).

[0054] Step S340, and so on, obtains k×m fault feature similarities for k separate signals in the first power data source. These k×m fault feature similarities are then output as the fault feature similarity calculation result, where k and m are both positive integers greater than or equal to 0. Specifically, step S330 is repeated for each power data source, that is, each separate signal in each power data source is used as a reference signal, and the fault feature similarity is calculated with the m separate signals from all remaining power data sources. All calculated fault feature similarities are collected to form a complete fault feature similarity matrix or list.

[0055] This implementation ensures that each isolated signal is compared with all other isolated signals without omission, providing comprehensive similarity information for fault analysis.

[0056] Step S400: Perform cross-source clustering based on the fault feature similarity calculation results until multiple power data source ends are traversed and multiple separate signal clusters are output. Input the multiple separate signal clusters into the power grid fault diagnosis model for analysis to obtain multiple fault types that correspond one-to-one with the multiple separate signal clusters. After jointly identifying the multiple fault types, send the power grid fault diagnosis results.

[0057] Specifically, cross-source clustering refers to the process of clustering separate signals from different power data sources according to their fault characteristic similarity. This is used to group signals with similar fault characteristics, even if they come from different devices or geographical locations, into the same cluster. This allows for a global analysis of fault conditions in the entire power network, uncovering potential associated and systemic fault risks, and providing strong support for accurate fault diagnosis and location. The power grid fault diagnosis model is a model built on certain theories and algorithms for analyzing and judging power grid faults. By learning and training on a large amount of historical fault data and normal operation data, it masters the characteristic patterns and rules of different fault types, enabling rapid and accurate fault type identification and location of input separate signal clusters.

[0058] Based on the obtained fault feature similarity matrix, it is transformed into a similarity graph, where nodes represent separated signals and edge weights represent similarity. A clustering algorithm is selected, and clustering parameters are determined, such as the neighborhood radius ε and minimum sample size MinPts in the DBSCAN algorithm. Clustering is performed on the similarity graph to obtain multiple separated signal clusters, each containing separated signals with high similarity. These separated signal clusters are then input into a power grid fault diagnosis model. The model, based on pre-trained fault feature knowledge, analyzes the fault type of each cluster and outputs the corresponding fault type for each cluster. Finally, the fault types of all separated signal clusters are jointly identified, and the final power grid fault diagnosis result is generated by combining the fault types of each cluster with their location and impact within the entire power network.

[0059] In one possible implementation, cross-source clustering is performed based on the fault feature similarity calculation results until multiple power data source ends are traversed and multiple separate signal clusters are output. Step S400 further includes step S410, setting a fault similarity threshold, and sequentially traversing the fault feature similarity of separate signal pairs from the multiple power data source ends according to the fault similarity threshold. Specifically, a fault similarity threshold is set according to actual needs and experience to determine whether the similarity between separate signal pairs is high enough to be classified into the same cluster. For example, assuming the threshold is set to 0.7, it means that only separate signal pairs with a fault feature similarity greater than 0.7 will be classified into the same cluster. In practical applications, the threshold can be adjusted according to different fault diagnosis scenarios and requirements for clustering results. For example, in scenarios with high requirements for fault type discrimination, the threshold can be appropriately increased; when more lenient clustering conditions are required, the threshold can be appropriately decreased. The fault feature similarity of all separate signal pairs is viewed sequentially according to the set order (e.g., starting from the first separate signal from the first power data source end). During the traversal, the similarity of each separated signal pair is marked, and it is recorded whether the signal pair has already been processed to avoid duplicate processing. For example, a two-dimensional array can be used to mark which signal pairs have been compared and processed.

[0060] Step S420: Separate signal pairs with a similarity greater than the fault similarity threshold are grouped into the same cluster, outputting multiple separate signal clusters. Specifically, a new cluster is created, and the first unprocessed separate signal is used as the initial member of this cluster. The similarity of other separate signals with the fault characteristics of existing members in the cluster is checked sequentially. If a separate signal has a similarity greater than a set threshold with at least one member in the cluster, and this separate signal has not yet been assigned to another cluster, it is added to that cluster. The remaining separate signals are checked until all separate signals have been processed. For each unassigned separate signal, the above process is repeated to create new clusters or add them to existing clusters. All formed clusters are organized into the final output result. Each cluster contains a set of separate signals with similar fault characteristics; these signals may correspond to the same fault type or have similar fault modes.

[0061] This implementation method can quickly and effectively cluster separated signals by setting a fault similarity threshold, avoiding complex clustering algorithm calculations and improving clustering efficiency.

[0062] In one possible implementation, step S400 further includes step S430, whereby the multiple output separate signal clusters are input into a power grid fault diagnosis model for analysis. The power grid fault diagnosis model includes a Softmax single-label classification layer. Specifically, multiple separate signal clusters are used as input data, each containing a group of separate signals with similar fault characteristics. Feature extraction is performed on each separate signal cluster; the extracted features may include time-frequency features, waveform features, statistical features, etc. These features are then used as input to the power grid fault diagnosis model. The power grid fault diagnosis model can be a deep learning-based model, with the last few layers including a Softmax single-label classification layer. The Softmax layer maps the model's output to a probability distribution, representing the probability of each fault type. The output of the Softmax function is a probability vector, with all elements summing to 1. For example, if there are 5 fault types, the output of the Softmax layer will be a 5-dimensional vector, with each element representing the probability of the corresponding fault type.

[0063] Step S440: Based on the fault type distribution probability output by the Softmax single-label classification layer for each separated signal cluster, the first fault type of each separated signal cluster is determined according to the fault type distribution probability, where the first fault type is the class label with the highest probability. Specifically, the Softmax layer outputs the probability distribution of each fault type. For example, if the output vector is [0.1, 0.3, 0.4, 0.1, 0.1], it indicates the probability that the separated signal cluster belongs to each fault type. For each separated signal cluster, its fault type distribution probability is analyzed, and the class label with the highest probability is selected as the first fault type of the separated signal cluster. For example, in the above probability distribution, the third fault type (probability of 0.4) is determined as the first fault type. The first fault type of each separated signal cluster is output as the final fault diagnosis result. In this implementation, the Softmax single-label classification layer can effectively map each separated signal cluster to the most likely fault type, improving the accuracy of fault classification.

[0064] This application's embodiments employ techniques such as constructing a power network data source and extracting multi-source power data, performing semi-supervised separation using a blind signal separation algorithm to obtain separated signals from each source, calculating the similarity of fault features across source ends and performing cluster analysis to form signal clusters of different categories, inputting these signal clusters into a power grid fault diagnosis model, and finally outputting comprehensive fault diagnosis results through joint identification and diagnostic analysis. These techniques solve the technical problem of insufficient accuracy and reliability in existing power grid fault diagnosis due to signal coupling in mixed fault scenarios. The application achieves the technical effect of improving the accuracy and reliability of fault diagnosis through accurate separation and identification of complex mixed faults.

[0065] In the above text, refer to Figures 1-2 A power grid fault diagnosis method based on power data analysis according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 3 This invention describes a power grid fault diagnosis system based on power data analysis according to an embodiment of the present invention.

[0066] According to an embodiment of the present invention, a power grid fault diagnosis system based on power data analysis is used to solve the technical problem that existing power grid fault diagnosis systems are prone to misjudgment due to signal coupling in mixed fault scenarios, resulting in insufficient accuracy and reliability of diagnosis. The system achieves the technical effect of improving the accuracy and reliability of fault diagnosis through precise separation and identification of complex mixed faults. The power grid fault diagnosis system based on power data analysis includes: a power data extraction module 10, a semi-supervised separation module 20, a fault feature similarity calculation module 30, and a power grid fault diagnosis module 40.

[0067] The power data extraction module 10 is used to construct multiple power data sources of the power network and extract multiple power data corresponding to the multiple power data sources; the semi-supervised separation module 20 is used to perform semi-supervised separation of the multiple power data according to the fault type using a blind signal separation algorithm to obtain multiple separated signals corresponding to each power data source; the fault feature similarity calculation module 30 is used to calculate the fault feature similarity of the separated signal pairs between the multiple power data sources and obtain the fault feature similarity calculation result; the power grid fault diagnosis module 40 is used to perform cross-source clustering according to the fault feature similarity calculation result until the multiple power data sources are traversed and multiple separated signal clusters are output. The output multiple separated signal clusters are respectively input into the power grid fault diagnosis model for analysis to obtain multiple fault types that correspond one-to-one with the multiple separated signal clusters. After jointly identifying the multiple fault types, the power grid fault diagnosis result is sent.

[0068] The specific configuration of the semi-supervised separation module 20 will be described in detail below. As mentioned above, a blind signal separation algorithm is used to perform semi-supervised separation of the multiple power data according to the fault type. The semi-supervised separation module 20 may further include: a power signal conversion unit for pre-setting a power signal conversion template, and performing signal resampling processing on the multiple power data through the power signal conversion template to obtain multiple resampled power signals; a signal mapping unit for mapping the multiple resampled power signals to a preset numerical range in the power signal conversion template to obtain multiple standardized power signals; and a semi-supervised separation unit for performing semi-supervised separation of the multiple standardized power signals according to the fault type using a blind signal separation algorithm to obtain multiple separated signals corresponding to each power data source.

[0069] The method employs a blind signal separation algorithm to perform semi-supervised separation of the multiple standardized power signals according to fault type. The semi-supervised separation unit may further include: a preset fault type set acquisition subunit for obtaining a preset fault type set corresponding to any power data source, wherein each fault type in the preset fault type set includes a fault signal feature template; a fault type filter bank training subunit for training a fault type filter bank according to the fault type set and the corresponding fault signal feature template set, wherein the fault type filter bank includes multiple fault type filters corresponding to the multiple power data sources, each fault type filter is a multi-channel filter, and each channel corresponds to a fault type; and a frequency band filtering subunit for calling the fault type filters to perform frequency band filtering on the standardized power signals corresponding to each power data source to obtain multiple separated signals corresponding to each power data source.

[0070] The preset fault type set acquisition subunit may further include: each fault type in the preset fault type set includes a fault signal feature template, and the fault signal feature template includes frequency range, main energy distribution, harmonic ratio characteristics, instantaneous jump intensity and reference waveform morphology.

[0071] Specifically, the fault type filter is invoked to perform frequency band filtering on the standardized power signal corresponding to each power data source. The frequency band filtering subunit may further include: a fault type filter matching component for obtaining the fault type filter matched for each power data source; a frequency band convolution processing component for performing frequency band convolution processing on the standardized power signal corresponding to each power data source by matching and invoking the fault type filter to obtain the frequency band response components corresponding to each fault type; and a separation signal acquisition component for obtaining multiple separation signals corresponding to each power data source according to the frequency band response components corresponding to each fault type.

[0072] The specific configuration of the fault feature similarity calculation module 30 will be described in detail below. As mentioned above, the fault feature similarity calculation module 30 may further include: a weighted calculation unit for calculating the fault feature similarity of the separated signal pairs between the multiple power data source ends, wherein the fault features include time-frequency features, waveform distance features, and fault mode embedding features. The time-frequency features are two-dimensional time-frequency spectra of the separated signal pairs based on time-frequency distribution, the waveform distance features are the DTW waveform distance of the separated signal pairs, and the fault mode embedding features are high-level semantic fault features representing the separated signal pairs. By performing weighted calculations on the time-frequency features, waveform distance features, and fault mode embedding features, the fault feature similarity of the separated signal pairs is obtained.

[0073] The fault feature similarity calculation module 30, which outputs the fault feature similarity calculation results, may further include: a fault feature similarity calculation unit for sequentially calculating the fault feature similarity between the first separated signal in the first power data source and all m separated signals in the remaining power data source; and a fault feature similarity calculation result output unit for similarly obtaining k×m fault feature similarities between k separated signals in the first power data source, and outputting the k×m fault feature similarities as the fault feature similarity calculation results, where k and m are both positive integers greater than or equal to 0.

[0074] The specific configuration of the power grid fault diagnosis module 40 will be described in detail below. As mentioned above, cross-source clustering is performed according to the fault feature similarity calculation results until multiple power data source ends are traversed and multiple separate signal clusters are output. The power grid fault diagnosis module 40 may further include: a cluster division unit for setting a fault similarity threshold, and according to the fault similarity threshold, sequentially traversing the fault feature similarity of the separate signal pairs of the multiple power data source ends, dividing the separate signal pairs with a similarity greater than the fault similarity threshold into the same cluster, and outputting multiple separate signal clusters.

[0075] The power grid fault diagnosis module 40 may further include: a power grid fault diagnosis unit for inputting multiple output separate signal clusters into a power grid fault diagnosis model for analysis, wherein the power grid fault diagnosis model includes a Softmax single-label classification layer, and the fault type distribution probability of each separate signal cluster is output by the Softmax single-label classification layer, and the first fault type of each separate signal cluster is determined according to the fault type distribution probability, wherein the first fault type is the class label with the highest probability.

[0076] The power grid fault diagnosis system based on power data analysis provided in this embodiment of the invention can execute the power grid fault diagnosis method based on power data analysis provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0077] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0078] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A power data analysis based power grid fault diagnosis method, characterized in that, The method comprises: A plurality of power data source ends of a power network are constructed, and a plurality of power data corresponding to the plurality of power data source ends are extracted; A blind signal separation algorithm is used to semi-supervisedly separate the plurality of power data according to fault types, to obtain a plurality of separated signals corresponding to each power data source end; Fault feature similarities between separated signal pairs of the plurality of power data source ends are calculated, to obtain fault feature similarity calculation results; Cross-source end clustering is performed according to the fault feature similarity calculation results, until a plurality of separated signal clusters are output after traversing the plurality of power data source ends, the plurality of separated signal clusters are input into a power grid fault diagnosis model for analysis, a plurality of fault types corresponding to the plurality of separated signal clusters are obtained, and after the plurality of fault types are jointly identified, a power grid fault diagnosis result is sent; The blind signal separation algorithm is used to separate the plurality of power data according to fault types, and the method further comprises: A power signal conversion template is preset, and signal resampling processing is performed on the plurality of power data through the power signal conversion template, to obtain a plurality of resampled power signals; The plurality of resampled power signals are mapped in a numerical interval preset in the power signal conversion template, to obtain a plurality of standardized power signals; The blind signal separation algorithm is used to semi-supervisedly separate the plurality of standardized power signals according to fault types, to obtain a plurality of separated signals corresponding to each power data source end; The blind signal separation algorithm is used to semi-supervisedly separate the plurality of standardized power signals according to fault types, and the method comprises: A preset fault type set corresponding to any power data source end is obtained, wherein each fault type in the preset fault type set comprises a fault signal feature template; A fault type filter set is trained according to the fault type set and a corresponding fault signal feature template set, the fault type filter set comprises a plurality of fault type filters corresponding to the plurality of power data source ends, each fault type filter is a multi-channel filter, and each channel corresponds to a fault type; A fault type filter is called to perform frequency band filtering on the standardized power signal corresponding to each power data source end, to obtain a plurality of separated signals corresponding to each power data source end; Fault feature similarities between separated signal pairs of the plurality of power data source ends are calculated, wherein the fault features include time-frequency features, waveform distance features, and fault mode embedding features, the time-frequency features are two-dimensional time-frequency graphs of the separated signal pairs based on time-frequency distribution, the waveform distance features are DTW waveform distances of the separated signal pairs, and the fault mode embedding features are high-level semantic fault features representing the separated signal pairs; The fault feature similarities of the separated signal pairs are obtained by weighted calculation on the time-frequency features, the waveform distance features, and the fault mode embedding features.

2. The power data analysis based power grid fault diagnosis method of claim 1, wherein, Each fault type in the preset fault type set corresponds to a fault signal feature template, and the fault signal feature template comprises a frequency range, a main energy distribution, a harmonic bit feature, an instantaneous jump intensity, and a reference waveform form.

3. The power data analysis based power grid fault diagnosis method of claim 1, wherein, The fault type filter is called to perform frequency band filtering on the standardized power signal corresponding to each power data source end, and the method comprises: Obtain a fault type filter matched with each power data source end; The matched fault type filter performs band convolution processing on the standardized power signal corresponding to each power data source end to obtain a frequency band response component corresponding to each fault type. According to the frequency band response component corresponding to each fault type, a plurality of separated signals corresponding to each power data source end are obtained.

4. The power data analysis based power grid fault diagnosis method of claim 1, wherein, The method for outputting the fault feature similarity calculation result comprises: In turn, m fault feature similarities between the first separated signal in the first power data source end and all m separated signals in the remaining power data source ends are calculated. In this way, k*m fault feature similarities of k separated signals in the first power data source end are obtained, and the k*m fault feature similarities are output as the fault feature similarity calculation result. k and m are positive integers greater than or equal to 0.

5. The power data analysis based power grid fault diagnosis method of claim 1, wherein, According to the fault feature similarity calculation result, cross-source end clustering is performed until a plurality of separated signal clusters are output after traversing the plurality of power data source ends. The method comprises: A fault similarity threshold is set, and the fault feature similarities of the separated signal pairs of the plurality of power data source ends are traversed in turn according to the fault similarity threshold; Separated signal pairs greater than the fault similarity threshold are divided into the same clustering cluster, and a plurality of separated signal clusters are output.

6. The power data analysis based power grid fault diagnosis method of claim 1, wherein, The output plurality of separated signal clusters are input into a power grid fault diagnosis model for analysis, wherein the power grid fault diagnosis model comprises a Softmax single-label classification layer. According to the fault type distribution probability output by the Softmax single-label classification layer, the first fault type of each separated signal cluster is determined, and the first fault type is the class label with the maximum probability.

7. A power data analysis based power grid fault diagnosis system characterized in that, The system is used to implement the power data analysis-based power grid fault diagnosis method of any one of claims 1-6, and the system comprises: A power data extraction module is configured to construct a plurality of power data source ends of a power network and extract a plurality of power data corresponding to the plurality of power data source ends. A semi-supervised separation module is configured to perform semi-supervised separation on the plurality of power data according to fault types by using a blind signal separation algorithm to obtain a plurality of separated signals corresponding to each power data source end. A fault feature similarity calculation module is configured to calculate fault feature similarities of separated signal pairs between the plurality of power data source ends to obtain fault feature similarity calculation results. A power grid fault diagnosis module is configured to perform cross-source end clustering according to the fault feature similarity calculation results until a plurality of separated signal clusters are output after traversing the plurality of power data source ends. The output plurality of separated signal clusters are input into a power grid fault diagnosis model for analysis to obtain a plurality of fault types corresponding to the plurality of separated signal clusters. After joint identification of the plurality of fault types, a power grid fault diagnosis result is sent.

Images