Fault monitoring method, device, equipment and storage medium for power distribution cabinet
Through wavelet packet decomposition and XGBoost integrated learning algorithm, combined with gray correlation, efficient detection and positioning of distribution cabinet fault monitoring is achieved, and the problems of low fault detection efficiency and poor accuracy in the existing technology are solved, and the fault of distribution cabinet is improved Reliability and timeliness of processing.
Patent Information
- Application Number
- CN202410827964.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-25
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-06-25
AI Technical Summary
The existing distribution cabinet fault monitoring methods rely on manual inspection and traditional protection devices, which are inefficient and difficult to detect early micro faults in a timely manner, leading to the accumulation of fault hazards and affecting the reliability and safety of the power system.
Discrete wavelet transformation and soft threshold denoising methods remove high-frequency noise in historical current signals, extract spectral energy and current direction characteristics through wavelet packet decomposition, combine with XGBoost integrated learning algorithm to build a fault detection model, realize real-time fault detection, and pass grey Correlation degree accurately locates the fault point.
It improves the accuracy and robustness of the fault detection of the distribution cabinet, realizes remote fault monitoring of the distribution cabinet, promptly detects faults, and improves the accuracy and timelyness of fault handling.
Smart Images

Figure CN118468087B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of fault detection technology, and in particular to a method, device, equipment and storage medium for monitoring faults in a power distribution cabinet. Background Art
[0002] The distribution cabinet is responsible for the distribution and transmission of electrical energy in the power system. However, due to its complex structure and operating environment, the distribution cabinet is susceptible to various faults, such as overload, short circuit and aging of electrical equipment. These faults will not only cause unplanned shutdown of the power system and affect the reliability of power supply, but may also cause safety accidents.
[0003] At present, fault monitoring of distribution cabinets mainly relies on manual inspections and traditional protection devices. These methods have obvious limitations. Manual inspections have long cycles, low efficiency, and difficulty in timely detection of faults. Traditional protection devices can usually only detect more serious faults, and are difficult to detect for some early, minor faults, leading to the accumulation of potential fault hazards, which may eventually lead to major accidents. When faced with complex current signals, existing technologies are often unable to effectively distinguish between normal fluctuations and fault signals, and are prone to false alarms or missed alarms, affecting the accuracy and reliability of fault monitoring. Summary of the invention
[0004] The present application provides a distribution cabinet fault monitoring method, device, equipment and storage medium for improving the accuracy of distribution cabinet fault detection and thereby realizing remote fault monitoring of the distribution cabinet.
[0005] In a first aspect, the present application provides a method for monitoring a fault in a power distribution cabinet, the method comprising:
[0006] Collecting historical current signals of the power distribution cabinet, and removing high-frequency noise and harmonic interference from the historical current signals to obtain standard current signals;
[0007] Performing wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales, and calculating the spectrum energy characteristics and current direction characteristics of each sub-band signal to generate a full-frequency domain fault feature vector of each sub-band signal;
[0008] A comprehensive fault detection criterion of spectrum energy and current direction is established and used as the objective function, the full-frequency domain fault feature vector is input into the XGBoost integrated learning algorithm for model training, and a distribution cabinet fault detection model is constructed;
[0009] Acquire a real-time current signal of the power distribution cabinet, and input the real-time current signal into the power distribution cabinet fault detection model to perform fault detection, thereby obtaining a real-time fault detection result;
[0010] A fault feature sequence is constructed according to the real-time fault detection result, and the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point is calculated to generate fault point location information.
[0011] In a second aspect, the present application provides a distribution cabinet fault monitoring device, the distribution cabinet fault monitoring device comprising:
[0012] The acquisition module is used to collect the historical current signal of the power distribution cabinet, and remove the high-frequency noise and harmonic interference of the historical current signal to obtain a standard current signal;
[0013] A decomposition module, used to perform wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales, and calculate the spectrum energy characteristics and current direction characteristics of each sub-band signal to generate a full-frequency domain fault feature vector of each sub-band signal;
[0014] A training module is used to establish a comprehensive fault detection criterion of spectrum energy and current direction and use it as an objective function, input the full-frequency domain fault feature vector into the XGBoost integrated learning algorithm for model training, and build a distribution cabinet fault detection model;
[0015] A detection module, used to obtain a real-time current signal of the power distribution cabinet, and input the real-time current signal into the power distribution cabinet fault detection model to perform fault detection, so as to obtain a real-time fault detection result;
[0016] The generation module is used to construct a fault feature sequence according to the real-time fault detection result, and calculate the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point to generate fault point location information.
[0017] The third aspect of the present application provides a distribution cabinet fault monitoring device, including: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the distribution cabinet fault monitoring device executes the above-mentioned distribution cabinet fault monitoring method.
[0018] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned distribution cabinet fault monitoring method.
[0019] In the technical solution provided by the present application, the discrete wavelet transform and soft threshold denoising methods are used to effectively remove high-frequency noise in the historical current signal and reduce harmonic interference. The standard current signal is decomposed into sub-band signals of different frequency scales through wavelet packet decomposition technology, and the spectrum energy characteristics and current direction characteristics of each sub-band signal are extracted to generate a full-frequency domain fault feature vector. The full-frequency domain fault feature vector is trained using the XGBoost integrated learning algorithm to build a distribution cabinet fault detection model. The accuracy and robustness of fault detection are improved by combining the comprehensive fault detection criteria of spectrum energy and current direction, multi-level thresholds and fuzzy processing. The detection model is parameterized and cross-validated to ensure the reliability and generalization ability of the model under different fault conditions. The current signal of the distribution cabinet is sampled in real time, and wavelet packet decomposition and spectrum analysis are performed on it to extract the real-time fault feature vector. Using the trained fault detection model, the real-time current signal is quickly detected, the fault is discovered in time, and the real-time fault detection result is provided. The fault feature sequence is constructed according to the real-time fault detection result, and the gray correlation degree between the reference feature sequence of the candidate fault point is calculated. Through weight coefficient calculation and fuzzy cluster analysis, the fault point can be accurately located, and the fault point location information can be generated, which can improve the accuracy and timeliness of fault handling, thereby realizing remote fault monitoring of the distribution cabinet. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0021] Figure 1 A schematic diagram of an embodiment of a method for monitoring a fault of a power distribution cabinet in an embodiment of the present application;
[0022] Figure 2 It is a schematic diagram of an embodiment of a distribution cabinet fault monitoring device in an embodiment of the present application. DETAILED DESCRIPTION
[0023] Embodiments of the present application provide a method, device, equipment and storage medium for monitoring faults in a power distribution cabinet. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0024] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the method for monitoring faults of a power distribution cabinet in an embodiment of the present application includes:
[0025] Step S101, collecting historical current signals of the power distribution cabinet, and removing high-frequency noise and harmonic interference from the historical current signals to obtain standard current signals;
[0026] It is understandable that the execution subject of the present application may be a distribution cabinet fault monitoring device, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.
[0027] Specifically, the historical current signal of the distribution cabinet is collected, and the historical current signal is subjected to discrete wavelet transform to obtain multi-scale wavelet coefficients. Discrete wavelet transform can effectively decompose the signal into different frequency scales, so that high-frequency noise can be separated at a high scale. The threshold of each scale is calculated according to the multi-scale wavelet coefficients, and the multi-scale wavelet coefficients are subjected to soft threshold denoising to obtain the denoised wavelet coefficients. The soft threshold denoising suppresses the noise component by setting the threshold and retains the main component of the signal. The denoised wavelet coefficients are subjected to wavelet reconstruction to obtain the first current signal. The wavelet reconstruction recombines the multi-scale wavelet coefficients and restores them to the form of a time domain signal. The first current signal is subjected to fast Fourier transform to obtain a frequency domain current signal. Fast Fourier transform converts the time domain signal into a frequency domain signal, so that the harmonic component is more obvious and easy to identify. Harmonic components are identified according to the frequency domain current signal, and the harmonic components are subjected to amplitude attenuation processing to obtain a second current signal. The amplitude attenuation processing is used to reduce the influence of the harmonic components and obtain a smoother and more accurate current signal. The second current signal is subjected to an inverse fast Fourier transform to obtain a time domain current signal, and the time domain current signal is subjected to a dynamic state estimation to obtain a filtered current signal. The inverse fast Fourier transform reconverts the frequency domain signal into a time domain signal, and the dynamic state estimation removes residual noise and interference through a filtering algorithm, making the current signal more accurate and reliable. The filtered current signal is subjected to an empirical mode decomposition to obtain several intrinsic mode functions, and a standard current signal is obtained by reconstructing the several intrinsic mode functions. Empirical mode decomposition is an adaptive signal processing method that decomposes a complex signal into several physically meaningful intrinsic mode functions, each of which represents a different frequency component of the signal. By reconstructing the intrinsic mode functions, noise and interference are effectively removed to obtain a standard current signal.
[0028] Step S102, performing wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales, and calculating the spectrum energy characteristics and current direction characteristics of each sub-band signal to generate a full-frequency domain fault feature vector of each sub-band signal;
[0029] Specifically, the standard current signal is decomposed by multi-layer wavelet packet to obtain sub-band signals of different frequency scales. Through multi-layer wavelet packet decomposition, the standard current signal is decomposed into different frequency scales, each scale corresponds to a sub-band signal, and these sub-band signals contain information of the original signal in different frequency ranges. The energy entropy of each sub-band signal is calculated to obtain the sub-band energy entropy feature. Energy entropy is an effective feature used to measure the energy distribution of a signal at different frequency scales. By calculating the energy entropy of each sub-band signal, the energy characteristics of the signal at different frequency scales are extracted. Hilbert transform is performed on each sub-band signal to obtain the instantaneous phase and instantaneous amplitude of each sub-band signal. Hilbert transform can convert a real signal into an analytical signal, thereby obtaining the instantaneous phase and instantaneous amplitude of the signal. According to the instantaneous phase and instantaneous amplitude, the sub-band instantaneous frequency feature of each sub-band signal is calculated. The instantaneous frequency is the frequency value of the signal at a certain moment, reflecting the local frequency change of the signal. Short-time Fourier transform is performed on each sub-band signal to obtain a time-frequency distribution matrix. Short-time Fourier transform is a commonly used time-frequency analysis method. Through the sliding window function, the distribution of the signal at different times and frequencies is obtained. According to the time-frequency distribution matrix, the sub-band spectrum shape characteristics of each sub-band signal are calculated. The spectrum shape characteristics reflect the distribution shape of the signal in the frequency domain and are important spectrum characteristics of the signal. Zero-crossing rate analysis is performed on each sub-band signal to obtain the sub-band zero-crossing rate characteristics. The zero-crossing rate refers to the number of times a signal passes through the zero point per unit time, reflecting the frequency component and waveform characteristics of the signal. The waveform characteristics of the signal are obtained by calculating the zero-crossing rate. The peak factor and margin factor of each sub-band signal are calculated to obtain the sub-band waveform characteristics. The peak factor is the ratio of the maximum value of the signal to the root mean square value, and the margin factor is the ratio of the peak value of the signal to the average value. They are both waveform characteristics of the signal. Singular value decomposition is performed on each sub-band signal to obtain the sub-band singular value characteristics. Singular value decomposition is a matrix decomposition method that extracts the main component characteristics of the signal and obtains the singular value characteristics of the signal through singular value decomposition.
[0030] Step S103: Establish a comprehensive fault detection criterion of spectrum energy and current direction and use it as the objective function, input the full-frequency domain fault feature vector into the XGBoost integrated learning algorithm for model training, and build a distribution cabinet fault detection model;
[0031] Specifically, the full-frequency fault feature vector is normalized to obtain a standardized feature vector. Normalization can eliminate the differences between different feature dimensions, so that all feature values are in the same order of magnitude, which helps to improve the effect of model training. Then, the spectrum energy density and current direction angle are calculated according to the standardized feature vector to obtain the spectrum energy feature and current direction feature respectively. The spectrum energy density reflects the energy distribution of the signal in the frequency domain, and the current direction angle reflects the phase characteristics of the current signal. The spectrum energy feature and the current direction feature are weighted fused to obtain a comprehensive fault detection index. Multiple features are combined according to certain weights to comprehensively reflect the characteristics of the fault. Multi-level thresholds are set according to the comprehensive fault detection index to construct a multi-category fault classification criterion. The multi-level threshold can classify faults into different levels, thereby classifying the fault types. The multi-category fault classification criterion is fuzzified to obtain a comprehensive fault detection criterion. Fuzzification can make the classification criterion more flexible and reduce the possibility of misclassification. The objective function is constructed according to the comprehensive fault detection criterion, and the objective function is solved by convex optimization to obtain the optimal parameters. The objective function is the core of model optimization. Through convex optimization, we can find the parameter combination that makes the objective function reach the optimal value, thereby improving the accuracy of the fault detection model. The standardized feature vectors are randomly sampled to construct multiple sub-training sets. Random sampling can increase the diversity of data and avoid overfitting. Multiple decision trees are trained according to multiple sub-training sets to obtain a set of base learners. Decision tree is a commonly used classification algorithm. By training multiple decision trees, the generalization ability of the model is improved. The base learner set is integrated to construct the XGBoost ensemble learning algorithm. XGBoost is an efficient gradient boosting tree algorithm. By integrating multiple base learners, the prediction performance of the model is improved. According to the optimized parameters, the XGBoost ensemble learning algorithm is parameter tuned and cross-validated to obtain the distribution cabinet fault detection model. Parameter tuning is to find the most suitable parameter combination for the model, while cross-validation is to evaluate the performance of the model and prevent overfitting.
[0032] Step S104, obtaining a real-time current signal of the power distribution cabinet, and inputting the real-time current signal into a fault detection model of the power distribution cabinet to perform fault detection, thereby obtaining a real-time fault detection result;
[0033] Specifically, the current signal of the distribution cabinet is sampled in real time to obtain the real-time current signal. The real-time current signal is decomposed by wavelet packets to obtain multiple sub-band signals of the real-time current signal. Wavelet packet decomposition can decompose the real-time current signal into sub-band signals of different frequency scales and capture the characteristics of the signal in each frequency band. Spectral analysis is performed on the multiple sub-band signals of the real-time current signal to calculate the spectrum energy feature vector of each sub-band signal in the real-time current signal. Spectral analysis can reveal the energy distribution of the signal in the frequency domain and extract the spectrum energy characteristics of each sub-band signal. At the same time, instantaneous phase analysis is performed on the multiple sub-band signals of the real-time current signal to calculate the current direction feature vector of each sub-band signal in the real-time current signal. Instantaneous phase analysis helps to capture the phase change characteristics of the signal in the time domain, which reflect the direction change information of the current signal. The spectrum energy feature vector and the current direction feature vector are combined to obtain the real-time fault feature vector. The real-time fault feature vector is input into the distribution cabinet fault detection model, and the forward reasoning process is started. The real-time fault feature vector is branched and judged by each decision tree in the distribution cabinet fault detection model to obtain the sub-fault detection result of each decision tree. Each decision tree is inferred according to the real-time fault feature vector and outputs the corresponding sub-fault detection result. The sub-fault detection result of each decision tree is subjected to fuzzy reasoning and weighted fusion to obtain the real-time fault detection result. Fuzzy reasoning can handle the uncertainty in the sub-fault detection result, making the final fault detection result more robust and accurate. Weighted fusion is to integrate each sub-fault detection result according to a certain weight to obtain the final fault detection result.
[0034] Extract the first eigenvalue from the real-time fault feature vector, compare the first eigenvalue with the split threshold of the root node of the decision tree, and select to enter the left subtree or the right subtree according to the comparison result. In the selected subtree, extract the next eigenvalue of the real-time fault feature vector, compare the eigenvalue with the split threshold of the current node, and recursively select to enter the left subtree or the right subtree until the leaf node is reached. Through recursive comparison, ensure that the eigenvalue is branched along the correct path of the decision tree. When the leaf node of the decision tree is reached, the fault type corresponding to the leaf node is used as the first fault detection result of the decision tree. For each eigenvalue in the real-time fault feature vector, repeat the above steps, pass through each node of the decision tree in turn, and obtain the second fault detection result of the decision tree. Through repeated operations, ensure that each eigenvalue passes through the complete path of the decision tree and obtain an accurate detection result. According to the structure of the decision tree and the node splitting rule, construct the logical judgment expression of the decision tree, and substitute each eigenvalue of the real-time fault feature vector into the logical judgment expression, calculate and obtain the third fault detection result of the decision tree, simplify the branch judgment process of the eigenvalue, and improve the efficiency and accuracy of detection. Perform feature importance analysis on the real-time fault feature vector, calculate the influence weight of each eigenvalue and generate the feature importance weight vector of the decision tree. Feature importance analysis can reveal the contribution of each eigenvalue in the fault detection process and provide a basis for subsequent weighted combination. Perform weighted combination on the third fault detection result of the decision tree and the feature importance weight vector to obtain the sub-fault detection result of the decision tree. Through weighted combination, the influence of each eigenvalue is comprehensively considered to obtain more accurate detection results. Repeat the above steps for each decision tree in the distribution cabinet fault detection model to obtain the sub-weighted fault detection results of all decision trees. Through repeated operations, ensure that the detection results of each decision tree are fully utilized to improve the detection accuracy of the overall model.
[0035] Step S105: construct a fault feature sequence according to the real-time fault detection result, and calculate the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point to generate fault point location information.
[0036] Specifically, the real-time fault detection results are subjected to time series analysis and feature extraction to obtain a fault feature sequence. Through time series analysis, the changing law of the fault signal in the time dimension is captured, and feature extraction is to extract representative fault features from it to form a fault feature sequence. A reference feature sequence of candidate fault points is constructed according to the historical fault database. The historical fault database contains a large amount of fault data. By analyzing these data, feature sequences of different fault points are extracted as reference feature sequences of candidate fault points. The grey correlation coefficient matrix between the fault feature sequence and the reference feature sequence of the candidate fault point is calculated. Grey correlation is a method to measure the similarity between two sequences. By calculating the grey correlation coefficient matrix, the correlation between the fault feature sequence and the reference feature sequence of the candidate fault point is quantified. The weight coefficient of the candidate fault point in the grey correlation coefficient matrix is calculated to obtain the weighted correlation of each candidate fault point. The calculation of the weight coefficient is based on the importance of each candidate fault point in fault detection, and different weights are given to obtain a more accurate weighted correlation. The correlation threshold is set according to the weighted correlation, and the candidate fault points are screened to obtain the initial positioning result. The initial positioning result is subjected to fuzzy clustering analysis to obtain the cluster center of the fault point. Fuzzy cluster analysis is a method based on similarity measurement. It clusters the fault points in the initial location results to obtain several fault point cluster centers, which represent the concentrated distribution areas of the fault points. The fault point cluster centers are matched with the fault feature sequence to locate the position of the fault point and generate the fault point location information.
[0037] In the embodiment of the present application, the discrete wavelet transform and soft threshold denoising method are used to effectively remove the high-frequency noise in the historical current signal and reduce the harmonic interference. The standard current signal is decomposed into sub-band signals of different frequency scales by wavelet packet decomposition technology, and the spectrum energy characteristics and current direction characteristics of each sub-band signal are extracted to generate a full-frequency domain fault feature vector. The full-frequency domain fault feature vector is trained by the XGBoost integrated learning algorithm to build a distribution cabinet fault detection model. The accuracy and robustness of fault detection are improved by combining the comprehensive fault detection criteria of spectrum energy and current direction, multi-level thresholds and fuzzy processing. The detection model is parameterized and cross-validated to ensure the reliability and generalization ability of the model under different fault conditions. The current signal of the distribution cabinet is sampled in real time, and wavelet packet decomposition and spectrum analysis are performed on it to extract the real-time fault feature vector. Using the trained fault detection model, the real-time current signal is quickly detected, the fault is discovered in time, and the real-time fault detection result is provided. The fault feature sequence is constructed according to the real-time fault detection result, and the gray correlation degree between the reference feature sequence of the candidate fault point is calculated. Through weight coefficient calculation and fuzzy cluster analysis, the fault point can be accurately located, and the fault point location information can be generated, which can improve the accuracy and timeliness of fault handling, thereby realizing remote fault monitoring of the distribution cabinet.
[0038] In a specific embodiment, the process of executing step S101 may specifically include the following steps:
[0039] (1) Collect the historical current signal of the distribution cabinet and perform discrete wavelet transform on the historical current signal to obtain multi-scale wavelet coefficients;
[0040] (2) Calculate the threshold of each scale according to the multi-scale wavelet coefficients, perform soft threshold denoising on the multi-scale wavelet coefficients, and obtain the denoised wavelet coefficients;
[0041] (3) performing wavelet reconstruction on the denoised wavelet coefficients to obtain a first current signal, and performing fast Fourier transform on the first current signal to obtain a frequency domain current signal;
[0042] (4) Identifying harmonic components according to the frequency domain current signal, performing amplitude attenuation processing on the harmonic components, and obtaining a second current signal;
[0043] (5) performing an inverse fast Fourier transform on the second current signal to obtain a time domain current signal, and performing dynamic state estimation on the time domain current signal to obtain a filtered current signal;
[0044] (6) Performing empirical mode decomposition on the filtered current signal to obtain several intrinsic mode functions, and reconstructing the standard current signal based on the several intrinsic mode functions.
[0045] Specifically, the historical current signal of the power distribution cabinet is collected. Assume that within a period of time, the collected current signal is , which is a function that changes over time. The discrete wavelet transform is performed on this signal. The discrete wavelet transform is a tool for analyzing the multi-scale characteristics of the signal. The discrete wavelet transform can decompose the signal into wavelet coefficients of different frequency scales. The formula is as follows:
[0046] ;
[0047] in, Representative The wavelet coefficients of the scale, is the wavelet basis function. Through discrete wavelet transform, we get multi-scale wavelet coefficients of different scales. We calculate the threshold of each scale based on the multi-scale wavelet coefficients. The selection of threshold is usually based on the statistical characteristics of noise. For example, assuming The wavelet coefficients of the scale are It can be calculated by the following formula:
[0048] ;
[0049] in, represents the standard deviation of the noise, Represents the length of the signal. Next, the multi-scale wavelet coefficients are subjected to soft threshold denoising to obtain the denoised wavelet coefficients. The formula for soft threshold processing is:
[0050] ;
[0051] in, Represents the wavelet coefficient after denoising. Through soft threshold processing, the noise component in the signal is effectively removed. The wavelet coefficient after denoising is reconstructed by wavelet to obtain the first current signal . Recombine the wavelet coefficients of different scales and restore them to the form of time domain signals. The formula is as follows:
[0052] ;
[0053] Perform fast Fourier transform on the first current signal to obtain a frequency domain current signal Fast Fourier transform can convert time domain signals into frequency domain signals. The formula is:
[0054] ;
[0055] in, Indicates frequency, is an imaginary unit. Harmonic components are identified based on the frequency domain current signal. Harmonic components are the larger amplitude parts of certain specific frequency components in the frequency domain signal. After identifying these harmonic components, they are subjected to amplitude attenuation processing to obtain the second current signal The formula for amplitude attenuation processing is:
[0056] ;
[0057] in, is an attenuation function, which is usually designed to have a larger attenuation at harmonic frequencies and a smaller attenuation at other frequencies. Perform an inverse fast Fourier transform on the second current signal to obtain a time domain current signal The formula for the inverse fast Fourier transform is:
[0058] ;
[0059] The dynamic state estimation is performed on the time domain current signal to obtain the filtered current signal. Dynamic state estimation is a real-time filtering technology that can effectively remove the dynamic noise components in the signal. Assume that the filtered signal is . Perform empirical mode decomposition on the filtered current signal to obtain several intrinsic mode functions . Empirical mode decomposition is an adaptive signal processing method that can decompose complex signals into several physically meaningful mode functions. The formula is:
[0060] ;
[0061] in, is the residual component. The standard current signal is reconstructed based on several intrinsic mode functions. The reconstructed formula is:
[0062] ;
[0063] Through the above steps, a standard current signal is finally obtained.
[0064] In a specific embodiment, the process of executing step S102 may specifically include the following steps:
[0065] (1) Perform multi-layer wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales;
[0066] (2) Calculate the energy entropy of each subband signal to obtain the subband energy entropy feature, and perform Hilbert transform on each subband signal to obtain the instantaneous phase and instantaneous amplitude of each subband signal;
[0067] (3) Calculate the sub-band instantaneous frequency characteristics of each sub-band signal based on the instantaneous phase and instantaneous amplitude;
[0068] (4) Perform short-time Fourier transform on each sub-band signal to obtain a time-frequency distribution matrix, and calculate the sub-band spectrum shape characteristics of each sub-band signal based on the time-frequency distribution matrix;
[0069] (5) Perform zero-crossing rate analysis on each sub-band signal to obtain sub-band zero-crossing rate characteristics, calculate the peak factor and margin factor of each sub-band signal to obtain sub-band waveform characteristics, and perform singular value decomposition on each sub-band signal to obtain sub-band singular value characteristics;
[0070] (6) Generate a full-frequency domain fault feature vector for each sub-band signal based on the sub-band energy entropy characteristics, sub-band instantaneous frequency characteristics, sub-band spectrum shape characteristics, sub-band zero-crossing rate characteristics, sub-band waveform characteristics, and sub-band singular value characteristics.
[0071] Specifically, the standard current signal is decomposed by multi-layer wavelet packet. Assume that the standard current signal is , through wavelet packet decomposition, Decompose the signal into subbands at multiple frequency scales ,in represents the decomposition level, Indicated in The first subbands. The mathematical expression of wavelet packet decomposition is as follows:
[0072] ;
[0073] in, is the wavelet packet basis function. By decomposing, the subband signal on each frequency scale is obtained. The energy entropy of each subband signal is calculated separately. Energy entropy is an indicator to measure the complexity of the signal, and the calculation formula is:
[0074] ;
[0075] in, Indicates The energy entropy of the subband signal is Indicates The subband signal is Energy distribution at each moment. Perform Hilbert transform on each subband signal to obtain the instantaneous phase and instantaneous amplitude. Hilbert transform converts the real signal into an analytical signal, which can be expressed as:
[0076] ;
[0077] Among them, H yes By analyzing the signal, we can get the instantaneous phase and the instantaneous amplitude
[0078] ;
[0079] According to the instantaneous phase and instantaneous amplitude, the instantaneous frequency characteristics of the sub-band are calculated. The calculation formula of the instantaneous frequency is:
[0080] ;
[0081] Perform short-time Fourier transform (STFT) on each subband signal to obtain the time-frequency distribution matrix. The formula of STFT is:
[0082] ;
[0083] in, is a window function, represents the time offset, Represents frequency. The sub-band spectrum shape characteristics are calculated through the time-frequency distribution matrix. Zero crossing rate analysis is performed on each sub-band signal. The zero crossing rate is the number of times the signal passes through the zero point per unit time. The calculation formula is:
[0084] ;
[0085] in, is the zero crossing rate, is the total duration of the signal, sign yes At the moment The sign function of . Calculate the peak factor and margin factor of the subband signal. The peak factor is the ratio of the signal peak value to the root mean square value, and the margin factor is the ratio of the signal peak value to the average value. The calculation formulas are:
[0086] ;
[0087] ;
[0088] Where PF is the peak factor, MF is the margin factor, is the maximum value of the signal, is the number of samples of the signal, is the RMS value of the signal, is the average value of the signal. Perform singular value decomposition (SVD) on each subband signal to obtain the subband singular value features. The formula of SVD is:
[0089] ;
[0090] in, is the matrix representation of the subband signal, and is an orthogonal matrix, It is a diagonal matrix, and the diagonal elements are singular values. Through the above steps, the sub-band energy entropy characteristics, sub-band instantaneous frequency characteristics, sub-band spectrum shape characteristics, sub-band zero crossing rate characteristics, sub-band waveform characteristics and sub-band singular value characteristics of each sub-band signal are calculated respectively. Based on these characteristics, the full frequency domain fault feature vector of each sub-band signal is generated.
[0091] In a specific embodiment, the process of executing step S103 may specifically include the following steps:
[0092] (1) Normalize the full-frequency fault feature vector to obtain a standardized feature vector, and calculate the spectrum energy density and current direction angle based on the standardized feature vector to obtain the spectrum energy feature and current direction feature;
[0093] (2) Perform weighted fusion of spectrum energy features and current direction features to obtain comprehensive fault detection indicators;
[0094] (3) According to the comprehensive fault detection index, multi-level thresholds are set to construct multi-category fault classification criteria, and the multi-category fault classification criteria are fuzzy processed to obtain the comprehensive fault detection criteria;
[0095] (4) Construct an objective function based on the comprehensive fault detection criteria, and perform convex optimization on the objective function to obtain the optimal parameters;
[0096] (5) Randomly sample the standardized feature vectors, construct multiple sub-training sets, and train multiple decision trees based on the multiple sub-training sets to obtain a base learner set;
[0097] (6) The base learner set is integrated to construct the XGBoost ensemble learning algorithm. The XGBoost ensemble learning algorithm is then parameter tuned and cross-validated according to the optimized parameters to obtain the distribution cabinet fault detection model.
[0098] Specifically, the full-frequency fault feature vector is normalized to eliminate the differences between different feature dimensions, so that all feature values are at the same order of magnitude, and a standardized feature vector is obtained. The spectrum energy density and current direction angle are calculated based on the standardized feature vector. The spectrum energy density is the distribution of frequency domain signal energy within the unit frequency range, and the calculation formula is:
[0099] ;
[0100] in, represents the spectral energy density, is the Fourier transform of the normalized eigenvector, is the length of the signal. The current direction angle can be calculated by normalizing the instantaneous phase of the eigenvector, assuming that the instantaneous phase is , then the current direction angle It can be expressed as:
[0101] ;
[0102] in, is the period of the signal. The spectrum energy characteristics and current direction characteristics are weighted and fused to obtain the comprehensive fault detection index. The formula for weighted fusion is:
[0103] ;
[0104] in, represents the comprehensive fault detection index, and are the weights of the spectrum energy feature and the current direction feature, respectively. According to the comprehensive fault detection index, a multi-level threshold is set, a multi-category fault classification criterion is constructed, and the multi-category fault classification criterion is fuzzy processed. Assume that the multi-level threshold is , then the multi-category fault classification criterion can be expressed as:
[0105] ;
[0106] Through fuzzification, these classification criteria are smoothed to make the classification results more robust. For example, a fuzzy membership function is used to describe the membership of each classification. The objective function is constructed according to the comprehensive fault detection criteria, and the objective function is solved by convex optimization to obtain the optimal parameters. Assume that the objective function is ,in is the weight vector to be optimized, and the objective function can be in the form of:
[0107] ;
[0108] in, is the loss function, which represents the comprehensive fault detection index With the true label The error between is the regularization parameter. Through convex optimization, find the objective function The smallest weight vector w. Randomly sample the standardized feature vectors to construct multiple sub-training sets. By randomly sampling the original training set, multiple sub-training sets are obtained, which are used to train multiple decision trees. Each sub-training set contains the feature vectors and corresponding labels of some samples in the original training set. Train multiple decision trees based on multiple sub-training sets to obtain a base learner set. Assume that M decision trees are trained, and each decision tree forms a base learner by learning its sub-training set. A decision tree is a classifier based on feature partitioning. It constructs a tree-structured model by recursively partitioning the feature space. Integrate the base learner set to construct the XGBoost ensemble learning algorithm. XGBoost is an efficient gradient boosting tree algorithm. By integrating multiple base learners, the generalization ability and prediction performance of the model can be improved. XGBoost gradually constructs decision trees, and each decision tree fits the residual of the previous tree, thereby continuously reducing the error of the model. Parameter tuning and cross-validation of the XGBoost ensemble learning algorithm are performed according to the optimized parameters. The purpose of parameter tuning is to find the most suitable hyperparameter combination for the model, such as learning rate, tree depth, regularization parameter, etc. Cross-validation is a method for evaluating model performance by dividing the training set into multiple subsets, using one of the subsets as the validation set and the other subsets as the training set, training multiple models and calculating their average performance.
[0109] In a specific embodiment, the process of executing step S104 may specifically include the following steps:
[0110] (1) Sampling the current signal of the power distribution cabinet in real time to obtain the real-time current signal, and performing wavelet packet decomposition on the real-time current signal to obtain multiple sub-band signals of the real-time current signal;
[0111] (2) Performing spectrum analysis on multiple sub-band signals of the real-time current signal and calculating the spectrum energy feature vector of each sub-band signal in the real-time current signal;
[0112] (3) Perform instantaneous phase analysis on multiple sub-band signals of the real-time current signal and calculate the current direction characteristic vector of each sub-band signal in the real-time current signal;
[0113] (4) Combining the spectrum energy feature vector and the current direction feature vector to obtain the real-time fault feature vector;
[0114] (5) Input the real-time fault feature vector into the distribution cabinet fault detection model, start the forward reasoning process, and perform branch judgment on the real-time fault feature vector through each decision tree in the distribution cabinet fault detection model to obtain the sub-fault detection result of each decision tree;
[0115] (6) Perform fuzzy reasoning and weighted fusion on the sub-fault detection results of each decision tree to obtain real-time fault detection results.
[0116] Specifically, the current signal of the distribution cabinet is sampled in real time to obtain the real-time current signal Real-time sampling can be achieved through a high-frequency data logger to ensure that subtle changes in the current signal can be captured. Next, the real-time current signal is decomposed by wavelet packets to obtain sub-band signals at multiple frequency scales. ,in represents the decomposition level, Indicated in The first subbands. The formula for wavelet packet decomposition is:
[0117] ;
[0118] in, Represents the original real-time current signal, is the wavelet packet basis function, is the decomposed sub-band signal. Perform spectrum analysis on multiple sub-band signals of the real-time current signal and calculate the spectrum energy feature vector of each sub-band signal. Spectral analysis is achieved by calculating the energy density of each sub-band signal. The formula is as follows:
[0119] ;
[0120] in, Indicates Tier The spectral energy density of the sub-band signal, F is the subband signal The Fourier transform of is the number of samples of the signal. Perform instantaneous phase analysis on multiple sub-band signals of the real-time current signal and calculate the current direction eigenvector of each sub-band signal. Instantaneous phase analysis is achieved through Hilbert transform, and the formula is:
[0121] ;
[0122] in, is the analytical signal, H yes The Hilbert transform of is an imaginary unit. The instantaneous phase and instantaneous amplitude of the analytical signal are:
[0123] ;
[0124] According to the instantaneous phase and instantaneous amplitude, the instantaneous frequency characteristics of each sub-band signal are calculated, and the formula is as follows:
[0125] ;
[0126] in, It is Tier The instantaneous frequency of the subband signal, is the instantaneous phase, is time. The spectrum energy feature vector and the current direction feature vector are combined to obtain the real-time fault feature vector. The real-time fault feature vector is input into the distribution cabinet fault detection model to start the forward reasoning process. The distribution cabinet fault detection model uses multiple decision trees to make branch judgments on the real-time fault feature vector. Each decision tree makes judgments based on the feature splitting rules learned during its training process. Assume that the real-time fault feature vector is , the branch judgment process in the decision tree can be expressed as:
[0127] ;
[0128] in, Indicates A decision tree, Represents the leaf node of the decision tree, and the leaf node corresponds to the sub-fault detection result. The sub-fault detection result of each decision tree is subjected to fuzzy reasoning and weighted fusion to obtain the final real-time fault detection result. Fuzzy reasoning smoothes the sub-fault detection result through the membership function, and the weighted fusion formula is:
[0129] Final Result ;
[0130] Among them, Final Result is the final fault detection result. is the number of decision trees, It is The weight of a decision tree, It is the membership function, which is used to fuzzy process the sub-fault detection results.
[0131] In a specific embodiment, the execution step inputs the real-time fault feature vector into the distribution cabinet fault detection model, starts the forward reasoning process, and performs branch judgment on the real-time fault feature vector through each decision tree in the distribution cabinet fault detection model. The process of obtaining the sub-fault detection result of each decision tree can specifically include the following steps:
[0132] S1: Extract the first eigenvalue from the real-time fault feature vector, compare the first eigenvalue with the split threshold of the decision tree root node, and select to enter the left subtree or the right subtree according to the comparison result;
[0133] S2: In the selected subtree, continue to extract the next eigenvalue of the real-time fault feature vector, compare the next eigenvalue with the split threshold of the current node, and recursively select to enter the left subtree or the right subtree until reaching the leaf node;
[0134] S3: After reaching the leaf node of the decision tree, the fault type corresponding to the leaf node is used as the first fault detection result of the decision tree;
[0135] S4: for each characteristic value in the real-time fault characteristic vector, repeatedly execute steps S1 to S3, sequentially pass through each node of the decision tree, and obtain a second fault detection result of the decision tree;
[0136] S5: constructing a logical judgment expression of the decision tree according to the structure of the decision tree and the node splitting rule, substituting each characteristic value of the real-time fault characteristic vector into the logical judgment expression, and calculating and obtaining the third fault detection result of the decision tree;
[0137] S6: Perform feature importance analysis on the real-time fault feature vector, calculate the influence weight of each feature value and generate a feature importance weight vector of the decision tree;
[0138] S7: performing a weighted combination of the third fault detection result of the decision tree and the feature importance weight vector to obtain a sub-fault detection result of the decision tree;
[0139] S8: Repeat steps S1 to S7 for each decision tree in the distribution cabinet fault detection model to obtain sub-weighted fault detection results of all decision trees.
[0140] Specifically, the first eigenvalue is extracted from the real-time fault feature vector, the first eigenvalue is compared with the split threshold of the root node of the decision tree, and the left subtree or the right subtree is selected according to the comparison result. Assume that the real-time fault feature vector is ,in is the first eigenvalue. If the splitting threshold of the root node is , then by comparing and , decide whether to enter the left subtree or the right subtree. If , then enter the left subtree, otherwise enter the right subtree. Extract the next eigenvalue of the real-time fault feature vector in the selected subtree, compare the next eigenvalue with the split threshold of the current node, and recursively select to enter the left subtree or the right subtree until reaching the leaf node. Assume that the eigenvalue extracted on the second-layer node is , and then the comparison is performed again. According to the comparison result, the appropriate subtree is selected, and the process is continued until the leaf node is reached. When the leaf node is reached, the fault type corresponding to the leaf node is used as the first fault detection result of the decision tree. Each leaf node is pre-labeled with a fault type, which is learned during the model training process. When the real-time fault feature vector reaches the leaf node through the decision tree, the label of the node is used as the first fault detection result. For each eigenvalue in the real-time fault feature vector, the above steps are repeated, and each node of the decision tree is passed through in turn to obtain the second fault detection result of the decision tree. According to the structure of the decision tree and the node splitting rule, the logical judgment expression of the decision tree is constructed. Each eigenvalue of the real-time fault feature vector is substituted into the logical judgment expression to calculate the third fault detection result of the decision tree. By constructing the logical expression, the judgment process of the decision tree is simplified and the calculation efficiency is improved. For example, if the splitting rule of the decision tree is a simple threshold comparison, the logical expression can be simplified into a series of "if-else" statements. By nesting these statements, the fault type can be quickly determined. The feature importance analysis is performed on the real-time fault feature vector, the influence weight of each eigenvalue is calculated, and the feature importance weight vector of the decision tree is generated. By evaluating the contribution of each eigenvalue to the fault detection result, its importance in the decision tree is determined. The third fault detection result of the decision tree is weighted and combined with the feature importance weight vector to obtain the sub-fault detection result of the decision tree. The detection result is combined with the feature weight to make the final detection result more representative and accurate. Through this process, the sub-fault detection results of each decision tree are obtained, which comprehensively consider the importance of each eigenvalue. For each decision tree in the distribution cabinet fault detection model, the above steps are repeated to obtain the sub-weighted fault detection results of all decision trees. By weighted combination of the fault detection results of all decision trees in the model, a comprehensive fault detection result is finally formed.
[0141] In a specific embodiment, the process of executing step S105 may specifically include the following steps:
[0142] (1) Perform time series analysis and feature extraction on the real-time fault detection results to obtain the fault feature sequence, and construct a reference feature sequence of candidate fault points based on the historical fault database;
[0143] (2) Calculate the grey correlation coefficient matrix between the fault feature sequence and the reference feature sequence of the candidate fault point;
[0144] (3) Calculate the weight coefficients of the candidate fault points in the grey correlation coefficient matrix to obtain the weighted correlation degree of each candidate fault point;
[0145] (4) Set the correlation threshold according to the weighted correlation, and screen the candidate fault points to obtain the initial positioning result;
[0146] (5) Perform fuzzy clustering analysis on the initial positioning results to obtain the fault point cluster center, and match the fault point cluster center with the fault feature sequence to obtain the fault point positioning information.
[0147] Specifically, first, perform time series analysis and feature extraction on the real-time fault detection results to obtain the fault feature sequence. The real-time fault detection results are usually a series of data points that reflect the state of the system at different time points. Time series analysis is to process and analyze these data points to extract meaningful features. Assume that the feature sequence extracted from the real-time fault detection results is ,in, Indicates time, It is Features in time By analyzing these features, the laws and patterns of fault occurrence are identified. A reference feature sequence of candidate fault points is constructed based on the historical fault database. The historical fault database contains a large amount of fault data, which can be used to construct a reference feature sequence of candidate fault points. Assume that the reference feature sequence is ,in, It is Candidate fault points at time These reference feature sequences represent the characteristics of each fault point in history and are used to compare with the current fault feature sequence. The grey correlation coefficient matrix between the fault feature sequence and the reference feature sequence of the candidate fault point is calculated. The grey correlation coefficient is a method to measure the similarity between two sequences. By calculating the grey correlation coefficient matrix, the correlation between the fault feature sequence and the reference feature sequence of the candidate fault point can be quantified. Assume that the grey correlation coefficient matrix is ,in, Fault feature sequence With Candidate fault point reference feature sequence The correlation between them. The calculation formula is:
[0148] ;
[0149] in, is the resolution coefficient, usually taken as 0.5. is the first Features in time The value of It is The first candidate fault point Features in time The value of . The weight coefficients of the candidate fault points in the grey correlation coefficient matrix are calculated to obtain the weighted correlation of each candidate fault point. The weight coefficient reflects the importance of each candidate fault point in fault detection. The weighted correlation is obtained by weighting the grey correlation coefficient of each candidate fault point. The correlation threshold is set according to the weighted correlation, and the candidate fault points are screened to obtain the initial positioning result. The correlation threshold is used to filter out the candidate fault points with low correlation, and only those fault points that are highly correlated with the fault feature sequence are retained. The initial positioning result is subjected to fuzzy clustering analysis to obtain the cluster center of the fault point. Fuzzy clustering analysis is a method based on similarity measurement. By clustering the fault points in the initial positioning result, several cluster centers are obtained. These cluster centers represent the concentrated distribution areas of the fault points. The cluster centers of the fault points are matched with the fault feature sequence to obtain the fault point location information. By matching the cluster centers of the fault points with the fault feature sequence, the position of the fault point is located, and the fault point location information is generated.
[0150] The above describes the distribution cabinet fault monitoring method in the embodiment of the present application. The following describes the distribution cabinet fault monitoring device in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the power distribution cabinet fault monitoring device includes:
[0151] The acquisition module 201 is used to acquire the historical current signal of the power distribution cabinet, and remove the high-frequency noise and harmonic interference of the historical current signal to obtain a standard current signal;
[0152] The decomposition module 202 is used to perform wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales, and calculate the spectrum energy characteristics and current direction characteristics of each sub-band signal to generate a full-frequency domain fault feature vector of each sub-band signal;
[0153] The training module 203 is used to establish a comprehensive fault detection criterion of spectrum energy and current direction and use it as an objective function, input the full-frequency domain fault feature vector into the XGBoost integrated learning algorithm for model training, and build a distribution cabinet fault detection model;
[0154] The detection module 204 is used to obtain the real-time current signal of the power distribution cabinet, and input the real-time current signal into the power distribution cabinet fault detection model to perform fault detection, thereby obtaining a real-time fault detection result;
[0155] The generation module 205 is used to construct a fault feature sequence according to the real-time fault detection result, and calculate the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point, so as to generate the fault point location information.
[0156] Through the cooperation of the above components, the discrete wavelet transform and soft threshold denoising methods are used to effectively remove high-frequency noise in the historical current signal and reduce harmonic interference. The standard current signal is decomposed into sub-band signals of different frequency scales through wavelet packet decomposition technology, and the spectrum energy characteristics and current direction characteristics of each sub-band signal are extracted to generate full-frequency domain fault feature vectors. The full-frequency domain fault feature vectors are trained using the XGBoost ensemble learning algorithm to build a distribution cabinet fault detection model. The accuracy and robustness of fault detection are improved by combining the comprehensive fault detection criteria of spectrum energy and current direction with multi-level thresholds and fuzzy processing. The detection model is parameter tuned and cross-validated to ensure the reliability and generalization ability of the model under different fault conditions. The current signal of the distribution cabinet is sampled in real time, and wavelet packet decomposition and spectrum analysis are performed on it to extract the real-time fault feature vector. The trained fault detection model is used to quickly detect the real-time current signal, detect faults in time, and provide real-time fault detection results. The fault feature sequence is constructed based on the real-time fault detection results, and the gray correlation between it and the reference feature sequence of the candidate fault point is calculated. Through weight coefficient calculation and fuzzy cluster analysis, the fault point can be accurately located, and the fault point location information can be generated, which can improve the accuracy and timeliness of fault handling, thereby realizing remote fault monitoring of the distribution cabinet.
[0157] The present application also provides a distribution cabinet fault monitoring device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the distribution cabinet fault monitoring method in the above-mentioned embodiments.
[0158] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the distribution cabinet fault monitoring method.
[0159] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0160] If the integrated unit is implemented in the form of 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, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.
[0161] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A method for monitoring faults in a power distribution cabinet, characterized in that: The power distribution cabinet fault monitoring method comprises: Collecting historical current signals of the power distribution cabinet, and removing high-frequency noise and harmonic interference from the historical current signals to obtain standard current signals; The standard current signal is subjected to wavelet packet decomposition to obtain sub-band signals of different frequency scales, and the spectrum energy characteristics and current direction characteristics of each sub-band signal are calculated to generate a full-frequency domain fault feature vector of each sub-band signal; specifically, the method comprises: performing multi-layer wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales; performing energy entropy calculation on each sub-band signal to obtain sub-band energy entropy characteristics, and performing Hilbert transform on each sub-band signal to obtain instantaneous phase and instantaneous amplitude of each sub-band signal; calculating the sub-band instantaneous frequency characteristics of each sub-band signal according to the instantaneous phase and the instantaneous amplitude; performing Hilbert transform on each sub-band signal to obtain the instantaneous phase and instantaneous amplitude of each sub-band signal; performing Hilbert transform on each sub-band signal to obtain the instantaneous phase and the ... The signal is short-time Fourier transformed to obtain a time-frequency distribution matrix, and the sub-band spectrum shape characteristics of each sub-band signal are calculated according to the time-frequency distribution matrix; each sub-band signal is subjected to zero-crossing rate analysis to obtain a sub-band zero-crossing rate characteristic, and the peak factor and margin factor of each sub-band signal are calculated to obtain a sub-band waveform characteristic, and each sub-band signal is subjected to singular value decomposition to obtain a sub-band singular value characteristic; a full-frequency domain fault feature vector of each sub-band signal is generated according to the sub-band energy entropy characteristic, the sub-band instantaneous frequency characteristic, the sub-band spectrum shape characteristic, the sub-band zero-crossing rate characteristic, the sub-band waveform characteristic and the sub-band singular value characteristic; A comprehensive fault detection criterion of spectrum energy and current direction is established and used as the objective function. The full-frequency domain fault feature vector is input into the XGBoost integrated learning algorithm for model training to build a distribution cabinet fault detection model; specifically, the full-frequency domain fault feature vector is normalized to obtain a standardized feature vector, and the spectrum energy density and current direction angle are calculated according to the standardized feature vector to obtain spectrum energy features and current direction features; the spectrum energy features and the current direction features are weightedly fused to obtain a comprehensive fault detection index; multi-level thresholds are set according to the comprehensive fault detection index to build a multi-category fault Classification criteria, and fuzzify the multi-category fault classification criteria to obtain comprehensive fault detection criteria; construct an objective function according to the comprehensive fault detection criteria, and perform convex optimization on the objective function to obtain optimal parameters; randomly sample the standardized feature vectors to construct multiple sub-training sets, and train multiple decision trees based on the multiple sub-training sets to obtain a base learner set; integrate the base learner set to construct an XGBoost ensemble learning algorithm, and perform parameter tuning and cross-validation on the XGBoost ensemble learning algorithm according to the optimal parameters to obtain a distribution cabinet fault detection model; Acquire a real-time current signal of the power distribution cabinet, and input the real-time current signal into the power distribution cabinet fault detection model to perform fault detection, thereby obtaining a real-time fault detection result; A fault feature sequence is constructed according to the real-time fault detection result, and the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point is calculated to generate fault point location information.
2. The method for monitoring faults in a power distribution cabinet according to claim 1, characterized in that: The collecting of historical current signals of the power distribution cabinet and removing high-frequency noise and harmonic interference from the historical current signals to obtain standard current signals include: Collecting historical current signals of the power distribution cabinet, and performing discrete wavelet transform on the historical current signals to obtain multi-scale wavelet coefficients; Calculating a threshold value of each scale according to the multi-scale wavelet coefficients, and performing a soft threshold denoising process on the multi-scale wavelet coefficients to obtain denoised wavelet coefficients; Performing wavelet reconstruction on the denoised wavelet coefficients to obtain a first current signal, and performing fast Fourier transform on the first current signal to obtain a frequency domain current signal; Identify harmonic components according to the frequency domain current signal, perform amplitude attenuation processing on the harmonic components, and obtain a second current signal; Performing an inverse fast Fourier transform on the second current signal to obtain a time-domain current signal, and performing dynamic state estimation on the time-domain current signal to obtain a filtered current signal; The filtered current signal is subjected to empirical mode decomposition to obtain a number of intrinsic mode functions, and a standard current signal is obtained by reconstructing the number of intrinsic mode functions.
3. The method for monitoring faults in a power distribution cabinet according to claim 1, characterized in that: The acquiring of the real-time current signal of the power distribution cabinet and inputting the real-time current signal into the power distribution cabinet fault detection model for fault detection to obtain a real-time fault detection result includes: Performing real-time sampling on the current signal of the power distribution cabinet to obtain a real-time current signal, and performing wavelet packet decomposition on the real-time current signal to obtain a plurality of sub-band signals of the real-time current signal; Performing spectrum analysis on multiple sub-band signals of the real-time current signal, and calculating a spectrum energy feature vector of each sub-band signal in the real-time current signal; Performing instantaneous phase analysis on multiple sub-band signals of the real-time current signal, and calculating a current direction characteristic vector of each sub-band signal in the real-time current signal; Combining the spectrum energy feature vector and the current direction feature vector to obtain a real-time fault feature vector; Input the real-time fault feature vector into the distribution cabinet fault detection model, start the forward reasoning process, and perform branch judgment on the real-time fault feature vector through each decision tree in the distribution cabinet fault detection model to obtain a sub-fault detection result of each decision tree; The sub-fault detection results of each decision tree are subjected to fuzzy reasoning and weighted fusion to obtain real-time fault detection results.
4. The method for monitoring faults in a power distribution cabinet according to claim 3, characterized in that: The real-time fault feature vector is input into the distribution cabinet fault detection model, a forward reasoning process is started, and branch judgment is performed on the real-time fault feature vector through each decision tree in the distribution cabinet fault detection model to obtain a sub-fault detection result of each decision tree, including: S1: extracting a first eigenvalue from the real-time fault feature vector, comparing the first eigenvalue with the split threshold of the root node of the decision tree, and selecting to enter the left subtree or the right subtree according to the comparison result; S2: In the selected subtree, continue to extract the next eigenvalue of the real-time fault feature vector, compare the next eigenvalue with the split threshold of the current node, and recursively select to enter the left subtree or the right subtree until reaching the leaf node; S3: After reaching a leaf node of the decision tree, taking the fault type corresponding to the leaf node as the first fault detection result of the decision tree; S4: for each characteristic value in the real-time fault characteristic vector, repeatedly executing steps S1 to S3, sequentially passing through each node of the decision tree, to obtain a second fault detection result of the decision tree; S5: constructing a logical judgment expression of the decision tree according to the structure and node splitting rule of the decision tree, substituting each characteristic value of the real-time fault characteristic vector into the logical judgment expression, and calculating to obtain a third fault detection result of the decision tree; S6: performing feature importance analysis on the real-time fault feature vector, calculating the influence weight of each feature value and generating a feature importance weight vector of a decision tree; S7: performing a weighted combination of the third fault detection result of the decision tree and the feature importance weight vector to obtain a sub-fault detection result of the decision tree; S8: Repeat steps S1 to S7 for each decision tree in the distribution cabinet fault detection model to obtain sub-weighted fault detection results of all decision trees.
5. The method for monitoring faults in a power distribution cabinet according to claim 1, characterized in that: The method of constructing a fault feature sequence according to the real-time fault detection result, calculating the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point, and generating the fault point location information includes: Performing time series analysis and feature extraction on the real-time fault detection results to obtain a fault feature sequence, and constructing a reference feature sequence of candidate fault points based on a historical fault database; Calculating a grey correlation coefficient matrix between the fault feature sequence and a reference feature sequence of the candidate fault point; Calculating weight coefficients for candidate fault points in the grey correlation coefficient matrix to obtain a weighted correlation degree for each candidate fault point; Setting a correlation threshold according to the weighted correlation, and screening the candidate fault points to obtain an initial positioning result; Perform fuzzy clustering analysis on the initial positioning result to obtain the fault point cluster center, and match the fault point cluster center with the fault feature sequence to obtain fault point positioning information.
6. A distribution cabinet fault monitoring device, characterized in that: The power distribution cabinet fault monitoring device comprises: The acquisition module is used to collect the historical current signal of the power distribution cabinet, and remove the high-frequency noise and harmonic interference of the historical current signal to obtain a standard current signal; The decomposition module is used to perform wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales, and calculate the spectrum energy characteristics and current direction characteristics of each sub-band signal to generate a full-frequency domain fault feature vector of each sub-band signal; specifically, the decomposition module comprises: performing multi-layer wavelet packet decomposition on the standard current signal to obtain sub-band signals of different frequency scales; performing energy entropy calculation on each sub-band signal to obtain sub-band energy entropy characteristics, and performing Hilbert transform on each sub-band signal to obtain instantaneous phase and instantaneous amplitude of each sub-band signal; calculating the sub-band instantaneous frequency characteristics of each sub-band signal according to the instantaneous phase and the instantaneous amplitude; performing Hilbert transform on each sub-band signal to obtain the instantaneous phase and instantaneous amplitude of each sub-band signal; performing Hilbert transform on each sub-band signal to obtain the instantaneous phase and the ... The sub-band signals are subjected to short-time Fourier transform to obtain a time-frequency distribution matrix, and the sub-band spectrum shape characteristics of each sub-band signal are calculated according to the time-frequency distribution matrix; each sub-band signal is subjected to zero-crossing rate analysis to obtain a sub-band zero-crossing rate characteristic, and the peak factor and margin factor of each sub-band signal are calculated to obtain a sub-band waveform characteristic, and each sub-band signal is subjected to singular value decomposition to obtain a sub-band singular value characteristic; a full-frequency domain fault feature vector of each sub-band signal is generated according to the sub-band energy entropy characteristic, the sub-band instantaneous frequency characteristic, the sub-band spectrum shape characteristic, the sub-band zero-crossing rate characteristic, the sub-band waveform characteristic and the sub-band singular value characteristic; The training module is used to establish a comprehensive fault detection criterion of spectrum energy and current direction and use it as an objective function, input the full-frequency domain fault feature vector into the XGBoost integrated learning algorithm for model training, and build a distribution cabinet fault detection model; specifically, it includes: normalizing the full-frequency domain fault feature vector to obtain a standardized feature vector, and calculating the spectrum energy density and the current direction angle according to the standardized feature vector to obtain the spectrum energy feature and the current direction feature; weighted fusion of the spectrum energy feature and the current direction feature to obtain a comprehensive fault detection index; setting a multi-level threshold according to the comprehensive fault detection index to build a multi-level fault detection model. A multi-category fault classification criterion is obtained by fuzzifying the multi-category fault classification criterion to obtain a comprehensive fault detection criterion; an objective function is constructed according to the comprehensive fault detection criterion, and the objective function is solved by convex optimization to obtain the optimal parameters; the standardized feature vector is randomly sampled to construct multiple sub-training sets, and multiple decision trees are trained according to the multiple sub-training sets to obtain a base learner set; the base learner set is integrated to construct an XGBoost integrated learning algorithm, and the XGBoost integrated learning algorithm is parameterized and cross-validated according to the optimal parameters to obtain a distribution cabinet fault detection model; A detection module, used to obtain a real-time current signal of the power distribution cabinet, and input the real-time current signal into the power distribution cabinet fault detection model to perform fault detection, so as to obtain a real-time fault detection result; The generation module is used to construct a fault feature sequence according to the real-time fault detection result, and calculate the grey correlation between the fault feature sequence and the reference feature sequence of the candidate fault point to generate fault point location information.
7. A distribution cabinet fault monitoring device, characterized in that: The power distribution cabinet fault monitoring device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the power distribution cabinet fault monitoring device executes the power distribution cabinet fault monitoring method according to any one of claims 1 to 5.
8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the power distribution cabinet fault monitoring method according to any one of claims 1 to 5 is implemented.
