Power distribution cabinet arc discharge defect detection method

Abstract

The invention provides a power distribution cabinet arc discharge defect detection method, and belongs to the technical field of power distribution cabinet arc discharge defect detection, and the method comprises the steps: building an electric field intensity dense matrix through analyzing historical data and a simulation model, and carrying out the Fourier transform to obtain a frequency domain feature set; and a multi-branch hybrid neural network model is trained by comparing feature differences of normal and defect samples. And performing the same data processing and feature extraction on the power distribution cabinet to be tested, inputting an obtained difference feature set into the trained model, and predicting a possible arc discharge position. According to the method, the technical advantages of modern signal processing and machine learning are fully utilized, simulation analysis and experimental testing are combined, rich feature information contained in electric field data is deeply excavated, accurate identification and quantitative evaluation of arc discharge defects are achieved through deep analysis of the electric field data in the power distribution cabinet, and the accuracy and the reliability of the arc discharge defects are improved. The problems of low detection precision and high cost in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of arc discharge defect detection in power distribution cabinets, and in particular, relates to a method for detecting arc discharge defects in power distribution cabinets. Background Art

[0002] The normal operation of the power system is crucial to the social and economic development. As an indispensable and important equipment in the power system, the safe and stable operation of the distribution cabinet is directly related to the quality of power consumption and the reliability of power supply. However, in the actual operation process, the distribution cabinet is prone to defects such as arc discharge, which not only causes equipment damage, but also may cause serious safety accidents. Therefore, how to effectively monitor and diagnose the arc discharge hazards of the distribution cabinet has always been an important technical problem that needs to be solved in the power industry.

[0003] The existing distribution cabinet arc discharge detection methods mainly include: partial discharge detection method, infrared thermal imaging detection method and electromagnetic radiation detection method. The partial discharge detection method can indirectly reflect the occurrence of arc discharge by measuring the partial discharge signal inside the distribution cabinet. However, this method is limited by the layout position and number of partial discharge sensors, and it is difficult to fully reflect the arc discharge state of the entire distribution cabinet. The infrared thermal imaging detection method can find abnormal hot spots on the surface of the distribution cabinet and find out the area where arc discharge may occur. However, this method cannot deeply analyze the internal mechanism of arc discharge and can only play a preliminary warning role. The electromagnetic radiation detection method infers the occurrence of arc discharge by monitoring the changes in the electromagnetic environment of the distribution cabinet. This method has high accuracy, but it requires the installation of a large number of electromagnetic sensors, which is costly and sensitive to environmental interference.

[0004] In general, although the existing arc discharge detection methods can realize the monitoring of distribution cabinets to a certain extent, they often have problems such as low detection accuracy and high cost. These limitations restrict their wide application in power systems, and it is urgent to develop a more intelligent and efficient arc discharge diagnosis solution. Summary of the invention

[0005] In view of this, the present invention provides a method for detecting arc discharge defects in a distribution cabinet, which can solve the problems of low detection accuracy and high cost in the prior art.

[0006] The present invention is achieved in that:

[0007] The present invention provides a method for detecting arc discharge defects in a power distribution cabinet, comprising the following steps:

[0008] S10, determining multiple key locations where arc discharge occurred in the distribution cabinet based on the collected distribution cabinet arc discharge damage records;

[0009] S20. For each of the multiple key positions, establish a simulation model of the internal structure of the power distribution cabinet, including different types of electrical components, conductors, and insulating materials; perform electric field analysis on the simulation model using simulation software to simulate the electric field distribution under different voltages;

[0010] S30, according to the electric field analysis, adopting the method of adaptive grid refinement and interpolation to obtain a dense matrix of electric field intensity; and performing Fourier transform on the dense matrix of electric field intensity to obtain a frequency domain feature set including multiple complex matrices, recorded as a first feature set;

[0011] S40, obtaining a plurality of distribution cabinets known to have arc discharge defects, and after powering on, continuously increasing the voltage, performing multiple experiments, and recording experimental data of each experiment;

[0012] S50, constructing an electric field intensity dense matrix using the electric field data, recorded as a defect electric field intensity dense matrix; and performing Fourier transform on the defect electric field intensity dense matrix to obtain a frequency domain feature set including multiple complex matrices, recorded as a second feature set;

[0013] S60, screening out features in the second feature set whose difference from the first feature set is greater than a preset similarity threshold, to form a difference feature set;

[0014] S70, using the difference feature set as training input and the arc data corresponding to the difference feature set as training output, training a multi-branch hybrid neural network model to obtain an arc discharge defect model;

[0015] S80, obtaining the electric field data of the power distribution cabinet to be tested, and using the method of S50 to obtain a frequency domain feature set containing multiple complex matrices corresponding to the electric field data of the power distribution cabinet to be tested, recorded as a third feature set; screening the features in the third feature set whose difference from the first feature set is greater than a preset similarity threshold, to form a difference feature set of the power distribution cabinet to be tested;

[0016] S90. Input the difference feature set of the power distribution cabinet to be tested into the arc discharge defect model to obtain the arc position of the arc data of the power distribution cabinet to be tested as the arc discharge defect detection result of the power distribution cabinet to be tested.

[0017] The electric field data include electric field intensity distribution, electric field gradient, potential distribution, charge density distribution and dielectric polarization intensity.

[0018] Furthermore, the arc data includes arc position, arc starting voltage, arc extinguishing voltage, partial discharge at 10 kV, arc duration, arc current peak value and arc energy.

[0019] Furthermore, the preset similarity threshold is 0.7.

[0020] Among them, the multi-branch hybrid neural network model includes a convolutional branch network, a recurrent branch network, a fully connected branch network and a fusion branch network.

[0021] Furthermore, the convolution branch network is used to extract spatial features from the difference feature set, the input is a two-dimensional representation of the difference feature set, the output is a spatial feature vector of the difference feature set, and the structure is a combination of multiple convolutional layers and pooling layers.

[0022] Furthermore, the recurrent branch network is used to extract temporal features from a difference feature set, the input is a temporal representation of the difference feature set, the output is a temporal feature vector of the difference feature set, and the structure is a long short-term memory (LSTM) or a gated recurrent unit (GRU) network.

[0023] Furthermore, the fully connected branch network is used to extract global features from the difference feature set, the input is a one-dimensional flattened representation of the difference feature set, the output is a global feature vector of the difference feature set, and the structure is a multi-layer fully connected layer.

[0024] Furthermore, the fusion branch network is used to integrate the output features of each branch network. The input is the concatenation of the spatial feature vector, the temporal feature vector and the global feature vector. The output is the prediction result of the arc data. The structure is a multi-layer fully connected layer and an output layer.

[0025] Among them, the steps of conducting multiple experiments and recording the experimental data of each experiment are specifically: selecting multiple distribution cabinet samples with known arc discharge defects, connecting them to a high-voltage power supply, starting to power on from a low voltage, gradually increasing the voltage with a fixed step size, staying at each voltage level for a fixed time and collecting data; the collected data include electric field data such as electric field intensity distribution, electric field gradient, potential distribution, charge density distribution, dielectric polarization intensity, and arc data such as arc position, arc starting voltage, arc extinction voltage, partial discharge at 10kV, arc duration, arc current peak, and arc energy; repeating the experiment for each distribution cabinet sample at least 3 times; after each experiment, slowly reducing the voltage to zero and ensuring complete discharge.

[0026] Specifically, the steps of the experiment include:

[0027] 1. Preparation: Select several distribution cabinet samples with known arc discharge defects; prepare the measurement equipment required for the experiment, such as high-voltage power supply, voltmeter, ammeter, oscilloscope, electric field strength meter, partial discharge meter, etc.; ensure the safety of the experimental site and take protective measures;

[0028] 2. Power on the power distribution cabinet: Connect the power distribution cabinet to the high voltage power supply; start from low voltage and power on slowly;

[0029] 3. Voltage increase: gradually increase the voltage with a fixed step size (such as 500V each time); stay at each voltage level for a fixed time (such as 5 minutes) for observation and data collection;

[0030] 4. Data collection: At each voltage level, record the following data:

[0031] Electric field data: Use an electric field strength meter to measure the electric field strength distribution; calculate the electric field gradient; measure the potential distribution; measure the charge density distribution; measure the dielectric polarization strength;

[0032] Arc data: observe and record the location of the arc; record the arc starting voltage (the voltage when the arc starts to appear); record the arc extinction voltage (the voltage when the arc disappears); measure the amount of partial discharge at 10kV; record the arc duration with a high-speed camera; measure the arc current peak with an oscilloscope; calculate the arc energy;

[0033] 5. Repeat the experiment: Repeat the above steps 3-4 for each distribution cabinet sample at least 3 times to ensure the reliability of the data;

[0034] 6.Safety handling: After each experiment, slowly reduce the voltage to zero; disconnect the power supply and ensure that the distribution cabinet is fully discharged.

[0035] Compared with the prior art, the arc discharge defect detection method for distribution cabinet provided by the present invention fully utilizes the technical advantages of modern signal processing and machine learning, and realizes accurate identification and quantitative evaluation of arc discharge defects through in-depth analysis of the electric field data inside the distribution cabinet.

[0036] Compared with the traditional arc discharge detection method, the scheme of the present invention has the following significant advantages:

[0037] 1. High detection accuracy. This solution combines simulation analysis and experimental testing to deeply mine the rich feature information contained in the electric field data and construct a multi-branch hybrid neural network model that can comprehensively extract the spatial, temporal and global characteristics of arc discharge, greatly improving the accuracy of defect identification.

[0038] 2. Cost-effective. This method does not require large-scale transformation of the existing power distribution cabinet. It only requires the deployment of a small number of electric field sensors at key locations to complete online monitoring, greatly reducing the detection cost. At the same time, this method does not require the participation of professionals, and realizes fully automatic arc discharge diagnosis, further saving labor costs.

[0039] 3. Strong applicability. The solution of the present invention is based on the analysis of electric field information, is not affected by the external environment, and can be applied to the stock distribution equipment that has been in operation for many years without large-scale transformation, and has strong wide applicability.

[0040] 4. High level of intelligence. This method fully integrates modern signal processing and machine learning technologies, realizes automatic detection and diagnosis of arc discharge defects, and provides strong support for intelligent status monitoring and preventive maintenance of power distribution equipment.

[0041] In summary, the present invention solves the problems of low detection accuracy and high cost in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0044] like Figure 1 FIG. 1 is a flow chart of a method for detecting arc discharge defects in a power distribution cabinet provided by the present invention. The method comprises the following steps:

[0045] S10, determining multiple key locations where arc discharge occurred in the distribution cabinet based on the collected distribution cabinet arc discharge damage records;

[0046] S20. For each of the multiple key positions, establish a simulation model of the internal structure of the power distribution cabinet, including different types of electrical components, conductors, and insulating materials; use simulation software to perform electric field analysis on the simulation model to simulate the electric field distribution under different voltages;

[0047] S30, according to the electric field analysis, using adaptive grid refinement and interpolation to obtain a dense matrix of electric field intensity; and performing Fourier transform on the dense matrix of electric field intensity to obtain a frequency domain feature set including multiple complex matrices, recorded as a first feature set;

[0048] S40, obtaining a plurality of distribution cabinets known to have arc discharge defects, and after powering on, continuously increasing the voltage, performing multiple experiments, and recording experimental data of each experiment;

[0049] S50, constructing an electric field intensity dense matrix using electric field data, recorded as a defect electric field intensity dense matrix; and performing Fourier transform on the defect electric field intensity dense matrix to obtain a frequency domain feature set including multiple complex matrices, recorded as a second feature set;

[0050] S60, screening the features in the second feature set whose difference from the first feature set is greater than a preset similarity threshold to form a difference feature set;

[0051] S70, using the difference feature set as training input and the arc data corresponding to the difference feature set as training output, training a multi-branch hybrid neural network model to obtain an arc discharge defect model;

[0052] S80, obtaining electric field data of the distribution cabinet to be tested, and using the method of S50 to obtain a frequency domain feature set containing multiple complex matrices corresponding to the electric field data of the distribution cabinet to be tested, recorded as a third feature set; screening features in the third feature set whose difference from the first feature set is greater than a preset similarity threshold to form a difference feature set of the distribution cabinet to be tested;

[0053] S90, inputting the difference feature set of the power distribution cabinet to be tested into the arc discharge defect model, and obtaining the arc position of the arc data of the power distribution cabinet to be tested as the arc discharge defect detection result of the power distribution cabinet to be tested.

[0054] The specific implementation methods of the above steps are described in detail below:

[0055] The specific implementation method of step S10 is to first collect historical records of arc discharge damage in the distribution cabinet. According to these recorded data, the key positions where arc discharge occurs inside the distribution cabinet are determined. Usually, these key positions may appear at places such as busbar joints, switch contacts, insulators, etc. The specific position coordinate information of these key positions is determined as the basis for the subsequent establishment of the simulation model.

[0056] The specific implementation method of step S20 is that for each determined key position, a three-dimensional simulation model of its internal structure needs to be established. The simulation model needs to include various electrical components, metal conductors, and insulating materials inside the distribution cabinet. These simulation models can be drawn using some general three-dimensional modeling software such as SolidWorks or AutoCAD. Next, these simulation models are imported into finite element analysis software such as ANSYS, COMSOL, etc., and electric field analysis simulation is performed on them. By setting different voltage boundary conditions, the electric field distribution inside the distribution cabinet under different voltages can be simulated. These electric field analysis results lay the foundation for subsequent arc discharge feature extraction.

[0057] The specific implementation method of step S30 is to first perform adaptive grid refinement and interpolation processing on the electric field analysis data obtained in step S20. In this way, denser electric field intensity matrix data can be obtained. Next, these electric field intensity matrix data are subjected to Fourier transform processing. Fourier transform can convert the electric field intensity data in the time domain into a complex matrix form in the frequency domain. These frequency domain features can better reflect the spectral characteristics of the electric field distribution and provide valuable feature information for subsequent arc discharge defect identification. The complex matrix set composed of these frequency domain features is recorded as the first feature set.

[0058] The specific implementation method of step S50 is that first, a batch of distribution cabinet samples with known arc discharge defects need to be obtained. Experimental tests are conducted on these samples, and arc discharge is induced in these distribution cabinets by gradually increasing the working voltage. During each arc discharge experiment, a large amount of relevant experimental data needs to be recorded, including electric field data such as electric field intensity distribution, electric field gradient, potential distribution, charge density distribution, dielectric polarization intensity, and arc data such as arc position, arc starting voltage, arc extinguishing voltage, partial discharge amount at 10kV, arc duration, arc current peak, and arc energy. These rich experimental data provide important samples for subsequent defect recognition model training.

[0059] The specific implementation method of step S60 is to use the electric field data obtained in step S50 to construct an electric field intensity dense matrix that characterizes the arc discharge defect characteristics. This matrix is ​​recorded as the defect electric field intensity dense matrix. Next, the defect electric field intensity dense matrix is ​​also Fourier transformed to obtain a frequency domain feature set containing multiple complex matrices, which is recorded as the second feature set.

[0060] The specific implementation method of step S70 is to design and train a multi-branch hybrid neural network model with the difference feature set as the training input and the arc data as the training output. The model has a complex structure, including a convolutional neural network (CNN) branch to process frequency domain features, a long short-term memory network (LSTM) branch to analyze time series information, and a fully connected network branch to process scalar features. The outputs of each branch are integrated through a feature fusion layer, and finally the arc discharge defect model is obtained using a multi-task output layer. During the training process, the Adam optimizer is used, combined with a mixed loss function of mean square error and cross entropy, and a small batch gradient descent method is used for optimization. To prevent overfitting, early stopping method, learning rate decay strategy, Dropout and L2 regularization are implemented. Iterative learning is performed on the training set, and the model performance is evaluated through the validation set. Finally, the model with the best comprehensive performance on the validation set is selected as the final arc discharge defect model. The entire training process focuses on the generalization ability of the model, and the model performance is continuously improved by adjusting the network structure, hyperparameter optimization and other methods to ensure that the model can accurately predict the arc discharge characteristics of unseen samples.

[0061] Specifically, the model includes four main branch networks:

[0062] 1) Convolutional branch network: This network is used to extract spatial features from the difference feature set. The input is the two-dimensional representation of the difference feature set, and the output is the spatial feature vector. The network structure is a combination of multiple convolutional layers and pooling layers.

[0063] 2) Cyclic branch network: This network is used to extract the temporal features in the difference feature set. The input is the temporal representation of the difference feature set, and the output is the temporal feature vector. The network structure is long short-term memory (LSTM) or gated recurrent unit (GRU).

[0064] 3) Fully connected branch network: This network is used to extract global features from the difference feature set. The input is the one-dimensional flattened representation of the difference feature set, and the output is the global feature vector. The network structure is a multi-layer fully connected layer.

[0065] 4) Fusion branch network: This network is used to integrate the output features of the first three branch networks. The input is the concatenation of spatial feature vectors, temporal feature vectors and global feature vectors, and the output is the final prediction result of arc data. The network structure is a multi-layer fully connected layer and an output layer.

[0066] This multi-branch hybrid neural network model is trained end-to-end with the difference feature set as input and the corresponding arc data as output. After the training is completed, the model can be used to calculate the difference features based on the electric field data of the distribution cabinet to predict the arc discharge characteristic parameters of the distribution cabinet to be tested, so as to evaluate its arc discharge defects. This multi-branch neural network model can fully mine the spatial, temporal and global features contained in the electric field data, and improve the accuracy and reliability of arc discharge defect identification.

[0067] The specific implementation method of step S80 is to collect the electric field data of the distribution cabinet to be tested, ensure that the collection conditions are consistent with the training data, and construct a high-resolution electric field intensity dense matrix. Perform a multidimensional fast Fourier transform (FFT) on the matrix to obtain a frequency domain feature set containing multiple complex matrices, which is recorded as the third feature set. By calculating the similarity between the third feature set and the first feature set (normal distribution cabinet features), the features with significant differences are screened out using a preset similarity threshold. Combined with ensemble learning methods such as random forest or XGBoost to evaluate the importance of features, retain the features with the highest importance ranking, and form a difference feature set of the distribution cabinet to be tested. Standardize the screened features to eliminate the dimensionality effect. If the number of features is too large, methods such as principal component analysis (PCA) or t-SNE can be selected for dimensionality reduction to reduce the computational complexity. Perform unique hot encoding or embedded encoding on the categorical features to ensure that the encoding method is consistent with the training model. Finally, convert the processed feature data into the input format required by the model such as tensor to ensure that the data dimension matches the model input layer and prepare for subsequent model reasoning.

[0068] The specific implementation method of step S90 is to load the trained arc discharge defect model, ensure version consistency, and preheat the model to stabilize performance when necessary. Input the difference feature set of the distribution cabinet to be tested obtained in S80 into the model, and use the forward propagation function of the model for prediction. Post-process the original output of the model, including denormalization and converting the softmax output into a specific category label. Extract the arc position information from the output, and select the position with the highest probability as the prediction result. Calculate the confidence of the prediction result, and mark or filter the prediction with low confidence. Visually mark the predicted arc position on the distribution cabinet structure diagram, and use a heat map to show the spatial distribution of the arc occurrence probability. Generate a detailed inspection report, including the predicted arc position, confidence, potential risks and other information, and give corresponding maintenance suggestions. Establish a feedback mechanism to allow engineers to evaluate and correct the prediction results, and collect this feedback information for subsequent model updates and optimization. Implement an incremental learning strategy, regularly update the model with newly collected data, so that it can adapt to the dynamic changes in the state of the distribution cabinet, and continuously improve the accuracy and practicality of the model. Through this series of steps, the reliability and practical application value of arc discharge defect detection results are ensured.

[0069] Through the above detailed step implementation method, a distribution cabinet arc discharge defect detection method based on electric field feature analysis can be established. This method can more accurately identify the arc discharge hidden dangers in the distribution cabinet, and provide strong support for the status monitoring and fault diagnosis of the distribution equipment. Compared with the traditional arc discharge detection method, this method can more deeply analyze the rich information contained in the electric field data, and improve the accuracy and reliability of the detection. At the same time, this method does not require large-scale modification and upgrading of the distribution cabinet, and only requires a small number of sensors to be deployed at key positions to complete the online monitoring of arc discharge defects. This is a simple and effective optimization solution for distribution equipment that has been in operation for many years. This distribution cabinet arc discharge defect detection method based on electric field feature analysis makes full use of the technical advantages of modern signal processing and machine learning, and greatly reduces the detection cost while improving the detection accuracy.

[0070] Specifically, the principle of the present invention is that the arc discharge process is essentially a highly nonlinear electromagnetic phenomenon, which will produce complex electric field distribution changes inside the distribution cabinet. These electric field changes contain rich arc discharge characteristic information and are an important basis for reflecting and identifying arc discharge defects.

[0071] Specifically, the scheme of the present invention first determines the key positions where arc discharge occurs inside the power distribution cabinet through simulation analysis, and establishes three-dimensional simulation models of these key positions. Then, the electric field simulation of these simulation models is performed using finite element analysis software to obtain the electric field distribution inside the power distribution cabinet under different working voltages. These electric field analysis data lay the foundation for the subsequent arc discharge feature extraction.

[0072] Next, the electric field analysis data is digitally processed. First, a denser electric field intensity matrix is ​​obtained through adaptive grid refinement and interpolation. Then, these electric field intensity matrices are Fourier transformed to convert the time domain electric field data into a complex matrix representation in the frequency domain. This can better reflect the spectral characteristics of the electric field distribution and provide valuable feature information for subsequent arc discharge defect identification.

[0073] In order to train a model that can accurately identify arc discharge defects, a large number of experimental data samples need to be collected. By gradually increasing the working voltage of the distribution cabinet, arc discharge is induced in samples with known defects, and rich electric field data and arc data are recorded at the same time. Using these experimental data, the features in the electric field frequency domain that are significantly different from the normal state and the defective state can be screened out as training input. The arc data is used as the training output to build an arc discharge defect recognition model.

[0074] Finally, a multi-branch hybrid neural network model is designed to realize defect recognition. This model can fully extract the spatial, temporal and global features contained in the electric field data. Through end-to-end training, it can eventually predict the arc discharge characteristic parameters of the distribution cabinet to be tested based on the electric field data, and then evaluate its defect status. Compared with traditional detection methods, this defect diagnosis method based on electric field feature analysis can greatly improve the detection accuracy and greatly reduce the detection cost, providing reliable technical support for intelligent status monitoring and preventive maintenance of distribution equipment.

[0075] In order to better understand and implement the present invention, a specific embodiment 1 of the present invention is provided below, and the steps included are described in detail as follows:

[0076] The specific implementation of step S10 is as follows:

[0077] First, it is necessary to determine the key locations where arc discharge occurs inside the distribution cabinet based on the historical damage record data of the distribution cabinet. These key locations may appear in places such as busbar joints, switch contacts, insulators, etc. Record the three-dimensional coordinate information of these key locations. key =(x key ,y key ,z key ), which provides a reference for the subsequent establishment of simulation models.

[0078] The specific implementation of step S20 is as follows:

[0079] For each determined key position r key , use 3D modeling software to build a simulation model of its internal structure. The simulation model needs to include various electrical components, metal conductors and insulating materials inside the distribution cabinet. The set of these models is denoted as Then these simulation models are imported into the finite element analysis software to perform electric field analysis simulation. By applying different voltage boundary conditions V boundary , the electric field distribution E(r; M, V boundary ), where r represents the spatial coordinate and M represents the simulation model set M. These electric field analysis results lay the foundation for the subsequent arc discharge feature extraction.

[0080] The specific implementation of step S30 is as follows:

[0081] First, the electric field analysis data E(r; M, V) obtained in step S20 is analyzed. boundary ) to perform adaptive grid refinement and interpolation processing to obtain a denser electric field intensity matrix E dense Next, these electric field strength matrices E dense Perform Fourier transform to convert the electric field intensity data in the time domain into a complex matrix form F(E dense ). These frequency domain features can better reflect the spectral characteristics of the electric field distribution and provide valuable feature information for subsequent arc discharge defect identification. The complex matrix set composed of these frequency domain features is recorded as the first feature set

[0082] The specific implementation of step S50 is as follows:

[0083] First, obtain a set of distribution cabinet samples known to have arc discharge defects. These samples were experimentally tested by gradually increasing the operating voltage V test In each experiment, a large amount of relevant electric field data and arc data are recorded to form a sample set. Where E exp Represents the electric field data obtained from the experiment, including the electric field intensity distribution E, the electric field gradient Potential distribution φ, charge density distribution ρ, dielectric polarization intensity P, etc.; A exp Represents the corresponding arc data, including the arc position r arc , arc starting voltage V onset , arc extinction voltage V extinction、Partial discharge Q under 10kV PD , arc duration t arc , Arc current peak value I arc,peak , Arc energy E arc Etc. These rich experimental data provide important samples for subsequent defect recognition model training.

[0084] The specific implementation of step S60 is as follows:

[0085] The electric field data E obtained in step S50 is used exp , a dense matrix E of electric field intensity that characterizes the characteristics of arc discharge defects can be constructed dense,defect Next, we also calculate the defect electric field intensity dense matrix E dense,defect Perform Fourier transform to obtain a frequency domain feature set containing multiple complex matrices

[0086] The specific implementation of step S70 is as follows:

[0087] Need to filter out the second feature set The first feature set There are large differences in features. The specific approach is to calculate the first feature set and the second feature set The similarity between Select features whose similarity is lower than a preset threshold θ to form a difference feature set This difference feature set It reflects the difference in electric field spectrum characteristics between the distribution cabinet under normal working condition and the distribution cabinet with arc discharge defects. Through a large number of experimental tests, it is found that when the similarity threshold θ is set to 0.7, valuable difference features can be extracted better. Next, a multi-branch hybrid neural network model is constructed to realize the identification of arc discharge defects. The model includes four main branch networks:

[0088] 1) Convolutional branch network: This network is used to extract the difference feature set The spatial characteristics in Represented as a two-dimensional matrix Input into the network. The output of the network is the spatial feature vector v spatial ,The network structure is a combination of multiple convolutional layers and pooling layers.

[0089] 2) Cyclic branch network: This network is used to extract the difference feature set The timing characteristics in Represented as a time series Input into the network. The output of the network is the time series feature vector v temporal,The network structure is long short term memory LSTM or gated recurrent unit GRU.

[0090] 3) Fully connected branch network: This network is used to extract the difference feature set The global features in Represented as a one-dimensional vector Input into the network. The output of the network is the global feature vector v global ,The network structure is a multi-layer fully connected layer.

[0091] 4) Fusion branch network: This network is used to integrate the output features of the first three branch networks. The input is the spatial feature vector v spatial , time series feature vector v temporal and the global eigenvector v global The final prediction result of arc data is output The network structure consists of multiple fully connected layers and one output layer.

[0092] Differentiation feature set As input, the corresponding arc data A exp As output, this multi-branch hybrid neural network model is trained end-to-end. After the training is completed, the model can be used to calculate the difference features based on the electric field data of the distribution cabinet to be tested. Predict the arc discharge characteristic parameters of the distribution cabinet to be tested Thus, the arc discharge defects can be evaluated. This multi-branch neural network model can fully mine the spatial, temporal and global features contained in the electric field data, and improve the accuracy and reliability of arc discharge defect identification.

[0093] The variables and constants in the above formula are explained as follows:

[0094] r key =(x key ,y key ,z key ): The three-dimensional coordinates of the key position where arc discharge occurs inside the distribution cabinet.

[0095] A collection of simulation models of the internal structure of a power distribution cabinet.

[0096] V boundary : Voltage boundary conditions imposed on the simulation model.

[0097] E(r;M,V boundary ): According to the simulation model M and voltage boundary conditions V boundary Calculated electric field distribution inside the power distribution cabinet.

[0098] E dense: A denser electric field intensity matrix obtained after adaptive grid refinement and interpolation.

[0099] F(E dense ): dense The frequency domain feature set obtained by Fourier transform.

[0100] The first feature set is composed of F(E dense )composition.

[0101] Calculate the first feature set and the second feature set The similarity between .

[0102] θ: Similarity threshold, used to filter out the difference feature set

[0103] The difference feature set Represented as a two-dimensional matrix and used as the input of the convolutional branch network.

[0104] The difference feature set Represented as a time series form and used as the input of the recurrent branch network.

[0105] The difference feature set Represented as a one-dimensional vector and used as the input of the fully connected branch network.

[0106] v spatial ,v temporal ,v global : Output feature vector of each branch network.

[0107] The output of the fused branch network is the prediction result of the arc data.

[0108] The specific implementation of S80 and S90 adopts the steps similar to the above, and inputs the obtained results into the model to obtain the final detection results, which will not be described in detail here.

[0109] In general, the arc discharge defect detection method for distribution cabinets based on electric field characteristic analysis in this embodiment 1 fully utilizes the technical advantages of modern signal processing and machine learning. While improving the detection accuracy, it also greatly reduces the detection cost, providing a reliable solution for intelligent status monitoring and preventive maintenance of distribution equipment.

[0110] In order to better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A power operation and maintenance team is responsible for managing an important 110kV substation, which supplies power to most local industrial parks and residential areas, and has a direct impact on regional economic development and the quality of life of residents. In order to ensure the safe and stable operation of the substation equipment, the power operation and maintenance team decided to use the distribution cabinet arc discharge defect detection method based on electric field feature analysis proposed in the present invention to comprehensively monitor the key distribution cabinets in the substation.

[0111] First, the engineering and technical personnel conducted a preliminary investigation and on-site survey of the 20 main distribution cabinets in the substation. By reviewing the equipment operation records, it was found that 5 of the distribution cabinets had experienced serious arc discharge accidents, resulting in equipment damage and even power outages in a short period of time. After further analysis, the arc discharge of these 5 problematic distribution cabinets mainly occurred at the busbar joints, switch contacts and the surface of some insulators. Therefore, the engineering and technical personnel identified 10 key locations of these 5 distribution cabinets as the focus of subsequent monitoring.

[0112] Based on this key location information, the engineers used the 3D modeling software Solidworks to build simulation models of the five distribution cabinets. These simulation models not only include the detailed geometric dimensions of various electrical components, metal conductors, and insulating materials, but also consider the physical connection relationship between them. The simulation models of these distribution cabinets were imported into the ANSYS finite element analysis software, and the engineers conducted a detailed simulation analysis of the electric field distribution.

[0113] Taking one of the problematic distribution cabinets as an example, when applying the 110kV working voltage boundary condition, ANSYS simulation calculated the electric field distribution inside the distribution cabinet, as shown in Table 1. As can be seen from the table, the electric field strength at the busbar joint, the switch contact area and the surface of some insulators is significantly higher than that in other areas, which also corresponds to the location of the previous arc discharge fault. Through further analysis of the electric field data in these key areas, the engineering and technical personnel found that the electric field distribution in these high electric field strength areas showed obvious non-uniform characteristics.

[0114] Location Maximum electric field strength (kV / m) Busbar connector 85 Switch contacts 92 Insulator surface 78

[0115] Table 1 Simulation results of electric field distribution inside the power distribution cabinet

[0116] In order to better characterize the characteristics of these electric field distributions, the engineering and technical personnel performed Fourier transform processing on the electric field intensity matrix data obtained by simulation. Through Fourier transform, the time domain data of the electric field intensity is converted into a complex matrix form in the frequency domain. These frequency domain characteristics can more completely reflect the spectral characteristics of the electric field distribution. Taking the electric field intensity at the busbar joint as an example, its frequency domain characteristics are shown in Table 2. It can be seen from the table that the electric field intensity spectrum in this area shows obvious nonlinear distribution characteristics, and there are more prominent peaks in certain frequency bands. This nonlinear spectrum feature is the typical electromagnetic signal feature generated during the arc discharge process.

[0117] Frequency (kHz) Electric field strength amplitude (kV / m) 2 58.7 5 41.2 10 27.4 20 14.5 50 6.8 100 3.1

[0118] Table 2 Frequency domain characteristics of electric field strength at busbar joints

[0119] With these electric field characteristic data from simulation analysis, the next step is to combine the actual arc discharge fault data to establish an intelligent diagnosis model that can identify arc discharge defects. To this end, the engineering and technical personnel specially selected 3 of the 5 problematic distribution cabinets and conducted arc discharge characteristic tests in a laboratory environment. By gradually increasing the working voltage, arc discharge faults were purposefully induced in these distribution cabinets. During each experiment, not only the key parameters such as the arc position, arc starting voltage, and duration were recorded, but also rich electric field data such as electric field intensity distribution, electric field gradient, potential distribution, charge density distribution, and dielectric polarization intensity were collected.

[0120] Through the analysis of these experimental data, the engineering and technical personnel found that when arc discharge occurs, the electric field distribution inside the distribution cabinet does change significantly. First, at the key position where the arc discharge occurs, the electric field strength will jump sharply, as shown in Table 3. This is because after the arc is formed, a large number of ions and electrons will be generated at these positions, resulting in a sharp increase in local charge density, which in turn causes a sharp increase in electric field strength. At the same time, the electric field gradient near the arc discharge position will also increase sharply, as shown in Table 4. This abnormal change in the electric field gradient reflects the non-uniform characteristics of charge movement during the arc discharge process.

[0121]

[0122]

[0123] Table 3 Changes in electric field intensity during arc discharge

[0124] Time (ms) <![CDATA[Electric field gradient (kV / m 2 )]]> 0 12 2 15 4 28 6 42 8 35 10 22

[0125] Table 4 Changes of electric field gradient during arc discharge

[0126] In addition, arc discharge will also cause local disturbances in the potential distribution inside the distribution cabinet, as shown in Table 5. When an arc occurs, the potential distribution shows obvious asymmetric characteristics, which is related to the uneven distribution of space charge generated when the arc is formed. At the same time, the polarization intensity of the medium will also change significantly due to the appearance of the arc, as shown in Table 6. The strong ionization caused by the arc discharge makes the polarization state of the medium disordered, so that the polarization intensity fluctuates violently near the arc.

[0127] Time (ms) Potential distribution (kV) 0 110 2 108 4 102 6 90 8 85 10 92

[0128] Table 5 Changes in potential distribution during arc discharge

[0129] Time (ms) <![CDATA[Polarization intensity (C / m 2 )]]> 0 0.16 2 0.18 4 0.24 6 0.32 8 0.27 10 0.20

[0130] Table 6 Changes in dielectric polarization intensity during arc discharge

[0131] Through in-depth analysis of these experimental data, engineers and technicians found that arc discharge not only causes drastic changes in local electric field distribution, but also causes significant disturbances in potential distribution, charge density, and dielectric polarization state. The abnormal changes in these electric field characteristic parameters provide valuable basis for identifying arc discharge defects.

[0132] Based on the above analysis results, the engineering and technical personnel constructed a multi-branch hybrid neural network model to realize the intelligent diagnosis of arc discharge defects in the distribution cabinet. The overall architecture of the model is shown in Table 7, which mainly includes the following four branch networks:

[0133] 1) Convolutional branch network: This network takes the electric field intensity distribution matrix as input and extracts the spatial characteristics of the electric field distribution through multi-layer convolution and pooling operations.

[0134] 2) Recurrent branch network: This network takes the electric field time series data as input and uses the LSTM or GRU structure to extract the time series characteristics of the dynamic changes of the electric field.

[0135] 3) Fully connected branch network: This network flattens the electric field data and inputs it, extracting global features through multiple layers of fully connected layers.

[0136] 4) Fusion branch network: This network integrates the output features of the first three branch networks, and predicts the various parameters of arc discharge through further fully connected layers and output layers.

[0137] Network Branch enter Output Convolution branch Electric Field Strength Matrix Spatial eigenvector Loop Branch Electric field time series Time series feature vector Fully connected branches Flattened electric field data Global eigenvector Fusion Branch Feature vector concatenation Arc parameter prediction results

[0138] Table 7 Multi-branch hybrid neural network model architecture

[0139] During the model training process, the engineers first constructed the first feature set and the second feature set using the experimental data obtained in steps S50 and S60. By comparing the differences between the two feature sets, the features that are significantly different from the normal state and the defective state were screened out to form a difference feature set. Then, the multi-branch neural network model was trained end-to-end with these difference features as input and the corresponding arc parameter data as output. After repeated optimization and debugging, the model achieved good prediction performance in the arc discharge defect recognition task.

[0140] With the trained neural network model, engineers can use it to perform intelligent diagnosis of arc discharge hazards in the other 17 distribution cabinets in the substation. First, corresponding electric field sensors are installed at key locations of these 17 distribution cabinets to monitor their internal electric field distribution in real time. The collected electric field data is input into the trained neural network model, which can quickly calculate the arc discharge characteristic parameters of each distribution cabinet, including arc position, arc starting voltage, duration, energy release, etc.

[0141] After a month of monitoring, the engineering and technical personnel found that among the 17 distribution cabinets, 3 had arc discharge hazards to varying degrees. Among them, the arc discharge characteristics of 2 distribution cabinets were more serious, such as arc energy release of up to 300kJ, duration of 15ms, and arc starting voltage dropped below 90kV. The other distribution cabinet was relatively mild, with an arc energy of only 150kJ, duration of less than 10ms, and arc starting voltage maintained at around 100kV. Based on these diagnostic results, the power operation and maintenance team immediately conducted key inspections and maintenance on the 3 distribution cabinets with hidden dangers. After careful investigation by professional maintenance personnel, it was found that the loose busbar joints and serious surface contamination of some insulators were the main causes of arc discharge.

[0142] After repairing and dealing with these problems, the engineering and technical personnel monitored the three distribution cabinets again. The results showed that after improving the quality of busbar connection and cleaning the surface of insulators, the arc discharge characteristics of the three distribution cabinets were significantly improved, the arc energy and duration were greatly reduced, and the arc starting voltage returned to normal levels.

[0143] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting arc discharge defects in a power distribution cabinet, characterized in that: The following steps are involved: S10, determining multiple key locations where arc discharge occurred in the distribution cabinet based on the collected distribution cabinet arc discharge damage records; S20. For each of the multiple key positions, establish a simulation model of the internal structure of the power distribution cabinet, including different types of electrical components, conductors, and insulating materials; perform electric field analysis on the simulation model using simulation software to simulate the electric field distribution under different voltages; S30, according to the electric field analysis, adopting the method of adaptive grid refinement and interpolation to obtain a dense matrix of electric field intensity; and performing Fourier transform on the dense matrix of electric field intensity to obtain a frequency domain feature set including multiple complex matrices, recorded as a first feature set; S40, obtaining a plurality of distribution cabinets known to have arc discharge defects, and after powering on, continuously increasing the voltage, performing multiple experiments, and recording experimental data of each experiment; S50, constructing an electric field intensity dense matrix using the electric field data, recorded as a defect electric field intensity dense matrix; and performing Fourier transform on the defect electric field intensity dense matrix to obtain a frequency domain feature set including multiple complex matrices, recorded as a second feature set; S60, screening out features in the second feature set whose difference from the first feature set is greater than a preset similarity threshold, to form a difference feature set; S70, using the difference feature set as training input and the arc data corresponding to the difference feature set as training output, training a multi-branch hybrid neural network model to obtain an arc discharge defect model; S80, obtaining the electric field data of the power distribution cabinet to be tested, and using the method of S50 to obtain a frequency domain feature set containing multiple complex matrices corresponding to the electric field data of the power distribution cabinet to be tested, recorded as a third feature set; screening the features in the third feature set whose difference from the first feature set is greater than a preset similarity threshold, to form a difference feature set of the power distribution cabinet to be tested; S90. Input the difference feature set of the power distribution cabinet to be tested into the arc discharge defect model to obtain the arc position of the arc data of the power distribution cabinet to be tested as the arc discharge defect detection result of the power distribution cabinet to be tested.

2. A method for detecting arc discharge defects in a power distribution cabinet according to claim 1, characterized in that: The electric field data include electric field intensity distribution, electric field gradient, potential distribution, charge density distribution and dielectric polarization intensity.

3. A method for detecting arc discharge defects in a power distribution cabinet according to claim 2, characterized in that: The arc data include arc position, arc starting voltage, arc extinction voltage, partial discharge at 10 kV, arc duration, arc current peak value and arc energy.

4. A method for detecting arc discharge defects in a power distribution cabinet according to claim 3, characterized in that: The preset similarity threshold is 0.

7.

5. The method for detecting arc discharge defects in a power distribution cabinet according to claim 1, characterized in that: The multi-branch hybrid neural network model includes a convolutional branch network, a cyclic branch network, a fully connected branch network and a fusion branch network.

6. A method for detecting arc discharge defects in a power distribution cabinet according to claim 5, characterized in that: The convolution branch network is used to extract spatial features in the difference feature set, the input is a two-dimensional representation of the difference feature set, the output is the spatial feature vector of the difference feature set, and the structure is a combination of multiple convolutional layers and pooling layers.

7. A method for detecting arc discharge defects in a power distribution cabinet according to claim 6, characterized in that: The cyclic branch network is used to extract the temporal features in the difference feature set, the input is the temporal representation of the difference feature set, the output is the temporal feature vector of the difference feature set, and the structure is a long short-term memory LSTM or a gated recurrent unit GRU network.

8. A method for detecting arc discharge defects in a power distribution cabinet according to claim 7, characterized in that: The fully connected branch network is used to extract global features from the difference feature set, the input is a one-dimensional flattened representation of the difference feature set, the output is a global feature vector of the difference feature set, and the structure is a multi-layer fully connected layer.

9. A method for detecting arc discharge defects in a power distribution cabinet according to claim 8, characterized in that: The fusion branch network is used to integrate the output features of each branch network. The input is the concatenation of spatial feature vectors, temporal feature vectors and global feature vectors. The output is the prediction result of arc data. The structure is a multi-layer fully connected layer and an output layer.

10. The method for detecting arc discharge defects in a power distribution cabinet according to claim 1, characterized in that: The steps of conducting multiple experiments and recording the experimental data of each experiment are specifically: selecting multiple distribution cabinet samples with known arc discharge defects, connecting them to a high-voltage power supply, starting to power on from a low voltage, gradually increasing the voltage with a fixed step size, staying at each voltage level for a fixed time and collecting data; the collected data include electric field intensity distribution, electric field gradient, potential distribution, charge density distribution, electric field data of dielectric polarization intensity, and arc data of arc position, arc starting voltage, arc extinction voltage, partial discharge at 10kV, arc duration, arc current peak, and arc energy; repeating the experiment for each distribution cabinet sample at least 3 times; and slowly reducing the voltage to zero after each experiment to ensure complete discharge.

Images