A switch cabinet partial discharge fault diagnosis method based on PSO-DBN neural network

Abstract

The present application relates to the technical field of switch cabinet fault diagnosis, and particularly relates to a switch cabinet partial discharge fault diagnosis method based on a PSO-DBN neural network. The present application can obtain a diagnosis conclusion by analyzing characteristic gas content through a first diagnosis model in view of whether partial discharge occurs; in view of a specific fault type, the fault type cannot be determined simply from the characteristic gas content, through experimental data analysis, it is found that the variation of the characteristic gas includes the characteristics of fault diagnosis, therefore, a second diagnosis model is used to analyze the variation of the characteristic gas at a specific time, so that the specific fault type can be obtained, and the monitoring, fault classification and positioning of switch cabinet partial discharge are realized.

Description

Technical Field

[0001] This invention relates to the field of switchgear fault diagnosis technology, and in particular to a method for diagnosing partial discharge faults in switchgear based on PSO-DBN neural network. Background Technology

[0002] Air circuit breaker cabinets are common power transmission and distribution equipment, mainly used in various large substations and distribution rooms, providing certain control and protection functions for the operation of power systems. During manufacturing, distribution, installation, operation, and maintenance, various insulation defects inevitably occur within the air circuit breaker cabinet, leading to partial discharge phenomena. Even small discharge defects can further accelerate equipment deterioration and may even cause fire accidents.

[0003] Currently, the main methods for detecting partial discharge faults include pulse current method, ultra-high frequency method, ultrasonic method, and transient voltage to ground method. However, these methods have limitations such as low identification efficiency, poor sensitivity, low identification accuracy, poor anti-interference ability, and inability to accurately locate the discharge position. Summary of the Invention

[0004] This invention discloses a method for diagnosing partial discharge faults in switchgear based on PSO-DBN neural networks. The specific method is as follows:

[0005] Collect the characteristic gas content of the insulating oil in the switchgear to be diagnosed;

[0006] Input the content of the characteristic gas into the first diagnostic model to determine whether partial discharge has occurred;

[0007] When partial discharge is detected, the change in characteristic gas over time is periodically read.

[0008] The change in characteristic gas over time is input into the second diagnostic model to diagnose the specific fault type and location.

[0009] The advantage of this embodiment is that, in order to determine whether partial discharge has occurred, a diagnostic conclusion can be obtained by analyzing the content of characteristic gases through the first diagnostic model; however, for specific fault types, it is not possible to determine the fault type simply from the content of characteristic gases. Through analysis of experimental data, it was found that the change in characteristic gases includes the characteristics of fault diagnosis. Therefore, by using the second diagnostic model to analyze the change in characteristic gases at a specific time, the specific fault type can be obtained, thereby realizing the monitoring, fault classification, and location of partial discharge in the switchgear.

[0010] Furthermore, the characteristic gases include NO2, O3, and CO, which are collected in real time by three gas sensors installed at the gas intake positions.

[0011] Furthermore, the first diagnostic model is a Support Vector Machine (SVM), and the specific construction method is as follows:

[0012] Simulate partial discharges in combination of possible discharge locations and possible fault types within the switchgear to obtain characteristic gas data;

[0013] Preprocess the characteristic gas data and construct the dataset;

[0014] Initialize the SVM diagnostic model and train it with the dataset.

[0015] Furthermore, the Support Vector Machine (SVM) includes the following settings during construction:

[0016] Kernel functions: Gaussian sum function RBF

[0017] RBF function parameters: kernel width σ = 3

[0018]

[0019] The optimal penalty coefficient C and slack variable l are: C = 2 and l = 0.003.

[0020] Furthermore, the fault types include: air gap discharge, surface discharge, floating potential discharge, and corona discharge.

[0021] The advantages of this embodiment are that the air gap discharge is characterized by irregular discharge intervals and numerous discharges; the surface discharge is characterized by regular discharge intervals and numerous discharges; the floating potential discharge is characterized by the shortest discharge intervals and few discharges; and the corona discharge is characterized by the longest discharge intervals and few discharges. These four types of discharge faults exhibit different discharge behaviors, and their discharge characteristics can all be reflected by changes in the characteristic gas over a characteristic time, making them suitable for classification using the same diagnostic model.

[0022] Furthermore, the change in the characteristic gas over time includes: the duration of the change in the characteristic gas content and the number of changes in the characteristic gas content.

[0023] Furthermore, the changes in the characteristic gas over time are periodically read, as follows:

[0024] The preset reading period T and the preset reading duration are used to read the change of characteristic gas over time at the initial period interval;

[0025] If the accuracy of the diagnostic result is less than the preset target after the reading result is input into the second diagnostic model, then adjust the preset reading period T and the preset duration until the accuracy of the diagnostic result is greater than or equal to the preset target.

[0026] The advantage of this embodiment is that by adjusting the reading period T and the preset duration, the data that best represents the fault characteristics can be selected. During the adjustment, it can be adjusted manually based on experience or by using an optimization algorithm. The fitness function during adjustment is that the accuracy of the diagnostic result is greater than or equal to the preset target.

[0027] Furthermore, the second diagnostic model is a PSO-DBN neural network model, and the specific construction method is as follows:

[0028] Inside the switchgear, several simulations were conducted to detect combinations of partial discharges at possible locations and with possible fault types.

[0029] Obtain the time-varying characteristic gas of combined partial discharge;

[0030] The changes in characteristic gases over time are normalized.

[0031] The normalized information is then subjected to T-SNE dimensionality reduction.

[0032] The training set is constructed using the dimensionality-reduced data;

[0033] Set the hyperparameters of the DBN neural network;

[0034] Perform DBN neural network testing;

[0035] Perform PSO (Particle Swarm Optimization) algorithm initialization;

[0036] Determine the optimal position and fitness of the particle swarm;

[0037] Iteratively update the particle velocity and position until the preset target or the number of iterations is reached;

[0038] The PSO optimization results are output and assigned to the DBN neural network to complete the construction of the PSO-DBN neural network model.

[0039] Furthermore, when establishing the PSO-DBN neural network diagnostic model, the parameters and hyperparameters are designed as follows:

[0040] The particle swarm population size m is set to 8;

[0041] The number of iterations, n, is set to 20;

[0042] The learning factors C1 and C2 are set to C1 = C2 = 1.495;

[0043] The particle length l is set to 4;

[0044] The maximum particle velocity Vmax = k·Vmax, 0.1≤k≤1. For the dimension of optimizing the number of hidden layer neurons, we take k=1 and Vmax=1; for the dimension of optimizing the learning rate, we take k=0.1 and Vmax=0.5.

[0045] The particle range [Xmax, Xmin] is set to [0.001, 0.5].

[0046] Linear weights w: w max =0.8, w min =0.2, t max =20

[0047] ω=ω max -(ω max -ω min )[2t / t max -(t / t max ) 2 ]

[0048] The initial learning rate of the DBN neural network model was set to 0.1;

[0049] The initial momentum of the DBN neural network model is set to 0.1;

[0050] The retention rate parameter for Dropout in the DBN neural network model is 0.95;

[0051] The number of hidden neuron nodes is set to a constant value, with 100 hidden neurons per layer;

[0052] The number of hidden layers is set to 3.

[0053] The advantage of this embodiment is that it further trains and optimizes the diagnostic model using detected historical data, improving the speed and accuracy of partial discharge fault identification by the PSO-DBN neural network diagnostic model. Maintenance personnel can use the diagnostic results to understand the potential sources of partial discharge within the switchgear and address any safety hazards, reducing the number of switchgear trips caused by partial discharge and ensuring the safe and stable operation of the switchgear, the reliability of power supply, and the safe operation of the power grid. Attached Figure Description

[0054] The accompanying drawings of this invention are described below.

[0055] Figure 1 This is a schematic diagram of the process of the present invention.

[0056] Figure 2 A schematic diagram of the support vector machine construction process.

[0057] Figure 3 A schematic diagram of the process for building a PSO-DBN neural network.

[0058] Figure 4 This is a schematic diagram of the partial discharge location.

[0059] Figure 5 This is a schematic diagram of partial discharge zones. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] When partial discharge occurs in an air switchgear cabinet, the air undergoes a chemical reaction and decomposes into a series of characteristic gases, the concentration of which can be detected by a gas sensor. Based on this fundamental principle, the following method is proposed.

[0062] A method for partial discharge fault diagnosis in switchgear based on PSO-DBN neural network, such as Figure 1 As shown, the specific steps are as follows:

[0063] S1. Collect the characteristic gas content of the insulating oil in the switchgear to be diagnosed.

[0064] Specifically, the characteristic gases include NO2, O3, and CO, which are collected in real time by three gas sensors installed at the gas intake positions.

[0065] S2. Input the characteristic gas content into the first diagnostic model to determine whether partial discharge has occurred.

[0066] Specifically, the first diagnostic model is a Support Vector Machine (SVM), such as... Figure 2 As shown, the specific construction method is as follows:

[0067] S21. Simulate partial discharges of possible discharge locations and possible fault types within the switchgear to obtain characteristic gas data;

[0068] S22. Preprocess the characteristic gas data and construct the dataset.

[0069] S23. Initialize the SVM diagnostic model and train the SVM diagnostic model with the dataset.

[0070] In step S23, when constructing the Support Vector Machine (SVM), the kernel function is the Gaussian sum function (RBF), and the kernel width σ of the RBF function parameter is 3. The optimal penalty coefficient C and slack variable l are: C = 2 and l = 0.003.

[0071] In steps S3 and S4, the dependent variable is the fault type, including air gap discharge, surface discharge, floating potential discharge, and corona discharge. The variable is the change in characteristic gas over time, including the duration of characteristic gas content changes and the number of times characteristic gas content changes.

[0072] S3. When partial discharge is detected, the change in characteristic gas over time is periodically read.

[0073] Specifically, the change in the characteristic gas over time is periodically read, and the specific method is as follows:

[0074] S31. Preset reading cycle T, with each reading duration being the preset duration, to read the change in characteristic gas over time at the initial cycle interval.

[0075] S32. If the accuracy of the diagnostic result is less than the preset target after the reading result is input into the second diagnostic model, then adjust the preset reading period T and the preset duration until the accuracy of the diagnostic result is greater than or equal to the preset target.

[0076] S4. Input the change in characteristic gas over time into the second diagnostic model to diagnose the specific fault type and location, such as discharge location. Figure 4 As shown, to facilitate positioning, the discharge positioning can be divided into zones, as illustrated in the zoning diagram. Figure 5 As shown.

[0077] Specifically, the second diagnostic model is the PSO-DBN neural network model, such as... Figure 3 As shown, the specific construction method is as follows:

[0078] S41. In the switchgear, simulate partial discharges of possible locations and possible fault types several times;

[0079] S42. Obtain the change in characteristic gas of combined partial discharge over time;

[0080] S43. Normalize the changes in characteristic gases over time;

[0081] S44. Perform T-SNE dimensionality reduction on the normalized information.

[0082] S45. Construct a training set using the dimensionality-reduced data;

[0083] S46. Set the hyperparameters of the DBN neural network;

[0084] S47. Perform DBN neural network testing;

[0085] S48. Initialize the PSO (Particle Swarm Optimization) algorithm.

[0086] S49. Determine the optimal position and fitness of the particle swarm;

[0087] S410: Iteratively update particle velocity and position until the preset target or the number of iterations is reached;

[0088] S411. Output the PSO optimization result and assign it to the DBN neural network to complete the construction of the PSO-DBN neural network model.

[0089] The specific parameters and hyperparameters are designed as follows:

[0090] The particle swarm population size m is set to 8;

[0091] The number of iterations, n, is set to 20;

[0092] The learning factors C1 and C2 are set to C1 = C2 = 1.495;

[0093] The particle length l is set to 4;

[0094] The maximum particle velocity Vmax = k·Vmax, 0.1≤k≤1. For the dimension of optimizing the number of hidden layer neurons, we take k=1 and Vmax=1; for the dimension of optimizing the learning rate, we take k=0.1 and Vmax=0.5.

[0095] The particle range [Xmax, Xmin] is set to [0.001, 0.5].

[0096] Linear weights w: w max =0.8, w min =0.2, t max =20

[0097] ω=ω max -(ω max -ω min )[2t / t max -(t / t max ) 2 ]

[0098] The initial learning rate of the DBN neural network model was set to 0.1;

[0099] The initial momentum of the DBN neural network model is set to 0.1;

[0100] The retention rate parameter for Dropout in the DBN neural network model is 0.95;

[0101] The number of hidden neuron nodes is set to a constant value, with 100 hidden neurons per layer;

[0102] The number of hidden layers is set to 3.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for diagnosing partial discharge faults in switchgear based on PSO-DBN neural networks, characterized in that, The specific method is as follows: Collect the characteristic gas content of the insulating oil in the switchgear to be diagnosed; The content of characteristic gases is input into the first diagnostic model to determine whether partial discharge has occurred; the first diagnostic model is a support vector machine (SVM). When partial discharge is detected, the change in characteristic gas over time is periodically read. Input the change of characteristic gas over time into the second diagnostic model to diagnose the specific fault type and location; The characteristic gases include NO2, O3 and CO, which are collected in real time by three gas sensors installed at the gas intake positions. The fault types include: air gap discharge, surface discharge, floating potential discharge, and corona discharge. The change in the characteristic gas over time includes: the duration of the change in the characteristic gas content and the number of changes in the characteristic gas content; The method for periodically reading the changes in characteristic gas over time is as follows: The preset reading period T and the preset reading duration are used to read the change of characteristic gas over time at the initial period interval; If the accuracy of the diagnostic result is less than the preset target after the reading result is input into the second diagnostic model, then adjust the preset reading period T and the preset duration until the accuracy of the diagnostic result is greater than or equal to the preset target. The second diagnostic model is a PSO-DBN neural network model, and the specific construction method is as follows: Inside the switchgear, several simulations were conducted to detect combinations of partial discharges at possible locations and with possible fault types. Obtain the time-varying characteristic gas of combined partial discharge; The changes in characteristic gases over time are normalized. The normalized information is then subjected to T-SNE dimensionality reduction. The training set is constructed using the dimensionality-reduced data; Set the hyperparameters of the DBN neural network; Perform DBN neural network testing; Perform PSO (Particle Swarm Optimization) algorithm initialization; Determine the optimal position and fitness of the particle swarm; Iteratively update the particle velocity and position until the preset target or the number of iterations is reached; The PSO optimization results are output and assigned to the DBN neural network to complete the construction of the PSO-DBN neural network model.

2. The method for partial discharge fault diagnosis of switchgear based on PSO-DBN neural network as described in claim 1, characterized in that, The specific construction method of the first diagnostic model is as follows: Simulate partial discharges in combination of possible discharge locations and possible fault types within the switchgear to obtain characteristic gas data; Preprocess the characteristic gas data and construct the dataset; Initialize the SVM diagnostic model and train it with the dataset.

3. The method for partial discharge fault diagnosis of switchgear based on PSO-DBN neural network as described in claim 2, characterized in that, The Support Vector Machine (SVM) includes the following settings during construction: Kernel functions: Gaussian sum function RBF RBF function parameters: kernel width σ=3 The optimal penalty coefficient C and slack variable l are: C=2 and l=0.

003.

4. The method for partial discharge fault diagnosis of switchgear based on PSO-DBN neural network as described in claim 1, characterized in that, When establishing the PSO-DBN neural network diagnostic model, the parameters and hyperparameters are designed as follows: The particle swarm population size m is set to 8; The number of iterations, n, is set to 20; Learning factors C1 and C2 are set to C1=C2=1.495; The particle length l is set to 4; The maximum particle velocity Vmax = k·Vmax, 0.1≤k≤1. For the dimension of optimizing the number of neurons in the hidden layer, we take k=1 and Vmax=1; for the dimension of optimizing the learning rate, we take k=0.1 and Vmax=0.

5. The particle range [Xmax, Xmin] is set to [0.001, 0.5]. Linear weights w: w max =0.8, w min =0.2, t max =20 The initial learning rate of the DBN neural network model was set to 0.1; The initial momentum of the DBN neural network model is set to 0.1; The retention rate parameter for Dropout in the DBN neural network model is 0.95; The number of hidden neuron nodes is set to a constant value, with 100 hidden neurons per layer; The number of hidden layers is set to 3.

Images