Convolutional neural network-based high-voltage switch cabinet partial discharge identification method

Abstract

The invention discloses a high-voltage switch cabinet partial discharge identification method based on a convolutional neural network, and relates to the technical field of high-voltage equipment live detection, a TEV method and an ultrasonic detection method are used to perform combined sampling on partial discharge signals, and a combined optimization algorithm is used to perform sampling optimization on the TEV method and the ultrasonic detection method on the partial discharge signals, so that the partial discharge identification accuracy is improved. A convolutional neural network is used for carrying out feature extraction and sample training on signals, finally, a model of the convolutional neural network is used for predicting and recognizing unknown signals, and comprehensive judgment is given to the operation condition of the switch cabinet. According to the method, the advantages of the TEV method and the ultrasonic detection method can be integrated, and meanwhile, the signal quality is improved by utilizing the adaptive filtering and preprocessing technology. As a powerful pattern recognition tool, the convolutional neural network can effectively classify and recognize complex partial discharge signals. Through the combined sampling, filtering, preprocessing and training method, the detection accuracy and reliability of the partial discharge signal can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of live detection of high voltage equipment, and in particular to a method for identifying partial discharge of a high voltage switch cabinet based on a convolutional neural network. Background Art

[0002] 10kV and 35kV high-voltage switchgear equipment is a very important electrical equipment in the power system. Its reliable operation is directly related to the power quality and power supply reliability of the power system. High-voltage switchgear is mainly composed of cabinet, circuit breaker, disconnector, lightning arrester, transformer and other components, used to transmit electric energy. Its structure and functional characteristics are prone to common fault types including: insulation fault, mechanical fault, current-carrying fault and others. The main manifestations of insulation fault are: external flashover breakdown, internal fault, porcelain bottle flashover breakdown explosion, phase-to-phase flashover breakdown, flashover breakdown caused by overvoltage, etc. Most of the high-voltage switchgear accidents that have occurred in recent years are caused by insulation degradation and other reasons. Insulation fault has become the main cause of switchgear accidents. Therefore, early detection of internal insulation faults in switchgear can effectively reduce the occurrence of switchgear accidents and improve the safety and stability of power grids and equipment.

[0003] Small electrical sparks are caused by local breakdown in a certain area of ​​the insulating medium of the power equipment. They are usually prone to occur in solid air gaps, liquid bubbles, interfaces between different insulating materials, sharp edges of metal conductors or semiconductor electrodes, and extremely uneven electric fields. If the air gap exists in organic solids or liquids, the organic insulating material will be corroded and the performance will gradually decline. It can be seen that the significance of partial discharge detection is to identify the early signs of insulation defects inside the high-voltage switchgear equipment as early as possible, so that countermeasures can be taken in time to avoid complete breakdown of the insulating medium of the switchgear and eliminate equipment safety hazards.

[0004] For partial discharge detection of high-voltage switchgear, commonly used technical methods include: pulse current method, ultrasonic method, transient low voltage method and ultra-high frequency method. At present, the commonly used feature extraction methods mostly obtain feature quantities through the design of artificial feature engineering, mainly including feature extraction methods based on spectrograms, feature extraction methods based on pulse waveforms, and feature extraction methods based on time-frequency analysis.

[0005] Although the above-mentioned feature extraction methods have achieved certain results in the identification of partial discharge types, they all extract the characteristics of the signal in one or several aspects through complex artificial design, which cannot well represent the overall change of the discharge signal characteristics, and are prone to omission or loss of feature information. Moreover, the artificial features extracted by feature engineering mostly represent the shallow information of the signal. In the subsequent identification, the ability of this information to express the characteristics of the partial discharge signal is very high, and it is difficult to ensure that the various discharge types are always maximally separable.

[0006] There are many methods for detecting partial discharge signals. However, due to the strong randomness of partial discharge signals and the presence of other complex characteristics when a fault occurs, it is difficult to extract partial discharge signals with high precision.

[0007] Therefore, a local discharge identification method for high-voltage switchgear based on convolutional neural networks has become an urgent problem to be solved. Summary of the invention

[0008] The purpose of the present invention is to provide a method for identifying partial discharge of high-voltage switchgear based on convolutional neural network, which uses TEV (transient voltage to earth) method and ultrasonic detection method to jointly sample partial discharge signals, uses joint optimization algorithm to optimize the sampling of partial discharge signals by TEV method and ultrasonic detection method, uses convolutional neural network to extract features and train samples of signals, and finally, uses its model to predict and identify unknown signals, and gives a comprehensive judgment on the operating status of the switchgear. This method can combine the advantages of TEV method and ultrasonic detection method, and at the same time use adaptive filtering and preprocessing technology to improve signal quality. As a powerful pattern recognition tool, convolutional neural network can effectively classify and identify complex partial discharge signals. Through this method of joint sampling, filtering, preprocessing and training, the detection accuracy and reliability of partial discharge signals can be improved.

[0009] To achieve the above object, the present invention provides a technical solution: a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network, comprising the following steps:

[0010] S1. Data preparation

[0011] Collect and pre-process TEV signals and ultrasonic signals in switchgear;

[0012] S2. Feature extraction

[0013] Design and apply a CNN model to extract key PD features from pre-processed signal data; these features include spectral features, time domain features, and waveform shape and pattern;

[0014] S3. Model training

[0015] The preprocessed data set is input into the CNN model for training. During the training process, CNN optimizes its internal weights and structure by learning the complex patterns and features of the signal data to improve its ability to recognize partial discharge features.

[0016] S4, verification and tuning

[0017] Use the validation set to verify and tune the trained CNN model to ensure its generalization ability and accuracy on unknown data;

[0018] S5. Eigenvalue extraction and classification

[0019] Use the trained CNN model to extract features and classify the newly collected partial discharge signals; the CNN model automatically identifies partial discharge events of different types and severity, providing accurate data support for the evaluation of equipment operating status.

[0020] Further, the TEV and ultrasonic combined sampling method is as follows:

[0021] (1) Initialize weights: First, randomly initialize the layers and connection weights of the neural network;

[0022] (2) Forward propagation: Through the neural network, the signal data of the TEV branch and the ultrasonic branch are input into their respective processing layers for feature extraction and representation learning;

[0023] (3) Signal fusion: Dynamically adjust the weights of the TEV branch and ultrasonic branch outputs by setting learnable parameters in the fusion layer; these parameters are updated according to the error back propagation process;

[0024] (4) Back propagation: Based on the error output by the neural network, the back propagation algorithm is used to calculate the gradient, and the weights of each layer and the parameters of the fusion layer are updated based on the gradient;

[0025] (5) Training and optimization: Through repeated forward propagation and backpropagation, the neural network gradually adjusts its weights so that its performance on the training set continues to improve. This process is tuned by monitoring the performance on the validation set to ensure that the neural network can generalize to unseen data sets.

[0026] Furthermore, the determination method based on the comprehensive evaluation of TEV and ultrasonic test data is as follows:

[0027] (1) Ultrasonic detection data determination process

[0028] Sound identification: judging whether there is partial discharge by sound characteristics;

[0029] Fixed value discrimination: set a threshold value, and if it exceeds the threshold value, it is considered an abnormal signal;

[0030] Longitudinal analysis: Analyze the time series changes of the signal to detect whether the signal is continuous or gradually strengthened;

[0031] (2) TEV test data determination process:

[0032] Fixed value judgment: set the threshold of transient ground voltage;

[0033] Horizontal analysis: compare voltage data at different time points to detect whether there are significant changes;

[0034] Longitudinal analysis: Analyze the time series of voltage signals to observe their changing trends and frequencies;

[0035] (3) Comprehensive evaluation process:

[0036] Normal state: All characteristic values ​​show no abnormality and the equipment operates well;

[0037] Abnormal (concern): Some characteristic values ​​show abnormal signs and require further observation and analysis;

[0038] Abnormal (warning): Multiple characteristic values ​​show obvious abnormalities, indicating that there may be problems with the equipment and detailed inspection and maintenance are required;

[0039] Dangerous state: The characteristic value shows serious abnormality, the operation of the equipment may have been affected, and it needs to be shut down for maintenance immediately.

[0040] The advantages of the present invention compared with the prior art are:

[0041] 1. Comprehensive performance: Ultrasonic detection and TEV partial discharge detection each have a good response to different types of partial discharge faults. Joint sampling can comprehensively utilize the advantages of the two methods to improve the detection capabilities of different types of faults.

[0042] 2. Accuracy: CNN can extract local features in the signal through multi-layer convolution operations, thereby more accurately describing the characteristics of partial discharge signals and helping to distinguish different types of discharge signals and fault modes.

[0043] 3. High efficiency: Adaptive filtering technology can effectively process a large amount of sampled data and improve the efficiency and accuracy of data processing. CNN has high computing efficiency and processing speed when processing large-scale data. It can quickly and accurately extract features from partial discharge signals, which is suitable for real-time monitoring and fault diagnosis, and improves the efficiency of data processing.

[0044] 4. Cost savings: By jointly sampling TEV and ultrasonic partial discharge signals, the time and cost of equipment maintenance and troubleshooting can be reduced, the maintainability and reliability of equipment can be improved, and the equipment operation risk and maintenance cost can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is an ultrasonic judgment flow chart; in the figure, P is each ultrasonic test value of the switch cabinet, and Δq% is the relative deviation of the average value of the switch cabinet test.

[0046] Figure 2It is a transient ground voltage judgment flow chart; in the figure, A is the average value of the transient ground voltage test (the same voltage level in the same switch room), P is each transient ground voltage test value of the switch cabinet, Δm% is the offset of each test value of the switch cabinet, and Δq% is the relative deviation of the average value of the switch cabinet test.

[0047] Figure 3 The present invention is a flow chart of a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network.

[0048] Figure 4 It is a flow chart of the feature value extraction method of convolutional neural network. DETAILED DESCRIPTION

[0049] The following is a further detailed description of a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network according to the present invention in conjunction with the accompanying drawings.

[0050] Combined with Figure 1-4 The specific implementation process of the method for identifying partial discharge of high-voltage switchgear based on convolutional neural network of the present invention is as follows:

[0051] A method for identifying partial discharge of high-voltage switchgear based on convolutional neural network is proposed. It uses a joint sampling method of TEV and ultrasonic wave based on joint optimization algorithm technology, a partial discharge feature value extraction method based on convolutional neural network technology, and a partial discharge fault constant value discrimination method to evaluate the operating status of the switchgear and give a comprehensive evaluation. Among them, the joint optimization algorithm refers to a type of algorithm that optimizes the complementarity and synergy of TEV method and ultrasonic detection method in partial discharge signal detection.

[0052] The method for identifying partial discharge of high-voltage switchgear is as follows:

[0053] Weighted Optimization of Signal Fusion:

[0054] By weighted fusion of the signals collected by the TEV method and the ultrasonic detection method, the weight is dynamically adjusted according to the performance and importance of each signal under different conditions. The weight optimization can be based on real-time feedback and historical data analysis, so that the final fused signal has the best partial discharge signal detection effect.

[0055] For example, the back-propagation algorithm can be used to learn and adjust the weights of different signal sources in a neural network to maximize the performance of the overall system. Suppose we have a deep neural network for processing data from two different signal sources, TEV and ultrasonic detection, and the goal is to effectively fuse them to improve the accuracy of partial discharge signal detection.

[0056] Example scenario:

[0057] Neural Network Architecture:

[0058] We designed a neural network consisting of two branches:

[0059] TEV branch: processes signal data from the TEV method.

[0060] Ultrasonic branch: processes signal data from ultrasonic testing methods.

[0061] The outputs of these two branches will be fed into a fusion layer, which uses the back-propagation algorithm to learn and adjust the weights of different signal sources to maximize the performance of the overall system, i.e., improve the detection accuracy and sensitivity of partial discharge signals.

[0062] Steps and methods:

[0063] 1. Initialize weights: First, the layers and connection weights of the neural network are randomly initialized.

[0064] 2. Forward propagation: Through the neural network, the signal data of the TEV branch and the ultrasonic branch are input into their respective processing layers for feature extraction and representation learning.

[0065] 3. Signal fusion: In the fusion layer, a mechanism is designed to dynamically adjust the weights of the TEV branch and the ultrasonic branch outputs. This can be achieved by setting learnable parameters in the fusion layer, which are updated according to the error back propagation process.

[0066] 4. Back propagation: Based on the error output by the neural network, the back propagation algorithm is used to calculate the gradient, and the weights of each layer and the parameters of the fusion layer are updated according to the gradient. This process enables the neural network to learn the best weight combination of the TEV branch and the ultrasonic branch to optimize the detection effect of the partial discharge signal.

[0067] 5. Training and optimization: Through repeated forward propagation and back propagation, the neural network gradually adjusts its weights so that its performance on the training set continues to improve. This process can be tuned by monitoring the performance on the validation set to ensure that the neural network can generalize to unseen data sets.

[0068] In this way, the neural network can gradually adjust the weights of the TEV method and the ultrasonic detection method during the learning process, thereby achieving the optimal signal fusion strategy. This method not only improves the sensitivity and accuracy of partial discharge signals, but also can adapt to different working conditions and signal characteristics. It is a typical application example of using the back propagation algorithm to optimize signal fusion.

[0069] Implementation of partial discharge feature value extraction method based on convolutional neural network technology:

[0070] The partial discharge feature value extraction method based on convolutional neural network (CNN) technology is usually used to process partial discharge monitoring data of high-voltage switchgear. The following are the implementation steps of this method:

[0071] 1. Data preparation:

[0072] Collect and pre-process TEV and ultrasonic signals in switch cabinets. Ensure that the signals undergo pre-processing steps such as amplification, filtering and denoising to improve the effect and accuracy of subsequent processing.

[0073] 2. Feature extraction:

[0074] A convolutional neural network model is designed and applied to extract key partial discharge features from the preprocessed signal data. These features can include spectral features, time domain features, and the shape and pattern of the waveform.

[0075] 3. Model training:

[0076] The preprocessed data set is input into the CNN model for training. During the training process, CNN optimizes its internal weights and structure by learning the complex patterns and features of the signal data to improve its ability to recognize partial discharge features.

[0077] 4. Verification and tuning:

[0078] The validation set is used to verify and tune the trained CNN model to ensure its generalization ability and accuracy on unknown data.

[0079] 5. Feature extraction and classification:

[0080] The trained CNN model is used to extract features and classify newly acquired partial discharge signals. The model can automatically identify partial discharge events of different types and severity, providing accurate data support for the evaluation of equipment operating status.

[0081] Through the above steps, the partial discharge feature value extraction method based on convolutional neural network technology can effectively extract key feature information from complex monitoring data and realize accurate identification and classification of partial discharge activities in switch cabinets. This method not only improves the response speed and accuracy of the monitoring system, but also provides strong support for equipment maintenance and fault warning.

[0082] Comprehensive evaluation of switchgear operating status:

[0083] The transient and ultrasonic detection methods can refer to the judgment logic in the attached page, and no detailed description is given here.

[0084] The judgment process of comprehensive evaluation of the switchgear operating status is based on the extracted feature values ​​and is mainly divided into four levels: normal, abnormal (concern), abnormal (warning), and dangerous. The following are the main judgment methods based on transient ground voltage and ultrasonic detection data:

[0085] 1. Ultrasonic detection data determination process:

[0086] Sound identification: Determine whether partial discharge occurs by sound characteristics.

[0087] Fixed value judgment: set a threshold value, and if it exceeds the threshold value, it will be identified as an abnormal signal.

[0088] Longitudinal analysis: Analyze the time series changes of the signal to detect whether the signal is continuous or gradually strengthened.

[0089] 2. Transient ground voltage detection data determination process:

[0090] Constant value judgment: set the threshold of transient ground voltage.

[0091] Horizontal analysis: Compare voltage data at different time points to detect whether there are significant changes.

[0092] Longitudinal analysis: Analyze the time series of voltage signals and observe their changing trends and frequencies.

[0093] Comprehensive evaluation process:

[0094] Normal status: All characteristic values ​​show no abnormality and the equipment operates well.

[0095] Abnormal (concern): Some characteristic values ​​show abnormal signs and require further observation and analysis.

[0096] Abnormal (warning): Multiple characteristic values ​​show obvious abnormalities, indicating that there may be problems with the equipment and detailed inspection and maintenance are required.

[0097] Dangerous state: The characteristic value shows serious abnormality, the operation of the equipment may have been affected, and it needs to be shut down for maintenance immediately.

[0098] Through such a comprehensive evaluation method, problems in equipment operation can be discovered in time, and corresponding maintenance and repair measures can be taken to ensure the safe operation of the equipment and extend the service life of the equipment. The judgment process is shown in the figure below, where horizontal analysis is applicable when the test data of a single or multiple switch cabinets is significantly different from the data of other switch cabinets; vertical analysis is applicable to the detection judgment of shortening the cycle of switch cabinets judged to be abnormal (concerned).

[0099] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A method for identifying partial discharge in high-voltage switchgear based on convolutional neural network. Features: The following steps are included: S1. Data preparation Collect and pre-process TEV signals and ultrasonic signals in switchgear; S2. Feature extraction Design and apply a CNN model to extract key PD features from preprocessed signal data; These features include spectral characteristics, time domain characteristics, and the shape and pattern of the waveform; S3. Model training Input the preprocessed dataset into the CNN model for training; During the training process, CNN optimizes its internal weights and structure by learning the complex patterns and features of signal data to improve its ability to identify partial discharge features; S4, verification and tuning Use the validation set to verify and tune the trained CNN model to ensure its generalization ability and accuracy on unknown data; S5. Eigenvalue extraction and classification Use the trained CNN model to extract features and classify the newly collected partial discharge signals; the CNN model automatically identifies partial discharge events of different types and severity, providing accurate data support for the evaluation of equipment operating status.

2. According to claim 1, a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network, Features: The specific method of data preparation in step S1 is as follows: By weighted fusion of the signals collected by TEV and ultrasonic detection methods, the weights are dynamically adjusted according to the performance and importance of each signal under different conditions; the weight optimization is based on real-time feedback and historical data analysis, so that the final fused signal has the best partial discharge signal detection effect.

3. According to claim 2, a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network, Features: The back-propagation algorithm is used to learn and adjust the weights of different signal sources in the neural network to maximize the performance of the overall system.

4. According to claim 3, a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network, Features: The combined TEV and ultrasonic sampling method is as follows: (1) Initialize weights: First, randomly initialize the layers and connection weights of the neural network; (2) Forward propagation: Through the neural network, the signal data of the TEV branch and the ultrasonic branch are input into their respective processing layers for feature extraction and representation learning; (3) Signal fusion: Dynamically adjust the weights of the TEV branch and ultrasonic branch outputs by setting learnable parameters in the fusion layer; these parameters are updated according to the error back propagation process; (4) Back propagation: Based on the error output by the neural network, the back propagation algorithm is used to calculate the gradient, and the weights of each layer and the parameters of the fusion layer are updated based on the gradient; (5) Training and optimization: Through repeated forward propagation and backpropagation, the neural network gradually adjusts its weights so that its performance on the training set continues to improve. This process is tuned by monitoring the performance on the validation set to ensure that the neural network can generalize to unseen data sets.

5. According to claim 4, a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network, Features: The operating status of the switchgear is evaluated based on the extracted feature values, and a comprehensive evaluation is given, which is divided into four levels: normal, abnormal (concern), abnormal (warning), and dangerous.

6. According to claim 5, a method for identifying partial discharge of a high-voltage switch cabinet based on a convolutional neural network, Features: The judgment method based on the comprehensive evaluation of TEV and ultrasonic test data is as follows: (1) Ultrasonic detection data determination process Sound identification: judging whether there is partial discharge by sound characteristics; Fixed value discrimination: set a threshold value, and if it exceeds the threshold value, it is considered an abnormal signal; Longitudinal analysis: Analyze the time series changes of the signal to detect whether the signal is continuous or gradually strengthened; (2) TEV test data determination process: Fixed value judgment: set the threshold of transient ground voltage; Horizontal analysis: compare voltage data at different time points to detect whether there are significant changes; Longitudinal analysis: Analyze the time series of voltage signals to observe their changing trends and frequencies; (3) Comprehensive evaluation process: Normal state: All characteristic values ​​show no abnormality and the equipment operates well; Abnormal (concern): Some characteristic values ​​show abnormal signs and require further observation and analysis; Abnormal (warning): Multiple characteristic values ​​show obvious abnormalities, indicating that there may be problems with the equipment and detailed inspection and maintenance are required; Dangerous state: The characteristic value shows serious abnormality, the operation of the equipment may have been affected, and it needs to be shut down for maintenance immediately.

Images