A three-phase traveling wave fault diagnosis model construction method and fault diagnosis method

Abstract

The present invention relates to a method for constructing a three-phase traveling wave fault diagnosis model, including obtaining a traveling wave signal of a three-phase signal, preprocessing the traveling wave signal; converting the preprocessed traveling wave signal into a GASF image; inputting the GASF image into a multi-channel convolutional neural network, each channel corresponding to a traveling wave signal of a different phase of the GASF image; optimizing the hyperparameters of the multi-channel convolutional neural network using a whale optimization algorithm; and training the multi-channel convolutional neural network to obtain a fault diagnosis model according to fault information corresponding to the traveling wave signal. The present invention automatically optimizes the hyperparameters of the multi-channel convolutional neural network through a whale optimization algorithm, realizes the search for the global optimal hyperparameter combination, improves the training efficiency, training effect and generalization ability of the neural network; realizes the complete process from data acquisition, preprocessing, feature extraction, model optimization to fault identification, and realizes efficient and accurate diagnosis of traveling wave signal faults.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a three-phase traveling wave fault diagnosis model construction method and a fault diagnosis method. Background Art

[0002] In the power system, traveling wave fault diagnosis and location is a key technology to ensure the safe and stable operation of the power system. Rapidly and accurately locating the fault point can not only reduce the fault outage time and improve the reliability of the power system, but also reduce maintenance costs and economic losses.

[0003] Traveling wave fault diagnosis technology is extremely important in power systems. Through the analysis of traveling wave signals, it is possible to quickly and accurately locate and classify faults, thereby ensuring the safe operation of the power system and improving the reliability and economic benefits of the power system. Although current technical applications have achieved certain results, there are still many problems and challenges that need to be solved.

[0004] In practical applications, traveling wave signals are often interfered by various noises, such as electromagnetic interference and other random noises. Some existing methods perform poorly in dealing with noise and lack effective noise filtering mechanisms. Many existing technologies rely on manually adjusting hyperparameters (such as learning rate, number of neurons, etc.) when using neural networks or deep learning models. In a three-phase system, signals of different phases carry different information, and existing methods are often unable to efficiently process multi-phase signals at the same time. This leads to poor detection of complex faults such as cross-phase faults, increasing the risk of misdiagnosis and missed diagnosis. These defects limit the application and effectiveness of traditional fault diagnosis techniques in modern complex power systems. Therefore, it is necessary to develop more intelligent, efficient and robust fault diagnosis methods to cope with the growing complexity and reliability requirements of power systems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a three-phase traveling wave fault diagnosis model construction method and a fault diagnosis method to overcome the deficiencies in the above-mentioned prior art.

[0006] The technical solution of the present invention to solve the above technical problem is as follows: A method for constructing a three-phase traveling wave fault diagnosis model comprises the following steps:

[0007] Step S01: obtaining a traveling wave signal of a three-phase signal from an existing traveling wave signal library, and preprocessing the traveling wave signal; wherein the preprocessing includes standardization processing and noise filtering processing;

[0008] Step S02: converting the preprocessed traveling wave signal into a GASF image;

[0009] Step S03: inputting the GASF image into a multi-channel convolutional neural network, where each channel corresponds to a traveling wave signal of a different phase of the GASF image;

[0010] Step S04: Optimizing the hyperparameters of the multi-channel convolutional neural network using the whale optimization algorithm;

[0011] Step S05: According to the fault information corresponding to the traveling wave signal, a multi-channel convolutional neural network is trained to obtain a fault diagnosis model; wherein the fault information includes the fault type, fault phase and fault location.

[0012] The beneficial effects of the present invention are as follows: the present invention utilizes the Gram angle cosine field method to convert the time series data of the traveling wave signal of the three-phase signal into a two-dimensional image; this image representation not only retains the time series characteristics of the original signal, but also improves the accuracy of fault signal analysis through image feature extraction; for the traveling wave signal of the three-phase signal, a multi-channel convolutional neural network structure is designed, each channel input corresponds to a Gram angle cosine field image of one phase, and the multi-dimensional features of the signal are extracted through multi-channel parallel processing; this architecture effectively enhances the model's recognition ability for different fault signals; the hyperparameters of the multi-channel convolutional neural network are automatically optimized through the whale optimization algorithm, the algorithm simulates the foraging behavior of whales, realizes the search for the global optimal hyperparameter combination, and improves the training efficiency, training effect and generalization ability of the neural network.

[0013] Combining deep learning technology with traditional signal processing methods, especially feature fusion technology after multiple convolution and activation function processing, further improves the depth and breadth of signal analysis, thereby achieving more accurate fault classification and positioning; the present invention provides a complete process from data acquisition, preprocessing, feature extraction, model optimization to fault identification, and through the integrated application of the above innovative technologies, efficient and accurate diagnosis of traveling wave signal faults is achieved.

[0014] Based on the above technical solution, the present invention can also be improved as follows.

[0015] Further, step S01 specifically includes the following steps:

[0016] Step S11: Check the timestamp or sampling point of the traveling wave signal. If the sampling points are inconsistent, perform alignment processing. If there is a time misalignment or a signal with different sampling frequencies, use interpolation to compensate.

[0017] Step S12: performing standardization processing on the traveling wave signal to eliminate the amplitude difference between different signals.

[0018] Furthermore, step S01 further includes the following steps:

[0019] Step S13: Design a first-order low-pass filter to perform noise filtering on the normalized traveling wave signal. The transfer function of the first-order low-pass filter is Defined as:

[0020] ,

[0021] in, is the time constant, which determines the cutoff frequency of the filter; is a complex frequency domain variable.

[0022] Further, step S03 specifically includes the following steps:

[0023] Step S31: each channel of the input layer of the multi-channel convolutional neural network corresponds to GASF image signal data of one phase;

[0024] Convolution first applies a convolution kernel to each channel of the input layer to generate a corresponding number of intermediate feature maps with the same number of channels. The mathematical expression of deep convolution is:

[0025] ,

[0026] in, Indicates the channel, Indicates convolution kernels, is the bias term;

[0027] Step S32: The feature data of the traveling wave signal is spliced ​​through a multi-channel convolutional neural network, and then goes through two convolutional layers and one fully connected layer, and finally a Softmax layer for classification output.

[0028] Further, step S04 specifically includes:

[0029] The whale optimization algorithm is used to iteratively optimize the hyperparameters of the multi-channel convolutional neural network. If the maximum number of iterations is reached or the preset convergence conditions are met, the iteration is stopped and the optimal hyperparameters are output.

[0030] Furthermore, the hyperparameters of the multi-channel convolutional neural network in step S04 include convolution kernel size, number of network layers, and learning rate.

[0031] The present invention also discloses a fault diagnosis method, which deploys the fault diagnosis model constructed by the above-mentioned three-phase traveling wave fault diagnosis model construction method into an actual fault diagnosis system, monitors the traveling wave signal in real time and performs fault classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a method for constructing a fault diagnosis model of the present invention;

[0033] Figure 2 It is a flow chart of the traveling wave signal preprocessing of the present invention;

[0034] Figure 3 The overall flow chart of multi-channel neural network feature extraction of the present invention;

[0035] Figure 4 This is a multi-channel neural network feature fusion output flow chart of the present invention. DETAILED DESCRIPTION

[0036] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0037] like Figure 1 to Figure 4 As shown, in embodiment 1, a method for constructing a three-phase traveling wave fault diagnosis model comprises the following steps:

[0038] Step S01: Obtain the traveling wave signal of the three-phase signal from the existing traveling wave signal library, and pre-process the traveling wave signal; wherein the pre-processing includes standardization processing and noise filtering processing; in specific implementation, the existing traveling wave signal library can be the traveling wave signal of the three-phase signal obtained from the sensor or data acquisition device in the power system; process the three-phase signal (phase A, phase B, phase C) separately to ensure the independence of each phase signal; when performing subsequent analysis, maintain the synchronization and integrity of the three-phase signal. There may be noise in the traveling wave signal, such as environmental noise, electromagnetic interference, etc. The collected raw data is subjected to noise filtering processing to remove unnecessary high-frequency noise or interference of specific frequencies. In order to avoid the order of magnitude difference between features affecting the training effect of the model, the traveling wave signal is standardized;

[0039] Step S02: converting the preprocessed traveling wave signal into a GASF image (Gram angle cosine field image); the GASF image can convert the one-dimensional time series data into a two-dimensional image form, retain the time series characteristics and enhance the visual representation of the signal pattern;

[0040] Gramian Angular Summation Field (GASF) is a method that converts one-dimensional time series data into a two-dimensional matrix representation, which is often used in image recognition and signal processing. GASF conversion not only retains the original information of the time series, but also can intuitively display the characteristics of the data in the form of images. The basic principles of GASF are as follows:

[0041] GASF is calculated based on the angle encoding of the time series and the cosine function. The core idea is to map the time series data to the polar coordinate system and then use the cosine function to construct a two-dimensional matrix.

[0042] Map the normalized time series to the polar coordinate system. , the angle at each time point can be calculated :

[0043] ,

[0044] By calculating the cosine function, the angle information is converted into a two-dimensional matrix representation, namely the GASF matrix. This matrix retains the temporal relationship and amplitude information of the time series. Converting the time series data from one dimension to two dimensions allows it to be input as an image into the deep learning model for processing.

[0045] Step S03: Input the GASF image into a multi-channel convolutional neural network, where each channel corresponds to a traveling wave signal of a different phase of the GASF image; such a design allows the network to simultaneously learn and extract information from each channel, making full use of the interrelationships of multi-phase signals;

[0046] Step S04: Optimize the hyperparameters of the multi-channel convolutional neural network using the whale optimization algorithm to improve the performance and accuracy of the model;

[0047] Step S05: According to the fault information corresponding to the traveling wave signal, a multi-channel convolutional neural network is trained to obtain a fault diagnosis model; wherein the fault information includes the fault type (short circuit, ground fault), fault phase (phase A, phase B, phase C) and fault location.

[0048] The present invention utilizes the Gram angle cosine field method to convert the time series data of the traveling wave signal of the three-phase signal into a two-dimensional image; this image representation not only retains the time series characteristics of the original signal, but also improves the accuracy of fault signal analysis through image feature extraction; for the traveling wave signal of the three-phase signal, a multi-channel convolutional neural network structure is designed, each channel input corresponds to a phase of the Gram angle cosine field image, and the multi-dimensional features of the signal are extracted through multi-channel parallel processing; this architecture effectively enhances the model's recognition ability for different fault signals; the hyperparameters of the multi-channel convolutional neural network are automatically optimized through the whale optimization algorithm, the algorithm simulates the foraging behavior of whales, realizes the search for the global optimal hyperparameter combination, and improves the training efficiency, training effect and generalization ability of the neural network.

[0049] Combining deep learning technology with traditional signal processing methods, especially feature fusion technology after multiple convolution and activation function processing, further improves the depth and breadth of signal analysis, thereby achieving more accurate fault classification and positioning; the present invention provides a complete process from data acquisition, preprocessing, feature extraction, model optimization to fault identification, and through the integrated application of the above innovative technologies, efficient and accurate diagnosis of traveling wave signal faults is achieved.

[0050] Embodiment 2: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0051] Step S01 specifically includes the following steps:

[0052] Step S11: Check the timestamp or sampling point of the traveling wave signal. If the sampling points are inconsistent, perform alignment processing. If there is a time misalignment or a signal with different sampling frequencies, use interpolation to compensate. To ensure the time synchronization of the traveling wave signal and avoid the time misalignment problem caused by sampling differences, first check the timestamp or sampling point of the traveling wave signal to ensure that the signal data is sampled under the same time reference. If the sampling points are found to be inconsistent, alignment processing is required. For signals with time misalignment or different sampling frequencies, use interpolation to compensate and align feature points to further ensure synchronization.

[0053] Step S12: Standardize the traveling wave signal to eliminate the amplitude differences between different signals, so that all signals can be analyzed on the same scale.

[0054] Data standardization is a data preprocessing technique that aims to scale data features to a standard range so that different features can be compared on the same scale. This processing method is particularly important for datasets with large feature differences because it can prevent certain features from dominating other features in model training. Standardization helps speed up model convergence and improves model performance and stability.

[0055] For a given input dataset , calculate the feature mean of each feature , the calculation formula is:

[0056] ,

[0057] in, is the sample size, It is The original eigenvalues ​​of samples;

[0058] Calculate the raw eigenvalues ​​for each feature Standard Deviation , used to measure the original eigenvalue The degree of dispersion; standard deviation The calculation formula is:

[0059] ,

[0060] The original feature value of each sample After standardization, the eigenvalues ​​after standardization The calculation formula is:

[0061] ,

[0062] Through the above steps, data standardization can scale different features to the same scale, thereby eliminating the dimensional differences between features and ensuring that each feature is treated fairly in model training.

[0063] Embodiment 3, this embodiment is a further improvement on the basis of embodiment 2, and its details are as follows:

[0064] Step S01 also includes the following steps:

[0065] Step S13: Design a first-order low-pass filter to perform noise filtering on the normalized traveling wave signal. The transfer function of the first-order low-pass filter is Defined as:

[0066] ,

[0067] in, is the time constant, which determines the cutoff frequency of the filter; is a complex frequency domain variable.

[0068] Since the noise in the traveling wave signal is mainly high-frequency components, the low-pass filter is a commonly used signal processing tool. Its main function is to allow low-frequency signals to pass through while suppressing high-frequency noise. In order to improve the design and application accuracy of the filter, the present invention uses a first-order low-pass filter for data processing.

[0069] Embodiment 4: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0070] Step S03 specifically includes the following steps:

[0071] Step S31: each channel of the input layer of the multi-channel convolutional neural network corresponds to GASF image signal data of one phase;

[0072] Convolution first applies a convolution kernel to each channel of the input layer to generate a corresponding number of intermediate feature maps with the same number of channels. The mathematical expression of deep convolution is:

[0073] ,

[0074] in, Indicates the channel, Indicates convolution kernels, is the bias term;

[0075] The three-convolution layer + LReLU structure enables the model to extract and learn more complex and abstract features layer by layer. Compared with shallow networks, this deep structure can better capture the patterns and regularities in the signal. Using GASF images as input can fully utilize the advantages of convolutional neural networks in image processing and capture the spatiotemporal characteristics of time series data, thereby improving the accuracy and robustness of fault diagnosis.

[0076] LReLU stands for Leaky ReLU, which is a variant of the rectified linear unit (ReLU) with better performance. Its function expression is:

[0077] ,

[0078] Step S32: The feature data of the traveling wave signal is spliced ​​through a multi-channel convolutional neural network, and then goes through two convolutional layers and one fully connected layer, and finally a Softmax layer for classification output.

[0079] Through the above steps, the multi-channel convolutional neural network can extract and fuse the features of different signals from the traveling wave signal. After deep convolution and nonlinear activation, fault classification is performed through the fully connected layer and Softmax output. This structure can make full use of the powerful feature extraction ability of the deep learning model, extract meaningful information from complex multi-phase signals, and achieve high-precision fault diagnosis.

[0080] Compared with single-channel convolutional neural networks, the traveling wave signal of three-phase signals can simultaneously process signal data from multiple phases, adapting to the characteristics of multi-phase signals in power systems; it can simultaneously process data from multiple input channels. For the traveling wave signal of three-phase signals in power systems, each channel represents the signal of one phase; the multi-channel neural network can extract the characteristics of each phase signal in parallel, while capturing the relationships and interactions between them.

[0081] The features extracted by the convolutional layer are fused in subsequent layers to form a higher-level comprehensive feature representation. This fusion process allows the network to understand the relationship between signals from different channels, which helps to improve the accuracy of fault classification.

[0082] Embodiment 5, this embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0083] Step S04 specifically includes:

[0084] The whale optimization algorithm is used to iteratively optimize the hyperparameters of the multi-channel convolutional neural network. If the maximum number of iterations is reached or the preset convergence conditions are met, the iteration is stopped and the optimal hyperparameters are output.

[0085] The Whale Optimization Algorithm (WOA) is a global optimization algorithm based on whale foraging behavior, simulating the process of humpback whales encircling prey. The algorithm performs well in solving complex optimization problems, especially in multidimensional nonlinear problems, with good search ability and convergence. In the present invention, the Whale Optimization Algorithm is used to optimize the hyperparameters of the multi-channel convolutional neural network to improve the performance and generalization ability of the model. The application steps of the Whale Optimization Algorithm in the present invention are as follows:

[0086] Randomly initialize a set of "whales" (candidate solutions) within the domain of hyperparameters. Each "whale" represents a set of hyperparameter combinations, such as learning rate, convolution kernel size, number of network layers, etc., and set the control parameters of the algorithm, such as population size, maximum number of iterations, etc.

[0087] Use the loss function value on the cross-validation or training set (such as cross entropy loss) as the fitness function to evaluate the quality of each candidate solution, train the model for each "whale" hyperparameter combination, and calculate its performance index on the validation set as the fitness value;

[0088] According to the position of the current optimal solution, other "whales" move closer to the optimal solution through the shrinking circle encirclement mechanism, and the update position formula is:

[0089] ,

[0090] in, is the current global optimal solution, is the current solution, and are all random vectors,

[0091] Used to control the search scope and direction;

[0092] Simulate the spiral hunting behavior of whales, calculate the distance D between the "whale" and the prey, and update the position using the following formula:

[0093] ,

[0094] in, is the absolute value of the distance vector, is a constant, yes A random number between

[0095] If the random vector If the absolute value of is greater than 1, the algorithm tends to explore new areas to avoid falling into the local optimal solution;

[0096] During the exploration, the position of the "whale" is updated as follows:

[0097] ,

[0098] in, is a randomly selected solution from the population;

[0099] If the maximum number of iterations is reached or the preset convergence condition is met, the iteration is stopped and the optimal hyperparameter combination is output. The optimal hyperparameter combination obtained by the whale optimization algorithm is applied to the final training and testing of the multi-channel convolutional neural network to obtain the best fault diagnosis performance.

[0100] Embodiment 6: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0101] The hyperparameters of the multi-channel convolutional neural network in step S04 include convolution kernel size, number of network layers, and learning rate.

[0102] Embodiment 7, a fault diagnosis method, deploys a fault diagnosis model constructed by a three-phase traveling wave fault diagnosis model construction method of any one of embodiments 1 to 6 into an actual fault diagnosis system, monitors the traveling wave signal in real time and performs fault classification. In specific implementation, it can be integrated with other monitoring and control equipment of the power system to provide fault alarms and recommended measures.

[0103] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for constructing a three-phase traveling wave fault diagnosis model, characterized in that: The steps include: Step S01: obtaining a traveling wave signal of a three-phase signal from an existing traveling wave signal library, and preprocessing the traveling wave signal; wherein the preprocessing includes standardization processing and noise filtering processing; Step S02: converting the preprocessed traveling wave signal into a GASF image; Step S03: inputting the GASF image into a multi-channel convolutional neural network, wherein each channel corresponds to a traveling wave signal of a different phase of the GASF image; Step S04: Optimizing the hyperparameters of the multi-channel convolutional neural network using the whale optimization algorithm; Step S05: training the multi-channel convolutional neural network to obtain a fault diagnosis model according to the fault information corresponding to the traveling wave signal; wherein the fault information includes the fault type, fault phase and fault location; The step S01 also includes designing a first-order low-pass filter to perform noise filtering on the normalized traveling wave signal, wherein the transfer function of the first-order low-pass filter is Defined as: ， in, is the time constant, which determines the cutoff frequency of the filter; is a complex frequency domain variable; The step S03 specifically includes the following steps: Step S31: each channel of the input layer of the multi-channel convolutional neural network corresponds to GASF image signal data of one phase; Convolution first applies a convolution kernel to each channel of the input layer to generate a corresponding number of intermediate feature maps with the same number of channels. The mathematical expression of deep convolution is: ， in, Indicates the channel, Indicates convolution kernels, is the bias term; Step S32: The feature data of the traveling wave signal is spliced ​​through a multi-channel convolutional neural network, and then goes through two convolutional layers and one fully connected layer, and finally a Softmax layer for classification output.

2. A three-phase traveling wave fault diagnosis model construction method according to claim 1, characterized in that: The step S01 specifically includes the following steps: Step S11: Check the timestamp or sampling point of the traveling wave signal, and perform alignment processing if the sampling points are inconsistent. If there is a time misalignment or a signal with different sampling frequencies, use interpolation to compensate; Step S12: performing standardization processing on the traveling wave signal to eliminate the amplitude difference between different signals.

3. A three-phase traveling wave fault diagnosis model construction method according to claim 1, characterized in that: The step S04 specifically includes: The whale optimization algorithm is used to iteratively optimize the hyperparameters of the multi-channel convolutional neural network. If the maximum number of iterations is reached or the preset convergence conditions are met, the iteration is stopped and the optimal hyperparameters are output.

4. A three-phase traveling wave fault diagnosis model construction method according to claim 1, characterized in that: The hyperparameters of the multi-channel convolutional neural network in step S04 include convolution kernel size, number of network layers and learning rate.

5. A fault diagnosis method, characterized in that: The fault diagnosis model constructed by the three-phase traveling wave fault diagnosis model construction method described in any one of claims 1 to 4 is deployed in an actual fault diagnosis system to monitor the traveling wave signal in real time and perform fault classification.

Images