Training method and use method of traveling wave fault classification model and related device

Abstract

The application relates to the technical field of power failure processing, and discloses a traveling wave fault classification model training method, a traveling wave fault classification model use method and related devices, the training method comprising the following steps: acquiring first time-frequency feature data of sample traveling wave signals of different fault types; performing clustering training on a preset self-organizing neural network by using the first time-frequency feature data, obtaining a target self-organizing neural network after training and a plurality of first clustering clusters; determining real fault type labels of the first clustering clusters; performing fault type prediction training by using the first clustering clusters, the real fault type labels and a preset deep learning model, and obtaining a target deep learning model for predicting the real fault type labels of the clustering clusters. In the manner, the dynamic characteristics of different types of faults are extracted by clustering and extracting time-frequency features by using a self-organizing neural network, and then the deep learning model learns the dynamic characteristics of different types of faults to predict the fault types, thereby improving the diagnosis capability for hidden faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power failure processing, and in particular to a traveling wave fault classification model training method, a traveling wave fault classification model using method and related devices. BACKGROUND

[0002] Hidden faults (such as tree barriers, line house faults, insulator breakdown, lightning arrester breakdown, etc.) are a type of fault in overhead line power systems that is difficult to directly detect, with weak and complex signals. Traditional fault identification methods cannot accurately classify hidden faults, and the diagnostic ability of hidden faults is poor. SUMMARY

[0003] The main purpose of the present application is to provide a traveling wave fault classification model training method, a traveling wave fault classification model using method and related devices, which can solve the problem of poor diagnostic ability of hidden faults in the prior art.

[0004] To achieve the above-mentioned purpose, the first aspect of the present application provides a traveling wave fault classification model training method, which comprises:

[0005] Obtaining first time-frequency feature data of sample traveling wave signals of different fault types;

[0006] Using the first time-frequency feature data to perform clustering training on a preset self-organizing neural network to obtain a trained target self-organizing neural network and a plurality of first clustering clusters;

[0007] Determining the real fault type label of each first clustering cluster;

[0008] Using the first clustering cluster, the real fault type label and a preset deep learning model to perform fault type prediction training to obtain a trained target deep learning model, the target deep learning model being used to predict the real fault type label corresponding to the clustering cluster, and the traveling wave fault classification model comprising the target self-organizing neural network and the target deep learning model.

[0009] In a feasible implementation manner, the self-organizing neural network comprises a plurality of nodes, the dimension of the weight vector of each node being the same as the feature dimension of the first time-frequency feature data, and the using the first time-frequency feature data to perform clustering training on the preset self-organizing neural network to obtain the trained target self-organizing neural network and the plurality of first clustering clusters comprises:

[0010] Initializing the weight vector of each node of the self-organizing neural network;

[0011] Randomly selecting one of the plurality of first time-frequency feature data as an input vector for the t-th round of training, t being the training round number, and t having an initial value of 1;

[0012] calculating similarity between the input vector and the current weight vector of each node;

[0013] determining the node with the highest similarity to the input vector among the nodes as a winning node, and mapping the input vector to the winning node to obtain an initial first clustering cluster centered on each node;

[0014] updating the current weight vector of the winning node and the nodes in the neighborhood of the winning node to obtain an updated weight vector;

[0015] judging whether the self-organizing neural network converges or not by using a preset convergence condition and current characteristic parameters of the self-organizing neural network;

[0016] if the self-organizing neural network does not converge, setting t = t + 1 and returning to the step of randomly selecting one of the first time-frequency feature data as an input vector for the t-th round of training;

[0017] if the self-organizing neural network converges, obtaining a trained target self-organizing neural network and a plurality of first clustering clusters.

[0018] In a feasible implementation manner, the first time-frequency feature data of the sample traveling wave signals of different fault types is obtained by:

[0019] electromagnetic transient waveform signals of different fault types of an overhead line power system are obtained from a traveling wave monitoring device or a sensor, and the sample traveling wave signals include the electromagnetic transient waveform signals;

[0020] time-frequency transformation is performed on the collected electromagnetic transient waveform signals to obtain time-frequency feature vectors of the electromagnetic transient waveform signals;

[0021] the time-frequency feature vectors are normalized by using a preset normalization algorithm to obtain normalized time-frequency feature vectors, and the first time-frequency feature data includes the normalized time-frequency feature vectors.

[0022] In a feasible implementation manner, the time-frequency transformation is performed on the collected electromagnetic transient waveform signals to obtain time-frequency feature vectors of the electromagnetic transient waveform signals, and includes:

[0023] each of the electromagnetic transient waveform signals is divided into a plurality of time window signals according to a preset time window;

[0024] fast Fourier transformation is performed on each time window signal to obtain the time-frequency feature vectors of the electromagnetic transient waveform signals.

[0025] In a feasible implementation manner, the updating of the current weight vector of the winning node and the nodes in the neighborhood of the winning node comprises the following mathematical expression:

[0026] ;

[0027] wherein, is an updated weight vector of the node in the i-th row and the j-th column at the next training round t+1; is a current weight vector of the node in the i-th row and the j-th column at the current training round t; is a current learning rate at the current training round t; is a field function of the node in the i-th row and the j-th column; is an input vector.

[0028] To achieve the above object, the second aspect of the present application provides a use method of the traveling wave fault classification model, and the use method comprises:

[0029] obtaining second time-frequency feature data of a current traveling wave signal;

[0030] performing clustering on the second time-frequency feature data by using a target self-organizing neural network to determine a second clustering cluster of the second time-frequency feature data, wherein the target self-organizing neural network is obtained by training the traveling wave fault classification model by using the training method according to the first aspect and any feasible implementation manner;

[0031] performing fault type prediction by using the second clustering cluster and a target deep learning model to obtain a target fault type of the second clustering cluster, wherein the target deep learning model is obtained by training the traveling wave fault classification model by using the training method according to the first aspect and any feasible implementation manner, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model.

[0032] To achieve the above object, the third aspect of the present application provides a training device of a traveling wave fault classification model, and the training device comprises:

[0033] a sample collection module configured to obtain first time-frequency feature data of sample traveling wave signals of different fault types;

[0034] a first training module configured to perform clustering training on a preset self-organizing neural network by using the first time-frequency feature data to obtain a target self-organizing neural network trained and a plurality of first clustering clusters;

[0035] a label determination module configured to determine a real fault type label of each of the first clustering clusters;

[0036] The second training module is configured to perform fault type prediction training by using the first clustering cluster, the real fault type label, and a preset deep learning model, to obtain a trained target deep learning model, and the target deep learning model is configured to predict the real fault type label corresponding to the clustering cluster, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model.

[0037] To achieve the above object, the fourth aspect of the present application provides a use device of the traveling wave fault classification model, and the use device comprises:

[0038] The data acquisition module is configured to acquire second time-frequency feature data of a current traveling wave signal.

[0039] The clustering analysis module is configured to perform clustering on the second time-frequency feature data by using a target self-organizing neural network, to determine a second clustering cluster of the second time-frequency feature data, and the target self-organizing neural network is trained by using the training method of the traveling wave fault classification model according to the first aspect and any feasible implementation manner.

[0040] The fault prediction module is configured to perform fault type prediction by using the second clustering cluster and a target deep learning model, to obtain a target fault type of the second clustering cluster, and the target deep learning model is trained by using the training method of the traveling wave fault classification model according to the first aspect and any feasible implementation manner, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model.

[0041] To achieve the above object, the fifth aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the training method according to the first aspect and any feasible implementation manner or the steps of the use method according to the second aspect.

[0042] To achieve the above object, the sixth aspect of the present application provides a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the training method according to the first aspect and any feasible implementation manner or the steps of the use method according to the second aspect.

[0043] By using the embodiment of the present application, the following beneficial effects are achieved:

[0044] The application provides a training method of a traveling wave fault classification model, which comprises the following steps: obtaining first time-frequency feature data of sample traveling wave signals of different fault types; performing clustering training on a preset self-organizing neural network by using the first time-frequency feature data, obtaining a target self-organizing neural network after training and a plurality of first clustering clusters; determining real fault type labels of each first clustering cluster; performing fault type prediction training by using the first clustering cluster, the real fault type labels and a preset deep learning model, obtaining a target deep learning model after training, and the target deep learning model is used for predicting real fault type labels corresponding to the clustering cluster, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model.

[0045] In the above manner, the self-organizing neural network (SONN) is used for time-frequency feature clustering, dynamic features of different types of faults are extracted, and then the deep learning model is used for learning the dynamic features of different types of faults to predict the fault types, thereby improving the diagnosis capability for hidden faults. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0047] Among them:

[0048] Figure 1 It is a flow chart of the training method of the traveling wave fault classification model in the embodiment of the present application.

[0049] Figure 2 It is a flow chart of the use method of the traveling wave fault classification model in the embodiment of the present application.

[0050] Figure 3 It is a structure block diagram of the training device of the traveling wave fault classification model in the embodiment of the present application.

[0051] Figure 4 It is a structure block diagram of the use device of the traveling wave fault classification model in the embodiment of the present application.

[0052] Figure 5 It is a structure block diagram of the computer device in the embodiment of the present application. DETAILED DESCRIPTION

[0053] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0054] Please refer to Figure 1 , Figure 1 The flowchart of the training method of the traveling wave fault classification model in an embodiment of the present application, the training method of the traveling wave fault classification model shown in the present application is applied to a training system of a traveling wave fault classification model. The training system includes a terminal and a server. The terminal and the server are connected through a network, and the terminal can be a desktop terminal or a mobile terminal, and the mobile terminal can be at least one of a mobile phone, a tablet computer, a notebook computer, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers. The method can be applied to the terminal and the server, and the present embodiment is exemplified by the application to the terminal, as shown in the method Figure 1 The method includes the following steps:

[0055] 101. Obtain first time-frequency feature data of sample traveling wave signals of different fault types.

[0056] It should be noted that the present application proposes an innovative scheme, and provides a traveling wave concealed fault diagnosis method based on a self-organizing neural network and deep learning, wherein the self-organizing neural network can be a self-organizing mapping network (SOM, Self-Organizing Map). Based on the SOM and deep learning classification technology, combined with efficient feature extraction and analysis means, the diagnosis capability of concealed faults is improved. The present application is applicable to overhead line power systems, smart grids, railway transportation, track power supply systems, etc., especially to scenes requiring efficient diagnosis of concealed faults.

[0057] In order to let the self-organizing neural network and the deep learning model learn the related features of the concealed fault, so as to train a traveling wave fault classification model with the related features of the concealed fault, a large number of training samples are first needed for model learning. The learning samples can be time-frequency feature data of sample traveling wave signals of different fault types, and the sample traveling wave signals can be various fault type traveling wave signals collected historically.

[0058] Moreover, since the traveling wave signal is weak and complex, the collected traveling wave signal can be subjected to time-frequency transformation, and the time-frequency feature data of the traveling wave signal is taken as the training sample.

[0059] In an implementable manner, the traveling wave signal can be preprocessed to obtain time-frequency feature data, and the data preprocessing includes signal collection, time-frequency feature extraction and normalization processing, that is, step 101 can include steps A01 to A03:

[0060] A01, obtaining electromagnetic transient waveform signals of different fault types of the overhead line power system from a traveling wave monitoring device or a sensor, the sample traveling wave signal including the electromagnetic transient waveform signal;

[0061] First, signal collection is performed to obtain traveling wave signals of different fault types, specifically, electromagnetic transient waveform signals of different fault types of the overhead line power system are obtained from a traveling wave monitoring device or a sensor, and the above sample traveling wave signal includes the electromagnetic transient waveform signal, wherein a traveling wave monitoring device is installed at a key node (such as a substation, a branch point, etc.) of an overhead line power line, and the monitoring device collects high-frequency signals (usually in the range of 10 kHz to 20 kHz) in real time.

[0062] A02, performing time-frequency transformation on the collected electromagnetic transient waveform signal to obtain a time-frequency feature vector of the electromagnetic transient waveform signal;

[0063] Further, after obtaining the sample traveling wave signal, time-frequency transformation can be performed on the collected electromagnetic transient waveform signal to obtain a time-frequency feature vector of the electromagnetic transient waveform signal.

[0064] In an implementable manner, the time-frequency transformation can be fast Fourier transform, and step A02 includes: dividing each of the electromagnetic transient waveform signals into a plurality of time window signals according to a preset time window; performing fast Fourier transform on each time window signal to obtain a time-frequency feature vector of the electromagnetic transient waveform signal, and the time window can be a Hanning window.

[0065] An exemplary time-frequency feature extraction process is as follows: performing time-frequency transformation on the collected waveform data to convert the signal into a time-frequency feature vector to form a high-dimensional feature data set, specifically, the collected traveling wave signal is divided into a plurality of windows according to a fixed length, a Hanning window is selected as the window function to avoid spectral leakage, and fast Fourier transform (FFT) is performed on each window, and the expression is as follows:

[0066] ;

[0067] wherein, is a window function for segmenting the signal at time t; x(t) is an original time-domain signal, that is, the collected electromagnetic transient waveform signal; j is an imaginary unit; π is a circular constant; is the amplitude of the frequency at time t; and the result is a two-dimensional time-frequency matrix , the storage time t, the frequency f and the corresponding amplitude value to obtain a time-frequency feature vector.

[0068] A03, normalizing the time-frequency feature vector by using a preset normalization algorithm to obtain a normalized time-frequency feature vector, and the first time-frequency feature data comprises the normalized time-frequency feature vector.

[0069] After obtaining the time-frequency feature vector, the time-frequency feature vector needs to be normalized, specifically by using a preset normalization algorithm to normalize the time-frequency feature vector to obtain a normalized time-frequency feature vector, wherein the first time-frequency feature data comprises the normalized time-frequency feature vector.

[0070] An exemplary normalization process is as follows: normalize each feature to map the data range to [0, 1] or [-1, 1] to eliminate the influence of the dimensional difference of the feature value on the model, specifically: linearly map the feature value to the range of [0, 1] or [-1, 1], and the normalization algorithm is as follows:

[0071] ;

[0072] wherein, is the original feature value; is the maximum value and the minimum value of the feature; is the normalized feature value.

[0073] 102, using the first time-frequency feature data to cluster train a preset self-organizing neural network to obtain a trained target self-organizing neural network and a plurality of first clustering clusters;

[0074] Further, the first time-frequency feature data is used to cluster train a preset self-organizing neural network to obtain a trained target self-organizing neural network and a plurality of first clustering clusters.

[0075] Wherein, the preset self-organizing neural network comprises a plurality of nodes, each node is the center of a clustering cluster, in order to realize the clustering of time-frequency feature data of different fault types, the dimension of the weight vector of each node is the same as the feature dimension of the first time-frequency feature data, and step 102 comprises steps B01 to B08:

[0076] B01, initializing the weight vector of each node of the self-organizing neural network;

[0077] It should be noted that the self-organizing neural network is initialized, specifically including:

[0078] 1) Determine the network structure: Select a two-dimensional grid (such as a rectangle or hexagon) as the output layer of the SOM, and set the number of nodes according to the complexity of the task (such as a 10x10 or 20x20 node matrix), specifically: Select a hexagonal network as the two-dimensional network structure, and the number of nodes determines the resolution of the SOM, so set the number of nodes according to the input data, specifically: Set the number of nodes M according to the input data set N and the feature dimension d, the specific formula is , take the square root of M, and construct a hexagonal network, then arrange the nodes of the hexagonal network according to the two-dimensional coordinates (i, j), where i and j are row and column indices, and finally output a hexagonal network structure with fixed node distribution.

[0079] 2) Initialize the weight: The weight vector of each node is randomly initialized, and the dimension of the weight vector is the same as the feature dimension of the input data, specifically: The weight vector of each node The dimension of the weight vector is equal to the feature dimension of the input data : Then , where i, j are node indices, and d is the feature dimension; When randomly initializing the weight, in order to ensure that the weight initialization range is consistent with the feature distribution of the input data, , is the minimum and maximum value of the kth feature of the input data, and for each node (i, j), a random number is generated for each weight vector dimension k and assigned.

[0080] B02, randomly select one from a plurality of first time-frequency feature data as an input vector of the tth training, t is the training round number, t initial value is 1;

[0081] B03, calculate the similarity between the input vector and the current weight vector of each node;

[0082] B04, determine the node with the highest similarity to the input vector among the nodes as the winning node, and map the input vector to the winning node to obtain an initial first clustering cluster centered on each node;

[0083] Further, the input data is gradually input into the network, specifically, in each round of clustering training, randomly select one from a plurality of first time-frequency feature data as an input vector X of the tth training, t is the training round number, t initial value is 1, calculate the similarity between the input vector X and the current weight vector of each node, determine the node with the highest similarity to the input vector among the nodes as the winning node c, and map the input vector to the winning node c to obtain an initial first clustering cluster centered on each node.

[0084] For example, randomly select an input vector from the feature data set ;

[0085] Similarity of the computing nodes: compute the distance of the input vector to each node weight vector (usually using Euclidean distance), the mathematical expression is as follows: , wherein, is the weight vector of node ;

[0086] Determine the winning node: find the node most similar to the input vector , the most similar node satisfies: ; that is, the node with the smallest distance.

[0087] B05, update the current weight vector of the winning node and the nodes in the neighborhood of the winning node to obtain an updated weight vector;

[0088] Further, each execution of B04 is equivalent to changing the cluster corresponding to the node, and the network parameters need to be self-organized, specifically, the current weight vector of the winning node and the nodes in the neighborhood of the winning node needs to be updated to obtain an updated weight vector, so as to better express the dynamic characteristics of the cluster corresponding to the node.

[0089] Illustratively, the self-organizing adjustment includes the following contents:

[0090] Adjust the weight vector: update the weight vector of the winning node and the nodes in the neighborhood of the winning node ; wherein, is the current learning rate at the current training round t is the neighborhood function of the node in the ith row and jth column ( ,j) and the winning node , usually in the form of a Gaussian function: ; wherein, and are the position coordinates of node (i,j) and winning node c respectively; is the neighborhood width, which decreases over time; then gradually reduce the learning rate and the neighborhood width, the initial learning rate is set to 0.1-0.5, gradually reduced to close to 0; the initial neighborhood width is set to 1 / 3~1 / 2 of the network radius, which decreases over time.

[0091] That is, the update of the current weight vector of the winning node and the nodes in the neighborhood of the winning node includes the following mathematical expression:

[0092] ;

[0093] wherein, is the updated weight vector of the node in the i-th row and j-th column at the next training round t+1; is the current weight vector of the node in the i-th row and j-th column at the current training round t; is the current learning rate at the current training round t; is the field function of the node in the i-th row and j-th column; is the input vector.

[0094] B06, using the preset convergence condition and the current characteristic parameter of the self-organizing neural network, judging whether the self-organizing neural network converges or not;

[0095] B07, if the self-organizing neural network does not converge, setting t = t + 1, and returning to execute the step of randomly selecting one from the plurality of first time-frequency feature data as the input vector of the t-th training round;

[0096] B08, if the self-organizing neural network converges, obtaining the trained target self-organizing neural network and the plurality of first clustering clusters.

[0097] Further, further iteration training is needed, and the iteration training is specifically: repeating B02 to B07 until the preset training round or the convergence condition is reached.

[0098] The convergence condition includes but is not limited to: the training reaches the preset maximum round, the change amount of the weight vector is less than the set threshold, and the classification result of the input vector is stable, so the current characteristic parameter includes the current training round, the current weight vector of each node, and the classification result of the current input vector.

[0099] In a feasible implementation manner, the iteration training implementation method can refer to the following contents:

[0100] In order to train the SOM network, the process of "winning node selection of competition layer" and "weight update" needs to be repeated until the preset convergence condition is reached. The following is a detailed implementation description:

[0101] 1. Iteration process: constantly adjust the weight of the SOM network, so that it can adaptively capture the distribution characteristics of the input data, and through the competition and weight adjustment process, the network gradually converges, an input data sample (selected in order or randomly from the training data set) is input, the winning node is selected by calculating the Euclidean distance (steps B03 to B04), and the weights of the winning node and its neighborhood nodes are adjusted according to the learning rate and the neighborhood function (step B05),

[0102] Check if the convergence condition is reached:

[0103] Convergence condition 1: The training reaches the preset maximum number of rounds;

[0104] Convergence condition 2: The change in the weight vector is less than a set threshold;

[0105] Convergence condition 3: The classification results of the input data are stable (each sample is mapped to the same node, and the mapping no longer changes).

[0106] If convergence is not achieved, return to step B02.

[0107] 2. Training Termination Conditions

[0108] Condition 1: Reach the preset maximum number of rounds The number of training epochs is usually set based on the amount of data and the size of the network, for example... In each round of training, all training data is input into the network one by one;

[0109] Condition 2: Weight change is less than a threshold. After each training round, calculate the overall change in the weight vector. : ;

[0110] in, For the first Nodes in round training The weight vector, if (like If the weight changes converge, then it is considered that the weight changes have converged.

[0111] Condition 3: The classification results are stable. For two consecutive training rounds, check whether the mapping nodes of the input data are stable.

[0112] ;

[0113] Training stops if stability reaches 100% or exceeds a set threshold (e.g., 99%).

[0114] This process is used to cluster all time-frequency feature data, resulting in several first clusters and a trained target self-organizing neural network.

[0115] An example of the iterative training process for a self-organizing neural network:

[0116] 1) Initialization parameters: learning rate The initial value is usually 0.1~0.5, and decreases with each training round:

[0117] ;

[0118] in, Current training round Maximum number of training rounds, neighborhood radius : Initial value is generally equal to half of the grid size, decreasing with training rounds:

[0119] ;

[0120] The specific training steps are:

[0121] Select input samples: In each training round, select a sample vector from the input data set ; Select in order, which can ensure uniform data coverage; Random selection: suitable for large-scale data sets.

[0122] 2) Select the winning node: Calculate the Euclidean distance between the sample and all node weights , select the node with the smallest distance as the winning node :

[0123] ;

[0124] 3) Update the weight: update the weight according to the neighborhood function and the learning rate :

[0125] ;

[0126] Neighborhood function:

[0127] ;

[0128] where : the grid coordinates of the node , c: the grid coordinates of the winning node , d: the grid distance between nodes;

[0129] 4) Update parameters: update the learning rate and neighborhood radius ;

[0130] With each training, the learning rate and neighborhood radius will gradually decrease to ensure that the network gradually converges and avoid over-adjustment:

[0131] ;

[0132] where and are the initial learning rate and neighborhood radius, is the maximum training round.

[0133] 5) Check the convergence condition: calculate the weight change , check the stability of the classification result. ​

[0134] Convergence verification involves calculating the amount of change in network weights at each training iteration and determining if the network has converged based on this change. Specifically, the amount of change in the weight vectors of all nodes after each training round is compared. If the amount of change is less than a set threshold, it indicates that the network has converged.

[0135] Implementation steps:

[0136] Weight change calculation: For the weight vector after each training round and the weight vector after the previous training round , calculate the difference between them:

[0137] ;

[0138] Then, calculate the sum of the weight changes of all nodes:

[0139] ;

[0140] If is less than a set threshold , it is considered that the network has converged.

[0141] Mapping stability: Compare the mapping results (clustered clusters) after each training round to verify if the mapping points are fixed.

[0142] By dynamically adjusting the learning rate, neighborhood function, and weights, the SOM network can gradually learn the distribution characteristics of the input data. The convergence conditions and parameter adjustments during training are critical and directly affect the performance and efficiency of the model.

[0143] Finally, result visualization and clustering can also be performed. To better understand the features learned by the SOM, we can visualize the mapping results to help identify different types of fault patterns. This process involves two parts: visualization of mapping results and fault classification and identification.

[0144] Mapping result visualization can be: After training is completed, we need to visualize the mapping results of each input sample into a two-dimensional graph. Each node represents a neuron in the SOM network, and these nodes will form a grid structure. When an input sample is mapped to a node, we can distinguish the fault type corresponding to each node by color or shape. Through the visualization results, users can intuitively observe the clustering distribution of different fault types in the SOM network.

[0145] Implementation steps:

[0146] Input sample mapping to nodes: For each input sample, it is mapped to the output layer of the SOM network. The mapping process involves calculating the distance between the input vector and the weight vector of each node to find the best-matching node (i.e., the winning node).

[0147]

[0148] in, It is a node The weight vector, It is the input sample. It is the node to which it is mapped.

[0149] Visualizing the output layer: The network's output layer is laid out in a two-dimensional grid, and the color of each node represents the type or category of the input sample corresponding to that node.

[0150] Finally, combining the visualization of the mapping results and fault classification, the final process can be summarized as follows:

[0151] Training Phase: Traveling wave signals from overhead power lines are collected, time-frequency features are extracted, normalized, and input into the SOM network for training. Weights are continuously adjusted during training to form clusters for different fault types. The mapping results are visualized to observe the clustering of different fault types.

[0152] 103. Determine the true fault type label for each of the first clusters;

[0153] 104. Using the first cluster, the real fault type label, and the preset deep learning model, perform fault type prediction training to obtain the trained target deep learning model. The target deep learning model is used to predict the real fault type label corresponding to the cluster. The traveling wave fault classification model includes the target self-organizing neural network and the target deep learning model.

[0154] Furthermore, to better identify fault types, this application combines a deep learning model for fault type classification and fault type prediction. First, for each of the first clusters, a true fault type label is determined. This label can be manually labeled or automatically labeled and manually verified; no limitation is made here. Next, the first clusters, the true fault type labels, and a pre-defined deep learning model are used for fault type prediction training to obtain a trained target deep learning model. The first cluster is used as input to the deep learning model, and the deep learning model outputs the predicted fault type of the first cluster. This predicted fault type is compared with the pre-determined true fault type label of the first cluster. The loss value between the two is used to determine whether the deep learning model has converged. If it has not converged, it is retrained; otherwise, it has converged, and the target deep learning model is obtained.

[0155] The application provides a training method of a traveling wave fault classification model, which comprises the following steps: obtaining first time-frequency feature data of sample traveling wave signals of different fault types; performing clustering training on a preset self-organizing neural network by using the first time-frequency feature data, to obtain a target self-organizing neural network and a plurality of first clustering clusters after training; determining real fault type labels of each first clustering cluster; performing fault type prediction training by using the first clustering clusters, the real fault type labels and a preset deep learning model, to obtain a target deep learning model after training, wherein the target deep learning model is used to predict real fault type labels corresponding to the clustering clusters, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model. In the foregoing manner, the SOM is used to cluster time-frequency features, and dynamic features of different types of faults are extracted, and then the dynamic features of different types of faults are learned by the deep learning model to predict the fault types, thereby improving the diagnosis capability for hidden faults.

[0156] Please refer to Figure 2 , Figure 2 A flowchart of a use method of a traveling wave fault classification model in an embodiment of the application, the use method of the traveling wave fault classification model shown in the application is applied to a use system of a traveling wave fault classification model. The use system comprises a terminal and a server. The terminal and the server are connected through a network, and the terminal can be a desktop terminal or a mobile terminal, and the mobile terminal can be at least one of a mobile phone, a tablet computer, a notebook computer and the like. The server can be implemented by an independent server or a server cluster composed of multiple servers. The method can be applied to the terminal and the server, and the embodiment is exemplarily explained by application to the terminal, as shown in the use method shown in Figure 2 The use method comprises the following steps:

[0157] 201. Obtain second time-frequency feature data of a current traveling wave signal.

[0158] 202. Cluster the second time-frequency feature data by using a target self-organizing neural network, to determine a second clustering cluster of the second time-frequency feature data, wherein the target self-organizing neural network is obtained by training the training method of the traveling wave fault classification model. Figure 1

[0159] 203. Perform fault type prediction by using the second clustering cluster and a target deep learning model, to obtain a target fault type of the second clustering cluster, wherein the target deep learning model is obtained by training the training method of the traveling wave fault classification model, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model. Figure 1

[0160] ​​It can be understood that the trained model can classify the current traveling wave signal collected in real time. Specifically, the second time-frequency feature data of the current traveling wave signal is obtained, and the acquisition of the second time-frequency feature data can refer to step 101, which will not be repeated here. Further, the target self-organizing neural network is used to cluster the second time-frequency feature data, and the second clustering cluster of the second time-frequency feature data is determined. The target fault type of the second clustering cluster is obtained by using the second clustering cluster and the target deep learning model for fault type prediction. The target fault type is used to reflect which fault the current traveling wave signal is, such as which hidden fault.

[0161] For example, the identification stage: the same preprocessing is performed on the newly input traveling wave data, which is mapped to the SOM network. The input sample is classified into a corresponding fault cluster, and the fault type is identified. Through the above method, the SOM network can automatically classify and identify hidden faults in the overhead line power system, thereby providing important support for fault diagnosis of the overhead line power system.

[0162] Further, the online learning and dynamic adjustment of SOM are as follows:

[0163] In order to adapt to the changes of the overhead line power system operation condition, the SOM needs to have the ability of online learning and dynamic adjustment;

[0164] Real-time data input: the system continuously monitors the running state of the overhead line power system, and inputs the newly collected traveling wave data into the SOM network in real time.

[0165] Dynamic weight adjustment: based on the new data, the weight vectors of the winning node and the nodes in its neighborhood are adjusted to maintain the adaptability of the model.

[0166] Fault mode update: when a new fault type appears, the clustering structure of the network is automatically adjusted through online learning, and the classification ability of the SOM is expanded.

[0167] Specifically: in order to realize the online learning and dynamic adjustment ability of the SOM (self-organizing mapping) network, especially in the application of processing real-time data in the overhead line power system, some strategies and methods are needed to support this dynamic nature. The following are the detailed steps to achieve this goal:

[0168] 1. Real-time data input: configure the data acquisition system to continuously monitor the traveling wave signal in the overhead line power system, and input the collected real-time signal (electromagnetic transient waveform) to the SOM network through the data acquisition device or sensor, ensuring the timeliness of data transmission. The delay from data collection to input to the SOM network should be as small as possible to maintain the real-time performance of the system.

[0169] 2. Dynamic weight adjustment:

[0170] Step 2.1: After obtaining real-time data, input it into the SOM network and calculate the most similar node ("winning node").

[0171] Step 2.2: Based on the learning rules of SOM, adjust the weights of the winning node and its neighborhood to enable the network to adapt to new data.

[0172] Weight update formula:

[0173] where: w is the weight vector of the node. is the input sample vector. is the learning rate, controlling the update step size. is the neighborhood function, representing the similarity between the node and the winning node , usually a Gaussian function.

[0174] Neighborhood function: The neighborhood function determines the influence of neighboring nodes during the update process. Over time, the neighborhood range gradually decreases, prompting the network to approach global optimization.

[0175] Dynamic adjustment of learning rate: Gradually reduce the learning rate as training progresses to avoid excessive updates to weights when approaching convergence.

[0176] 3. The purpose of fault mode update is to dynamically update the clustering structure of the SOM network through online learning when new fault modes appear in the power system of overhead lines, so as to cope with these new modes.

[0177] Add new nodes: When new fault types appear and existing nodes cannot handle them, new nodes can be added to the SOM network or the network dimension can be expanded. Decide whether to add nodes through algorithms, or optimize the network by merging similar fault types.

[0178] Re-train part of the network: For new fault modes, the network can selectively retrain part of the nodes. This not only maintains the stability of the existing network structure, but also has stronger adaptability under new fault modes.

[0179] 4. Implementation steps

[0180] Step 4.1: Real-time data acquisition and input configuration Sensors and monitoring devices are configured to ensure real-time data acquisition and transmission.

[0181] Connect the data acquisition system to the SOM network to ensure seamless data input into the network.

[0182] Step 4.2: Real-time training and dynamic adjustment

[0183] For each new input data, the winning node is first calculated.

[0184] Then, the weights are dynamically adjusted according to the weight update formula of the winning node.

[0185] With each new data input, the network adjusts the weights to adapt to the operating conditions of the power system.

[0186] The beneficial effects of the present application are:

[0187] 1. The present application utilizes the online learning ability of self-organizing mapping (SOM) network, enabling the power system to adapt to new operating conditions and fault modes in real time. Through dynamic adjustment of network weights, the SOM network can automatically update the fault classification model according to the changing power system data, without manual intervention or retraining. The combination of real-time data input and dynamic weight adjustment ensures efficient response and adaptability of the system to new data.

[0188] 2. The present application automatically identifies and classifies various fault modes in the power system through the SOM network, and improves the accuracy and stability of the system through continuous learning. When new fault types appear, the SOM network can automatically expand its clustering structure and update the fault mode in a timely manner, ensuring that the system can accurately identify various fault types at all times.

[0189] 3. By combining real-time data input with the fault classification ability of the SOM network, the present application can quickly judge the hidden faults in overhead line power systems, significantly improving the response speed and accuracy of the system. Through dynamic adjustment and visualization of clustering results, the system can provide early warning at the initial stage of fault occurrence, helping maintenance personnel to quickly prepare for maintenance measures and reduce downtime.

[0190] Please refer to Figure 3 , Figure 3 for the structure block diagram of the training device of a kind of traveling wave fault classification model in the embodiment of the present application, as Figure 3 shown in the training device of the traveling wave fault classification model, the training device includes:

[0191] Sample collection module 301: for obtaining the first time-frequency feature data of sample traveling wave signal of different fault types;

[0192] First training module 302: for clustering training of the first time-frequency feature data to the preset self-organizing neural network, obtaining the target self-organizing neural network of training completion and a plurality of first clustering clusters;

[0193] Label determination module 303: for determining the real fault type label of each first clustering cluster;

[0194] The second training module 304 is configured to perform fault type prediction training by using the first clustering cluster, the real fault type label, and a preset deep learning model, to obtain a trained target deep learning model, and the target deep learning model is used to predict the real fault type label corresponding to the clustering cluster, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model.

[0195] It should be noted that, Figure 3 The functions of each module of the device are similar to those of the device Figure 1 The contents of each step in the method are similar to those of the method Figure 1 The contents of each step in the method are similar to those of the method

[0196] The present application provides a kind of training method device of traveling wave fault classification model, training device includes: sample collection module: for obtaining the first time-frequency feature data of sample traveling wave signal of different fault types;First training module: for using first time-frequency feature data to preset self-organizing neural network is clustered training, obtains the target self-organizing neural network and several first clustering clusters of training completion;Label determination module: for determining the real fault type label of each first clustering cluster;Second training module: for using first clustering cluster, real fault type label and preset deep learning model are trained to predict fault type, to obtain the target deep learning model of training completion, target deep learning model is used to predict the real fault type label corresponding to clustering cluster, traveling wave fault classification model includes target self-organizing neural network and target deep learning model.By the above device, using self-organizing neural network (SOM) carries out time-frequency feature clustering, extracts the dynamic characteristics of different types of faults, and then learns the dynamic characteristics of different types of faults by deep learning model to predict fault type, improve the diagnosis ability of hidden fault.

[0197] Please refer to Figure 4 , Figure 4 The structure block diagram of a kind of using device of traveling wave fault classification model in the embodiment of the present application, as shown in Figure 4 The device comprises:

[0198] Data acquisition module 401: for obtaining the second time-frequency feature data of current traveling wave signal;

[0199] Clustering analysis module 402: for using target self-organizing neural network to cluster the second time-frequency feature data, to determine the second clustering cluster of the second time-frequency feature data, and the target self-organizing neural network is trained by the training method of the traveling wave fault classification model; Figure 1 The traveling wave fault classification model;

[0200] The fault prediction module 403 is configured to perform fault type prediction by using the second clustering cluster and a target deep learning model, to obtain a target fault type of the second clustering cluster, and the target deep learning model is obtained by training the traveling wave fault classification model. Figure 1 The traveling wave fault classification model includes the target self-organizing neural network and the target deep learning model.

[0201] It should be noted that, Figure 4 The functions of each module in the device are similar to those of the method. Figure 2 The contents of each step in the method are similar to those of the device, and to avoid repetition, the specific contents can be referred to in the method. Figure 2 The contents of each step in the method are similar to those of the device.

[0202] The application provides a traveling wave fault classification model using device. Through the device, time-frequency feature clustering is performed by using a self-organizing neural network (SOM), dynamic features of different types of faults are extracted, and a deep learning model is used to learn the dynamic features of different types of faults to predict the fault type, thereby improving the diagnosis capability of hidden faults.

[0203] Figure 5 An internal structure diagram of a computer device in an embodiment is shown. The computer device can be a terminal or a server. As shown in the figure, Figure 5 The computer device includes a processor, a memory and a network interface connected through a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and can also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. The internal memory can also store a computer program. When the computer program is executed by the processor, the processor can execute the above method. Those skilled in the art can understand that Figure 5 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0204] In one embodiment, a computer device is provided, including a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the method as shown in Figure 1 or Figure 2 The method shown in the figure.

[0205] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the method as shown inFigure 1 or Figure 2 the steps of the method shown.

[0206] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0207] The technical features of the above embodiments can be combined in any way. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0208] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of protection of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.

Claims

1. A training method for a traveling wave fault classification model, characterized in that, The method comprises: obtaining first time-frequency feature data of sample traveling wave signals of different fault types; performing clustering training on a preset self-organizing neural network by using the first time-frequency feature data to obtain a trained target self-organizing neural network and a plurality of first clustering clusters, the clustering clusters being used to reflect clustering of time-frequency feature data of different fault types, the fault types including hidden faults; determining real fault type labels of each of the first clustering clusters; performing fault type prediction training by using the first clustering clusters, the real fault type labels, and a preset deep learning model to obtain a trained target deep learning model, the target deep learning model being used to predict real fault type labels corresponding to the clustering clusters, the traveling wave fault classification model comprising the target self-organizing neural network and the target deep learning model; wherein the self-organizing neural network comprises a plurality of nodes, a dimension of a weight vector of each node being the same as a feature dimension of the first time-frequency feature data, and the clustering training on the preset self-organizing neural network by using the first time-frequency feature data to obtain the trained target self-organizing neural network and the plurality of first clustering clusters comprises: initializing weight vectors of each node of the self-organizing neural network; randomly selecting one of the plurality of first time-frequency feature data as an input vector for t-th training, t being a training round number, t initial value being 1; calculating similarity between the input vector and current weight vectors of each node; determining a node with the highest similarity to the input vector among the nodes as a winning node, and mapping the input vector to the winning node to obtain an initial first clustering cluster centered on each node; updating the current weight vectors of the winning node and nodes in a neighborhood of the winning node to obtain updated weight vectors; judging whether the self-organizing neural network converges by using a preset convergence condition and current characteristic parameters of the self-organizing neural network; if the self-organizing neural network does not converge, setting t = t + 1 and returning to the step of randomly selecting one of the plurality of first time-frequency feature data as an input vector for t-th training; if the self-organizing neural network converges, obtaining the trained target self-organizing neural network and the plurality of first clustering clusters.

2. The method of claim 1, wherein, The obtaining of the first time-frequency feature data of sample traveling wave signals of different fault types comprises: obtaining electromagnetic transient waveform signals of different fault types of an overhead line power system from a traveling wave monitoring device or a sensor, the sample traveling wave signals comprising the electromagnetic transient waveform signals; performing time-frequency transformation on the collected electromagnetic transient waveform signals to obtain time-frequency feature vectors of the electromagnetic transient waveform signals; performing normalization processing on the time-frequency feature vectors by using a preset normalization algorithm to obtain normalized time-frequency feature vectors, the first time-frequency feature data comprising the normalized time-frequency feature vectors.

3. The method of claim 2, wherein, The time-frequency transformation on the collected electromagnetic transient waveform signals to obtain time-frequency feature vectors of the electromagnetic transient waveform signals comprises: Divide each of the electromagnetic transient waveform signals into a plurality of time window signals according to a preset time window; Perform fast Fourier transform on each time window signal to obtain a time-frequency feature vector of the electromagnetic transient waveform signal.

4. The method of claim 1, wherein, The updating of the current weight vector of the winning node and the nodes in the neighborhood of the winning node includes the following mathematical expression: ； wherein is the updated weight vector for the node in row i and column j at the next training epoch t+1; is the current weight vector for the node in row i and column j at the current training epoch t; is the current learning rate at the current training epoch t; is the neighborhood function for the node in row i and column j; is the input vector.

5. A method of using a traveling wave fault classification model, the method comprising: The use method comprises: Obtaining second time-frequency feature data of a current traveling wave signal; Clustering the second time-frequency feature data by using a target self-organizing neural network to determine a second clustering cluster of the second time-frequency feature data, wherein the target self-organizing neural network is obtained by training the traveling wave fault classification model according to the training method in any one of claims 1 to 4; Performing fault type prediction by using the second clustering cluster and a target deep learning model to obtain a target fault type of the second clustering cluster, wherein the target deep learning model is obtained by training the traveling wave fault classification model according to the training method in any one of claims 1 to 4, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model. 6.A device for training a traveling wave fault classification model, characterized in that, The training device comprises: a sample collection module configured to obtain first time-frequency feature data of sample traveling wave signals of different fault types; a first training module configured to perform clustering training on a preset self-organizing neural network by using the first time-frequency feature data to obtain a target self-organizing neural network after training and a plurality of first clustering clusters, wherein the clustering cluster is used to reflect the clustering of time-frequency feature data of different fault types, and the fault types include hidden faults; a label determination module configured to determine a real fault type label of each of the first clustering clusters; a second training module configured to perform fault type prediction training by using the first clustering clusters, the real fault type labels, and a preset deep learning model to obtain a target deep learning model after training, wherein the target deep learning model is used to predict a real fault type label corresponding to a clustering cluster, and the traveling wave fault classification model comprises the target self-organizing neural network and the target deep learning model. The self-organizing neural network includes a plurality of nodes, and a dimension of a weight vector of each node is the same as a feature dimension of the first time-frequency feature data.

7. A use of a traveling wave fault classification model, the use comprising: The using device includes: A data acquisition module is configured to acquire second time-frequency feature data of a current traveling wave signal. A clustering analysis module is configured to perform clustering on the second time-frequency feature data by using a target self-organizing neural network to determine a second clustering cluster of the second time-frequency feature data, the target self-organizing neural network being trained by using the training method of the traveling wave fault classification model according to any one of claims 1 to 4. A fault prediction module is configured to perform fault type prediction by using the second clustering cluster and a target deep learning model to obtain a target fault type of the second clustering cluster, the target deep learning model being trained by using the training method of the traveling wave fault classification model according to any one of claims 1 to 4, the traveling wave fault classification model including the target self-organizing neural network and the target deep learning model.

8. A computer readable storage medium storing a computer program, characterized in that, The computer program, when executed by the processor, causes the processor to perform the steps of the training method according to any one of claims 1 to 4 or the steps of the using method according to claim 5. 9.A computer device, comprising a memory and a processor, and characterized in that, The memory stores a computer program, and the computer program, when executed by the processor, causes the processor to perform the steps of the training method according to any one of claims 1 to 4 or the steps of the using method according to claim 5.

Images