Tree line discharge fault identification method based on big data fault traveling wave current characteristics

Abstract

The invention discloses a tree line discharge fault identification method based on big data fault traveling wave current characteristics, and the method comprises the steps: S1, collecting fault traveling wave current waveform data and environment data, and enabling the environment data to comprise wind speed, humidity and temperature data; s2, performing adaptive discrete wavelet transform on the fault traveling wave current waveform to obtain a fault wavelet coefficient, obtaining an enhanced fault wavelet coefficient through an improved dynamic attention network based on the fault wavelet coefficient, and obtaining the optimal frequency band energy of the enhanced fault wavelet coefficient; splicing the enhanced fault wavelet coefficient, the optimal frequency band energy and the environment data into a comprehensive feature vector; s3, introducing a self-organizing mapping network based on the comprehensive feature vector to construct an optimized state space, collecting traveling wave current waveform data in real time, and obtaining waveform time sequence similarity between the traveling wave current waveform data and the optimized state space by using an improved dynamic time warping algorithm; when real-time data is input, subspace matching of corresponding environmental conditions can be quickly positioned.

Description

Technical Field

[0001] The present invention relates to the technical field of current fault identification, and in particular to a tree-line discharge fault identification method based on big data fault traveling wave current characteristics. Background Art

[0002] With the expansion and intelligent development of power grids, the safe operation of transmission lines is of paramount importance. Tree-line discharge faults pose a serious threat to the safe operation of transmission lines. The traveling current waveform is affected by multiple factors such as tree sway, air humidity, and temperature, showing complex characteristics such as strong nonlinearity and time-varying characteristics. However, traditional fault identification methods are difficult to adapt to this complex characteristic.

[0003] At present, traditional fault identification methods only focus on analyzing the traveling wave current waveform itself, fail to fully recognize the close relationship between various environmental factors and tree-line discharge faults, and find it difficult to fully capture the complex similarities of traveling wave current waveform characteristics in time series, resulting in inaccurate fault judgment. Therefore, a tree-line discharge fault identification method based on big data fault traveling wave current characteristics is proposed. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention proposes the following technical solutions:

[0005] The tree-line discharge fault identification method based on big data fault traveling wave current characteristics includes:

[0006] S1: Collecting fault traveling wave current waveform data and environmental data, the environmental data including wind speed, humidity, and temperature data;

[0007] S2: Perform adaptive discrete wavelet transform on the fault traveling current waveform to obtain the fault wavelet coefficient. Based on the fault wavelet coefficient, the enhanced fault wavelet coefficient is obtained by improving the dynamic attention network. The optimal frequency band energy of the enhanced fault wavelet coefficient is obtained. The enhanced fault wavelet coefficient, the optimal frequency band energy and the environmental data are spliced into a comprehensive feature vector.

[0008] Among them, the improved dynamic attention network is achieved by adding an adaptive mechanism to the dynamic changes of time series to the traditional attention mechanism to enhance the core feature representation of the fault wavelet coefficients;

[0009] S3: Based on the comprehensive eigenvector, a self-organizing map network is introduced to construct an optimized state space. The traveling wave current waveform data collected in real time is used to obtain the waveform timing similarity between the traveling wave current waveform data and the optimized state space using an improved dynamic time warping algorithm;

[0010] Among them, the improved dynamic time warping algorithm is extended to five implementations based on the traditional dynamic programming algorithm which usually only considers three adjacent directions when finding the optimal path;

[0011] S4: Setting a fault similarity threshold, when the waveform time sequence similarity of the real-time collected traveling wave current waveform data is greater than or equal to the fault similarity threshold, marking the corresponding traveling wave current waveform data as abnormal.

[0012] The process of obtaining the fault wavelet coefficients is as follows:

[0013] On the basis of traditional discrete wavelet transform, an adaptive wavelet basis selection mechanism is introduced to add scale parameters and translation parameters to perform adaptive wavelet basis selection to obtain fault wavelet coefficients.

[0014] The process of obtaining the enhanced fault wavelet coefficients is as follows:

[0015] Obtain feature representations of the fault wavelet coefficients, perform linear transformation on the feature representations, and obtain attention scores between different feature representations;

[0016] Based on the attention scores between different feature representations, an optimized time series dynamic weight matrix is introduced to adjust the attention scores between different feature representations according to the time series change trend between different feature representations;

[0017] The element values in the obtained optimized time series dynamic weight matrix are used to adjust the corresponding attention scores according to the changing trend of the signal in the time series;

[0018] The attention scores corresponding to the characteristic representation of the fault wavelet coefficients are optimized by optimizing the time series dynamic weight matrix;

[0019] The optimized attention weight matrix is used to directly perform weighted summation on the feature representation of the fault wavelet coefficients to obtain the enhanced feature representation of the fault wavelet coefficients.

[0020] The process of obtaining the optimized time series dynamic weight matrix is as follows:

[0021] Obtain the characteristic representation of the fault wavelet coefficients to obtain the rate of change sequence in different time windows;

[0022] The rate of change sequence is used as the input of a convolutional neural network, where the convolutional network consists of L convolutional layers. The parameters of each convolutional layer include the convolution kernel and bias. After processing by L convolutional layers, the feature map is output. Then, a fully connected layer is used to map the feature map into an optimized time series dynamic weight matrix with the same dimension as the attention score matrix.

[0023] The optimization state space construction process is as follows:

[0024] Introducing the self-organizing map network, which includes an input layer and a competition layer;

[0025] Input the comprehensive feature vector into the input layer of the self-organizing map network, obtain the distance between the input data and the weight vector of each neuron in the competition layer, and obtain the winning neuron based on the neuron with the smallest distance;

[0026] The neurons in the competition layer are divided into P subspaces. For any specific combination of environmental condition intervals, the neurons that respond more to the data under the environmental conditions during training are divided into a subspace. Each subspace stores the comprehensive feature vector under the corresponding environmental conditions to obtain the optimized state space.

[0027] The waveform timing similarity is obtained by improving the dynamic programming algorithm according to the cumulative distance corresponding to the optimal path.

[0028] The process of calculating the waveform timing similarity based on the cumulative distance corresponding to the optimal path is:

[0029] Define a dynamic programming table, where the elements of the dynamic programming table represent the minimum cumulative distance from the starting point to the coordinate position;

[0030] Set a boundary condition and obtain the boundary values of the edge and column, where the cumulative distance of the starting point is the weighted distance of the starting point;

[0031] After determining the boundary values of the edges and columns, a dynamic programming table is filled in. Based on the filled dynamic programming table, the real-time collected traveling wave current waveform data and the dimension of the comprehensive feature vector in the optimized state space after dimensionality reduction are obtained. The reduced dimension is used as the judgment length in the dynamic programming table, and the waveform timing similarity is obtained based on the judgment length.

[0032] The minimum cumulative distance is obtained by adding the minimum value of the weighted distance and the minimum cumulative distances of the five adjacent direction positions;

[0033] The weighted distance is obtained by obtaining the distance between the elements in the traveling wave current waveform data and the elements of the comprehensive characteristic vector in the optimized state space through the Euclidean distance.

[0034] The present invention has the following beneficial effects:

[0035] In this invention, firstly, the optimal wavelet basis is selected by dynamically adjusting the scale and translation parameters, which solves the problem that the fixed wavelet basis is insufficiently adaptable to complex waveforms and effectively extracts fault features in different frequency bands. At the same time, a time series dynamic weight matrix is generated through a convolutional neural network, and the attention score is dynamically adjusted according to the signal change rate to highlight key features such as the rising edge and peak of the waveform.

[0036] Secondly, the enhanced fault wavelet coefficients, optimal frequency band energy, and environmental data (wind speed, humidity, and temperature) are spliced into a comprehensive feature vector to construct a multidimensional input space that includes electrical characteristics and environmental conditions. This space comprehensively reflects the inducing conditions and waveform characteristics of tree-line discharge faults. At the same time, a self-organizing map network is used to cluster the comprehensive feature vector and divide it into multiple subspaces according to environmental condition intervals. This enables refined classification and management of historical data. When real-time data is input, it can quickly locate the subspace matching the corresponding environmental conditions.

[0037] Finally, by expanding the path search direction to five directions (up, left, upper left, upper right, and down), it better fits the complex deformation pattern of the traveling wave current waveform. Compared with the traditional three-directional search, the accuracy of waveform timing matching is higher. Combined with the optimized state space statistics, the similarity distribution between real-time data and fault data can be accurately achieved for fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a method step diagram for the tree-line discharge fault identification method based on big data fault traveling wave current characteristics proposed by the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] Example 1

[0041] like Figure 1 As shown, the tree-line discharge fault identification method proposed in the present invention based on big data fault traveling wave current characteristics includes:

[0042] S1: Collect fault traveling wave current waveform data and environmental data, including wind speed, humidity, and temperature data;

[0043] A Rogowski coil current sensor is installed on a transmission line tower near the tree-line discharge fault area. The sensor samples the traveling wave current at a sampling frequency of 100 kHz, which ensures that the waveform details of the traveling wave current are fully recorded. The traveling wave current data obtained at each sampling is then represented as a discrete time series I.

[0044] In fault detection, environmental factors also have a significant impact on the occurrence of treeline discharge faults. Therefore, it is necessary to collect environmental data such as wind speed, humidity, and temperature. The collection frequency of environmental data is consistent with the collection frequency of traveling wave current waveform data, both of which are 100kHz. Let the environmental data be H = [h1, h2, h3], where h1 represents wind speed data, h2 represents humidity data, and h3 represents temperature data.

[0045] Use a three-cup wind speed sensor to measure wind speed. The rotation of the wind cup drives the shaft to rotate, and then the wind speed is calculated based on the shaft rotation speed to obtain the wind speed data h1;

[0046] Specifically, high wind speeds can cause trees to sway, increasing the frequency of contact with power lines, affecting the stability of wind speed arcs, and changing discharge characteristics;

[0047] A capacitive humidity sensor is used to measure the ambient humidity. The capacitance of the humidity-sensitive capacitor changes with the change of ambient humidity, and the relative humidity data h is obtained.

[0048] Specifically, humidity changes will change the conductivity and dielectric strength of the air. High humidity will reduce the breakdown voltage of the air, making discharge more likely to occur, while increasing leakage current and affecting the amplitude and waveform duration of the traveling wave current.

[0049] Use thermocouple temperature sensor to measure temperature, use the thermoelectric effect of two different metal conductors to sense temperature changes and obtain temperature data T;

[0050] Specifically, temperature changes will affect the thermal expansion and contraction of the wires, changing the contact pressure between the wires and the trees. High temperatures will cause the wires to expand and sag, increasing the probability of contact. Temperature will also affect the density and degree of ionization of the air. In a high temperature environment, the air is more easily ionized, making the air gap between the wires and the trees more easily broken down, thereby causing discharge.

[0051] S2: Perform adaptive discrete wavelet transform on the fault traveling current waveform to obtain the fault wavelet coefficient. Based on the fault wavelet coefficient, the enhanced fault wavelet coefficient is obtained by improving the dynamic attention network. The optimal frequency band energy of the enhanced fault wavelet coefficient is obtained. The enhanced fault wavelet coefficient, the optimal frequency band energy and the environmental data are spliced into a comprehensive feature vector.

[0052] Discrete wavelet transform is a powerful mathematical tool that can decompose a signal into different frequency components, thereby effectively extracting the characteristics of the signal;

[0053] The traditional discrete wavelet transform selects a fixed wavelet basis function, such as the common db4 wavelet basis. Regardless of the characteristics of the input signal, this fixed wavelet basis is used to decompose the signal. However, different signals have different characteristics in terms of frequency components, waveform changes, etc., and the fixed wavelet basis cannot adapt to all signal conditions.

[0054] An adaptive discrete wavelet transform is performed on the collected fault traveling current waveform I. Based on the traditional discrete wavelet transform, an adaptive wavelet basis selection mechanism is introduced (adding scale parameters and translation parameters). Let ψa,b(t) be the wavelet basis function, where a is the scale parameter and b is the translation parameter. The expression of the adaptive discrete wavelet transform is:

[0055]

[0056] Among them, F is the converted fault wavelet coefficient, t is the time index, ψ(t) is the mother wavelet function, It is the conjugate function of the wavelet basis function ψa,b(t), and the adaptive selection of the wavelet basis is achieved by dynamically adjusting the scale parameter and the translation parameter;

[0057] According to the characteristics of the fault traveling current waveform itself, the most appropriate wavelet basis function is selected by dynamic scale parameters and translation parameters;

[0058] Specifically, the scale parameter a and the translation parameter b are both discrete values that are directly inserted into the mother wavelet function ψ(t). The scale parameter a determines the degree of expansion and contraction of the mother wavelet function. A larger a value is used for the traveling wave current waveform of a low-frequency fault, and a smaller a value is used for the traveling wave current waveform of a high-frequency fault. The translation parameter b is used to move the mother wavelet function on the time axis so that the signal can be analyzed at different positions.

[0059] The fault wavelet coefficients are input into an improved dynamic attention network;

[0060] Traditional attention mechanisms have limitations when directly processing complex time-varying signals such as traveling current waveforms. To better capture the core features of the signal, an adaptive mechanism for dynamically changing time series is added to the fault wavelet coefficients obtained through adaptive discrete wavelet transform. This mechanism focuses on the key features of the fault wavelet coefficients to enhance the core feature representation of the fault wavelet coefficients.

[0061] Get the characteristic representation X of the fault wavelet coefficient F , the feature representation X is represented by the learning matrix mechanism in the basic attention network F Perform linear transformation and obtain attention scores between different feature representations (attention scores represent the closeness of the association);

[0062] Introduce the optimized time series dynamic weight matrix D, and adjust the attention scores between different feature representations according to the time series change trend between different feature representations.

[0063] Specifically, for the position corresponding to the time window with a larger rate of change, a relatively larger element value is assigned in the matrix D, and vice versa;

[0064] The element value d in the optimized time series dynamic weight matrix D is obtained i ∈D, adjust the corresponding attention score S according to the changing trend of the signal in the time series i :S * =S i *d i ;

[0065] The attention scores corresponding to the feature representation of the fault wavelet coefficients are optimized by optimizing the time series dynamic weight matrix, so that the improved dynamic attention network can adaptively highlight the key features in the signal and improve the ability to extract the tree line discharge fault features. The optimized attention weight matrix is used to directly represent the feature representation X of the fault wavelet coefficients. F Perform weighted summation to obtain the enhanced characteristic representation of the fault wavelet coefficients

[0066] Specifically, the process of optimizing the time series dynamic weight matrix D is as follows:

[0067] Get the characteristic representation X of the fault wavelet coefficient F The sequence of rates of change in different time windows:

[0068] Assume the length of the time window is w, for a series of time windows y i ,y i+1 ,..,y i+w-1 , calculate the slope S between the two feature representations:

[0069]

[0070] Among them, y j is an arbitrary time series in the time window, Δt is the time interval;

[0071] Calculate the average value of the absolute value of the slope S within the time window as the rate of change r, which is expressed as follows:

[0072]

[0073] Through the above calculation, we can obtain the rate of change sequence R=[r1, r2, .., r i+w-1 ];

[0074] The rate of change sequence R containing the rate of change information is used as the input of a convolutional neural network. The convolutional network consists of L convolutional layers. The parameters of each convolutional layer include the convolution kernel and the bias. After processing by L convolutional layers, the output feature map h L , and then pass the feature map h through a fully connected layer L Mapped to an optimized time series dynamic weight matrix D of the same dimension as the attention score matrix;

[0075] From the enhanced fault wavelet coefficients characteristic representation Obtain the frequency band energy vector, and set the enhanced feature representation of the fault wavelet coefficient There are m optimal frequency bands closely related to the fault. For the wavelet coefficients in each frequency band, the frequency band energy is calculated according to the formula:

[0076]

[0077] Where E is the frequency band energy;

[0078] Calculate the energy-enhanced feature representation in m frequency bands The frequency band energy is obtained to obtain the optimal frequency band energy vector

[0079] The enhanced fault wavelet coefficients are represented by Optimal frequency band energy vector After normalization with the environmental coefficient, a comprehensive feature vector is formed, which is expressed as

[0080] Specifically, from the adaptive discrete wavelet transform that selects the wavelet basis according to the waveform characteristics, to the improved dynamic attention network that adjusts the weight according to the changing trend of the signal time series, it can perform adaptive processing according to the different characteristics and changes of the fault traveling wave current waveform, and realize the accurate extraction and analysis of fault characteristics. At the same time, it not only considers the wavelet coefficients and frequency band energy characteristics of the fault traveling wave current waveform, but also incorporates environmental data to achieve deep fusion of multi-source information. The constructed comprehensive feature vector more comprehensively reflects the overall state when the fault occurs.

[0081] S3: Based on the comprehensive eigenvector, a self-organizing map network is introduced to construct an optimized state space. The traveling wave current waveform data collected in real time is used to obtain the waveform timing similarity between the traveling wave current waveform data and the optimized state space using an improved dynamic time warping algorithm;

[0082] Based on the comprehensive feature vector, a self-organizing map network is introduced, which includes an input layer and a competition layer.

[0083] The number of neurons in the input layer is consistent with the dimension of the comprehensive feature vector. If the dimension of the comprehensive feature vector is u, then the input layer has u neurons;

[0084] The competitive layer is a two-dimensional grid structure containing A×B neurons (A and B are the number of grid rows and columns, determined by the amount of historical data and computing resources). The input layer and the competitive layer neurons are connected through the weight vector ω of the competitive layer neurons.

[0085] The comprehensive feature vector Input the data to the input layer of the self-organizing map network and obtain the distance between the input data and the weight vector ω of each neuron in the competition layer. The neuron with the smallest distance is the winning neuron c. The distance calculation formula is:

[0086]

[0087] The winning neuron is represented as c = argmin(d). After determining the winning neuron c, the weight vectors of the winning neuron and the neurons in its neighborhood are updated. The neighborhood is defined as a local area centered on the winning neuron c. As the training progresses, the neighborhood radius gradually decreases. The weight update formula is:

[0088]

[0089] Among them, h c is the domain function, θ is the number of training times, α θ is the learning rate for the number of training times θ, ω θ is the weight of the training times θ, ω θ+1 is the updated weight of the weight after training times θ;

[0090] After multiple rounds of training (set the number of training rounds to Continuously select data from the comprehensive feature vector to perform the above training steps) The competition layer neurons of the self-organizing map network cluster the historical comprehensive feature vectors. Each competition layer neuron represents a class of comprehensive feature vectors with similar characteristics. According to different environmental conditions (different interval combinations of wind speed, humidity, and temperature environmental factors), the competition layer neurons are divided into P subspaces, represented as [Ω1, Ω2, .., Ω P ], where Ω represents a subspace;

[0091] Specifically, by dividing the wind speed, humidity, and temperature into intervals, different environmental condition combination categories are formed. For example, the wind speed is divided into p1 intervals (assuming that p1 interval includes low speed, low humidity, and low temperature), the wind speed is divided into p2 intervals (assuming that p2 intervals include low speed, low humidity, and normal temperature), and the wind speed is divided into p3 intervals (assuming that p3 intervals include low speed, low humidity, and high temperature). Then the total number of environmental condition combination categories is P = p1 + p2 + p3, and P different environmental condition categories are finally determined.

[0092] For any specific combination of environmental conditions, the neurons that respond more to the data under the environmental conditions during training are divided into a subspace Ω P In each subspace Ω P Store the comprehensive feature vectors under the corresponding environmental conditions to obtain the optimized state space;

[0093] The process of obtaining waveform timing similarity is as follows:

[0094] Through an improved dynamic programming algorithm, traditional dynamic programming algorithms usually only consider three adjacent directions (up, left, and upper-left) when backtracking to find the optimal path. This solution expands the search to consider five directions (up, left, upper-left, upper-right, and lower) to find the optimal path. The waveform timing similarity is calculated based on the cumulative distance corresponding to the optimal path.

[0095] Specifically, traditional dynamic programming algorithms typically only consider three adjacent directions when backtracking to find the optimal path. This scheme expands this to consider five directions (up, left, upper left, upper right, and lower). This is more consistent with the complex and changeable characteristics of traveling wave current waveforms, can fully capture the similarity of waveforms in time series, and more accurately calculate waveform timing similarity.

[0096] The process of calculating the waveform timing similarity based on the cumulative distance corresponding to the optimal path is:

[0097] Real-time collection of traveling wave current waveform data, assuming that the real-time collection of traveling wave current waveform data is

[0098] Define a dynamic programming table [DP];

[0099] The elements in the dynamic programming table represent the minimum cumulative distance from the starting point to the coordinate position, that is, the elements in the table Represents the minimum cumulative distance from the starting point to the coordinate location. The minimum cumulative distance is calculated by weighted distance. And the minimum value of the cumulative distance between adjacent positions in the dynamic programming table is jointly determined. The formula for the minimum cumulative distance from the starting point to the coordinate position is expressed as:

[0100] This formula shows that the minimum cumulative distance of the current position is obtained by adding the weighted distance of the current position and the minimum cumulative distance of the five adjacent directions;

[0101] Weighted distance By introducing the Euclidean distance, The traveling wave current waveform data is Middle elements and the comprehensive eigenvector in the optimization state space Middle The weighted distance between elements (obtained directly through Euclidean distance);

[0102] Specifically, Indicates the upward direction, Indicates the left direction, Indicates the upper left direction, Indicates the upper right direction, Indicates the downward direction. Different directions reflect the cumulative distance from different locations to the current location.

[0103] At the same time, set a boundary condition;

[0104] The starting value of the boundary condition is It is clear that the cumulative distance of the starting point is the weighted distance of the starting point, and the boundary values of the edges and columns are obtained:

[0105] For example: For the first column Elements on Starting from the starting point and moving in the left direction, the weighted distance of each step is accumulated in turn to obtain That is, the column boundary values, and for the first row Elements on Starting from the starting point and moving upward, the weighted distance of each step is accumulated in sequence to obtain That is, the row boundary value;

[0106] After determining the two boundary values, for all positions in the dynamic programming table ( and ) elements, and calculate and fill the entire dynamic programming table based on the recursive formula;

[0107] Specifically, DP [1][1] For example, you need to first obtain View its five adjacent directions at the same time (top: DP [0][1] , Left: DP [1][0] , Upper left: DP [0][0] , Upper right: DP [2][0] , Next: DP [0][2] ) position has been calculated cumulative distance value, and then select the minimum value from it, and then Add them together to get DP [1][1] The value of is calculated from the beginning of the table and gradually moves to other positions. Each calculation of a position depends on the value of the adjacent position that has been calculated before. This process is repeated continuously to gradually fill the entire dynamic programming table.

[0108] Based on the dynamic programming table after filling, the principal component analysis (PCA) method is used to analyze the traveling current waveform data. And the comprehensive eigenvector in the optimized state space Perform dimensionality reduction to obtain the traveling wave current waveform data after dimensionality reduction and the comprehensive eigenvector in the optimized state space And the dimension of the traveling wave current waveform data after dimensionality reduction is The dimension of the comprehensive feature vector after dimensionality reduction in the optimization state space is

[0109] The dimension of the traveling wave current waveform data after dimension reduction is The dimension of the comprehensive feature vector after dimensionality reduction in the optimized state space is As the judgment length in the dynamic programming table, when calculated into the table Position (here The length of the traveling current waveform data collected in real time after dimensionality reduction, The length of the comprehensive feature vector in the optimized state space after dimensionality reduction) is obtained. The cumulative distance obtained at this time is the waveform time series similarity S sim ;

[0110] Specifically, the comprehensive feature vectors are clustered through the self-organizing mapping network, and the competition layer neurons are divided into multiple subspaces according to environmental conditions to construct an optimized state space, so that the comprehensive feature vectors with similar characteristics are classified together, so that when calculating the waveform timing similarity, they can be compared in a more accurately matched historical data set. At the same time, the waveform timing similarity calculated by the improved dynamic time warping algorithm provides a quantitative basis for fault judgment. When the similarity between the real-time collected traveling wave current waveform data and the historical fault waveform data in the optimized state space exceeds the set threshold, the occurrence of tree line discharge fault can be accurately judged.

[0111] S4: Setting a fault similarity threshold, when the waveform time sequence similarity of the real-time collected traveling wave current waveform data is greater than or equal to the fault similarity threshold, marking the corresponding traveling wave current waveform data as abnormal.

[0112] The waveform timing similarity S of the traveling wave current waveform data and the optimized state space is obtained in real time based on the improved dynamic time warping algorithm. sim , the calculated waveform timing similarity S sim Compare with the pre-set fault similarity threshold σ;

[0113] Specifically, the fault similarity threshold σ can be preset by collecting historical fault data. By drawing similarity distribution histogram curves of historical normal data and historical fault data, the statistics of the similarity of each type of data are calculated. Through these statistics and distribution graphs, the distribution characteristics of the similarity between normal data and fault data can be intuitively obtained, thereby determining the fault similarity threshold σ;

[0114] If S sim≥σ, the corresponding traveling current waveform data is marked as abnormal, and it is determined that a tree line discharge fault has occurred in the real-time data. At this time, the alarm mechanism can be triggered to send fault warning information to the operation and maintenance personnel;

[0115] If S sim <σ, the current traveling wave current waveform data is considered normal, and the subsequent traveling wave current waveform data are monitored in real time, and the above-mentioned process of acquisition, processing, similarity calculation and judgment is repeated.

[0116] In the application, several formulas involved are calculated by taking their numerical values after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent real situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they will not be elaborated here.

[0117] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0118] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A tree-line discharge fault identification method based on big data fault traveling wave current characteristics is characterized by: include: S1: Collecting fault traveling wave current waveform data and environmental data, the environmental data including wind speed, humidity, and temperature data; S2: Perform adaptive discrete wavelet transform on the fault traveling current waveform to obtain the fault wavelet coefficient. Based on the fault wavelet coefficient, the enhanced fault wavelet coefficient is obtained by improving the dynamic attention network. The optimal frequency band energy of the enhanced fault wavelet coefficient is obtained. The enhanced fault wavelet coefficient, the optimal frequency band energy and the environmental data are spliced into a comprehensive feature vector. Among them, the improved dynamic attention network is achieved by adding an adaptive mechanism to the dynamic changes of time series to the traditional attention mechanism to enhance the core feature representation of the fault wavelet coefficients; S3: Based on the comprehensive eigenvector, a self-organizing map network is introduced to construct an optimized state space. The traveling wave current waveform data collected in real time is used to obtain the waveform timing similarity between the traveling wave current waveform data and the optimized state space using an improved dynamic time warping algorithm; Among them, the improved dynamic time warping algorithm is extended to five implementations based on the traditional dynamic programming algorithm which usually only considers three adjacent directions when finding the optimal path; S4: Setting a fault similarity threshold, when the waveform time sequence similarity of the real-time collected traveling wave current waveform data is greater than or equal to the fault similarity threshold, marking the corresponding traveling wave current waveform data as abnormal.

2. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 1 is characterized in that: The process of obtaining the fault wavelet coefficients is as follows: On the basis of traditional discrete wavelet transform, an adaptive wavelet basis selection mechanism is introduced to add scale parameters and translation parameters to perform adaptive wavelet basis selection to obtain fault wavelet coefficients.

3. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 2 is characterized in that: The process of obtaining the enhanced fault wavelet coefficients is as follows: Obtain feature representations of the fault wavelet coefficients, perform linear transformation on the feature representations, and obtain attention scores between different feature representations; Based on the attention scores between different feature representations, an optimized time series dynamic weight matrix is introduced to adjust the attention scores between different feature representations according to the time series change trend between different feature representations; The element values in the obtained optimized time series dynamic weight matrix are used to adjust the corresponding attention scores according to the changing trend of the signal in the time series; The attention scores corresponding to the characteristic representation of the fault wavelet coefficients are optimized by optimizing the time series dynamic weight matrix; The optimized attention weight matrix is used to directly perform weighted summation on the feature representation of the fault wavelet coefficients to obtain the enhanced feature representation of the fault wavelet coefficients.

4. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 3 is characterized in that: The process of obtaining the optimized time series dynamic weight matrix is as follows: Obtain the characteristic representation of the fault wavelet coefficients to obtain the rate of change sequence in different time windows; The rate of change sequence is used as the input of a convolutional neural network, where the convolutional network consists of L convolutional layers. The parameters of each convolutional layer include the convolution kernel and bias. After processing by L convolutional layers, the feature map is output. Then, a fully connected layer is used to map the feature map into an optimized time series dynamic weight matrix with the same dimension as the attention score matrix.

5. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 4 is characterized in that: The optimization state space construction process is as follows: Introducing the self-organizing map network, which includes an input layer and a competition layer; Input the comprehensive feature vector into the input layer of the self-organizing map network, obtain the distance between the input data and the weight vector of each neuron in the competition layer, and obtain the winning neuron based on the neuron with the smallest distance; The neurons in the competition layer are divided into P subspaces. For any specific combination of environmental condition intervals, the neurons that respond more to the data under the environmental conditions during training are divided into a subspace. Each subspace stores the comprehensive feature vector under the corresponding environmental conditions to obtain the optimized state space.

6. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 5 is characterized in that: The waveform timing similarity is obtained by improving the dynamic programming algorithm according to the cumulative distance corresponding to the optimal path.

7. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 6 is characterized in that: The process of calculating the waveform timing similarity based on the cumulative distance corresponding to the optimal path is: Define a dynamic programming table, where the elements of the dynamic programming table represent the minimum cumulative distance from the starting point to the coordinate position; Set a boundary condition and obtain the boundary values of the edge and column, where the cumulative distance of the starting point is the weighted distance of the starting point; After determining the boundary values of the edges and columns, a dynamic programming table is filled in. Based on the filled dynamic programming table, the real-time collected traveling wave current waveform data and the dimension of the comprehensive feature vector in the optimized state space after dimensionality reduction are obtained. The reduced dimension is used as the judgment length in the dynamic programming table, and the waveform timing similarity is obtained based on the judgment length.

8. The tree-line discharge fault identification method based on big data fault traveling wave current characteristics according to claim 7 is characterized in that: The minimum cumulative distance is obtained by adding the minimum value of the weighted distance and the minimum cumulative distances of the five adjacent direction positions; The weighted distance is obtained by obtaining the distance between the elements in the traveling wave current waveform data and the elements of the comprehensive characteristic vector in the optimized state space through the Euclidean distance.

Images