Power transmission line fault diagnosis method based on time-frequency multistage fusion and bidirectional time sequence enhancement network

Abstract

The invention relates to a power transmission line fault diagnosis method based on time-frequency multistage fusion and a bidirectional time sequence enhancement network, and belongs to the field of power system power transmission line fault diagnosis. The method comprises the following steps: firstly, constructing a time-frequency multi-level fusion and bidirectional time sequence enhancement network: converting input power transmission line partial discharge signal data into a time-frequency spectrogram; constructing a multi-scale time-frequency feature joint extraction module, and performing feature extraction on the time-frequency spectrogram by dual offset branches; constructing a dynamic weighted fusion module, and carrying out adaptive fusion on the extracted time-frequency features; a bidirectional time sequence information enhancement module is constructed, modeling is performed on the fusion features through a bidirectional long-short-term memory network, and feature expression of key discharge moments is enhanced by adopting an attention guiding mechanism; and finally, training the time-frequency multi-level fusion and bidirectional time sequence enhancement network by using the training set, and finally realizing fault type identification through a full connection layer and a Softmax classifier. According to the invention, the accuracy of power transmission line fault diagnosis and the model robustness are improved.

Description

Technical Field

[0001] The present invention relates to a transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network, and belongs to the technical field of transmission line fault diagnosis in power systems. Background Art

[0002] As a key component of the power system, the operating state of the transmission line is directly related to the stability of the power grid and the power supply safety. However, due to the long-term exposure of the transmission line to a complex natural environment, it is vulnerable to various external factors such as lightning strikes, pollution, tree contact, and wind deflection, which may cause various faults, such as corona discharge, surface discharge, floating discharge, and partial arc discharge. If these faults cannot be detected and processed in a timely and accurate manner, it may cause serious power accidents and even large-scale power outages. Therefore, studying efficient and accurate transmission line fault diagnosis methods has important engineering value and practical application significance.

[0003] Traditional transmission line fault diagnosis methods mainly include methods based on physical models, signal processing, and machine learning. Physical models rely on electromagnetic field and circuit theories, but due to the complex structure of the transmission line, it is difficult to establish an accurate mathematical model. Although signal processing methods (such as wavelet transform and short-time Fourier transform) can extract partial discharge characteristics, their performance is limited by the selection of preset basis functions and has poor adaptability. In recent years, with the development of deep learning technology, fault diagnosis methods based on deep learning have received extensive attention. Such methods can automatically extract high-dimensional features in data and have strong feature expression ability and generalization ability. However, existing deep learning methods still face certain challenges. On the one hand, the partial discharge signals of the transmission line have complex time-frequency characteristics, and it is difficult to comprehensively characterize the fault information only by relying on a single feature extraction method. On the other hand, the traditional attention mechanism has a high computational complexity and is difficult to efficiently capture the key features in the discharge signal. In addition, there is usually an obvious time series dependence relationship in the partial discharge signal. How to effectively model its time series information on the basis of feature extraction to improve the accuracy and robustness of fault diagnosis is still an urgent problem to be solved. Summary of the Invention

[0004] Aiming at the above existing problems, the present invention proposes a transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network. The method of the present invention can efficiently mine the time-frequency characteristics of partial discharge signals and improve the accuracy and robustness of transmission line fault diagnosis by integrating advanced technologies such as multi-scale time-frequency feature extraction, dynamic weighted fusion, time series information modeling, and guided attention mechanism.

[0005] The technical solution of the present invention is: a transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network. The method includes: constructing a time-frequency multi-level fusion and bidirectional time series enhancement network model, training the model using a training set, and finally realizing fault type recognition through a fully connected layer and a Softmax classifier. The specific process of constructing the time-frequency multi-level fusion and bidirectional time series enhancement network model includes:

[0006] Step1: Use the short-time Fourier transform to convert the input partial discharge signal data of the transmission line into a time-frequency spectrogram;

[0007] Step2: Construct a multi-scale time-frequency feature joint extraction module, and use high-frequency, medium-frequency, and low-frequency dual offset branches to extract time-frequency features from the time-frequency spectrogram respectively;

[0008] Step3: Construct a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features;

[0009] Step4: Construct a bidirectional time series information enhancement module, model the time series information through a bidirectional long short-term memory network BiLSTM, and use a guided attention mechanism to enhance the feature expression of key discharge moments.

[0010] Furthermore, in the Step1, when constructing the time-frequency multi-level fusion and bidirectional time series enhancement network model, the data set acquisition method is as follows:

[0011] Step 1: Collect a data set of partial discharge signals of transmission lines with labeled fault types;

[0012] Step 2: Use wavelet transform to denoise the partial discharge signal data of the transmission line, and use the Z-score normalization method to normalize the partial discharge signal data of the transmission line;

[0013] Step 3: Divide the processed partial discharge signal data of the transmission line, and randomly select 80% as the training set and 20% as the test set.

[0014] Furthermore, in the Step 1, the partial discharge signal data set of the transmission line includes 6 fault types: corona discharge, surface discharge, internal discharge, suspended discharge, crack discharge, and arc discharge; the data source of the partial discharge signal data set of the transmission line includes public data sets, databases, and partial discharge signal data collected from relevant literatures.

[0015] Furthermore, in the Step 2, the wavelet transform separates low-frequency useful signals and high-frequency noise through multi-scale decomposition, and uses a threshold method to suppress noise; the Z-score normalization method calculates the mean μ and standard deviation σ of the partial discharge signal data of the transmission line, according to the formula The transmission line partial discharge signal data is normalized, where X is the original transmission line partial discharge signal data point.

[0016] Furthermore, in Step 1, the short-time Fourier transform is specifically implemented as follows:

[0017] First, a sliding window is selected to segment the partial discharge signal data of the transmission line, and Fourier transform is performed on the partial discharge signal data of each segment of the transmission line. Then, the spectrum diagram of each time window is synthesized into a two-dimensional time-spectrogram to show the dynamic changes of the signal in the time and frequency domains.

[0018] Furthermore, in Step 2, the multi-scale time-frequency feature joint extraction module includes three parallel branches: high-frequency, medium-frequency, and low-frequency dual offset branches. The high-frequency, medium-frequency, and low-frequency dual offset branches use 3×3, 5×5, and 7×7 convolution kernels to extract high-frequency, medium-frequency, and low-frequency features, respectively. Except for the different convolution kernel sizes, the high-frequency, medium-frequency, and low-frequency dual offset branch structures remain consistent. The specific operations involved in each branch are:

[0019] The time-frequency spectrum X generated in Step 1 m First, a 3×3 / 5×5 / 7×7 convolution layer is used to extract the initial features. Then, the Sigmoid activation function is used to smooth the feature map. Then, the smoothed features are subtracted and added respectively to obtain the reverse offset features X. r and the positive offset feature X f ; Feature X will then be offset in reverse r and the positive offset feature X f Respectively with the time-spectrum diagram X m Perform matrix multiplication and add the results; finally, pass through the average pooling layer, 1×1 convolution layer and normalization layer in sequence to obtain the time-frequency features Where i = 1, 2, 3, represents the i-th scale branch.

[0020] Furthermore, in Step 3, the dynamic weighted fusion module uses channel attention to perform weighted fusion on the time-frequency features of branches at different scales; the specific steps include:

[0021] Step 3.1, time-frequency features of the input Perform global average pooling to extract global information; where C is the number of channels, H and W are the spatial dimensions of the feature map;

[0022]

[0023] Among them, i represents the i-th scale branch, G i ∈R CRepresents the global feature of the i-th scale branch;

[0024] Step3.2. Calculate the weight ω of the global feature of the i-th scale branch through two fully connected layers, ReLU activation function, and Sigmoid activation function i , and normalize the weight ω of the global feature of the i-th scale branch i to obtain the normalized value α of the corresponding weight i ;

[0025] ω i = σ(W2δ(W1G i ))

[0026]

[0027] where W1 and W2 are fully connected layers of trainable parameters, δ(·) is the ReLU activation function, and σ(·) is the Sigmoid activation function, which is used to normalize the weight to [0, 1];

[0028] Step3.3. Calculate the fused feature through weighted fusion

[0029] Furthermore, the said Step4 includes: The bidirectional long short-term memory network BiLSTM is used to capture the temporal dependence of the partial discharge signal data of the transmission line, and combined with the improved guided attention mechanism to focus on the key temporal information in the partial discharge signal data of the transmission line; The specific steps include:

[0030] Step4.1. The bidirectional long short-term memory network BiLSTM propagates information bidirectionally by combining the forward long short-term memory network and the backward long short-term memory network to more comprehensively mine the temporal dependence relationship of the partial discharge signal data of the transmission line;

[0031] Step4.2. Based on the temporal feature H output by the bidirectional long short-term memory network BiLSTM, introduce a guided attention mechanism to calculate the attention weight of each time step, so that the key time steps contribute more to the fault classification; The guided attention mechanism includes two self-attention stages, where:

[0032] The first stage: Introduce a guided token G, where G is generated by performing a pooling operation on the query matrix Q and serves as the global representation of the original query; Use this guided token G to replace the original query matrix Q and guide the key matrix K and value matrix V for feature aggregation to generate the guided feature H G ;

[0033] The second stage: Use the guided token G as the key and the guided feature H GAs a value, broadcast global information, transfer global information from the guiding feature to each query, and finally generate the optimized attention-enhanced feature H GA 。

[0034] Further, the Step4.1 includes:

[0035] Step4.1.1. Given the fused feature X F ={x1, x2, …, x t …, x T}, x t ∈R d , where T is the time step and d is the dimension of the fused feature. Each time step corresponds to a d-dimensional feature vector;

[0036] The forward long short-term memory network processes the fused feature X in sequence from t = 1 to t = T F and calculates the forward hidden state

[0037]

[0038] where, represents the forward hidden state at time t, represents the hidden state of the previous time step, that is, the forward hidden state at time t - 1, combines with the current input x t to calculate the state of the current time step; x t represents the current input, that is, the fused feature X F the fused feature vector at time t, depends on the hidden state of the previous time step and the current input x t ;

[0039] Step4.1.2. The backward long short-term memory network processes the fused feature X in the reverse order, from t = T to t = 1 F the fused feature vector x t at time t, calculates the backward hidden state

[0040]

[0041] where, depends on the backward hidden state at time t + 1 and the fused feature X F the fused feature vector x t at time t;

[0042] Step4.1.3. Concatenate the forward hidden state at time t and the backward hidden state Obtain the final bidirectional hidden state h t :

[0043]

[0044] where h t is the final bidirectional hidden state, containing past and future information;

[0045] Step4.1.4. The bidirectional long short-term memory network finally aggregates the bidirectional hidden states of all time steps to output the temporal feature H; the temporal feature H is expressed as:

[0046] H = {h1, h2, … h t …, h T}, H ∈ R T×2d .

[0047] Furthermore, the said Step4.2 includes:

[0048] Step4.2.1. First, perform a linear transformation on the temporal feature H to obtain the query matrix Q, the key matrix K, and the value matrix V respectively:

[0049] Q = W Q H, K = W K H, V = W V H

[0050] where W Q , W K , W V are trainable parameter matrices;

[0051] Step4.2.2. Perform a global summary on the query matrix Q through global average pooling operation to obtain the guiding token G:

[0052] G = GAP(Q)

[0053] where GAP(·) represents the global average pooling operation, and the guiding token G is used as the global representation of the entire temporal feature, capable of capturing global feature information;

[0054] Step4.2.2. Use the guiding token G to replace the original query matrix Q, and guide the key matrix K and the value matrix V to perform feature aggregation to generate the guiding feature H G , the process of generating the guiding feature H G is expressed as:

[0055]

[0056] where Softmax(·) represents the softmax activation function, d k is the dimension of the key matrix;

[0057] Step 4.2.3: Using the guiding token G as the key and the guiding feature H G as the value, broadcast the global information, and transfer the global information from the guiding feature H G to each query, and finally generate the optimized attention-enhanced feature H GA ; The process of generating the optimized attention-enhanced feature H GA is expressed as:

[0058]

[0059] where Q is the query matrix;

[0060] Step 4.2.4: Add a residual connection to the optimized attention-enhanced feature H GA :

[0061] H′ GA = H GA + H.

[0062] The beneficial effects of the present invention are:

[0063] 1. The present invention constructs a novel time-frequency multi-level fusion and bidirectional time-series enhancement network architecture, realizes efficient time-frequency domain feature mining, and significantly improves the performance of the transmission line fault diagnosis task;

[0064] 2. The present invention proposes a time-frequency multi-level fusion mechanism, combines the short-time dynamic features of the time-domain signal and the global spectral characteristics of the frequency-domain signal, realizes more comprehensive fault feature extraction, and improves the accuracy of fault diagnosis compared with the single-domain feature extraction method;

[0065] 3. The present invention introduces the BiLSTM architecture, fully captures the forward and backward dependencies of the discharge signal, effectively models the time-series characteristics of the fault signal, enhances the separability of fault patterns, and improves the stability of the diagnosis model;

[0066] 4. The present invention proposes a guiding attention mechanism, which can effectively reduce the computational complexity of the traditional self-attention mechanism, dynamically allocate the weights of different features, strengthen the contribution of key fault features to the classification decision, and thus improve the discrimination ability of the model;

[0067] 5. The present invention improves the feature expression ability, can effectively improve the accuracy, stability and adaptability of transmission line fault diagnosis, and has wide engineering application value. Description of the Drawings

[0068] Figure 1 is the overall process schematic diagram of the present invention;

[0069] Figure 2It is a schematic structural diagram of the time-frequency multi-level fusion and bidirectional time series enhancement network in the present invention;

[0070] Figure 3 It is a schematic structural diagram of the dynamic weighted fusion module in the present invention;

[0071] Figure 4 It is a schematic structural diagram of the long short-term memory network in the present invention;

[0072] Figure 5 It is a schematic structural diagram of the guided attention mechanism in the present invention. Detailed implementation manners

[0073] Example 1: As Figures 1 - 5 shown, a transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network, the method steps include:

[0074] (1) Data collection: Collect a partial discharge signal dataset of transmission lines with labeled types; the partial discharge signal dataset of the transmission lines includes 6 types of fault types, namely corona discharge, surface discharge, internal discharge, suspended discharge, crack discharge, and arc discharge; the dataset sources include public datasets and transmission line partial discharge signal data collected from databases and relevant literatures.

[0075] (2) Data preprocessing: Use wavelet transform to denoise the partial discharge signal data of the transmission line, and use the Z-score normalization method to normalize the partial discharge signal data of the transmission line;

[0076] In step 2, the wavelet transform separates low-frequency useful signals and high-frequency noise through multi-scale decomposition, and uses the threshold method to suppress noise. Through threshold processing, high-frequency noise is removed at each scale, while retaining the useful information in the signal; the Z-score normalization method calculates the mean (Mean) μ and standard deviation (Standard Deviation) σ of the partial discharge signal data of the transmission line, and normalizes the partial discharge signal data of the transmission line according to the formula where X is the original partial discharge signal data point of the transmission line.

[0077] (3) Data selection and division: Randomly sample 80% of the data processed in Step 2 as the training set and 20% as the test set to verify the accuracy and rationality of the analysis model.

[0078] (4) Construct a time-frequency multi-level fusion and bidirectional time series enhancement network model, and use the training set for training to extract feature representations from the data. At the same time, adjust the importance of the features to make the model more focused on key features. Input the preprocessed and labeled data into the time-frequency multi-level fusion and bidirectional time series enhancement network to learn discriminative features, referring to Figure 2 , the time-frequency multi-level fusion and bidirectional time series enhancement network model includes a multi-scale time-frequency joint feature extraction module, a dynamic weighted fusion module, and a bidirectional time series information enhancement module; the output of the model passes through a fully connected layer and is classified by the Softmax function for faults; the specific steps for constructing the time-frequency multi-level fusion and bidirectional time series enhancement network model are as follows:

[0079] Step1. Use the Short-Time Fourier Transform (STFT) to convert the input partial discharge signal data of the transmission line into a time-frequency spectrogram X m ;

[0080] The specific implementation method of the short-time Fourier transform is as follows:

[0081] First, select a sliding window to segment the partial discharge signal data of the transmission line, and perform Fourier transform on each segment of the partial discharge signal data of the transmission line. The Fourier transform result of each segment will provide the frequency information of that segment of the signal; then, synthesize the spectrograms of each time window into a two-dimensional time-frequency spectrogram to show the dynamic changes of the signal in the time and frequency domains.

[0082] Use STFT to convert the partial discharge signal of the transmission line into the corresponding time-frequency spectrogram and judge the fault type by fully mining its time-frequency features, effectively improving the accuracy of fault diagnosis; the fault type judgment method is shown in the following table:

[0083]

[0084] Step2. Construct a multi-scale time-frequency feature joint extraction module, and use high-frequency, medium-frequency, and low-frequency dual offset branches to extract time-frequency features from the time-frequency spectrogram X m respectively;

[0085] Furthermore, in Step2, the multi-scale time-frequency feature joint extraction module includes three parallel branches: high-frequency, medium-frequency, and low-frequency dual offset branches. The high-frequency, medium-frequency, and low-frequency dual offset branches use 3×3, 5×5, and 7×7 convolutional kernels respectively to extract high-frequency, medium-frequency, and low-frequency features; except for the different convolutional kernel sizes, the structures of the high-frequency, medium-frequency, and low-frequency dual offset branches are the same; the specific operations involved in each branch are as follows:

[0086] The time-frequency spectrogram X generated in Step1m Firstly, initial feature extraction is performed through a 3×3 / 5×5 / 7×7 convolutional layer; then the Sigmoid activation function is used to smooth the feature map. The Sigmoid activation function maps the feature values to a fixed range, making the feature distribution smoother, preventing the influence of extreme values, and at the same time increasing the sensitivity of the model to low-amplitude discharge patterns; then subtraction by one and addition by one operations are respectively performed on the smoothed features to obtain the reverse-offset feature X r and the forward-offset feature X f , thus effectively expanding the representation range of the features, enhancing the feature expression ability, enabling the model to capture richer time-frequency features in the partial discharge signal, and improving the classification accuracy and robustness; subsequently, the reverse-offset feature X r and the forward-offset feature X f are respectively multiplied by the time-frequency spectrogram X m in matrix multiplication and the results are added up, which not only retains the original distribution information of the features, but also enhances the non-linear expression ability of the features, and at the same time introduces new feature information to achieve information enhancement; finally, through an average pooling layer, a 1×1 convolutional layer and a normalization layer in sequence, the time-frequency features are obtained where i = 1, 2, 3, representing the i-th scale branch.

[0087] Step3. Construct a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features;

[0088] Furthermore, in the Step3, the dynamic weighted fusion module uses channel attention to perform weighted fusion on the time-frequency features of different scale branches; according to the characteristics of different fault signals, the weights of the time-frequency features of high, medium and low frequencies are adaptively adjusted to improve the flexibility of feature fusion and enhance the fusion effect; the specific steps include:

[0089] Step3.1. The dynamic weighted fusion module receives three features from the multi-scale time-frequency feature joint extraction module, and performs global average pooling (GAP) on the input time-frequency features to extract global information of the time-frequency features representing high, medium and low frequencies respectively; where C is the number of channels, and H, W are the spatial dimension sizes of the feature map;

[0090]

[0091] where i represents the i-th scale branch, and G i ∈R C represents the global feature of the i-th scale branch;

[0092] Step3.2. Calculate the weight ω of the global feature of the i-th scale branch through two fully connected layers, the ReLU activation function and the Sigmoid activation functioni The weight ω of the global feature of the i-th scale branch i is normalized to obtain the normalized value α of the corresponding weight i ;

[0093] ω i = σ(W2δ(W1G i ))

[0094]

[0095] where W1 and W2 are fully connected layers of trainable parameters, δ(·) is the ReLU activation function, and σ(·) is the Sigmoid activation function, which is used to normalize the weight to [0,1]; calculate the normalized value of ω i to ensure that the sum of the weights is 1;

[0096] Step3.3. Calculate the fused feature through weighted fusion Ensure that the network can adaptively focus on the most discriminative features under different fault modes through dynamic weighted fusion.

[0097] Step4. Construct a bidirectional temporal information enhancement module, and model the temporal information through a Bidirectional Long Short-Term Memory (BiLSTM) and enhance the feature expression at the key discharge moment by using a guided attention mechanism.

[0098] Furthermore, the Step4 includes: The bidirectional long short-term memory network BiLSTM is used to capture the temporal dependence of the partial discharge signal data of the transmission line, model the temporal dependence of the partial discharge signal, enhance the expression ability of the fault features, and combine the improved guided attention mechanism to focus on the key temporal information in the partial discharge signal data of the transmission line, which can effectively reduce the computational complexity of the traditional self-attention mechanism, strengthen the contribution of the key temporal information to the classification decision, and improve the discrimination ability of the model;

[0099] The bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, BiLSTM) is an extension of the traditional long short-term memory network (Long Short-Term Memory, LSTM), which includes a forward LSTM (Forward LSTM) and a backward LSTM (Backward LSTM) for capturing the forward and backward dependencies of time series data;

[0100] BiLSTM is extended on the basis of the traditional LSTM. The core advantage of the bidirectional long short-term memory network BiLSTM is that it can capture temporal information from both the past and the future simultaneously, enhancing the model's ability to represent complex time series. In the task of transmission line fault diagnosis, the temporal characteristics of partial discharge signals are particularly important.

[0101] The specific steps of Step 4 include:

[0102] Step 4.1: The bidirectional long short-term memory network BiLSTM propagates information bidirectionally by combining the forward long short-term memory network (Forward LSTM) and the backward long short-term memory network (Backward LSTM), more comprehensively mining the temporal dependence relationship of the partial discharge signal data of the transmission line, and improving the accuracy of fault classification.

[0103] Based on the temporal features H output by the bidirectional long short-term memory network BiLSTM, a guided attention mechanism is introduced to calculate the attention weights at each time step, making the contribution of key time steps to fault classification greater. The guided attention mechanism includes two self-attention stages, extracting key temporal features with the global guiding token G and enhancing the global information propagation ability, thereby optimizing the final attention-enhanced features H. GA ;

[0104] By introducing a global guiding token, the guided attention mechanism converts the pairwise interaction of all time steps in the traditional self-attention mechanism into two low-dimensional attention stages, significantly reducing the computational complexity. Compared with the computational complexity of the traditional self-attention mechanism O(T 2 d), the guided attention mechanism only requires O(nTd) (where n << T), effectively reducing the computational amount and memory occupancy, making it more efficient in processing long-sequence fault signals. In addition, while reducing the computational complexity, the guided attention mechanism can retain the information of key time steps and enhance the feature expression ability through global guidance, thereby improving the computational efficiency and training stability of the model while ensuring the classification accuracy.

[0105] Based on the temporal features H output by BiLSTM, a guided attention mechanism is introduced to calculate the attention weights at each time step, emphasizing the contribution of key time steps to fault classification. The guided attention mechanism consists of two self-attention stages, extracting key temporal features with the global guiding token G and enhancing the global information propagation ability, thereby optimizing the final attention-enhanced features H. GA ;

[0106] The two self-attention stages specifically include:

[0107] The first stage: Introduce the guiding token G, where G is generated by performing a pooling operation on the query matrix Q and serves as the global representation of the original query; use this guiding token G to replace the original query matrix Q and guide the key matrix K and value matrix V to perform feature aggregation to generate the guiding feature H G ;

[0108] The second stage: Use the guiding token G as the key and the guiding feature H G as the value to broadcast the global information, transfer the global information from the guiding feature to each query, and finally generate the optimized attention-enhanced feature H GA .

[0109] Furthermore, the Step4.1 includes:

[0110] Step4.1.1. The bidirectional long short-term memory network BiLSTM receives the fused feature X F ∈R T×d output by the dynamic weighted fusion module. For the fused feature X F ={x1, x2, … x t …, x T}, x t ∈R d , where T is the time step and d is the dimension of the fused feature. Each time step corresponds to a d-dimensional feature vector;

[0111] The forward long short-term memory network processes the fused feature X in sequence from t = 1 to t = T F and calculates the forward hidden state

[0112]

[0113] where, represents the forward hidden state at time t, represents the hidden state of the previous time step, that is, the forward hidden state at time t - 1, combines with the current input x t to calculate the state of the current time step; x t represents the current input, that is, the fused feature vector of the fused feature X F at time t, depends on the hidden state of the previous time step and the current input x t ;

[0114] Step4.1.2. The backward long short-term memory network processes the fused feature X in the opposite order, from t = T to t = 1 F The fused feature vector x t at time t, calculates the backward hidden state

[0115]

[0116] Among them, depending on the backward hidden state at time t+1 and the fused feature X F the fused feature vector x at time t t ;

[0117] Step4.1.3. Concatenate the forward hidden state at time t and the backward hidden state at time t to obtain the final bidirectional hidden state h t :

[0118]

[0119] Among them, h t is the final bidirectional hidden state, containing past and future information;

[0120] Step4.1.4. The bidirectional long short-term memory network finally aggregates the bidirectional hidden states of all time steps to output the temporal feature H; the temporal feature H is expressed as:

[0121] H = {h1, h2, … h t …, h T}, H ∈ R T×2d .

[0122] Furthermore, the said Step4.2 includes:

[0123] Step4.2.1. First, perform a linear transformation on the temporal feature H to obtain the query matrix Q, the key matrix K, and the value matrix V respectively:

[0124] Q = W Q H, K = W K H, V = W V H

[0125] Among them, W Q 、W K 、W V are trainable parameter matrices;

[0126] Step4.2.2. Perform a global summary on the query matrix Q through global average pooling operation to obtain the guiding token G:

[0127] G = GAP(Q)

[0128] Among them, GAP(·) represents the global average pooling operation, which is used to guide the token G as the global representation of the entire temporal feature and can capture global feature information;

[0129] Step 4.2.2. Use the guiding token G to replace the original query matrix Q, and guide the key matrix K and value matrix V to perform feature aggregation to generate the guiding feature H G , generating the guiding feature H G The process is expressed as:

[0130]

[0131] where Softmax(·) represents the softmax activation function, and d k is the dimension of the key matrix;

[0132] Step 4.2.3. Use the guiding token G as the key and the guiding feature H G as the value to broadcast the global information, and transfer the global information from the guiding feature H G to each query, and finally generate the optimized attention-enhanced feature H GA ; The process of generating the optimized attention-enhanced feature H GA is expressed as:

[0133]

[0134] where Q is the query matrix;

[0135] Step 4.2.4. Add a residual connection to the optimized attention-enhanced feature H GA :

[0136] H′ GA = H GA + H.

[0137] Furthermore, the self-attention calculation process in the guiding attention mechanism includes the following steps:

[0138] Perform a linear transformation on the input feature H to obtain the query matrix Q, key matrix K, and value matrix V respectively:

[0139] Q = W Q H, K = W K H, V = W V H

[0140] where W Q , W K , W V are trainable parameter matrices;

[0141] Calculate the attention score matrix:

[0142]

[0143] where d kis the dimension of the key matrix;

[0144] Perform Softmax normalization on the attention score matrix A:

[0145] A′ = Softmax(A)

[0146] Calculate the aggregated feature:

[0147] H′ = A′V

[0148] where H′ is the final output of the aggregated feature;

[0149] After the second-stage calculation, the finally generated attention-enhanced feature H GA is fused with the original input feature through a residual connection to retain key information and enhance stability.

[0150] The method of the present invention first constructs a transmission line partial discharge signal dataset containing various labeled fault types through data collection, and performs noise reduction and normalization processing on the original signal; then designs a deep learning model based on a time-frequency multi-level fusion mechanism and a bidirectional time series enhancement network, and optimizes network parameters using the training dataset; finally, realizes intelligent classification of typical fault types such as corona discharge, surface discharge, floating discharge, and local arc discharge through the trained model. The present invention innovatively integrates the feature extraction ability of the Convolutional Neural Network (CNN), the bidirectional time series modeling advantage of the Bidirectional Long Short-Term Memory (BiLSTM), and the key feature dynamic focusing ability of the attention mechanism, effectively improving the accuracy and robustness of transmission line fault diagnosis, and having important engineering application value in the field of power system condition monitoring and fault warning.

[0151] The specific embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network, characterized in that: The method includes: constructing a time-frequency multi-level fusion and bidirectional time-series enhancement network model, training the model using a training set, and finally realizing fault type recognition through a fully connected layer and a Softmax classifier; specifically, the construction process of the time-frequency multi-level fusion and bidirectional time-series enhancement network model includes: Step1. Use the short-time Fourier transform to convert the input partial discharge signal data of the transmission line into a time-frequency spectrogram; Step2. Construct a multi-scale time-frequency feature joint extraction module, and use high-frequency, medium-frequency, and low-frequency dual offset branches to extract time-frequency features from the time-frequency spectrogram respectively; Step3. Construct a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features; Step4. Construct a bidirectional time-series information enhancement module, model the time-series information through a bidirectional long short-term memory network BiLSTM, and use a guided attention mechanism to enhance the feature expression at the key discharge moment.

2. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 1, characterized in that: In the above Step1, when constructing the time-frequency multi-level fusion and bidirectional time-series enhancement network model, the data set acquisition method is as follows: Step 1. Collect a data set of partial discharge signals of transmission lines with labeled fault types; Step 2. Use wavelet transform to denoise the partial discharge signal data of the transmission line, and use the Z-score normalization method to normalize the partial discharge signal data of the transmission line; Step 3: Divide the processed partial discharge signal data of the transmission line, and randomly select 80% as the training set and 20% as the test set.

3. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 2, wherein: In the above Step1, the partial discharge signal data set of the transmission line includes 6 fault types: corona discharge, surface discharge, internal discharge, floating discharge, crack discharge, and arc discharge; the data source of the partial discharge signal data set of the transmission line includes public data sets, databases, and partial discharge signal data collected from relevant literatures.

4. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 2, characterized in that: In the said step 2, the wavelet transform separates the low-frequency useful signal and high-frequency noise through multi-scale decomposition, and the threshold method is adopted to suppress the noise; the Z-score normalization method calculates the mean μ and standard deviation σ of the partial discharge signal data of the transmission line, and according to the formula normalizes the partial discharge signal data of the transmission line, where X is the original partial discharge signal data point of the transmission line.

5. The power transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 1, wherein: In the above Step1, the specific implementation method of the short-time Fourier transform is as follows: First, select a sliding window to segment the partial discharge signal data of the transmission line, and perform Fourier transform on each segment of the partial discharge signal data of the transmission line; then synthesize the spectrograms of each time window into a two-dimensional time-frequency spectrogram to show the dynamic changes of the signal in the time and frequency domains.

6. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 1, wherein: In the above Step2, the multi-scale time-frequency feature joint extraction module includes three parallel branches: high-frequency, medium-frequency, and low-frequency dual offset branches. The high-frequency, medium-frequency, and low-frequency dual offset branches use 3×3, 5×5, and 7×7 convolution kernels respectively to extract high-frequency, medium-frequency, and low-frequency features; Except for the different convolution kernel sizes, the structures of the high-frequency, medium-frequency, and low-frequency dual offset branches are the same; the specific operations involved in each branch are: The time-frequency spectrogram X generated in Step1 m First, initial feature extraction is performed through a 3×3 / 5×5 / 7×7 convolutional layer; then the Sigmoid activation function is used to smooth the feature map; next, subtraction by one and addition by one operations are respectively performed on the smoothed features to obtain the reverse offset feature X r and the forward offset feature X f ; subsequently, the reverse offset feature X r and the forward offset feature X f are respectively multiplied by the time-frequency spectrogram X m in matrix multiplication and the results are added together; finally, through an average pooling layer, a 1×1 convolutional layer, and a normalization layer in sequence, the time-frequency feature where i = 1, 2, 3, represents the i-th scale branch.

7. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 1, wherein: In the above Step3, the dynamic weighted fusion module uses channel attention to perform weighted fusion on the time-frequency features of different scale branches; The specific steps include: Step 3.

1. Perform global average pooling on the input time-frequency features to extract global information. Here, C is the number of channels, and H and W are the spatial dimension sizes of the feature map; where \(i\) represents the \(i\)-th scale branch, and \(G\) i \(\in \mathbb{R}\) C represents the global feature of the \(i\)-th scale branch; Step 3.2: Calculate the weight ω of the global feature of the i-th scale branch through two fully connected layers, ReLU activation function, and Sigmoid activation function i , and normalize the weight ω of the global feature of the i-th scale branch i to obtain the normalized value α of the corresponding weight i ; ω i = σ(W2δ(W1G i )) Among them, W1 and W2 are fully connected layers of trainable parameters, δ(·) is the ReLU activation function, and σ(·) is the Sigmoid activation function, which is used to normalize the weight to [0,1]; Step 3.

3. Calculate the fused feature through weighted fusion 8. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 1, wherein: Step 4 includes: The Bidirectional Long Short-Term Memory Network (BiLSTM) is used to capture the temporal dependence of the partial discharge signal data of the transmission line, and combined with an improved guided attention mechanism to focus on the key temporal information in the partial discharge signal data of the transmission line. The specific steps are as follows: Step 4.1: The Bidirectional Long Short-Term Memory Network (BiLSTM) propagates information bidirectionally by combining the forward and backward long short-term memory networks to more comprehensively mine the temporal dependence relationship of the partial discharge signal data of the transmission line. Step 4.2: Based on the temporal feature H output by the Bidirectional Long Short-Term Memory Network (BiLSTM), a guided attention mechanism is introduced to calculate the attention weights at each time step, making the contribution of the key time steps to fault classification greater. The guided attention mechanism includes two self-attention stages, where: The first stage: Introduce the guiding token G, where G is generated by performing a pooling operation on the query matrix Q and serves as the global representation of the original query; use this guiding token G to replace the original query matrix Q and guide the key matrix K and the value matrix V to perform feature aggregation to generate the guiding feature H G ; Second stage: Using the guiding token G as the key and the guiding feature H G as the value, broadcast the global information, transfer the global information from the guiding feature to each query, and finally generate the optimized attention-enhanced feature H GA .

9. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 8, characterized in that: Step 4.1 includes: Step 4.1.

1. Given the fused feature X F ={x1, x2, … x t …, x T}, x t ∈R d , where T is the time step and d is the dimension of the fused feature. Each time step corresponds to a d-dimensional feature vector; The forward long short-term memory network processes the fused feature X sequentially from t = 1 to t = T in the forward direction F , and calculates the forward hidden state Among them, represents the forward hidden state at time t, represents the hidden state of the previous time step, that is, the forward hidden state at time t - 1, combined with the current input x t to calculate the state of the current time step; x t represents the current input, that is, the fused feature X F is the fused feature vector at time t, depends on the hidden state of the previous time step and the current input x t ; Step 4.1.

2. The backward long short-term memory network processes the fused feature X in the reverse order, from t = T to t = 1 F The fused feature vector x at time t t , and calculate the backward hidden state Among them, depending on the backward hidden state at time t + 1 and the fused feature X F the fused feature vector x at time t t ; Step 4.1.3, concatenate the forward hidden state at time t and the backward hidden state at time t to obtain the final bidirectional hidden state h t : where h t is the final bidirectional hidden state, containing information from both the past and the future; Step 4.1.4: The Bidirectional Long Short-Term Memory Network finally aggregates the bidirectional hidden states at all time steps to output the temporal feature H. The temporal feature H is expressed as: H = {h1, h2, … h t …, h T}, H ∈ R T×2d 。 10. The transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time series enhancement network according to claim 8, wherein: Step 4.2 includes: Step 4.2.1: First, perform a linear transformation on the temporal feature H to obtain the query matrix Q, the key matrix K, and the value matrix V respectively: Q = W Q H, K = W K H, V = W V H Among them, W Q , W K , W V are trainable parameter matrices; Step 4.2.2: Perform a global average pooling operation on the query matrix Q to obtain the guiding token G: G = GAP(Q) where GAP(·) represents the global average pooling operation, and the guiding token G is used as the global representation of the entire temporal feature, which can capture the global feature information; Step 4.2.

2. Replace the original query matrix Q with the guiding token G, and guide the key matrix K and the value matrix V to perform feature aggregation to generate the guiding feature H G , generate the guiding feature H G The process is expressed as: where Softmax(·) represents the softmax activation function, and d k is the dimension of the key matrix; Step4.2.3: Using the guiding token G as the key and the guiding feature H G as the value, broadcast the global information, and transfer the global information from the guiding feature H G to each query, and finally generate the optimized attention-enhanced feature H GA ; The process of generating the optimized attention-enhanced feature H GA is expressed as: where Q is the query matrix; Step4.2.

4. Add a residual connection to the optimized attention-enhanced feature H GA : H′ GA = H GA + H。

Images