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

Abstract

The application relates to a power transmission line fault diagnosis method based on time-frequency multi-level fusion and a bidirectional time sequence enhancement network and belongs to the field of power system power transmission line fault diagnosis. First, a time-frequency multi-level fusion and bidirectional time sequence enhancement network is constructed, including: converting input power transmission line partial discharge signal data into a time-frequency spectrum; constructing a multi-scale time-frequency feature joint extraction module to extract features of the time-frequency spectrum through a pair of offset branches; constructing a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features; constructing a bidirectional time sequence information enhancement module to model the fused features through a bidirectional long short-term memory network and adopt a guide attention mechanism to enhance feature expression of a key discharge moment; finally, the time-frequency multi-level fusion and bidirectional time sequence enhancement network is trained by using a training set, and finally, fault type recognition is realized through a full connection layer and a Softmax classifier. The application improves the accuracy and model robustness of power transmission line fault diagnosis.

Description

Technical Field

[0001] This invention relates to a method for diagnosing transmission line faults based on time-frequency multi-level fusion and bidirectional time-series enhancement networks, belonging to the field of power system transmission line fault diagnosis technology. Background Technology

[0002] As a critical component of the power system, the operating status of transmission lines directly affects the stability of the power grid and the security of power supply. However, due to their long-term exposure to the complex natural environment, transmission lines are susceptible to various external factors such as lightning strikes, pollution, tree contact, and wind deflection, leading to various faults, including corona discharge, surface discharge, floating discharge, and localized arc discharge. If these faults are not detected and handled promptly and accurately, they can cause serious power accidents or even large-scale power outages. Therefore, researching efficient and accurate fault diagnosis methods for transmission lines has significant engineering value and practical application implications.

[0003] Traditional methods for fault diagnosis in transmission lines mainly include those based on physical models, signal processing, and machine learning. Physical models rely on electromagnetic field and circuit theory, but due to the complexity of transmission line structures, it is difficult to establish accurate mathematical models. Signal processing methods (such as wavelet transform and short-time Fourier transform) can extract partial discharge features, but their performance is limited by the choice of preset basis functions, resulting in poor adaptability. In recent years, with the development of deep learning technology, fault diagnosis methods based on deep learning have received widespread attention. These methods can automatically extract high-dimensional features from data and have strong feature representation and generalization capabilities. However, existing deep learning methods still face certain challenges. On the one hand, partial discharge signals in transmission lines have complex time-frequency characteristics, making it difficult to comprehensively characterize fault information using only a single feature extraction method. On the other hand, traditional attention mechanisms have high computational complexity, making it difficult to efficiently capture key features in discharge signals. Furthermore, partial discharge signals often exhibit significant temporal dependencies; how to effectively model their temporal information based on feature extraction to improve the accuracy and robustness of fault diagnosis remains a problem to be solved. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a transmission line fault diagnosis method based on time-frequency multi-level fusion and bidirectional time-series enhancement network. This method integrates advanced technologies such as multi-scale time-frequency feature extraction, dynamic weighted fusion, time-series information modeling, and guided attention mechanism to efficiently mine the time-frequency features of partial discharge signals, thereby improving the accuracy and robustness of transmission line fault diagnosis.

[0005] The technical solution of this 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 identifying the fault type 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] Step 1: Use short-time Fourier transform to convert the input partial discharge signal data of the transmission line into a time spectrum diagram;

[0007] Step 2: Construct a multi-scale time-frequency feature joint extraction module, and use high-frequency, mid-frequency and low-frequency dual offset branches to extract time-frequency features from the time spectrogram;

[0008] Step 3: Construct a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features;

[0009] Step 4: Construct a bidirectional temporal information enhancement module. The temporal information is modeled by a bidirectional long short-term memory network (BiLSTM), and a guided attention mechanism is used to enhance the feature representation of key discharge moments.

[0010] Furthermore, in Step 1, the dataset acquisition method used when constructing the time-frequency multi-level fusion and bidirectional temporal enhancement network model is as follows:

[0011] Step 1: Collect a dataset of partial discharge signals from 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 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 into two parts, randomly selecting 80% as the training set and 20% as the test set.

[0014] Furthermore, in step 1, the partial discharge signal dataset of the transmission line includes six types of faults: corona discharge, surface discharge, internal discharge, floating discharge, crack discharge, and arc discharge; the sources of the partial discharge signal dataset of the transmission line include public datasets, databases, and partial discharge signal data of the transmission line collected from relevant literature.

[0015] Furthermore, in step 2, the wavelet transform separates the low-frequency useful signal from the high-frequency noise through multi-scale decomposition and uses a threshold method 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, according to the formula... The partial discharge signal data of the transmission line is normalized, where X is the original partial discharge signal data point of the transmission line.

[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 a Fourier transform is performed on each segment of the partial discharge signal data. Then, the spectrum of each time window is combined into a two-dimensional time spectrum 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, mid-frequency, and low-frequency dual offset branches. These branches employ 3×3, 5×5, and 7×7 convolutional kernels, respectively, to extract high-frequency, mid-frequency, and low-frequency features. Except for the different kernel sizes, the structures of the high-frequency, mid-frequency, and low-frequency dual offset branches remain consistent. The specific operations involved in each branch are as follows:

[0019] The time-frequency spectrum X generated in Step 1 m First, a preliminary feature extraction is performed using a 3×3 / 5×5 / 7×7 convolutional layer; then, a sigmoid activation function is used to smooth the feature mapping; finally, the smoothed features are subtracted and incremented by one to obtain the inverse offset feature X. r and positive offset feature X f ;Then the reverse offset feature X r and positive offset feature X f Compared with the time spectrum diagram X respectively m Matrix multiplication is performed, and the results are summed. Finally, the time-frequency features are obtained by sequentially passing the data through an average pooling layer, a 1×1 convolutional layer, and a normalization layer. 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 of the time-frequency features of different scale branches; the specific steps include:

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

[0022]

[0023] Where i represents the i-th scale branch, G i ∈R CThe global feature representing the i-th scale branch;

[0024] Step 3.2: Calculate the weight ω of the global feature of the i-th scale branch using two fully connected layers, ReLU activation function, and Sigmoid activation function. i The weights ω of the global features of the i-th scale branch are... i After normalization, the normalized values ​​α of the corresponding weights are obtained. i ;

[0025] ω i =σ(W2δ(W1G) i ))

[0026]

[0027] Where W1 and W2 are fully connected layers with trainable parameters, δ(·) is the ReLU activation function, and σ(·) is the Sigmoid activation function, used to normalize the weights to [0,1].

[0028] Step 3.3: Calculate the fusion features through weighted fusion.

[0029] Further, Step 4 includes: using a bidirectional long short-term memory network (BiLSTM) to capture the temporal dependencies of partial discharge signal data from transmission lines, and combining this with an improved guided attention mechanism to focus on key temporal information in the partial discharge signal data from transmission lines; the specific steps include:

[0030] Step 4.1: Bidirectional Long Short-Term Memory (BiLSTM) network combines forward and backward long short-term memory networks to propagate information bidirectionally, thus more comprehensively exploring the temporal dependencies of partial discharge signal data in transmission lines.

[0031] Step 4.2: Based on the temporal features H output by the Bidirectional Long Short-Term Memory (BiLSTM) network, a guided attention mechanism is introduced to calculate the attention weights at each time step, making the key time steps contribute more to fault classification. The guided attention mechanism includes two self-attention stages, where:

[0032] Phase 1: Introduce a bootstrap token G, which is generated by pooling the query matrix Q and serves as the global representation of the original query. Use this bootstrap token G to replace the original query matrix Q, and perform feature aggregation on the key matrix K and value matrix V to generate the bootstrap feature H. G ;

[0033] Phase Two: Using the bootstrap token G as the key, bootstrap feature H GAs a value, global information is broadcast, passing global information from the guiding feature to each query, ultimately generating the optimized attention-enhanced feature H. GA .

[0034] Further, Step 4.1 includes:

[0035] Step 4.1.1: Given the fusion 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 features, with each time step corresponding to a d-dimensional feature vector;

[0036] The feedforward long short-term memory network processes the fused feature X sequentially from t=1 to t=T. F Calculate the forward hidden state

[0037]

[0038] in, This represents the forward hidden state at time t. This represents the hidden state at the previous time step, i.e., the forward hidden state at time t-1, combined with the current input x. t Calculate the state at the current time step; x t This represents the current input, i.e., 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] 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 Calculate the backward hidden state

[0040]

[0041] in, Depends on the backward hidden state at time t+1 and fusion feature X F The fused feature vector x at time t t ;

[0042] Step 4.1.3, Forward Hidden State at Assembly Time t Backward hidden state at time t Obtain the final bidirectional hidden state h t :

[0043]

[0044] Among them, h t It is the final two-way hidden state, containing information about the past and the future;

[0045] Step 4.1.4: The final set of the bidirectional long short-term memory network contains the bidirectional hidden state output temporal features H for all time steps; the temporal features H are represented as:

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

[0047] Further, Step 4.2 includes:

[0048] Step 4.2.1: First, perform a linear transformation on the time-series feature H to obtain the query matrix Q, key matrix K, and value matrix V:

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

[0050] Among them, W Q W K W V It is a trainable parameter matrix;

[0051] Step 4.2.2: Perform a global digest of the query matrix Q using global average pooling to obtain the bootstrap token G.

[0052] G = GAP(Q)

[0053] Wherein, GAP(·) represents the global average pooling operation, which is used to guide token G as a global representation of the entire time series features, and can capture global feature information;

[0054] Step 4.2.2: Replace the original query matrix Q with the guiding token G, and perform feature aggregation on the guiding key matrix K and value matrix V to generate the guiding feature H. G Generate guiding feature H G The process is represented as:

[0055]

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

[0057] Step 4.2.3: Using the bootstrap token G as the key and the bootstrap feature H... G As a value, broadcast global information from the guiding feature H G Global information is passed to each query, ultimately generating the optimized attention-enhanced feature H. GA Generate optimized attention-enhanced features H GA The process is represented as:

[0058]

[0059] Where Q is the query matrix;

[0060] Step 4.2.4: Optimize the attention enhancement feature H GA Add residual join to:

[0061] H′ GA =H GA +H.

[0062] The beneficial effects of this invention are:

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

[0064] 2. This invention proposes a time-frequency multi-level fusion mechanism, which combines the short-time dynamic characteristics of time-domain signals with the global spectral characteristics of frequency-domain signals to achieve more comprehensive fault feature extraction. Compared with single-domain feature extraction methods, it improves the accuracy of fault diagnosis.

[0065] 3. This invention introduces a BiLSTM architecture, which fully captures the sequential dependencies of discharge signals, effectively models the temporal characteristics of fault signals, enhances the separability of fault modes, and improves the stability of the diagnostic model.

[0066] 4. This invention proposes a guided attention mechanism, which can effectively reduce the computational complexity of the traditional self-attention mechanism, while dynamically allocating the weights of different features, strengthening the contribution of key fault features to classification decisions, thereby improving the model's discriminative ability.

[0067] 5. This invention improves the feature representation capability, effectively enhancing the accuracy, stability, and adaptability of transmission line fault diagnosis, and has broad engineering application value. Attached Figure Description

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

[0069] Figure 2This is a schematic diagram of the structure of the time-frequency multi-level fusion and bidirectional temporal enhancement network in this invention;

[0070] Figure 3 This is a schematic diagram of the dynamic weighted fusion module in this invention;

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

[0072] Figure 5 This is a schematic diagram of the attention guidance mechanism in this invention. Detailed Implementation

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

[0074] (1) Data collection: Collect partial discharge signal datasets of transmission lines with labeled types; the partial discharge signal datasets of transmission lines include six types of faults: corona discharge, surface discharge, internal discharge, floating discharge, crack discharge and arc discharge; the dataset sources include public datasets and partial discharge signal data of transmission lines collected from databases and related literature.

[0075] (2) Data preprocessing: Wavelet transform is used to denoise the partial discharge signal data of the transmission line, and Z-score normalization method is used to normalize the partial discharge signal data of the transmission line.

[0076] In step 2, the wavelet transform separates the low-frequency useful signal from the high-frequency noise through multi-scale decomposition and uses a thresholding method to suppress noise. Through thresholding, high-frequency noise is removed at each scale while retaining the useful information in the signal. 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 partial discharge signal data of the transmission line is normalized, 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 temporal series enhancement network model, and train it using the training set. Extract feature representations from the data, and adjust the importance of 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 temporal series enhancement network to learn discriminative features, referring to... Figure 2 The time-frequency multi-level fusion and bidirectional temporal enhancement network model includes a multi-scale time-frequency joint feature extraction module, a dynamic weighted fusion module, and a bidirectional temporal information enhancement module; the model output passes through a fully connected layer and is classified for faults using the Softmax function; the specific steps for constructing the time-frequency multi-level fusion and bidirectional temporal enhancement network model include:

[0079] Step 1: Use Short-Time Fourier Transform (STFT) to convert the input partial discharge signal data of the transmission line into a time-spectrum graph X. m ;

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

[0081] First, a sliding window is selected to segment the partial discharge signal data of the transmission line, and a Fourier transform is performed on each segment of the partial discharge signal data. The Fourier transform result of each segment will provide the frequency information of the signal. Then, the spectrum of each time window is combined into a two-dimensional time spectrum to show the dynamic changes of the signal in the time and frequency domains.

[0082] The partial discharge signal of the transmission line is converted into a corresponding time-frequency spectrum using STFT, and the fault type is determined by fully exploiting its time-frequency characteristics, effectively improving the accuracy of fault diagnosis. The fault type determination method is shown in the table below:

[0083]

[0084] Step 2: Construct a multi-scale time-frequency feature joint extraction module, using high-frequency, mid-frequency, and low-frequency dual offset branches to extract time-frequency spectrograms X. m Perform time-frequency feature extraction;

[0085] Furthermore, in Step 2, the multi-scale time-frequency feature joint extraction module includes three parallel branches: high-frequency, mid-frequency, and low-frequency dual offset branches. These branches employ 3×3, 5×5, and 7×7 convolutional kernels, respectively, to extract high-frequency, mid-frequency, and low-frequency features. Except for the different kernel sizes, the structures of the high-frequency, mid-frequency, and low-frequency dual offset branches remain consistent. The specific operations involved in each branch are as follows:

[0086] The time-frequency spectrum X generated in Step 1m First, a preliminary feature extraction is performed using a 3×3 / 5×5 / 7×7 convolutional layer. Then, a sigmoid activation function is used to smooth the feature mapping. The sigmoid activation function maps feature values ​​to a fixed range, making the feature distribution smoother, preventing the influence of extreme values, and improving the model's sensitivity to low-amplitude discharge patterns. Next, the smoothed features are subtracted and incremented by one to obtain the inverse offset feature X. r and positive offset feature X f This effectively expands the representation range of features and enhances feature expressive power, enabling the model to capture richer time-frequency features in partial discharge signals and improve classification accuracy and robustness; subsequently, the reverse-shifted feature X... r and positive offset feature X f Compared with the time spectrum diagram X respectively m Performing matrix multiplication and summing the results not only preserves the original distribution information of the features but also enhances their nonlinear expressive power, while introducing new feature information, thus achieving information enhancement. Finally, the time-frequency features are obtained by sequentially passing them through an average pooling layer, a 1×1 convolutional layer, and a normalization layer. Where i = 1, 2, 3 represents the i-th scale branch.

[0087] Step 3: Construct a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features;

[0088] Furthermore, in Step 3, the dynamic weighted fusion module uses channel attention to weight and fuse the time-frequency features of different scale branches; based on the characteristics of different fault signals, it adaptively adjusts the weights of high, medium, and low-frequency time-frequency features to improve the flexibility of feature fusion and enhance the fusion effect; the specific steps include:

[0089] Step 3.1: The dynamic weighted fusion module receives three features from the multi-scale time-frequency feature joint extraction module and processes the input time-frequency features. Global average pooling (GAP) is used to extract global information from time-frequency features representing high, mid, and low frequencies, respectively; where C is the number of channels, and H and W are the spatial dimensions of the feature map.

[0090]

[0091] Where i represents the i-th scale branch, G i ∈R C The global feature representing the i-th scale branch;

[0092] Step 3.2: Calculate the weight ω of the global feature of the i-th scale branch using two fully connected layers, ReLU activation function, and Sigmoid activation function.i The weights ω of the global features of the i-th scale branch are... i After normalization, the normalized values ​​α of the corresponding weights are obtained. i ;

[0093] ω i =σ(W2δ(W1G) i ))

[0094]

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

[0096] Step 3.3: Calculate the fusion features through weighted fusion. The dynamic weighted fusion method ensures that the network can adaptively focus on the most discriminative features under different fault modes.

[0097] Step 4: Construct a bidirectional temporal information enhancement module. Model the temporal information using a bidirectional long short-term memory (BiLSTM) network and employ a guided attention mechanism to enhance the feature representation of key discharge moments.

[0098] Furthermore, Step 4 includes: using a bidirectional long short-term memory network (BiLSTM) to capture the temporal dependencies of partial discharge signal data of transmission lines, modeling the temporal dependencies of partial discharge signals, enhancing the expressive power of fault features, and combining with an improved guided attention mechanism to focus on key temporal information in the partial discharge signal data of transmission lines, which can effectively reduce the computational complexity of traditional self-attention mechanisms, strengthen the contribution of key temporal information to classification decisions, and improve the model's discriminative ability;

[0099] Bidirectional Long Short-Term Memory (BiLSTM) is an extension of the traditional Long Short-Term Memory (LSTM) network. It includes forward LSTM and backward LSTM to capture the dependencies between time series data.

[0100] BiLSTM is an extension of the traditional LSTM. The core advantage of the Bidirectional Long Short-Term Memory Network (BiLSTM) is that it can capture both past and future temporal information 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: Bidirectional Long Short-Term Memory Network (BiLSTM) combines forward LSTM and backward LSTM to propagate information bidirectionally, thus more comprehensively mining the temporal dependencies of partial discharge signal data in transmission lines and improving the accuracy of fault classification.

[0103] Step 4.2: Based on the temporal features H output by the Bidirectional Long Short-Term Memory (BiLSTM) network, a guided attention mechanism is introduced to calculate the attention weights at each time step, making the key time steps contribute more to fault classification. The guided attention mechanism includes two self-attention stages, using a global guided token G to extract key temporal features and enhance the global information propagation capability, thereby optimizing the final attention-enhanced feature H. GA ;

[0104] The guided attention mechanism, by introducing a global guiding token, transforms the pairwise interactions of the traditional self-attention mechanism across all time steps into two low-dimensional attention stages, significantly reducing computational complexity; compared to the traditional self-attention mechanism's O(T)... 2 The computational cost of d) is only O(nTd) (where n << T), which effectively reduces the computational cost and memory usage, making it more efficient in long sequence fault signal processing. In addition, while reducing computational complexity, the guided attention mechanism can retain key time step information and enhance feature expression capabilities through global guidance, thereby improving the computational efficiency and training stability of the model while ensuring 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: a global guided token G is used to extract key temporal features and enhance the global information propagation capability, thereby optimizing the final attention-enhanced feature H. GA ;

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

[0107] Phase 1: Introduce a bootstrap token G, which is generated by pooling the query matrix Q and serves as the global representation of the original query. Use this bootstrap token G to replace the original query matrix Q, and perform feature aggregation on the key matrix K and value matrix V to generate the bootstrap feature H. G ;

[0108] Phase Two: Using the bootstrap token G as the key, bootstrap feature H G As a value, global information is broadcast, passing global information from the guiding feature to each query, ultimately generating the optimized attention-enhanced feature H. GA .

[0109] Further, Step 4.1 includes:

[0110] Step 4.1.1: The Bidirectional Long Short-Term Memory (BiLSTM) network receives the fused feature X output by the dynamically weighted fusion module. F ∈R T×d 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 features, with each time step corresponding to a d-dimensional feature vector;

[0111] The feedforward long short-term memory network processes the fused feature X sequentially from t=1 to t=T. F Calculate the forward hidden state

[0112]

[0113] in, This represents the forward hidden state at time t. This represents the hidden state at the previous time step, i.e., the forward hidden state at time t-1, combined with the current input x. t Calculate the state at the current time step; x t This represents the current input, i.e., 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 ;

[0114] 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 Calculate the backward hidden state

[0115]

[0116] in, Depends on the backward hidden state at time t+1 and fusion feature X F The fused feature vector x at time t t ;

[0117] Step 4.1.3, Forward Hidden State at Assembly Time t Backward hidden state at time t Obtain the final bidirectional hidden state h t :

[0118]

[0119] Among them, h t It is the final two-way hidden state, containing information about the past and the future;

[0120] Step 4.1.4: The final set of the bidirectional long short-term memory network contains the bidirectional hidden state output temporal features H for all time steps; the temporal features H are represented as:

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

[0122] Further, Step 4.2 includes:

[0123] Step 4.2.1: First, perform a linear transformation on the time-series feature H to obtain the query matrix Q, key matrix K, and value matrix V:

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

[0125] Among them, W Q W K W V It is a trainable parameter matrix;

[0126] Step 4.2.2: Perform a global digest of the query matrix Q using global average pooling to obtain the bootstrap token G.

[0127] G = GAP(Q)

[0128] Wherein, GAP(·) represents the global average pooling operation, which is used to guide token G as a global representation of the entire time series features, and can capture global feature information;

[0129] Step 4.2.2: Replace the original query matrix Q with the guiding token G, and perform feature aggregation on the guiding key matrix K and value matrix V to generate the guiding feature H. G Generate guiding feature H G The process is represented as:

[0130]

[0131] Where Softmax(·) represents the softmax activation function, d k Let be the dimension of the key matrix;

[0132] Step 4.2.3: Using the bootstrap token G as the key and the bootstrap feature H... G As a value, broadcast global information from the guiding feature H G Global information is passed to each query, ultimately generating the optimized attention-enhanced feature H. GA Generate optimized attention-enhanced features H GA The process is represented as:

[0133]

[0134] Where Q is the query matrix;

[0135] Step 4.2.4: Optimize the attention enhancement feature H GA Add residual join to:

[0136] H′ GA =H GA +H.

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

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

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

[0140] Among them, W Q W K W V It is a trainable parameter matrix;

[0141] Calculate the attention score matrix:

[0142]

[0143] Where, d kLet be the dimension of the key matrix;

[0144] Softmax normalization is applied to the attention score matrix A:

[0145] A′=Softmax(A)

[0146] Calculate aggregate features:

[0147] H′=A′V

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

[0149] After the second stage of computation, the final attention-enhanced feature H is generated. GA By fusing residual connections with the original input features, key information can be preserved and stability enhanced.

[0150] This invention first constructs a partial discharge signal dataset for transmission lines containing various labeled fault types through data acquisition, and then performs noise reduction and normalization processing on the original signals. Subsequently, a deep learning model based on a time-frequency multi-level fusion mechanism and a bidirectional temporal enhancement network is designed, and the network parameters are optimized using the training dataset. Finally, the trained model achieves intelligent classification of typical fault types such as corona discharge, surface discharge, floating discharge, and partial arc discharge. This invention innovatively integrates the feature extraction capabilities of Convolutional Neural Networks (CNNs), the bidirectional temporal modeling advantages of Bidirectional Long Short-Term Memory (BiLSTM) networks, and the key feature dynamic focusing capabilities of attention mechanisms, effectively improving the accuracy and robustness of transmission line fault diagnosis, and has significant engineering application value in the field of power system condition monitoring and fault early warning.

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

Claims

1. A method for fault diagnosis of transmission lines 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 temporal enhancement network model, training the model using a training set, and finally achieving fault type identification through a fully connected layer and a Softmax classifier; the specific construction process of the time-frequency multi-level fusion and bidirectional temporal enhancement network model includes: Step 1: Use short-time Fourier transform to convert the input partial discharge signal data of the transmission line into a time spectrum diagram; Step 2: Construct a multi-scale time-frequency feature joint extraction module, and use high-frequency, mid-frequency and low-frequency dual offset branches to extract time-frequency features from the time spectrogram; Step 3: Construct a dynamic weighted fusion module to adaptively fuse the extracted time-frequency features; Step 4: Construct a bidirectional temporal information enhancement module. The temporal information is modeled through a bidirectional long short-term memory network (BiLSTM), and a guided attention mechanism is used to enhance the feature representation of key discharge moments. In Step 2, the multi-scale time-frequency feature joint extraction module includes three parallel branches: high-frequency, mid-frequency, and low-frequency dual offset branches. These branches employ 3×3, 5×5, and 7×7 convolutional kernels, respectively, to extract high-frequency, mid-frequency, and low-frequency features. Except for the different kernel sizes, the structures of the high-frequency, mid-frequency, and low-frequency dual offset branches remain consistent. The specific operations involved in each branch are as follows: The time spectrum generated in Step 1 First, a preliminary feature extraction is performed using a 3×3 / 5×5 / 7×7 convolutional layer; then, the feature mapping is smoothed using a sigmoid activation function; finally, the smoothed features are subtracted and incremented by one to obtain the inverse offset features. and positive offset features Then the reverse offset feature will be used. and positive offset features Compared with the time spectrum diagram respectively Matrix multiplication is performed, and the results are summed. Finally, the time-frequency features are obtained by sequentially passing the data through an average pooling layer, a 1×1 convolutional layer, and a normalization layer. ,in , indicating the first One scale branch; Step 4 includes: using a bidirectional long short-term memory (BiLSTM) network to capture the temporal dependencies of partial discharge signal data from transmission lines, and combining this with an improved guided attention mechanism to focus on key temporal information in the partial discharge signal data; specific steps include: Step 4.1: Bidirectional Long Short-Term Memory (BiLSTM) network combines forward and backward long short-term memory networks to propagate information bidirectionally, thus more comprehensively exploring the temporal dependencies of partial discharge signal data in transmission lines. Step 4.2: Based on the temporal features H output by the Bidirectional Long Short-Term Memory (BiLSTM) network, a guided attention mechanism is introduced to calculate the attention weights at each time step, making the key time steps contribute more to fault classification. The guided attention mechanism includes two self-attention stages, where: Phase 1: Introduce a bootstrap token G, which is generated by pooling the query matrix Q and serves as the global representation of the original query. Use this bootstrap token G to replace the original query matrix Q, and perform feature aggregation on the key matrix K and value matrix V to generate bootstrap features. ; Phase Two: Using the bootstrap token G as the key, bootstrap features are established. As a value, global information is broadcast, passing global information from the guiding features to each query, ultimately generating optimized attention-enhanced features. .

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 Step 1, the dataset acquisition method used when constructing the time-frequency multi-level fusion and bidirectional temporal enhancement network model is as follows: Step 1: Collect a dataset of partial discharge signals from transmission lines with labeled fault types; Step 2: Use wavelet transform to denoise the partial discharge signal data of the transmission line, and use 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 into two parts, randomly selecting 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, characterized in that: In step 1, the partial discharge signal dataset of the transmission line includes six types of faults: corona discharge, surface discharge, internal discharge, floating discharge, crack discharge, and arc discharge. The sources of the partial discharge signal dataset of the transmission line include public datasets, databases, and partial discharge signal data of the transmission line collected from relevant literature.

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 step 2, the wavelet transform separates the low-frequency useful signal from the high-frequency noise through multi-scale decomposition and uses a threshold method to suppress the noise; the Z-score normalization method calculates the mean of the partial discharge signal data of the transmission line. and standard deviation According to the formula The partial discharge signal data of the transmission line is normalized, among which, These are the partial discharge signal data points of the original transmission line.

5. 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 Step 1, the short-time Fourier transform is specifically implemented as follows: First, a sliding window is selected to segment the partial discharge signal data of the transmission line, and a Fourier transform is performed on each segment of the partial discharge signal data. Then, the spectrum of each time window is combined into a two-dimensional time spectrum 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, characterized in that: In Step 3, the dynamic weighted fusion module uses channel attention to perform weighted fusion of time-frequency features of different scale branches; The specific steps include: Step 3.1: Analyze the time-frequency characteristics of the input. Global average pooling is used to extract global information; where C is the number of channels, and H and W are the spatial dimensions of the feature map. ； in, Indicates the first One scale branch, Representing the Global features of each scale branch; Step 3.2: Calculate the first activation function after two fully connected layers, ReLU activation function, and Sigmoid activation function. Weights of global features in each scale branch , will the Weights of global features in each scale branch After normalization, the normalized values ​​of the corresponding weights are obtained. ; ； ； in, and It is a fully connected layer with trainable parameters. It is the ReLU activation function. It is the Sigmoid activation function, used to normalize the weights to [0, 1]. 1]; Step 3.3: Calculate the fusion features through weighted fusion. .

7. 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: Step 4.1 includes: Step 4.1.1: Given fusion features , where T is the time step and d is the dimension of the fused features, with each time step corresponding to a d-dimensional feature vector; The feedforward long short-term memory network processes fused features sequentially from t=1 to t=T. Calculate the forward hidden state : ； in, This represents the forward hidden state at time t. This represents the hidden state of the previous time step, i.e. The forward hidden state at each moment, combined with the current input Calculate the state at the current time step; This represents the current input, i.e., the fused features. exist The fused feature vector at time step Depends on the hidden state of the previous time step and current input ; Step 4.1.2: The backward long short-term memory network processes the fused features in the reverse order, from t=T to t=1. exist Fusion feature vector at time step Calculate the backward hidden state : ； in, Depends on Backward hidden state at time and fusion features exist Fusion feature vector at time step ; Step 4.1.3, Forward Hidden State at Assembly Time t Backward hidden state at time t Obtain the final bidirectional hidden state. : ； in, It is the final two-way hidden state, containing information about the past and the future; Step 4.1.4: The final set of the bidirectional long short-term memory network contains the bidirectional hidden state output temporal features H for all time steps; the temporal features H are represented as: 。 8. 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: Step 4.2 includes: Step 4.2.1: First, perform a linear transformation on the time-series feature H to obtain the query matrix Q, key matrix K, and value matrix V: ； in, , , It is a trainable parameter matrix; Step 4.2.2: Perform a global digest of the query matrix Q using global average pooling to obtain the bootstrap token G. ； in, This represents a global average pooling operation, used to guide token G as a global representation of the entire time series features, capable of capturing global feature information; Step 4.2.3: Replace the original query matrix Q with the guiding token G, and perform feature aggregation on the guiding key matrix K and value matrix V to generate guiding features. Generate guiding features The process is represented as: ； in, This represents the softmax activation function. Let be the dimension of the key matrix; Step 4.2.4: Using the bootstrap token G as the key, and the bootstrap feature... As a value, broadcast global information, from guiding features Global information is passed to each query, ultimately generating optimized attention-enhanced features. Generate optimized attention-enhanced features The process is represented as: ； Where Q is the query matrix; Step 4.2.5: Optimized attention enhancement features Add residual join to: 。

Images