Gating bidirectional convolutional recurrent network-based fault diagnosis method for excitation system

Abstract

The application discloses a kind of excitation system fault diagnosis methods based on gated bidirectional convolution recurrent network, and the excitation recording wave signal data of synchronous generator excitation system under different operating conditions is collected, and signal is preprocessed to obtain standardized excitation timing sample;Standardized excitation timing sample is input into double-flow parallel feature extraction network, and local transient feature and long-range timing feature are extracted respectively;Two kinds of features are respectively input into gated attention fusion module through independent bypass to carry out adaptive weighted fusion, and obtain fusion fault feature;Finally, classification output module is input to obtain excitation system fault class diagnosis result.The application solves the problem that transient feature and long-range timing feature are mixed and complex in excitation system recording wave signal, which makes it difficult to capture fault features comprehensively, reduces feature attenuation caused by serial transmission and feature redundancy caused by simple splicing, effectively reduces information dilution and feature distortion, improves the accuracy and stability of excitation system fault identification under complex conditions.

Description

Technical Field

[0001] This invention relates to the field of excitation system fault diagnosis technology, specifically to a method for extracting time-series signal features and identifying faults based on deep learning, and more particularly to a method for excitation system fault diagnosis based on a gated bidirectional convolutional recurrent network. Background Technology

[0002] The excitation system of a synchronous generator is a crucial control unit in a power system, and its operating status directly affects generator terminal voltage regulation, reactive power distribution, and system stability. In actual operation, faults such as thyristor breakdown, winding short circuits, firing angle deviation, trigger pulse loss, and control circuit abnormalities can lead to excitation system failure, weak excitation, or abnormal output, thereby affecting the safe and stable operation of the generator and the power system. Therefore, accurately identifying excitation system fault types has significant engineering value.

[0003] Excitation system fault signals typically contain both local transient anomalies and long-term temporal evolution features. While existing threshold comparison methods, spectral analysis methods, and model-driven methods can achieve fault diagnosis to some extent, they suffer from insufficient ability to recognize complex waveform patterns, difficulty in capturing fault evolution across time scales, and strong dependence on prior models. Existing deep learning-based methods typically employ a single neural network or cascade and fuse convolutional and recurrent networks. While these methods can extract certain local or temporal features, they still have shortcomings: Firstly, transient anomalies such as short-duration spikes, high-frequency glitches, and local abrupt changes in excitation waveform signals have short durations and are easily smoothed or attenuated after serial temporal modeling. Secondly, faults such as trigger angle offset and trigger pulse loss exhibit cross-cycle evolution characteristics, which simple local convolution cannot fully represent. Furthermore, direct splicing or fixed-weighted fusion makes it difficult to dynamically determine the contribution ratio of local transient features and long-term temporal features based on different fault samples, easily leading to feature redundancy or insufficient contribution of key fault information.

[0004] Therefore, it is necessary to propose a feature extraction and fusion method for excitation system fault diagnosis. While maintaining the independent expression of local transient features and long-range time-series features, the method can achieve adaptive fusion of the two types of features, thereby reducing information dilution and feature distortion during feature transmission and improving the accuracy and stability of multi-category fault diagnosis of excitation systems under complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to propose a fault diagnosis method for excitation systems based on gated bidirectional convolutional recurrent networks, in order to solve the problems of difficulty in co-extracting local transient features and long-range temporal features in traditional excitation system fault diagnosis, easy information dilution during feature transfer, and difficulty in adaptively allocating the contributions of different features, thereby improving the accuracy and stability of excitation system fault identification under complex operating conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A fault diagnosis method for an excitation system based on a gated bidirectional convolutional recurrent network includes the following steps:

[0008] S1: Collect multi-channel excitation waveform data of synchronous generator excitation system under different operating conditions, and preprocess it to obtain standardized excitation timing samples.

[0009] S2, the standardized excitation timing sample is input into a dual-stream parallel feature extraction network, which includes a mutually independent CNN transient feature branch and a BiLSTM temporal feature branch to extract local transient features and long-range temporal features respectively;

[0010] S3, the local transient features and the long-term temporal features are respectively input into the gated attention fusion module through independent bypasses, so that the two types of features do not undergo serial transmission to each other before entering the gated attention fusion module;

[0011] S4, the gated attention fusion module generates branch-level gated weights corresponding to the CNN transient feature branch and the BiLSTM temporal feature branch based on the joint representation of local transient features and long-term temporal features, and performs adaptive weighted fusion to obtain fused fault features;

[0012] S5 integrates fault characteristics into the classification output module to obtain the fault category diagnosis results of the excitation system.

[0013] A further technical solution is that the multi-channel excitation waveform signal in S1 includes at least one or more of the following: generator three-phase voltage, generator current and excitation current, generator terminal voltage, and state variables related to excitation trigger control, expressed as:

[0014]

[0015] In the formula, This indicates the sampling sequence length of a single excitation waveform signal. Indicates the number of signal channels. Indicates the first The multi-channel excitation state quantity corresponding to each sampling time.

[0016] Furthermore, the multi-channel excitation recording signal described in S1 is normalized by mean-standard deviation for each channel and divided into time-series samples according to a preset sampling length; sample labels are constructed based on at least one of the following operating conditions: normal operation, thyristor short circuit, thyristor open circuit, abnormal firing angle, and firing pulse loss.

[0017]

[0018] In the formula, Represents the standardized eigenvalues ​​after normalization. Represents the original signal value. Indicates the first The average of each channel, Indicates the first The standard deviation of each channel.

[0019] Furthermore, the preprocessing described in S1 is not merely a numerical normalization of the original signal, but rather, in conjunction with the operating mechanism of the excitation system, it performs channel unification, time window division, normalization processing, and fault label construction on different recorded waveform samples. Channel unification ensures that different recorded waveform samples have a consistent input channel structure; time window division allows excitation recorded waveform signals of different lengths to form time-series samples with uniform sample lengths; normalization processing reduces the impact of differences in the dimensions and amplitudes of different channels on model training; and sample label construction matches standardized excitation time-series samples with corresponding fault categories, providing a unified input for subsequent dual-stream parallel feature extraction and fault classification diagnosis.

[0020] A further technical solution is that the transient feature extraction branch described in S2 is a one-dimensional convolutional neural network branch. Compared with the single-scale convolution method, this method adopts a multi-scale one-dimensional convolution structure, which can simultaneously extract local abnormal features of different durations in the excitation waveform signal through convolution kernels of different sizes. Specifically, small-scale convolution kernels focus on capturing transient abrupt changes such as short-term spikes, narrow pulses, and high-frequency glitches, while larger-scale convolution kernels focus on capturing local waveform distortions, amplitude changes, and non-stationary segments with longer durations. Since the excitation system fault signal may simultaneously contain short-term abrupt changes and local persistent distortions, multi-scale convolution can reduce the feature omission problem caused by a fixed receptive field, improve the completeness of transient feature extraction, and enhance adaptability to complex fault waveforms. Multi-scale convolution is represented as:

[0021]

[0022] In the formula, Indicates the first The convolutional kernel of scale at the th scale in the th... Features output at each position The kernel size is the convolution kernel size. As weight, For bias terms, It is a non-linear activation function. The input excitation timing sample corresponds to the sampled feature values. Features output from convolutional kernels of different scales are concatenated along the channel dimension and pooled to obtain a local transient feature vector. .

[0023] A further technical solution is that the time-series evolution feature extraction branch mentioned in S2 is a bidirectional long short-term memory network branch, which is used to extract the long-range time-series features of the excitation recording signal in order to capture the slow evolution, cumulative changes and cross-time correlation of the fault state.

[0024] Furthermore, the bidirectional long short-term memory network branch does not rely solely on historical states in a single time direction for feature transfer, but simultaneously models the excitation waveform sequence from both forward and reverse time directions. For excitation system fault signals, some abnormal states are not only determined by waveform changes before the fault occurs, but may also be related to the recovery process after the fault occurs, the duration of waveform distortion, and state changes between adjacent periods. For example, faults such as trigger angle offset, trigger pulse loss, and thyristor open circuit may manifest as phase changes, amplitude drift, or continuous waveform distortion across periods. When extracting features only from a single time direction, the correlation between the current sampling point and subsequent waveforms is easily overlooked. By introducing a bidirectional temporal modeling structure, the model can simultaneously utilize contextual information before and after the current moment to more completely depict the dynamic process of excitation faults from generation and development to waveform response changes, thereby improving the ability to distinguish between similar fault categories and slowly evolving faults.

[0025] Furthermore, in the bidirectional long short-term memory network branch, its gating unit is calculated as follows:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] In the formula, Represents the Gate of Oblivion Indicates the input gate. Indicates the output gate. Indicates the current cell state. Indicates the state of candidate cells. Indicates the current hidden state. This is the weight matrix. For bias terms, This represents the Hadamard product. The forward and backward hidden states of the BiLSTM branch are concatenated as follows: Through feature aggregation function Obtain long-range time series feature vectors .

[0033] A further technical solution involves a dual-stream parallel structure composed of the CNN transient feature branch and the BiLSTM temporal feature branch described in S2. These branches receive the same standardized excitation timing sample as input and operate independently during the feature extraction stage. The output of the CNN transient feature branch is not used as the input to the BiLSTM temporal feature branch, and vice versa. The CNN branch specifically extracts local abrupt changes, short-term spikes, and high-frequency distortions caused by thyristor breakdown, open circuit, or triggering anomalies. The BiLSTM branch specifically extracts cross-cycle changes and slow drifts caused by trigger angle offset, trigger pulse loss, or control anomalies. This structure solves the problem that a single network cannot simultaneously handle both local transient anomalies and long-term temporal evolution.

[0034] A further technical solution is that the independent bypass in S3 includes a first bypass connecting the output of the CNN transient feature branch to the gated attention fusion module, and a second bypass connecting the output of the BiLSTM temporal feature branch to the gated attention fusion module; let the local transient features output by the CNN transient feature branch be... The long-range time series features output by the BiLSTM time series feature branch are: The features of the gated attention fusion module input via the first and second bypasses can be expressed as:

[0035]

[0036]

[0037] in, and These represent feature transfer or dimension alignment mapping in the first and second bypasses, respectively; when the feature dimensions of the two paths are consistent, the mapping is an identity mapping. Through the above independent bypasses, local transient features and long-range temporal features do not undergo mutual serial transfer or cascaded progressive processing before entering the gated attention fusion module.

[0038] A further technical solution is that the gated attention fusion module described in S4 concatenates the two bypass input features to obtain joint state features:

[0039]

[0040] In the formula, This represents the joint state representation. Indicates local transient characteristics, This represents long-range time series characteristics.

[0041] The joint state features are input into a learnable gating network to generate gating scores for the corresponding CNN transient feature branch and BiLSTM temporal feature branch. The gating scores are then normalized to obtain the first gating weights corresponding to the CNN transient feature branch and the second gating weights corresponding to the BiLSTM temporal feature branch.

[0042]

[0043] In the formula, The weights represent the weights corresponding to local transient features. This represents the weights corresponding to long-range time-series features.

[0044] The two features are weighted and fused based on the first and second gating weights to obtain the fused fault features:

[0045]

[0046] In the formula, This indicates the characteristics of a fusion failure.

[0047] Furthermore, this invention retains the original discriminative information of the two deep feature paths through independent bypasses, avoiding the smoothing, dilution, or coverage of transient fault features such as local abrupt changes, short-term spikes, and high-frequency distortions by long-range temporal features during multi-layer transmission; the gated attention module dynamically adjusts the contribution ratio of the two feature paths according to the feature performance of different fault samples. This technical solution solves the problems of information dilution, feature redundancy, and difficulty in dynamically allocating the contributions of different features in existing fusion methods.

[0048] A further technical solution involves integrating the fault characteristics in S4. The input classification and output module performs global average pooling and fully connected mapping to obtain the excitation system fault category diagnosis results:

[0049]

[0050] In the formula, and These are the weight matrix and bias term of the classification layer, respectively. GAP(⋅) represents the global average pooling operation on the feature vector in the spatial or channel dimension. This represents the multi-class fault diagnosis results output by the network, which are normalized using the Softmax function to obtain the probability of each class.

[0051] A fault diagnosis system for an excitation system based on a gated bidirectional convolutional recurrent network, comprising:

[0052] The data acquisition and processing module is used to acquire multi-channel excitation waveform signal data and preprocess the signals to obtain standardized excitation timing samples.

[0053] The dual-stream feature extraction module includes a CNN transient feature branch and a BiLSTM temporal feature branch, used to extract local transient features respectively. and long-range time series characteristics ;

[0054] The bypass fidelity module includes a first bypass and a second bypass, which are used to directly transmit local transient features and long-term temporal features to the gated attention fusion module, respectively.

[0055] The gated attention fusion module is used to generate branch-level gating weights based on the joint representation of local transient features and long-term temporal features, and to adaptively weight and fuse the two types of features according to the branch-level gating weights to obtain fused fault features;

[0056] The classification output module is used to output the fault category diagnosis result of the excitation system based on the fused fault characteristics.

[0057] Beneficial effects

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) The present invention obtains standardized excitation timing samples with consistent channel structure, uniform sample length and clear label through preprocessing, which provides a unified input for subsequent dual-stream parallel feature extraction and fault category diagnosis.

[0060] (2) The present invention adopts a dual-stream parallel structure composed of CNN transient feature branch and BiLSTM temporal feature branch, so that local transient anomalies and long-term temporal evolution are extracted independently by different branches, avoiding the problem that a single network cannot take into account the two types of fault features.

[0061] (3) The present invention sets up a first independent bypass and a second independent bypass so that the two deep features do not pass through each other serially before fusion, thereby reducing the attenuation, smoothing or coverage of transient features such as local mutations, short-term spikes and high-frequency spikes in the multilayer transmission process.

[0062] (4) The present invention adopts a branch-level gating attention fusion mechanism to dynamically determine the contribution weights of the CNN transient feature branch and the BiLSTM temporal feature branch based on the joint feature state of the current excitation fault sample, thereby avoiding feature redundancy and insufficient contribution of key features caused by simple splicing or fixed weight fusion.

[0063] (5) The present invention can dynamically adjust the feature focus direction according to different fault waveforms, improve the transient feature contribution for local sudden faults, and improve the temporal feature contribution for cross-cycle evolution faults, thereby improving the diagnostic accuracy and stability under complex working conditions and similar fault categories. Attached Figure Description

[0064] To more clearly illustrate the implementation of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0065] Figure 1 This is a schematic diagram of the process of the present invention;

[0066] Figure 2 This is a schematic diagram of the gated bidirectional convolutional recurrent network model provided in an embodiment of the present invention. Detailed Implementation

[0067] The specific embodiments of the present invention will be further described below in conjunction with the technical solution of the present invention. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0068] Example: Refer to Figure 1 This invention discloses a fault diagnosis method for an excitation system based on a gated bidirectional convolutional recurrent network, the main steps of which are as follows:

[0069] S101: Collect multi-channel excitation waveform data of synchronous generator excitation system under different operating conditions, and preprocess it to obtain standardized excitation timing samples.

[0070] S102, the standardized excitation timing sample is input into a dual-stream parallel feature extraction network, which includes a mutually independent CNN transient feature branch and a BiLSTM temporal feature branch to extract local transient features and long-range temporal features respectively;

[0071] S103, the local transient features and the long-range temporal features are respectively input into the gated attention fusion module through independent bypass, so that the two types of features do not undergo serial transmission to each other before entering the gated attention fusion module;

[0072] S104, the gated attention fusion module generates branch-level gate weights corresponding to the CNN transient feature branch and the BiLSTM temporal feature branch based on the joint representation of local transient features and long-range temporal features, and performs adaptive weighted fusion to obtain fused fault features.

[0073] S105 integrates fault characteristics into the classification output module to obtain the fault category diagnosis results of the excitation system.

[0074] In a preferred embodiment of the present invention, in step S101, the objects of the excitation waveform data acquisition include the three-phase voltage of the synchronous generator, generator current, excitation voltage, excitation current, generator terminal voltage, and state variables related to excitation trigger control. The acquired data comes from the excitation system factory test or field operation waveform recording.

[0075] In a preferred embodiment of the present invention, step S101 includes preprocessing the signals from different channels to achieve a uniform scale, thereby reducing the impact of different physical dimensions and numerical ranges on model training. Specifically, the original excitation waveform signal is normalized by mean-standard deviation:

[0076]

[0077] in,

[0078]

[0079]

[0080] In the formula, Indicates the original waveform signal number 1 The first channel in the The value at each sampling location, Represents the standardized eigenvalues ​​after normalization. Indicates the first The average of each channel, Indicates the first The standard deviation of each channel This indicates the number of sampling points. After normalization, the data from different channels are mapped to a relatively uniform numerical distribution range, which is beneficial for subsequent convolutional and recurrent networks to stably extract features.

[0081] As a preferred embodiment of the present invention, refer to Figure 2 In step S102, the gated bidirectional convolutional recurrent network model adopts a dual-stream parallel structure. Standardized excitation time-series samples are simultaneously input into the CNN transient feature branch and the BiLSTM temporal feature branch. The CNN transient feature branch is mainly used to extract local anomalous responses within a short time range, while the BiLSTM temporal feature branch is mainly used to extract state evolution relationships within a longer time range. The two branches operate independently during the feature extraction stage, without sequentially transmitting features, thus avoiding interference between one branch and the feature representation of the other.

[0082] In a preferred embodiment of the present invention, the CNN transient feature branch is used to extract local transient features from standardized excitation timing samples. When the excitation system experiences faults such as thyristor breakdown, open circuit, or triggering anomalies, local abrupt changes, short-term spikes, high-frequency distortions, or non-stationary waveform changes often appear in the recorded waveform signal. The CNN transient feature branch captures these local abnormal patterns by sliding a one-dimensional convolution kernel along the time direction. Furthermore, a multi-scale convolution structure can be used in the CNN transient feature branch. Convolution kernels of different scales correspond to different lengths of time receptive fields. Smaller-scale convolution kernels are more suitable for capturing narrow pulses, spikes, or local abrupt changes, while larger-scale convolution kernels are more suitable for capturing local waveform distortions with longer durations. The convolution calculation process can be expressed as follows:

[0083]

[0084] In the formula, Indicates the first The convolutional kernel of scale at the th scale in the th... Features output at each position Indicates the first Each convolutional kernel size, Represents the convolution weights. Indicates the bias term. Represents a non-linear activation function. This represents the feature value corresponding to the sampling position of the input excitation timing sample. Then, the feature maps output by convolutional kernels of different scales are concatenated along the channel dimension to obtain multi-scale transient features:

[0085]

[0086] In the formula, Indicates the number of convolution scales. These represent local transient features extracted at different convolutional scales. Subsequently, pooling and dimensionality reduction are performed on the multi-scale transient features to obtain the local transient feature vector:

[0087]

[0088] In the formula, This represents the local transient feature vector output by the transient feature branch of the CNN. This represents the pooling operation. Through this process, the model can reduce feature dimensionality and computational cost while retaining representative local anomaly responses for fault identification.

[0089] In a preferred embodiment of the present invention, the BiLSTM timing feature branch is used to extract long-range timing features from the excitation waveform signal. Some faults in the excitation system do not necessarily manifest as instantaneous abrupt changes, but may exhibit phenomena such as trigger angle shift, slow state drift, periodic waveform changes, or enhanced cross-time correlation. The BiLSTM timing feature branch can simultaneously utilize forward and backward time information to more fully characterize the evolution of the fault state in the time dimension. Furthermore, in the BiLSTM timing feature branch, the long short-term memory unit controls information transmission through forget gates, input gates, and output gates; its calculation process can be represented as follows:

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] In the formula, represents the forget gate. Indicates the input gate. Indicates the output gate. Indicates the current cell state. Indicates the state of candidate cells. Indicates the current hidden state. Represents the weight matrix. Indicates the bias term. This represents the Hadama product.

[0097] The BiLSTM branch includes a forward Long Short-Term Memory (LSTM) network and a backward LSTM network. The forward network extracts state changes from the past to the present in forward chronological order, while the backward network extracts context information from the future to the present in reverse chronological order. The two networks are concatenated at each time step's hidden state to obtain a bidirectional temporal output.

[0098]

[0099] In the formula, represents the forward hidden state. Indicates a backward hidden state. This indicates a feature concatenation operation. Subsequently, the outputs at each time step are projected or aggregated to obtain a long-range temporal feature vector:

[0100]

[0101] In the formula, This represents the long-range temporal feature vector output by the temporal feature branch of the BiLSTM. This represents the feature projection or temporal aggregation function. Through this structure, the model can capture the evolution of the excitation waveform signal over a longer time range, overcoming the limitation of simple convolutional structures in characterizing long-range dependencies.

[0102] In a preferred embodiment of the present invention, in step S103, the bypass fidelity structure is positioned between the CNN transient feature branch, the BiLSTM temporal feature branch, and the gated attention fusion module. Let the local transient features output by the CNN transient feature branch be... The long-range time series features output by the BiLSTM time series feature branch are: Among them, local transient features Long-term temporal features are obtained through a first independent bypass input gating attention fusion module. The bypass propagation process through the second independent bypass input gating attention fusion module can be represented as follows:

[0103]

[0104]

[0105] In the formula, This represents the local transient characteristics of the gated attention fusion module after passing through the first independent bypass input. This represents the long-range temporal characteristics of the gated attention fusion module via the second independent bypass input; and These represent feature transfer or dimension alignment mappings in the first and second independent bypasses, respectively. When the two feature dimensions are the same, the mapping can be an identity mapping; when the two feature dimensions are different, the mapping can use linear mapping, fully connected mapping, or one-dimensional convolutional mapping for dimension alignment. Through the aforementioned first and second independent bypasses, local transient features and long-range temporal features do not undergo mutual serial transfer, cascaded progressive processing, or the process of one feature overwriting another before entering the gated attention fusion module, thus preserving the discriminative information of each type of deep feature.

[0106] In a preferred embodiment of the present invention, step S104 specifically includes the following process:

[0107] By concatenating local transient features and long-range temporal features, a joint state representation is obtained:

[0108]

[0109] In the formula, The joint state representation is then input into a learnable gating network; in this embodiment, a multilayer perceptron is chosen to obtain the gating score.

[0110]

[0111] In the formula, represents the gate score. This represents the weight matrix in a multilayer perceptron. Indicates the bias term. This represents the activation function. The gate scores are normalized using the Softmax function to obtain the gate weights for different branches:

[0112]

[0113] In the formula, The weights represent the weights corresponding to local transient features. The weights represent the weights corresponding to long-range time-series features. The gated weights satisfy the following:

[0114]

[0115]

[0116] The two types of features are weighted and fused according to the gating weights to obtain the fused fault features:

[0117]

[0118] In the formula, This represents the fusion of fault features. Through the above fusion process, the gated attention fusion module can dynamically adjust the contribution level of different branches according to the feature state of the current input sample. When the fault signal mainly manifests as local abrupt changes, high-frequency spikes, or short-term distortions, the weight of local transient features can be increased accordingly; when the fault signal mainly manifests as slow drift, periodic evolution, or cross-time-series correlation changes, the weight of long-range time-series features can be increased accordingly. Thus, the fused fault features can simultaneously contain local anomaly information and overall evolutionary information, improving the distinguishability between fault modes.

[0119] In a preferred embodiment of the present invention, in step S105, the classification output module is positioned after the gated attention fusion module and is used to map the fused fault features to specific fault categories. Specifically, the classification output module first processes the fused fault features... Global average pooling is performed to compress feature dimensions and reduce the number of parameters, preserving key discriminative information while mitigating parameter redundancy and overfitting risks associated with traditional flattening operations. Subsequently, the pooled features are input into a fully connected layer to obtain classification scores for each fault category. Finally, the classification scores are normalized using the Softmax function to obtain the probability distribution of each fault category, and the category with the highest probability is taken as the fault category diagnosis result for the excitation system. The calculation process can be expressed as follows:

[0120]

[0121] In the formula, and These represent the weight matrix and bias term of the classification layer, respectively. This indicates a global average pooling operation. This indicates the fault category diagnosis result output by the network.

[0122] In one specific embodiment, the operating condition categories may include typical operating states of the excitation system such as normal operation, thyristor breakdown and short circuit, thyristor open circuit, multi-transistor fault (either in-phase or out-of-phase), trigger angle offset, and trigger pulse loss. After the model training is completed, inputting a new excitation waveform signal will output the corresponding fault category, thereby achieving automatic identification of the operating state of the excitation system.

[0123] This invention also provides a fault diagnosis system for excitation systems based on gated bidirectional convolutional recurrent networks, including a data acquisition and processing module, a dual-stream parallel feature extraction module, a bypass fidelity module, a gated attention fusion module, and a classification output module. The data acquisition and processing module acquires and preprocesses the excitation waveform signal, and its output is connected to the dual-stream feature extraction module. The dual-stream parallel feature extraction module includes a CNN transient feature branch and a BiLSTM temporal feature branch, which output local transient features and long-range temporal features, respectively. The bypass fidelity module includes a first bypass and a second bypass, which are used to directly transmit the local transient features and long-range temporal features to the gated attention fusion module, respectively. The gated attention fusion module generates branch-level gate weights based on the joint representation of the two types of features and performs adaptive weighted fusion. The classification output module outputs the fault category diagnosis result of the excitation system based on the fused fault features.

[0124] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A fault diagnosis method for an excitation system based on a gated bidirectional convolutional recurrent network, characterized in that, Includes the following steps: S1: Collect multi-channel excitation waveform data of synchronous generator excitation system under different operating conditions, and preprocess it to obtain standardized excitation timing samples. S2, the standardized excitation timing sample is input into a dual-stream parallel feature extraction network, which includes a mutually independent CNN transient feature branch and a BiLSTM temporal feature branch to extract local transient features and long-range temporal features respectively; S3, the local transient features and the long-term temporal features are respectively input into the gated attention fusion module through independent bypasses, so that the two types of features do not undergo serial transmission to each other before entering the gated attention fusion module; S4, the gated attention fusion module generates branch-level gated weights corresponding to the CNN transient feature branch and the BiLSTM temporal feature branch based on the joint representation of local transient features and long-term temporal features, and performs adaptive weighted fusion to obtain fused fault features; S5 integrates fault characteristics into the classification output module to obtain the fault category diagnosis results of the excitation system.

2. The method according to claim 1, characterized in that, In S1, the preprocessing includes channel unification, time window division, normalization processing, and sample label construction to obtain standardized excitation timing samples with consistent channel structure and uniform sample length. In S2, the CNN transient feature extraction branch includes at least one one-dimensional convolutional layer that slides along the time direction, used to extract local abrupt changes, short-time spikes, glitches and high-frequency distortion features in the excitation recording signal caused by thyristor breakdown, thyristor open circuit or triggering abnormality.

3. The method according to claim 2, characterized in that, The transient feature extraction branch employs a multi-scale one-dimensional convolutional structure, with different scale convolutional kernels corresponding to different lengths of temporal receptive fields. Small-scale convolutional kernels are used to extract narrow pulse and spike abrupt change features, while large-scale convolutional kernels are used to extract local waveform distortion features with longer durations. The multi-scale convolution is represented as follows: In the formula, This represents the feature output by the k-th scale convolution kernel at position t. The kernel size is the convolution kernel size. As weight, For bias terms, It is a non-linear activation function. The input excitation timing sample corresponds to the sampled feature value; The output features of convolutional kernels of different scales are concatenated along the channel dimension and then pooled to obtain local transient feature vectors.

4. The method according to claim 1, characterized in that, In S2, the time-series evolution feature extraction branch includes forward LSTM units and backward LSTM units, which extract the context information of the excitation recording signal from the forward and backward time directions, respectively, to characterize cross-cycle changes, slow drift and long-range time-series dependence caused by trigger angle offset, trigger pulse loss or control loop abnormality. The temporal feature extraction branch is a bidirectional long short-term memory network branch, and its gate unit is calculated as follows: In the formula, Represents the Gate of Oblivion Indicates the input gate. Indicates the output gate. Indicates the current cell state. Indicates the state of candidate cells. Indicates the current hidden state. , , , This is the weight matrix. , , , For bias terms, Represents the Hadama product; The BiLSTM temporal feature branch concatenates the forward and backward hidden states and aggregates or projects the bidirectional hidden states at each time step to obtain long-range temporal features.

5. The method according to claim 1, characterized in that, In S2, the CNN transient feature branch and the BiLSTM temporal feature branch receive the same standardized excitation time series sample as input and operate independently of each other in the feature extraction stage.

6. The method according to claim 1, characterized in that, In S3, the independent bypass includes a first bypass connecting the output of the CNN transient feature branch to the gated attention fusion module, and a second bypass connecting the output of the BiLSTM temporal feature branch to the gated attention fusion module; the first bypass and the second bypass directly transmit local transient features and long-range temporal features respectively, so that the two types of features do not undergo mutual serial transmission or cascaded progressive processing before entering the gated attention fusion module.

7. The method according to claim 1, characterized in that, In S4, the gated attention fusion includes: concatenating the local transient features and long-range temporal features along the feature dimension to obtain joint state features; inputting the joint state features into a multilayer perceptron to generate gate scores corresponding to the CNN transient feature branch and the BiLSTM temporal feature branch; and normalizing the gate scores to obtain the first gate weight corresponding to the CNN transient feature branch and the second gate weight corresponding to the BiLSTM temporal feature branch.

8. The method according to claim 7, characterized in that, When the fault waveform corresponding to the input excitation timing sample exhibits local abrupt changes, short-term spikes, or high-frequency distortion, the gated attention fusion module increases the first gate weight corresponding to the transient feature branch of the CNN; when the fault waveform corresponding to the input excitation timing sample exhibits slow drift, cross-cycle changes, or trigger angle shift, the gated attention fusion module increases the second gate weight corresponding to the timing feature branch of the BiLSTM; the local transient features and long-range timing features are weighted and fused according to the first gate weight and the second gate weight to obtain the fused fault features.

9. The method according to claim 1, characterized in that, In S5, the classification output module performs global average pooling on the fused fault features and inputs the pooled features into the fully connected layer. It outputs the probability corresponding to each fault category through the normalized classification function and takes the fault category with the highest probability as the fault category diagnosis result of the excitation system.

10. A fault diagnosis system for an excitation system based on a gated bidirectional convolutional recurrent network, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition and processing module is used to acquire multi-channel excitation waveform signal data and preprocess the signals to obtain standardized excitation timing samples. The dual-stream parallel feature extraction module implements the CNN transient feature branch and the BiLSTM temporal feature branch, which are used to extract local transient features and long-range temporal features respectively. The bypass fidelity module includes a first bypass and a second bypass, which are used to directly transmit local transient features and long-term temporal features to the gated attention fusion module, respectively. The gated attention fusion module is used to generate branch-level gate weights based on the joint representation of local transient features and long-term temporal features, and to adaptively weight and fuse the two types of features according to the branch-level gate weights to obtain fused fault features. The classification output module is used to output the fault category diagnosis result of the excitation system based on the fused fault characteristics.

Images