A synchronous generator electromechanical fault diagnosis method based on a digital twin model
By combining digital twin models with the synchronous acquisition and processing of voltage and vibration signals, the problems of insufficient information and instability in the diagnosis of complex faults in synchronous generators are solved, and high-precision fault identification and stable output are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTH CHINA ELECTRIC POWER UNIV
- Filing Date
- 2026-04-23
- Publication Date
- 2026-07-03
AI Technical Summary
Existing synchronous generator fault diagnosis schemes cannot effectively utilize the dynamic evolution characteristics and coupled response characteristics of the electrical and mechanical sides, resulting in insufficient accuracy and stability in complex fault diagnosis. Traditional methods also have limited ability to process non-stationary fault signals.
A digital twin-based approach is adopted, which simultaneously acquires three-phase voltage and dual-axis vibration signals, combines continuous wavelet transform and dual-branch network, utilizes voltage LSTM branch and vibration CNN branch, and introduces spatial and channel collaborative attention module to improve information coverage and feature extraction capability, thereby achieving fault mode differentiation and stability.
It improves the information coverage and diagnostic accuracy of complex electromechanical faults in synchronous generators, enhances the observability of local abnormal modes and the ability to distinguish fault modes, and improves the stability and accuracy of diagnosis.
Smart Images
Figure CN122332869A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of synchronous generator condition monitoring and intelligent diagnosis technology, specifically relating to a method for diagnosing electromechanical faults in synchronous generators based on a digital twin model. Background Technology
[0002] Synchronous generators, as key equipment in power systems, are prone to electromechanical coupling faults such as rotor-to-turn short circuits, stator-to-turn short circuits, and air gap eccentricity during long-term operation. If these faults are not identified in time, they can easily lead to abnormal temperature rise, insulation degradation, increased vibration, decreased efficiency, or even shutdown accidents.
[0003] Existing synchronous generator fault diagnosis schemes mostly use a single electrical signal or a single vibration signal for judgment. Although they can reflect local anomalies under specific operating conditions, because synchronous generator faults have both electrical side dynamic evolution characteristics and mechanical side coupled response characteristics, single-mode diagnosis methods are prone to problems such as insufficient feature information, insufficient differentiation of complex fault boundaries, and weak generalization ability.
[0004] On the other hand, traditional manual feature extraction methods rely heavily on experience and are difficult to balance local time-frequency features and long-term sequence dependencies in non-stationary fault signals. Furthermore, if only ordinary convolutional networks or ordinary recurrent networks are used, they are still easily affected by redundant textures, irrelevant responses, and weak fault feature masking, making it difficult to further improve the accuracy and stability of complex electromechanical fault diagnosis of synchronous generators. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a method for diagnosing electromechanical faults in synchronous generators based on digital twin models, thereby resolving the issues in the prior art. The technical solution adopted by this invention is as follows: The present invention has the following beneficial effects: (1) This invention achieves the coordinated use of electrical side dynamic information and mechanical side response information by synchronously collecting three-phase voltage and biaxial vibration signals, thereby improving the information coverage of complex electromechanical faults of synchronous generators; (2) This invention converts non-stationary fault sequences into time-frequency diagrams by performing continuous wavelet transform on vibration signals, which is beneficial to improving the observability of local abnormal modes; (3) This invention constructs a dual-branch network that combines voltage LSTM branch and vibration CNN branch, while taking into account both long-term time-dependent features and local time-frequency texture features; (4) By introducing a spatial and channel collaborative attention module into the vibration branch, the present invention can enhance key time-frequency regions and key channels in a targeted manner, thereby improving the ability to distinguish complex fault modes and the stability of diagnosis. (5) By constructing a digital twin model corresponding to the operating state of the synchronous generator, the present invention can realize the digital mapping output of fault categories and confidence levels, which facilitates subsequent display, alarm and linkage control. Attached Figure Description
[0006] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the overall structure of the present invention; Figure 3 This is a schematic diagram of the structure of the digital twin model of the present invention; Figure 4 This is a schematic diagram of the SCSA module of the present invention; Figure 5 This is a schematic diagram of the continuous wavelet transform principle of the present invention; Figure 6 The confusion matrix of CNN-LSTM-SCSA; Figure 7 A comparison of diagnostic performance metrics for different models; Figure 8 Performance comparison of CNN-LSTM-SCSA under different dataset partitioning ratios. Detailed Implementation
[0007] The following will be described in conjunction with embodiments of the present invention. Figures 1-8 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.
[0008] A method for diagnosing electromechanical faults in synchronous generators based on digital twin models includes the following steps: Step 1: During the operation of the synchronous generator, three-phase voltage signals are acquired through a voltage acquisition unit, and biaxial vibration signals are acquired through two vibration acquisition units arranged in mutually orthogonal directions. The three-phase voltage signals and biaxial vibration signals are time-synchronized and segmented by a sliding window to obtain multi-channel raw samples. Step 2: Standardize the signals of each channel of each original sample, and perform continuous wavelet transform on the standardized biaxial vibration signal to obtain the corresponding vibration time-frequency feature map; Step 3: Construct a digital twin model, which includes a voltage time-series branch, a vibration time-frequency branch, a feature fusion layer, and a classification output layer. The voltage time-series branch is used to model the three-phase voltage sequence using a long short-term memory network to extract time-series features. The vibration time-frequency branch is used to extract convolutional features from the vibration time-frequency feature map. In the process of convolutional feature extraction, a spatial and channel collaborative attention module is introduced to obtain enhanced vibration features. Step 4: Input the labeled samples into the model for supervised training to obtain the trained digital twin model; Step 5: Input the sample to be tested into the digital twin model, fuse the voltage time series features and the enhanced vibration features, and output the corresponding fault category and fault confidence. Step 6: Generate the status identification result of the synchronous generator based on the fault category and confidence level. The status identification result can be used for display, archiving, alarm, and linkage control.
[0009] Furthermore, in step one, the multi-channel original sample consists of a three-phase voltage channel and a biaxial vibration channel; the sliding window segmentation uses a fixed window length T and a preset overlap rate. This allows continuously sampled data to be divided into multiple time window samples.
[0010] Furthermore, in step two, the standardization process is as follows: for the j-th channel signal of any time window sample... ,according to Standardization is carried out, among which Let be the mean value of the j-th channel signal. Let be the standard deviation of the j-th channel signal, and ε be a constant to prevent the denominator from being zero.
[0011] Furthermore, in step two, regarding the vibration signal... Its continuous wavelet transform is defined as Where a and b represent the scaling factor and translation factor, respectively. The wavelet mother function is used. The continuous wavelet transform employs Morlet wavelets to map the biaxial vibration signal in both the scale domain and the time domain, respectively, to generate a two-dimensional time-frequency feature map characterizing local time-domain variations and frequency-domain energy distribution.
[0012] Furthermore, in step three, the vibration time-frequency branch includes a convolutional layer, an activation layer, a pooling layer, and an adaptive global average pooling layer connected in sequence; the convolutional layer is used to extract local time-frequency texture features, the pooling layer is used to compress the feature map size and expand the receptive field, and the adaptive global average pooling layer is used to map high-dimensional convolutional features into vibration feature vectors of fixed length.
[0013] Furthermore, in step three, the Spatial and Channel Co-Attention (SCSA) module includes a shared multi-semantic spatial attention unit (SMSA) and a progressive channel self-attention unit (PCSA) arranged in series, and is embedded in the vibrating branch convolutional feature extraction process to collaboratively enhance the feature map in both spatial and channel dimensions. Let the output feature map of the convolutional layer be... Where B, C, H, and W represent batch size, number of channels, height, and width, respectively. The overall structure of SCSA consists of shared multi-semantic space attention (SMSA) and progressive channel self-attention (PCSA) concatenated, and its overall form can be expressed as: .
[0014] In the Shared Multi-Semantic Spatial Attention Unit (SMSA), input features are first aggregated along two spatial directions, and then combined with multi-scale depth-shared one-dimensional convolution to extract spatial semantic information under different receptive fields. Subsequently, spatial attention weights are generated in the two directions, and the input features are recalibrated. Let the spatial weights in the two directions be denoted as... and Then the output of SMSA can be written as Where ⊙ denotes element-wise multiplication. Equation This indicates that SMSA essentially weights convolutional features in two spatial directions to highlight fault-sensitive regions and suppress irrelevant background, thereby obtaining spatially enhanced features. Subsequently, PCSA further models the dependencies between different channels.
[0015] In the progressive channel self-attention unit PCSA, first for Progressive compression is performed, followed by linear projection to obtain the Query, Key, and Value, and a self-attention matrix is calculated along the channel dimension. Its core computation can be represented as follows: Where Q, K, and V represent Query, Key, and Value, respectively, and M is the channel relevance matrix. Unlike traditional channel attention, which assigns weights independently to each channel, the formula... It explicitly characterizes the correlation between channels, thus enabling more accurate modeling of the complementary and redundant relationships of fault information in different convolutional channels.
[0016] Furthermore, in step five, the feature fusion employs a feature concatenation method to combine the voltage time-series features. and enhanced vibration characteristics Perform fusion and combine fusion features Input a fully connected classifier, activate it with ReLU, regularize it with Dropout, and apply it with Softmax to obtain the fault category and its corresponding confidence level.
[0017] Furthermore, in step four, the digital twin model is trained under supervision using labeled samples, and the training termination condition is determined based on the validation set loss and classification accuracy; the labeled samples include at least the normal state, rotor inter-turn short circuit state, stator inter-turn short circuit state, and radial static air gap eccentricity state.
[0018] Based on the above method, this invention also proposes a synchronous generator electromechanical fault diagnosis system based on a digital twin model, comprising: a signal acquisition module for acquiring three-phase voltage signals and biaxial vibration signals of the synchronous generator; a preprocessing module for performing time synchronization, sliding window segmentation, standardization, and continuous wavelet transform on the acquired signals; a diagnosis module for calling the trained digital twin model to extract voltage time-series features and vibration time-frequency features, and realizing feature enhancement, fusion classification, and fault identification based on a spatial and channel collaborative attention module; and an output module for outputting fault category and confidence level, which can be used for display, archiving, alarm, and linkage control based on the state identification results.
[0019] The present invention provides the following specific embodiments: Example 1: As Figure 2 As shown, the diagnostic system in this embodiment includes a synchronous generator 201, a voltage sensor group 202, a dual-axis vibration sensor 203, a data acquisition and synchronization module 204, a preprocessing and diagnostic processor 205, a display terminal 206, and an alarm control interface 207. The voltage sensor group 202 is used to acquire the three-phase voltage of the synchronous generator; the dual-axis vibration sensor 203 is arranged in mutually orthogonal directions to acquire vibrations in the x and y directions respectively; the data acquisition and synchronization module 204 is used to complete synchronous sampling of multi-source signals; the preprocessing and diagnostic processor 205 is used to perform data preprocessing, model training, and online fault diagnosis; the display terminal 206 is used to display the fault type and confidence level; and the alarm control interface 207 is used to output an alarm or control command when a preset threshold is reached.
[0020] like Figure 1 As shown, the method in this embodiment first acquires three-phase voltage signals and biaxial vibration signals, and then performs sliding segmentation according to a preset window length T and overlap rate η to obtain time window samples of uniform length. For each time window sample, the signals of the five channels are standardized to reduce the impact of differences in amplitude magnitude between different channels on the stability of model training.
[0021] like Figure 5 As shown, after standardization, Morlet continuous wavelet transform is performed on the biaxial vibration signal to convert the one-dimensional vibration sequence into a two-dimensional time-frequency feature map. This time-frequency feature map retains both local time-domain variation information and frequency-domain energy distribution information, which can enhance the separability of weak faults, non-stationary faults, and electromechanical coupling faults.
[0022] like Figure 3 As shown, the digital twin model adopts a dual-branch structure. The three-phase voltage sequence is input into the voltage timing branch, which uses a long short-term memory network to extract the dynamic changes within the sample window and outputs high-level timing features. The dual-axis vibration sequence is first processed by Morlet-CWT time-frequency mapping and then input into the convolutional feature extraction module, before entering the SCSA collaborative attention module to obtain enhanced vibration features.
[0023] like Figure 4 As shown, the SCSA module consists of an SMSA spatial attention unit and a PCSA channel attention unit connected in series. The SMSA spatial attention unit generates attention weights in different spatial directions by performing semantic aggregation along two spatial directions of the feature map, which are used to highlight fault-sensitive areas and suppress irrelevant backgrounds. The PCSA channel attention unit generates Query, Key and Value based on linear projection, and enhances channels with discriminative capabilities through the channel correlation matrix, thereby improving the vibration branch's ability to extract complex fault features.
[0024] In the feature fusion stage, voltage features and enhanced vibration features are fused in the feature stitching module, and then input into the fully connected and Softmax classification module. The output includes multiple candidate categories and corresponding confidence levels for normal state, rotor inter-turn short circuit state, stator inter-turn short circuit state, and radial static air gap eccentricity state. The fault category output results are as follows: Figure 3 As shown in the figure.
[0025] Example 2: Verification experiments were conducted on a synchronous generator experimental platform. The experimental platform used a CS-5 type fault simulation generator set. The generator operated under rated conditions, maintaining a speed of 3000 r / min, and the ambient reference temperature was 22℃. Rotor inter-turn short-circuit faults were simulated using rotor winding inter-turn short-circuit taps, stator inter-turn short-circuit faults were simulated using stator winding inter-turn short-circuit taps, and radial static air gap eccentricity faults were simulated by adjusting the horizontal radial displacement of the stator relative to the rotor.
[0026] In this embodiment, the input samples consist of five channels: three-phase voltage and biaxial vibration. The voltage channel characterizes the dynamic changes on the electrical side, and the vibration channel characterizes the coupling response on the mechanical side. Labeled samples are used for supervised training of the digital twin model, and online diagnostics are performed on the test samples after training.
[0027] like Figure 6 As shown, in one set of embodiments, the prediction results of CNN-LSTM-SCSA are mainly concentrated on the main diagonal, with significantly fewer off-diagonal elements compared to other contrasting models. This indicates that the proposed model has clear class boundaries between the four states: normal state, rotor inter-turn short circuit, stator inter-turn short circuit, and radial static air gap eccentricity, effectively suppressing misclassification. Figure 7 As shown, the proposed solution, on a dataset containing four states—normal state, rotor inter-turn short circuit, stator inter-turn short circuit, and radial static air gap eccentricity—demonstrates superior diagnostic accuracy, error control, and training stability compared to a standalone CNN model, a standalone LSTM model, and a CNN-LSTM model without a collaborative attention module. Specifically, the accuracy reaches 0.9817, precision reaches 0.9849, recall reaches 0.9838, F1 score reaches 0.9843, and AUC reaches 0.9921, all outperforming the comparative models. This indicates that the proposed solution not only effectively improves the diagnostic rate of various electromechanical faults in synchronous generators but also reduces prediction bias and enhances the convergence stability of the training process. Furthermore, to further verify the stability and generalization ability of the proposed model under different training sample sizes, the diagnostic performance of CNN-LSTM-SCSA under different partition ratios was compared while maintaining the model structure and training settings. Figure 8 As shown, although the model's performance decreases with fewer training samples, the overall trend remains stable without any abnormal fluctuations or severe instability. This indicates that CNN-LSTM-SCSA can achieve high-precision fault diagnosis not only with sufficient training samples but also with a certain degree of robustness and usability even with relatively limited training samples.
[0028] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, alterations, or substitutions made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for diagnosing an electromechanical fault of a synchronous generator based on a digital twin model, characterized in that, Includes the following steps: Step 1: Acquire the three-phase voltage signal of the synchronous generator and acquire the biaxial vibration signal along two mutually orthogonal directions. Perform time synchronization and sliding window segmentation on the three-phase voltage signal and the biaxial vibration signal to obtain multi-channel raw samples; Step 2: Standardize the signals of each channel of each multi-channel original sample, and perform continuous wavelet transform on the standardized biaxial vibration signal to obtain the corresponding vibration time-frequency feature map; Step 3: Construct a digital twin model, which includes a voltage time-series branch, a vibration time-frequency branch, a feature fusion layer, and a classification output layer. The voltage time-series branch is used to model the three-phase voltage signal using a long short-term memory network to extract voltage time-series features. The vibration time-frequency branch is used to extract convolutional features from the vibration time-frequency feature map. In the process of convolutional feature extraction, a spatial and channel collaborative attention module is introduced to obtain enhanced vibration features. Step 4: Input the labeled samples into the digital twin model for supervised training to obtain the trained digital twin model; Step 5: Input the sample to be tested into the digital twin model, fuse the voltage timing features and the enhanced vibration features, and output the corresponding fault category and fault confidence. Step 6: Generate the state identification results of the synchronous generator based on the fault category and fault confidence level.
2. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 1, characterized in that, In step two, the standardization process includes: for the j-th channel signal of any time window sample ,according to Standardization is carried out, among which Let be the mean value of the j-th channel signal. Let be the standard deviation of the j-th channel signal, and ε be a constant.
3. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 1, characterized in that, Step two, when performing continuous wavelet transform, includes: for vibration signals Its continuous wavelet transform is defined as: Where a and b represent the scaling factor and translation factor, respectively. The wavelet mother function is used; the continuous wavelet transform uses Morlet wavelets to map the biaxial vibration signal in the scale domain and time domain respectively, so as to generate vibration time-frequency feature maps that characterize local time-domain changes and frequency-domain energy distribution.
4. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 1, characterized in that, In step three, the vibration time-frequency branch includes a convolutional layer, an activation layer, a pooling layer, and an adaptive global average pooling layer connected in sequence; the convolutional layer is used to extract local time-frequency texture features, the pooling layer is used to compress the feature map size and expand the receptive field, and the adaptive global average pooling layer is used to map high-dimensional convolutional features into vibration feature vectors of fixed length.
5. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 4, characterized in that, In step three, the spatial and channel collaborative attention module includes a shared multi-semantic spatial attention unit (SMSA) and a progressive channel self-attention unit (PCSA) set in series, which are embedded in the vibration time-frequency branch convolutional feature extraction process to collaboratively enhance the feature map in the spatial and channel dimensions. Let the output feature map of the convolutional layer be Where B, C, H, and W represent batch size, number of channels, height, and width, respectively; the spatial and channel collaborative attention module is represented as: in, This is the enhanced output feature map of the spatial and channel collaborative attention module. This describes the processing procedure of the spatial and channel-based collaborative attention module. This represents the processing procedure of attention units sharing a multi-semantic space. This represents the processing procedure of the progressive channel self-attention unit.
6. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 5, characterized in that, In the Shared Multi-Semantic Spatial Attention Unit (SMSA), the input features are first aggregated along two spatial directions, and then combined with multi-scale depth-shared one-dimensional convolution to extract spatial semantic information under different receptive fields. Subsequently, spatial attention weights in the two directions are generated, and the input features are recalibrated. The output of the Shared Multi-Semantic Spatial Attention Unit (SMSA) is represented as follows: ,in and These represent the spatial weights in two directions, and ⊙ denotes element-wise multiplication. In the progressive channel self-attention unit PCSA, first for Progressive compression is performed, followed by linear projection to obtain the query vector, key vector, and value vector. A self-attention matrix is then calculated along the channel dimension, as follows: , where Q, K and V represent the query vector, key vector and value vector respectively, and M is the channel relevance matrix.
7. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 1, characterized in that, In step five, feature fusion uses a feature splicing method to combine voltage time series features. and enhanced vibration characteristics Perform fusion and combine fusion features Input a fully connected classifier, activate it with ReLU, regularize it with Dropout, and apply it with Softmax to obtain the fault category and its corresponding confidence level.
8. The method for diagnosing electromechanical faults of a synchronous generator based on a digital twin model according to claim 1, characterized in that, In step four, the digital twin model is trained under supervision using labeled samples, and the training termination condition is determined based on the validation set loss and classification accuracy. The labeled samples include at least the normal state, rotor inter-turn short circuit state, stator inter-turn short circuit state, and radial static air gap eccentricity state.