Pumped storage group fault diagnosis model training method and fault diagnosis method

Abstract

The invention discloses a method for training a fault diagnosis model of a pumped storage unit, and also discloses a method for performing fault diagnosis by adopting the trained model. According to the model training method provided by the invention, the improved TimeGAN algorithm is adopted to generate new data on the basis of the existing data, and then the new data are input into the LightGBM model for training to finally obtain the model for fault diagnosis of the pumped storage unit. According to the TimeGAN algorithm improvement method provided by the invention, the JS divergence and the optimizer in the original model are replaced and a time attention mechanism is added, so that the generated data are closer to the actual data, and the LightGBM fault classification model is assisted for training, thereby improving the stability of fault diagnosis of the small sample pumped storage unit.

Description

Technical Field

[0001] This invention belongs to the field of pumped storage unit fault diagnosis technology, and relates to a method for training a fault diagnosis model for pumped storage units, as well as a fault diagnosis method for using the fault diagnosis model obtained by this method. Background Technology

[0002] Pumped storage hydroelectric units are key energy storage equipment for building new power systems. Their frequent start-stop operation makes condition monitoring and accurate early fault diagnosis crucial. However, the scarcity of unit fault samples and uneven data distribution severely limit the diagnostic performance of deep learning models.

[0003] Time-series generative adversarial networks (TimeGANs), as a method capable of learning the distribution of time-series data and generating synthetic samples, offer a new approach to alleviating the aforementioned data bottleneck. By jointly training a generator and a discriminator, it aims to capture the dynamic time dependencies in vibration signals, thereby generating data highly similar to real fault sequences to expand the sample size and improve the model's generalization ability.

[0004] However, traditional TimeGAN models still have key limitations when applied to vibration swing signals from pumped storage units. First, the adversarial loss function based on JS divergence in the TimeGAN model is prone to gradient instability and pattern collapse during training, affecting generation quality and stability. Second, although the model uses a GRU network to capture temporal dependencies, it lacks an explicit attention mechanism for key time steps, making it difficult to effectively capture subtle but critical faults such as transient shocks in the signal. Finally, the optimizers commonly used in model training suffer from convergence inconsistencies and late-stage oscillations, resulting in low training efficiency and requiring more iterations to stabilize. Based on these problems, fault diagnosis models trained using supplementary data from traditional TimeGAN models have low accuracy and require long training times, making them unsuitable for practical fault diagnosis needs. Summary of the Invention

[0005] The purpose of this invention is to provide a training method for a fault diagnosis model of pumped storage units, which solves the problems of low training efficiency and inaccurate model detection results in existing fault diagnosis models of pumped storage units due to scarce data samples.

[0006] Another objective of this invention is to provide a method for diagnosing faults in pumped storage units.

[0007] The technical solution adopted in this invention is a fault diagnosis model training method for pumped storage units, which is carried out according to the following steps: Step 1: Collect vibration and oscillation data of the pumped storage power station units; Step 2: Denoise the vibration swing data; Step 3: Extract time-domain and frequency-domain features from the denoised vibration swing data and divide it into training and testing sets; Step 4: Use the improved time-series generative adversarial network algorithm to generate simulated data based on the training set; Step 5: Train the LightGBM model using simulated data and verify the model performance using a test set. When the model performance meets the standard, use the model at this time as the fault diagnosis model.

[0008] The invention is further characterized by: In step 2, the singular value decomposition method is used to denoise the vibration swing data.

[0009] The time-domain features in step 3 include the maximum value, minimum value, mean, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, impulse factor, and margin factor of the signal. The frequency-domain features include the centroid frequency, average frequency, root mean square frequency, frequency standard deviation, frequency variance, root mean square frequency, spectral skewness, spectral kurtosis, spectral flatness, mean spectral kurtosis, standard deviation of spectral kurtosis, and skewness of spectral kurtosis.

[0010] The improvements to the TimeGAN algorithm in step 4 include replacing JS divergence with Wasserstein distance and gradient penalty; Introduce a time attention mechanism into the generator and discriminator; Replace the optimizer from Adam with RAdam.

[0011] When improving the TimeGAN algorithm by replacing JS divergence with Wasserstein distance and gradient penalty, the following modifications are required: First, remove the Sigmoid activation function from the original discriminator output layer and replace it with an unbounded real-valued score. Secondly, Wasserstein distance and gradient penalty are used instead of JS divergence, defined as follows:

[0012]

[0013] ,

[0014] Combining the above three equations yields a newly defined adversarial loss function:

[0015] In the formula, Let x be the real sample and y be the generated sample. express The lower bound of expectation For the edge distribution and The set of joint distributions, For gradient penalty, The coefficient of the gradient penalty term, To generate a linear interpolation between the sample and the real sample, These are the interpolation coefficients. For generator samples; When improving the TimeGAN algorithm by introducing a time attention mechanism into the generator and discriminator, the following modifications are required: After the three-layer GRU layer of generator and discriminator and before the output layer, a TA layer is inserted. The TA layer performs weighted aggregation of the hidden states of all time steps output by the GRU to generate a context vector to highlight the information of important time steps. For the generator, after the three-layer GRU, the context vector generated by the TA layer is fused with the original GRU output and used as the input of the fully connected layer. For the discriminator, after the three-layer GRU, the context vector generated by the TA layer is input into the fully connected layer of the discriminator for true / false classification.

[0016] The improvement to the TimeGAN algorithm directly replaces the optimizers of the generator and discriminator from Adam to RAdam, while retaining the original hyperparameters, which are consistent with the settings used by the original Adam optimizer.

[0017] In step 5, the model performance is verified using a test set. When the Pearson correlation coefficient between the predicted result and the actual result is greater than 0.95, the model performance meets the standard and the model at this time is used as the fault diagnosis model. Otherwise, the improved TimeGAN algorithm is used to regenerate data based on the existing data for model training until the obtained model meets the requirements.

[0018] Another technical solution adopted in this invention is a fault diagnosis method for pumped storage units, which is carried out according to the following steps: Step 1: The vibration swing data of the pumped storage power station unit to be diagnosed is processed for noise reduction using the singular value decomposition method; Step 2: Extract features from the denoised vibration swing data, extracting time-domain features and frequency-domain features respectively; Step 3: Input the extracted features into the fault diagnosis model to perform fault diagnosis and output the diagnosis results.

[0019] The beneficial effects of this invention are: This invention improves the TimeGAN algorithm. Firstly, the improved algorithm provides smoother vibration signals, significantly enhancing stability during the modeling phase. Secondly, by introducing a time attention mechanism, the improved algorithm can adaptively focus on key time segments in the vibration signal, enhancing its ability to capture transient fault features. Finally, modifications to the optimizer, through dynamic adjustment of the adaptive learning rate, alleviate fluctuations in the early stages of training, improving training efficiency while maintaining generation quality. The method provided by this invention can generate higher-quality data even with scarce samples and uneven data distribution, assisting in the training of the LightGBM fault classification model, thereby improving the stability of fault diagnosis for small-sample pumped storage units. Attached Figure Description

[0020] Figure 1 This is a structural diagram of the improved TimeGAN model in this invention; Figure 2 This is a graph showing the test results of performance test two. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: The training method for the fault diagnosis model of pumped storage units shall be carried out according to the following steps: Step 1: Collect vibration and oscillation data of the pumped storage power station units; Step 2: Denoise the vibration swing data; Step 3: Extract time-domain and frequency-domain features from the denoised vibration swing data and divide it into training and testing sets; Step 4: Use the improved time-series generative adversarial network algorithm to generate simulated data based on the training set; Step 5: Train the LightGBM model using simulated data and verify the model performance using a test set. When the model performance meets the standard, use the model at this time as the fault diagnosis model.

[0023] Example 2: The training method for the fault diagnosis model of pumped storage units shall be carried out according to the following steps: Step 1: Collect vibration and oscillation data of the pumped storage power station units; Step 2: Use singular value decomposition to denoise the vibration swing data; Step 3: Extract time-domain and frequency-domain features from the denoised vibration swing data and divide it into training and testing sets; Step 4: Use the improved time-series generative adversarial network algorithm to generate simulated data based on the training set; Step 5: Train the LightGBM model using simulated data and verify the model performance using a test set. When the model performance meets the standard, use the model at this time as the fault diagnosis model.

[0024] Example 3: The training method for the fault diagnosis model of pumped storage units shall be carried out according to the following steps: Step 1: Collect vibration and oscillation data of the pumped storage power station units; Step 2: Denoise the vibration swing data; Step 3: Extract time-domain and frequency-domain features from the denoised vibration swing data. The time-domain features include the maximum value, minimum value, mean, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, impulse factor, and margin factor of the signal. The frequency-domain features include the centroid frequency, average frequency, root mean square frequency, frequency standard deviation, frequency variance, root mean square frequency, spectral skewness, spectral kurtosis, spectral flatness, mean spectral kurtosis, standard deviation spectral kurtosis, and skewness of spectral kurtosis. The time-domain and frequency-domain features are then used to divide the data into training and testing sets. Step 4: Use the improved time-series generative adversarial network algorithm to generate simulated data based on the training set; Step 5: Train the LightGBM model using simulated data and verify the model performance using a test set. When the model performance meets the standard, use the model at this time as the fault diagnosis model.

[0025] Example 4: The training method for the fault diagnosis model of pumped storage units shall be carried out according to the following steps: Step 1: Collect vibration and oscillation data of the pumped storage power station units; Step 2: Denoise the vibration swing data; Step 3: Extract time-domain and frequency-domain features from the denoised vibration swing data and divide it into training and testing sets; Step 4: Use the improved time-series generative adversarial network algorithm to generate simulated data based on the training set; Step 5: Train the LightGBM model using simulated data and verify the model performance using a test set. When the model performance meets the standard, use the model at this time as the fault diagnosis model.

[0026] The improvements to the TimeGAN algorithm in step 4 include replacing JS divergence with Wasserstein distance and gradient penalty; Introduce a time attention mechanism into the generator and discriminator; Replace the optimizer from Adam with RAdam.

[0027] Example 5: Based on Example 4: When improving the TimeGAN algorithm by replacing JS divergence with Wasserstein distance and gradient penalty, the following modifications are required: First, remove the Sigmoid activation function from the original discriminator output layer and replace it with an unbounded real-valued score. Secondly, Wasserstein distance and gradient penalty are used instead of JS divergence, defined as follows:

[0028]

[0029] ,

[0030] Combining the above three equations yields a newly defined adversarial loss function:

[0031] In the formula, Let x be the real sample and y be the generated sample. express The lower bound of expectation For the edge distribution and The set of joint distributions, For gradient penalty, The coefficient of the gradient penalty term, To generate a linear interpolation between the sample and the real sample, These are the interpolation coefficients. This is a generator sample.

[0032] When using the original JS divergence to generate data, training instability issues such as vanishing and exploding gradients are prone to occur. Therefore, this application replaces JS divergence with Wasserstein distance and gradient penalty. Wasserstein distance, as a measure of distributional difference, ensures the existence and directionality of gradients at the objective function level. Meanwhile, the gradient penalty technique, at the optimization process level, enforces Lipschitz constraints to ensure that the discriminator can stably and accurately approximate this distance, thereby reliably transmitting smooth gradient signals to the generator. Providing smoother gradient signals significantly improves training stability.

[0033] When improving the TimeGAN algorithm by introducing a time attention mechanism into the generator and discriminator, the following modifications are required: After the three-layer GRU layer of generator and discriminator and before the output layer, a TA layer is inserted. The TA layer performs weighted aggregation of the hidden states of all time steps output by the GRU to generate a context vector to highlight the information of important time steps. For the generator, after the three-layer GRU, the context vector generated by the TA layer is fused with the original GRU output and used as the input of the fully connected layer. For the discriminator, after the three-layer GRU, the context vector generated by the TA layer is input into the fully connected layer of the discriminator for true / false classification.

[0034] The temporal attention mechanism re-aggregates the hidden state sequence to reveal the importance of capturing historical moments to the current moment, thereby enhancing the long-range consistency of the generated sequence. The core idea of ​​this mechanism is to allow the network to "look back" at the entire historical sequence at each time step, evaluate the importance of each time step using a learnable scoring function, and then perform a weighted summation of the historical hidden states to construct a context vector focused on key dynamics. This method can enhance the continuity of the data.

[0035] The improvement to the TimeGAN algorithm directly replaces the optimizers of the generator and discriminator from Adam to RAdam, while retaining the original hyperparameters, which are consistent with the settings used by the original Adam optimizer.

[0036] Compared with the original Adam algorithm, the RAdam used in this application no longer uses fixed bias correction, but introduces a "rectification" mechanism: in the warm-up stage, the adaptive learning rate is turned off first, and the variance is reduced by updating the degenerate momentum; once the second moment estimation is reliable enough, the complete adaptive step size is dynamically restored, thus combining the advantages of the Adam algorithm and the SGD algorithm, with excellent convergence stability and robustness, and can significantly alleviate the large oscillations in the early stage of training.

[0037] Example 6: Step 1: The vibration swing data of the pumped storage power station unit to be diagnosed is processed for noise reduction using the singular value decomposition method; Step 2: Extract features from the denoised vibration swing data, extracting time-domain features and frequency-domain features respectively; Step 3: Input the extracted features into the fault diagnosis model to perform fault diagnosis and output the diagnosis results.

[0038] Performance Test 1: The fault diagnosis model trained in this application was tested. The difference between the two control groups and the fault diagnosis model provided in this application is that the control groups used the SMOTE algorithm and the original TimeGAN algorithm respectively when generating data. Different models were trained by changing the ratio of the true parameters and generated parameters of the training parameters, and the performance of different models was tested using a test set. The test results are shown in the table below:

[0039] As shown in the table above, the model trained using the improved TimeGAN proposed in this application outperforms the model trained on the control group in all groups, including both real and generated data. This demonstrates that the model trained using the method proposed in this application yields more accurate results in fault diagnosis prediction.

[0040] Meanwhile, as shown in the table above, with the increase in the proportion of generated data, the accuracy, recall and F1-score of the model trained in this application gradually increase. Therefore, the training method provided in this application can effectively solve the problems of scarce fault samples and uneven data distribution.

[0041] Performance Test 2: In addition to the model trained in this application, the comparative models used BP, Support Vector Machine, and Random Forest algorithms instead of the LightGBM model. All models were input with 50 sets of real data and 150 sets of generated data, and underwent the same preprocessing. The parameters of each model in the comparative experiments were set to default. After training, the model performance was tested using a test set, and the statistical results were recorded as follows: Figure 2 As shown in the figure, compared with other algorithms, LightGBM maintains a balance in recognition performance across categories while demonstrating a significant advantage in handling difficult samples, making it the most effective solution for solving the problem of fault diagnosis with small samples.

Claims

1. A method for training a pumped storage unit fault diagnosis model, characterized in that, According to the following steps: Step 1: Collect the vibration displacement data of the pumped storage power station unit; Step 2: Perform noise reduction processing on the vibration displacement data; Step 3: Extract time domain features and frequency domain features from the noise-reduced vibration displacement data, and divide them into a training set and a test set; Step 4: Use the improved TimeGAN to generate simulated data based on the training set; Step 5: Train the LightGBM model using the simulated data and validate the model performance using the test set. When the model performance meets the standard, use the model at that time as the fault diagnosis model.

2. The method of claim 1, wherein the method further comprises: In step 2, the singular value decomposition method is used to perform noise reduction processing on the vibration displacement data.

3. The method of claim 1, wherein the method further comprises: In step 3, the time domain features include the maximum value, minimum value, mean value, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, waveform factor, peak factor, pulse factor, and margin factor of the signal. The frequency domain features include the center of gravity frequency, average frequency, root mean square frequency, frequency standard deviation, frequency variance, root mean square frequency, spectral skewness, spectral kurtosis, spectral flatness, spectral kurtosis mean, spectral kurtosis standard deviation, and spectral kurtosis skewness.

4. The method of claim 1, wherein the method further comprises: In step 4, the improvements to the TimeGAN algorithm include using Wasserstein distance and gradient penalty instead of JS divergence; Introduce time attention mechanism in the generator and discriminator; Replace the optimizer from Adam to RAdam.

5. The method of claim 4, wherein the method further comprises: In the improvement of the TimeGAN algorithm, when using Wasserstein distance and gradient penalty instead of JS divergence, the following changes need to be made: First, remove the Sigmoid activation function in the original discriminator output layer and change it to output an unbounded real value fraction; Second, use Wasserstein distance and gradient penalty instead of JS divergence, which is defined as follows: ， The above three formulas are combined to obtain the newly defined adversarial loss function: wherein is, x is a real sample, y is a generated sample, denotes the desired lower bound, is the marginal distribution and a set of joint distributions is the gradient penalty, is the gradient penalty term coefficient, is a linear interpolation between the generated sample and the real sample, is the interpolation coefficient, is the generator sample.

6. The method of claim 4, wherein the method further comprises: In the improvement of the TimeGAN algorithm, when introducing the time attention mechanism in the generator and discriminator, the following changes need to be made: After the three-layer GRU layer in the generator and discriminator, before the output layer, insert a TA layer. The TA layer weights and aggregates all the time step hidden states output by the GRU to generate a context vector, highlighting the information of important time steps; For the generator, after the three-layer GRU, the context vector generated by the TA layer is fused with the original GRU output as the input of the fully connected layer; For the discriminator, after the three-layer GRU, the context vector generated by the TA layer is input to the fully connected layer of the discriminator for true or false classification.

7. The method of claim 4, wherein the method further comprises: In the improvement of the TimeGAN algorithm, the optimizer of the generator and discriminator is directly replaced from Adam to RAdam, and the original hyperparameters are retained, consistent with the settings used by the original Adam optimizer.

8. The method of claim 1, wherein the method further comprises: The model performance is verified in step 5, and when the Pearson correlation coefficient between the predicted result and the actual result is greater than 0.95, the model performance meets the standard and the model at this time is taken as the fault diagnosis model, otherwise the improved TimeGAN algorithm is used to regenerate data according to the existing data for model training until the obtained model meets the requirements.

9. A method for diagnosing a fault of a pumped storage unit, characterized by, The fault diagnosis model of the pumped storage unit obtained by claim 5 is used for fault diagnosis, and the following steps are specifically performed: Step 1: The vibration swing data of the pumped storage power station unit to be diagnosed is denoised by using the singular value decomposition method; Step 2: The denoised vibration swing data is subjected to feature extraction, and time domain features and frequency domain features are extracted respectively; Step 3: The extracted features are input into the fault diagnosis model for fault diagnosis, and the diagnosis result is output.

Images