A pump cavitation pressure fluctuation signal data enhancement method based on SCWGAN

Abstract

This invention discloses a data augmentation method for pump cavitation pressure pulsation signals based on SCWGAN, belonging to the field of mechanical equipment fault diagnosis technology. First, raw pressure pulsation signals of pumps under different cavitation states are collected and preprocessed to construct a sample set. Then, an SCWGAN model with an explicit conditional injection mechanism for statistical features is built, which combines the conditional control capability of CGAN with the training stability advantages of WGAN-GP. A two-stage training strategy is used to train and optimize the model. The trained model is then used to generate original samples, and a time-domain-frequency domain post-processing mechanism is applied. Finally, the generated high-fidelity samples are fused with the original dataset to achieve data augmentation. The samples generated by this invention have high similarity to real samples, effectively alleviating the class imbalance problem of pump cavitation samples, improving the recognition accuracy and generalization ability of downstream fault diagnosis models, and providing reliable data support for accurate diagnosis and early warning of pump cavitation faults.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, and particularly relates to pump cavitation status monitoring and early warning technology, specifically to a method for enhancing pump cavitation pressure pulsation signal data based on SCWGAN. Background Technology

[0002] Pumps, as core equipment for fluid transportation, play an irreplaceable role in key fields such as petrochemicals, nuclear power, water conservancy, and shipbuilding. Cavitation is one of the main failure modes affecting the safe and stable operation of pumps. Pressure pulsation signal monitoring has become an important means of monitoring pump cavitation faults due to its advantages such as non-intrusiveness, rapid response, and ability to reflect the cavitation evolution process in real time.

[0003] However, pump cavitation pressure pulsation signals exhibit strong non-stationarity and non-linearity, with complex frequency components. Furthermore, real samples for different cavitation stages (especially severe cavitation stages) are extremely scarce, and the class distribution is severely imbalanced, leading to insufficient training and poor generalization ability of traditional data-driven deep learning diagnostic models. To address the problems of insufficient samples and class imbalance, Generative Adversarial Networks (GANs) have been widely applied in the field of fault diagnosis signal data augmentation, but traditional GAN ​​methods still have many shortcomings.

[0004] In existing technologies, the one-dimensional bearing fault signal data generation method proposed by Liu Tao et al., which integrates VAE and WGAN-GP, improves the quality of generated samples and training stability by extracting latent variables through VAE and introducing Wasserstein distance and gradient penalty terms. However, this method directly drives the generation with latent variables, resulting in insufficient fidelity of the generated samples in the time domain, coarse frequency domain distribution matching, and no targeted post-processing mechanism. This leads to significant deviations between the generated signals and the real signals in the fine-grained time domain structure. In highly non-stationary signal scenarios with complex evolution processes, such as pump cavitation pressure pulsation, the physical authenticity and diagnostic usability of the generated samples are limited, making it difficult to meet the data augmentation needs under scarce cavitation conditions.

[0005] In addition, traditional GAN ​​methods have problems when processing pump pressure pulsation signals, such as insufficient temporal fidelity of generated samples, large deviation of key statistical features, poor physical authenticity of scarce cavitation state samples, and inaccurate control of class conditions. As a result, the usability of generated data in downstream diagnostic models is limited, and it is difficult to effectively alleviate the dilemma of extreme imbalance of minority class samples such as severe cavitation. Summary of the Invention

[0006] Purpose of the invention: To address the aforementioned deficiencies in existing technologies, this invention provides a data enhancement method for pump cavitation pressure pulsation signals based on SCWGAN. By designing explicit conditional injection of statistical features, a two-stage training strategy, and a time-domain-frequency domain post-processing mechanism, high-fidelity and highly physically consistent pump cavitation pressure pulsation signals are generated. This alleviates the problem of sample class imbalance, improves the recognition accuracy and generalization ability of downstream fault diagnosis models, and provides reliable data support for accurate diagnosis and early warning of pump cavitation faults.

[0007] Technical solution: The present invention provides a method for enhancing pump cavitation pressure pulsation signal data based on SCWGAN, comprising the following steps:

[0008] Step 1: Sensor signal acquisition. The original pressure pulsation signal is acquired using a pressure pulsation sensor under four states of pump: no cavitation, initial cavitation, critical cavitation, and severe cavitation.

[0009] Step 2: Signal preprocessing and sample construction. The original pressure pulsation signal is preprocessed by low-pass filtering and normalization. The long time series signal is divided into non-overlapping samples using the random starting point fixed-length segmentation method. The real sample set is divided according to the cavitation state label, and the statistical characteristics of each sample are calculated as a condition vector.

[0010] Step 3: Build the SCWGAN model. Based on the conditional generative adversarial network CGAN, add an explicit conditional injection mechanism for statistical features. Inject the statistical feature vectors into the generator and discriminator through a dedicated projection network. Introduce a self-attention layer in the generator and use Wasserstein distance combined with gradient penalty term in the discriminator. Add temporal loss, spectral loss, multi-resolution STFT loss and statistical feature matching loss to the generator.

[0011] Step 4, Model Training and Optimization, adopts a two-stage training strategy: In the first stage, real samples and conditional vectors are input, and the discriminator and generator are trained alternately, using the Adam optimizer and learning rate decay scheduling; in the second stage, the discriminator is fixed, and only the generator is fine-tuned to minimize temporal loss.

[0012] Step 5: Sample generation and post-processing. Call the trained generator, input noise and statistical condition vector to generate original samples, and apply a time-domain to frequency-domain post-processing mechanism that includes cross-correlation alignment, envelope matching, local window scaling, and frequency domain amplitude replacement.

[0013] Step 6: Data augmentation application. The generated high-fidelity samples are merged with the original dataset to form a class-balanced augmented dataset, which is used to train the downstream pump cavitation fault diagnosis model.

[0014] Furthermore, the normalization preprocessing described in step two unifies the signal amplitude within the range of [-1, 1], and the fixed-length segmented sample window contains 1860 sampling points.

[0015] Furthermore, the statistical characteristics described in step two include the sample mean, standard deviation, skewness, and kurtosis.

[0016] Furthermore, step three specifically involves:

[0017] The core of the SCWGAN model consists of a generator G and a discriminator D. First, the environment and parameters are initialized. Then, the generator is constructed, which adopts an architecture of noise-conditional feature input-fully connected mapping-1D convolutional upsampling-self-attention-conditional injection-output. Next, the discriminator is constructed, which adopts an architecture of 1D convolutional downsampling-adaptive pooling-conditional feature fusion-discriminative output. Finally, the training process is designed to capture long-distance dependencies of the signal and avoid fragmentation of local features in the generated signal.

[0018] The Wasserstein distance, combined with a gradient penalty term, approximates the Wasserstein distance by maximizing the discriminator's output on real samples and minimizing its output on generated samples. The gradient penalty (GP) prevents gradient explosion in the discriminator. ; In the formula: To calculate the mean for all real cavitation signal samples; To traverse all noise, the mean of the generated samples output by the discriminator is calculated; To calculate the mean of the gradient penalty term; This is a real data sample; For discriminator output; Output to the generator; This is the penalty coefficient; Interpolation between real samples and generated samples; The L2 norm of the discriminator for the interpolated samples; Temporal loss: ; In the formula: For L1 loss; This is the correlation loss; For envelope loss; For zero-crossing rate loss; This is a local extremum loss; Spectral loss: ; In the formula: This is the amplitude L1 loss; For logarithmic amplitude spectrum loss; For power spectral loss; This refers to the amplitude envelope loss; Multi-resolution STFT loss: ; Where: M is the number of STFT resolutions; For the amplitude spectrum loss of STFT; The logarithmic amplitude spectrum loss of the STFT; Statistical feature loss: ; In the formula: This is a true cavitation signal; To generate a cavitation signal; , These are the mean values ​​of the actual cavitation signal and the generated cavitation signal, respectively. , These are the standard deviations of the actual cavitation signal and the generated cavitation signal, respectively. , These are the skewnesses of the actual cavitation signal and the generated cavitation signal, respectively. , These are the kurtosis of the actual cavitation signal and the generated cavitation signal, respectively.

[0019] Furthermore, in step four, during the first stage of training, the training ratio of the discriminator to the generator is 5:1, the learning rate of the generator in the Adam optimizer is 0.0001, and the learning rate of the discriminator is 0.0004; the learning rate of the second stage of training is 0.00001.

[0020] Furthermore, step five specifically involves: Based on the convergent conditional WGAN-GP generator, original candidate samples are generated by inputting random noise and statistical condition vectors of real samples. Then, a multi-dimensional post-processing with time-domain matching and frequency-domain constraints is constructed. Cross-correlation time-domain alignment, Hilbert envelope amplitude matching, and sliding window local scaling calibration are performed in sequence. Finally, phase preservation and accurate amplitude matching are achieved by frequency-domain amplitude spectrum replacement. This ensures that the generated samples are completely consistent with the real cavitation signal in terms of time-domain morphology, envelope features, local waveforms, and frequency-domain energy distribution, which significantly improves the engineering fidelity of the samples. ; In the formula: The optimal time-domain offset; It is a cross-correlation function; This is a sample of a real cavitation signal; To generate cavitation signal samples; Candidate offset; ; ; In the formula: For the signal envelope; For Hilbert transform; These are the generated samples after envelope matching; It is a numerically stable term; This is a real signal; This is the initial signal generation; ; ; In the formula: These are real samples; The frequency domain amplitude spectrum of a real sample; The imaginary unit represents the phase. The frequency domain phase angle of the signal generated after local calibration; Frequency domain signal after amplitude replacement; This is the inverse fast Fourier transform.

[0021] Furthermore, the SCWGAN model is a statistical conditional Wasserstein generative adversarial network that combines the conditional control capabilities of CGAN with the training stability of WGAN-GP, thus avoiding the mode collapse problem.

[0022] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0023] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0024] The present invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method of the present invention.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0026] 1. High fidelity of generated samples: Through the combination of explicit injection of statistical features, two-stage training strategy and time-frequency domain post-processing mechanism, the average of the five similarity indicators between the generated samples and the real samples exceeds 0.95, and the time domain correlation and envelope correlation mostly exceed 0.992. The generated samples are highly consistent with the real samples in terms of overall waveform structure, envelope shape and amplitude dynamics, and have strong physical consistency.

[0027] 2. Precise category control: The explicit conditional injection mechanism of statistical features can precisely control the statistical characteristics of generated samples, effectively improve the ability to generate scarce cavitation states (such as severe cavitation), alleviate the problem of sample class imbalance, and provide sufficient high-availability samples for downstream diagnostic models.

[0028] 3. Stable and reliable training: By integrating the loss function of WGAN-GP with a two-stage training strategy, the model avoids the mode collapse problem of traditional GANs, improves the stability and convergence efficiency of model training, and reduces the difficulty of model training.

[0029] 4. High practicality: The generated high-fidelity samples can be directly used for training downstream pump cavitation fault diagnosis models, improving the model's recognition accuracy and generalization ability. They can be deployed in pump online monitoring systems to achieve dynamic data enhancement and cavitation early warning, improving the intelligence level and reliability of pump equipment operation and maintenance. They are suitable for pump cavitation monitoring scenarios in multiple key fields such as petrochemicals, nuclear power, and water conservancy. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the pump cavitation pressure pulsation signal data enhancement method based on SCWGAN of the present invention;

[0031] Figure 2 The training history curve of the SCWGAN model of this invention shows the changing trends of multiple dimensions of loss, such as discriminator loss, generator loss, temporal domain loss, and spectral loss.

[0032] Figure 3 The T-SNE distribution diagram of the generated signal and the original signal in this invention intuitively shows the consistency of the distribution between the generated sample and the real sample.

[0033] Figure 4 This is a time-domain comparison diagram of the signal amplification results based on SCWGAN in different cavitation states according to the present invention, including a comparison of the original signal and the amplified signal in four states: no cavitation, primary cavitation, critical cavitation, and severe cavitation.

[0034] Figure 5 This is a frequency domain comparison diagram of the signal amplification results based on SCWGAN under different cavitation states according to the present invention, including a comparison of the spectra of the original signal and the amplified signal under four states: no cavitation, initial cavitation, critical cavitation, and severe cavitation. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] The specific technical solution of the present invention is as follows:

[0037] Step 1: Sensor Signal Acquisition

[0038] The original pressure pulsation signal of the pump was collected by a pressure pulsation sensor under different cavitation states. The different cavitation states include four types: non-cavitation, incipient cavitation, critical cavitation, and severe cavitation. This ensures that the collected signal can fully cover the entire evolution process of pump cavitation and provide basic data for subsequent sample construction and model training.

[0039] Step 2: Signal preprocessing and sample construction

[0040] The raw signal acquired in step one undergoes low-pass filtering and normalization preprocessing: low-pass filtering removes high-frequency noise to ensure a clean signal; normalization unifies the signal amplitude within the [-1, 1] interval, eliminating the influence of different amplitude signals on model training. A random starting point fixed-length segmentation method is used, dividing the long-time-series signal into multiple samples with 1860 sampling points as a sample window, ensuring non-overlapping and diversity among samples and avoiding sample redundancy. Simultaneously, the real sample set is divided according to the cavitation state label, and the statistical characteristics of each sample (including mean, standard deviation, skewness, and kurtosis) are calculated as a condition vector for subsequent model condition control.

[0041] Step 3: Build the SCWGAN model

[0042] Based on Conditional Generative Adversarial Network (CGAN), the SCWGAN (Statistical Conditional Wasserstein GAN) model is built, and the specific optimization design is as follows:

[0043] 1. An explicit conditional injection mechanism for statistical features is introduced, in which the statistical feature vector calculated in step two is injected into the generator and discriminator through a dedicated projection network, thereby achieving precise constraints on the statistical characteristics of the generated samples and improving the model's ability to generate conditions for scarce cavitation states and its physical consistency.

[0044] 2. A self-attention layer is introduced into the generator to enhance the model's ability to capture the dependencies of long sequence signals. At the same time, conditional information is repeatedly injected in multiple upsampling stages to improve the temporal consistency of generated samples.

[0045] 3. The discriminator uses Wasserstein distance to measure the distribution distance between real samples and generated samples, and combines it with gradient penalty (GP) to ensure the stability of model training and avoid the mode collapse problem common in traditional GANs;

[0046] 4. Add multi-dimensional loss functions to the generator, including time-domain loss, spectral loss, multi-resolution STFT loss, and statistical feature matching loss, to achieve multi-dimensional fidelity constraints on the generated samples and improve sample quality.

[0047] Step 4: Model Training and Optimization

[0048] A two-stage training strategy was adopted to train and optimize the SCWGAN model. The specific process is as follows:

[0049] 1. First Stage: Input the real samples and their statistical conditional vectors obtained in Step 2 into the SCWGAN model, and train the discriminator and generator alternately, with a training ratio of 5:1 (n_critic=5). The Adam optimizer is used, with the generator learning rate set to 0.0001 and the discriminator learning rate to 0.0004. A step-decay scheduling mechanism for the learning rate is introduced to ensure stable adversarial training and initial distribution matching.

[0050] 2. Second Stage: With the discriminator parameters fixed, only the generator is fine-tuned using a low learning rate (0.00001) to further minimize temporal loss and improve the fidelity of temporal waveform details in the generated samples. During training, the temporal metrics of the generated samples (correlation, envelope correlation, etc.) are monitored in real time. The generator parameters are saved when the model reaches its optimal state for subsequent sample generation.

[0051] Step 5: Sample Generation and Performance Evaluation

[0052] The trained generator is invoked, and the statistical condition vectors corresponding to random noise and real samples are input to generate original samples. Then, a targeted time-domain-frequency domain post-processing mechanism is applied to perform personalized corrections on the generated samples, specifically including:

[0053] 1. Cross-correlation alignment: Corrects the temporal offset between generated samples and real samples, ensuring consistency between the two on the time axis;

[0054] 2. Envelope matching: Adjust the envelope shape of the generated sample to make it consistent with the envelope features of the real sample;

[0055] 3. Local window scaling: Fine-tunes the local waveforms of the generated samples to improve fine-grained temporal structure matching.

[0056] 4. Frequency Domain Amplitude Replacement: The phase information of the generated sample is retained and replaced with the frequency domain amplitude characteristics of the real sample to improve the physical authenticity of the generated sample.

[0057] The performance of the generated samples and real samples after processing is evaluated. Five indicators are used to measure the similarity between the two: correlation, envelope correlation, dynamic range, zero crossing rate, and local extrema. Time-domain comparison diagram, distribution visualization diagram and training history curve are plotted to intuitively show the generation quality and physical consistency.

[0058] Step Six: Data Augmentation Application Phase

[0059] The high-fidelity samples generated in step five are fused with the original dataset to form a class-balanced augmented dataset, which is used to train the downstream pump cavitation fault diagnosis model. This improves the model's recognition accuracy and generalization ability for different cavitation states (especially scarce states such as severe cavitation). This method can be deployed in pump online monitoring systems to generate auxiliary samples in real time based on the statistical characteristics of the current pressure pulsation signal, supporting dynamic data augmentation and issuing timely warnings when severe cavitation risks are detected, thereby improving the intelligence and reliability of pump equipment operation and maintenance.

[0060] Technical improvements and innovations:

[0061] 1. CGAN introduces explicit conditional injection of statistical features to improve category control and statistical fidelity: Traditional CGANs typically rely solely on implicit conditions or random noise to drive generation, making it difficult to accurately control the key statistical characteristics of the generated samples. This invention directly injects the mean, standard deviation, skewness, and kurtosis of real samples as conditional vectors into the generator and discriminator, achieving precise constraints on the statistical characteristics of the generated samples and improving the model's ability to generate conditional samples in scarce vacuolated states and its physical consistency.

[0062] 2. A two-stage training strategy enhances accuracy and training stability: Traditional GANs often suffer from distortion of time-domain waveform details, pattern collapse, or unstable convergence when processing highly non-stationary pressure pulsation signals due to single-stage training. The two-stage training strategy of this invention achieves stable adversarial training and initial distribution matching in the first stage, and precise fine-tuning in the second stage to improve time-domain fidelity, significantly improving the quality of generated samples, training stability, and convergence efficiency.

[0063] 3. Introducing a targeted time-frequency post-processing mechanism to enhance physical realism: Although the original samples generated by traditional GANs have similar distributions, they often suffer from time-domain offsets, envelope distortion, or spectral amplitude deviations, resulting in poor physical interpretability and diagnostic effectiveness. The post-processing mechanism of this invention enables personalized correction for each sample, effectively improving the physical consistency of the generated signals, fine-grained time-domain structure matching, and downstream diagnostic usability.

[0064] Example

[0065] I. Test Instruments and Equipment

[0066] 1. Pump test bench: The Centrifugal pump test bench, model IS65-50-160, is used, with a rated flow rate of 25 m³ / h, a rated head of 32 m, and a rated speed of 2900 r / min;

[0067] 2. Pressure pulsation sensor: Model PCB113B22, measurement range 0-10MPa, sensitivity 10mV / MPa, sampling frequency set to 10kHz;

[0068] 3. Data acquisition card: Model NI cDAQ-9178, sampling accuracy 16-bit, sampling rate supports continuous acquisition of 10kHz;

[0069] 4. Computer: Configured with an Intel Core i7-12700H CPU, 32GB DDR5 memory, and an NVIDIA RTX3080Ti graphics card (16GB VRAM) for model training, data processing, and analysis;

[0070] 5. Software environment: The operating system is Windows 11, the programming language is Python 3.8, the deep learning framework is PyTorch 1.12.0, and the data processing libraries include NumPy, SciPy, and Matplotlib, which are used for signal preprocessing, model building, and result visualization.

[0071] II. Experimental Procedure

[0072] 1. Sensor Installation and Signal Acquisition: Install the pressure pulsation sensor on the pump outlet pipe, ensuring a tight fit between the sensor and the pipe to avoid signal interference. Adjust the inlet and outlet valves of the pump test bench to simulate four states: no cavitation, initial cavitation, critical cavitation, and severe cavitation. Run each state stably for 30 minutes, and acquire the raw pressure pulsation signal using a data acquisition card at a sampling frequency of 10kHz. The data acquired for each state should have no less than 1,000,000 sampling points.

[0073] 2. Signal Preprocessing and Sample Construction: A Butterworth low-pass filter (cutoff frequency 1kHz) was used to filter the original signal to remove high-frequency noise. The min-max normalization method was used to normalize the signal amplitude to the [-1,1] interval. A random starting point fixed-length segmentation method was used, with 1860 sampling points as a sample window, dividing the long-series signal of each state into 500 non-overlapping samples to form a true sample set. The mean, standard deviation, skewness, and kurtosis of each sample were calculated as conditional vectors and stored corresponding to the sample data.

[0074] 3. Model Setup and Parameter Settings: An SCWGAN model was built. The generator uses a convolutional neural network structure, containing 3 downsampling layers, 4 upsampling layers, and 1 self-attention layer. Each upsampling layer injects a statistical feature conditional vector. The discriminator uses a convolutional neural network structure, containing 5 convolutional layers, and introduces Wasserstein distance and gradient penalty terms. The model parameters are set as follows: batch size is 64, total training epochs are 5000, the first stage discriminator-to-generator training ratio is 5:1, the Adam optimizer generator learning rate is 0.0001, and the discriminator learning rate is 0.0004, with the learning rate decaying to 1 / 10 of its original value after 3000 epochs; the second stage learning rate is 0.00001, with 2000 training epochs.

[0075] 4. Model Training and Optimization: The real sample set and corresponding conditional vectors are input into the SCWGAN model, and training is performed according to a two-stage training strategy. The first stage involves 3000 epochs of training to achieve stable adversarial training and initial distribution matching. In the second stage, the discriminator is fixed, and only the generator is trained for 2000 epochs to minimize temporal loss. During training, the generator parameters are saved every 100 epochs, and the similarity index between the generated samples and the real samples is calculated. The generator parameters with the optimal index are selected as the final model parameters.

[0076] 5. Sample Generation and Post-processing: The trained generator is invoked, and random noise (following a normal distribution N(0,1)) and the statistical condition vector of each cavitation state are input. 500 samples are generated for each state. A time-domain to frequency-domain post-processing mechanism is applied to the generated samples: cross-correlation alignment uses a normalized cross-correlation method to correct time-domain offset; envelope matching uses Hilbert transform to extract the envelope and adjust the generated sample envelope to match the real sample envelope; local window scaling uses a sliding window (100 sampling points) for amplitude fine-tuning; frequency domain amplitude replacement uses FFT transform to retain the phase of the generated sample and replace it with the frequency domain amplitude of the corresponding real sample.

[0077] 6. Performance Evaluation and Data Application: Five indicators were calculated between the generated samples and the real samples: correlation, envelope correlation, dynamic range, zero-crossing rate, and local extrema. The results are shown in Table 1 below:

[0078] Table 1 Performance metrics of the SCWGAN model

[0079]

[0080] As shown in Table 1, the SCWGAN-based generated samples exhibit extremely high fidelity in the temporal quality assessment, with the overall average similarity of the five indicators exceeding 0.95. This indicates that the generated samples are highly consistent with the real samples in terms of overall waveform structure, envelope shape, and amplitude dynamics. Specifically, the temporal and envelope correlations of most categories (such as Non, Incipient, and Severe) exceed 0.992, and the dynamic range similarity is also higher than 0.982. Among them, the Non category performs best in most indicators (correlation reaches 0.9979), while the Severe category has the highest zero-crossing rate similarity (0.954), demonstrating the method's excellent ability to generate scarce and severely cavitation states. However, the local extremum similarity of the Critical category (0.861) and the zero-crossing rate similarity of the Incipient category (0.871) are relatively low, indicating that there is still room for optimization in capturing transition states and local transient features, which may be related to the complex impulse characteristics of these signal categories. Overall, this method significantly improves the physical consistency and fine-grained temporal fidelity of generated samples through the effective combination of explicit conditional injection of statistical features, a two-stage training strategy, and a targeted time-frequency post-processing mechanism. In particular, its robust performance in scarce categories provides highly usable synthetic samples for data augmentation of pump cavitation pressure pulsation signals and downstream fault diagnosis.

[0081] like Figure 2 As shown, this GAN training history chart illustrates the model's loss changes over 5000 epochs. The generator loss drops sharply from an initial high value of about 300 and gradually stabilizes, indicating that the generator is gradually optimized. The discriminator loss remains low and close to 0, possibly suggesting that the discriminator dominates the training. The auxiliary loss converges to near 0, showing that the model stabilizes quickly in the time domain, frequency domain, and statistical matching.

[0082] like Figure 3 As shown, this t-SNE visualization chart displays the distribution of generated data (red) and real data (blue) in the two-dimensional embedding space. The point clusters are highly mixed with no obvious separation or clustering differences, indicating that the generated data successfully captures the diversity and distribution characteristics of the real data, avoiding model collapse. The overall scatter plot coverage ranges from -150 to 150 in t-SNE1 and from -150 to 100 in t-SNE2, showing good data overlap and model generation quality.

[0083] like Figure 4 As shown, this waveform comparison plot illustrates the comparison between the original signal and the generated signal. The generated signal and the original signal are highly matched in shape, peak and valley positions, frequency and amplitude, with almost no obvious deviation. This indicates that the model performs well in time domain reconstruction, avoiding noise or distortion problems, and the overall generation quality is high.

[0084] like Figure 5 As shown, this spectrum comparison plot illustrates the comparison between the original signal and the generated signal. The generated signal is highly similar to the original signal in terms of peak position, intensity, and overall shape. The main peaks are concentrated in the low frequency band and gradually decay, with almost no significant differences or noise interference. This indicates that the model performs well in frequency domain reconstruction, successfully capturing the spectral characteristics of real data, and reflecting high overall generation quality.

[0085] The above embodiments are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and equivalent substitutions without departing from the principle of the present invention. All such improvements and equivalent substitutions to the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A method for enhancing pump cavitation pressure pulsation signal data based on SCWGAN, characterized in that, Includes the following steps: Step 1: Sensor signal acquisition. The original pressure pulsation signal is acquired using a pressure pulsation sensor under four states of pump: no cavitation, initial cavitation, critical cavitation, and severe cavitation. Step 2: Signal preprocessing and sample construction. The original pressure pulsation signal is preprocessed by low-pass filtering and normalization. The long time series signal is divided into non-overlapping samples using the random starting point fixed-length segmentation method. The real sample set is divided according to the cavitation state label, and the statistical characteristics of each sample are calculated as a condition vector. Step 3: Build the SCWGAN model. Based on the conditional generative adversarial network CGAN, add a statistical feature explicit conditional injection mechanism to inject statistical feature vectors into the generator and discriminator through a dedicated projection network. The generator incorporates a self-attention layer, the discriminator uses Wasserstein distance combined with a gradient penalty term, and the generator incorporates temporal loss, spectral loss, multi-resolution STFT loss, and statistical feature matching loss. Step 4, Model Training and Optimization, adopts a two-stage training strategy: In the first stage, real samples and conditional vectors are input, and the discriminator and generator are trained alternately, using the Adam optimizer and learning rate decay scheduling. The second stage fixes the discriminator and only fine-tunes the generator to minimize temporal loss; Step 5: Sample generation and post-processing. Call the trained generator, input noise and statistical condition vector to generate original samples, and apply a time-domain to frequency-domain post-processing mechanism that includes cross-correlation alignment, envelope matching, local window scaling, and frequency domain amplitude replacement. Step 6: Data augmentation application. The generated high-fidelity samples are merged with the original dataset to form a class-balanced augmented dataset, which is used to train the downstream pump cavitation fault diagnosis model.

2. The method according to claim 1, characterized in that, The normalization preprocessing described in step two unifies the signal amplitude within the range of [-1, 1], and the fixed-length segmented sample window has 1860 sampling points.

3. The method according to claim 1, characterized in that, The statistical characteristics mentioned in step two include the sample mean, standard deviation, skewness, and kurtosis.

4. The method according to claim 1, characterized in that, Step three specifically involves: The core of the SCWGAN model consists of a generator G and a discriminator D. First, the environment and parameters are initialized. Then, the generator is constructed. The generator adopts an architecture of noise-conditional feature input-fully connected mapping-1D convolutional upsampling-self-attention-conditional injection-output. Next, a discriminator is constructed, which adopts an architecture of 1D convolutional downsampling-adaptive pooling-conditional feature fusion-discriminative output; Then, the training process is designed to capture long-distance dependencies in the signal and avoid fragmentation of local features in the generated signal. The Wasserstein distance, combined with a gradient penalty term, approximates the Wasserstein distance by maximizing the discriminator's output on real samples and minimizing its output on generated samples. The gradient penalty (GP) prevents gradient explosion in the discriminator. ； In the formula: To calculate the mean for all real cavitation signal samples; To traverse all noise, the mean of the generated samples output by the discriminator is calculated; To calculate the mean of the gradient penalty term; This is a real data sample; For discriminator output; Output to the generator; This is the penalty coefficient; Interpolation between real samples and generated samples; The L2 norm of the discriminator for the interpolated samples; Temporal loss: ； In the formula: For L1 loss; This is the correlation loss; For envelope loss; For zero-crossing rate loss; This is a local extremum loss; Spectral loss: ； In the formula: This is the amplitude L1 loss; For logarithmic amplitude spectrum loss; For power spectral loss; This refers to the amplitude envelope loss; Multi-resolution STFT loss: ； In the formula: M represents the number of STFT resolutions; For the amplitude spectrum loss of STFT; For the logarithmic amplitude spectrum loss of the STFT; Statistical feature loss: ； In the formula: This is a true cavitation signal; To generate a cavitation signal; , These are the mean values ​​of the actual cavitation signal and the generated cavitation signal, respectively. , These are the standard deviations of the actual cavitation signal and the generated cavitation signal, respectively. , These are the skewnesses of the actual cavitation signal and the generated cavitation signal, respectively. , These are the kurtosis of the actual cavitation signal and the generated cavitation signal, respectively.

5. The method according to claim 1, characterized in that, In step four, during the first stage of training, the training ratio of the discriminator to the generator is 5:1, the learning rate of the generator in the Adam optimizer is 0.0001, and the learning rate of the discriminator is 0.0004; the learning rate of the second stage of training is 0.00001.

6. The method according to claim 1, characterized in that, Step five is as follows: Based on the convergent conditional WGAN-GP generator, original candidate samples are generated by inputting random noise and statistical condition vectors of real samples. Then, a multi-dimensional post-processing with time-domain matching and frequency-domain constraints is constructed. Cross-correlation time-domain alignment, Hilbert envelope amplitude matching, and sliding window local scaling calibration are performed in sequence. Finally, phase preservation and accurate amplitude matching are achieved by frequency-domain amplitude spectrum replacement. This ensures that the generated samples are completely consistent with the real cavitation signal in terms of time-domain morphology, envelope features, local waveforms, and frequency-domain energy distribution, which significantly improves the engineering fidelity of the samples. ； In the formula: The optimal time-domain offset; It is a cross-correlation function; This is a sample of a real cavitation signal; To generate cavitation signal samples; Candidate offset; ； ； In the formula: For the signal envelope; For Hilbert transform; These are the generated samples after envelope matching; It is a numerically stable term; This is a real signal; This is the initial signal generation; ； ； In the formula: These are real samples; The frequency domain amplitude spectrum of a real sample; The imaginary unit represents the phase. The frequency domain phase angle of the signal generated after local calibration; Frequency domain signal after amplitude replacement; This is the inverse fast Fourier transform.

7. The method according to claim 1, characterized in that, The SCWGAN model is a statistical conditional Wasserstein generative adversarial network that combines the conditional control capabilities of CGAN with the training stability of WGAN-GP, thus avoiding the mode collapse problem.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images