A Fault Diagnosis Method for Vacuum Pumps in the Strong Electromagnetic Environment of Fusion Reactors

Abstract

The present invention relates to the technical field of vacuum pump fault diagnosis, and particularly to a vacuum pump fault diagnosis method for the strong electromagnetic environment of a fusion reactor. Its technical solution includes the following steps: building a vacuum pump fault test bench with adjustable electromagnetic field strength, collecting the vibration acceleration signals, three-phase current signals and electromagnetic field strength signals of the vacuum pump under different electromagnetic field strengths through the test bench to form a simulation environment fault data set; during the operation of the fusion reactor, collecting the operation data and electromagnetic interference spectrum of the vacuum pump in real time. By simulating the strong electromagnetic environment of the fusion reactor and combining with real operation data, the present invention innovatively solves the problem of the failure of traditional vacuum pump fault diagnosis methods under electromagnetic interference, significantly improves the adaptability and reliability of the diagnosis model, effectively overcomes the limitations of small sample data by using data augmentation and cross-domain joint training techniques, realizes the accurate identification of various fault types, and provides an important technical guarantee for the safe and stable operation of the fusion device.

Description

Technical Field

[0001] The present invention relates to the technical field of vacuum pump fault diagnosis, and particularly to a vacuum pump fault diagnosis method for a strong electromagnetic environment of a fusion reactor. Background Art

[0002] The vacuum pump group is one of the auxiliary key devices of the fusion device. Whether it is the EAST (Experimental Advanced Superconducting Tokamak) fusion device in our country, or the largest fusion device ITER (International Thermonuclear Experimental Reactor) under construction internationally, including the future CFETR (China Fusion Engineering Test Reactor) fusion device in our country, a huge vacuum pump group is required to provide a good vacuum environment for the fusion device. The total operating power of the fusion device and system usually exceeds 30 MW, and the pulse experiment time for each time is 10 - 1000 s, or even longer. As a result, an extremely complex and harsh electromagnetic field environment will be generated. The electromagnetic field in the fusion reactor is mainly generated by the interaction of strong currents and magnetic fields. These environmental factors not only affect the stability of the fusion reaction, but also pose a threat to the safety of the equipment. Many measurement and control devices are around or even inside the fusion device, very close to the electromagnetic interference source, and are extremely vulnerable to electromagnetic interference, such as many diagnostic devices, solenoid valves for controlling valves, probes for electromagnetic measurement, and devices of the quench protection system, etc. Among them, the influence of electromagnetic interference on the vacuum pump is significant, which may cause problems such as abnormal solenoid valves of control components, degradation of equipment performance, or stoppage of work. In such a complex electromagnetic environment, the vacuum pump group, as the core device for maintaining the internal vacuum state of the fusion reactor, plays a crucial role. The performance of the vacuum pump group is directly related to the stability and reaction efficiency of the plasma. Therefore, the monitoring and fault diagnosis of its operating state are particularly important. Therefore, when using the vacuum pump, it is necessary to fully consider the influence of electromagnetic interference and take corresponding measures to reduce the influence of interference on the equipment and system. The GTEM cell is a test system designed based on the principle of coaxial and asymmetric matrix transmission lines, which can generate a relatively large radio frequency electromagnetic field intensity and is suitable for simulating the strong electromagnetic interference environment of the fusion reactor.

[0003] Therefore, the present application proposes a vacuum pump fault diagnosis method for a strong electromagnetic environment of a fusion reactor. Summary of the Invention

[0004] The object of the present invention is to propose a vacuum pump fault diagnosis method for a strong electromagnetic environment of a fusion reactor in view of the problem that there is no monitoring and fault diagnosis method for the operating state of the vacuum pump group in the background art.

[0005] Technical solution of the present invention: A method for diagnosing faults of a vacuum pump facing the strong electromagnetic environment of a fusion reactor, comprising the following steps:

[0006] (1) Build a vacuum pump fault test bench with adjustable electromagnetic field strength, and collect the vibration acceleration signals, three-phase current signals and electromagnetic field strength signals of the vacuum pump under different electromagnetic field strengths through the test bench to form a simulation environment fault data set;

[0007] (2) During the operation of the fusion reactor, collect the operation data and electromagnetic interference spectrum of the vacuum pump in real time, and combine the on-site fault data of the fusion reactor to construct a fusion environment fault data set;

[0008] (3) Perform data augmentation on the fusion environment fault data set: Generate synthetic data for minority class samples based on the k-nearest neighbor algorithm, and generate noise-resistant samples through a generative adversarial network (GAN) to expand the sample quantity and diversity of the fusion environment fault data set;

[0009] (4) Perform domain adaptation joint training on the simulation environment fault data set and the enhanced fusion environment fault data set: Reduce the distribution difference between the two types of data sets through feature extraction and feature space mean alignment methods, and design a joint loss function for model optimization;

[0010] (5) Based on the jointly trained model, it is used for fault diagnosis of the vacuum pump under the strong electromagnetic environment of the fusion reactor.

[0011] Optionally, the collection of the simulation environment fault data set in step (1) specifically includes: placing the faulty vacuum pump in a GTEM cell (GTEM cell: a test device for simulating the strong electromagnetic environment of a fusion reactor), applying continuous wave electromagnetic interference of 100 Hz - 50 MHz, and changing the vacuum pump load through pulse width speed regulation, and continuously collecting acceleration waveforms, current waveforms and electromagnetic field strength data at a sampling rate of 30 kHz.

[0012] Optionally, the data augmentation based on the k-nearest neighbor algorithm in step (3) is specifically: Calculate the k-nearest neighbors of the minority class samples, determine the sampling multiple according to the quantity ratio of the majority class to the minority class, and generate new samples through linear interpolation:

[0013] Wherein, are the horizontal and vertical coordinates of the new sample point, are the horizontal and vertical coordinates of the original sample, are the horizontal and vertical coordinates of the neighboring points, is a random number between zero and one.

[0014] Optionally, the loss function of the feature space mean alignment in step (4) is:

[0015] Among them, is the fusion environment fault data set, is the simulation environment fault data set, Feature extraction operation, is the original data of the th sample in the fusion environment data set, is the original data of the th sample in the simulation environment data set, is the Frobenius norm.

[0016] Optionally, the fault types of the vacuum pump include at least one of drive motor fault, pump body air leakage fault, blade wear fault, bearing damage fault, and pump oil shortage fault.

[0017] Optionally, the generator of the generative adversarial network (GAN) is used to synthesize noisy fault signals, and the discriminator is used to distinguish real samples from synthetic samples, and the noise resistance of the generated samples is improved through adversarial training.

[0018] Optionally, the joint loss function includes a multi-classification loss term and a domain difference loss term, specifically: Among them, is the multi-classification loss of the vacuum pump fault data set, is the difference loss between the fusion environment data and the simulation environment data. Through joint training, the total loss value is reduced to complete the training of the fault diagnosis model, is a hyperparameter. In the domain adaptation joint training, the optimal value of the hyperparameter is determined through grid search to balance the weights of the multi-classification loss and the domain difference loss.

[0019] Optionally, the test bench includes multi-class fault vacuum pumps, triaxial acceleration sensors, current transformers, broadband electromagnetic sensors, and acquisition boards.

[0020] Optionally, the sampling rate of the acquisition board is not less than 10 kHz, and it is used to continuously collect the vibration, current, and electromagnetic field signals of the vacuum pump.

[0021] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0022] By simulating the electromagnetic environment of the fusion reactor through the GTEM cell and combining with real data training, the robustness of the model under complex electromagnetic conditions is significantly improved.

[0023] Using k-nearest neighbor interpolation and GAN to generate noise-resistant samples effectively expands the small sample data and improves the generalization ability of the model.

[0024] Through domain adaptation joint training, the distribution difference between simulation and real data is reduced, and the cross-scene diagnosis accuracy is improved.

[0025] Multi-sensor synchronous acquisition with a sampling rate of ≥10 kHz is adopted to ensure signal integrity under strong interference and enhance the ability to extract fault features.

[0026] Through simulating the strong electromagnetic environment of a fusion reactor and combining with real operation data, the present invention innovatively solves the problem of the failure of traditional vacuum pump fault diagnosis methods under electromagnetic interference, significantly improves the adaptability and reliability of the diagnosis model, and uses data augmentation and cross-domain joint training technologies to effectively overcome the limitations of small-sample data, realizes the accurate identification of various fault types, and provides an important technical guarantee for the safe and stable operation of the fusion device. Brief Description of the Drawings

[0027] Figure 1 It is a flowchart of a vacuum pump fault diagnosis method for the strong electromagnetic environment of a fusion reactor. Detailed Embodiments

[0028] The technical solutions of the present invention will be further described below in conjunction with the drawings and specific embodiments.

[0029] Embodiment

[0030] As Figure 1 shown, a vacuum pump fault diagnosis method for the strong electromagnetic environment of a fusion reactor proposed by the present invention will be described in detail below.

[0031] I. Construction of the fault test bench: Design and manufacture multiple faulty vacuum pumps, including various types of faulty vacuum pumps such as drive motor faults, pump body air leakage faults, blade wear faults, bearing damage faults, pump oil shortage faults, etc. Build a fault test bench, equipped with two three-axis acceleration sensors and three current transformers for measuring the vibration acceleration signals of the vacuum pump body and the drive motor part, as well as the three-phase current signals. Equip a broadband electromagnetic sensor for collecting the electromagnetic field strength near the vacuum pump. Equip a collection host, and perform continuous waveform signal collection with a sampling rate of at least 10 kHz in combination with a 24-bit high-precision collection board. Design vacuum pumps with various types of faults such as drive motor faults, pump body air leakage, and blade wear to ensure that the fault diagnosis method can adapt to various failure modes that may occur in actual applications and improve the comprehensiveness of diagnosis. Equip three-axis acceleration sensors, current transformers, broadband electromagnetic sensors and high-precision collection equipment to synchronously obtain multi-modal signals such as vibration, current, and electromagnetic environment, providing rich data support for subsequent fault feature analysis and avoiding the one-sidedness of single-signal diagnosis. Adopt a 24-bit high-precision collection board and a sampling rate of ≥10 kHz to ensure that signal details are accurately captured, reduce data distortion, and lay a foundation for identifying subtle differences in fault features.

[0032] II. Collection of the vacuum pump fault dataset in the fusion environment: During the steady-state operation of the fusion reactor and under other working conditions, collect the operation datasets of various faulty vacuum pumps nearby. At the same time, deploy broadband electromagnetic sensors at key positions of the vacuum pumps to record the electromagnetic interference spectrum in real time and construct a dynamic baseline database of the electromagnetic environment. Through the collected vacuum pump fault data and normal datasets in the fusion experiments, obtain the fusion environment fault dataset. Collect data directly under the steady-state and other working conditions of the fusion reactor to ensure that the dataset reflects the complex electromagnetic interference and load changes in the actual operating environment and improve the adaptability of the model to the real scenario. Record the interference spectrum in real time through the broadband electromagnetic sensors and establish a baseline database, which can be used to distinguish the characteristics of normal electromagnetic fluctuations and fault-related signals and enhance the anti-interference ability of fault identification. Collect fault and normal operation data to form a complete dataset, providing positive and negative samples for subsequent model training and avoiding diagnostic misjudgments caused by data deviation.

[0033] III. Collection of the vacuum pump fault dataset in the simulation environment: Place the vacuum pump fault test bench in the GTEM cell and use the GTEM cell test system to apply electromagnetic environments with different field strengths in a closed space. Before each experiment, the GTEM cell is not started and runs without electromagnetic interference for 10 minutes. At the same time, vacuum pumps of different fault types are running, and acceleration waveform data and current waveform data are continuously collected at a sampling rate of 30 kHz. Subsequently, the directional coupler of the GTEM cell starts to generate continuous waves of 100 Hz - 50 MHz in sequence, with a single duration of 10 minutes, and collect the operation data of different faulty vacuum pumps. Change the load of the vacuum pump through pulse width speed regulation and obtain the corresponding vibration and current data. Through the above operations, obtain the simulation fault dataset of vacuum pumps under different working conditions in different electromagnetic environments. Using the GTEM cell to apply electromagnetic interference with different field strengths and frequency bands (100 Hz - 50 MHz) has strong controllability and can systematically analyze the influence of electromagnetic interference on different fault signals, filling the deficiency that electromagnetic interference is difficult to artificially adjust in the real fusion environment.

[0034] In this embodiment, by changing the load through pulse width speed regulation and combining the comparative collection of the no-interference (initial 10 minutes) and interference scenarios, obtain the fault data under different loads and different electromagnetic intensities, greatly increasing the diversity of the dataset and enhancing the generalization ability of the model to complex working conditions. The 30 kHz sampling rate ensures that high-frequency signals (such as transient vibrations of bearing damage and blade wear) are effectively recorded, providing a guarantee for the extraction of high-frequency fault characteristics.

[0035] IV. Enhancement of Fusion Environment Samples: Considering that the number of fusion environment samples is much less than that of simulation environment samples, it is necessary to expand the number of samples here and enhance the noise to ensure sample diversity. To this end, a fusion environment sample space is defined. By calculating the distances of all samples in the minority class sample set, the k-nearest neighbors are obtained. According to the ratio N of the number of majority class to minority class, the sampling multiple N is obtained. For each minority class sample , several samples are randomly selected from its k-nearest neighbors. Suppose the selected ones are . By calculating any point between the sample point and the selected point as the new sample point, the calculation is as follows

[0036] where are the horizontal and vertical coordinates of the new sample point, are the horizontal and vertical coordinates of the original sample, are the horizontal and vertical coordinates of the neighboring point, is a random number between zero and one. Repeat the calculation multiple times until the samples are balanced. For the minority fault samples (such as rare fault types) in the fusion environment, new samples are generated through k-nearest neighbor interpolation to balance the data volume of each category, avoid the "majority class bias" caused by sample imbalance in the model, and improve the recognition accuracy of minority fault categories. The new samples are generated on the line connecting the original sample and its neighbor, ensuring that the newly added data conforms to the manifold structure of the original data and avoiding the distribution deviation caused by randomly generated samples.

[0037] Construct a general generative adversarial network GAN, where the generator is G and the discriminator is D. The generator generates signals with noise and the fusion environment dataset and inputs them into the discriminator to train the GAN. Through the trained GAN network, samples with noise resistance ability are generated and expanded into the fusion environment dataset.

[0038] Domain Adaptive Training: Use the simulation environment fault dataset as the auxiliary dataset , and perform joint training with the fusion environment fault dataset . First, feature extraction operations need to be performed on the two datasets to obtain the feature vectors in the feature space. By aligning the means of the feature vectors of the fusion environment and simulation environment datasets in the feature space, the distribution difference between the two types of datasets is reduced.

[0039] where is the fusion environment fault dataset, is the simulation environment fault dataset, is the feature extraction operation, is any sample in the dataset, is the feature space.

[0040] where is the sum of the number of categories of manufactured faulty vacuum pumps and normal vacuum pumps, represents the sample belongs to the true label of the class, represents the sample belongs to the predicted probability of the class.

[0041] Among them, is the multi-classification loss of the vacuum pump fault dataset, is the difference loss between the fusion environment data and the simulation environment data. Through joint training, the total loss value is reduced to complete the training of the fault diagnosis model, is a hyperparameter, and the optimal value is obtained through grid search. Through GAN training, the generator learns the noise distribution law of the real signal, generates enhanced samples containing actual electromagnetic noise, and improves the robustness of the model under strong electromagnetic interference. Supplement the extreme noise scenario data that is difficult to directly collect in the fusion environment, expand the noise coverage of the dataset, and enhance the generalization ability of the model. Through feature space mean alignment (calculate ), the distribution difference between the simulation environment and the fusion environment data is reduced, so that the knowledge of the simulation data (auxiliary dataset) can be effectively transferred to the real scenario, alleviating the problem of insufficient samples in the fusion environment. Combining the classification loss ( ) and the domain difference loss ( ), while ensuring the fault classification accuracy, the model is forced to learn domain-invariant features, improving the stability of cross-environment diagnosis. Determine the δ value through grid search to achieve the optimal balance of the two types of losses, avoid over-biasing towards simulation data or real data, and ensure the reliability of the model in practical applications.

[0042] The present invention significantly improves the robustness of the fault diagnosis model under complex electromagnetic interference by building a vacuum pump fault test bench with adjustable electromagnetic field strength (such as GTEM cell) to simulate the continuous wave electromagnetic interference environment of 100 Hz - 50 MHz in the fusion reactor and combining the joint training of real fusion reactor operation data and simulation data. Traditional methods are vulnerable to signal distortion interference due to not considering the influence of strong electromagnetic interference, while the present invention effectively suppresses the interference of electromagnetic noise on the diagnosis result through feature space mean alignment and anti-noise sample generation.

[0043] Aiming at the problems of few fault data samples and class imbalance in the fusion environment, the present invention proposes a linear interpolation data augmentation method based on the k-nearest neighbor algorithm and combines the generative adversarial network (GAN) to generate synthetic samples with anti-noise ability. By expanding the diversity and quantity of minority class samples, the data skew problem is significantly alleviated, enabling the model to more comprehensively learn the features of different fault modes during training and improving the generalization performance.

[0044] Through the domain adaptation joint training technology, the feature spaces of the simulation environment dataset (auxiliary domain) and the fusion environment dataset (target domain) are mean-aligned, and a joint loss function including multi-classification loss ( ) and domain difference loss ( ) is designed. This method significantly reduces the difference in the distributions of the two types of data, enabling the model to make full use of the richness of the simulation data and at the same time adapt to the complexity of the real fusion environment, improving the diagnostic accuracy in cross-domain scenarios.

[0045] A high-precision acquisition board with a sampling rate of not less than 10 kHz, combined with broadband electromagnetic sensors, triaxial acceleration sensors and current transformers, realizes the full-band synchronous acquisition of vibration, current and electromagnetic field signals. Even under strong electromagnetic interference, the signal integrity can still be ensured, providing a reliable data basis for fault feature extraction. Compared with traditional methods, it improves the signal resolution and effectively captures tiny fault features.

[0046] It supports the accurate diagnosis of multiple types of faults such as drive motor faults, pump body air leakage, blade wear, bearing damage and pump oil shortage. By constructing a test bench including various fault types and combining dynamic load regulation (pulse width speed regulation) to simulate actual working conditions, the model can identify the composite fault modes under complex working conditions and improve the diagnostic coverage rate.

[0047] The above specific embodiments are only several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for diagnosing faults of a vacuum pump for a strong electromagnetic environment of a fusion reactor, characterized in that, It includes the following steps: Build a vacuum pump fault test bench with adjustable electromagnetic field strength. Collect the vibration acceleration signals, three-phase current signals, and electromagnetic field strength signals of the vacuum pump under different electromagnetic field strengths through the test bench to form a simulation environment fault data set. The collection of the simulation environment fault data set specifically includes: Place the faulty vacuum pump in a GTEM cell, apply continuous wave electromagnetic interference of 100 Hz - 50 MHz, and change the vacuum pump load through pulse width speed regulation. Continuously collect the acceleration waveform, current waveform, and electromagnetic field strength data at a sampling rate of 30 kHz. During the operation of the fusion reactor, collect the operation data and electromagnetic interference spectrum of the vacuum pump in real time, and combine the on-site fault data of the fusion reactor to construct a fusion environment fault data set. Perform data augmentation on the fusion environment fault data set: Generate synthetic data for minority class samples based on the k-nearest neighbor algorithm, and generate noise-resistant samples through a generative adversarial network. Conduct domain adaptive joint training on the simulation environment fault data set and the augmented fusion environment fault data set: Narrow the distribution difference between the two types of data sets through feature extraction and feature space mean alignment methods, and design a joint loss function for model optimization. The joint loss function includes a multi-classification loss term and a domain difference loss term, specifically: Loss=Loss A +δLoss B Among them, Loss A is the multi-classification loss of the vacuum pump failure data set, and Loss B is the difference loss between the fusion environment data and the simulation environment data. Through joint training, the total loss value of Loss is reduced to complete the training of the fault diagnosis model. δ is a hyperparameter. In the domain adaptation joint training, the optimal value of the hyperparameter δ is determined through grid search to balance the weights of the multi-classification loss and the domain difference loss; Based on the model after joint training, it is used for fault diagnosis of the vacuum pump under the strong electromagnetic environment of the fusion reactor.

2. The method for diagnosing faults of a vacuum pump for a strong electromagnetic environment of a fusion reactor according to claim 1, wherein, The data augmentation based on the k-nearest neighbor algorithm is specifically: Calculate the k-nearest neighbors of minority class samples, determine the sampling multiple according to the quantity ratio of the majority class to the minority class, and generate new samples through linear interpolation. (x new ,y new ) = (x, y) + rand(0, 1) * ((x n -x), (y n -y)) Among them, (x new , y new ) are the horizontal and vertical coordinates of the new sample point, (x, y) are the horizontal and vertical coordinates of the original sample, (x n , y n ) are the horizontal and vertical coordinates of the neighboring point, and rand(0, 1) is a random number between zero and one.

3. A method for diagnosing faults of a vacuum pump facing the strong electromagnetic environment of a fusion reactor according to claim 1, characterized in that, The loss function of the feature space mean alignment is: Among them, N s is the fusion environment fault dataset, N t is the simulation environment fault dataset, F is the feature extraction operation, D s (x i ) is the original data of the i-th sample in the fusion environment dataset, D t (x i ) is the original data of the i-th sample in the simulation environment dataset, ||·|| F is the Frobenius norm.

4. A vacuum pump fault diagnosis method for a strong electromagnetic environment of a fusion reactor according to claim 1, characterized in that, The fault types of the vacuum pump include at least one of drive motor fault, pump body air leakage fault, blade wear fault, bearing damage fault, and pump oil shortage fault.

5. A method for diagnosing faults of a vacuum pump for a strong electromagnetic environment of a fusion reactor according to claim 1, characterized in that, The generator of the generative adversarial network is used to synthesize faulty signals with noise, and the discriminator is used to distinguish real samples from synthetic samples. The noise-resistant ability of the generated samples is improved through adversarial training.

6. The vacuum pump fault diagnosis method for the strong electromagnetic environment of a fusion reactor according to claim 1, characterized in that, The test bench includes various types of faulty vacuum pumps, triaxial acceleration sensors, current transformers, broadband electromagnetic sensors, and acquisition boards.

7. A method for diagnosing faults of a vacuum pump for a strong electromagnetic environment of a fusion reactor according to claim 6, characterized in that, The sampling rate of the acquisition board is not less than 10 kHz and is used to continuously collect the vibration, current, and electromagnetic field signals of the vacuum pump.

Images