A sample enhancement method, a fault diagnosis method and a device for fault diagnosis of a permanent magnet synchronous motor

Abstract

This invention belongs to the field of fault diagnosis technology, specifically relating to a sample augmentation method, fault diagnosis method, and apparatus for fault diagnosis of permanent magnet synchronous motors. The sample augmentation method includes: S1, acquiring real training samples for training a fault diagnosis model of a permanent magnet synchronous motor; S2, in the frequency domain, superimposing a random amplitude related to the amplitude of each frequency point of the real training samples as the amplitude of the generated sample, and using the phase of the real training samples as the phase of the generated sample to obtain the generated sample; if the difference between the generated sample and its corresponding real training sample does not meet the preset requirements, then S2 is repeated; otherwise, the generated sample is used as the first virtual training sample for training the fault diagnosis model of the permanent magnet synchronous motor. This invention solves the technical problem of poor generation performance in the field of small-sample fault diagnosis due to the lack of a theoretical mechanism for data augmentation in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a sample enhancement method, fault diagnosis method and device for fault diagnosis of permanent magnet synchronous motors. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in aerospace, intelligent manufacturing, rail transportation, and defense industries due to their high efficiency, high power density, and excellent control performance. As PMSMs operate under complex conditions such as high loads and overloads in various fields, motor redundancy gradually decreases, directly increasing the operational risk and failure probability of PMSM systems. Demagnetization faults, eccentricity faults, inter-turn short-circuit faults, and main insulation faults are common faults in PMSMs that can lead to serious consequences, causing not only economic losses but also catastrophic accidents. Fault diagnosis technology can effectively improve the operational reliability of PMSMs, thereby preventing fault escalation and reducing failure losses. Therefore, multi-fault diagnosis research targeting demagnetization faults, eccentricity faults, inter-turn short-circuit faults, and main insulation faults has significant theoretical and engineering application value.

[0003] Deep learning networks possess significant technical advantages in solving multi-fault diagnosis problems of permanent magnet synchronous motors due to their outstanding ability to extract high-dimensional features. However, deep learning networks are highly dependent on existing data, which leads to a significant decline in their diagnostic performance under small sample conditions, severely limiting the applicability of deep learning-based fault diagnosis methods.

[0004] Data augmentation is a primary technique for improving the diagnostic performance of deep learning networks with limited sample sizes. Its core lies in artificially and appropriately expanding the training dataset to compensate for the insufficient amount of original data, thereby enhancing the model's generalization ability and robustness. Currently, Generative Adversarial Networks (GANs) are commonly used tools for data augmentation in various application scenarios. GANs learn through adversarial training between the generator and discriminator, with the theoretical goal of converging to Nash equilibrium. However, their training dynamics lack a clear theoretical mechanism, and the generation quality is highly dependent on the training data, making it difficult to control stably. Although some data augmentation methods have clear theoretical support, they are essentially limited to the data level and do not address the underlying mathematical and physical mechanisms. This results in significant room for improvement in the generative performance of existing methods in the field of small-sample fault diagnosis of permanent magnet synchronous motors. Summary of the Invention

[0005] The purpose of this invention is to provide a sample enhancement method, a fault diagnosis method, and an apparatus for fault diagnosis of permanent magnet synchronous motors, in order to solve the technical problem that the lack of theoretical mechanism for data enhancement in the prior art leads to poor generation performance in the field of small sample fault diagnosis.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a sample enhancement method for fault diagnosis of permanent magnet synchronous motors, the method comprising:

[0007] S1. Obtain real training samples for training the fault diagnosis model of permanent magnet synchronous motor;

[0008] S2. In the frequency domain, a random amplitude related to the amplitude is superimposed on the amplitude of each frequency point of the real training sample as the amplitude of the generated sample, and the phase of the real training sample is used as the phase of the generated sample to obtain the generated sample.

[0009] If the difference between the generated sample and its corresponding real training sample does not meet the preset requirements, then S2 is executed again; otherwise, the generated sample is used as the first virtual training sample for training the permanent magnet synchronous motor fault diagnosis model.

[0010] In one possible implementation, the method further includes: S3, adjusting the frequency distribution of the first virtual training sample in the frequency domain so that the amplitude is not zero only at the frequency point corresponding to the target speed and the amplitude is zero at the other frequency points, so as to obtain a second virtual training sample for training the fault diagnosis model of the permanent magnet synchronous motor that simulates the signal at the target speed.

[0011] In one possible implementation, the magnitude of the generated sample in S2 is obtained as follows:

[0012]

[0013] in, To generate sample frequency The amplitude at that point; Frequency of real training samples The amplitude at that point; This is the amplitude adjustment factor; for A random number between [a certain number of points].

[0014] In one possible implementation, re-executing S2 includes: adjusting the amplitude adjustment coefficient. After obtaining the value, S2 will be executed again.

[0015] In one possible implementation, the frequency distribution of the first virtual training sample in S3 is adjusted as follows:

[0016]

[0017]

[0018] in, The fundamental frequency corresponding to the rotational speed of the real training sample; The fundamental frequency corresponding to the target rotational speed; These are the octaves of the real training samples and the second virtual training samples, respectively. These are the rotational speeds of the real training samples and the second virtual training samples, respectively. For the second virtual training sample frequency The amplitude at that point; For the second virtual training sample frequency Phase at; Frequency of the first virtual training sample Phase at; Frequency of the first virtual training sample The amplitude at that point.

[0019] In one possible implementation, the degree of difference between the generated sample and its corresponding real training sample is measured by the root mean square error.

[0020] In one possible implementation, the pre-set requirement is that the distance between the root mean square error of the generated sample and its corresponding real training sample and the threshold α satisfies the pre-set allowable error.

[0021] The threshold α is β times the average of the root mean square errors between all pairs of real training samples, where 0 < β < 1.

[0022] In one possible implementation, virtual training samples in the time domain that can be directly used for training are obtained as follows: the virtual training samples include a first virtual training sample and a second virtual training sample:

[0023]

[0024] in, These are virtual training samples within the time domain; The sampling rate of the sampling device for real training samples; The amplitude of the virtual training sample; The phase of the virtual training sample.

[0025] To address the aforementioned technical problems, a second aspect of the present invention provides a fault diagnosis method for a permanent magnet synchronous motor, the method comprising:

[0026] (1) Obtain the fault diagnosis parameters of the permanent magnet synchronous motor under test;

[0027] (2) Input the fault diagnosis parameters into the pre-trained fault diagnosis model to obtain the fault diagnosis results of the permanent magnet synchronous motor under test;

[0028] The training samples used to train the fault diagnosis model include training samples obtained by means of any possible implementation of the first aspect of the invention.

[0029] To address the aforementioned technical problems, a third aspect of the present invention provides a fault diagnosis device for a permanent magnet synchronous motor, comprising a processor for executing a computer program to implement the steps of the method in any possible implementation of the second aspect of the present invention.

[0030] The beneficial effects of this invention are as follows: Based on the underlying mechanism of the influence of small sample conditions on the fault characteristic quantities of permanent magnet synchronous motors, this invention generates enhanced samples by adjusting the time-frequency characteristics of the characteristic quantities. This improves the generation efficiency and performance of enhanced samples. Specifically, noise mainly affects the amplitude at each frequency point, and the simulation of noise is mainly achieved by adjusting the amplitude at each frequency point. A random number related to the amplitude is superimposed on the amplitude at each frequency point to simulate random noise. This invention solves the technical problem of poor generation performance in the field of small sample fault diagnosis due to the lack of theoretical mechanism for data augmentation in existing technologies.

[0031] In addition, since permanent magnet synchronous motors mostly adjust their speed by changing the power supply frequency, the fundamental frequency of the signal is different for different speeds. This makes the frequency domain characteristics of fault features at different speeds significantly different. To simulate virtual samples at different speeds, the frequency distribution of the signal needs to be adjusted: the amplitude of some frequency points is set to 0 and the amplitude of that point is superimposed on the target frequency point, so that the signal only has amplitude at the frequency point corresponding to the target speed, thus realizing the simulation of signals at different speeds. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the fault diagnosis method of the present invention for permanent magnet synchronous motors.

[0033] Figure 2 This is a flowchart of the sample enhancement method for an embodiment of the fault diagnosis method for permanent magnet synchronous motors of the present invention.

[0034] Figure 3 This is a schematic diagram of the amplitude adjustment method in an embodiment of the fault diagnosis method for permanent magnet synchronous motors of the present invention;

[0035] Figure 4 This is a schematic diagram of the frequency distribution adjustment method in an embodiment of the fault diagnosis method for permanent magnet synchronous motors of the present invention.

[0036] Figure 5The simulated samples for the fault diagnosis method of permanent magnet synchronous motor of the present invention only simulate the diagnostic results under different sample augmentation ratios when noise is involved.

[0037] Figure 6 The diagnostic accuracy of simulated noise samples with different augmentation ratios under 1% training samples in the implementation of the fault diagnosis method for permanent magnet synchronous motors of the present invention is shown.

[0038] Figure 7 This invention provides a method for fault diagnosis of permanent magnet synchronous motors using 1% of the training samples at different k values. d Diagnostic accuracy of simulated noise samples;

[0039] Figure 8 This invention provides a single-speed training sample with different k values ​​for the fault diagnosis method of permanent magnet synchronous motors. d The diagnostic accuracy of virtual samples;

[0040] Figure 9 This is a schematic diagram of the device architecture in an embodiment of the fault diagnosis device for permanent magnet synchronous motors of the present invention. Detailed Implementation

[0041] This invention is based on the underlying mechanism of the influence of small sample conditions on the fault characteristics of permanent magnet synchronous motors. It generates enhanced samples by adjusting the time-frequency characteristics of these characteristics. This improves the generation efficiency and performance of the enhanced samples. Specifically, noise mainly affects the amplitude at each frequency point, and noise simulation is achieved primarily by adjusting the amplitude at each frequency point. A random number related to the amplitude is superimposed on the amplitude at each frequency point to simulate random noise. This invention solves the technical problem of poor generation performance in the field of small sample fault diagnosis due to the lack of a theoretical mechanism for data augmentation in existing technologies.

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] Implementation methods for fault diagnosis of permanent magnet synchronous motors:

[0044] To address the technical problems existing in the background technology, this embodiment proposes a small-sample fault diagnosis method for permanent magnet synchronous motors based on time-frequency feature reconstruction and deep learning, such as... Figure 1 As shown, this method is based on the underlying mechanism of the influence of small sample conditions on the fault characteristics of permanent magnet synchronous motors, and generates enhanced samples by adjusting the time-frequency characteristics of the characteristics. This improves the generation efficiency and performance of enhanced samples. The meanings of the parameters in this implementation are shown in Table 1:

[0045] Table 1

[0046]

[0047]

[0048] The method includes steps 1 to 5, wherein steps 1 to 3 involve virtual sample reconstruction, and the detailed process is as follows: Figure 2 As shown:

[0049] Step 1: Perform time-frequency analysis on the existing real feature quantity samples. Using the electrical cycle as the sliding window width, perform Fourier analysis on the signal to obtain the amplitude and phase of the feature quantity at different frequencies for each electrical cycle. At the same time, calculate the root mean square error between the time-domain signals of different electrical cycles as an indicator to evaluate the degree of difference of the reconstructed virtual samples, and use its average value (denoted as RMSE_A) as the main reference to set the target value of the difference of the reconstructed virtual samples.

[0050] Common physical meanings of characteristic quantities include, but are not limited to, current, voltage, vibration, magnetic field strength, etc. The actual characteristic quantity used in this embodiment is the voltage signal output by the detection coil installed at a specific location.

[0051] First, Fourier decomposition of the existing real feature samples is performed using a sliding window to obtain the amplitude and phase parameters required in step 2. The Fourier transform considering the sliding window can be expressed as:

[0052]

[0053] in, It is time. It's frequency. It is a real sample signal. It is a window function, the window type is a rectangular window, and the window length is one electrical cycle. It is the kernel function of the Fourier transform.

[0054] After Fourier transform, a complex matrix is ​​obtained. The magnitude of each element in the matrix represents the amplitude of the signal at that time and frequency. The angle of each element represents the phase of the signal at that time and frequency, as shown in the following formula:

[0055]

[0056]

[0057] in, and Represent complex matrices respectively The real and imaginary parts.

[0058] Secondly, the average root mean square error between all pairwise electrical periodic signals needs to be calculated as the main basis for the target parameters of the virtual sample. The root mean square error between them is calculated as follows:

[0059]

[0060] in, For signal and The root mean square error between the two signals is denoted by n, where n is the number of signal sampling points. The smaller the root mean square error, the better the similarity between the two signals.

[0061] Step 2: Adjust the amplitude and phase at different frequencies. By changing the frequency distribution, the amplitude of the low-speed harmonic point is transferred to the high-speed harmonic point to simulate the high-speed sample. A random number related to the amplitude ratio is superimposed on the amplitude at each frequency point to simulate the effect of random noise on the feature quantity.

[0062] Step 2 mainly involves adjusting the amplitude and frequency distribution of the real signal to make it different from the real signal, thereby simulating the impact of real noise and rotation speed on the data.

[0063] Amplitude adjustment method is illustrated as follows: Figure 3 As shown: Since noise primarily affects the amplitude at each frequency point, noise simulation is mainly achieved by adjusting the amplitude at each frequency point. The amplitude adjustment method involves superimposing a random number related to the amplitude onto the amplitude at each frequency point to simulate random noise. This is done by adjusting the correlation coefficient. It can adjust the root mean square error between virtual and real samples, thereby controlling the noise level of virtual samples. This correlation coefficient... Defined as follows:

[0064]

[0065] in, This is the amplitude adjustment coefficient (i.e., the correlation coefficient). For the signal at frequency Amplitude at that point In frequency Add a random number value to the location. Adjust the frequency point. The amplitude and phase at a point can be expressed as:

[0066]

[0067]

[0068] in, for Random numbers between To reconstruct virtual samples at frequency The amplitude at that point. To reconstruct virtual samples at frequency The phase at that point, compared to the real sample at frequency The phase remains consistent at that point.

[0069] The frequency distribution adjustment method is illustrated as follows: Figure 4 As shown: Since permanent magnet synchronous motors mostly adjust their speed by changing the power supply frequency, the fundamental frequency of the signal differs at different speeds. This results in significant differences in the frequency domain characteristics of fault features at different speeds. Therefore, after amplitude adjustment, if it is necessary to simulate virtual samples at different speeds, the frequency distribution of the signal needs to be adjusted. For example... Figure 4 As shown, by setting the amplitude of some frequency points to 0 and superimposing the amplitude of those points onto the target frequency point, the signal will only exhibit amplitude at the frequency point corresponding to the target rotational speed. This allows for the simulation of signals at different rotational speeds. This adjustment method can be expressed as:

[0070]

[0071]

[0072] in, This is the fundamental frequency corresponding to the rotational speed of the actual sample. To reconstruct the fundamental frequency corresponding to the simulated rotational speed of the virtual sample, These are the octaves of the real samples and the reconstructed virtual samples, respectively. The rotation speeds are for the real sample and the reconstructed virtual sample, respectively.

[0073] It should be noted that the speed simulation is based on samples that have already been simulated for noise. This is because, under this method, the samples after speed simulation correspond one-to-one with the samples before simulation. Multiple sets of data need to be generated through the noise model so that multiple sets of data can also be generated after noise simulation.

[0074] The speed simulation targets samples that have already been simulated for noise. This is because, under this method, the samples after speed simulation correspond one-to-one with the samples before simulation. Multiple sets of data need to be generated through the noise model so that multiple sets of data can also be generated after noise simulation.

[0075] Step 3: Based on the adjusted frequency points, amplitudes, and phases, construct cosine signals corresponding to different frequency points. Superimposing the cosine signals from all frequency points reconstructs a set of virtual samples. Continuously change the scaling factor. Adjust the root mean square error between the reconstructed virtual samples and the real samples to control the noise level of the virtual samples, and finally achieve a value of about 0.15 times the average root mean square error obtained in step 1 (error of ±5%).

[0076]

[0077] in, For each reconstructed electrical cycle, a virtual sample, The sampling rate is the sampling rate of the real sample sampling device. By concatenating the virtual samples of each electrical cycle according to the time-domain relationships in the original signal, a complete virtual sample of the feature quantities can be obtained.

[0078] Step 4: Merge the reconstructed virtual samples with the real samples to form the augmented training sample dataset, and train the augmented training sample dataset using a deep learning network.

[0079] To verify whether using simulated noisy samples in training can effectively improve diagnostic performance, samples with different rotational speeds were not simulated when reconstructing virtual samples. Instead, random samples, comprising 1% of the training data, were used for verification. The diagnostic accuracy (average of diagnostic accuracy of ten training samples) and its fluctuation (fluctuation of ten samples) at different sample augmentation ratios when the value is 0.3 are shown below. Figure 5 As shown, the sample augmentation factor is the ratio between the training sample size after augmentation and the training sample size before augmentation. A sample augmentation factor of 1 indicates that no sample augmentation was performed.

[0080] Figure 5 In the study, the diagnostic accuracy was significantly improved after data augmentation using the proposed method. The diagnostic accuracy showed a non-monotonic relationship with the augmentation factor, indicating that simulated samples also interfered with the features of the original samples. The diagnostic accuracy reached its highest level of 72.06% with an augmentation factor of 8, an improvement of 12.33% compared to the method without augmentation. Simultaneously, the accuracy fluctuation showed a decreasing trend, indicating increased stability of the diagnostic results. This demonstrates that the virtual samples added to the training dataset effectively enhance the robustness of the comprehensive fault diagnosis method for permanent magnet synchronous motors to noise.

[0081] To further analyze the impact of noise-simulated samples on diagnostic performance, a comparison was made with 1% of random training samples under different conditions. Diagnostic performance after data augmentation: Figure 6 As shown. The diagnostic accuracy rate is the average of the diagnostic accuracy rates of ten training samples. Samples with different rotation speeds are not simulated when reconstructing virtual samples.

[0082] Figure 6 In the middle, virtual samples at different augmentation ratios The value has different effects on diagnostic accuracy. When the sample size is increased by 2 times, the diagnostic accuracy increases accordingly. The value shows a monotonically increasing trend. When the sample size is increased 8-fold, the diagnostic accuracy increases accordingly. The value exhibits a non-monotonic change, and in The value reaches its extreme at 0.3, at which point the root mean square error between the virtual and real samples is approximately 15% of the root mean square error between the real samples across different electrical cycles. This indicates that at different augmentation rates, the diagnostic accuracy is affected by the virtual sample value. The sensitivity of the value varies; the greater the augmentation factor, the better. The more sensitive the value, the easier it is to reach the extreme point.

[0083] Secondly, we need to verify whether using simulated rotational speed samples in training can effectively improve diagnostic performance. Since noise simulation is a prerequisite for rotational speed simulation in the proposed method, and virtual samples simulating different rotational speeds also include noise simulation, it is necessary to compare the diagnostic results of noise-simulated virtual samples and rotational speed-simulated virtual samples. The difference between the two represents the impact of rotational speed simulation on diagnostic performance. This verification was performed using a random sample representing 1% of the training data. The diagnostic accuracy at different sample augmentation folds when the value is 0.3 (the average diagnostic accuracy of ten training samples) is as follows. Figure 7 As shown.

[0084] Figure 7 In the training, the diagnostic accuracy of the simulated rotational speed virtual samples was slightly higher than that of the noise-simulated virtual samples, and both showed the same trend. However, as the augmentation ratio increased, the diagnostic accuracy advantage of the simulated rotational speed virtual samples gradually decreased compared to the noise-simulated virtual samples. Considering the large number of rotational speed types used in the database during the analysis, it is difficult for the expected value of the rotational speed type in the original samples to be 1 when randomly selecting training samples, thus failing to fully reflect the optimization effect of the simulated rotational speed samples. The training samples were artificially controlled to be of a single rotational speed and a single load to further analyze the impact of the simulated rotational speed samples on diagnostic performance. When the original samples were of a single rotational speed, the results were compared with those of different loads under 1% of the training samples. Diagnostic performance of virtual samples (sample augmentation 2x) as follows Figure 8 As shown.

[0085] Figure 8 In this case, only one set of virtual samples needs to be generated for each set of real samples. The value can be set to 0 to completely eliminate the impact of noise simulation on diagnostic accuracy. It can be observed that compared to the unamplified training samples, the virtual samples with simulated rotational speeds (…) When the value is 0, the diagnostic accuracy improves significantly to 9.58%. This proves that the speed simulation method proposed in this paper can effectively improve the diagnostic accuracy under single speed conditions. As the value increases, the effects of noise and rotational speed are simulated simultaneously. A correct diagnosis indicates that the value increases first and then decreases. The diagnostic accuracy peaked at a value of 0.3, reaching 52.2%.

[0086] Implementation method of fault diagnosis device for permanent magnet synchronous motor:

[0087] A schematic diagram of the fault diagnosis device for permanent magnet synchronous motors in this embodiment is shown below. Figure 9 As shown, the system includes a memory, a processor, a system bus, and a computer program stored in the memory. The processor and memory communicate and exchange data via the system bus. The processor executes the computer program to implement the steps of the fault diagnosis method for permanent magnet synchronous motors described in the embodiments of the present invention. The specific fault diagnosis method for permanent magnet synchronous motors has been described in sufficient detail in the above embodiments and will not be repeated here. The processor can be a microprocessor (MCU) or other processing device; the memory can be any type of memory that stores information using electrical energy, such as non-volatile storage media (including computer programs, databases), or other types of memory.

[0088] Implementation of a sample augmentation method for fault diagnosis of permanent magnet synchronous motors:

[0089] A sample enhancement method for fault diagnosis of permanent magnet synchronous motors, which is equivalent to steps 1-3 in the implementation of the fault diagnosis method for permanent magnet synchronous motors, and specifically refers to steps 1-3 in the implementation of the fault diagnosis method for permanent magnet synchronous motors, includes the following steps:

[0090] S1. Obtain real training samples for training the fault diagnosis model of permanent magnet synchronous motor.

[0091] S2. In the frequency domain, a random amplitude related to the amplitude is superimposed on the amplitude of each frequency point of the real training sample as the amplitude of the generated sample, and the phase of the real training sample is used as the phase of the generated sample to obtain the generated sample.

[0092] The amplitude of the generated sample is obtained as follows:

[0093]

[0094] in, To generate sample frequency The amplitude at that point; Frequency of real training samples The amplitude at that point; This is the amplitude adjustment factor; for A random number between [a certain number of points].

[0095] If the difference between the generated sample and its corresponding real training sample does not meet the preset requirements, the amplitude adjustment coefficient will be adjusted. After obtaining the value, S2 is executed again; otherwise, the generated sample is used as the first virtual training sample for training the permanent magnet synchronous motor fault diagnosis model.

[0096] The degree of difference between the generated sample and its corresponding real training sample is measured by the root mean square error.

[0097] The pre-set requirement is that the distance between the root mean square error of the generated sample and its corresponding real training sample and the threshold α meets the pre-set allowable error.

[0098] The threshold α is β times the average of the root mean square errors between all pairs of real training samples, where 0 < β < 1.

[0099] S3. In the frequency domain, adjust the frequency distribution of the first virtual training sample so that the amplitude is not zero only at the frequency point corresponding to the target speed and the amplitude is zero at the other frequency points, so as to obtain the second virtual training sample for training the fault diagnosis model of the permanent magnet synchronous motor that simulates the signal at the target speed.

[0100] Adjust the frequency distribution of the first virtual training sample as follows:

[0101]

[0102]

[0103] in, The fundamental frequency corresponding to the rotational speed of the real training sample; The fundamental frequency corresponding to the target rotational speed; These are the octaves of the real training samples and the second virtual training samples, respectively. These are the rotational speeds of the real training samples and the second virtual training samples, respectively. For the second virtual training sample frequency The amplitude at that point; For the second virtual training sample frequency Phase at; Frequency of the first virtual training sample Phase at; Frequency of the first virtual training sample The amplitude at that point.

[0104] Virtual training samples in the time domain that can be directly used for training are obtained in the following manner, wherein the virtual training samples include a first virtual training sample and a second virtual training sample:

[0105]

[0106] in, These are virtual training samples within the time domain; The sampling rate of the sampling device for real training samples; The amplitude of the virtual training sample; The phase of the virtual training sample.

[0107] This invention has the following characteristics:

[0108] This invention is based on the underlying mechanism of the influence of small sample conditions on the fault characteristics of permanent magnet synchronous motors. It generates enhanced samples by adjusting the time-frequency characteristics of these characteristics. This improves the generation efficiency and performance of the enhanced samples. Specifically, noise mainly affects the amplitude at each frequency point, and noise simulation is achieved primarily by adjusting the amplitude at each frequency point. A random number related to the amplitude is superimposed on the amplitude at each frequency point to simulate random noise. This invention solves the technical problem of poor generation performance in the field of small sample fault diagnosis due to the lack of a theoretical mechanism for data augmentation in existing technologies.

[0109] In addition, since permanent magnet synchronous motors mostly adjust their speed by changing the power supply frequency, the fundamental frequency of the signal is different for different speeds. This makes the frequency domain characteristics of fault features at different speeds significantly different. To simulate virtual samples at different speeds, the frequency distribution of the signal needs to be adjusted: the amplitude of some frequency points is set to 0 and the amplitude of that point is superimposed on the target frequency point, so that the signal only has amplitude at the frequency point corresponding to the target speed, thus realizing the simulation of signals at different speeds.

[0110] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still make modifications to the technical solutions described in the foregoing embodiments without creative effort, or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A sample augmentation method for fault diagnosis of permanent magnet synchronous motors, characterized in that, The method includes: S1. Obtain real training samples for training the fault diagnosis model of permanent magnet synchronous motor; S2. In the frequency domain, a random amplitude related to the amplitude is superimposed on the amplitude of each frequency point of the real training sample as the amplitude of the generated sample, and the phase of the real training sample is used as the phase of the generated sample to obtain the generated sample. If the difference between the generated sample and its corresponding real training sample does not meet the preset requirements, then S2 is executed again; otherwise, the generated sample is used as the first virtual training sample for training the permanent magnet synchronous motor fault diagnosis model.

2. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 1, characterized in that, The method further includes: S3, adjusting the frequency distribution of the first virtual training sample in the frequency domain so that the amplitude is not zero only at the frequency point corresponding to the target speed and the amplitude is zero at the other frequency points, so as to obtain a second virtual training sample for training the fault diagnosis model of the permanent magnet synchronous motor that simulates the signal at the target speed.

3. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 1, characterized in that, The amplitude of the generated sample in S2 is obtained in the following way: in, To generate sample frequency The amplitude at that point; Frequency of real training samples The amplitude at that point; This is the amplitude adjustment factor; for A random number between [a certain number of points].

4. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 3, characterized in that, Re-execution of S2 includes: adjusting the amplitude adjustment coefficient. After obtaining the value, S2 will be executed again.

5. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 2, characterized in that, In S3, the frequency distribution of the first virtual training sample is adjusted as follows: in, The fundamental frequency corresponding to the rotational speed of the real training sample; The fundamental frequency corresponding to the target rotational speed; These are the octaves of the real training samples and the second virtual training samples, respectively. These are the rotational speeds of the real training samples and the second virtual training samples, respectively. For the second virtual training sample frequency The amplitude at that point; For the second virtual training sample frequency Phase at; Frequency of the first virtual training sample Phase at; Frequency of the first virtual training sample The amplitude at that point.

6. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 1, characterized in that, The degree of difference between the generated sample and its corresponding real training sample is measured by the root mean square error.

7. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 6, characterized in that, The pre-set requirement is that the distance between the root mean square error of the generated sample and its corresponding real training sample and the threshold α meets the pre-set allowable error. The threshold α is β times the average of the root mean square errors between all pairs of real training samples, where 0 < β < 1.

8. The sample enhancement method for fault diagnosis of permanent magnet synchronous motors according to claim 2, characterized in that, Virtual training samples in the time domain that can be directly used for training are obtained in the following manner, wherein the virtual training samples include a first virtual training sample and a second virtual training sample: in, These are virtual training samples within the time domain; The sampling rate of the sampling device for real training samples; The amplitude of the virtual training sample; The phase of the virtual training sample.

9. A fault diagnosis method for permanent magnet synchronous motors, characterized in that, The method includes: (1) Obtain the fault diagnosis parameters of the permanent magnet synchronous motor under test; (2) Input the fault diagnosis parameters into the pre-trained fault diagnosis model to obtain the fault diagnosis results of the permanent magnet synchronous motor under test; The training samples used to train the fault diagnosis model include training samples obtained using the sample augmentation method for fault diagnosis of permanent magnet synchronous motors as described in any one of claims 1 to 8.

10. A fault diagnosis device for a permanent magnet synchronous motor, comprising a processor, characterized in that, The processor is used to execute a computer program to implement the steps of the fault diagnosis method for a permanent magnet synchronous motor as described in claim 9.

Images