A method and apparatus for enhanced diagnosis of motor faults based on recurrent generative adversarial networks

Abstract

This invention discloses a motor fault enhancement diagnosis method and apparatus based on a cyclic generative adversarial network. The method includes constructing a motor fault sample migration generation model based on a cyclic generative adversarial network to realize the migration from a similar sample domain to a target sample domain, thereby augmenting the target sample and achieving enhanced diagnosis of motor faults. The target sample domain is the motor fault sample domain under the operating conditions for which fault diagnosis is required, and the similar sample domain is the fault sample domain of the motor under other operating conditions with the same fault mode.

Description

Technical Field

[0001] This invention relates to the field of prognostics and health management (PHM) technology, and in particular to a method and apparatus for enhanced diagnosis of motor faults based on Cycle-consistent Generative Adversarial Network (CycleGAN). Background Technology

[0002] In circuit diagrams, an electric motor is represented by the letter M. It typically provides torque to mechanical devices, converting electrical energy into the mechanical energy required by the device. Its applications are widespread in various aspects of industrial production. Motor malfunctions can cause anything from minor property damage to serious safety hazards, making motor fault diagnosis crucial. Motor fault diagnosis technology can detect early-stage motor failures, allowing technicians to understand the motor's current operating status before serious consequences arise, facilitating timely and targeted repairs. Therefore, fault diagnosis during the daily use of motors is highly effective and necessary, possessing significant engineering importance.

[0003] Motors generally consist of magnetic poles, rotors, bearings, and housings. When faults occur in different parts, the vibration signals output by the motor exhibit different characteristics. Therefore, the different characteristics of these vibration signals are often used to detect motor fault modes and achieve fault diagnosis. To achieve high-accuracy motor fault diagnosis, traditional methods often require a large number of motor fault samples. However, in actual industrial production, obtaining effective motor fault samples is difficult and costly, resulting in a small number of target samples for fault diagnosis in many cases. This leads to poor diagnostic results using traditional methods. Therefore, how to achieve the best possible fault diagnosis effect under limited target sample conditions is a crucial problem that needs to be solved. Currently, methods based on Generative Adversarial Networks (GANs) to augment fault samples are the main means to address the problem of insufficient samples. However, traditional GANs can only utilize information from the target samples themselves to augment the samples, resulting in low-quality generated samples when the number of target samples is small, making it difficult to effectively enhance motor fault diagnosis. Considering that each fault mode often has fault samples under multiple operating conditions, and that motor fault samples under different operating conditions within the same fault mode share similar characteristics, utilizing the characteristics of these similar samples can improve the effectiveness of fault diagnosis. The current approach primarily involves using transfer learning to pre-train a diagnostic model using similar samples, and then fine-tuning the model using target samples. However, traditional transfer learning methods are highly sensitive to the similarity between samples and the transfer method used. When the similarity between the target and similar samples is insufficient, or when the transfer learning method is unsuitable, the transfer effect is poor, and sometimes even negative transfer occurs, which can actually reduce the diagnostic performance. Therefore, research on how to effectively utilize similar sample information for enhanced motor fault diagnosis is particularly important when the target sample is small and similar samples are available to assist in diagnosis. Summary of the Invention

[0004] The technical problem solved by the solution provided in the embodiments of the present invention is how to effectively utilize the information of similar samples to enhance the diagnosis of motor faults when the target sample is small and there are similar samples to assist in the diagnosis.

[0005] An enhanced fault diagnosis method for motors based on recurrent generative adversarial networks (GANs) is provided according to an embodiment of the present invention, comprising:

[0006] Obtain target motor fault samples and similar motor fault samples for each motor fault mode, and select the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples for each motor fault mode.

[0007] The target motor fault sample and the most similar motor fault sample are used to train the pre-constructed motor target sample transfer generation model based on recurrent generative adversarial network to obtain the trained motor target sample transfer generation model based on recurrent generative adversarial network.

[0008] By inputting the most similar motor fault sample for each motor fault mode into the trained motor target sample transfer generation model based on recurrent generative adversarial network, the generated target motor fault sample for each motor fault mode is obtained.

[0009] The pre-built motor fault enhancement diagnosis model based on convolutional neural network is trained using the target motor fault samples and the generated target motor fault samples to obtain the trained motor fault enhancement diagnosis model based on convolutional neural network.

[0010] The fault data of the motor to be diagnosed is obtained, and the fault data is input into the trained motor fault enhancement diagnosis model based on convolutional neural network for fault diagnosis processing to obtain the fault type of the motor to be diagnosed.

[0011] Preferably, the step of selecting the most similar motor fault sample with the highest similarity to the target motor fault sample from similar motor fault samples in each motor fault mode includes:

[0012] Multiple MMD values ​​are obtained by calculating the maximum mean difference (MMD) between each similar motor fault sample and the target motor fault sample under each motor fault mode;

[0013] The smallest MMD value is selected from the plurality of MMD values, and the similar motor fault sample corresponding to the smallest MMD value is taken as the most similar motor fault sample.

[0014] Preferably, the pre-built motor target sample transfer generation model based on a recurrent generative adversarial network includes:

[0015] Based on the target motor fault sample and the most similar motor fault sample, a motor target sample transfer generation model based on a recurrent generative adversarial network (RGAN) is constructed, which includes a recurrent generative adversarial network structure and a total loss function.

[0016] The recurrent generative adversarial network structure includes: a first generator G and a first discriminator D. y The first generative adversarial network; containing a second generator F and a second discriminator D. x The second generative adversarial network;

[0017] The total loss function includes: a first generative adversarial network loss function, a second generative adversarial network loss function, and a cycle consistency loss function.

[0018] Preferably, the step of training a pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network using the target motor fault sample and the most similar motor fault sample to obtain the trained motor target sample transfer generation model based on a recurrent generative adversarial network includes:

[0019] By converting the target motor fault sample and the most similar motor fault sample respectively, we obtain the target motor fault sample in the form of a two-dimensional grayscale image and the most similar motor fault sample in the form of a two-dimensional grayscale image.

[0020] The target motor fault sample in the form of a two-dimensional grayscale image and the most similar motor fault sample in the form of a two-dimensional grayscale image are used as inputs to train the motor target sample transfer generation model based on a recurrent generative adversarial network several times, so as to obtain the trained motor target sample transfer generation model based on a recurrent generative adversarial network.

[0021] Preferably, the pre-built motor fault enhancement diagnostic model based on convolutional neural networks includes:

[0022] Based on the target motor fault samples and the generated target motor fault samples, a motor fault enhancement diagnosis model based on convolutional neural networks, including a convolutional neural network structure and a loss function, is constructed.

[0023] Preferably, the step of training a pre-built motor fault enhancement diagnosis model based on a convolutional neural network using the target motor fault samples and the generated target motor fault samples to obtain a trained motor fault enhancement diagnosis model based on a convolutional neural network includes:

[0024] The target motor fault sample and the generated target motor fault sample are used as inputs to train the motor fault enhancement diagnosis model based on convolutional neural network several times to obtain a trained motor fault enhancement diagnosis model based on convolutional neural network.

[0025] According to an embodiment of the present invention, a motor fault enhancement and diagnosis device based on a recurrent generative adversarial network is provided, comprising:

[0026] The sample acquisition module is used to acquire target motor fault samples and similar motor fault samples for each motor fault mode, and to select the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples for each motor fault mode.

[0027] The first construction and training module is used to train a pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network using the target motor fault sample and the most similar motor fault sample, so as to obtain a trained motor target sample transfer generation model based on a recurrent generative adversarial network.

[0028] The sample generation module is used to obtain the generated target motor fault samples for each motor fault mode by inputting the most similar motor fault samples for each motor fault mode into the trained motor target sample transfer generation model based on recurrent generative adversarial network.

[0029] The second construction and training module is used to train the pre-constructed motor fault enhancement diagnosis model based on convolutional neural network using the target motor fault samples and the generated target motor fault samples, so as to obtain the trained motor fault enhancement diagnosis model based on convolutional neural network.

[0030] The fault diagnosis module is used to acquire fault data of the motor to be diagnosed, and input the fault data into the trained motor fault enhancement diagnosis model based on convolutional neural network for fault diagnosis processing to obtain the fault type of the motor to be diagnosed.

[0031] The solution provided by the embodiments of the present invention has the following beneficial effects:

[0032] (1) Using similar sample transfer to generate target samples, while effectively learning the feature information of the target samples, the effective feature information of the similar samples is also retained, which improves the quality of the generated motor fault samples.

[0033] (2) It is not sensitive to the similarity between the samples being transferred, and there will be no negative transfer phenomenon due to insufficient similarity in transfer learning, making it more practical.

[0034] (3) It can augment the target sample by using similar sample information under the condition of extremely small target sample size, effectively reducing the adverse effects of insufficient target sample size on data-driven motor fault diagnosis methods, and greatly improving the accuracy of motor fault diagnosis under extremely small target sample size. Attached Figure Description

[0035] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to understand the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0036] Figure 1 This is a flowchart of an enhanced motor fault diagnosis method based on CycleGAN provided in an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the motor target sample transfer generation model based on CycleGAN provided in an embodiment of the present invention;

[0038] Figure 3 These are grayscale images of target samples under various fault modes provided in the embodiments of the present invention;

[0039] Figure 4 It is a grayscale image of the most similar sample under each fault mode provided in the embodiments of the present invention;

[0040] Figure 5 This is a schematic diagram illustrating the changing trend of the MMD of the normal mode target sample and the generated sample with the number of training rounds, provided in an embodiment of the present invention.

[0041] Figure 6 This is a schematic diagram illustrating the changing trends of the target sample and generated sample MMD of the rotor shaft bending mode with the number of training rounds provided in this embodiment of the invention;

[0042] Figure 7 This is a schematic diagram illustrating the changing trends of the bearing failure mode target sample and generated sample MMD with the number of training rounds provided in this embodiment of the invention;

[0043] Figure 8 This is a schematic diagram illustrating the changing trends of the target sample and generated sample MMD of the rotor imbalance mode with the number of training rounds provided in this embodiment of the invention;

[0044] Figure 9 This is a schematic diagram illustrating the changing trend of the target sample and generated sample MMD of the rotor bar breakage mode with the number of training rounds provided in this embodiment of the invention;

[0045] Figure 10 This is a schematic diagram comparing the power spectra of the target sample and the generated sample in the normal mode provided in an embodiment of the present invention;

[0046] Figure 11 This is a schematic diagram comparing the power spectrum of the target sample and the generated sample in the rotor shaft bending mode provided in an embodiment of the present invention;

[0047] Figure 12 This is a schematic diagram comparing the power spectrum of the target sample and the generated sample of the bearing failure mode provided in an embodiment of the present invention;

[0048] Figure 13 This is a schematic diagram comparing the power spectrum of the target sample and the generated sample in the rotor imbalance mode provided in an embodiment of the present invention;

[0049] Figure 14 This is a schematic diagram comparing the power spectrum of the target sample and the generated sample in the rotor bar breakage mode provided in an embodiment of the present invention;

[0050] Figure 15This is a schematic diagram illustrating the change in training loss for fault diagnosis before target sample augmentation, provided in an embodiment of the present invention.

[0051] Figure 16 This is a schematic diagram illustrating the change in training loss for fault diagnosis after target sample augmentation, provided in an embodiment of the present invention.

[0052] Figure 17 This is a flowchart of a motor fault enhancement and diagnosis method based on a recurrent generative adversarial network provided in an embodiment of the present invention;

[0053] Figure 18 This is a schematic diagram of a motor fault enhancement and diagnosis device based on a cyclic generative adversarial network provided in an embodiment of the present invention. Detailed Implementation

[0054] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments described below are only for illustration and explanation of the present invention and are not intended to limit the present invention.

[0055] Figure 17 This is a flowchart of a motor fault enhancement and diagnosis method based on a recurrent generative adversarial network provided in an embodiment of the present invention, as shown below. Figure 17 As shown, it includes:

[0056] Step S101: Obtain the target motor fault sample and similar motor fault sample for each motor fault mode, and select the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples for each motor fault mode.

[0057] Step S102: Use the target motor fault sample and the most similar motor fault sample to train the pre-constructed motor target sample transfer generation model based on recurrent generative adversarial network to obtain the trained motor target sample transfer generation model based on recurrent generative adversarial network.

[0058] Step S103: By inputting the most similar motor fault sample for each motor fault mode into the trained motor target sample transfer generation model based on recurrent generative adversarial network, the generated target motor fault sample for each motor fault mode is obtained.

[0059] Step S104: Use the target motor fault sample and the generated target motor fault sample to train the pre-built motor fault enhancement diagnosis model based on convolutional neural network to obtain the trained motor fault enhancement diagnosis model based on convolutional neural network.

[0060] Step S105: Obtain the fault data of the motor to be diagnosed, and input the fault data into the trained motor fault enhancement diagnosis model based on convolutional neural network for fault diagnosis processing to obtain the fault type of the motor to be diagnosed.

[0061] Specifically, the step of selecting the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples under each motor fault mode includes: calculating the maximum mean difference (MMD) value between each similar motor fault sample and the target motor fault sample under each motor fault mode to obtain multiple MMD values; selecting the minimum MMD value from the multiple MMD values, and taking the similar motor fault sample corresponding to the minimum MMD value as the most similar motor fault sample.

[0062] Furthermore, the pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network includes: constructing a motor target sample transfer generation model based on a recurrent generative adversarial network, comprising a recurrent generative adversarial network structure and a total loss function, based on the target motor fault sample and the most similar motor fault sample; wherein the recurrent generative adversarial network structure comprises: a first generator G and a first discriminator D. y The first generative adversarial network; containing a second generator F and a second discriminator D. x The second generative adversarial network; the total loss function includes: the first generative adversarial network loss function, the second generative adversarial network loss function, and the cycle consistency loss function.

[0063] Further, the step of training the pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network using the target motor fault sample and the most similar motor fault sample to obtain the trained motor target sample transfer generation model based on a recurrent generative adversarial network includes: transforming the target motor fault sample and the most similar motor fault sample respectively to obtain the target motor fault sample in two-dimensional grayscale image form and the most similar motor fault sample in two-dimensional grayscale image form; using the target motor fault sample in two-dimensional grayscale image form and the most similar motor fault sample in two-dimensional grayscale image form as input to the motor target sample transfer generation model based on a recurrent generative adversarial network for several training iterations to obtain the trained motor target sample transfer generation model based on a recurrent generative adversarial network.

[0064] The pre-built motor fault enhancement diagnosis model based on convolutional neural networks includes: constructing a motor fault enhancement diagnosis model based on convolutional neural networks, which includes a convolutional neural network structure and a loss function, based on the target motor fault sample and the generated target motor fault sample.

[0065] Furthermore, the step of training the pre-constructed motor fault enhancement diagnosis model based on convolutional neural networks using the target motor fault samples and the generated target motor fault samples to obtain the trained motor fault enhancement diagnosis model based on convolutional neural networks includes: using the target motor fault samples and the generated target motor fault samples as inputs to the motor fault enhancement diagnosis model based on convolutional neural networks for several training iterations to obtain the trained motor fault enhancement diagnosis model based on convolutional neural networks.

[0066] Figure 18 This is a schematic diagram of a motor fault enhancement and diagnosis device based on a cyclic generative adversarial network provided in an embodiment of the present invention, as shown below. Figure 18 As shown, it includes: a sample acquisition module 201, used to acquire target motor fault samples and similar motor fault samples for each motor fault mode, and to select the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples for each motor fault mode; a first construction and training module 202, used to train a pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network using the target motor fault samples and the most similar motor fault samples to obtain a trained motor target sample transfer generation model based on a recurrent generative adversarial network; and a sample generation module 203, used to input the most similar motor fault samples for each motor fault mode into the model. The trained motor target sample transfer generation model based on recurrent generative adversarial networks is used to obtain generated target motor fault samples for each motor fault mode; the second construction and training module 204 is used to train a pre-built motor fault enhancement diagnosis model based on convolutional neural networks using the target motor fault samples and the generated target motor fault samples to obtain a trained motor fault enhancement diagnosis model based on convolutional neural networks; the fault diagnosis module 205 is used to acquire fault data of the motor to be diagnosed and input the fault data into the trained motor fault enhancement diagnosis model based on convolutional neural networks for fault diagnosis processing to obtain the fault type of the motor to be diagnosed.

[0067] It should be noted that the target motor fault sample is referred to as the target sample, and the generated target motor fault sample is referred to as the generated sample.

[0068] To improve the accuracy of motor fault diagnosis when the target sample for motor fault diagnosis is small and similar samples exist under other operating conditions, this invention proposes a motor fault enhancement diagnosis method and apparatus based on a recurrent generative adversarial network (RGAN). The advantage of RGANs is that they do not require paired training samples; only two domain samples are needed for transfer. This method constructs a motor fault sample transfer generation model based on RGANs to achieve transfer from the similar sample domain to the target sample domain, thus broadening the target sample and enhancing motor fault diagnosis. The target sample domain is the motor fault sample domain under the operating conditions requiring fault diagnosis, and the similar sample domain is the fault sample domain of the motor under other operating conditions with the same fault mode.

[0069] This invention proposes a motor fault enhancement and diagnosis method based on cyclic generative adversarial networks, such as... Figure 1 As shown, it includes:

[0070] Step 1: Based on the small sample data conditions of the target motor and the similar sample data conditions under various working conditions, conduct a transferability measurement of the target sample (target motor fault sample) and similar samples (similar motor fault samples), and select the most similar sample (most similar motor fault sample) to ensure that the selected similar sample and the target sample are similar enough to ensure the transfer quality.

[0071] Assume there are n motor fault modes, and for each fault mode, there are m similar samples under different operating conditions. For each type of motor fault mode, the target sample has a quantity of k0, the number of samples to be generated is k1, the number of each similar sample is k2, and the number of samples from the same source as the target sample as the test sample is k.

[0072] Furthermore, the transferability measure addresses the situation in actual industrial production where multiple similar samples often exist for each type of motor fault mode. It uses the Maximum Mean Discrepancy (MMD) method to select the most similar sample. For m similar samples, MMD analysis is performed with the target sample to obtain m MMD values. The sample with the smallest MMD value between the target and similar samples is selected as the most similar sample for transfer. In other words, the transferability measure is based on MMD, which measures the similarity between the similar sample domain and the target sample domain. The smaller the MMD value, the higher the similarity between the two domains, and the more suitable it is for sample domain transfer. In practical applications, for each type of motor fault mode, there are often multiple similar samples under various operating conditions. For these similar samples, their MMD values ​​with the target sample are calculated separately. The similar sample domain under the operating condition with the smallest MMD value is selected as the most similar sample domain for transfer, ensuring effective transfer.

[0073] The basic idea of ​​the MMD method used is as follows: based on samples from two distributions, find a continuous function f in the sample space, calculate the difference between the mean values ​​of the function on f for samples from different distributions, and obtain the mean difference (MD) between the two sample domains. Change f to maximize MD, and the MMD of the two sample domains can be obtained, thereby evaluating the similarity between the two domains. The specific calculation process is shown in formula (1):

[0074]

[0075] in, Represents the regenerated Hilbert space; P data(x) P data(y) These represent the distributions of similar samples and the target sample, respectively; n x n y , respectively, are the number of similar samples and the target sample; x and y are the samples in the similar sample domain X and the target sample domain Y, respectively; k(·,·) refers to the kernel function.

[0076] Step 2: Considering that the original motor fault signal is a one-dimensional vibration monitoring signal, the one-dimensional monitoring signals of the target sample and the most similar sample selected are converted into two-dimensional grayscale images, which are used as inputs for the subsequent motor sample migration generation model and motor fault diagnosis model.

[0077] Furthermore, in step 2, converting the one-dimensional vibration monitoring signal into a two-dimensional grayscale image first requires determining the side length M of the grayscale image. To ensure the distinguishability of fault features, M should not be too small. When using motor vibration signals as monitoring signals, M is generally taken as 64 based on experience. Then, the original one-dimensional monitoring signal of length l is segmented into segments of length M×M to obtain... One-dimensional samples, [·] represents rounding. Finally, each one-dimensional sample of length M×M is rearranged into a square two-dimensional signal with side length M, and the original signal monitoring value is converted into a gray value between 0 and 255 to obtain a two-dimensional gray map sample. That is to say, the specific method of converting the original one-dimensional vibration monitoring signal of the target sample and the most similar sample into a two-dimensional gray map is as follows: First, the original one-dimensional vibration monitoring signal is divided into one-dimensional vibration signals of equal length intervals according to the requirements; then, the segmented one-dimensional vibration signals are rearranged to convert into two-dimensional signals, and the original signal values ​​are converted into gray values. The original signal value of the two-dimensional signal is converted into gray value using the method of formula (2):

[0078]

[0079] Where P(j,k) is the transformed value of the coordinate (j,k) in the grayscale image, L(i) is the original intensity of each signal, round(·) represents rounding, and M is the side length of the transformed grayscale image. This method normalizes all intensity values ​​to between 0 and 255, converting them into pixel values ​​of the grayscale image, which facilitates subsequent network training.

[0080] Step 3: Construct a motor target sample transfer generation model based on CycleGAN to realize the transfer from the most similar sample domain to the target sample domain, and generate target samples from the most similar samples to achieve target sample augmentation.

[0081] Furthermore, a CycleGAN-based model for generating motor target samples through transfer is constructed, such as... Figure 2 As shown, CycleGAN contains two pairs of GANs. The generator of the first pair of GANs is G, and the discriminator is D. y The second pair of GANs uses F as the generator and D as the discriminator. x In the motor target sample transfer generation model, the first pair of GANs is mainly used to generate fault samples. The most similar sample x is generated by generator G to produce false target samples G(x), and then discriminator D... y The authenticity of G(x) is determined, and G(x) is then reconstructed back to the most similar sample F(G(x)) by the generator F. The cycle consistency loss between x and F(G(x)) is used to ensure that the fault sample G(x) generated by the transfer model retains the features of the original most similar sample, rather than being a fault sample randomly generated by the model. The motor fault sample in the form of the most similar sample domain X and the target sample domain Y in the form of a two-dimensional grayscale image after step 2 is used as the input of the motor target sample transfer generation model. After repeated debugging, the values ​​of each hyperparameter are obtained, and the generation of motor fault samples is realized.

[0082] Furthermore, after training the motor target sample transfer generation model, k1 motor fault samples from the most similar sample domain X are used as input to transfer and generate k1 motor fault samples from the target sample domain Y, thus achieving target sample augmentation. Using motor fault samples under different fault modes as input, motor fault samples for each fault mode can be generated. Empirically, when the target samples are small, k1 is more than 10 times the original number of target samples k0, resulting in better performance.

[0083] The specific steps for constructing a motor target sample transfer generation model based on CycleGAN are as follows: First, a CycleGAN structure is constructed based on the characteristics of the target sample and the most similar sample, determining the generator and discriminator structures of two pairs of GANs; then, the model's loss function is calculated, including the loss of the two pairs of GANs and the cycle consistency loss. The model's training process is then established through continuous iterations of gradient zeroing, forward propagation, error calculation, and backpropagation; finally, after repeated debugging to determine the hyperparameters for model training, the target sample is generated.

[0084] The basic principle of the motor target sample transfer generation model is as follows: Assuming the most similar sample domain and the target sample domain are X and Y respectively, the goal of the model is to generate samples in the Y domain from samples in the X domain. First, two pairs of GANs are constructed. The generator of the first pair of GANs is G, and the discriminator is D. y The second pair of GANs uses F as the generator and D as the discriminator. x Second, a sample x in the X domain is used by generator G to generate a fake sample G(x) in the Y domain; third, discriminator D... y The first step is to determine whether the forged sample G(x) is a genuine Y-domain sample. The second step involves reconstructing the forged sample G(x) into an X-domain sample F(G(x)) using a generator F. The cycle consistency loss between x and F(G(x)) is used to ensure that the transferred image G(x) retains the features of the X-domain sample, preventing it from being randomly generated. The process of transferring X-domain samples from Y-domain samples is similar, but since this method does not involve this process, it will not be described here.

[0085] The activation functions used in the target sample transfer generation model for motors include ReLU, Tanh, LeakyReLU, and Sigmoid.

[0086] The loss function in the motor target sample transfer generation model is constructed as follows:

[0087] The generator G aims to transform samples x in the most similar sample domain X into samples in the target sample domain Y, learning the X→Y mapping. Based on the cross-entropy loss, the loss function of equation (3) is constructed as follows:

[0088] L GAN (G,D y (X,Y)=E y～Y [ln D y (y)]+E x～X [ln(1-D y (G(x)))] (3)

[0089] Where E y～Y For all input target samples, E x～X D refers to all the most similar samples in the input.y (y) represents the probability that the discriminator considers the target sample y to belong to the target sample domain, D. y (G(x)) represents the probability that the discriminator of y determines that the motor fault sample G(x) generated by G after the most similar sample x belongs to the target sample domain.

[0090] The generator F aims to transform target samples in the target sample domain Y into the most similar samples in the most similar sample domain X, learning the mapping from Y to X. Based on the cross-entropy loss, the loss function of equation (4) is constructed as follows:

[0091] L GAN (F,D x (X,Y)=E x～X [ln D x (x)]+E y～Y [ln(1-D x (F(y)))] (4)

[0092] Where D x (x) represents the probability that the discriminator considers the most similar sample x to belong to the most similar sample domain.

[0093] D x (F(y)) represents the probability that the discriminator of x determines that the motor fault sample F(y) generated by F after the target sample y belongs to the most similar sample domain.

[0094] Since the two mappings between G and F are learned simultaneously, in order to prevent overlearning, an L1 loss is added to the loss function, as shown in equation (5) below:

[0095] L cyc (G,F)=E x～X [||F(G(x))-x||1]+E y～Y [||G(F(y))-y||1] (5)

[0096] Where ||F(G(x))-x||1 represents the difference between the reconstructed fault sample obtained by the most similar sample x after passing through generators G and F in sequence and the original fault sample x; ||G(F(y))-y||1 represents the difference between the reconstructed fault sample obtained by the target sample y after passing through generators F and G in sequence and the original fault sample y.

[0097] The final total loss function is shown in equation (6) below:

[0098] L(G,F,D x D y ) = L GAN (G,D y ,X,Y)+L GAN (F,D x,X,Y)+λL cyc (G,F) (6)

[0099] After determining the structure of the generator and discriminator of CycleGAN and constructing the loss function, the framework of the motor target sample transfer generation model was established. The next step is to train the constructed model. The training process includes two parts: first, selecting the CycleGAN hyperparameters, which refer to the manually input training parameters, not the weights and other parameters in the network; second, training the CycleGAN with the selected hyperparameters to establish the transfer generation capability. Finally, since the trained motor target sample transfer generation model requires a most similar sample as input to generate a fake target sample, multiple most similar samples are used as input to the trained transfer generation model to obtain multiple generated target samples, thereby broadening the target sample domain.

[0100] The steps for selecting hyperparameters are as follows: Under a specific set of hyperparameters, the network is trained using the most similar samples and target samples, and target samples are generated. Then, the similarity between the generated target samples and the real target samples is analyzed to determine the quality of the generated samples. The hyperparameters are continuously adjusted and the above process is repeated to find a set of hyperparameters that results in high-quality generated samples and a relatively stable generation process. The steps for training the CycleGAN network under the selected hyperparameters are as follows: Under the selected hyperparameters, the network is trained using the most similar samples and target samples, ensuring that the weights and other parameters in the network match the characteristics of the faulty samples used. The trained network can then transfer data from the most similar samples to generate target samples.

[0101] Step 4: Using the augmented target samples, construct a CNN-based motor fault enhancement and diagnosis model to achieve motor fault diagnosis under small sample conditions.

[0102] Furthermore, to achieve enhanced motor fault diagnosis, firstly, the enhanced motor fault diagnosis model is trained using n*(k0+k1) target samples of each fault mode after augmentation in step 3 as input, and the hyperparameter values ​​are obtained through repeated adjustments. Then, n*k test samples from the same source as the target samples are used to test the fault diagnosis effect. For each test sample, the enhanced motor fault diagnosis model identifies its fault mode and compares it with the known fault mode label of that sample. The total number of correctly identified test samples, k', is accumulated to obtain the accuracy of the enhanced motor fault diagnosis.

[0103] The specific steps for constructing a CNN-based motor fault diagnosis model are as follows: First, construct the structure of a convolutional neural network based on the characteristics of motor fault samples; then, determine the model's loss function, and build the model's training process through continuous iterations of gradient zeroing, forward propagation, error calculation, and backpropagation; finally, after repeated debugging to determine the model's final hyperparameters, perform motor fault diagnosis.

[0104] The CNN-based motor fault enhancement and diagnosis model comprises an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The number of convolutional and pooling layers can be determined according to requirements. Fully connected layers are generally located adjacent to the output layer. The network parameters are optimized through backpropagation. The input layer takes fault samples in the form of converted two-dimensional grayscale images. The main function of the convolutional layer is to perform convolution operations on the two-dimensional grayscale images of the fault samples to extract fault features. Finally, an activation function outputs a fault feature map to the next layer. Increasing the number of convolutional layers can achieve the extraction of more complex fault features. The pooling layer segments the motor fault feature map output by the convolutional layers, taking the maximum or average value to reduce the dimensionality of the fault features, facilitating better fault diagnosis.

[0105] Since this invention uses CNN for motor fault diagnosis, to better classify fault modes, the activation function used in the convolutional layer is ReLU, and the activation function used in the output layer is Softmax. Assume there is an array V of size k, where the i-th element is V. i The specific structure of the Softmax function is shown in equation (7):

[0106]

[0107] This function can establish a mapping between the output of a neuron and the interval (0,1), which can be directly regarded as the estimated probability that the input fault sample belongs to each fault category, thus realizing the fault diagnosis function. For example, when V is [1,3,6,9], the output of the Softmax function is [0.0003,0.0024,0.0472,0.9500].

[0108] The Cross Entropy Loss (CELoss) function is used as the loss function for the motor fault enhancement diagnosis model, as shown in Equation (8):

[0109]

[0110] Where N is the number of input motor fault samples, y (i) The label for faulty sample i. Let be the predicted probability of the i-th fault sample.

[0111] After determining the CNN network structure and constructing the loss function, the framework for the motor fault diagnosis model is established. The next step is to select hyperparameters for the constructed model. Under a specific set of hyperparameters, the CNN is trained using the augmented target samples from step 3 to obtain the diagnostic accuracy of the training samples (i.e., the augmented target samples) under this set of hyperparameters. The hyperparameters are continuously adjusted and the above process is repeated to find a set of hyperparameters that provides a high and stable diagnostic accuracy. After determining the required hyperparameters for the model, the CNN is trained again using the augmented target samples under this set of hyperparameters. This trained CNN can then be used as the motor fault diagnosis model to implement actual motor fault diagnosis. The actual motor fault diagnosis process is as follows: using motor fault data of unknown fault types as input to the trained motor fault diagnosis model, the motor fault diagnosis model will output the fault mode corresponding to the input data.

[0112] The technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0113] First, let's introduce the experimental data used in the case study:

[0114] The experimental platform used for data acquisition was the Mechanical Fault Simulation Test Bench (MFS), manufactured by Spectra Quest Corporation of the United States and belonging to the Predictive and Health Management National Defense Science and Technology Innovation Team of the School of Reliability and Systems Engineering, Beijing XX University. This test bench consists of components such as a driver, motor, hydraulic circuit and hydraulic pressure adjustment system, transmission shaft system, and vibration sensors. It can simulate common faults in mechanical products and be used for research such as fault diagnosis. The MFS motor controller has a speed control range of 0-60Hz, and the reluctance torque load is adjustable from 0 to 5 levels, where level 0 is no load, and the load torque range from 1 to 5 is 0.5 to 10 in-pounds (in·lbs), which is 0.0565 to 1.13 Nm.

[0115] Table 1 Fault modes and fault descriptions of sample data

[0116]

[0117] As shown in Table 1, five types of fault samples were selected: normal, bearing inner ring fault, central shaft bending, rotor imbalance, and rotor bar breakage. The fault descriptions for each type are shown in the table. Considering the motor's characteristic parameters, the experimental sampling frequency was set to 2.56 × 10⁻⁶. 4 Hz, with a sampling time of 192 seconds for each fault mode, totaling 4.96 × 10⁻⁶. 6 10 sampling points. Due to the high quality of the laboratory samples, and to simulate the characteristics of motor fault signals in actual industrial production, all raw signals were annotated with a mean of 0 and a variance of 2 × 10⁻⁶. -5The Gaussian distributed random signal is used as the noise signal.

[0118] Step 11: Based on the small sample data conditions of the target motor and the similar sample data conditions under various working conditions, conduct a transferability measurement of the target sample and similar samples, select the most similar samples, and ensure that the selected similar samples and the target samples are similar enough to ensure the quality of transfer.

[0119] The target samples are those under each fault mode with a motor speed of 30Hz and a load of 0. The similar samples are fault samples under other operating conditions under each fault mode, and there are four categories: motor speed of 30Hz and load 1, motor speed of 30Hz and load 3, motor speed of 35Hz and load 1, and motor speed of 40Hz and load 1. For each fault mode, the sample with the highest sample domain similarity, i.e., the smallest MMD, is selected as the most similar sample. The number of samples used for MMD analysis is 10 for the target sample and 10 for the most similar sample. The calculation process is based on equation (1). The final selection results and MMD values ​​are shown in Table 2.

[0120] Table 2. MMD results of similar samples and target samples under different failure modes.

[0121]

[0122] After performing MMD analysis and selecting the sample with the highest similarity to the target sample as the most similar sample for transfer, the quality of the generated sample can be effectively guaranteed, thereby ensuring the effect of enhanced diagnosis of motor faults.

[0123] Step 12: Considering that the original motor fault signal is a one-dimensional vibration monitoring signal, the one-dimensional monitoring signals of the target sample and the most similar sample selected are converted into two-dimensional grayscale images, which are used as inputs for the subsequent motor sample migration generation model and motor fault diagnosis model.

[0124] First, the sample format is as described in the experimental data introduction. For each type of fault mode, the target sample and the most similar sample are selected from the vibration signal in the X direction as the original signal, with a mean of 0 and a variance of 2 × 10⁻⁶. -5 After using the Gaussian distributed random signal as noise, the original signal was divided into 1200 groups, each with a length of 4096 sampling points, totaling 4.96 × 10⁻⁶. 6 One sampling point.

[0125] Second, the signals of each group of 4096 sampling points are converted into 64×64 grayscale images. The conversion process is carried out according to formula (2). First, the signals are rearranged into 64×64 form, and then the original vibration signal acceleration values ​​are converted into grayscale values.

[0126] Third, in order to simulate the difficulty and high cost of obtaining motor fault samples in actual industrial production, 10 sets were randomly selected from 1200 target samples in each fault mode as training sets, that is, only 10 sets of samples were used as target samples to train the transfer generation model, and the remaining 1190 sets were used as test samples; the 1200 most similar samples in each fault mode were used to train the transfer generation model.

[0127] After processing according to step 12 above, the grayscale images of the target sample and the most similar sample under each fault mode are as follows: Figure 3 and Figure 4 As shown, after completion, input the migration generation model established in step 13 to generate sample migration.

[0128] Step 13: Construct a motor target sample transfer generation model based on CycleGAN to realize the transfer from the most similar sample domain to the target sample domain, and generate target samples from the most similar samples to achieve target sample augmentation.

[0129] like Figure 2 As shown in the figure, the specific process of constructing a CycleGAN-based motor target sample transfer generation model based on the characteristics of the selected fault samples is as follows:

[0130] First, the generator and discriminator in CycleGAN are constructed. Since CycleGAN contains two pairs of GANs, there are two generators and two discriminators. The two generators have the same network structure. Based on the characteristics of the case samples, the final generator structure is shown in Table 3 below:

[0131] Table 3 CycleGAN Generator Structure

[0132]

[0133]

[0134] IN stands for Instance Normalization, used to reduce overfitting features during training. The generator consists of an input layer, two convolutional layers, six residual layers, two deconvolutional layers, and an output layer. The convolutional layers extract features from the input image, the residual layers transform and combine various image features, and the deconvolutional layers generate the desired image based on the features. The number of residual layers can be selected according to factors such as image resolution and complexity. The convolutional and deconvolutional layers use the ReLU activation function, which has fast convergence and does not suffer from gradient vanishing when the input is greater than 0, making it suitable for use in convolutional operations. The residual layers contain two convolutional processes, using ReLU and Identity activation functions respectively. The output layer uses the Tanh activation function, which has good stability and is suitable for use in the output layer.

[0135] The discriminator network structures in both pairs of GANs are identical. Based on the characteristics of the case samples, the final discriminator structures are shown in Table 4 below:

[0136] Table 4. Structure of CycleGAN Discriminator

[0137]

[0138] The discriminator consists of an input layer, convolutional layer 1, convolutional layer 2, convolutional layer 3, and an output layer. The input is a one-channel grayscale image, and the output is a single number between 0 and 1, representing the probability that the discriminator considers the input image to be a real sample rather than generated by the generator. The first four layers use the LeakyReLU activation function, the output layer uses the Sigmoid activation function, and IN is added in the second to fourth layers to prevent overfitting.

[0139] After building the generator and discriminator of the model, the training function of the model needs to be built. The pseudocode for the training process is as follows:

[0140] Input: most similar sample domain X, target sample domain Y, number of training rounds NUM_EPOCHS, number of generated samples NUM_SAMPLE, generated sample storage address SAVE_DIR;

[0141] Output: A specified number of generated grayscale images of the target samples;

[0142] 1.for epoch in range(NUM_EPOCHS)do

[0143] 2. For each data in the dataset, do

[0144] 3. Generate the forged and reconstructed images corresponding to the X and Y domains respectively;

[0145] 4. Construct discriminators D respectively. x and D y The loss function;

[0146] 5. Discriminator D x and D y The gradient is cleared to zero;

[0147] 6. Calculate the discriminator D x and D y The gradient;

[0148] 7. Update discriminator D x and D y parameter;

[0149] 8. Construct the loss functions for generators G and F respectively;

[0150] 9. Clear the gradients of generators G and F to zero;

[0151] 10. Calculate the gradients of generators G and F;

[0152] 11. Update generator parameters G and F;

[0153] 12. end for

[0154] 13. end for

[0155] 14.for i in range(NUM_SAMPLE)do

[0156] 15. Use the i-th most similar sample to generate one target sample using generator G;

[0157] 16. Save the generated sample number i under SAVE_DIR;

[0158] 17. end for

[0159] At this point, after constructing the model structure and determining the model training process, the motor target sample transfer generation model based on CycleGAN has been completed.

[0160] The training batch size (BATCH_SIZE) is set to 10, the learning rate (LEARNING_RATE) is set to 0.00002, the coefficient LAMBDA_CYCLE in the loss function is set to 15, and the number of training epochs (NUM_EPOCHS) is set to 25. After training is completed, the target sample is generated.

[0161] MMD was used to analyze the changes in the similarity between generated and target samples with the number of training epochs under various failure modes. The results are as follows: Figures 5 to 9 As shown in the figure, it can be seen that with the increase of training times, the MMD value between the generated samples and the real samples shows a decreasing trend, that is, the generated samples become closer and closer to the real samples, and tend to stabilize after 20 rounds of training; then, the power spectra of the generated samples and the target samples were compared using various fault modes, and the results are as follows. Figures 10 to 14 As shown, the transfer generation model can learn fault features well. The frequency domain features of the generated fault samples are obvious, and the similarity with the target samples is high. It has achieved the transfer generation of target samples quite well.

[0162] Step 14: Using the augmented target samples, construct a CNN-based motor fault enhancement and diagnosis model to achieve motor fault diagnosis under small sample conditions.

[0163] The fault diagnosis model is built based on a CNN network. The network structure constructed in combination with the characteristics of the case samples is shown in Table 5. The CNN consists of 9 layers, namely four convolutional layers, four pooling layers and one fully connected layer.

[0164] Table 5 CNN Diagnostic Model Structure

[0165]

[0166]

[0167] All convolutional layers use the ReLU activation function, and batch normalization (BN) is added for initialization to prevent overfitting; pooling layers use max pooling with a kernel size of 2×2; fully connected layers use the Softmax activation function, and for each input sample, the final output is 5 numbers, which correspond to the probability of the sample belonging to each fault category. The category with the highest probability is the final classification.

[0168] After constructing the diagnostic model network structure, it is necessary to construct the model training function. The pseudocode for the training function is as follows:

[0169] Input: Training set A, test set B, number of training epochs NUM_EPOCHS;

[0170] Output: Diagnostic accuracy of the test set samples;

[0171] 1.for epoch in range(NUM_EPOCHS)do

[0172] 2. For each data in A do

[0173] 3. Clear the gradients of the CNN network to zero;

[0174] 4. Calculate the error of the CNN network;

[0175] 5. Update the CNN network parameters;

[0176] 6. end for

[0177] 7. end for

[0178] 8. For each data in B do

[0179] 9. Obtain the predicted value for the sample's attribution;

[0180] 10. Cumulative test set diagnostic accuracy;

[0181] 11. end for

[0182] At this point, the motor fault diagnosis model has been completed. The following steps will use this model to verify its enhanced motor fault diagnosis effect. The motor fault diagnosis model constructed in step 14 will be trained using the target samples before and after augmentation as training sets. The same test samples will then be used for testing to compare the enhanced motor fault diagnosis effect.

[0183] First, the model is trained using the un-augmented target sample training set. As mentioned earlier in this section, there are 10 target samples for each fault mode (training set), totaling 50 sets, and 1190 samples for the test set, for a total of 5950 sets, to train and test the diagnostic model. The batch size is set to 64, the learning rate to 0.0005, and the number of training epochs to 10. The training set loss function changes as follows: Figure 15 As shown in the figure. Then, the augmented target samples were used for enhanced fault diagnosis. The original 10 sets of target samples for each fault mode were augmented, with 100 sets augmented for each fault mode, resulting in 110 sets of target samples per fault mode (i.e., the training set). The test set remained unchanged at 1190 sets. Simultaneously, the training parameters of the diagnostic model were kept constant, and the diagnostic performance was compared. The change in the loss function of the diagnostic model after augmentation is shown in the figure. Figure 16 As shown.

[0184] Five fault diagnosis verifications were performed using the target samples before and after augmentation, and the results are shown in Table 6 below:

[0185] Table 6 Comparison of Fault Diagnosis Accuracy Before and After Target Sample Augmentation

[0186]

[0187]

[0188] As can be seen, the average diagnostic accuracy of motor fault diagnosis using the target sample before augmentation is 87.44%, while the diagnostic accuracy after augmentation reaches 98.63%, an improvement of 11.19%. The enhanced diagnostic effect is very significant. Therefore, this invention has achieved enhanced diagnosis of motor faults well when there are few target samples and similar samples exist.

[0189] According to the solution provided in the embodiments of the present invention, motor fault enhancement diagnosis is well achieved when there are few target samples and similar samples exist.

[0190] Although the present invention has been described in detail above, it is not limited thereto, and those skilled in the art can make various modifications based on the principles of the present invention. Therefore, all modifications made in accordance with the principles of the present invention should be understood to fall within the protection scope of the present invention.

Claims

1. A motor fault enhancement diagnosis method based on recurrent generative adversarial networks, characterized in that, include: The process involves acquiring target motor fault samples and similar motor fault samples for each motor fault mode, and then selecting the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples under each motor fault mode. This includes: there are n motor fault modes; for each fault mode, the target sample has m similar samples under different operating conditions; for the m similar samples, the maximum mean difference (MMD) value between each similar motor fault sample and the target motor fault sample under each motor fault mode is calculated to obtain m MMD values; the minimum MMD value is selected from the m MMD values, and the similar motor fault sample corresponding to the minimum MMD value is taken as the most similar motor fault sample. The pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network (RBAN) is trained using the target motor fault sample and the most similar motor fault sample to obtain a trained motor target sample transfer generation model based on a RBAN. The pre-constructed motor target sample transfer generation model based on a RBAN includes: constructing a motor target sample transfer generation model based on a RBAN, comprising a RBAN structure and a total loss function, based on the target motor fault sample and the most similar motor fault sample; wherein the RBAN structure includes: a first generator... and the first discriminator The first generative adversarial network; containing a second generator Second discriminator The second generative adversarial network; the total loss function includes: the first generative adversarial network loss function, the second generative adversarial network loss function, and the cycle consistency loss function; By inputting the most similar motor fault sample for each motor fault mode into the trained motor target sample transfer generation model based on recurrent generative adversarial network, the generated target motor fault sample for each motor fault mode is obtained. The pre-built motor fault enhancement diagnosis model based on convolutional neural network is trained using the target motor fault samples and the generated target motor fault samples to obtain the trained motor fault enhancement diagnosis model based on convolutional neural network. The fault data of the motor to be diagnosed is obtained, and the fault data is input into the trained motor fault enhancement diagnosis model based on convolutional neural network for fault diagnosis processing to obtain the fault type of the motor to be diagnosed.

2. The method according to claim 1, characterized in that, The step of training a pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network using the target motor fault sample and the most similar motor fault sample to obtain the trained motor target sample transfer generation model based on a recurrent generative adversarial network includes: By converting the target motor fault sample and the most similar motor fault sample respectively, we obtain the target motor fault sample in the form of a two-dimensional grayscale image and the most similar motor fault sample in the form of a two-dimensional grayscale image. The target motor fault sample in the form of a two-dimensional grayscale image and the most similar motor fault sample in the form of a two-dimensional grayscale image are used as inputs to train the motor target sample transfer generation model based on a recurrent generative adversarial network several times, so as to obtain the trained motor target sample transfer generation model based on a recurrent generative adversarial network.

3. The method according to claim 1, characterized in that, The pre-built motor fault enhancement diagnostic model based on convolutional neural networks includes: Based on the target motor fault samples and the generated target motor fault samples, a motor fault enhancement diagnosis model based on convolutional neural networks, including a convolutional neural network structure and a loss function, is constructed.

4. The method according to claim 3, characterized in that, The step of training a pre-built motor fault enhancement and diagnosis model based on a convolutional neural network using the target motor fault samples and the generated target motor fault samples to obtain a trained motor fault enhancement and diagnosis model based on a convolutional neural network includes: The target motor fault sample and the generated target motor fault sample are used as inputs to train the motor fault enhancement diagnosis model based on convolutional neural network several times to obtain a trained motor fault enhancement diagnosis model based on convolutional neural network.

5. A motor fault enhancement and diagnostic device based on a recurrent generative adversarial network, characterized in that, include: The sample acquisition module is used to acquire target motor fault samples and similar motor fault samples for each motor fault mode, and to select the most similar motor fault sample with the highest similarity to the target motor fault sample from the similar motor fault samples under each motor fault mode. This includes: there are n motor fault modes, and the target sample under each fault mode has m similar samples under different operating conditions. For the m similar samples, the maximum mean difference (MMD) value between each similar motor fault sample and the target motor fault sample under each motor fault mode is calculated to obtain m MMD values; the minimum MMD value is selected from the m MMD values, and the similar motor fault sample corresponding to the minimum MMD value is taken as the most similar motor fault sample. The first construction and training module is used to train a pre-constructed motor target sample transfer generation model based on a recurrent generative adversarial network (RBAN) using the target motor fault samples and the most similar motor fault samples, to obtain a trained motor target sample transfer generation model based on a RBAN; wherein, the pre-constructed motor target sample transfer generation model based on a RBAN includes: constructing a motor target sample transfer generation model based on a RBAN, comprising a RBAN structure and a total loss function, based on the target motor fault samples and the most similar motor fault samples; wherein, the RBAN structure includes: a first generator and the first discriminator The first generative adversarial network; containing a second generator Second discriminator The second generative adversarial network; the total loss function includes: the first generative adversarial network loss function, the second generative adversarial network loss function, and the cycle consistency loss function; The sample generation module is used to obtain the generated target motor fault samples for each motor fault mode by inputting the most similar motor fault samples for each motor fault mode into the trained motor target sample transfer generation model based on recurrent generative adversarial network. The second construction and training module is used to train the pre-constructed motor fault enhancement diagnosis model based on convolutional neural network using the target motor fault samples and the generated target motor fault samples, so as to obtain the trained motor fault enhancement diagnosis model based on convolutional neural network. The fault diagnosis module is used to acquire fault data of the motor to be diagnosed, and input the fault data into the trained motor fault enhancement diagnosis model based on convolutional neural network for fault diagnosis processing to obtain the fault type of the motor to be diagnosed.

Images