Equipment fault diagnosis method and system based on Gramb angle field and VAE-GAN

Abstract

The invention belongs to the related technical field of intelligent fault diagnosis of equipment, and discloses an equipment fault diagnosis method and system based on a Gramb angle field and a VAE-GAN, and the method comprises the steps: an offline stage: collecting equipment vibration signals, and building an initial signal sample set; a GAF coding technology is used to code the signal sample into a Gramer angle field picture; establishing a sample augmentation model based on the GAN, and training by using the initial picture sample set; acquiring an extended picture sample set by using the trained sample augmentation model; establishing a picture classification model, and training based on the extended picture sample set; an on-line diagnosis stage: encoding a signal sample to be detected into a Gramer angle field picture; and inputting the trained picture classification model to obtain a fault mode classification result. According to the method, the equipment fault vibration signal is converted into the two-dimensional image, the advantages of the GAN in image generation can be fully played, and the problems of few fault samples and unbalanced sample distribution in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to equipment intelligent fault diagnosis, and more specifically, to an equipment fault diagnosis method and system based on Gram angular field and VAE-GAN. Background Art

[0002] In the field of mechanical engineering, ball screws are an important transmission element, and their operating status is directly related to the normal operation and production efficiency of the equipment. However, in actual use, ball screws may suffer from various faults, such as pitting, poor lubrication, non-parallelism between the screw and the guide rail, and ball wear. These problems not only affect the normal operation of the equipment, but may also lead to reduced accuracy, increased noise, and even serious mechanical failures. Therefore, it is particularly important to perform real-time and accurate fault diagnosis on ball screws. Traditional fault diagnosis methods mainly rely on manual experience and regular inspections, which is not only inefficient, but also difficult to accurately detect early faults.

[0003] With the development of sensor technology and data processing technology, fault diagnosis methods based on intelligent algorithms have gradually become a research hotspot. These methods monitor the vibration signal, current signal and other parameters of the ball screw in real time, use advanced signal processing technology to extract fault features, and then combine machine learning or deep learning algorithms to identify and classify fault modes. Although deep learning has shown great potential in improving fault detection accuracy and efficiency, with the development of technology, its limitations and challenges have gradually emerged, among which the difficulty of data acquisition and the impact of data quality are particularly prominent.

[0004] In practical applications, large amounts of high-quality data are essential for deep learning-based fault diagnosis methods. However, due to the randomness and complexity of ball screw faults, as well as maintenance and testing limitations, it is extremely difficult to obtain large amounts of high-quality data. This situation of small numbers of fault samples and unbalanced samples limits the performance improvement of deep learning models, making it difficult to effectively apply traditional machine learning and deep learning methods to ball screw fault diagnosis. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a device fault diagnosis method and system based on Gram angular field and VAE-GAN, which is used to solve the problems of poor model generalization ability and low diagnostic accuracy caused by the difficulty in obtaining a large number of high-quality ball screw fault samples in actual fault diagnosis applications, the small number of fault samples and the uneven distribution of samples.

[0006] To achieve the above object, according to one aspect of the present invention, a device fault diagnosis method based on Gram angle field and VAE-GAN is provided, comprising:

[0007] Offline data processing and model training phase:

[0008] S11, collecting vibration signals of the equipment under different failure modes and establishing an initial signal sample set;

[0009] S12, encoding the signal samples in the initial signal sample set into a Gram angular field picture by using a GAF coding technology to obtain an initial picture sample set;

[0010] S13, establishing a sample augmentation model based on GAN to generate new image samples, training the sample augmentation model using the initial image sample set, and obtaining a trained sample augmentation model;

[0011] S14, using the trained sample augmentation model to expand the initial image sample set to obtain an extended image sample set;

[0012] S15, establishing an image classification model for classifying fault modes based on image samples, training the image classification model based on the extended image sample set, and acquiring a trained image classification model;

[0013] Online diagnosis stage:

[0014] S21, obtaining an actual vibration signal of the device and generating a signal sample to be tested;

[0015] S22, encoding the signal sample to be tested into a Gram angular field picture by using a GAF coding technology, and obtaining a picture sample to be tested;

[0016] S23, inputting the image sample to be tested into the trained image classification model to obtain the fault mode classification result to realize fault diagnosis.

[0017] According to the equipment fault diagnosis method based on Gram angle field and VAE-GAN provided by the present invention, S12 specifically includes:

[0018] Performing normalization processing on the signal samples in the initial signal sample set;

[0019] Convert the sequence data formed after normalization into a polar coordinate system;

[0020] Based on the parameters of the sequence data in the polar coordinate system after normalization, GAF is used to reconstruct the time series and obtain the Gram angle field image.

[0021] According to the equipment fault diagnosis method based on Gram angle field and VAE-GAN provided by the present invention, GAF is used to reconstruct the time series to obtain a Gram angle field image. Specifically, the Gram angle field image can be obtained based on the Gram angle sum field or the Gram angle difference field. The Gram angle sum field and the Gram angle difference field are specifically shown in the following formula:

[0022]

[0023] Where I is the unit row vector; is the normalized sample signal time series; for The transpose of ; GADF represents the Gram angle sum field; GASF represents the Gram angle difference field; φ i for Corresponding angle in polar coordinates; is the data in the normalized time series; i=1…n, where n is the length of the time series.

[0024] According to the equipment fault diagnosis method based on Gram angle field and VAE-GAN provided by the present invention, S13 is specifically:

[0025] For different fault modes, sample augmentation models are established respectively, and the sample augmentation models of the corresponding fault modes are trained using the image samples belonging to the same fault mode in the initial image sample set to obtain the trained sample augmentation models corresponding to the different fault modes;

[0026] Correspondingly, S14 is specifically:

[0027] By using the trained sample augmentation models corresponding to different fault modes, the image samples belonging to different fault modes in the initial image sample set are expanded respectively to obtain an extended image sample set.

[0028] According to the equipment fault diagnosis method based on Gram angular field and VAE-GAN provided by the present invention, establishing a sample augmentation model based on GAN in S13 specifically includes:

[0029] Combine VAE and GAN, use the encoder output of VAE to replace the random noise input of the generator in GAN, and establish a VAE-GAN model as a sample augmentation model;

[0030] The VAE-GAN model includes an encoder, a generator and a discriminator. The encoder is used to obtain its hidden layer variables and their probability distribution based on the real picture samples in the initial picture sample set. The generator is used to output new generated picture samples based on the hidden layer variables of the real picture samples and random noise. The discriminator is used to distinguish between real picture samples and generated picture samples.

[0031] According to the device fault diagnosis method based on Gram angular field and VAE-GAN provided by the present invention, in S13, the sample augmentation model is trained using the initial image sample set to obtain the trained sample augmentation model, which specifically includes:

[0032] Optimize the encoder training according to the encoder's divergence loss function;

[0033] Based on the Wasserstein divergence, an objective function is established to optimize the training of the generator and discriminator; finally, the trained VAE-GAN model is obtained.

[0034] According to the device fault diagnosis method based on Gram angular field and VAE-GAN provided by the present invention, the encoder divergence loss function L E Specifically:

[0035]

[0036] Where: μ and σ 2 are the mean and variance of the hidden layer variables of the encoder network respectively; KL(·) is the divergence between two data samples; N(μ,σ 2 ) represents normal distribution; N(0,1) represents standard normal distribution.

[0037] According to the equipment fault diagnosis method based on Gram angular field and VAE-GAN provided by the present invention, the objective function established based on Wasserstein divergence is shown as follows:

[0038]

[0039] Among them, G represents the generator; D represents the discriminator; E represents the expected function; x represents the real sample; y represents the generated sample; P r represents the true sample distribution; P g represents the generated sample distribution; P u is the Radon measure; represents the generated samples of random noise z; is the gradient operator; D(·) is the output of the discriminator; the two parameters k and p are constants, requiring p≠1, k>0.

[0040] According to the device fault diagnosis method based on Gram's angle field and VAE-GAN provided by the present invention, a selective convolutional network is introduced into the generator and the discriminator to optimize the performance of the VAE-GAN model in generating new image samples.

[0041] According to another aspect of the present invention, a device fault diagnosis system based on Gram's angle field and VAE-GAN is provided, the system comprising a memory and a processor, the memory storing a computer program, and the processor executing any one of the above-mentioned device fault diagnosis methods based on Gram's angle field and VAE-GAN when executing the computer program.

[0042] In general, compared with the prior art, the above technical solutions conceived by the present invention provide a device fault diagnosis method and system based on Gram angle field and VAE-GAN:

[0043] 1. It is proposed to convert the one-dimensional time series signal of the equipment fault vibration signal into a two-dimensional image, and use the two-dimensional image samples to train the GAN-based sample augmentation model, which can give full play to the advantages of GAN in image generation, making the extended image sample set generated after training more reliable, and can better solve the problem that a large number of high-quality ball screw fault samples are difficult to obtain in actual fault diagnosis applications, and there are few fault samples and uneven sample distribution, which is conducive to improving the generalization ability and diagnostic accuracy of the diagnosis model trained based on the extended image sample set;

[0044] 2. The proposed VAE-GAN model establishes an objective function based on Wasserstein divergence. Compared with other existing generative models, such as deep convolutional generative adversarial networks, the quality of samples generated by the model and the stability of the training process are significantly improved, and the difficulty and time of network training are reduced, achieving effective data enhancement for unbalanced samples;

[0045] 3. The problem of uneven sample distribution is solved. While reducing the training time and difficulty of the GAN model, it also reduces the risk of gradient vanishing and explosion. The introduction of the selective convolutional network better adapts to the non-stationary nature of device signals such as ball screws, generates a large amount of high-quality data, and effectively improves the performance and diagnostic accuracy of the fault diagnosis algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a schematic diagram of GAF image coding in an embodiment of the present invention;

[0047] Figure 2 This is a structural diagram of an improved VAE-GAN model in an embodiment of the present invention;

[0048] Figure 3 This is an overall flow chart of ball screw fault diagnosis based on the GAF-IVAE-GAN method in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0050] See also Figure 1 , Figure 2 and Figure 3 This embodiment provides a device fault diagnosis method based on Gram angular field and VAE-GAN (variational autoencoder generative adversarial network), and the device fault diagnosis method includes:

[0051] Offline data processing and model training phase:

[0052] S11, collecting vibration signals of the equipment under different failure modes and establishing an initial signal sample set;

[0053] S12, encoding the signal samples in the initial signal sample set into a Gram angular field picture by using a GAF coding technology to obtain an initial picture sample set;

[0054] S13, establishing a sample augmentation model based on GAN (Generative Adversarial Network) for generating new image samples, training the sample augmentation model using the initial image sample set, and obtaining a trained sample augmentation model;

[0055] S14, using the trained sample augmentation model to expand the initial image sample set to obtain an extended image sample set;

[0056] S15, establishing an image classification model for classifying fault modes based on image samples, training the image classification model based on the extended image sample set, and acquiring a trained image classification model;

[0057] Online diagnosis stage:

[0058] S21, obtaining an actual vibration signal of the device and generating a signal sample to be tested;

[0059] S22, encoding the signal sample to be tested into a Gram angular field picture by using a GAF coding technology, and obtaining a picture sample to be tested;

[0060] S23, inputting the image sample to be tested into the trained image classification model to obtain the fault mode classification result to realize fault diagnosis.

[0061] Based on the current situation that there are few fault samples and unbalanced samples when training the existing fault diagnosis model, this embodiment proposes to enhance the data of the samples, and specifically proposes to use a generative adversarial network (GAN) to provide a new solution for data enhancement. GAN can learn the data distribution of samples and generate fake samples that are indistinguishable from the real ones, so as to achieve the purpose of expanding data. This embodiment further takes into account that the most widely used and mature application of GAN is two-dimensional image generation, but in the existing deep learning method for equipment such as ball screw fault diagnosis, the commonly used method is to process one-dimensional signals; when enhancing the fault sample data, it is difficult to give full play to the advantages of GAN in image generation, and the generation accuracy is low.

[0062] Based on this, this embodiment proposes to convert the one-dimensional time series signal of the equipment fault vibration signal into a two-dimensional image, and use the two-dimensional image samples to train the GAN-based sample augmentation model, which can give full play to the advantages of GAN in image generation, making the extended image sample set generated after training more reliable, and can better solve the problems of a large number of high-quality ball screw fault samples being difficult to obtain in actual fault diagnosis applications, and there being few fault samples and uneven sample distribution, which is beneficial to improving the generalization ability and diagnostic accuracy of the diagnostic model trained based on the extended image sample set; and GAF ​​encoding is used to generate Gram angular field images, which retains the time correlation of the signal compared with grayscale images, fast Fourier transforms and other methods.

[0063] Furthermore, S12 specifically includes:

[0064] Performing normalization processing on the signal samples in the initial signal sample set;

[0065] Convert the sequence data formed after normalization into a polar coordinate system;

[0066] Based on the parameters of the sequence data in the polar coordinate system after normalization, GAF is used to reconstruct the time series and obtain the Gram angle field image.

[0067] S13 is specifically:

[0068] For different fault modes, sample augmentation models are established respectively, and the sample augmentation models of the corresponding fault modes are trained using the image samples belonging to the same fault mode in the initial image sample set to obtain the trained sample augmentation models corresponding to the different fault modes;

[0069] Correspondingly, S14 is specifically:

[0070] By using the trained sample augmentation models corresponding to different fault modes, the image samples belonging to different fault modes in the initial image sample set are expanded respectively to obtain an extended image sample set.

[0071] Furthermore, establishing a sample augmentation model based on GAN in S13 specifically includes:

[0072] Combine VAE (Variational Autoencoder Network) and GAN, use the encoder output of VAE to replace the random noise input of the generator in GAN, and establish a VAE-GAN model as a sample augmentation model;

[0073] The VAE-GAN model includes an encoder, a generator and a discriminator. The encoder is used to obtain its hidden layer variables and their probability distribution based on the real picture samples in the initial picture sample set. The generator is used to output new generated picture samples based on the hidden layer variables of the real picture samples and random noise. The discriminator is used to distinguish between real picture samples and generated picture samples.

[0074] A selective convolutional network is introduced into the generator and the discriminator to optimize the performance of the VAE-GAN model in generating new image samples.

[0075] In some specific embodiments, the device fault diagnosis method can be used to diagnose the fault of a ball screw, that is, a ball screw fault diagnosis method based on Gram angular field and improved VAE-GAN is provided, comprising the following steps:

[0076] Step 1: Collect vibration signals at the fixed end bearing seat during the operation of the ball screw, obtain vibration data samples of normal, worn, pitted, and non-parallel ball screws, and divide them into training and test sets in a ratio of 4:1;

[0077] Step 2: Encode the vibration signal into a Gram angular field image using GAF coding technology;

[0078] Step 3: Establish IVAE-GAN (Improved Variational Autoencoder Generative Adversarial Network) model, combine VAE and GAN, introduce Wasserstein divergence with better smoothing performance to solve the problem of gradient disappearance or explosion. At the same time, add selective convolution network (SKNet) to VAE-GAN to dynamically adjust the receptive field size to improve the quality of generated samples;

[0079] Step 4: Input the Gram angle field image of the training set in step 2 into the IVAE-GAN model established in step 3, and train the model until the generator and the discriminator reach Nash equilibrium and then stop training; the number of training rounds can be preset, and when the number of training rounds is reached, analyze whether the model loss reaches Nash equilibrium, and comprehensively determine whether to end the training;

[0080] Step 5: Use the model trained in step 4 to perform sample enhancement and expand the training set data to make its sample distribution balanced;

[0081] Step 6: Use the sample-enhanced training set in step 5 to train an image classification model, such as a VGG16 classification model, and test it with the sample data of the test set in step 2.

[0082] Further, in the step 2, refer to Figure 1 In order to give full play to the image generation ability of GAN and make it more suitable for sample enhancement of ball screw, the vibration signal is converted into Gram angle field image. The specific method is as follows:

[0083] The ball screw vibration signal data is a time series X={x 1 ,x 2 ,…,x n}, perform normalization according to formula (1) or formula (2) to normalize the time series range to the interval [-1, 1] or [0, 1].

[0084]

[0085] Where: Indicates that x i Normalized to the range [-1,1], Indicates that x i Normalized to the range [0,1]; i = 1…n, n is the length of the time series.

[0086] The normalized sequence data is converted into or Convert to polar coordinate system.

[0087]

[0088] Where: is the normalized time series data; φ i for Corresponding to the angle in polar coordinates; r i represents the radius in polar coordinates; t i is the timestamp; N is the constant factor of the regularized polar coordinate system space, which is used to adjust the span of the polar coordinate system. The rescaled time series in different target intervals have different angle boundaries. If the data scaling range is [-1,1], the converted angle range is [0,π]; if the scaling range is [0,1], the converted angle range is [0,π / 2].

[0089] Finally, GAF is used to reconstruct the time series. GAF is used to reconstruct the time series to obtain the Gram angle field image. Specifically, two different images can be generated based on different formulas, namely the Gram angle sum field (GASF) and the Gram angle difference field (GADF). Their expressions are shown in equations (4) and (5):

[0090]

[0091] Where: I is the unit row vector; is the normalized time series; for The transpose of ; GADF represents the Gram angle sum field; GASF represents the Gram angle difference field; φ i for Corresponding angle in polar coordinates; is the data in the normalized time series; i=1…n.

[0092] Furthermore, in step 3, this embodiment takes into account that in traditional GAN, the input of the generator G is usually random noise, which may lead to unstable training of the generator G and the discriminator D in the early stage of model training, and even mode collapse. To solve this problem, this embodiment uses a variational autoencoder (VAE) to automatically learn the potential features of the image and its probability distribution, and improve the input noise of the generator. Therefore, this embodiment combines VAE and GAN, and the encoder output of VAE replaces the random noise input of the generator in GAN, while stabilizing the training process through the reconstruction loss of VAE, and improving the generation quality of samples.

[0093] In the VAE-GAN structure proposed in this embodiment, the input of the generator is determined by the encoded hidden variables of the real samples. The encoder network is finally composed of a fully connected layer, which maps the high-dimensional features of the real input samples to the mean and variance of the hidden variables and provides constraints on the distribution of the input noise. The optimization goal of the encoder network is to minimize the KL divergence between the distribution of the hidden variables and the standard normal distribution, that is, the divergence loss function of the encoder can be KL divergence, and the divergence loss function of the encoder L E The definition of is shown in formula (7):

[0094]

[0095] Where: μ and σ 2 are the mean and variance of the hidden layer variables of the encoder network respectively; KL(·) is the divergence between two data samples; N(μ,σ 2 ) represents normal distribution; N(0,1) represents standard normal distribution. The encoder's divergence loss function can also be a JS divergence loss function.

[0096] In order to avoid the gradient vanishing problem that may be caused by using KL divergence or JS divergence in the above VAE-GAN, the traditional method usually uses Wasserstein distance to measure the similarity between two samples. Its advantage is that even if the two sample distributions do not overlap, Wasserstein distance can still smoothly reflect the distance between the two samples, which alleviates the gradient vanishing problem to a certain extent and provides better generated sample quality and stability. Its definition is shown in formula (8).

[0097]

[0098] y=G(Z) (9)

[0099] Z=μ+e×σ(10)

[0100] Where: x represents the real sample; y represents the generated sample; G(·) is the output of the generator; P r represents the true sample distribution; P g represents the generated sample distribution; W(P r ,P g ) is the Wasserstein distance between the real sample x and the generated sample y; ∏(P r ,P g ) is the true sample distribution P r and generate sample distribution P g The joint distribution set of; E is the expected function; e~N(0,1); Z is the combination of μ and σ 2 Synthesized generator input noise; inf denotes infimum.

[0101] This embodiment takes into account that the 1-Lipschitz constraint will be introduced when the Wasserstein distance is introduced. A common practice is to introduce a gradient penalty mechanism (GP) to perform random interpolation between true and false samples to penalize, thereby ensuring that the transition area between true and false samples satisfies the L constraint. However, this approach is mechanical and it is difficult to simulate this full-space distribution with limited sampling. Therefore, the Wasserstein divergence is used to solve this problem. While retaining the advantages of the traditional Wasserstein distance, its Lipschitz constraint is removed. Based on the above improvements, the objective function of VAE-GAN training, that is, the objective function established based on the Wasserstein divergence, is shown in formula (11):

[0102]

[0103] Where: G represents the generator; D represents the discriminator; E represents the expected function; x represents the real sample; y represents the generated sample; P r represents the true sample distribution; P g represents the generated sample distribution; P u is the Radon measure; represents the generated samples of random noise z; is the gradient operator; D(·) is the output of the discriminator; the two parameters k and p are constants, p represents the norm type of the gradient penalty term, requiring p≠1, k>0.

[0104] Specifically, in S13, the sample augmentation model is trained using the initial image sample set to obtain a trained sample augmentation model, which specifically includes:

[0105] Optimize the encoder training according to the encoder's divergence loss function;

[0106] Based on the Wasserstein divergence, an objective function is established to optimize the training of the generator and discriminator; finally, the trained VAE-GAN model is obtained.

[0107] Furthermore, in step three, considering that the vibration signal during the movement of the ball screw is often in a non-stationary state, a selective convolutional network is introduced to dynamically adjust the receptive field and understand the dynamic changes of the signal, so as to better adapt to the non-stationarity of the signal. The selective convolutional network (SKNet) can take into account feature information at multiple scales and focus on key features in the image. This mechanism enables the generator to pay more attention to areas that are closely related to the target image features when generating images, thereby optimizing the performance of VAE-GAN in image generation. The selective convolutional network mainly includes the following three processes:

[0108] Separation: In order to capture multi-scale feature information, the separation operation first processes the input features through convolution kernels of different sizes, and then connects a batch normalization layer (BN) and a ReLU activation function after each convolution to form a set of convolution-batch normalization-activation (CBR) units. The input features are separated into feature maps of different scales, which are then concatenated together to form a richer feature representation.

[0109] Fusion: First, the feature maps of all branches, that is, the feature maps of different scales mentioned above, are added element by element to obtain:

[0110]

[0111] Where: and Feature graph representing a branch.

[0112] Then, the summed results are globally averaged and pooled to obtain the statistical information of each channel, i.e., a single feature vector s c :

[0113]

[0114] Where: C, H, W represent the size of the feature map, U c Represents the feature map corresponding to a single feature vector.

[0115] Then, through the fully connected layer, the pooled feature z is processed as follows:

[0116] z=F fc (s) = δ(B(Ws))(14)

[0117] Where: δ represents the ReLU activation function; B represents batch normalization; W∈R d×C ; The d value is controlled by the compression ratio r:

[0118] d=max(C / r,L)(15)

[0119] Where: L is the minimum value of d.

[0120] Selection: Get each feature scale through Softmax and The weight vectors a and b are obtained, and finally the fused feature map V is obtained:

[0121]

[0122] Where: V = [V 1 ,V 2 ,…,V c ]; V∈R H×W×C .

[0123] Specific examples:

[0124] This embodiment provides a ball screw fault diagnosis method based on the Gram angular field and improved variational autoencoder generative adversarial network. The specific implementation steps of the present invention are described in detail below with reference to the accompanying drawings.

[0125] Step 1: Collect the vibration signal of the fixed bearing seat during the operation of the ball screw and pre-process the data. In order to obtain the vibration signals of the ball screw in different operating states, a corresponding ball screw fault simulation experimental platform was built. The experimental platform includes a mobile platform, a guide rail, a ball screw pair, a fixed end bearing seat, a diaphragm coupling and a servo motor. The three-dimensional acceleration sensor is installed on the fixed end bearing seat with a sampling frequency of 10240HZ. The three-dimensional vibration signals of the ball screw in four states: normal, worn, pitting and non-parallel between the screw and the guide rail are collected and transmitted to the NI 9230 data acquisition card to convert the analog signal into a digital signal.

[0126] The collected vibration signals of the fixed-end bearing seat are truncated into time series with a length of 1024 data points. There are 50 samples for each fault condition and 250 samples for the normal condition, totaling 400 samples, which are divided into training set and test set in a ratio of 4:1.

[0127] Step 2: Convert each sample into a GADF image (Gram angular field image) through GAF image encoding. The image size is 224×224×3. The schematic diagram of GAF image encoding is as follows: Figure 1 shown.

[0128] Step 3: Establish the IVAE-GAN sample enhancement model to expand the fault class samples, including the encoder, the decoder as the generator and the discriminator. The specific network structure is as follows Figure 2 The encoder learns the effective features of the real image and maps them to hidden variables. The decoder uses the hidden variables to generate images. Finally, the discriminator determines whether the image is real. The network structure is as follows:

[0129] Encoder: The network consists of 6 convolutional layers and 1 fully connected layer. "Convolutional BN / LeakyReLU" in the network means: a BN layer and a LeakyReLU activation function are added after the convolutional layer. The network first encodes the one-dimensional time series signal into a 3-channel 224×224 real image through GAF, and then passes through multiple convolutional layers to increase the number of channels layer by layer and reduce the feature dimension layer by layer. A BN layer is used after each convolutional layer to increase the network training speed, and LeakyReLU is selected as the activation function. After passing through the encoder, the image is mapped to 1000-dimensional mean and variance data, which ultimately constitutes the input of the generator.

[0130] Generator: The network consists of 1 fully connected layer and 5 deconvolution layers. First, the 1000-dimensional noise signal input to the generator is mapped and reconstructed into a size of 7×7×256 using a fully connected layer. Then it is sent to 5 deconvolution layers, and the final output dimension is consistent with the size of the real image. Similarly, a BN layer is added after the first four deconvolution layers, the activation function is ReLU, and the activation function of the last layer is Sigmoid. At the same time, the SKNet module is placed after the first deconvolution layer.

[0131] Discriminator: The network consists of 4 convolutional layers and 2 fully connected layers. First, the generated and real images are feature extracted and dimension reduced through convolutional layers and max pooling layers. Then, they are mapped to 1D using fully connected layers. Except for the last layer, LeakyReLU is used as the activation function, and the SKNet module is placed after the first convolutional layer.

[0132] The detailed parameters of the network are as follows:

[0133] Table 1 IVAE-GAN network structure parameters

[0134]

[0135]

[0136] Step 4: Use the GADF encoding graph of the training set as the input of the IVAE-GAN model, save the model after 5000 iterations, and use the saved generator model to perform sample enhancement to generate samples similar to real samples. Select structural similarity (SSIM) as the basis for the quality of generated samples to quantitatively evaluate the quality of generated samples. The SSIM index is composed of the brightness, contrast, and structural differences between samples, and measures the similarity between images. The definition is shown in Equation 17. The larger the SSIM value, the higher the sample similarity.

[0137]

[0138] Where: μ x , μ y are the average values ​​of x and y respectively; σ x , σ y are the variances of x and y respectively; σ xy is the covariance of x and y; C 1 , C 2 is a constant.

[0139] Step 5: Figure 3As shown, it is the algorithm flow chart involved in this embodiment, using the trained generator model to perform sample enhancement and expand the training set. Finally, the sample-enhanced training set is used to train the VGG16 classification model, and the test set is used for verification to complete the fault diagnosis process.

[0140] Furthermore, this embodiment also provides an equipment fault diagnosis system based on Gram angular field and VAE-GAN, the system comprising a memory and a processor, the memory storing a computer program, and the processor executing any one of the above-mentioned equipment fault diagnosis methods based on Gram angular field and VAE-GAN when executing the computer program.

[0141] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A device fault diagnosis method based on Gram angle field and VAE-GAN, characterized in that: include: Offline data processing and model training phase: S11, collecting vibration signals of the equipment under different failure modes and establishing an initial signal sample set; S12, encoding the signal samples in the initial signal sample set into a Gram angular field picture by using a GAF coding technology to obtain an initial picture sample set; S13, establishing a sample augmentation model based on GAN to generate new image samples, training the sample augmentation model using the initial image sample set, and obtaining a trained sample augmentation model; S14, using the trained sample augmentation model to expand the initial image sample set to obtain an extended image sample set; S15, establishing an image classification model for classifying fault modes based on image samples, training the image classification model based on the extended image sample set, and acquiring a trained image classification model; Online diagnosis stage: S21, obtaining an actual vibration signal of the device and generating a signal sample to be tested; S22, encoding the signal sample to be tested into a Gram angular field picture by using a GAF coding technology, and obtaining a picture sample to be tested; S23, inputting the image sample to be tested into the trained image classification model to obtain the fault mode classification result to realize fault diagnosis.

2. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 1, characterized in that: S12 specifically includes: Performing normalization processing on the signal samples in the initial signal sample set; Convert the sequence data formed after normalization into a polar coordinate system; Based on the parameters of the sequence data in the polar coordinate system after normalization, GAF is used to reconstruct the time series and obtain the Gram angle field image.

3. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 2, characterized in that: GAF is used to reconstruct the time series to obtain the Gram angle field image. Specifically, the Gram angle field image can be obtained based on the Gram angle sum field or the Gram angle difference field. The Gram angle sum field and the Gram angle difference field are specifically shown in the following formula: Where I is the unit row vector; is the normalized sample signal time series; for The transpose of ; GADF represents the Gram angle sum field; GASF represents the Gram angle difference field; φ i for Corresponding angle in polar coordinates; is the data in the normalized time series; i=1…n, where n is the length of the time series.

4. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 1, characterized in that: S13 is specifically: For different fault modes, sample augmentation models are established respectively, and the sample augmentation models of the corresponding fault modes are trained using the image samples belonging to the same fault mode in the initial image sample set to obtain the trained sample augmentation models corresponding to the different fault modes; Correspondingly, S14 is specifically: By using the trained sample augmentation models corresponding to different fault modes, the image samples belonging to different fault modes in the initial image sample set are expanded respectively to obtain an extended image sample set.

5. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 1, characterized in that: The sample augmentation model based on GAN in S13 specifically includes: Combine VAE and GAN, use the encoder output of VAE to replace the random noise input of the generator in GAN, and establish a VAE-GAN model as a sample augmentation model; The VAE-GAN model includes an encoder, a generator and a discriminator. The encoder is used to obtain its hidden layer variables and their probability distribution based on the real picture samples in the initial picture sample set. The generator is used to output new generated picture samples based on the hidden layer variables of the real picture samples and random noise. The discriminator is used to distinguish between real picture samples and generated picture samples.

6. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 5, characterized in that: In S13, the sample augmentation model is trained using the initial image sample set to obtain a trained sample augmentation model, which specifically includes: Optimize the encoder training according to the encoder's divergence loss function; Based on the Wasserstein divergence, an objective function is established to optimize the training of the generator and discriminator; finally, the trained VAE-GAN model is obtained.

7. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 6, characterized in that: The encoder’s divergence loss function L E Specifically: Where: μ and σ 2 are the mean and variance of the hidden layer variables of the encoder network respectively; KL(·) is the divergence between two data samples; N(μ,σ 2 ) represents normal distribution; N(0,1) represents standard normal distribution.

8. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 6, characterized in that: The objective function established based on Wasserstein divergence is as follows: Among them, G represents the generator; D represents the discriminator; E represents the expected function; x represents the real sample; y represents the generated sample; P r represents the true sample distribution; P g represents the generated sample distribution; P u is the Radon measure; represents the generated samples of random noise z; is the gradient operator; D(·) is the output of the discriminator; the two parameters k and p are constants, requiring p≠1, k>

0.

9. The device fault diagnosis method based on Gram angle field and VAE-GAN according to claim 4, characterized in that: A selective convolutional network is introduced into the generator and the discriminator to optimize the performance of the VAE-GAN model in generating new image samples.

10. A device fault diagnosis system based on Gram angle field and VAE-GAN, characterized in that: The system includes a memory and a processor, the memory stores a computer program, and the processor executes the equipment fault diagnosis method based on Gram angle field and VAE-GAN as described in any one of claims 1 to 9 when executing the computer program.

Images