A fault diagnosis method for complex power supply and transmission mechanisms based on sparse autoencoder-assisted classification generative adversarial network

By combining sparse autoencoder and auxiliary classification generative adversarial network in the trouble diagnosis of complex supply and delivery mechanisms, the problems of insufficient sample data and unbalance are solved, and more efficient fault recognition accuracy and stability are achieved, especially in small samples. The recognition ability of most categories under conditions is improved.

CN114676733BInactive Publication Date: 2024-11-26NAVAL UNIV OF ENG PLA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210354689.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2022-01-17
Filing Date
2022-04-06
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

There are problems of insufficient sample data and uneven data in the fault diagnosis of complex supply and transportation mechanisms, resulting in poor performance of deep learning models under small sample conditions and poor recognition of most classes. Traditional data enhancement methods cannot effectively solve the problem of uneven data. .

Method used

Combining sparse autoencoder and auxiliary classification generative adversarial networks (SAE-ACGANs), the sparse autoencoder generates hidden variables and category information to form a joint feature, inputs the generator to generate images, and uses the discriminator's loss function to optimize network parameters. Achieve adversarial learning to improve feature learning ability and fault recognition accuracy.

Benefits of technology

The SAE-ACGANs framework can effectively capture the intrinsic distribution of wavelet time-frequency graphs, improve the quality of generated data and fault identification accuracy, improve the influence of classification preferences of most categories, and improve the convergence speed, training accuracy and stability of the model, especially in non-equilibrium Fault identification capability under small sample conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114676733B_ABST
    Figure CN114676733B_ABST
Patent Text Reader

Abstract

The present invention provides a method for diagnosing faults of complex supply and transmission mechanisms based on a sparse autoencoder-assisted classification generative adversarial network. The ACGANs framework is applied to the field of fault diagnosis. The trained discriminator can directly identify the truth or falsity of the input time-frequency feature graph and the category of the fault state to which it belongs, without mixing the generated data with the real data and adding another model for training and identification. From the perspective of increasing the image feature constraints on the latent variables, a SE is pre-trained to encode the original data, and noise is added to the encoding vector through a standard normal distribution to jointly constitute a latent variable containing the real image feature information, which is then input into the generator together with the category to which it belongs, so as to strengthen the ability of the latent variable to represent the features related to the category to which the image belongs, and reduce the scope of the generator to learn the real sample feature space, so that the image category features can be effectively maintained in both the model training and recognition stages, thereby further improving the performance of the discriminator.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a complex supply and transmission mechanism fault diagnosis method based on a sparse autoencoder-assisted classification generative adversarial network. Background Art

[0002] Complex supply and delivery mechanisms are important components of special equipment. Their reliability and stability directly affect the performance of special equipment functions. For example, the supply and delivery mechanism of artillery directly affects the movement state of ammunition, and its movement process is accompanied by violent impacts and vibrations, and the mechanical environment is very complex. Therefore, conducting fault diagnosis of complex supply and delivery mechanisms is of great significance to improving work efficiency and reliability, reducing maintenance time and costs, etc.

[0003] The complex mechanical environment and unique mechanical structure make the vibration signal of the supply and transportation mechanism have the characteristics of nonlinearity and non-stationarity. The time-frequency analysis method can reflect the energy intensity changes of the signal in two dimensions of time and frequency, effectively describe the subtle fault characteristics of the signal, and is one of the important methods for analyzing and processing non-stationary signals. The traditional short-time Fourier transform (STFT) can transform the one-dimensional fault vibration signal into a two-dimensional matrix, thereby obtaining a feature map containing time domain and frequency domain information. However, the size and shape of the STFT time window are relatively fixed, and it cannot meet the requirements of time resolution and frequency resolution at the same time. The continuous wavelet transform (CWT) with continuous scale transformation can overcome the defects of STFT. The sampling step size of different frequencies is adjustable using a wavelet basis similar to the characteristic distribution waveform. The frequency resolution is higher at the low frequency of the signal, and the time resolution is higher at the high frequency. Therefore, CWT can carefully depict the local morphology of the signal in the time-frequency feature map, has good time-frequency window characteristics and local resolution ability, and is a powerful tool for processing the sudden change part of the non-stationary signal. [5] .

[0004] Therefore, intelligent algorithms based on wavelet time-frequency graphs and deep learning are widely used in feature extraction and fault diagnosis of complex mechanical systems. Compared with traditional methods, deep learning can adaptively extract fault features from time-frequency graphs and achieve end-to-end fault diagnosis. Yuan Jianhu et al. input the compressed wavelet time-frequency graph into a convolutional neural network, which can effectively identify the fault type of rolling bearings; Cheng et al. proposed a data-driven rotating machinery fault diagnosis method based on CWT and local binary convolutional neural network model. Experiments show that this method has more stable and reliable prediction accuracy. However, there are two difficulties in using deep learning in fault diagnosis of complex supply and transportation mechanisms: ① Insufficient sample data: Due to the complexity of the mechanical structure and working mechanism of the supply and transportation mechanism, it is usually difficult to obtain sufficient samples. Since the deep learning model is prone to overfitting under small sample conditions, the model fault diagnosis effect is poor; ② Unbalanced sample data: In order to adapt to the working environment of special equipment, the supply and transportation mechanism must be in normal operation for a long time, and key components are frequently replaced within the preset life range. Therefore, the number of fault samples is very scarce, and the ratio of positive and negative samples in the actual data set is extremely disparate, which leads to the model classification results being biased towards the majority class and poor recognition performance for the minority class.

[0005] The traditional method to solve the above problems is to enhance the training data based on geometric transformation, noise perturbation and grayscale transformation. Although the above methods are simple to operate, the data distribution generated is too single, the problem of data imbalance is not solved essentially, and it is easy to cause the model to overfit on the target data set under small sample conditions. Goodfellow et al. proposed generative adversarial networks (GANs) in 2014, which achieves "Nash equilibrium" through mutual confrontation and learning between the generator and the discriminator, and finally makes the data distribution generated by the generator from simple hidden variables close to the real data distribution, solving the problem of insufficient sample data. However, GANs networks are only used in the field of computer graphics and cannot classify and identify labeled data. It was not until Augustus et al. proposed auxiliary classifier generative adversarial networks (ACGANs) in 2016 that they creatively believed that the discriminator has the ability to identify class labels, and divided the loss function into true and false discrimination loss and sample classification loss, so that the generator can generate corresponding samples according to the label, and the discriminator can achieve multi-classification functions while performing binary classification. As the research deepens, many scholars have optimized ACGANs in order to obtain better classification results. Sun Canfei et al. used a stacked contraction autoencoder network as the discriminator of ACGANs to improve the fault diagnosis performance of the network; Niu Weiyu et al. introduced a pooling layer in the discriminator to speed up its calculation speed in multi-classification tasks and prevent the model from overfitting; Zhu Kefan et al. introduced weight coefficients in the discriminant probability of the ACGANs loss function, and selected high-quality generated samples according to the weight size to further optimize the recognition network. In these ACGANs frameworks, image information can be learned based on labels, but when the characteristics of the sample data are not obvious, the image distribution characteristics cannot be effectively learned, and the interpretability of the input hidden variables of the generator for the generated data is not considered. Summary of the invention

[0006] The purpose of the present invention is to further improve the feature learning ability of ACGANs and better solve the fault identification problem of complex supply and transmission mechanisms under unbalanced small sample conditions. The present invention combines sparse autoencoder (SAE) with ACGANs and proposes a fault diagnosis method based on sparse autoencoder assisted classification generative adversarial network (SAE-ACGANs).

[0007] The technical solution adopted by the present invention is:

[0008] A complex supply and transmission mechanism fault diagnosis method based on sparse autoencoder assisted classification generative adversarial network includes the following steps:

[0009] Step 1: Collect the vibration signal of the supply and transmission mechanism, calculate the wavelet time-frequency diagram, and divide it into training samples and test samples;

[0010] Step 2: Input the training sample into the sparse autoencoder and encode it to generate the latent variable z r ;

[0011] Step 3: Combine category information c and latent variable z r Construct joint features, input generator, and obtain generated image

[0012] Step 4: Transform the real image and generate images The discriminator is input together to distinguish the authenticity and category of the samples. At the same time, the network loss value of the discriminator is calculated based on its loss function, and the network parameters are updated using the Adam optimizer. The parameters of the discriminator are fixed, and the parameters of the sparse autoencoder and generator are adjusted using the loss function.

[0013] Step 5: Based on the adversarial learning method, steps 2 to 4 are repeated until the network converges, completing the training of the sparse autoencoder-assisted classification generative adversarial network;

[0014] Step 6: Take out the discriminator as the supply and delivery mechanism fault identification network, input the test sample into it to realize fault diagnosis, and output the diagnosis result.

[0015] Preferably, in step 3, an image is generated It can be expressed as:

[0016]

[0017] In the formula, Represents the original image input and the generated image obtained by the category label c; f φ (·) represents the mapping function of the sparse autoencoder, φ is the encoding parameter set of the sparse autoencoder; the category information c is represented by the one-hot encoding of the label of the original image; the generator f θ The decoding parameter set of (·) is θ, which is obtained by combining the category information c and the latent variable z r Composition of joint features (z r , c) to generate images of a specific category Then the real image and generate images They are input into the discriminator together to distinguish the authenticity and category of the input image.

[0018] Preferably, in step 4, the loss function of the discriminator is:

[0019]

[0020] In the formula, the discriminant loss L S and classification loss L C The expression is as follows:

[0021]

[0022] Where P data (x) is the true distribution of the original image; q φ (z r |x) represents the encoding mapping distribution; is the parameter set of the discriminator; and are the probability outputs of the discriminator for the true or false input sample and the category to which it belongs; loss(·) is the classification loss between the sample category predicted by D and the true category, and the cross entropy loss function is adopted.

[0023] Preferably, the loss functions of the sparse autoencoder and generator are:

[0024]

[0025] In the formula, α is the coefficient for adjusting the loss size of the auxiliary classification generative adversarial network, and α is set to 50% during training; Represents the loss function of the sparse autoencoder. The specific expression is as follows:

[0026] L SAE (x, z) = l AE (x,z)+βΩ(ρ)

[0027] In the formula, x represents z means β is the weight coefficient of the sparse penalty term; the expression of the sparse penalty term Ω is:

[0028]

[0029] Where D is the number of neurons in the hidden layer; ρ is the sparsity parameter; when ρ j =ρ, the value of the sparse penalty term Ω is zero; when ρ j ≠ρ, Ω will increase with the degree of deviation; then the average activation degree ρj of the network is controlled by limiting the sparse parameter ρ;

[0030] Secondly, l AE The expression for (x, z) is as follows:

[0031]

[0032] In the formula, θ z To decode the parameter set, the expression is: in, is the decoder weight matrix; b z is the decoder bias vector; θ h For the encoding parameter set, the expression is: in, is the encoder weight matrix, b h is the encoder bias vector.

[0033] Preferably, the average activation ρ j The expression is as follows:

[0034]

[0035] In the formula, x i is the i-th input sample; is the weight matrix of the jth neuron; b j is the bias vector of the jth neuron.

[0036] Beneficial effects of the present invention:

[0037] The present invention adds a SAE structure as an image feature extractor in the ACGANs framework, encodes the original data, and adds noise to the encoding vector through the standard normal distribution to jointly constitute a hidden variable containing the real image feature information, and then inputs it together with the category information to the generator to enhance the ability of the hidden variable to represent the features related to the category to which the image belongs, so that the image category features can be effectively maintained in both the model training and recognition stages, thereby improving the performance of the discriminator. Therefore, SAE-ACGANs can capture the intrinsic distribution of the wavelet time-frequency graph, optimize the feature learning ability of the model, improve the quality of the generated data and the fault recognition accuracy, and effectively improve the influence of the majority class classification preference, and finally realize the fault recognition of complex supply and transportation mechanisms under the condition of unbalanced small samples. The experimental research results show that the SAE-ACGANs framework can fully learn the intrinsic distribution and deep features of the input samples, and improve the performance of the discriminator compared with the original ACGANs framework, and achieve the improvement of the model convergence speed, training accuracy and stability; compared with the traditional unbalanced data processing algorithm, the model effectively improves the influence of the majority class classification preference, and greatly improves the recognition ability of minority class fault samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 Structural diagram of GANs and ACGANs; (a) GANs; (b) ACGANs;

[0040] Figure 2 The schematic diagram of the fault diagnosis method based on wavelet time-frequency diagram and SAE-ACGANs;

[0041] Figure 3 It is a flow chart of the fault diagnosis method based on wavelet time-frequency diagram and SAE-ACGANs;

[0042] Figure 4 It is the time domain waveform of the vibration signal;

[0043] Figure 5 The time-frequency domain waveform of the RC unit period signal; (a) unit period time domain waveform; (b) unit period frequency domain waveform;

[0044] Figure 6 The time-frequency diagrams of unit periodic signals in different states; (a) time-frequency diagram of NM state signal; (b) time-frequency diagram of RC state signal; (c) time-frequency diagram of SW state signal;

[0045] Figure 7 Comparison of PCA visualization of extracted features; (a) PCA visualization of original time-frequency graph features; (b) PCA visualization of SAE-ACGANs extracted features;

[0046] Figure 8 Performance comparison between SAE-ACGANs and ACGANs; (a) generation loss comparison chart; (b) classification loss comparison chart; (c) classification accuracy comparison chart. DETAILED DESCRIPTION

[0047] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0048] 1. Sparse Autoencoder

[0049] A single-layer autoencoder (AE) is a symmetric neural network consisting of an encoder and a decoder. The encoder function and the decoder function are trained in a loop through the back-propagation algorithm. While the output layer z reconstructs the input layer x, the hidden layer h extracts and expresses the features of the input layer x.

[0050] For the original sample data set AE takes the input layer x m Mapped to hidden layer h m The expression is:

[0051]

[0052] Where: θ h is the encoding parameter set; is the encoder weight matrix, b h is the encoder bias vector; f(*) is the activation function. The present invention adopts the nonlinear Sigmoid activation function, and its mathematical expression is:

[0053]

[0054] The decoder converts the hidden layer h m The reverse output is the input layer x m The reconstructed expression z m , whose expression is:

[0055]

[0056] Where: θ z is the decoding parameter set; is the decoder weight matrix; b z is the decoder bias vector.

[0057] The encoding and decoding parameter sets {θ h ,θ z}, so that the input layer x m and reconstruct the output layer z m The loss function L between AE (x, z) minimizes:

[0058]

[0059] In order to improve the anti-noise ability of coding learning and realize deep mining of input data features, the sparse autoencoder introduces an additional sparsity constraint in the loss function of AE to control the activation degree of hidden neurons. For the jth neuron in the hidden layer, its average output activation degree ρ j for:

[0060]

[0061] In the formula, x i is the i-th input sample; is the weight matrix of the jth neuron; b j is the bias vector of the jth neuron.

[0062] KL divergence can be used to measure the difference between two asymmetric probability distributions, so KL divergence is introduced as a penalty term to constrain ρ j , so the expression of the sparse penalty term Ω is:

[0063]

[0064] Where D is the number of neurons in the hidden layer; ρ is the sparsity parameter; when ρ j =ρ, the value of the sparse penalty term Ω is zero; when ρ j ≠ρ, Ω will increase with the degree of deviation; then the average activation of the network is controlled by limiting the sparse parameter ρ j , in order to achieve the sparse penalty method of the network.

[0065] In summary, the loss function of SAE is SAE is defined as:

[0066] L SAE (x, z) = l AE (x,z)+βΩ(ρ) (7)

[0067] Where β is the weight coefficient of the sparse penalty term.

[0068] During the training process of SAE, the loss function is minimized through the back propagation of gradients, thereby updating the weight matrix and bias vector in the network:

[0069]

[0070] Where: ε is the learning rate.

[0071] If the dimension of the hidden layer is smaller than that of the input layer, a better hidden layer sparse representation of the input data can be obtained after training is completed, and the complex intrinsic features of the input data can be effectively learned after encoding.

[0072] 2. Auxiliary Classification Generative Adversarial Networks (ACGANs)

[0073] GANs consists of a generator G and a discriminator D. The structural framework is as follows Figure 1 As shown in (a), its core idea is based on the two-person zero-sum game in game theory. The random hidden variable z is input into G to generate a distribution P that obeys the real data distribution as much as possible. data The generated sample G(z) is G, which tries to make the distribution space P of the random latent variable z z Mapping to the real data distribution space P data , thereby generating a generated sample G(z), and D is used to determine whether the input sample is a real sample x or a generated sample G(z).

[0074] In the training process of GANs, D is trained by maximizing the difference between the data distribution of the real sample x and the generated sample G(z), and the generator is trained by minimizing this difference. The adversarial mechanism is adopted to make G and D continuously optimize their own performance and finally reach the "Nash equilibrium", that is, the generated sample G(z) successfully misleads D, so that G can achieve the goal of P. data An approximate estimate of . Therefore, the loss function of GANs is:

[0075]

[0076] Where: E[·] is the expectation of the corresponding distribution.

[0077] ACGANs is an improvement on GANs. Its structure is as follows: Figure 1 (b) As shown. There are two main improvements to ACGANs: inputting the latent variable z and the class label c into G together, thereby guiding G to generate class conditional samples G(c, z); secondly, adding a softmax classifier in D so that it can calculate the probability distribution of the class label while judging whether the input sample is real data. Therefore, the loss function of ACGANs is composed of the discriminant loss L S and classification loss L C constitute:

[0078]

[0079] Where: L S is the loss function for judging the authenticity of data; L C is the loss function of the probability distribution of data category labels; D S (x) is the probability that D judges x to be a true sample; D S (G(c, z)) is the probability that D judges that the generated sample G(c, z) is a true sample; L D (c x |x) is the classification loss of D for x, c x is the true category of x; L D (c|G(c,z)) is the classification loss of D on the generated sample G(c,z), and c is the label of G(c,z).

[0080] Since D should distinguish between generated samples and real samples as much as possible and classify them effectively; G should simulate the intrinsic distribution of real samples as much as possible and also be effectively classified.

[0081] So the loss function of D is as follows:

[0082]

[0083] The loss function of G is as follows:

[0084]

[0085] Based on the above analysis, this embodiment proposes a fault diagnosis method based on sparse coding assisted generative adversarial network.

[0086] 3.1 Principle of the method

[0087] The principle of the fault diagnosis method for supply and transportation mechanism based on wavelet time-frequency diagram and SAE-ACGANs is as follows Figure 2 As shown in Figure 1, this method is divided into the discriminator training and testing stages. The SAE-ACGANs framework consists of a sparse encoder SE, a generator G, and a discriminator D. The decoder of SAE is designated as the generator of ACGANs, and the generated image It can be expressed as:

[0088]

[0089] Where: Represents the original image input and the generated image obtained by the category label c; f φ (·) represents the mapping function of the sparse autoencoder, φ is the encoding parameter set of the sparse autoencoder; the category information c is represented by the one-hot encoding of the label of the original image; the generator f θ The decoding parameter set of (·) is θ, which is obtained by combining the category information c and the latent variable z r Composition of joint features (z r , c) to generate images of a specific category Then the real image and generate images They are input into the discriminator together to distinguish the authenticity and category of the input image.

[0090] From the above, we can see that SE converts the original data Sparse coding is the latent variable z r , fully express the characteristics of the original image; specify the decoder of SAE as G of ACGANs to fully optimize the objective function of SAE. Therefore, the complete objective function of SAE-ACGANs is:

[0091]

[0092] Where: represents the loss function of the SAE module, and its specific expression is shown in formula (7); α is the coefficient for adjusting the loss size of ACGANs, and α is set to 50% during training; J ACGANs Represents the loss function of the ACGANs module, which is composed of the discriminative loss L S and classification loss L C composition:

[0093]

[0094] Where P data (x) is the true distribution of the original image; q φ (z r |x) represents the encoding mapping distribution; is the parameter set of the discriminator; and are the probability outputs of the discriminator for the true or false input sample and the category it belongs to. loss(·) is the classification loss between the sample category predicted by D and the true category. This paper adopts the cross entropy loss function.

[0095] Therefore, the loss functions of SE and G are:

[0096]

[0097] The loss function of D is:

[0098]

[0099] 3.2 Method flow

[0100] The implementation process of the complex supply and transmission mechanism fault diagnosis method based on wavelet time-frequency diagram and SAE-ACGANs is as follows: Figure 3 , the specific steps are as follows:

[0101] Step 1: Collect the vibration signal of the supply and transmission mechanism, calculate the wavelet time-frequency diagram, and divide it into training samples and test samples;

[0102] Step 2: Input the training sample into the sparse autoencoder and encode it to generate the latent variable z r ;

[0103] Step 3: Combine category information c and latent variable z r Construct joint features, input generator, and obtain generated image

[0104] Step 4: Transform the real image and generate images The discriminator is input together to distinguish the authenticity and category of the samples. At the same time, the network loss value of the discriminator is calculated based on its loss function, and the network parameters are updated using the Adam optimizer. The parameters of the discriminator are fixed, and the parameters of the sparse autoencoder and generator are adjusted using the loss function.

[0105] Step 5: Based on the adversarial learning method, steps 2 to 4 are repeated until the network converges, completing the training of the sparse autoencoder-assisted classification generative adversarial network;

[0106] Step 6: Take out the discriminator as the supply and delivery mechanism fault identification network, input the test sample into it to realize fault diagnosis, and output the diagnosis result.

[0107] 4. Experimental verification

[0108] 4.1 Experimental design and signal acquisition

[0109] The basic structure of a certain type of supply and delivery mechanism test platform includes a power unit, a test bench frame, a control device, a swing mechanism and a test system. The test platform uses a mobile hydraulic station as the power. After the manual recoil cylinder returns to its position, the trigger releases the tail iron block to move backward quickly. The tail iron block stores energy for the swing mechanism during the recoil process.

[0110] The tail iron block moves backwards with the roller as support under the drive of the slide, which is used to simulate the mechanism's recoil process. Then the swing mechanism drives the swing arm to swing from the vertical direction to the horizontal direction, which is used to simulate the mechanism's swing process. During the working process, the operating status of the slide and roller directly affects whether the tail iron block can move in place, and then affects whether the swing arm can swing in place normally. Therefore, the slide and roller are key components that determine whether the supply and transportation mechanism can work normally. Since the slide and roller cannot be directly installed with sensors, the experiment arranged 6 vibration acceleration sensors near the swing machine of the bench device and near the press machine above the roller, numbered A1~A6. The sensor type is an ICP acceleration sensor with a sampling frequency of 10kHz, and a 32-channel LMS signal acquisition system is used. Repeated experiments show that the data collected by the measuring point A1 directly above the roller can effectively reflect the operating status of the supply and transportation mechanism, so the vibration signal of the measuring point A1 is selected as the analysis object.

[0111] In order to simulate the failure of key components of the supply and transportation mechanism, the skateboard and roller were artificially processed to form wear and crack damage. The maximum damage size is shown in Table 1. The skateboard and roller in normal state, the worn skateboard and the roller with cracks can form three operating states: ① State 1—Normal state (Normal, NM), all key parts are not damaged; ② State 2—Skateboard wear (Skateboard Wear, SW), only one skateboard has long strip wear; ③ State 3—Roller Crack (Roller Crack, RC), only one roller has multiple cracks.

[0112] Table 1 Maximum damage size of key components

[0113] Tab.1 Maximum damage size of key components

[0114]

[0115] The control device is adjusted to the cyclic working mode, and the vibration acceleration signals are collected respectively under the above three states. The time domain waveforms are as follows: Figure 4 The test was performed in 11 cycles in the NM state, each cycle included 20 actions, and in 8 cycles in the RC and SW states, each cycle included 20 actions, so the total number of samples collected was 540.

[0116] 4.2 Signal analysis and parameter selection

[0117] Depend on Figure 4 It can be seen that there is a large impact vibration in the vibration acceleration signal measured by the supply and transmission mechanism. It is difficult to effectively distinguish the fault types based on the time domain waveform of the vibration signal. In addition, the test platform uses an automatic control system to control the running time of each action. A complete unit cycle time is about 6.8s. The vibration acceleration signal of the unit cycle under the RC state is extracted, and its time domain waveform and frequency domain waveform are as follows: Figure 5 This cycle mainly includes four actions: energy storage - recoil - swing up - swing back. Figure 5 (b) It can be seen that the vibration acceleration signal of the supply and transmission mechanism is a typical non-stationary signal, and traditional spectrum analysis is difficult to extract the characteristics of the mechanism fault category.

[0118] The unit period vibration signals under three states are subjected to CWT, and cmor3-3 wavelet is selected as the wavelet basis function of CWT. The obtained wavelet time-frequency diagram is shown in the figure below: Figure 6 shown.

[0119] Depend on Figure 6 It can be seen that the energy of the vibration signal of the supply and transmission mechanism is distributed in almost all frequency bands, and the energy in the medium and low frequency bands is higher than that in the high frequency band. Within a unit cycle, the signal shows obvious energy fluctuations as time changes, and the transient characteristics of the signal are obvious. Comparing the time-frequency diagrams of the NM and RC and SW states, it can be seen that although the energy of the fault state signal is slightly higher than that of the normal state, it is impossible to accurately judge the fault state of the supply and transmission mechanism based on the difference in the signal transient process.

[0120] When using the SAE-ACGANs method, the network hyperparameters involved are mainly the number of iterations, batch size, and network structure. Inappropriate hyperparameters will lead to insufficient generalization of fault identification and cause overfitting of the model. Considering the size of the input time-frequency graph and the sample size, a single factor analysis method is adopted.

[15] Analyze and select the parameters that make the model fault diagnosis accuracy the highest, as shown in Table 2.

[0121] Table 2 Model parameters

[0122] Tab.2 Model parameters table

[0123]

[0124] 4.3 Analysis of diagnostic results

[0125] 4.3.1 Feature Visualization Analysis

[0126] In order to verify the effectiveness of the SAE-ACGANs method in extracting deep features from small input samples, principal component analysis (PCA) is used to extract two principal components, and a visual comparison is made with the two principal components extracted from the time-frequency graph features. The resulting scatter plot is shown in Figure 2. Figure 7 As shown in the figure, the time-frequency graph features of the original signal have a wide distribution range and overlap with each other. After SAE-ACGANs fully learns the intrinsic distribution of the input samples, it can effectively distinguish the deep features of different fault types and has better clustering performance, which verifies the effectiveness of SAE-ACGANs in extracting deep features.

[0127] 4.3.2 Analysis of diagnostic accuracy

[0128] To simulate the unbalanced small sample condition, 40 NM samples, 10 RC samples and 10 SW samples were randomly selected from the 540 test samples as training data, and 40 NM, RC and SW samples were used as test data. The model parameters were set according to Table 2 for the experiment, and compared with the original ACGANs framework with the same parameters and network structure, and the model performance was recorded. The results are shown in Figure 8 shown.

[0129] Figure 8 Respectively represent the ability of SAE-ACGANs and original ACGANs to learn the intrinsic distribution of real samples and predict the category of input samples. The generator loss represents the probability loss of the generated samples to deceive the discriminator, and the discriminator loss represents the loss of judging whether the input sample is a generated sample. Figure 8 (a) and (b) show that SAE-ACGANs enhances the ability of latent variables to represent features related to the category to which the image belongs by adding a sparse encoder. The generator and encoder form an autoencoder, which can learn the feature distribution of real samples more quickly ( Figure 8 (a) As the number of iterations increases, the generation and classification losses of ACGANs are close to those of SAE-ACGANs, but the variation is relatively large and cannot be stabilized, resulting in large performance fluctuations when the discriminator performs fault diagnosis.

[0130] Figure 8 (c) is a comparison of the classification accuracy of the two frameworks on the test set. When the number of iterations is close to 160 and 350, ACGANs shows a large fluctuation ( Figure 8 (c) The arrow indicates that the performance of ACGANs is unstable. In addition, in the later stage of training, the generator learns the deep features of real samples, and the classification accuracy of SAE-ACGANs can be slightly improved.

[0131] In summary, the SAE-ACGANs framework constrains the image features of the input latent variables of the generator, and the model can converge at a faster speed. At the same time, the classification accuracy and stability are better than the original ACGANs framework.

[0132] 4.3.3 Sample Imbalance Analysis

[0133] The fault diagnosis problem of the supply and transportation mechanism for the unbalanced data set needs to pay more attention to the recognition performance of the minority class samples, so it is necessary to calculate the precision rate P of the prediction class. i and recall R i , the calculation formulas are shown in Equation 18 and Equation 19.

[0134]

[0135]

[0136] Among them, L is the number of categories in the data set; n ii is the number of samples of the i-th category correctly predicted as category i, n ij is the number of samples of the i-th category that are incorrectly predicted as category j.

[0137] Precision is a measure of classification accuracy, and recall is a measure of classification comprehensiveness. The performance evaluation criteria for models that process unbalanced data sets need to consider the comprehensive performance of the two, so as to achieve recognition effects that take into account both the majority class and the minority class. Therefore, in order to comprehensively and effectively evaluate the classification performance of the model in the case of class imbalance, the classification accuracy λ is selected. acc , F1 metric λ F1 and G-mean index λ G-mean As an evaluation index of the comprehensive classification performance of the model, the calculation formulas are shown in Equations 20 to 22.

[0138]

[0139]

[0140]

[0141] To further verify the superiority of the SAE-ACGANs framework in processing unbalanced data, random oversampling (ROS), random undersampling (RUS), synthetic minority oversampling (SMOTE) and adaptive sample synthesis (ADASYN) are selected.

[19] Four traditional algorithms for processing unbalanced samples are compared and analyzed. The above methods all guarantee the generation of data that makes the category distribution in the training set completely balanced, that is, the ratio of the number of samples in the NM, RC and SW states in the training set is 1:1:1. Then, CWT is performed on the balanced data set, and the time-frequency diagram is used as the training set to train a CNN with the same network structure and hyperparameter settings as the discriminator, and then the test set is input into it. The fault diagnosis results are shown in Table 3.

[0142] Table 3 Comparison of fault diagnosis results under different algorithms

[0143] Tab.3 Comparison table of fault diagnosis results under different algorithms

[0144]

[0145] As shown in Table 3, compared with the original data set, the diagnostic performance of the models after optimization by the five algorithms has been improved. The diagnostic performance of the ROS algorithm is less improved than that of the RUS algorithm combined with the CNN model, indicating that compared with the information loss caused by undersampling, the overfitting of the model caused by oversampling has a greater impact on the diagnostic performance. SMOTE and its derivative algorithm ADASYN can artificially synthesize new samples based on a small number of samples, which solves the problem of sample imbalance to a certain extent, but increases the possibility of overlap between samples of different categories. The λ of the SAE-ACGANs algorithm acc , F1 , G-mean The indicators are improved by 49.17%, 0.3604 and 0.7416 compared with the original data. Compared with the traditional unbalanced data processing algorithm, the method proposed in this paper can learn the actual distribution characteristics of fault samples, effectively solve the problem of unbalanced distribution of training set sample categories, improve the comprehensive classification ability of the discriminator, and reduce the missed judgment rate of power supply and transmission mechanism faults.

[0146] in conclusion

[0147] Aiming at the problem of few fault samples and serious imbalance between categories in intelligent fault diagnosis of complex supply and transportation mechanisms, this paper proposes a supply and transportation mechanism fault diagnosis method based on wavelet time-frequency diagram and SAE-ACGANs, and conducts experimental verification. The results show that:

[0148] (1) SAE-ACGANs can fully learn the intrinsic distribution of input samples, effectively distinguish the deep features of different fault types, and have better aggregation performance.

[0149] (2) Compared with the original ACGANs framework, SAE-ACGANs adds an SE structure to ACGANs, which together with the generator forms SAE, thereby imposing image feature constraints on the input latent variables, further improving the performance of the discriminator, and achieving improvements in model convergence speed, training accuracy, and stability.

[0150] (3) Compared with traditional unbalanced data processing algorithms such as ROS and RUS, the SAE-ACGANs framework can more effectively improve the classification preference problem of the diagnosis model for majority category samples, and the model's recognition ability for minority category fault samples is greatly improved.

[0151] The above description is only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A complex supply and transmission mechanism fault diagnosis method based on sparse autoencoder assisted classification generative adversarial network, characterized in that: The following steps are involved: Step 1: Collect the vibration signal of the supply and transmission mechanism, calculate the wavelet time-frequency diagram, and divide it into training samples and test samples; Step 2: Input the training sample into the sparse autoencoder and encode it to generate the latent variable z r ; Step 3: Combine category information c and latent variable z r Construct joint features, input generator, and obtain generated image Step 4: Transform the real image and generate images The discriminator is input together to distinguish the authenticity and category of the samples. At the same time, the network loss value of the discriminator is calculated based on its loss function, and the network parameters are updated using the Adam optimizer. The parameters of the discriminator are fixed, and the parameters of the sparse autoencoder and generator are adjusted using the loss function. Step 5: Based on the adversarial learning method, steps 2 to 4 are repeated until the network converges, completing the training of the sparse autoencoder-assisted classification generative adversarial network; Step 6: Take out the discriminator as the supply and delivery mechanism fault identification network, input the test sample into it to realize fault diagnosis, and output the diagnosis result.

2. According to claim 1, a complex supply and transportation mechanism fault diagnosis method based on sparse autoencoder assisted classification generative adversarial network is characterized in that: In step 3, generate the image It can be expressed as: In the formula, Represents the original image input and the generated image obtained by the category label c; f φ (·) represents the mapping function of the sparse autoencoder, φ is the encoding parameter set of the sparse autoencoder; the category information c is represented by the one-hot encoding of the label of the original image; the generator f θ The decoding parameter set of (·) is θ, which is obtained by combining the category information c and the latent variable z r Composition of joint features (z r , c) to generate images of a specific category Then the real image and generate images They are input into the discriminator together to distinguish the authenticity and category of the input image.

3. According to claim 1, a complex supply and transmission mechanism fault diagnosis method based on sparse autoencoder assisted classification generative adversarial network is characterized in that: In step 4, the loss function of the discriminator is: In the formula, the discriminant loss L S and classification loss L C The expression is as follows: Where P data (x) is the true distribution of the original image; q φ (z r |x) represents the encoding mapping distribution; is the parameter set of the discriminator; and are the probability outputs of the discriminator for the true or false input sample and the category to which it belongs, respectively; loss(·) is the classification loss between the sample category predicted by D and the true category, and the cross entropy loss function is adopted.

4. The method for diagnosing complex supply and transportation mechanism faults based on sparse autoencoder-assisted classification generative adversarial network according to claim 3 is characterized in that: The loss functions of the sparse autoencoder and generator are: In the formula, α is the coefficient for adjusting the loss size of the auxiliary classification generative adversarial network, and α is set to 50% during training; Represents the loss function of the sparse autoencoder. The specific expression is as follows: L SAE (x, z)=l AE (x,z)+βΩ(ρ) In the formula, x represents z means β is the weight coefficient of the sparse penalty term; the expression of the sparse penalty term Ω is: Where D is the number of neurons in the hidden layer; ρ is the sparsity parameter; when ρ j =ρ, the value of the sparse penalty term Ω is zero; when ρ j ≠ρ, Ω will increase with the degree of deviation; then the average activation of the network is controlled by limiting the sparse parameter ρ j ; Secondly, l AE The expression for (x, z) is as follows: In the formula, θ z To decode the parameter set, the expression is: in, is the decoder weight matrix; b z is the decoder bias vector; θ h For the encoding parameter set, the expression is: in, is the encoder weight matrix, b h is the encoder bias vector.

5. A complex supply and transportation mechanism fault diagnosis method based on sparse autoencoder assisted classification generative adversarial network according to claim 4, characterized in that: Average activation ρ j The expression is as follows: In the formula, x i is the i-th input sample; is the weight matrix of the jth neuron; b j is the bias vector of the jth neuron.