Rotating machinery fault diagnosis method based on zero-trial learning and feature extraction
Through the combined feature refinement module of generalized zero-test learning and VAEGAN generation model, the problem of data scarcity and cross-data set deviation in rotary machinery fault diagnosis is solved, and high-accuracy fault diagnosis with very few samples is achieved. It is suitable for rotary machinery. real-time monitoring and potential hazard prediction.
Patent Information
- Application Number
- CN202210076016.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-23
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-01-23
AI Technical Summary
In rotary machinery fault diagnosis, the existing technology faces data scarcity and cross-data set deviation problems, resulting in increased difficulty in troubleshooting, especially in insufficient diagnostic accuracy under very few samples.
Using generalized zero-trial learning (GZSL) method, combined with VAEGAN generation model and feature refinement module, VAE and WGAN are generated through feature, and two-stage learning is carried out, feature extraction and classification are optimized, and edge center loss and cyclic consistency loss functions are used to solve the problem. The problem of insufficient data volume and cross-data set deviations achieve high accuracy identification of new categories of samples.
High accuracy diagnosis of rotary mechanical failures is achieved under very few samples, reducing dependence on large amounts of data, improving the adaptability and diagnostic accuracy of the model, able to monitor mechanical equipment in real time, prevent potential hazards and reduce financial and material loss .
Smart Images

Figure CN114492618B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of rotating machinery fault diagnosis, in particular to a rotating machinery fault diagnosis method of zero-trial learning and feature extraction. Background Art
[0002] Since the beginning of the 21st century, the use of mechanical equipment has promoted the development of the national economy and brought great convenience to production and life. At present, mechanical equipment is developing towards intelligence and precision. However, due to the correlation between parts, a small failure may cause serious consequences and cause huge losses of manpower and material resources.
[0003] In actual environments, mechanical fault diagnosis faces huge challenges. Since mechanical equipment works in harsh environments such as high temperature and high pressure, the mechanical load will change, and the data types are rich, but the single data type is small, which increases the difficulty of fault detection. At the same time, due to the long equipment monitoring cycle and many monitoring points, a huge amount of data will be obtained, but finding useful data in the massive data increases the difficulty of mechanical fault diagnosis.
[0004] In order to solve the problem of scarce fault data, researchers have simulated fault samples in the laboratory, but it is difficult to compare simulated fault samples with real data. Faced with practical problems, scholars aim to achieve fault diagnosis through training with very few samples. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a rotating machinery fault diagnosis method with zero-shot learning and feature extraction, and to establish a rotating machinery fault diagnosis method that learns existing data types to detect new categories by using a very small number of sample learning methods. In addition, in order to make the model have better detection performance for rotating machinery such as bearings or gears, we propose a new generalized zero-shot learning (GZSL) method, which can not only solve the problem of insufficient data volume, but also further correct the problem of cross-dataset deviation, and finally identify rotating machinery faults with high accuracy for samples with the same distribution as training and samples of new categories.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a rotating machinery fault diagnosis method based on zero-trial learning and feature extraction, comprising the following steps:
[0007] S1. Obtain a data set: Obtain a data set and divide it into a training set, a validation set, and a test set;
[0008] S2, data preprocessing and feature extraction: the acquired data set is divided into training data and test data and preprocessed, and then feature extraction is performed on the preprocessed data;
[0009] S3. Establish a feature refinement model based on generalized zero-trial learning: The feature refinement model includes integrating VAE and GAN into a unified VAEGAN generation model, a feature refinement module and a classifier. The VAEGAN generation model includes feature generation VAE and feature generation WGAN. The training process is divided into two stages of learning. The first stage: using the VAEGAN generation model as a framework, the encoder E, decoder G and discriminator D are used to learn the mapping relationship between the feature information and semantic information of the fault information. Then, the feature refinement module is used to conduct adversarial learning on the real fault information and the false fault information, and the real / false fault information is distinguished and represented. The edge center loss and cycle consistency are used for optimization, and the parameters of the feature refinement module are updated. The second stage: the trained feature refinement module is used to refine the feature information of the training data and the test data, and then the refined real fault feature information and the refined synthetic new category fault feature information are sent to the classifier for classification and identification. The model performance is tested using the validation set data. If the detection effect is good, the network parameters are saved, otherwise the hyperparameters are adjusted until the optimal model is obtained.
[0010] S4. After training is completed, the test set data is tested: the test set data is sent to the network, and the features after feature extraction and feature refinement are sent to the trained model to evaluate the final performance of the model.
[0011] A further improvement of the technical solution of the present invention is that: feature extraction is performed on the data set in step S2, and the extracted results are sent to the VAEWAN generation model and feature refinement module established in step S3 for training.
[0012] A further improvement of the technical solution of the present invention is that the feature generation VAE in step S3 includes an encoder E(x, a) and a decoder G(z, a), the encoder E(x, a) encodes the input visible visual feature x into a latent code z, and the decoder G(z, a) reconstructs z into a visual feature Feature generation VAE through a VAE loss To optimize:
[0013]
[0014] in, is the KL divergence, p(z|a) follows the N(0,1) distribution, It is the feature reconstruction loss represented by -logG(z,a).
[0015] A further improvement of the technical solution of the present invention is that the feature generation WGAN in step S3 includes a decoder G(z, a) and a discriminator D(x, a), the decoder G(z, a) synthesizes a visual feature x1 from the random input noise z, the discriminator D(x, a) accepts a real visual feature x or a synthesized visual feature x1 and outputs a real value indicating the degree of truth or falsehood, G(z, a) and D(x, a) are both conditioned on the embedding a, and optimized by the WGAN loss:
[0016]
[0017] Where x′=τx+(1-τ)x1, τ~U(0,1), λ is the weight coefficient.
[0018] A further improvement of the technical solution of the present invention is that: in the step S3, in the feature refinement module, the feature refinement module is optimized using the center loss and the cycle consistency loss function, and The statement is as follows:
[0019]
[0020] Among them, C y is the class prototype of the semantic information of the yth training class embedded in the feature space, C y′ is the y′th (except for the randomly selected class labels of class y) class prototype, Δ represents the marginal information controlling the distance between classes, μ is the encoded feature in the feature refinement module, and γ∈[0,1] is a parameter used to balance the separability of the distance between classes;
[0021]
[0022] in, is the feature information x of the fault sample in the training data using FR i Synthesized semantically relevant features, is the semantically relevant feature synthesized from x1, a is the semantic embedding corresponding to the visual feature x or x1.
[0023] A further improvement of the technical solution of the present invention is that: in step S3, the characteristic information x of the fault sample of the training data is i After feature optimization with pseudo feature information x1, the embedding vector h of the corresponding latent space in the feature refinement module i and h j , semantically related embedding vectors and Connect to form a completely refined feature, i, j are sample data in the training set and validation set respectively, expressed as:
[0024]
[0025]
[0026] Among them, h i and h j All∈H, and All∈A, is a connector, and
[0027] A further improvement of the technical solution of the present invention is that the overall optimization function of the feature refinement model in step S3 is:
[0028]
[0029] Among them, λ SAMC and λ R_a is the weight of the associated loss term.
[0030] A further improvement of the technical solution of the present invention is that: the specific process of step S3 is: the input signal is the input signal x sent to the network after feature extraction i (i=1,2,......,M),x i is the feature information of the fault samples in the training data. In the feature generation VAE, the encoder E(x i ,a) x i The code vector z encoded in the latent space i , and the decoder G(z i ,a) z i Can be reconstructed into visual features In the feature generation WGAN, the decoder G(z i ,a) From the code vector z in the latent space i and semantic information a to synthesize pseudo feature information The discriminator D(x i ,a) Receive a real fault feature information x i Or synthetic pseudo-feature information And output a true value, indicating the degree of truth or falsehood, and extract the fully refined features in the feature refinement module and The real training set feature information x i and the real / pseudo-new category features in the validation set Refine into identification features, and Used to train classifiers for classification.
[0031] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:
[0032] 1. The present invention performs fault diagnosis of rotating machinery with very few samples. Based on VAEGAN, a feature refinement (FR) module is used. This module integrates the semantic information and original feature information of the fault signal into a unified generation model to refine the visual features of known classes and new classes of faults to be tested, and can effectively solve the problem of cross-dataset deviation.
[0033] 2. The present invention introduces adaptive edge center loss in FR to ensure that different data sets can clarify the differences between drum data, so that the network model can better adapt to different data sets. At the same time, in order to better learn the semantic information of mechanical signals and the real feature information of vibration signals, the cycle consistency loss is added for constraint, so that FR can learn more distinguishable classes and related representations;
[0034] 3. The present invention performs fault diagnosis on unknown data when a small amount of known data is needed, which not only reduces the dependence of deep learning methods on large amounts of data, but also realizes the detection of unknown faults. The most important thing is that the labeled data is used reasonably, and mechanical equipment can be monitored in real time to prevent potential dangers. The prediction accuracy is high, which reduces the loss of financial and material resources in this regard. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a process flow chart of the present invention;
[0036] Figure 2 It is the network structure of the feature refinement module of the present invention. DETAILED DESCRIPTION
[0037] The present invention is further described in detail below in conjunction with embodiments:
[0038] The rotating machinery fault diagnosis method based on zero-trial learning and feature extraction includes the following steps:
[0039] S1. Obtain the data sets of bearings and gears required for the experiment, and divide the corresponding data sets into training set, validation set and test set.
[0040] S2. Data preprocessing and feature extraction: To normalize the data, we first need to determine the length of a single sample. After experimental comparison, we found that the effect is better when the one-dimensional data is cut into a length of 2048. Therefore, in the preprocessing process, the data samples are first cut into a sequence of length 2048. In the process of feature processing of the data samples, the WDCNN network is used for feature extraction. The network structure has a total of 5 convolutional layers, and finally the output of the fully connected layer is input into the feature refinement model based on generalized zero-trial learning for processing. The structure of the convolutional layer is conv—>BN—>ReLU—>Maxpool. One convolution layer is a large wide kernel convolution layer, using 16 convolution kernels of size 64×1, and the moving steps of convolution and pooling are 16×1 and 2×1 respectively. The sizes of the convolution kernels from the second to the fifth are 3×1, and the steps are 1. The size of the pooling kernel is 2×1, and the number of convolution kernels is 32, 64, 64, and 64 respectively. Finally, the feature-extracted signal is sent to the network with VAEGAN as the framework through the fully connected layer.
[0041] S3. Establish a feature refinement model based on generalized zero-trial learning: This model consists of a feature generation model VAEGAN, a feature refinement module (FR) and a classifier. VAEGAN integrates VAE and GAN into a unified generation model. It learns the semantic information of vibration signals and the feature mapping of vibration signals through encoder E, decoder G and discriminator D, and synthesizes a large number of feature samples of new categories. The specific structure is shown in Figure 1 In terms of the overall structure, the encoder, generator, and discriminator are multi-layer perceptrons (MLPs) with one hidden layer of 4096 units and LeakyReLU activation function. The feature refinement module is also a type of MLP with two hidden layers with 4096 units and LeakyReLU activation function. units, followed by an encoding layer where the second hidden layer is encoded as two units of size The feature vector of its output layer It is learned via the reparameterization trick and corresponds to the semantic vector of the dataset.
[0042] The feature generation VAE consists of an encoder E(x,a) and a decoder G(z,a). The encoder E(x,a) encodes the input visible visual feature x into a potential code z, while the decoder G(z,a) reconstructs z into a visual feature Feature generation VAE through a VAE loss To optimize:
[0043]
[0044] in, is the KL divergence, p(z|a) follows the N(0,1) distribution, It is the feature reconstruction loss represented by -logG(z,a).
[0045] The feature generation WGAN includes a decoder G(z,a) and a discriminator D(x,a). The decoder G(z,a) synthesizes a visual feature x1 from the random input noise z. The discriminator D(x,a) accepts a real visual feature x or a synthesized visual feature x1 and outputs a real value indicating the degree of truth or falsehood. Both G(z,a) and D(x,a) are conditioned on the embedding a and optimized by the WGAN loss:
[0046]
[0047] Where x′=τx+(1-τ)x1, τ~U(0,1), λ is the weight coefficient.
[0048] After establishing the VAEGAN network based on the feature generation model, construct the feature refinement module and perform optimization training: the composition of the FR module is shown in Figure 2 , which aims to learn the distinguishing representation of real / pseudo signal visual features and perform constrained optimization through center loss and semantic cycle consistency loss. In the feature refinement module, the center loss and cycle consistency loss functions are used to optimize the feature refinement module. and The statement is as follows:
[0049]
[0050] Among them, C y is the class prototype of the semantic information of the yth training class embedded in the feature space, C y′ is the y′th (except for the randomly selected class labels of class y) class prototype, Δ represents the marginal information controlling the distance between classes, μ is the encoded feature in the feature refinement module, and γ∈[0,1] is a parameter used to balance the separability of the distance between classes;
[0051]
[0052] in, Is to use FR by x i Synthesized semantically relevant features, is the semantically relevant feature synthesized from x1, a is the semantic embedding corresponding to the visual feature x or x1.
[0053] The feature information x of the fault samples in the training data i After feature optimization with pseudo feature information x1, the embedding vector h of the corresponding latent space in the feature refinement module iand h j , semantically related embedding vectors and Connect to form a completely refined feature, i, j are sample data in the training set and validation set respectively, expressed as:
[0054]
[0055]
[0056] Among them, h i and h j All∈H, and All∈A, is a connector, and
[0057] The input signal is x after feature extraction and sent to the network. i (i=1,2,......,M),x i is the feature information of the fault samples in the training data. In the feature generation VAE, the encoder E(x i ,a) x i The code vector z encoded in the latent space i , and the decoder G(z i ,a) z i Can be reconstructed into visual features In the feature generation WGAN, the generator G(z i ,a) From the code vector z in the latent space i and semantic information a to synthesize a new feature The discriminator D(x i ,a) Receive a real fault feature information x i Or synthetic pseudo-feature information And output a true value, indicating the degree of truth or falsehood, and extract the fully refined features in the feature refinement module and The real training set feature information x i and the real / pseudo-new category features in the validation set Refine into identification features, and Used to train classifiers for classification.
[0058] The present invention performs fault diagnosis of rotating machinery under very few samples. Based on VAEGAN, a feature refinement (FR) module is used. The module integrates the semantic information and original feature information of the fault signal into a unified generation model to refine the visual features of known classes and new categories of faults to be tested, and can effectively solve the problem of cross-dataset deviation. The adaptive edge center loss is introduced in FR to ensure that different data sets are used to clarify the differences between drum data, so that the network model can better adapt to different data sets. At the same time, in order to better learn the semantic information of mechanical signals and the real feature information of vibration signals, the cycle consistency loss is added for constraint, so that FR can learn more distinguished classes and related representations. The present invention performs fault diagnosis on unknown data when a small amount of known data is required, which not only reduces the dependence of deep learning methods on a large amount of data, but also realizes the detection of unknown faults. The most important thing is that the labeled data is reasonably used, and the mechanical equipment can be monitored in real time to prevent potential dangers from occurring. The prediction accuracy is high, and the loss of financial and material resources in this regard is reduced.
Claims
1. A method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction, characterized in that: The steps include: S1. Obtain a data set: Obtain a data set and divide it into a training set, a validation set, and a test set; S2, data preprocessing and feature extraction: the acquired data set is divided into training data and test data and preprocessed, and then feature extraction is performed on the preprocessed data; S3. Establish a feature refinement model based on generalized zero-trial learning: The feature refinement model includes integrating VAE and GAN into a unified VAEGAN generation model, a feature refinement module and a classifier. The VAEGAN generation model includes feature generation VAE and feature generation WGAN. The training process is divided into two stages of learning. The first stage: using the VAEGAN generation model as a framework, the encoder E, decoder G and discriminator D are used to learn the mapping relationship between the feature information and semantic information of the fault information. Then, the feature refinement module is used to conduct adversarial learning on the real fault information and the false fault information, and the real / false fault information is distinguished and represented. The edge center loss and cycle consistency are used for optimization, and the parameters of the feature refinement module are updated. The second stage: the trained feature refinement module is used to refine the feature information of the training data and the test data, and then the refined real fault feature information and the refined synthetic new category fault feature information are sent to the classifier for classification and identification. The model performance is tested using the validation set data. If the detection effect is good, the network parameters are saved, otherwise the hyperparameters are adjusted until the optimal model is obtained. S4. After training is completed, the test set data is tested: the test set data is sent to the network, and the features after feature extraction and feature refinement are sent to the trained model to evaluate the final performance of the model.
2. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 1, characterized in that: Feature extraction is performed on the data set in step S2, and the extracted results are sent to the VAEWAN generation model and feature refinement module established in step S3 for training.
3. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 1, characterized in that: The feature generation VAE in step S3 includes an encoder E(x, a) and a decoder G(z, a), wherein the encoder E(x, a) encodes the input visible visual feature x into a latent code z, and the decoder G(z, a) reconstructs z into a visual feature Feature generation VAE through a VAE loss To optimize: in, is the KL divergence, p(z|a) follows the N(0,1) distribution, It is the feature reconstruction loss represented by -logG(z,a).
4. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 1, characterized in that: The feature generation WGAN in step S3 includes a decoder G(z,a) and a discriminator D(x,a). The decoder G(z,a) synthesizes a visual feature x1 from the random input noise z. The discriminator D(x,a) accepts a real visual feature x or a synthesized visual feature x1 and outputs a real value indicating the degree of truth or falsehood. Both G(z,a) and D(x,a) are conditioned on the embedding a and optimized by the WGAN loss: Where x′=τx+(1-τ)x1, τ~U(0,1), λ is the weight coefficient.
5. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 4, characterized in that: In the step S3, in the feature refinement module, the center loss and the cycle consistency loss function are used to optimize the feature refinement module. and The statement is as follows: Among them, C y is the class prototype of the semantic information of the yth training class embedded in the feature space, C y′ is the y′th class prototype, y′ is a randomly selected class label other than class y, Δ represents the marginal information controlling the intra-class and inter-class distances, μ is the encoded feature in the feature refinement module, and γ∈[0,1] is a parameter used to balance the separability of intra-class and inter-class distances; in, is the feature information x of the fault sample in the training data using FR i Synthesized semantically relevant features, is the semantically relevant feature synthesized from x1, a is the semantic embedding corresponding to the visual feature x or x1.
6. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 5, characterized in that: In step S3, the characteristic information x of the fault sample of the training data is i After feature optimization with pseudo feature information x1, the embedding vector h of the corresponding latent space in the feature refinement module i and h j , semantically related embedding vectors and Connect to form a completely refined feature, i, j are sample data in the training set and validation set respectively, expressed as: Among them, h i and h j All∈H, and All∈A.
7. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 6, characterized in that: The overall optimization function of the feature refinement model in step S3 is: Among them, λ SAMC and λ R_a is the weight of the associated loss term.
8. The method for fault diagnosis of rotating machinery based on zero-trial learning and feature extraction according to claim 7, characterized in that: The specific process of step S3 is: the input signal is the input signal x sent to the network after feature extraction. i (i=1,2,......,M),x i is the feature information of the fault samples in the training data. In the feature generation VAE, the encoder E(x i ,a) x i The code vector z encoded in the latent space i , and the decoder G(z i ,a) z i Can be reconstructed into visual features In the feature generation WGAN, the decoder G(z i ,a) From the code vector z in the latent space i and semantic information a to synthesize pseudo feature information The discriminator D(x i ,a) Receive a real fault feature information x i Or synthetic pseudo-feature information And output a true value, indicating the degree of truth or falsehood, and extract the fully refined features in the feature refinement module and The real training set feature information x i and the real / pseudo-new category features x in the validation set j / Refine into identification features, and Used to train classifiers for classification.
