A rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and transfer

Through the semi-supervised model comparison migration method, the improved VGG-16 network and self-supervised SimCLR framework are used to solve the problems of poor generalization ability and insufficient label data of migration among different rotating mechanical components, and the failure state with high accuracy is achieved Identify.

CN113723491BActive Publication Date: 2024-12-06HARBIN UNIV OF SCI & TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202110976619.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-08-24
Publication Date
2024-12-06
Estimated Expiration
2041-08-24

AI Technical Summary

Technical Problem

The existing fault diagnosis model is prone to the problem of not converging due to poor generalization capabilities of the model and the lack of label data to reduce the diagnosis rate.

Method used

Using the semi-supervised model comparison migration method, through the improved VGG-16 deep convolutional network and self-supervised SimCLR framework, combined with the SAdam gradient optimization algorithm and the Swish activation function, the source domain data is used for pre-training and fine-tuning in the target domain to achieve different results. Fault status identification between components.

Benefits of technology

The diagnostic accuracy between rotating mechanical components is improved, with an average accuracy of up to 96%, and the efficient diagnostic ability is maintained under a small number of training samples, which is better than the traditional model parameter migration method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113723491B_ABST
    Figure CN113723491B_ABST
Patent Text Reader

Abstract

A rolling bearing and gear fault diagnosis method based on semi-supervised model contrast transfer relates to the technical field of mechanical component fault diagnosis, and is used to solve the problem that the existing fault diagnosis model does not take into account the differences between vibration signals of different components or there are very few labeled samples in the actual training samples, resulting in poor fault diagnosis rate of rotating mechanical components. The present invention is based on the VGG‑16 deep convolutional network training fault diagnosis model. Firstly, it is proposed to use the SAdam gradient descent algorithm to optimize the VGG‑16 deep convolutional network in the source domain to form a pre-training model to prevent negative transfer; then, self-supervision is combined with deep transfer learning, and a method for improving the recognition ability of the transferred neural network model using an improved SimCLR contrast learning framework is proposed, and the projection head activation function in the SimCLR contrast learning framework is changed to a Swish function; further, a small amount of labeled training samples are combined for fine-tuning and joint training to obtain a fault diagnosis model. The present invention can be applied to the classification and judgment of the fault state of mechanical components.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical component fault diagnosis, and in particular to a rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration. Background Art

[0002] Rolling bearings and gears are key components in rotating machinery and are widely used in industrial production. Effective fault diagnosis can prevent serious accidents. [1,2] The working conditions of rolling bearings and gears are complex and changeable, resulting in a lack of labeled training data for certain components. Although there is a frequency difference in fault frequency or meshing frequency between rolling bearings and gears, they are similar. Therefore, it is of great practical significance to obtain a large amount of labeled data from key components of different types of rotating machinery to establish a fault diagnosis model for key components of another type of rotating machinery. [3] .

[0003] Traditional mechanical fault diagnosis methods rely on manual feature extraction. The quality of feature extraction will directly affect the recognition effect of the entire model. [4] In recent years, deep learning has gradually shown its superiority in the application of mechanical intelligent fault diagnosis technology. By providing end-to-end diagnosis, it can avoid the operational errors caused by manual extraction of fault features to a certain extent. Reference [5] proposed a bearing fault diagnosis method based on kernel function and denoising autoencoder, which improved the ability of deep autoencoder network to handle nonlinear problems. Reference [6] directly used the original vibration signal as the input of the bidirectional long short-term memory network diagnosis model to automatically extract the rolling bearing fault features and realize the fault identification of rolling bearings with different fault types and different damage degrees. Reference [7] used deep coupled dense convolutional network to integrate information fusion, feature extraction and fault classification to realize intelligent fault diagnosis of planetary gearboxes. Reference [8] introduced residual learning mechanism to train one-dimensional convolutional autoencoder, which was successfully applied to unsupervised learning and fault feature extraction of vibration signals, significantly improving the fault diagnosis rate of gearboxes. Reference [9] extracted features from time domain and frequency domain signals, used local and global principal component analysis to fuse multi-source domain deep features into inherent low-dimensional features, and used integrated kernel extreme learning machine to achieve good diagnostic results on gearboxes and rolling bearings.

[0004] The successful development of a deep learning model framework often requires a large amount of historical fault data support. In actual work, some rotating machinery parts have poor working conditions, and some parts cannot be installed with sensors for a long time due to physical limitations, resulting in high data acquisition costs. It is difficult to obtain a large amount of labeled vibration data of some rotating machinery parts, and there is a lack of sufficient label information to establish a more effective deep learning fault diagnosis model.

[0005] As a major breakthrough in the field of artificial intelligence, transfer learning has received widespread attention in the academic community in recent years. Its purpose is to use the existing knowledge in the source domain to solve problems in the related but not identical target domain. Current transfer learning strategies are mainly divided into four categories: instance-based transfer, relationship-based transfer, feature-based transfer, and model-based transfer.

[10] At present, many scholars have successfully applied transfer learning methods to the fault diagnosis of key components of rotating machinery and achieved good results. Reference

[11] used the TrAdaBoost algorithm based on instance transfer, and optimized the training set by improving its iterative process and output mechanism. It was successfully applied to multi-classification tasks and improved the bearing fault diagnosis rate. Reference

[12] quoted the relationship-based transfer knowledge to extract the relationship between the two types of faults in the source domain gearbox, and applied it to the target domain, which can identify the faulty gear of the target domain machine. However, the above algorithms have the disadvantages of being limited by sample features and quantity. Domain adaptation research based on feature transfer is the most popular research method in the field of transfer learning, and has been successfully applied to fault diagnosis scenarios under different working conditions such as different loads and different equipment. Reference

[13] proposed a fault diagnosis method for deep feature transfer of rolling bearings of different models, introduced a conditional adversarial mechanism, realized the simultaneous adaptation of source domain and target domain features and labels, and solved the problem of missing label information in the target domain data. Reference

[14] used the sufficient motor bearing and gearbox data in the laboratory to identify the health status of locomotive equipment in the actual working environment, and adopted a feature transfer neural network to obtain a high fault diagnosis accuracy.

[0006] The prerequisite for the successful application of domain adaptation methods is that the number of categories in the source domain and the target domain are the same. However, when migrating between key components of different rotating machinery, the number of categories in the source domain and the target domain may be different, which makes it difficult to map features between classes and easily leads to overfitting. The model-based migration method refers to building a parameter-sharing model. It does not require the source domain and the target domain to have exactly the same number of categories. It can effectively avoid overfitting during migration. Reference

[15] uses the AlexNet model as a pre-trained model to migrate its weights and only replaces the last fully connected layer to achieve rolling bearing fault diagnosis, reducing prior knowledge and the time required to establish a new model. Reference

[16] constructs a boosted deep autoencoder and uses the source domain data for pre-training. The obtained parameters are used as the initialization parameters of the target model. Only one training sample in the target domain is used to fine-tune the target model to achieve migration diagnosis of rolling bearings between different mechanical equipment. Reference [3] proposes a fault diagnosis method based on a one-dimensional convolutional neural network. It achieves migration between different rotating machinery components by fine-tuning parameters, which has good results and robustness.

[0007] Since there are large differences between the vibration signals of different components, if the traditional model parameter migration strategy is directly applied to the diagnostic migration between different components, it is easy to cause non-convergence due to the poor generalization ability of the model. In actual engineering, it is difficult to obtain labeled data. When the target field contains no labels or very few labels, the direct application of the migration model will lead to a decrease in the diagnostic rate. Summary of the invention

[0008] In view of the above problems, the present invention proposes a rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration, so as to solve the problem that the existing fault diagnosis model has a poor fault diagnosis rate for rotating machinery components because the differences between vibration signals of different components are not considered or the actual training samples contain very few labeled samples.

[0009] A rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration, comprising the following steps:

[0010] Step 1: Obtain vibration signals of all states of mechanical component A as source domain data; obtain a small amount of vibration signals with state labels of other mechanical components different from mechanical component A as target domain data;

[0011] Step 2: Preprocess the source domain and target domain data to obtain a source domain image dataset and a target domain image dataset;

[0012] Step 3: Input the source domain image dataset into the improved source domain VGG-16 deep convolutional network for pre-training, obtain the pre-training model, and save the pre-training model network parameters;

[0013] Step 4: Migrate the pre-trained model network parameters to the target domain VGG-16 deep convolutional network to initialize the model network parameters, input the target domain image dataset into the target domain VGG-16 deep convolutional network and introduce the self-supervised SimCLR framework for training, and adaptively update the network parameters of the target domain VGG-16 deep convolutional network through the SimCLR loss function to obtain the fault state classification model;

[0014] Step 5: Input the vibration signal of the unknown state of the mechanical component to be diagnosed into the fault state classification model after the preprocessing in the above step 2 to obtain the state recognition result of the mechanical component to be diagnosed.

[0015] Furthermore, the state in step one includes a normal state and a fault state, and the fault state includes an inner ring fault, a rolling element fault, an outer ring fault and states of different damage degrees.

[0016] Furthermore, the preprocessing in step 2 includes performing short-time Fourier transform on the source domain and target domain data to construct a multi-state two-dimensional source domain image dataset and a target domain image dataset, respectively.

[0017] Furthermore, the improvements of the improved VGG-16 deep convolutional network described in step 3 include: using the SAdam gradient optimization algorithm to update the network parameters to accelerate the convergence of the VGG-16 deep convolutional network; wherein the network parameter update process is expressed as:

[0018]

[0019] Where D represents bounded space; x t represents a decision selected from D; represents the deformation of the squared second-order momentum of the gradient after adding the vanishing factor; α represents the learning rate; t represents the time; represents the first-order momentum.

[0020] Furthermore, in step 4, a model migration parameter fine-tuning strategy is adopted to reduce the number of training parameters, specifically: when migrating the pre-trained model network parameters to the target domain VGG-16 deep convolutional network, the fully connected network parameters in the pre-trained model network parameters are not migrated, but new fully connected network parameters are added to the target domain VGG-16 deep convolutional network for classification; during the training process of the target domain VGG-16 deep convolutional network, the convolutional layers are gradually unfrozen and fine-tuned using a small amount of target domain image data sets carrying state labels.

[0021] Furthermore, in step 4, the process of inputting the target domain image dataset into the target domain VGG-16 deep convolutional network and introducing the self-supervised SimCLR framework for training includes: first, all input images are randomly transformed for data enhancement to generate two related enhanced images from the same image; after data enhancement, the basic encoder network, i.e., the target domain VGG-16 deep convolutional network, is connected to extract the representative feature vector from the two enhanced data images; then, the representative feature vector is mapped to the vector space where the contrast loss is applied, and the projection vector is obtained using a fully connected linear neural network with one hidden layer; finally, the overall loss is calculated by calculating the cosine similarity between the two enhanced images projected into the vector space, and the contrast loss function used by SimCLR is the normalized temperature-scale cross entropy loss.

[0022] Furthermore, during the training process of introducing the self-supervised SimCLR framework, the ReLU activation function is replaced by the Swish activation function when mapping the feature vector to the vector space where the contrastive loss is applied.

[0023] Furthermore, the loss function of the SimCLR is the average of the losses of all different image pairs, and the calculation formula is:

[0024]

[0025] Where N is the batch size and l is the loss of all distinct image pairs.

[0026] The beneficial technical effects of the present invention are:

[0027] Based on the characteristics of powerful fitting ability and simple training of VGG-16 deep convolutional network, firstly, the present invention proposes to use SAdam gradient descent algorithm to optimize the VGG-16 deep convolutional network of the source domain to form a pre-training model, and uses its strong convexity to prevent negative transfer. The experimental results of source domain performance show that the improved source domain training process has stable convergence and high recognition accuracy; then, the source domain network model parameters are fixed and then migrated to the target domain, which can effectively utilize the source domain knowledge and reduce the modeling time. The present invention combines self-supervision with deep transfer learning, and proposes a method for improving the recognition ability of the transferred neural network model by using the improved SimCLR contrastive learning framework, and changes the projection head activation function in the SimCLR contrastive learning framework to the Swish function to cope with the complex features generated by the enhanced image; further, after the self-supervised learning mines its own supervision information, it is fine-tuned in combination with a small number of labeled training samples to jointly train a high-quality model to achieve transfer diagnosis under semi-supervised conditions.

[0028] The present invention proposes migration diagnosis between key components of rotating machinery, and uses sufficient label information of a certain component to assist other components in identifying their health status. Experimental results show that the method of the present invention has a high diagnostic accuracy even with a small number of training samples, and is at least 5% higher than the compared model parameter migration method. The average diagnostic accuracy can reach 96%, proving the effectiveness of the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification, and are used to further illustrate the preferred embodiments of the present invention and explain the principles and advantages of the present invention.

[0030] Figure 1 Schematic diagram of the network structure of the VGG-16 convolutional neural network in the present invention;

[0031] Figure 2 Schematic diagram of the network parameter migration process of the VGG-16 convolutional neural network from the source domain to the target domain in the present invention;

[0032] Figure 3 It is a comparison diagram of the ReLU activation function and the Swish activation function in the present invention;

[0033] Figure 4 It is a flowchart of the method of the present invention;

[0034] Figure 5 is a schematic diagram of a bearing test bench in an embodiment of the present invention;

[0035] Figure 6 This is a graph of the training set results using the SAdam optimization algorithm in an embodiment of the present invention;

[0036] Figure 7 This is a test set result graph using the SAdam optimization algorithm in an embodiment of the present invention;

[0037] Figure 8 This is a graph of network characteristics results before improvement in an embodiment of the present invention;

[0038] Fig. 9 : is a visualization result diagram of improved network features in an embodiment of the present invention (unfreezing the last layer of fine-tuning);

[0039] Fig.10 This is a visualization result diagram of improved network features in an embodiment of the present invention (unfreezing the last two layers of fine-tuning);

[0040] Fig.11 This is a visualization result diagram of improved network features in an embodiment of the present invention (unfreezing the last three layers of fine-tuning);

[0041] Fig.12 is a comparison chart of diagnostic accuracy of training sample sets with different ratios in an embodiment of the present invention;

[0042] Fig.13 It is a graph showing experimental comparison results between the method of the present invention and other methods. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the scheme of the present invention, exemplary implementations or embodiments of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described implementations or embodiments are only implementations or embodiments of a part of the present invention, not all of them. Based on the implementations or embodiments of the present invention, all other implementations or embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0044] In view of the negative transfer phenomenon between different components and the lack of sufficient labeled samples in the target domain, which leads to the weakening of model generalization ability and the decrease of diagnosis rate, the present invention proposes a rolling bearing and gear fault diagnosis method with semi-supervised model comparison and migration. This method constructs a two-dimensional image time-frequency spectrum atlas from the vibration signals of key components of rotating machinery. First, the SAdam gradient optimization algorithm is used to accelerate the convergence of the source domain VGG-16 network to form a pre-training model. After the pre-training model parameters are migrated to the target domain network, the proposed improved self-supervised SimCLR framework is used to learn the unlabeled sample features and adaptively update the target domain network parameters; finally, a small amount of labeled training samples in the target domain are used for fine-tuning to achieve migration diagnosis between rolling bearings and gears of different components under semi-supervised conditions. The method of the present invention is described in detail below.

[0045] 1. Model transfer training based on VGG-16

[0046] VGG-16 is a deep convolutional neural network improved on the basis of AlexNet. It focuses on building a concise and regular structural network with simple and efficient training. VGG-16 uses several consecutive 3×3 convolution kernels to replace the larger convolution kernels (11×11, 7×7, 5×5) in AlexNet. This replacement of multiple 3×3 convolution kernels can obtain the same receptive field as the large convolution kernel, reducing the number of network parameters to be trained. The increased network depth will also deepen the learning of samples; at the same time, three nonlinear activation layers are integrated to replace a single nonlinear activation layer to enhance the nonlinear expression ability of the model. Take the 3×3 convolution kernel instead of the 5×5 convolution kernel as an example. Figure 1 As shown in the figure, for two 3×3 convolution kernels, the total number of parameters used is 2×(3×3)=18, and for a 5×5 convolution kernel, the total number of parameters used is 5×5=25, so the training parameters can be reduced.

[0047] Establishing a pre-trained model with strong generalization ability in the source domain helps to migrate excellent model parameters. The deeper network layers of VGG-16 have great advantages in mining deep features of data. It is basically the same as the typical CNN, consisting of convolutional layers, pooling layers, and fully connected layers. Each convolution kernel of the convolutional layer performs convolution calculations with the data in the convolutional layer with a fixed size and generates corresponding features. At the same time, a nonlinear activation function is used to enhance the generalization ability and prevent the gradient from disappearing. The convolution operation is:

[0048]

[0049] Where: represents the j-th output of the l-th layer, represents the i-th input of the l-1th layer, k is the convolution kernel, b is the network bias, and the multiplication sign represents the convolution operation; f is the activation function, ReLU is commonly used, and ReLU represents a nonlinear activation function, that is, a corrected linear unit, to prevent the occurrence of overfitting problems.

[0050] The output obtained after convolution will maintain a high dimension, which will result in a large amount of calculation if directly applied to classification tasks. Therefore, a pooling layer is usually required after each convolution layer to perform data scaling mapping. The maximum pooling process formula is expressed as:

[0051]

[0052] Where: represents the value of the tth neuron in the jth feature vector of the lth layer, W is the width of the pooling area, Represents the value corresponding to the neuron in the l+1th layer.

[0053] After a series of convolutions and dimensionality reduction, it enters the output part of CNN, that is, the output of the last pooling layer is fully connected head to tail and the Softmax classifier is used to classify the abstract high-level features, thereby solving the multi-classification problem:

[0054] H=f(b h +w h f v ) (3)

[0055] Where: f v is the feature vector, b h is the deviation vector, w h is the weight matrix.

[0056] Traditional deep network training often uses stochastic gradient descent (SGD) or adaptive moment estimation (Adam) to optimize network parameters. SGD has no concept of momentum, so it is difficult to choose a suitable learning rate. The uniformly scaled gradient in all directions is not suitable for sparse data, and the descent speed is slow and may fall into a local minimum and fail to converge. Adam combines first-order momentum and second-order momentum. Compared with SGD, it can automatically adjust the parameter learning rate, but Adam may also fail to converge in some cases. SAdam is a variant of the Adam optimization algorithm. The basic idea is to use strong convexity to maintain a faster convergence speed while controlling the step size and to obtain huge gains when facing sparse data.

[17] .

[0057] During the pre-training model training process, the SAdam optimization algorithm of the improved traditional optimizer is used to speed up the gradient convergence, and the optimization algorithm expression is:

[0058]

[0059]

[0060] V t =β 2t V t-1 +(1-β 2t )diag(g t g t T ) (6)

[0061]

[0062] Where: β is the attenuation rate; g t is the gradient value; is the first-order momentum; V t is the second-order momentum of the squared gradient; Deformation after adding the vanishing factor δ to the gradient squared second-order momentum.

[0063] Network model parameter update process:

[0064]

[0065] Where: D is the bounded space; α is the learning rate.

[0066] The purpose of model migration is to find parameter information that can be shared between the source domain and the target domain to assist the training of the target domain. Deep convolutional networks can be used to mine image features. Shallow convolutional layers learn the features of local areas and can extract basic features such as curves and contours. Deep convolutional layers have a larger receptive field and can learn abstract features. The parameters embedded in the deep convolutional network can be migrated to the target task. For tasks with small amounts of data and low data similarity, lower-level representations can be migrated and higher-level representations can be learned from new data sets.

[0067] The parameters of the pre-trained model trained with source domain data may not have strong generalization ability after migration, and it is difficult to be used for direct classification of target domain data. Fine-tuning is a commonly used parameter adjustment method that can automatically correct network parameters for specific tasks. Using a small amount of labeled data from the target domain to fine-tune the model during training can improve the diagnostic effect of the model after migration. The key to the success of fine-tuning depends on the similarity between the source domain data and the target domain data. For similar data sets, only the output category of the fully connected layer needs to be modified, while for data sets with large differences, only a few convolution blocks need to be fine-tuned. Compared with training from scratch, fine-tuning essentially reduces the number of parameters that need to be trained, saves computing time, and improves computing efficiency and accuracy.

[0068] The model migration parameter fine-tuning strategy adopted by the present invention is: the parameters of the pre-trained model (except the fully connected network) are fixedly migrated to the target domain network to initialize the network parameters, a new fully connected network is added to the target domain for classification, the convolutional layers are gradually unfrozen and a small amount of labeled data from the target domain is used to fine-tune the model effect. The model parameter migration process is as follows: Figure 2 shown.

[0069] 2. Self-supervised contrastive learning framework

[0070] Self-supervised learning trains the network by mining its own supervision information, thereby learning valuable representations for downstream tasks. Contrastive learning is a representative method of self-supervised learning, which aims to train machine learning models to distinguish similar and different objects. The present invention expounds and improves the relevant theories of SimCLR, and proposes a new fault diagnosis method that combines deep model migration with the SimCLR framework to improve the recognition ability of the migrated model.

[0071] The proposed visual representation contrastive learning framework (SimCLR) improves the existing self-supervised methods and has performance comparable to supervised training. The framework mainly consists of the following four parts:

[18] :

[0072] 1) Data enhancement: The data enhancement module randomly transforms all input images to generate two related enhanced images from the same image. and Image augmentation methods with random transformations applied: random cropping (cropping, resizing, and flipping) followed by resizing to the original size, random color distortion, and random Gaussian blur.

[0073] 2) Basic encoder: After data enhancement, the basic encoder network f(·) is connected to extract the representative feature vector from the enhanced data image. The basic encoder of this framework selects VGG-16 to obtain the representative feature vector:

[0074]

[0075] Where: h i ∈Rd is the output after the average pooling layer.

[0076] 3) Projection head: The neural network projection head g(·) maps the representation feature vector to a vector space where contrastive loss is applied. The projection vector is obtained using a fully connected linear neural network with one hidden layer:

[0077] z i =g(h i )=W (2) σ(W (1) h i ) (10)

[0078] Where: σ represents the activation function, and the traditional SimCLR framework uses the ReLU activation function.

[0079] 4) Contrastive loss function: Calculate the cosine similarity between two enhanced images projected into the vector space:

[0080]

[0081] Where: ||·|| is the norm of the vector.

[0082] The overall loss is calculated using the cosine similarity formula. The contrast loss function used by SimCLR is NT-Xent loss (normalized temperature-scale cross entropy loss). The loss function for each pair of different images is calculated as follows:

[0083]

[0084] Where: sim(z i ,z j ) represents the cosine similarity between two vectors, and τ represents the temperature parameter.

[0085] The final loss of the target model is the average of the losses of all mutually different image pairs:

[0086]

[0087] The overall loss function is minimized through formula calculation. As time goes by, the representations of the encoder and projection head place similar images in closer positions in the comparison space, expanding the similarities between images of the same type and the differences between images of different types, and the recognition ability of the neural network model gradually improves.

[0088] Choosing a suitable activation function in a neural network has a significant impact on the training process and task performance. Currently, the most widely used activation function is the linear rectification function. The ReLU activation function expression is:

[0089] y=max(0,x) (14)

[0090] The ReLU activation function has the characteristics of having an upper bound but no lower bound and being smooth overall, and it solves the gradient vanishing problem that exists in the back propagation of the sigmoid and tanh activation functions.

[19] Like ReLU, it has no lower bound but an upper bound and is smooth, but the Swish activation function is non-monotonic, suitable for local response normalization, and helps the model converge quickly. It is superior to traditional ReLU in deep models and complex data sets. The expression of the Swish activation function is:

[0091] y=x·sigmoid(βx) (15)

[0092] sigmoid(z)=(1+exp(-z)) -1 (16)

[0093] β can be either a constant or a trainable parameter. If β=1, Swish is equivalent to the sigmoid weighted linear unit commonly used in reinforcement learning. If β=0, Swish is equivalent to the scaled linear function f(x)=x / 2. If β tends to infinity, Swish is equivalent to the ReLU function. Swish can perform nonlinear interpolation between linear functions and ReLU functions. The degree of interpolation is controlled by β, which has higher controllability and flexibility. The image of the Swish activation function when ReLU and β=1 is as follows: Figure 3 shown.

[0094] The present invention adopts the Swish function to replace the ReLU activation function used in the projection head of the traditional SimCLR framework to cope with the complex features generated after image enhancement and help the deep network model converge quickly. The pseudo code of the SimCLR algorithm process is shown below.

[0095]

[0096] 3. Fault diagnosis methods for key components of rotating machinery

[0097] The fault diagnosis process of key components of rotating machinery based on semi-supervised model comparison and migration is shown in Figure 2. Figure 4 As shown, the specific steps are:

[0098] 1) Data selection and preprocessing.

[0099] All state vibration signals of a known component A (rolling bearing or gear) are obtained as source domain data. A small number of known label vibration signals of other components (gear or rolling bearing) are used as target domain data. Short-time Fourier transform is performed on the source domain and target domain data to construct a multi-state two-dimensional image dataset, which is input into the source domain network and the target domain network.

[0100] 2) Source domain model training.

[0101] The VGG-16 network is trained iteratively multiple times using source domain data to form a pre-trained model. During the training process, the SAdam gradient optimization algorithm is used to accelerate network convergence and prevent negative transfer.

[0102] 3) Model parameter migration and target domain network model training.

[0103] The fixed pre-trained model parameters (except the fully connected network) are migrated to the target domain network to initialize the network parameters. The self-supervised SimCLR framework is introduced to assist the target domain network in training. The projection head activation function is changed to the Swish function. By mining the deep information of unlabeled data, the model can learn to better distinguish different types of images. The target domain network parameters are adaptively updated through the SimCLR loss function to improve the generalization ability of the model after migration.

[0104] 4) Multi-state recognition.

[0105] A new classification convolutional block is added for downstream classification tasks. After a small amount of labeled data in the target domain is used to fine-tune the overall network, a semi-supervised model is finally established to compare and transfer the multi-state recognition model of key components of rotating machinery. The predicted labels of the target domain are compared with the true labels to obtain the accuracy of multi-state recognition and measure the classification performance of the model.

[0106] 4. Application and analysis

[0107] The rolling bearing experimental data set used in the embodiment of the present invention comes from the Electrical Engineering Laboratory of Case Western Reserve University in the United States.

[20] The schematic diagram of the test bench is as follows: Figure 5 As shown, the deep groove ball bearing model at the motor drive end is SKF6205, and the deep groove ball bearing model at the fan end is SKF6203. A 16-channel data recorder is used to collect vibration signals with a sampling frequency of 12kHz.

[0108] The experiment collected vibration signals of rolling bearings of different models under four different loads of 0hp, 1hp, 2hp, and 3hp, including normal state, inner ring fault, rolling element fault, outer ring fault and different damage degrees, a total of 40 states. For the convenience of description, the fault states of the inner ring damage diameter of the rolling bearing under 0hp of 0.1778mm, 0.3556mm, and 0.5334mm are represented by 0-IR07, 0-IR14, and 0-IR21, respectively, the different fault states of the rolling element are represented by 0-B07, 0-B14, and 0-B21, respectively, the different fault states of the outer ring are represented by 0-OR07, 0-OR14, and 0-OR21, respectively, and the normal state is represented by 0-N.

[0109] The embodiment of the present invention selects 6 sets of data sets. Data set A is the vibration data of the rolling bearing of model SKF6205, and data set B is the vibration data of the rolling bearing of model SKF6203. Each data set contains 40 states, 300 samples for each state, and a total of 12,000 samples. Data set C is the vibration data of the model 6307E rolling bearing collected by the laboratory test bench, with a sampling frequency of 8.192kHz and a speed of 680rps. The experiment uses three state data of inner ring fault, outer ring fault and normal. 300 samples for each state, a total of 900 samples. Data set D is the vibration data of the NICE bearing in the United States, with a ball diameter of 5.969mm, a pitch circle diameter of 31.623mm, a sampling frequency of 97656Hz, and a speed of 25rps. The experiment uses vibration data of three states of inner ring fault, outer ring fault and normal. 300 samples for each state, a total of 900 samples. Dataset E is the gear vibration data from the QPZZ-II rotating machinery vibration analysis and fault diagnosis test platform system, with a sampling frequency of 5.12kHz and a rotation speed of 880rps. Gear dataset E includes six states under each load. The experiment uses a 0.2A load, 300 samples for each state, and a total of 1800 samples. Dataset F is a gear dataset with a primary meshing frequency of 250Hz, a secondary meshing frequency of 50Hz, a sampling frequency of 5.12kHz, a rotation speed of 750rps, and the shaft frequencies of the three shafts where the gears are located are 12.5Hz, 3.125Hz, and 0.61Hz respectively. The fault occurred on the gear with a shaft frequency of 12.5Hz, including three states of broken teeth, cracks, and normal, with 300 samples for each state, and a total of 900 samples. The specific experimental data sets are shown in Tables 1 and 2.

[0110] Table 1 Bearing experimental data set

[0111]

[0112] Table 2 Gear experimental data set

[0113]

[0114] According to the fault diagnosis problem of different parts of rotating machinery proposed by the present invention, migration tasks such as E→C, F→C, E→D, F→D, A→E, B→E, A→F, B→F, C→F and D→F are used to verify the effectiveness of the method of the present invention, where E→C represents the knowledge migration of the source domain data set E to the target domain data set C. In the experiment, the ratio of the training set, validation set and test set of the source domain and the target domain is set to 70:15:15, and in order to adapt to the data imbalance under actual working conditions, the number of samples of each category in the training set, validation set and test set of the target domain adopts a random allocation and unequal distribution strategy. Taking the rolling bearing data set C as an example, the division of the target domain sample set is shown in Table 3.

[0115] Table 3 Target domain sample set of rolling bearing dataset C

[0116]

[0117] The source domain was compared before and after the SAdam optimization algorithm was used, before and after fine-tuning, and before and after the target domain SimCLR framework projection head used the Swish activation function. Keras was used as the learning framework in the experiment. According to multiple experimental results and experience, the learning rate was uniformly set to 0.001. In order to reduce the impact of randomly initialized parameters to be trained on the diagnostic results of the proposed method, the experiment was repeated five times with a five-fold cross validation and the average was taken. The classification task used accuracy and F1 value as evaluation indicators:

[0118]

[0119]

[0120]

[0121] The experiment takes the migration task gear dataset E→rolling bearing dataset C as an example. During the source domain network training process, the SAdam optimization algorithm is used to accelerate the gradient convergence speed. The experimental results compared with the traditional optimization algorithm are as follows: Figure 6 and Figure 7 As shown. Figure 6 and Figure 7 It can be seen that when using the traditional optimization algorithm, the loss function gradually stabilizes after about 30 iterations; when using the SAdam optimization algorithm, the loss function remains stable after about 10 iterations. The SAdam optimization algorithm saves about 60% of the training time and can effectively speed up the convergence speed.

[0122] Under the premise that the source domain refers to the SAdam optimization algorithm and the target domain SimCLR framework projection head adopts the Swish activation function, the target domain adopts the strategy of unfreezing the last three layers of convolution fine-tuning. The experimental results are compared with the target domain directly using Softmax classification after transferring parameters, as shown in Table 4.

[0123] Table 4 Experimental results before and after using the frozen fine-tuning strategy (accuracy % / F1 value %)

[0124]

[0125] In order to intuitively prove the effectiveness of unfreezing the last three layers of convolution fine-tuning and the proposed method, the t-distribution stochastic neighbor embedding (t-SNE) algorithm is used.

[21] , the features of the last fully connected layer of the improved network are reduced to two dimensions and presented in the form of a scatter plot, as shown in Figure 8 , 9 , 10, and 11. It can be seen that the model before the improvement has obvious misclassification and unclear category boundaries; the improved model has clear category boundaries, less sample aliasing, and Fig.11 As shown in the figure, the strategy of unfreezing the last three layers of convolution fine-tuning makes the samples clustered tightly and can better distinguish the categories. It further proves that the proposed method can be applied to the migration diagnosis between different components, bearings and gears, and can obtain a higher classification accuracy. In the future, the strategy of unfreezing the three layers of fine-tuning will be used to test the classification accuracy of different migration tasks.

[0126] Under the premise of using the SAdam optimization algorithm in the source domain and performing fine-tuning freezing, the activation function of the projection head in the SimCLR framework uses Swish, which is compared with the ReLU activation function. The experimental results are shown in Table 5.

[0127] Table 5 Experimental results before and after changing the projection head activation function

[0128]

[0129] It can be seen from Table 5 that after changing the projection head activation function, the fault recognition accuracy is improved. The reason is that the Swish activation function can effectively process the complex features generated after image enhancement, so that the complex features can express more useful information.

[0130] In the process of model parameter migration experiment, SimCLR framework was introduced to improve the recognition ability of basic model. Swish was used as the projection head activation function. The experiment was repeated five times with five-fold cross validation and the average was taken to achieve multi-classification. The classification task used accuracy and F1 value as evaluation indicators. The data sets used are shown in Tables 1 and 2. The experimental results of different migration tasks are shown in Table 6.

[0131] Table 6 Experimental results of different migration tasks

[0132]

[0133] As can be seen from Table 6, the average accuracy of the method recognition is over 96%, and the average F1 value is over 96%, which fully demonstrates that the knowledge of rolling bearing status and the knowledge of gear status can be transferred to each other, that is, the mutual diagnosis of rolling bearings and gears can be achieved through knowledge transfer, and it has a very good effect.

[0134] In order to fully demonstrate the applicability and generalization of the proposed method, 5%, 20% and 100% of the target domain training sample sets of different migration tasks are used as new target domain training sample sets, and the number of validation set and test set samples is still the same as the previous experiment. Taking the migration task E→C as an example, the specific comparison experiment average effect is as follows Fig.12 It can be seen that when VGG-16 is used alone for migration classification without the SimCLR framework, a small number of training samples will significantly reduce the diagnosis rate, while the improved method proposed in the present invention also achieves good results with a small number of samples, further proving that the method of the present invention can be applied to the fault diagnosis of rolling bearings and gears of different rotating machinery.

[0135] To further illustrate that the proposed method has great advantages in the migration fault diagnosis of rolling bearings and gears of rotating machinery components, VGG16 deep model (without migration, only for the target domain dataset classification), VGG-16 migration model (without SIMCLR framework), reference [3], and reference

[15] were selected for comparative experimental research. During the experiment, the same source domain and target domain datasets and the same migration task were set. The comparative experimental results are shown in Figure 2. Fig.13 As shown. Fig.13 It can be seen that the average accuracy of the method of the present invention can reach 96% in the experiments of 10 different migration tasks, and the accuracy is improved by at least 5% compared with traditional deep learning classification, direct classification of VGG-16 migration model without SimCLR framework and other model migration methods. Therefore, the method based on semi-supervised model comparison migration proposed in the present invention can effectively solve the fault diagnosis problem of rolling bearings and gears, which are key components of rotating machinery.

[0136] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments are contemplated within the scope of the invention thus described. The disclosure of the present invention is intended to be illustrative rather than restrictive of the scope of the invention, which is defined by the appended claims.

[0137] The documents cited in the present invention are as follows:

[0138] [1] Kang Shouqiang, Zou Jiayue, Wang Yujing, et al. Rolling bearing fault diagnosis method under variable load based on unsupervised feature alignment[J]. Proceedings of the CSEE, 2020, 40(1): 274-281.

[0139] [2] Li Yibing, Huang Dinghong, Ma Jianbo, et al. Gear fault diagnosis method based on deep belief network and information fusion [J]. Vibration and Shock, 2021, 40(8): 62-69.

[0140] [3] Li Xudong, Hu Yang, Li Mingtao, et al. fault diagnostics between different type of components: a transfer learning approach [J]. IEEE Access, 2019: 42373-42383.

[0141] [4] Yang Yi, Fan Dongchen, Yin Haoran, et al. Research on fault phase selection model and transferability of transmission lines based on deep transfer learning [J]. Electric Power Automation Equipment, 2020, 40(10): 165-172.

[0142] [5] Wang Fengtao, Liu Xiaofei, Dun Bosen, et al. Application of nuclear autoencoder based on firefly optimization in intermediate bearing fault diagnosis[J]. Journal of Mechanical Engineering, 2019, 55(7): 58-64.

[0143] [6] Zhao Zhihong, Zhao Jingjiao, Wei Ziyang, et al. Research on rolling bearing fault diagnosis based on BiLSTM[J]. Vibration and Shock, 2021, 40(1): 95-101.

[0144] [7]Jiao Jinyang, Zhao Ming, Lin Jing, et al. Deep coupled denseconvolutional network with complementary data for intelligent fault diagnosis[J]. IEEE Transactions on Industrial Electronics, 2019, 66(12): 9858-9867.

[0145] [8] Zhou Xingkang, Yu Jianbo. Gearbox fault diagnosis based on deep one-dimensional residual convolutional autoencoder network[J]. Journal of Mechanical Engineering, 2020, 56(7): 96-108.

[0146] [9] Pang Shan, Yang Xinyi, Zhang Xiaofeng, et al. Fault diagnosis ofrotating machinery with ensemble kernel extreme learning machine based on fused multi-domain features[J]. ISA Transactions, 2020, 98: 320-337.

[0147]

[10] Pan SJ, Yang Qiang.A survey on transfer learning[J].IEEE Transactions on Knowledge and Data Engineering, 2010, 22(10): 1345-1359.

[0148]

[11] Chen Renxiang, Chen Siyang, Yang Lixia, et al. Rolling bearing fault diagnosis based on improved TrAdaBoost multi-classification algorithm [J]. Vibration and Shock, 2019, 38(15): 36-48.

[0149]

[12] Yan Ruqiang, Shen Fei, Sun Chuang, et al. Knowledge transfer for rotary machine fault diagnosis [J]. IEEE Sensors Journal, 2020, 20(15): 8374-8393.

[0150]

[13] Wang Yujing, Lü Haiyan, Kang Shouqiang, et al. Fault diagnosis methods for rolling bearings of different types[J]. Proceedings of the CSEE, 2021, 41(1): 267-276.

[0151]

[14] Yang Bin, Lei Yaguo, Jia Feng, et al. An intelligent fault diagnosis approach based on transfer learning from laboratory bearings to locomotivebearings[J]. Mechanical Systems and Signal Processing, 2019, 122: 692-706.

[0152]

[15] Wang Jianyu, Mo Zhenling, Zhang Heng, et al. A deep learning method for bearing fault diagnosis based on time-frequency image [J]. IEEE Access, 2019, 7: 42373-42383.

[0153]

[16] Shao Haidong, Zhang Xiaoyang, Cheng Junsheng, et al. Intelligent fault diagnosis of bearings based on boosted deep migration autoencoder[J]. Journal of Mechanical Engineering, 2020, 56(9): 84-90.

[0154]

[17] Wang Guanghui, Lu Shiyin, Cheng Quan, et al.SAdam: A variant of Adam for strongly convex functions[C].ICLR 2020.

[0155]

[18] Chen Ting, Kornblith S, Norouzi M, et al.A simple framework for contrastive learning of visual representations[C].ICML 2020.

[0156]

[19] Ramachandran P, Zoph B, Le Q V.Searching for Activation Functions[J].arXiv preprint arXiv:1710.05941,2017.

[0157]

[20] LOPARO, K A. Bearing data center [M / OL]. Cleveland: Case Western Reserve University, 2013. http: / / csegroups.case.edu / bearingdatacenter / home.

[0158]

[21] Van Der Maaten L, Hinton G.Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(11): 2579-2625.

Claims

1. A rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration, characterized in that: The following steps are involved: Step 1: Obtain vibration signals of all states of mechanical component A as source domain data; obtain a small amount of vibration signals with state labels of other mechanical components different from mechanical component A as target domain data; Step 2: preprocessing the source domain and target domain data to obtain a source domain image dataset and a target domain image dataset; the preprocessing includes performing short-time Fourier transform on the source domain and target domain data to construct a multi-state two-dimensional source domain image dataset and a target domain image dataset respectively; Step 3: Input the source domain image dataset into the improved source domain VGG-16 deep convolutional network for pre-training, obtain the pre-training model, and save the pre-training model network parameters; wherein, the improvements of the improved VGG-16 deep convolutional network include: using the SAdam gradient optimization algorithm to update the network parameters to accelerate the convergence of the VGG-16 deep convolutional network; the network parameter update process is expressed as: Where D represents bounded space; x t represents a decision selected from D; represents the deformation of the squared second-order momentum of the gradient after adding the vanishing factor; α represents the learning rate; t represents the time; represents the first-order momentum; Step 4: Migrate the pre-trained model network parameters to the target domain VGG-16 deep convolutional network to initialize the model network parameters, input the target domain image dataset into the target domain VGG-16 deep convolutional network and introduce the self-supervised SimCLR framework for training, and adaptively update the network parameters of the target domain VGG-16 deep convolutional network through the SimCLR loss function to obtain the fault state classification model; Step 5: Input the vibration signal of the unknown state of the mechanical component to be diagnosed into the fault state classification model after the preprocessing in the above step 2 to obtain the state recognition result of the mechanical component to be diagnosed.

2. The rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration according to claim 1 is characterized in that: The states in step 1 include a normal state and a fault state, and the fault state includes an inner ring fault, a rolling element fault, an outer ring fault, and states of different damage degrees.

3. The rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration according to claim 1 is characterized in that: In step 4, a model migration parameter fine-tuning strategy is adopted to reduce the number of training parameters. Specifically, when migrating the pre-trained model network parameters to the target domain VGG-16 deep convolutional network, the fully connected network parameters in the pre-trained model network parameters are not migrated. Instead, new fully connected network parameters are added to the target domain VGG-16 deep convolutional network for classification; during the training process of the target domain VGG-16 deep convolutional network, the convolutional layers are gradually unfrozen and fine-tuned using a small amount of target domain image data sets carrying state labels.

4. The rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration according to claim 3 is characterized in that: In step 4, the process of inputting the target domain image dataset into the target domain VGG-16 deep convolutional network and introducing the self-supervised SimCLR framework for training includes: first, all input images are randomly transformed for data enhancement to generate two related enhanced images from the same image; after data enhancement, the basic encoder network, i.e., the target domain VGG-16 deep convolutional network, is connected to extract the representative feature vector from the two enhanced data images; then, the representative feature vector is mapped to the vector space where the contrast loss is applied, and the projection vector is obtained using a fully connected linear neural network with one hidden layer; finally, the overall loss is calculated by calculating the cosine similarity between the two enhanced images projected into the vector space. The contrast loss function used by SimCLR is the normalized temperature-scale cross entropy loss.

5. The rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration according to claim 4 is characterized in that: In the process of introducing the self-supervised SimCLR framework for training, the ReLU activation function is replaced by the Swish activation function when mapping the feature vector to the vector space where the contrastive loss is applied.

6. The rolling bearing and gear fault diagnosis method based on semi-supervised model comparison and migration according to claim 5, characterized in that: The loss function of the SimCLR is the average loss of all different image pairs, and the calculation formula is: Where N is the batch size and l is the loss of all distinct image pairs.

Citation Information

Cited By

  • Fault diagnosis method and system for gearbox of transmission reducer of ship engine

    CN122365041A