A rolling bearing fault identification method based on improved relational network
By improving the relational network method, combining meta-learning and residual shrinkage modules, the identification problem of rolling bearing fault classification model under a small number of labeled samples and multiple samples is solved, and the accurate fault identification of different models and load conditions is achieved, and the classification is improved Accuracy and model generalization ability.
Patent Information
- Application Number
- CN202110975013.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-24
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-08-24
AI Technical Summary
Existing rolling bearing fault classification models do not perform well under small numbers of labeled samples, especially under different models of bearings and different load conditions, making it difficult to achieve accurate fault identification.
Using an improved relational network method, combined with meta-learning training strategy, by introducing residual shrinkage module and SELU activation function, we build embedded modules and relationship modules, perform small sample meta-learning, extract features and classify them, which is suitable for different models and load conditions. fault identification.
It realizes effective fault classification of different models of rolling bearings under a small number of marked samples, which improves the generalization ability and classification accuracy of the model, and is suitable for diagnosis under multiple samples and multi-load conditions.
Smart Images

Figure CN113723489B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of bearing fault classification, and in particular to a rolling bearing fault identification method based on an improved relationship network. Background Art
[0002] Rolling bearings are important components of rotating machinery. Once a failure occurs, it will directly affect the performance of the mechanical equipment and even cause life-threatening accidents. Therefore, it is of great significance to accurately diagnose the health status of bearings. [1] Rolling bearings usually work under different loads, and the bearing models in different mechanical equipment may be different, which increases the difficulty of bearing fault diagnosis. [2] In practical engineering, the scarcity of available bearing data with sufficient health marker information makes bearing fault diagnosis increasingly challenging. [3] .
[0003] Fault diagnosis plays an important role in the health management of mechanical equipment. Traditional fault diagnosis methods mainly rely on the rich experience and expertise of engineers. However, in engineering, people expect the fault diagnosis process to be intelligent enough to automatically detect and identify the health status of the machine. [4] Applying machine learning theory to the mechanical field can realize intelligent fault diagnosis. Traditional machine learning theories mainly include artificial neural networks, support vector machines, etc. [5] First, some commonly used features are extracted from the collected vibration data, such as time domain features, frequency domain features, and time-frequency domain features, and then the health status is identified using traditional machine learning theory.
[0004] Although traditional machine learning theory can intelligently identify the health status of a machine, the feature extraction stage still relies mainly on manual work. In addition, the generalization performance of traditional machine learning theory is low, which reduces the diagnostic accuracy. Deep learning is a new topic in the field of machine learning and has received widespread attention. Deep learning can automatically learn fault features from collected data and can provide an end-to-end diagnostic model. [6]. Reference [7] uses EMD to process the signal to obtain a multi-channel one-dimensional signal, and constructs a multi-channel one-dimensional convolutional neural network model to extract features from the multi-channel one-dimensional signal. A stacked denoising autoencoder layer is connected after the fully connected layer to further reduce the dimension and extract features and realize feature classification. Reference [8] proposes a rolling bearing state recognition method based on ensemble empirical mode decomposition-Hilbert envelope spectrum and deep belief network. Experimental results show that this method can well realize the multi-state recognition of rolling bearings under variable load. Reference [9] proposes an adaptive one-dimensional convolutional neural network fault diagnosis algorithm based on "end-to-end", adaptively hierarchically extracts features from the original vibration signal, and uses the Softmax classifier to output the diagnosis results. Reference
[10] proposes two multi-wavelet coefficient fusion methods for a ResNet-based diagnosis model, which helps to learn features that are easier to identify from the input data. Reference
[11] proposes a data-driven planetary gearbox fault diagnosis model that combines time-frequency analysis and ResNet. Although deep learning has achieved remarkable success in machine fault diagnosis, the actual operating environment of rolling bearings is complex and changeable, and there are distribution differences between data obtained under different loads, which will cause the diagnostic model to have poor generalization ability. Transfer learning is an emerging method that uses existing knowledge in the source domain to solve problems in the target domain, which can reduce data differences between domains. In recent years, transfer learning has been increasingly used in rotating machinery fault diagnosis.
[12] . Reference
[13] proposed a migration principal component analysis method, which reduces the distribution difference between training data and test data by mapping cross-domain data to a shared latent space, and uses the features composed of common latent principal components to train the model. Experimental results show that this method can effectively diagnose asynchronous motor faults under various working conditions. Although the existing rolling bearing intelligent diagnosis technology is becoming more and more mature, these methods are mostly applied to bearings of the same model, and few studies focus on fault diagnosis between bearings of different models. There are many types of rolling bearings, and the fault diagnosis model applicable to a certain type of bearing may not be applicable to another type of bearing. Reference
[14] proposed a fault diagnosis method for deep feature migration of rolling bearings of different models. This method constructs a domain-shared improved AlexNet deep convolutional network, introduces a conditional adversarial mechanism, improves the optimization method of the joint distribution of features and labels to a random linear combination, extracts deep features, and realizes the simultaneous adaptation of source domain and target domain features and labels to achieve the purpose of migration.
[0005] In engineering practice, the available data of mechanical equipment is scarce, making it difficult to train and obtain an intelligent diagnosis model with high accuracy in identifying the health status of equipment. [3]In this case, small sample learning came into being. The specific task of small sample learning is that for categories that have not appeared in the training process, even if only a very small number of labeled samples are given, the model can correctly identify them. This method has great application value in the industrial field. The current mainstream solutions for small sample learning tasks include transfer learning, metric learning, and meta-learning.
[15] . Transfer learning uses the knowledge learned in a certain field to solve problems in different but related fields. This technology can solve the problem of scarce available training samples. Reference
[16] extracts shared learnable filter kernels from multiple auxiliary tasks and adaptively transfers them to the target task. Then, through multi-classifier integration, it realizes machine fault identification when there is less available fault data. In the metric learning method, the more representative research results include matching networks, prototype networks and relational network models. Reference
[17] maps the vibration signal to the fault feature metric space through the prototype network, and then constructs a hybrid self-attention module to learn the metric prototypes of the wind turbine gearbox under various health states. Finally, the metric classifier is used for pattern recognition to realize the fault diagnosis of the wind turbine gearbox under small sample conditions. The meta-learning-based method is a very popular method in the field of small sample learning in recent years. This method usually designs some iterative optimization strategies so that the model can learn some general knowledge related to model optimization from a large number of tasks. When encountering new categories, the model can use this general knowledge to quickly optimize and iterate its parameters for the new category, thereby adapting to the new category.
[18] .
[0006] In actual engineering, different mechanical equipment may use different rolling bearing models, and the working loads often change, which will lead to large distribution differences between bearing vibration signals. If only the data of bearings of the same model are used to train the fault classification model, the diagnostic effect is often not ideal. Summary of the invention
[0007] In view of the above problems, the present invention proposes a rolling bearing fault identification method based on an improved relational network to solve the problem that the existing rolling bearing fault classification model has poor fault classification and identification effect in the case of a small number of labeled samples.
[0008] A rolling bearing fault identification method based on an improved relationship network comprises the following steps:
[0009] Step 1: Obtain training data and test data of the original vibration signal of the rolling bearing;
[0010] Step 2: preprocess the training data and the test data to obtain a source domain dataset and a target domain dataset respectively;
[0011] Step 3: Input the source domain data set into the meta-learning training model based on the improved relational network to obtain a trained rolling bearing fault classification model; wherein the improved relational network includes an improved embedding module and a relational module, and the improved embedding module and the relational module perform small sample meta-learning in an end-to-end manner;
[0012] Step 4: Input the target domain data set into the trained rolling bearing fault classification model to obtain the rolling bearing fault identification result.
[0013] Furthermore, the original vibration signal of the rolling bearing in step one is collected under 4 load conditions, and the load types include 0hp, 1hp, 2hp and 3hp; the training data and test data include the original vibration signals of the rolling bearing in 10 states, including normal state, inner ring, outer ring, rolling element failure and different degrees of damage.
[0014] Furthermore, the preprocessing in step 2 includes performing Fourier transform on the training data and the test data to obtain a frequency domain amplitude sequence.
[0015] Furthermore, the improved embedding module described in step three is an improvement on the original embedding module, which is composed of four convolution blocks. The improved embedding module replaces the second and third convolution blocks with residual shrinkage modules, and replaces the fourth convolution block with a residual module with a SELU activation function.
[0016] Furthermore, the specific steps of step three include:
[0017] Step 31: Randomly select C categories from the source domain data set, select K samples from each category as the source domain support set, and select N samples from the remaining samples of each category as the source domain query set. The source domain support set samples and the source domain query set samples together constitute a C-way K-shot task;
[0018] Step 32: Input the source domain support set samples and the source domain query set samples into the improved embedding module, and concatenate the output source domain support set features and source domain query set features to form feature pairs;
[0019] Step 33: Input the feature pair into the relationship module to calculate the relationship score, and determine which category the source domain query set sample belongs to by the relationship score; wherein the relationship score is used to represent the similarity between the samples in the source domain support set and the samples in the source domain query set;
[0020] Step 34: Iterate and loop through steps 31 to 33, optimize network parameters using mean square error loss, and stop executing when the maximum number of iterations is reached to obtain a rolling bearing fault classification model.
[0021] Furthermore, in step 3, the residual shrinkage module in the improved embedding module adds the soft threshold as a nonlinear transformation layer to the residual module, and introduces an attention mechanism; specifically:
[0022] First, the absolute value of the feature is input into the global average pooling layer to obtain a one-dimensional vector;
[0023] Then, the one-dimensional vector is input into two fully connected layers, and the number of neurons in the second fully connected layer is equal to the number of channels of the input feature;
[0024] Then, the normalized weight of each feature channel is obtained by the Sigmoid activation layer;
[0025] Finally, the Scale operation is used to add the normalized weights to each channel of the feature to obtain the threshold of each channel of the feature.
[0026] Furthermore, the threshold calculation formula for each channel of feature x is:
[0027]
[0028] Among them, τ c represents the threshold of the cth channel of the feature; α c represents the cth attention weight; i, j, c are the width, height, and channel index of feature x, respectively.
[0029] Furthermore, the specific steps of step 4 include:
[0030] Step 41: Randomly select C categories from the target domain data set, select K samples from each category as the target domain support set, and select N samples from the remaining samples of each category as the target domain query set. The target domain support set samples and the target domain query set samples together constitute a C-way K-shot task.
[0031] Step 42: Input the target domain support set samples and the target domain query set samples into the trained rolling bearing fault classification model, use the improved embedding module in the rolling bearing fault classification model to extract the target domain support set and target domain query set sample features, and concatenate the output target domain support set features and target domain query set features and input them into the relationship module in the rolling bearing fault classification model to calculate the relationship score, classify the target domain query set samples according to the relationship score, and obtain the rolling bearing fault classification result.
[0032] The beneficial technical effects of the present invention are:
[0033] The present invention combines a meta-learning training strategy with a relational network model, and can train a fault classification model using only a small amount of labeled samples, thereby solving the problem of insufficient rolling bearing fault data in practice; wherein, the relational network is improved: a residual shrinkage module and a SELU activation function are introduced into an embedding module of the relational network, so that the network is more suitable for processing bearing vibration signals and can extract more useful features from the data; the improved relational network can well classify rolling bearing faults of different models under different load conditions, thereby solving the problem that the various bearing models in different mechanical equipment result in poor generalization ability of a model trained on bearing data of a certain model and cannot be well applied to bearings of other models. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] 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.
[0035] Figure 1 It is a schematic diagram of the small sample classification task in the present invention;
[0036] Figure 2 It is a schematic diagram of the relationship network model in the present invention;
[0037] Figure 3 is a graph of the soft threshold function in the present invention;
[0038] Figure 4 is a derivative graph of the soft threshold function in the present invention;
[0039] Figure 5 Schematic diagram of the residual module structure in the present invention;
[0040] Figure 6 It is a schematic diagram of the structure of the residual shrinkage module in the present invention;
[0041] Figure 7 It is the SELU activation function diagram in the present invention;
[0042] Figure 8 1 is a structural diagram of the relational network embedding module before and after the improvement in the present invention; wherein, FIG. (a) represents the relational network embedding module before the improvement; FIG. (b) represents the relational network embedding module after the improvement;
[0043] Fig. 9 is a flow chart of the method of the present invention;
[0044] Fig.10 is a schematic diagram of a test bench in an embodiment of the present invention;
[0045] Fig.111 is a comparison chart of classification accuracy of three groups of experiments in an embodiment of the present invention;
[0046] Fig.12 1 is a feature visualization effect diagram in an embodiment of the present invention; wherein, FIG (a) represents a 5-way 5-shot task in scheme 1; FIG (b) represents a 5-way 5-shot task in scheme 3; FIG (c) represents a 5-way 10-shot task in scheme 1; FIG (d) represents a 5-way 10-shot task in scheme 3;
[0047] Fig.13 3 is a comparison chart of classification accuracy before and after the relationship network is improved in an embodiment of the present invention. DETAILED DESCRIPTION
[0048] 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.
[0049] The present invention proposes a rolling bearing fault identification method based on an improved relational network, which is used to classify the faults of rolling bearings of different models under different loads with only a small number of samples. The method first divides the data set according to the meta-learning training strategy, and introduces the residual shrinkage module and the SELU activation function into the embedding module of the relational network. The residual shrinkage module automatically determines the threshold using the attention mechanism to remove the redundant information in the signal; the embedding module is used to extract sample features, and the support set sample features and the query set features are spliced and input into the relational module; finally, the query set samples are classified according to the relational score to achieve the classification of rolling bearing faults. The various steps of the present invention are described in detail below.
[0050] 1. Meta-learning training strategy
[0051] Meta-learning is also called learning to learn.
[19] , is a cutting-edge research framework in the field of machine learning, which aims to solve the problem of how models learn. The purpose of meta-learning is to enable the model to acquire a learning ability that allows the model to automatically learn some meta-knowledge. Meta-knowledge refers to knowledge that can be learned outside the model training process, such as the model's hyperparameters, the initial parameters of the neural network, the structure of the neural network, and the optimizer.
[20] .
[0052] Meta-learning currently focuses on small sample learning problems. In small sample learning, meta-learning specifically refers to learning meta-knowledge from a large number of prior tasks, using previous prior knowledge to guide the model to learn faster in new tasks.
[21] . The data sets in meta-learning are usually divided into meta-training sets and meta-test sets, and the categories of data in the meta-training set and meta-test set are different. Both the meta-training and meta-testing stages use small sample tasks as the basic units, and each task has its own training data set and test data set, also called support set and query set. Meta-learning uses a large number of small sample tasks for training in the training phase, and regards the entire task set as training examples. Each task-based training will sample different meta-tasks, and then train different category combinations. This mechanism enables the model to learn their common parts from different meta-tasks. The model learned through this learning mechanism has strong generalization ability. When facing unseen meta-tasks in the meta-testing phase, it does not need to adjust the existing model and can still perform classification well. Small sample classification tasks such as Figure 1 Shown
[22] .
[0053] Meta-learning research is currently divided into four relatively independent directions: metric-based methods, initialization-based methods, optimizer-based methods, and external storage-based methods.
[23] The metric-based approach tends to extract the features contained in the task samples to the greatest extent possible, and uses feature comparison to determine the type of the samples. Therefore, how to extract the features that best represent the characteristics of the samples has become the research focus of this approach. The present invention adopts a metric-based meta-learning method, in which metric learning uses a relational network.
[0054] 2. Relationship network and improvement
[0055] Relation Network (RN) is an architecture for small sample learning proposed by Flood Sung et al. in 2018.
[24] , which has the advantages of simple concept, flexibility and versatility. The relation network consists of two parts: the embedding module and the relation module. First, the embedding module extracts the features of the input sample, and then the relation module compares these features to determine whether they belong to the same category. The embedding module and the relation module perform small sample meta-learning in an end-to-end manner. The relation network structure is as follows Figure 2 As shown in the figure, the embedded module is Indicates that the relation module is represented by g k express.
[0056] Use the meta-learning training strategy described above for small sample learning. In each training iteration, randomly select C categories from the dataset and select K samples from each category as the support set of the model. The remaining samples in C classes are used as the query set of the model The sample x in the support set i and the sample x in the query set j Input to the embedding module to get the features and The obtained features are cascaded along the channel through the splicing module C(·,·), and we get The combined features of the support sample and the query sample are input into the relation module g k In the output, a scalar in the (0,1) interval is called the relation score r i,j , used to represent x i and x j In the case of a single sample, the relationship score r i,j Specifically:
[0057]
[0058] For the small sample case, the features of the support set samples of each class are element-wise added to form the features of this class, and then combined with the query set sample features mentioned above. Therefore, in the single-shot or small-shot setting, the number of relationship scores obtained for each query sample is always C.
[0059] The mean square error loss function is used to train the model in the relational network. Assuming that when the support sample and the query sample belong to the same category, their relation score is 1, and when the support and query samples do not belong to the same category, the relation score is 0. In this case, the classification problem can be regarded as a regression problem. The objective function is:
[0060]
[0061] The residual shrinkage module introduces soft threshold and attention mechanism on the basis of the residual module.
[25] ,Through the attention mechanism to automatically determine the threshold, more discriminative features of the vibration signal can be ,extracted.
[0062] 1) Soft Threshold
[0063] The soft threshold method can retain useful information features in the vibration signal and is used for feature activation of the network model. The function expression of the soft threshold is:
[0064]
[0065] Among them, x is the input feature, y is the output feature, and τ is the threshold. The soft threshold sets the features close to zero to zero, retaining the useful negative features, unlike the ReLU activation function, which directly sets the negative features to zero.
[0066] The soft thresholding process is as follows Figure 3As shown. It can be seen that the derivative of the output to the input is 1 or 0, which can effectively prevent the gradient disappearance and gradient explosion problems, such as Figure 4 As shown. The soft threshold derivative can be expressed as:
[0067]
[0068] Setting an appropriate threshold often requires a lot of signal processing expertise, but the deep architecture in the residual shrinkage module can be used to automatically determine the threshold.
[0069] 2) Residual shrinkage module
[0070] The residual shrinkage module adds the soft threshold as a nonlinear transformation layer to the residual module, and introduces an attention mechanism to estimate the threshold. First, the absolute value of the feature x is input into the global average pooling layer to obtain a one-dimensional vector. Then the one-dimensional vector is input into two fully connected layers, and the number of neurons in the second fully connected layer is equal to the number of channels of the input feature. After that, the normalized weight of each feature channel is obtained by the Sigmoid activation layer, and the final Scale operation adds the normalized weight to each channel of the feature to achieve the improvement of important channels. It can be expressed as:
[0071]
[0072] Among them, z c is the characteristic of the cth neuron, α c is the cth attention weight. Then, the threshold is calculated as follows:
[0073]
[0074] Among them, τ c is the threshold of the cth channel of the feature, i, j, c are the width, height, and channel index of feature x respectively. The threshold should be a positive number within a reasonable range to prevent the output feature from being all zero. The residual module and residual shrinkage module structures are as follows: Figure 5 and Figure 6 shown.
[0075] In 2017, the paper
[26] proposed an activation function called scaled exponential linear unit (SELU). SELU not only has the advantages of ReLU and can overcome the gradient vanishing and gradient exploding problems, but also, unlike the sparsity of ReLU, SELU retains the calculation of the input part less than 0, which can provide richer features. The function expression is as follows:
[0076]
[0077] Among them, λ≈1.05 and α≈1.67.
[0078] SELU sets a λ slightly larger than 1 to keep the input greater than 0 non-saturated and retains the calculation of the input less than 0. It sets a constant α to ensure that the neuron output distribution can be automatically normalized to zero mean and unit variance, thereby extracting more high-dimensional features. SELU has the dual characteristics of nonlinear mapping activation and self-normalization. Therefore, SELU has certain advantages in theory. Figure 7 A diagram of the SELU activation function is shown.
[0079] The embedding module of the relational network consists of four convolution blocks. The improved relational network replaces the second and third convolution blocks of the embedding module with residual shrinkage modules, and replaces the fourth convolution block with a residual module with a SELU activation function. The improved embedding module only adds a small number of hidden layers, which does not increase the complexity of the network. At the same time, due to the introduction of the attention mechanism and SELU activation function, the relational network can focus on more useful features in the vibration signal. The residual shrinkage module itself is used to process vibration signals, so it is introduced into the relational network, which is suitable for rolling bearing fault classification. The specific structure of the relational network before and after improvement is as follows: Figure 8 As shown in (a) and (b).
[0080] 3. Rolling bearing fault identification method
[0081] Using the improved relational network, the inner ring, outer ring, and rolling element faults and normal states of different types of rolling bearings are classified. The diagnosis process includes data preprocessing, meta-training, and meta-testing. The specific process is as follows: Fig. 9 shown.
[0082] (1) Data preprocessing
[0083] Fourier transform is performed on the original vibration signals of rolling bearings of different models to obtain frequency domain amplitude sequences, which constitute the source domain dataset and the target domain dataset respectively for meta-training and meta-testing stages. C categories are randomly selected from the source domain, K samples are selected from each category as the support set, and N samples are selected from the remaining samples of each category as the query set. The support set and query set samples together constitute a C-way K-shot task, and different tasks are selected for each iterative training. The target domain constructs tasks in the same way for testing.
[0084] (2) Meta-training stage
[0085] The support set samples and query set samples of the source domain are input into the embedding module of the improved relational network. The improved embedding module automatically determines the threshold through the attention mechanism and introduces the SELU activation function to pay attention to more details and extract more useful features of the samples. The obtained support set features and query set features are concatenated to form feature pairs. The feature pairs are input into the relational module to calculate the relation scores. The relation scores are used to determine which category the query set samples belong to. The mean square error loss is used to optimize the parameters of the network and obtain the rolling bearing fault classification model.
[0086] (3) Meta-testing phase
[0087] The support set and query set samples of the target domain are input into the trained classification model, and the features of the support set and query set samples are extracted using the embedding module of the model. The features are concatenated and input into the relationship module, and the target domain query set samples are classified according to the relationship scores. At this time, there is no need to update the network, and the classification results of bearing faults of different models can be obtained directly.
[0088] 4. Application and analysis
[0089] The experimental data of the embodiment of the present invention comes from the public bearing data set of Case Western Reserve University in the United States
[27] . Schematic diagram of the bearing test bench is as follows Fig.10 As shown. The drive end bearing model is SKF6205, and the fan end bearing model is SKF6203. The bearing vibration signal data is collected by a 16-channel acceleration sensor with a sampling frequency of 12kHz. The bearing vibration signal is collected under four different loads: 0, 1, 2, and 3hp (hp is British horsepower, 1hp = 0.75kw).
[0090] The embodiment of the present invention uses two types of rolling bearing data, SKF6205 and SKF6203. Each type of bearing data contains 10 states, including normal state, inner ring, outer ring, rolling element fault and different damage degrees. IR(7,14,21) indicates that the inner ring damage diameter is 7mils, 14mils and 21mils respectively. Similarly, OR(7,14,21) and B(7,14,21) indicate different fault degrees of the outer ring and rolling element respectively. The vibration signal in the data set is truncated into samples with a length of 2048 points, and each sample is fast Fourier transformed to reduce the sample length to half of the original. The two types of rolling bearings have 10 states, each with 16 samples, which are used to solve the small sample bearing classification problem. The experimental data set is shown in Table 1. Each experiment divides the data according to the meta-learning training strategy, selects 5 categories in each task (C=5), and the number of support set samples includes 1, 3, 5, and 10 samples (K=1, 3, 5, 10). The Adam optimization algorithm was used to train the network end-to-end, the learning rate was set to 0.001, the maximum number of episode iterations was set to 50, and each experiment was repeated 5 times.
[0091] Table 1 Experimental dataset
[0092]
[0093] Combining the relational network model with the meta-learning training strategy, fault diagnosis is performed on rolling bearings of different models under different load conditions. The two types of bearing data used in the embodiment have 10 states respectively, and there are four load conditions during operation. The change of load may increase the difficulty of rolling bearing fault classification. Three groups of experiments are set up to classify the faults of bearings of different models under different load conditions. For example: 6205_1hp / 6203_023hp means that the classification model is trained using the SKF6205 bearing data collected when the workload is 1hp, and the SKF6203 bearing data collected under 0, 2, and 3hp loads are classified. The experimental results are shown in Table 2.
[0094] Table 2 Fault classification results of rolling bearings of different types under different loads
[0095]
[0096] Fig.11 It represents the classification accuracy of three groups of experiments under the 5-way 1-shot task. Fig.11The first and third experimental results show that more load conditions in the target domain will increase the difficulty of fault classification. The second experimental results show that the increase in load conditions in the source domain does not improve the classification effect of the model. As can be seen from Table 2, under the other three sample settings, load changes will also have the same impact on the classification results. When the target domain data has only a single load, the classification effect is better. The experimental results show that the mixing of data under multiple loads will cause certain interference to the classification of bearing faults of different models, but the relationship network can solve this problem better.
[0097] Further comparative experiments were conducted on the classification effects before and after the improvement of the relational network. Scheme 1 used the original relational network, Scheme 2 only introduced the residual shrinkage module into the embedding module of the relational network, and Scheme 3 was the improved relational network of the present invention, that is, the residual shrinkage module and the SELU activation function were simultaneously introduced into the embedding module. Multiple groups of experiments were set up under different loads, and the experimental results are shown in Table 3.
[0098] Table 3 Fault classification results of rolling bearings of different models before and after the relationship network improvement
[0099]
[0100] From the experimental results, it can be seen that under different loads, the classification effect of bearing fault classification using the relational network that only introduces the residual shrinkage module is improved. On this basis, the SELU activation function is further introduced, and the classification accuracy is increased by up to 3.44% compared with the original relational network. In the case of a single sample, the classification accuracy can also be above 86%, proving that the improved relational network can more accurately classify the faults of rolling bearings of different models.
[0101] In order to more intuitively reflect the classification effect of the improved relationship network, in the 6205_0123hp / 6203_0123hp experiment, the t-SNE algorithm is used
[28] The query set sample features in the 5-way 5-shot and 5-way 10-shot tasks are visualized, and a bar chart is drawn according to the rolling bearing fault classification results, as shown in the figure. Fig.12 and Fig.13 shown.
[0102] Depend on Fig.12 It can be seen that the multi-state features of rolling bearings extracted by the improved relational network (Scheme 1) overlap and are difficult to distinguish, which will have a negative impact on the classification effect. The improved relational network (Scheme 3) introduces the attention mechanism and the nonlinear SELU activation function, which can focus on the more useful parts of the data. The extracted features have smaller intra-class spacing and larger inter-class spacing, making it easier to distinguish different states of bearings. The classification effect on most tasks is relatively ideal. In addition, Fig.13It can be seen that when the support set samples in the task increase from 1-shot to 10-shot, the fault diagnosis accuracy of the model is improved.
[0103] In order to further prove the effectiveness of the method of the present invention, the laboratory data set of bearing model 6307E is supplemented as experimental data to perform fault diagnosis on bearings of different models. The bearing data set of model 6307E has a rotation speed of 680r / s and a sampling frequency of 8192Hz. It contains three types of state data: inner ring fault, outer ring fault and normal state, and each state has 16 samples. The fault classification results of rolling bearings of different models are shown in Table 4.
[0104] Table 4 Fault classification results of rolling bearings of different types
[0105]
[0106] It can be seen from Table 4 that the accuracy of classifying rolling bearing faults of different types can reach more than 90%, which proves that the method of the present invention can better realize bearing fault diagnosis under small samples.
[0107] The documents cited in the present invention are as follows:
[0108] [1] Yang B, Lei Y, Jia F, et al. An intelligent fault diagnosis approach based on transfer learning from laboratory bearings to locomotive bearings[J]. Mechanical Systems and Signal Processing, 2019, 122: 692-706.
[0109] [2]Chen C, Shen F, Xu J, et al. Model Parameter Transfer for Gear FaultDiagnosis under Varying Working Conditions[J]. Chinese Journal of MechanicalEngineering, 2021.
[0110] [3] Lei Yaguo, Yang Bin, Du Zhaojun, et al. Deep migration diagnosis method for mechanical equipment faults under big data[J]. Journal of Mechanical Engineering, 2019, 55(07): 1-8.
[0111] [4]Hoang DT,Kang H JA survey on Deep Learning based bearing faultdiagnosis[J].Neurocomputing,2019,335:327-335.
[0112] [5]Liu R, Yang B, Zio E, et al.Artificial intelligence for faultdiagnosis of rotating machinery:A review[J].Mechanical Systems&SignalProcessing,2018,108:33-47.
[0113] [6] Lei Y, Yang B, Jiang X, et al.Applications of machine learning tomachine fault diagnosis: A review and roadmap[J]. Mechanical Systems and SignalProcessing, 2020,138:106587-.
[0114] [7] Ye Zhuang, Yu Jianbo. Gearbox fault diagnosis method based on multi-channel one-dimensional convolutional neural network feature learning[J]. Vibration and Shock, 2020, 39(20): 55-66.
[0115] [8] Wang Yujing, Na Xiaodong, Kang Shouqiang, et al. State identification method of rolling bearing under variable load based on EEMD-Hilbert envelope spectrum and DBN[J]. Proceedings of the CSEE, 2017, 37(23): 6943-6950+7085.
[0116] [9] Qu Jianling, Yu Lu, Yuan Tao, et al. Adaptive fault diagnosis algorithm for rolling bearings based on one-dimensional convolutional neural network[J]. Chinese Journal of Scientific Instrument, 2018, 39(07): 134-143.
[0117]
[10] Zhao M, Kang M, Tang B, et al.Multiple Wavelet Coefficients Fusionin Deep Residual Networks for Fault Diagnosis[J]. IEEE Transactions on Industrial Electronics, 2019, 66(6): 4696-4706.
[0118]
[11] Ma S, F Chu, Han Q. Deep residual learning with demodulated time-frequency features for fault diagnosis of planetary gearbox undernonstationary running conditions [J]. Mechanical Systems and Signal Processing, 2019, 127: 190-201.
[0119]
[12] Yan R, Shen F, Sun C, et al. Knowledge Transfer for Rotary MachineFault Diagnosis[J]. IEEE Sensors Journal, 2020, 20(15):8374-8393.
[0120]
[13] Yan R, F Shen, Zhou M. Induction Motor Fault Diagnosis Based on Transfer Principal Component Analysis[J]. Chinese Journal of Electronics, 2021, 30(1).
[0121]
[14] Wang Yujing, Lü Haiyan, Kang Shouqiang, et al. Fault diagnosis methods for rolling bearings of different types[J]. Proceedings of the CSEE, 2021, 41(01): 267-276+416.
[0122]
[15] Jin Lu, Liu Shijian, Wang Xiao, et al. Small sample infrared aerial target classification method based on improved relational network[J]. Acta Optica Sinica, 2020, 40(08): 87-96.
[0123]
[16] Xing S, Lei Y, Yang B, et al. Adaptive Knowledge Transfer by Continual Weighted Updating of Filter Kernels for Few-shot Fault Diagnosis of Machines[J]. IEEE Transactions on Industrial Electronics, 2021, (99): 1-1.
[0124]
[17] Yu Haoshuai, Tang Baoping, Zhang Kai, et al. Wind turbine gearbox fault diagnosis method based on hybrid self-attention prototype network under small sample size [J / OL]. China Mechanical Engineering: 1-8 [2021-06-02]. http: / / kns.cnki.net / kcms / detail / 42.1294.th.20210330.1320.006.html.
[0125]
[18] Wu Ziyang. Research on small sample learning algorithm based on metric learning[D]. South China University of Technology, 2020.
[0126]
[19] Thrun S, Pratt L. Learning to learn: Introduction and overview[M]. Springer US, 1998.3-17.
[0127]
[20] Vilalta R,Drissi YA Perspective View and Survey of Meta-Learning[J].Artificial Intelligence Review, 2002,18(2):77-95.
[0128]
[21] Zhao Kailin, Jin Xiaolong, Wang Yuanzhuo. A review of small sample learning research[J]. Journal of Software, 2021, 32(02): 349-369.
[0129]
[22] Ravi S,Larochelle H.Optimization as a model for few-shot learning[C].International Conference on Learning Representations,Toulon,France,2017:1-3.
[0130]
[23] Li Fanchang, Liu Yang, Wu Pengxiang, et al. A review of meta-learning research[J]. Chinese Journal of Computers, 2021, 44(02): 422-446.
[0131]
[24] Sung F, Yang Y, Li Z, et al. Learning to Compare:Relation Network for Few-Shot Learning[C]. 2018 IEEE / CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2018.
[0132]
[25] Zhao M, Zhong S, Fu X, et al. Deep Residual Shrinkage Networks for Fault Diagnosis[J]. IEEE Transactions on Industrial Informatics, 2020, (99): 1-1.
[0133]
[26] klambauer G, Unterthiner T, Mayr A, et al. Self-normalizing neural networks [C]. 31st Conference on Neural Information Processing Systems. Long Beach, CA: NIPS Foundation. 2017: 971-980.
[0134]
[27] Case Western Reserve University Bearing Data Center, http: / / csegroups.case.edu / bearingdatacenter / home .
[0135]
[28] Der Maaten LV, Hinton G E.Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, (9): 2579-2605.
[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.
Claims
1. A rolling bearing fault identification method based on improved relational network, It is characterized in that The steps include: Step 1, obtaining training data and test data of the original vibration signal of the rolling bearing; the original vibration signal of the rolling bearing is collected under 4 load conditions, and the load types include 0hp, 1hp, 2hp and 3hp; the training data and test data include the original vibration signal of the rolling bearing in 10 states, including normal state, inner ring, outer ring, rolling element failure and different damage degrees; Step 2: preprocess the training data and the test data to obtain a source domain dataset and a target domain dataset respectively; Step 3: Input the source domain data set into the meta-learning training model based on the improved relational network to obtain a trained rolling bearing fault classification model; wherein the improved relational network includes an improved embedding module and a relational module, and the improved embedding module and the relational module perform small sample meta-learning in an end-to-end manner; the improved embedding module is an improvement on the original embedding module, which is composed of four convolution blocks. The improved embedding module replaces the second and third convolution blocks with residual shrinkage modules, and replaces the fourth convolution block with a residual module with a SELU activation function; the residual shrinkage module adds a soft threshold as a nonlinear transformation layer to the residual module, and introduces an attention mechanism; specifically: input the absolute value of the feature into the global average pooling layer to obtain a one-dimensional vector; input the one-dimensional vector into two fully connected layers, and the number of neurons in the second fully connected layer is equal to the number of channels of the input feature; obtain the normalized weight of each feature channel by the Sigmoid activation layer; use the Scale operation to weight the normalized weight to each channel of the feature to obtain the threshold of each channel of the feature; the threshold calculation formula for each channel of the feature x is: Among them, τ c represents the threshold of the cth channel of the feature; α c represents the cth attention weight; i, j, c are the width, height, and channel index of feature x respectively; Step 4: Input the target domain data set into the trained rolling bearing fault classification model to obtain the rolling bearing fault identification result.
2. According to the rolling bearing fault identification method of the improved relationship network of claim 1, It is characterized in that The preprocessing in step 2 includes performing Fourier transform on the training data and the test data to obtain a frequency domain amplitude sequence.
3. The rolling bearing fault identification method based on an improved relational network according to claim 1, It is characterized in that The specific steps of step three include: Step 31: Randomly select C categories from the source domain data set, select K samples from each category as the source domain support set, and select N samples from the remaining samples of each category as the source domain query set. The source domain support set samples and the source domain query set samples together constitute a C-way K-shot task; Step 32: Input the source domain support set samples and the source domain query set samples into the improved embedding module, and concatenate the output source domain support set features and source domain query set features to form feature pairs; Step 33: Input the feature pair into the relationship module to calculate the relationship score, and determine which category the source domain query set sample belongs to by the relationship score; wherein the relationship score is used to represent the similarity between the samples in the source domain support set and the samples in the source domain query set; Step 34: Iterate and loop through steps 31 to 33, use mean square error loss to optimize network parameters, stop executing when the maximum number of iterations is reached, and obtain a rolling bearing fault classification model.
4. The rolling bearing fault identification method based on an improved relational network according to claim 3, It is characterized in that The specific steps of step 4 include: Step 41: Randomly select C categories from the target domain data set, select K samples from each category as the target domain support set, and select N samples from the remaining samples of each category as the target domain query set. The target domain support set samples and the target domain query set samples together constitute a C-way K-shot task. Step 42: Input the target domain support set samples and the target domain query set samples into the trained rolling bearing fault classification model, use the improved embedding module in the rolling bearing fault classification model to extract the target domain support set and target domain query set sample features, and concatenate the output target domain support set features and target domain query set features and input them into the relationship module in the rolling bearing fault classification model to calculate the relationship score, classify the target domain query set samples according to the relationship score, and obtain the rolling bearing fault classification result.
Citation Information
Patent Citations
Mechanical fault diagnosis method based on migration relation network
CN111709448A
Rotating part fault diagnosis method and device based on improved CNN and relation module
CN113111820A
