CN113723489A

CN113723489A - Rolling bearing fault identification method for improved relation network

Info

Publication number: CN113723489A
Application number: CN202110975013.0A
Authority: CN
Inventors: 梁欣涛; 王玉静; 乔春阳; 康守强; 王庆岩; 兰朝凤
Original assignee: Harbin University of Science and Technology
Current assignee: Harbin University of Science and Technology
Priority date: 2021-08-24
Filing date: 2021-08-24
Publication date: 2021-11-30

Abstract

A rolling bearing fault identification method for improving a relational network relates to the technical field of bearing fault identification and is used for solving the problem that an existing rolling bearing fault identification model is poor in fault identification effect under the condition that a small number of marked samples exist. The technical points of the invention comprise: dividing a data set according to a meta-learning training strategy, introducing a residual shrinkage module and an SELU (self-adaptive selection unit) activation function into an embedding module of a relational network, automatically determining a threshold value by using an attention mechanism through the residual shrinkage module, and removing redundant information in a signal; extracting sample characteristics by using an embedding module, splicing the sample characteristics of the support set and the characteristics of the query set, and inputting the spliced sample characteristics and the characteristics of the query set into a relation module; and finally classifying the query set samples according to the relation scores to realize the fault identification of the rolling bearing. The fault recognition method can train the fault recognition model by using a small number of marked samples, and solves the problems that the model trained on the bearing data of a certain model is poor in generalization capability and cannot be effectively used for fault recognition of bearings of other models.

Description

Rolling bearing fault identification method for improved relation network

Technical Field

The invention relates to the technical field of bearing fault classification, in particular to a rolling bearing fault identification method for an improved relational network.

Background

Rolling bearings are important parts of rotating machines, and in case of failure, they directly affect the performance of the mechanical equipment and even cause life-threatening accidents. Therefore, it is of great importance to accurately diagnose the health of the bearing^[1]. Rolling bearings typically operate under different loads and the bearing types may differ in different mechanical equipment, which can increase the difficulty of bearing fault diagnosis^[2]. In practical engineering, the scarcity of available bearing data with sufficient health marker information has made bearing fault diagnosis increasingly challenging^[3]。

Fault diagnosis plays an important role in the health management of mechanical equipment. The traditional fault diagnosis method mainly depends on abundant experience and professional knowledge of engineers, however, in engineering, people expect that the fault diagnosis process is intelligent enough to automatically detect and identify the health state of a machine^[4]. The intelligent fault diagnosis can be realized by applying the machine learning theory to the mechanical field. The traditional machine learning theory mainly comprises an artificial neural network, a support vector machine and the like^[5]. Firstly, some common features such as time domain features, frequency domain features and time-frequency domain features are extracted from collected vibration data, and then the traditional machine learning theory is used for health state identification.

Although the traditional machine learning theory can intelligently identify the health state of the machine, the feature extraction stage still mainly depends on manual work, and in addition, the generalization performance of the traditional machine learning theory is low, so that the diagnosis accuracy rate is reduced. Deep learning is a new subject in the field of machine learning, and has attracted extensive attention. Deep learning can automatically learn fault characteristics from collected data and can provide an end-to-end diagnostic model^[6]. Document [7]]And processing the signals by using EMD to obtain multi-channel one-dimensional signals, constructing a multi-channel one-dimensional convolution neural network model, and extracting the characteristics of the multi-channel one-dimensional signals. And stacking a noise reduction self-encoder layer behind the full-connection layer, further performing dimension reduction and feature extraction, and realizing feature classification. Document [8]A rolling bearing state identification method based on ensemble empirical mode decomposition-Hilbert envelope spectrum and deep belief network is provided. Experimental results show that the method can well realize multi-state identification of the rolling bearing under variable load. Document [9]]An adaptive one-dimensional convolutional neural network fault diagnosis algorithm based on end-to-end is provided, features are extracted from the original vibration signals in an adaptive hierarchical mode, and a diagnosis result is output by using a Softmax classifier. Document [10 ]]Two multi-wavelet coefficient fusion methods have been proposed for ResNet-based diagnostic models, which help to learn more easily identifiable features from the input data. Document [11]A data driven planetary gearbox fault diagnosis model combining frequency analysis and ResNet is presented. Although the current deep learning achieves remarkable success in the aspect of machine fault diagnosis, the actual operating environment of the rolling bearing is complex and changeable, the data obtained under different loads have distribution difference, and the diagnosis model has the problem of poor generalization capability. Transfer learning is a new method for solving the problem of a target domain by using the existing knowledge of a source domain, and can reduce the data difference between domains. In recent years, transfer learning is increasingly applied to fault diagnosis of rotary machines^[12]. Document [13 ]]A migration principal component analysis method is provided, the distribution difference between training data and test data is reduced by mapping cross-domain data to a shared potential space, and a model is trained by using features composed of common potential principal components. Experimental results show that the method can effectively diagnose the faults of the asynchronous motor under various working conditions. Although the existing rolling bearing intelligent diagnosis technology is mature day by day, the methods are mostly applied to the bearings of the same model, and less research is focused on fault diagnosis among the bearings of different models. Of rolling bearingsThe types are many, and the fault diagnosis model suitable for one type of bearing is not necessarily suitable for another type of bearing. Document [14]A fault diagnosis method for deep characteristic migration of rolling bearings of different models is provided. The method constructs a field-shared improved AlexNet deep convolution network, introduces a conditional countermeasure mechanism, improves an optimization method of feature and label joint distribution into random linear combination, extracts deep features, realizes simultaneous self-adaptation of source domain and target domain features and labels, and achieves the purpose of migration.

Available data of mechanical equipment in engineering practice is scarce, training is difficult to achieve, and an intelligent diagnosis model with high equipment health state recognition precision is obtained^[3]. In this case, small sample learning arises. The specific task of small sample learning is that for classes which do not appear in the training process, even if only a small amount of labeled samples are given, the model can be correctly identified. The method has great application value in the industrial field. The solutions of the small sample learning task of the current mainstream comprise transfer learning, metric learning and meta learning^[15]. The transfer learning solves the problems of different but related fields by applying the knowledge learned in a certain field, and the technology can solve the problem of scarcity of available training samples. Document [16 ]]The shared learnable filtering kernel is extracted from the multiple auxiliary tasks and adaptively migrated to the target task, and then machine fault identification is realized under the condition of less available fault data through multi-classifier integration. In the metric learning method, the comparative representative research results comprise a matching network, a prototype network and a relationship network model. Document [17]]The method comprises the steps of mapping vibration signals to a fault feature measurement space through a prototype network, then constructing a hybrid self-attention module, learning measurement prototypes of the wind power gear box under various health states, and finally performing mode recognition through a measurement classifier to realize fault diagnosis of the wind power gear box under the condition of small samples. The meta learning based method is a very popular method in the field of small sample learning in recent years. This approach is usually to design some strategy of iterative optimization such that the model learns some common knowledge related to model optimization from a large number of tasks, and when a new class is encountered, the modelCan utilize the general knowledge to carry out rapid optimization iteration on the parameters of the new category so as to adapt to the new category^[18]。

In actual engineering, rolling bearing models used by different mechanical equipment may be different, and working loads are changed frequently, which causes large distribution difference among bearing vibration signals, and the diagnosis effect is often not ideal enough only by training a fault classification model with the bearing data of the same model.

Disclosure of Invention

In view of the above problems, the invention provides a rolling bearing fault identification method for improving a relational network, which is used for solving the problem that the existing rolling bearing fault classification model has poor fault classification identification effect under the condition of a small number of marked samples.

A rolling bearing fault identification method for improving a relational network comprises the following steps:

firstly, acquiring original vibration signal training data and test data of a rolling bearing;

preprocessing the training data and the test data to respectively obtain a source domain data set and a target domain data set;

inputting the source domain data set into a meta-learning training model based on an improved relationship network to obtain a trained rolling bearing fault classification model; the improved relation network comprises an improved embedding module and a relation module, and the improved embedding module and the relation module perform small sample element learning in an end-to-end mode;

and step four, inputting the target domain data set into the trained rolling bearing fault classification model to obtain a rolling bearing fault identification result.

Further, in the first step, original vibration signals of the rolling bearing are collected under 4 load conditions, wherein the load types comprise 0hp, 1hp, 2hp and 3 hp; the training data and the testing data comprise rolling bearing original vibration signals of 10 states including a normal state, inner ring faults, outer ring faults and rolling body faults and different damage degrees.

Further, the preprocessing in the second step includes performing fourier transform on the training data and the test data to obtain a frequency domain amplitude sequence.

Further, the improved embedding module in step three is to improve the original embedding module, where the original embedding module is composed of four convolution blocks, the improved embedding module replaces the second and third convolution blocks with the residual puncturing module, and replaces the fourth convolution block with the residual puncturing module with the SELU activation function.

Further, the third step specifically comprises:

step three, randomly selecting C categories from a source domain data set, selecting K samples from each category as a source domain support set, selecting N samples from the rest samples of each category as a source domain query set, and enabling the source domain support set samples and the source domain query set samples to jointly form a C-way K-shot task;

inputting the source domain support set sample and the source domain query set sample into an improved embedding module, and splicing the output source domain support set characteristics and the source domain query set characteristics to form a characteristic pair;

inputting the feature pairs into a relation module to calculate a relation score, and judging which category the source domain query set sample belongs to according to the relation score; wherein the relationship score is used to represent a similarity between samples in the source domain support set and samples in the source domain query set;

step three, performing the step three to the step three in an iterative cycle manner, optimizing network parameters by using mean square error loss, and stopping performing when the maximum iteration times is reached to obtain a rolling bearing fault classification model.

Further, the residual shrinkage module in the improved embedding module in the third step adds a soft threshold as a nonlinear transformation layer into the residual module, and introduces an attention mechanism; the method specifically comprises the following steps:

firstly, inputting the absolute value of the feature into a global average pooling layer to obtain a one-dimensional vector;

then, inputting the one-dimensional vector into two fully-connected layers, wherein the number of the neurons in the second fully-connected layer is equal to the number of channels of the input features;

then, obtaining the normalization weight of each characteristic channel by a Sigmoid activation layer;

and finally, weighting the normalized weight to each channel of the features by using Scale operation to obtain the threshold value of each channel of the features.

Further, the threshold value calculation formula of each channel of the feature x is as follows:

wherein, tau_cA threshold value representing a characteristic c channel; alpha is alpha_cRepresents the c-th attention weight; i. j, c are the width, height, index of the channel, respectively, of feature x.

Further, the specific steps of the fourth step include:

randomly selecting C categories from the target domain data set, selecting K samples from each category as a target domain support set, selecting N samples from the rest samples of each category as a target domain query set, and enabling the target domain support set samples and the target domain query set samples to jointly form a C-way K-shot task;

and step two, inputting the target domain support set samples and the target domain query set samples into a trained rolling bearing fault classification model, extracting the target domain support set and the target domain query set sample characteristics by utilizing an improved embedded module in the rolling bearing fault classification model, splicing the output target domain support set characteristics and the target domain query set characteristics, inputting the spliced target domain support set characteristics and the target domain query set characteristics into a relation module in the rolling bearing fault classification model to calculate relation scores, and classifying the target domain query set samples according to the relation scores to obtain rolling bearing fault classification results.

The beneficial technical effects of the invention are as follows:

according to the invention, the meta-learning training strategy is combined with the relational network model, and the fault classification model can be trained by using a small amount of labeled samples, so that the problem of insufficient fault data of the rolling bearing in practice is solved; wherein, the relation network is improved: a residual shrinkage module and an 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 more useful features in data can be extracted; the improved relation network can well classify the faults of the rolling bearings of different models under different load conditions, and the problems that the models trained on bearing data of a certain model are poor in generalization capability and cannot be well applied to bearings of other models due to the fact that the models of the rolling bearings in different mechanical equipment are various are solved.

Drawings

The present invention may be better understood by reference to the following description taken in conjunction with the accompanying drawings, which are incorporated in and form a part of this specification, and which are used to further illustrate preferred embodiments of the present invention and to explain the principles and advantages of the present invention.

FIG. 1 is a schematic diagram of a small sample classification task in the present invention;

FIG. 2 is a schematic diagram of a relational network model in accordance with the present invention;

FIG. 3 is a diagram of the soft threshold function of the present invention;

FIG. 4 is a graph of the derivative of the soft threshold function of the present invention;

FIG. 5 is a schematic diagram of a residual module according to the present invention;

FIG. 6 is a block diagram of a residual shrinkage module according to the present invention;

FIG. 7 is a diagram of SELU activation function in the present invention;

FIG. 8 is a diagram of a relational network embedding module before and after improvement in the present invention; wherein, the graph (a) represents a relational network embedding module before improvement; FIG. (b) shows the improved relational network embedding module;

FIG. 9 is a flow chart of the method of the present invention;

FIG. 10 is a schematic illustration of a test stand in an embodiment of the present invention;

FIG. 11 is a comparison of classification accuracy for three experiments in accordance with an embodiment of the present invention;

FIG. 12 is a graph of the effect of feature visualization in an embodiment of the invention; wherein, the diagram (a) represents a 5-way 5-shot task in the first scheme; FIG. (b) shows a 5-way 5-shot task in scenario three; FIG. (c) shows a 5-way 10-shot task in scenario one; FIG. (d) shows the 5-way 10-shot task in scenario three;

FIG. 13 is a comparison graph of classification accuracy before and after improvement of the relational network in an embodiment of the invention.

Detailed Description

In order that those skilled in the art will better understand the disclosure, exemplary embodiments or examples of the disclosure are described below with reference to the accompanying drawings. It is obvious that the described embodiments or examples are only some, but not all embodiments or examples of the invention. All other embodiments or examples obtained by a person of ordinary skill in the art based on the embodiments or examples of the present invention without any creative effort shall fall within the protection scope of the present invention.

The invention provides a rolling bearing fault identification method for improving a relational network, which is used for carrying out fault classification on rolling bearings of different models under different loads under the condition of only a small number of samples. Dividing a data set according to a meta-learning training strategy, introducing a residual shrinkage module and an SELU (self-adaptive selection unit) activation function into an embedding module of a relational network, automatically determining a threshold value by using an attention mechanism through the residual shrinkage module, and removing redundant information in a signal; extracting sample characteristics by using an embedding module, splicing the sample characteristics of the support set and the characteristics of the query set, and inputting the spliced sample characteristics and the characteristics of the query set into a relation module; and finally classifying the query set samples according to the relation scores to realize fault classification of the rolling bearings. The respective steps of the present invention will be explained in detail below.

1. Meta-learning training strategy

Meta-learning (meta-learning) is also called learning to learn (learning to learn)^[19]The method is a leading research framework in the field of machine learning and aims at solving the problem of how to learn the model. The purpose of meta-learning is to make the model acquire a learning ability that allows the model to automatically learn some meta-knowledge. Meta-knowledge refers to knowledge that can be learned outside of the model training process, such as hyper-parameters of the model, initial parameters of the neural network, and,Neural network structure and optimizer^[20]。

Meta-learning is currently mainly directed to the problem of small sample learning. In small sample learning, meta-learning specifically refers to learning meta-knowledge from a large number of prior tasks, and the prior knowledge is utilized to guide a model to learn more quickly in a new task^[21]. The data sets in meta-learning are generally divided into meta-training sets and meta-testing sets, and the categories of data in the meta-training sets and the meta-testing sets are different. In the meta-training and meta-testing stages, small sample tasks are used as basic units, and each task has a training data set and a testing data set, which are also called a support set and a query set. Meta-learning is trained using a large number of small sample tasks in the training phase, and the whole task set is regarded as a training sample. Different meta-tasks are sampled in each task type training, different category combinations are trained, and the mechanism enables the models to learn the common parts of the meta-tasks from the different meta-tasks. The model learned by the learning mechanism has strong generalization capability, and can be well classified without adjusting the existing model when the unseen meta-tasks are faced in the meta-testing stage. The small sample classification task is shown in FIG. 1^[22]。

Meta-learning studies are currently roughly divided into four more independent directions: metric-based method, initialization-based method, optimizer-based method, and external storage-based method^[23]. The direction based on the measurement tends to extract the features contained in the task sample to the maximum extent, and the type attribution of the sample is determined by using a feature comparison mode, so how to extract the features which can represent the features of the sample most becomes the research focus of the direction. The invention adopts a metric-based meta-learning method, wherein the metric learning adopts a relationship network.

2. Relationship network and improvements

Relationship Network (RN) is an architecture for small sample learning proposed by Flood Sung et al in 2018^[24]The method has the advantages of simple concept, and strong flexibility and universality. The relation network consists of an embedded module and a relation module. First, the embedding module extracts features of the input sample, and thenThe relationship module compares the features to determine if they belong to the same category. The embedding module and the relation module carry out small sample element learning in an end-to-end mode. The relational network structure is shown in fig. 2. In the figure, for embedded modules

Representing, the relationship module by g_kAnd (4) showing.

Small sample learning was performed using the meta-learning training strategy described above. In each training iteration, C classes are randomly extracted from the data set, and K samples are extracted from each class to serve as a support set of the model

A part of samples in C classes are used as a query set of the model

Support concentrated samples x_iAnd sample x in the query set_jInputting to an embedding module to obtain features

And

the resulting features are cascaded along the channel by means of a concatenation module C (·,) to yield

Inputting the combined features of the support sample and the query sample to the relationship module g_kIn (2), a scalar of one (0,1) interval is output, called the relationship score r_i,jIs used for representing x_iAnd x_jThe similarity between them. In the single sample case, the relationship score r_i,jThe method specifically comprises the following steps:

for the small sample case, the features of the support set samples of each class are added element by element to form the features of this class, which are then combined with the query set sample features described above. Thus, in a single sample or small sample setting, the number of relationship scores obtained per query sample is always C.

The model is trained in the relational network using a mean square error loss function. The classification problem may be considered a regression problem when the support sample and the query sample assume a relationship score of 1 when they belong to the same class and a relationship score of 0 when they do not belong to the same class. The objective function is:

the residual shrinking module introduces a soft threshold and an attention mechanism on the basis of the residual module^[25]And the attention mechanism automatically determines the threshold value, so that the characteristic with higher vibration signal discriminability can be extracted.

1) Soft threshold

The soft threshold method can retain useful information features in the vibration signal for feature activation of the network model. The functional expression of the soft threshold is:

where x is the input characteristic, y is the output characteristic, and τ is the threshold. The soft threshold zeroes out features that are close to zero, preserving useful negative features, unlike the direct zeroing of negative features in the ReLU activation function.

The soft threshold process is shown in figure 3. It can be seen that the derivative of the output to the input is 1 or 0, which effectively prevents the gradient extinction and gradient explosion problems, as shown in fig. 4. The soft threshold derivative may be expressed as:

setting a suitable threshold value often requires a lot of expertise in signal processing, and the depth framework in the residual puncturing module can be used to automatically determine the threshold value.

2) Residual shrinking module

The residual puncturing module adds soft thresholds as a non-linear transform layer to the residual module and introduces an attention mechanism for estimating the thresholds. Firstly, the absolute value of the feature x is input into a global average pooling layer to obtain a one-dimensional vector. The one-dimensional vectors are then input into two fully-connected layers, the number of neurons in the second fully-connected layer being equal to the number of channels of the input features. And then, obtaining the normalized weight of each characteristic channel by the Sigmoid activation layer, and weighting the normalized weight to each characteristic channel by the last Scale operation to realize the promotion of the important channel. Can be expressed as:

wherein z is_cIs a characteristic of the c-th neuron, α_cIs the c-th attention weight. Then, the threshold is calculated as follows:

wherein, tau_cFor the threshold of the c-th channel of the feature, i, j, c are the width, height, index of the channel, respectively, of the feature x. The threshold should be a positive number that remains within a reasonable range to prevent the output characteristic from being all zeros. The residual block and residual shrink block structures are shown in fig. 5 and 6, respectively.

In 2017, document [26] proposes an activation function named Scaled Exponential Linear Unit (SELU). SELU not only has the advantages of ReLU, and can overcome the problems of gradient disappearance and gradient explosion, but also, unlike the sparsity of ReLU, SELU retains the calculation of input less than 0 part, and can provide richer features. The function is expressed as follows:

wherein lambda is approximately equal to 1.05, alpha is approximately equal to 1.67.

The SELU keeps the non-saturation of the part with the input larger than 0 by setting a lambda slightly larger than 1, keeps the calculation of the part with the input smaller than 0, and ensures that the output distribution of the neuron can be automatically normalized to zero mean and unit variance by setting a constant alpha, thereby extracting more high-dimensional features. The SELU has the dual characteristics of nonlinear mapping activation and self-normalization, so the SELU has certain advantages in theory. Fig. 7 shows a diagram of the SELU activation function.

The improved relational network replaces the second and the third convolution blocks of the embedding module with a residual shrinkage module, and simultaneously replaces the fourth convolution block with a residual module with a SELU activation function. Only a few hidden layers are added in the improved embedded module, the complexity of the network is not increased, and meanwhile, due to the introduction of an attention mechanism and an SELU activation function, the relation network can focus on more useful characteristics in the vibration signal. The residual shrinkage module is used for processing the vibration signal, so that the residual shrinkage module is introduced into a relation network and is suitable for rolling bearing fault classification. The specific structure of the improved context network is shown in fig. 8(a) and (b).

3. Rolling bearing fault identification method

The method comprises the steps of classifying faults and normal states of inner rings, outer rings and rolling bodies of rolling bearings of different models by utilizing an improved relation network, wherein the diagnosis process comprises a data preprocessing stage, a meta-training stage and a meta-testing stage, and the specific flow is shown in fig. 9.

(1) Data pre-processing

Carrying out Fourier transform on original vibration signals of rolling bearings of different models to obtain frequency domain amplitude sequences, and respectively forming a source domain data set and a target domain data set for the meta-training and meta-testing stages. C categories are randomly selected from a source domain, K samples are selected from each category to serve as a support set, and N samples are selected from the remaining samples of each category to serve as a query set. The support set and the query set sample jointly form a C-way K-shot task, and different tasks are taken in each iterative training. The target domain builds tasks in the same way for testing.

(2) Meta training phase

And inputting the support set samples and the query set samples of the source domain into an embedding module of the improved relational network, wherein the improved embedding module automatically determines a threshold value through an attention mechanism, and simultaneously introduces a SELU (self-adaptive optimization) activation function, focuses on more details and extracts more useful characteristics of the samples. And splicing the obtained support set characteristics and the query set characteristics to form a characteristic pair. The feature pairs are input into a relationship module to calculate a relationship score. And judging which category the query set sample belongs to by the relationship score. And (4) obtaining a rolling bearing fault classification model by using parameters of the mean square error loss optimization network.

(3) Meta test phase

Inputting the support set and the query set samples of the target domain into a trained classification model, extracting the characteristics of the support set and the query set samples by using an embedding module of the model, inputting the characteristics into a relation module after splicing, classifying the query set samples of the target domain according to the relation scores, and directly obtaining the fault classification results of bearings of different models without updating the network.

4. Application and analysis

Experimental data of the embodiment of the invention is from a public bearing data set of the university of Kaiser storage^[27]. A schematic view of the bearing test stand is shown in fig. 10. The model of the drive end bearing is SKF6205, and the model of the fan end bearing is SKF 6203. The bearing vibration signal data is collected by a 16-channel acceleration sensor, the sampling frequency is 12kHz, and the bearing vibration signals are collected under four different loads of 0,1, 2 and 3hp (hp is horsepower manufactured by English, and 1hp is 0.75 kw).

The embodiment of the invention uses two types of rolling bearing data of SKF6205 and SKF6203, wherein each type of rolling bearing data comprises a normal state, and 10 states of faults and different damage degrees of an inner ring, an outer ring and a rolling body. IR (7,14,21) indicates inner ring damage diameters of 7mils,14mils and 21mils, respectively, and similarly, OR (7,14,21) and B (7,14,21) indicate different failure degrees of the outer ring and the rolling elements, respectively. The vibration signals in the data set are truncated into 2048-point samples, and each sample is subjected to fast Fourier transform, so that the length of the sample is reduced to half of the original length. The rolling bearings of two models have 10 states respectively, and each state has 16 samples and is used for solving the problem of classification of small-sample bearings. The experimental data set is shown in table 1. Each experiment divides data according to a meta-learning training strategy, 5 categories (C is 5) are selected in each task, and the number of samples in the support set comprises 1,3, 5 and 10 samples (K is 1,3, 5 and 10). And (3) performing end-to-end training on the network by adopting an Adam optimization algorithm, setting the learning rate to be 0.001, setting the maximum epsilon iteration number to be 50, and repeating the experiment for 5 times.

TABLE 1 Experimental data set

And combining the relational network model with the meta-learning training strategy, and carrying out fault diagnosis on rolling bearings of different models under different load conditions. The two types of bearing data used in the embodiment respectively have 10 states, the load conditions are respectively four in the working process, and the difficulty of classifying the faults of the rolling bearing can be increased due to the change of the load. Three sets of experiments are set, and different types of bearing faults are classified under different load conditions, for example: 6205_1hp/6203_023hp indicates that model SKF6205 bearing data collected under the loads of 0, 2 and 3hp are used for fault classification by training a classification model by utilizing model SKF6205 bearing data collected when the working load is 1 hp. The results of the experiment are shown in table 2.

TABLE 2 Fault classification results for different types of rolling bearings under different loads

FIG. 11 shows the classification accuracy of three sets of experiments under the 5-way 1-shot task. As can be seen from the first and third sets of experimental results in fig. 11, the greater the load condition of the target domain, the greater the difficulty of fault classification, and the second set of experimental results shows that the load condition in the source domain is increased and the classification effect of the model is not improved. As can be seen from table 2, load variations also have the same effect on the classification results at the other three sample settings. When the target domain data only has a single load, the classification effect is good. Experimental results show that data under various loads are mixed together, certain interference can be caused to fault classification of bearings of different models, and the problem can be well solved by a relational network.

And further performing a comparison experiment on classification effects before and after the relational network is improved, wherein the first scheme uses the original relational network, the second scheme only introduces the residual error shrinkage module into an embedded module of the relational network, and the third scheme is the improved relational network of the invention, namely simultaneously introducing the residual error shrinkage module and the SELU activation function into the embedded module. Multiple sets of experiments were set up at different loads and the results are shown in table 3.

TABLE 3 results of classification of rolling bearing fault of different models before and after improvement of relational network

According to experimental results, under different loads, the bearing fault is classified by using the relation network only introducing the residual shrinkage module, and the classification effect is improved. On the basis, an SELU activation function is further introduced, the classification accuracy is improved by 3.44% at most compared with that of the original relational network, and the classification accuracy can be over 86% under the condition of a single sample, so that the improved relational network can accurately classify the faults of rolling bearings of different models.

In order to more intuitively express the classification effect of the improved relational network, in 6205_0123hp/6203_0123hp experiment, a t-SNE algorithm is utilized^[28]And visualizing the sample characteristics of the query set in the 5-way 5-shot task and the 5-way 10-shot task, and simultaneously drawing a histogram according to the fault classification result of the rolling bearing, wherein the histogram is respectively shown in fig. 12 and fig. 13.

As can be seen from fig. 12, there is an overlap between the rolling bearing multi-state features extracted by the relationship network before the improvement (solution one), which is not easy to distinguish, and this may negatively affect the classification effect. The improved relation network (scheme III) introduces an attention mechanism and a nonlinear SELU activation function, can focus on more useful parts in data, has small inter-class spacing and large inter-class spacing of extracted features, can distinguish different states of the bearing more easily, and has ideal classification effect on most tasks. In addition, as can be seen from fig. 13, when the number of support set samples in the task is increased from 1-shot to 10-shot, the fault diagnosis accuracy of the model is improved.

In order to further prove the effectiveness of the method, a laboratory collected data set with the bearing model of 6307E is supplemented as experimental data, and fault diagnosis is carried out on bearings of different models. The model number of the bearing data set is 6307E, the rotating speed is 680r/s, the sampling frequency is 8192Hz, the bearing data set comprises 3 state data including inner ring faults, outer ring faults and normal states, and each state has 16 samples. The results of the classification of the faults of the rolling bearings of different models are shown in table 4.

TABLE 4 Fault Classification results for different types of rolling bearings

As can be seen from Table 4, the accuracy of classifying the faults of the rolling bearings of different models can reach more than 90%, and the method disclosed by the invention can better realize the fault diagnosis of the bearings under small samples.

The documents cited in the present invention are as follows:

[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.

[2]Chen C,Shen F,Xu J,et al.Model Parameter Transfer for Gear Fault Diagnosis under Varying Working Conditions[J].Chinese Journal of Mechanical Engineering,2021.

[3] deep migration diagnosis method of mechanical equipment fault under big data [ J ] mechanical engineering report, 2019,55(07):1-8.

[4]Hoang D T,Kang H J.A survey on Deep Learning based bearing fault diagnosis[J].Neurocomputing,2019,335:327-335.

[5]Liu R,Yang B,Zio E,et al.Artificial intelligence for fault diagnosis of rotating machinery:A review[J].Mechanical Systems&Signal Processing,2018,108:33-47.

[6]Lei Y,Yang B,Jiang X,et al.Applications of machine learning to machine fault diagnosis:A review and roadmap[J].Mechanical Systems and Signal Processing,2020,138:106587-.

[7] Robust leaves, residual wave, a gearbox fault diagnosis method based on multi-channel one-dimensional convolutional neural network feature learning [ J ] vibration and impact 2020,39(20):55-66.

[8] Wangyizhi, Nazidong, Kongxian, etc. the rolling bearing state identification method under variable load based on EEMD-Hilbert envelope spectrum and DBN [ J ] China Motor engineering report, 2017,37(23): 6943-.

[9] A rolling bearing self-adaptive fault diagnosis algorithm based on a one-dimensional convolutional neural network [ J ] is reported by instruments and meters, 2018,39(07):134-143.

[10]Zhao M,Kang M,Tang B,et al.Multiple Wavelet Coefficients Fusion in Deep Residual Networks for Fault Diagnosis[J].IEEE Transactions on Industrial Electronics,2019,66(6):4696-4706.

[11]Ma S,F Chu,Han Q.Deep residual learning with demodulated time-frequency features for fault diagnosis of planetary gearbox under nonstationary running conditions[J].Mechanical Systems and Signal Processing,2019,127:190-201.

[12]Yan R,Shen F,Sun C,et al.Knowledge Transfer for Rotary Machine Fault Diagnosis[J].IEEE Sensors Journal,2020,20(15):8374-8393.

[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).

[14] Wangyun, Luhai rock, Consumer, etc. the fault diagnosis method of rolling bearings of different models [ J ] China Motor engineering report 2021,41(01):267 and 276+416.

[15] Xylonite, Liushijian, Wangzao, and the like, a small sample infrared air target classification method [ J ] based on an improved relationship network, 2020,40(08):87-96.

[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.

[17] Yuhaoyaoshuai, Tangbaoping, Zhang regular script, and the like.wind power gear box fault diagnosis method mixed from an attention prototype network under small samples [ J/OL ]. China mechanical engineering: 1-8[2021-06-02]. http:// kns.cnki.net/kcms/detail/42.1294.th.20210330.1320.006.html.

[18] Wu Ziyang, a small sample learning algorithm based on metric learning study [ D ]. southern China university of science 2020.

[19]Thrun S,Pratt L.Learning to learn:Introduction and overview[M].Springer US,1998.3-17.

[20]Vilalta R,Drissi Y.A Perspective View and Survey of Meta-Learning[J].Artificial Intelligence Review,2002,18(2):77-95.

[21] Zhao Kaili, long and slender, Wangyuan Zhuo, small sample study reviews [ J ] software newspapers, 2021,32(02):349 and 369.

[22]Ravi S,Larochelle H.Optimization as a model for few-shot learning[C].International Conference on Learning Representations,Toulon,France,2017:1-3.

[23] Li Fang Chang, Liuyang, Wupeng Xiang, etc. Yuan learning research review [ J ] The computer science report 2021,44(02): 422-.

[24]Sung F,Yang Y,Li Z,et al.Learning to Compare:Relation Network for Few-Shot Learning[C].2018IEEE/CVF Conference on Computer Vision and Pattern Recognition.IEEE,2018.

[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.

[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.

[27]Case Western Reserve University Bearing Data Center,http:// csegroups.case.edu/bearingdatacenter/home.

[28]Der Maaten L V,Hinton G E.Visualizing data using t-SNE[J].Journal of Machine Learning Research,2008,(9):2579-2605.

While the invention has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of this description, will appreciate that other embodiments can be devised which do not depart from the scope of the invention as described herein. The present invention has been disclosed in an illustrative rather than a restrictive sense, and the scope of the present invention is defined by the appended claims.

Claims

1. A rolling bearing fault identification method for improving a relational network is characterized by comprising the following steps:

2. The method for identifying the fault of the rolling bearing with the improved relational network as claimed in claim 1, wherein in the step one, original vibration signals of the rolling bearing are collected under 4 load conditions, and the load types comprise 0hp, 1hp, 2hp and 3 hp; the training data and the testing data comprise rolling bearing original vibration signals of 10 states including a normal state, inner ring faults, outer ring faults and rolling body faults and different damage degrees.

3. The method for identifying the rolling bearing fault of the improved relational network as claimed in claim 2, wherein the preprocessing in the second step comprises performing fourier transform on the training data and the test data to obtain a frequency domain amplitude sequence.

4. The method according to claim 3, wherein in step three, the improved embedding module is an improvement on an original embedding module, the original embedding module is composed of four convolution blocks, the improved embedding module replaces the second and third convolution blocks with a residual puncturing module, and replaces the fourth convolution block with a residual puncturing module with a SELU activation function.

5. The method for identifying the rolling bearing fault of the improved relationship network as claimed in claim 4, wherein the concrete steps of the third step comprise:

6. The method for identifying the rolling bearing fault of the improved relationship network as claimed in claim 5, wherein the residual shrinkage module in the improved embedding module in the third step adds a soft threshold as a nonlinear transformation layer to the residual module and introduces an attention mechanism; the method specifically comprises the following steps:

7. The method for identifying the rolling bearing fault of the improved relational network according to claim 6, wherein the threshold calculation formula of each channel of the characteristic x is as follows:

8. The method for identifying the rolling bearing fault of the improved relationship network as claimed in claim 7, wherein the concrete steps of the fourth step comprise: