CN114861778A - Method for rapidly classifying rolling bearing states under different loads by improving width transfer learning - Google Patents

Abstract

A method for rapidly classifying rolling bearing states under different loads by improving width transfer learning relates to the technical field of rolling bearing fault classification, and is used for solving the problems that deep learning network training is time-consuming and the distribution difference of source domain data and target domain data of a rolling bearing under different loads is large. The technical points of the invention comprise: establishing an enhanced node window of a width learning system (BLS) in a cyclic extension mode, introducing a Maxout activation function into the enhanced layer to construct an improved BLS network, introducing a genetic algorithm to optimize a network node structure, and establishing a pre-training model based on source domain data; and migrating the network parameters of the pre-training model, the weight parameters of the characteristic layer and the enhancement layer to a target domain network, and fine-tuning the network by using a small number of target domain training samples to establish a state classification model. The method can establish a classification model in a shorter time and obtain good classification accuracy, and is suitable for quickly classifying the fault states of the rolling bearing under different loads.

Description

Method for rapidly classifying rolling bearing states under different loads by improving width transfer learning

Technical Field

The invention relates to the technical field of rolling bearing fault classification, in particular to a method for rapidly classifying rolling bearing states under different loads by improving width migration learning.

Background

Rolling bearings are important parts of rotating mechanical equipment and are very vulnerable to damage due to high-intensity motion and friction loss ^[1,2] . Once the rolling bearing is in failure, mechanical equipment is shut down to influence production, and the personal safety of staff can be seriously injured ^[3,4] . The intelligent fault diagnosis technology of the existing rolling bearing is mostly based on deep learning, the training consumes a long time, and in addition, the rolling bearing is in actual work, the load is often changed, so that the rolling bearing state diagnosis under different loads is realized quickly, and the intelligent fault diagnosis method has important significance.

Deep learning receives more and more attention in the field of intelligent fault diagnosis of rolling bearings ^[5] . Document [6]The method has the advantages that the super-parameters of a Deep Belief Network (DBN) are optimized by adopting a salp algorithm, manual selection is avoided, and experiments prove that the method has higher diagnosis precision in fault diagnosis of the rolling bearing; document [ 7]]The feature extraction method based on the deep automatic encoder and the ensemble learning is provided, and the generalization performance of the deep automatic encoder on the fault diagnosis of the rolling bearing is improved; document [8]An attention mechanism is introduced into a one-dimensional Convolutional Neural Network (CNN), so that the learning capability of the CNN on the fault characteristics of the rolling bearing is further enhanced, and higher learning capability is obtainedAnd (4) classification accuracy.

Although the deep learning method has a good effect in fault diagnosis of the rolling bearing, the following limitations exist in the condition diagnosis of the rolling bearing with different loads: (1) the deep learning network generally needs larger training samples, while in engineering practice, vibration data with marking information is scarce, and the sample size obtained in a fault state is limited; (2) the data distribution difference of the rolling bearings under different loads is large, and the single deep learning network has poor adaptability to the rolling bearing state diagnosis under different loads.

The transfer learning can be realized by mining the internal knowledge relation between the source domain and the target domain, and assisting the network learning of the target domain by virtue of the source domain knowledge ^[9,10] . The model-based migration learning strategy is simple and easy to implement, the fine tuning algorithm can complete network training by means of a small amount of labeled data of a target domain, and meanwhile the model migration effect is improved. Document [11]Based on the CNN and the model migration idea, the migration gradient characteristics are proposed to a target domain to obtain a state recognition model, so that the residual life prediction of the rolling bearing under different loads is realized; document [12]]A model migration learning method combining Long Term and Short Term Memory (LSTM) and migration learning is provided, fault diagnosis of the rolling bearing under different loads is achieved, and high accuracy is obtained.

The deep migration learning method breaks through 2 limitations existing in the rolling bearing state diagnosis under different loads, but the deep migration learning method is based on a deep learning network, and the deep learning network is complex in structure, multiple in training parameters and slow in training process, so that the deep migration learning network has the problem of time consumption in training.

Document [13] proposes a "flat" neural network structure, namely the Broad Learning System (BLS), which has the advantages of simple network structure, few calculation parameters, and fast training speed. The document [14] combines CNN and BLS, proposes a deep-width learning framework based on domain knowledge, and tests the performance of the deep-width learning framework on a rolling bearing and a pipeline defect data set, and the result shows that the framework has superior performance compared with the traditional learning method; document [15] proposes an adaptive BLS that enables a fault diagnosis of a rolling bearing quickly and with high accuracy. From the above, BLS has been applied to the field of fault diagnosis of rolling bearings.

The model migration learning method based on the BLS network is an idea for rapidly solving the fault diagnosis of the rolling bearing under different loads, and is inspired by a deep model migration learning method. However, the number of layers of the BLS network is small, and data features are extracted depending on feature nodes and enhanced nodes, so that the feature extraction capability and generalization capability of the BLS network need to be further improved, and meanwhile, the adaptive determination of the network node structure needs to be further studied.

Disclosure of Invention

In view of the above problems, the invention provides a method for rapidly classifying rolling bearing states under different loads by improving width migration learning, which is used for solving the problems that the deep learning network training consumes time and the distribution difference of source domain data and target domain data of the rolling bearing under different loads is large.

A method for improving width transfer learning and quickly classifying rolling bearing states under different loads comprises the following steps:

firstly, acquiring a rolling bearing time domain vibration signal under a load A as source domain data, and acquiring a rolling bearing time domain vibration signal under a load B as target domain data;

secondly, preprocessing the source domain data and the target domain data;

inputting the preprocessed source domain data into a source domain pre-training model based on an improved width learning system for training to obtain source domain pre-training model parameters; the source domain pre-training model parameters comprise the number N of characteristic nodes in each characteristic node window in the width learning system network ₁ Number of feature node windows N ₂ The number N of enhanced nodes in each enhanced node window ₃ Enhancing the number N of node windows ₄ And a feature layer weight parameter W _e And the enhancement layer weight parameter W _h ；

Transferring the source domain pre-training model parameters to a target domain network based on an improved width learning system to serve as initial network parameters; inputting the preprocessed target domain data into a target domain network based on an improved width learning system for training, and acquiring a rolling bearing state classification model based on the improved width model transfer learning;

and fifthly, inputting the preprocessed rolling bearing data to be classified into the rolling bearing state classification model based on the improved width transfer learning to obtain a rolling bearing state classification result.

Further, in the first step, the load a and the load B are different and both include one or more load types.

Further, the rolling bearing state comprises a normal state and a fault state, and the fault state comprises multiple fault damage degrees of inner ring fault, outer ring fault and rolling body fault.

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

Further, in the third step, a source domain pre-training model network node structure based on an improved width learning system is determined in a self-adaptive manner by combining a genetic algorithm GA, and the method specifically comprises the following steps:

step three, initializing GA: setting population size, cross probability, mutation probability, chromosome coding mode, next generation population selection mode and iteration stop conditions of the algorithm;

step three, defining a search space: n in source domain pre-training model parameters ₁ 、N ₂ 、N ₃ And N ₄ Forming a search space of the GA;

step three, obtaining the fitness value of individuals in the GA population: the source domain training sample passes through the feature layer weight parameter W _e And linear activation function to obtain mapping characteristics, and passing the weight parameter W of the enhancement layer according to the mapping characteristics _h Obtaining enhanced features with the Maxout activation function, using the mapping features and the enhanced features as the input of the output layer of the source domain pre-training model based on the improved width learning system, and using a ridge regression algorithmCalculating the weight W of the output layer of the source domain pre-training model to obtain the classification accuracy;

step four, comparing fitness values of individuals in the population, and obtaining a next generation population according to GA initialization parameters;

step three, judging whether an iteration stopping condition is reached, if so, obtaining a source domain pre-training model, and if not, repeating the step three to the step three;

and step three, inputting the source domain test sample into the source domain pre-training model, and verifying the source domain pre-training model.

Further, in step three, the weight parameter W of the enhancement layer is passed according to the mapping characteristics _h And the specific process of obtaining the enhanced features by the Maxout activating function comprises the following steps:

assume that each enhanced node window contains N ₃ Enhancement mapping, then the kth enhancement feature is represented as:

in the formula: z represents a mapping feature;

representing a kth weight parameter of the enhancement layer;

representing the k-th bias parameter of the enhancement layer; 1, 2, N ₃ ；

Definition H _n For an enhanced feature in an enhanced node window, obtaining the enhanced feature by adopting a Maxout activation function, then H _n ＝max{h ₁ ，h ₂ ，...，h _N3 }，n＝1，2，...，N ₄ (ii) a Enhancement layer N ₄ The enhanced features of each enhanced node window are expressed as:

further, in the fourth step, the target domain network based on the improved width learning system is combined with a ridge regression method to fine tune the output layer weight of the target domain network in the training process.

The beneficial technical effects of the invention are as follows:

the invention provides a rolling bearing state rapid classification method based on improved width transfer learning, aiming at the problems that deep learning network training is time-consuming and the distribution difference of source domain data and target domain data of a rolling bearing under different loads is large. Firstly, carrying out fast Fourier transform on vibration signals of a rolling bearing under different loads to construct a frequency domain amplitude sequence data set, and selecting a certain load data set or some load data sets as a source domain and other load data sets as a target domain; secondly, establishing an enhanced node window of a Broad Learning System (BLS) in a cyclic extension mode, introducing a Maxout activation function into the enhanced layer to construct an improved BLS network, introducing a genetic algorithm to optimize a network node structure, and establishing a pre-training model based on source domain data; and finally, migrating the network parameters of the pre-training model, the weight parameters of the characteristic layer and the enhancement layer to a target domain network, and finely adjusting the network by using a small amount of target domain samples to establish a state classification model. The experimental result shows that the average training time of the method is 32.6s, and the average test accuracy is 98.9%. Compared with other methods, the method provided by the invention can establish a classification model in a shorter time and obtain good classification accuracy.

Drawings

Fig. 1 is a schematic diagram of an unmodified BLS network.

FIG. 2 is a schematic diagram of a process for improving BLS network generation enhancement features in an embodiment of the present invention;

FIG. 3 is a schematic diagram of a parameter migration strategy adopted in the embodiment of the present invention to improve the width migration learning;

FIG. 4 is a block diagram illustrating an overall flowchart of a method for rapidly classifying rolling bearing states under different loads according to an embodiment of the present invention, with improved width migration learning;

FIG. 5 is a schematic view of an experimental apparatus according to an embodiment of the present invention;

FIG. 6 is a time domain signal diagram (fig. (a)) and a frequency domain amplitude signal diagram (fig. (B)) of the rolling bearing B07 according to the embodiment of the present invention;

FIG. 7 is a visualization diagram of comparison experiment results before and after optimization of network parameters in an embodiment of the present invention;

FIG. 8 is a graph showing the results of comparative experiments for 7 different methods in example of the present invention;

FIG. 9 is a graph of experimental results of 5 experiments performed on different migration tasks according to an embodiment of the present invention;

fig. 10 is a graph of ablation experiment results in an embodiment of the present 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 state fast classification method under different loads for improving width transfer learning, which comprises the steps of establishing an enhanced node of a BLS network in a cyclic extension mode, introducing a Maxout activation function into the enhanced layer, constructing the improved BLS network, and further providing the improved width transfer learning network by combining a transfer learning method on the basis of the improved BLS network so as to realize the fast classification of the rolling bearing states under different loads; meanwhile, aiming at the problem of how to adaptively determine the network node structure, the method provides the method for adaptively determining the network node structure by combining a Genetic Algorithm (GA) so as to improve the generalization capability of the improved breadth migration learning network.

A first embodiment of the present invention proposes an improved width learning network (BLS). Fig. 1 is a schematic diagram of an unmodified BLS network. The BLS network is a shallow neural network, has the characteristics of less hyper-parameters, less network layers and simple structure, and provides an efficient method for solving the classification problem. In the BLS method, firstly, input data X is subjected to linear transformation to form mapping characteristics; then, the mapping feature is connected to the enhancement layer through a nonlinear activation function to form an enhancement feature; finally, the enhanced features and the mapping features are merged and input to an output layer together.

Specific training process of BLS network: given a set of training data { X, Y }, X ∈ R ^N×M Representing that the training data set has N training samples, each training sample has M-dimensional characteristics, and Y belongs to R ^N×C Indicating that there are C class labels for the N training samples. Suppose a BLS network has N per characteristic node window ₁ And mapping features, the ith mapping feature is expressed as:

in the formula: weight of

And deviation of

Randomly generated and optimized by a sparse auto-encoder,

typically a linear transformation.

Defining:

for a mapped feature within a feature node window, the final feature level N ₂ The mapping characteristic of the characteristic node window can be expressed as

The mapping feature Z is connected to the enhancement layer by a non-linear transformation to form an enhancement feature, assuming that the BLS network contains N ₃ And the j enhanced feature is expressed as:

in the formula: the activation function ξ ═ tanh (x),

and

and (4) randomly generating.

The enhancement features of the enhancement layer are represented as

Mapping feature Z and enhancement feature H are input together to the output layer of the BLS network, and given that the label of input data X is Y, the relationship of mapping feature Z, enhancement feature H and Y can be expressed as:

Y＝[Z|H]W (3)

in the formula: w is the weight connecting the feature layer and the enhancement layer to the output layer, and the way W is calculated as shown in equation (4):

W＝[Z|H] ⁺ Y (4)

in the formula: let the feature matrix a ═ Z | H]When W is equal to A ⁺ Y，A ⁺ Is a pseudo-inverse matrix of a. The pseudo-inverse is the least-squares estimator of a linear equation, the objective of which is to obtain the output weight W with the minimum training error, and in particular, when the feature matrix A is not a full-rank matrix, calculate A ⁺ An error is made. Therefore, a ridge regression algorithm is introduced, and W is solved by using the formula (5).

In the formula: sigma ₁ >0，σ ₂ >0, u, v denotes norm regularization, by σ ₁ ＝σ ₂ Setting the optimization problem of equation (5) to L2 norm regularization, where u-v-2; lambda [ alpha ]>0, represents a limit on the weight.

The regularization function represented by the formula (5) is a convex function, and has better generalization performance. Theoretically, if λ is 0, equation (5) evolves to solve the least squares problem; if λ → ∞, the solution of equation (5) is severely limited and will tend to 0. Therefore, λ → 0 is usually set. Taking the derivative of equation (5) to equal 0, one obtains:

W＝(λI+AA ^T ) ^-1 A ^T Y (6)

in the formula: i denotes a unit matrix.

Because the BLS network only contains one enhancement layer, the BLS network randomly generates the image with the dimension of N and meets the standard normal distribution at one time ₃ The time complexity is reduced by the enhanced feature weight matrix, but the nonlinear structure of the BLS network is mainly embodied in the enhanced layer. In order to further enhance the nonlinear expression capability of the BLS network, the invention designs N on the basis of the BLS network ₄ And each enhanced node window randomly generates an enhanced feature weight matrix in each enhanced node window, and further adopts a Maxout activation function to generate enhanced features to construct an improved BLS network.

The Maxout activation function is a function which can be learned according to weight parameters, the core of the Maxout activation function is to take the maximum value of a series of linear functions as the activation value of the activation function, and when the number of the linear functions tends to be infinite, the function with any dimensionality can be fitted. Therefore, Maxout has a stronger fitting capability compared to the traditional Sigmoid and ReLU activation functions.

The process proposed by the present invention to improve the BLS network generation enhancement features is illustrated in fig. 2. Assume that each enhanced node window contains N ₃ Enhancement mapping, then the kth enhancement feature is represented as:

in the formula:

and

and (4) randomly generating.

Definition H _n For an enhanced node windowEnhancing features in the mouth by using Maxout activation function

Finally, enhancement layer N ₄ The enhanced features of each enhanced node window may be expressed as:

the algorithm for improving the BLS network to obtain the enhanced feature H is shown as algorithm 1 below.

The second embodiment of the invention provides a fast classification method for improving the transfer learning of a width model.

Currently, the migration learning method is mainly classified into 4 types: sample-based migration, feature-based migration, model-based migration, and relationship-based migration ^[16] . Compared with other methods, the model-based migration learning method has the advantages that the relevant knowledge in the source domain is migrated to the target domain in a parameter transfer mode, the method is simple and easy to implement, in addition, the successful application of the fine tuning algorithm in the deep model migration learning network greatly improves the space of the model-based migration strategy research ^[17] 。

Compared with a deep learning network, the width learning network is more prone to be constructed in the width direction, and the improved width model transfer learning embeds the transfer learning idea into the improved width learning network. The field is an important concept in the transfer learning, data and probability distribution thereof are represented, a source domain and a target domain are respectively marked by S and T generally, and the task of the transfer learning is to assist the target domain learning by virtue of source domain knowledge. In the method for improving the migration learning of the width model, the source domain data is recorded as

Wherein n is _S Is the number of samples in the source domain,

for the ith sample in the source domain,

is its corresponding tag; noting the target domain data as

Wherein n is _T Is the number of samples of the target domain,

for the ith sample in the target domain,

is its corresponding tag. The core of improving the width model transfer learning is to utilize D ^S Training the improved width learning network by using the training samples to obtain a pre-training model, selecting appropriate parameters to migrate to a target domain network, and using D ^T The target domain network is finely adjusted by a small amount of samples, the output layer parameters of the target domain network are further optimized to adapt to the target domain task, and the problem that the data distribution of a source domain and the data distribution of a target domain are different is solved.

Jason Yosinski et al ^[18] It is pointed out that the first few layers of the migration neural network have better parameter effect, so the method selects the characteristic layer weight parameter W in the migration improved width learning network _e And the enhancement layer weight parameter W _h And fine-tuning the output layer weight W of the target domain network by using a ridge regression method. A schematic diagram of a parameter migration strategy adopted by the improved width model migration learning is shown in fig. 3.

The performance of improving the width model transfer learning depends on the following 4 parameters: the number of the feature nodes in each feature node window, the number of the feature node windows, the number of the enhanced nodes in each enhanced node window and the number of the enhanced node windows are respectively N ₁ 、N ₂ 、N ₃ 、N ₄ And (4) showing. GA is introduced to optimize network parameters for improving the width model transfer learning, and the improved width model transfer learning is improvedThe generalization performance of the conventional method can obtain higher classification accuracy. The overall flow chart of the rolling bearing state rapid classification method based on the improved width model transfer learning under different loads is shown in fig. 4. The specific process steps of the classification method are as follows:

(1) data pre-processing

And performing fast Fourier transform on the original vibration signal of the rolling bearing to obtain a frequency domain amplitude sequence, wherein the amplitude sequence under a certain load (certain loads) is used as a source domain sample, and the amplitude sequence under other loads is used as a target domain sample.

(2) Constructing a pre-training model

Firstly, initializing GA: setting population size, cross probability, mutation probability, chromosome coding mode, next generation population selection mode and iteration stop conditions of the algorithm;

define the search space: network parameter N in source domain improved breadth learning network ₁ 、N ₂ 、N ₃ And N ₄ Forming a search space of the GA, wherein the variation range of the search space is 1-50;

obtaining the fitness value of the individuals in the GA population: the source domain training sample passes through the weight parameter W _e And linear activation function to obtain mapping characteristics, and the mapping characteristics are further processed by weight parameters W _h Obtaining an enhanced feature by the Maxout activation function, taking the mapping feature and the enhanced feature as the input of a source domain improved width learning network output layer, and calculating the weight W of the source domain network output layer by using a ridge regression algorithm to obtain the classification accuracy;

comparing the fitness values of the individuals in the population, and obtaining the next generation of population according to GA initialization parameters;

judging whether an iteration stopping condition is reached, if so, obtaining a pre-training model, and if not, repeating the third step and the fifth step;

and sixthly, inputting the source domain test sample into the pre-training model, and verifying the pre-training model.

(3) Building state classification models

Pre-training N in the model ₁ 、N ₂ 、N ₃ And N ₄ And W _e And W _h And (4) transferring to a target domain network, using a small number of training samples in the target domain, and finely adjusting the output layer weight W of the target domain improved width learning network by combining a ridge regression method to construct a rolling bearing state classification model.

(4) Rolling bearing condition testing under different loads

Through the steps, the rolling bearing state diagnosis models under different loads based on the improved width transfer learning are obtained, and the target domain test set samples are input into the models to obtain the classification results.

The technical effect of the invention is further verified through experiments.

The experimental data is from CWRU database of Kaiser university ^[19] The test system for collecting vibration data of the rolling bearing comprises a driving motor, a load and a control circuit, and the schematic diagram of the experimental device is shown in fig. 5. The embodiment of the invention uses the state data of an SKF6205 model rolling bearing with the sampling frequency of 48kHz and the driving end, the data set comprises load data of 0hp, 1hp, 2hp and 3hp (hp is English horsepower, 1hp is 0.75kW), and the corresponding motor rotating speeds are 1797rpm, 1772rpm, 1750rpm and 1730rpm respectively.

The rolling bearing has 10 types of states, specifically: b07, B14, B21, IR07, IR14, IR21, OR07, OR14, OR21 and N. Where N denotes a normal state, B denotes a failure position at the rolling elements, IR denotes a failure position at the inner ring, OR denotes a failure position at the outer ring, and 07, 14, and 21 denote failures in which the rolling bearings generate diameters of 0.1778mm, 0.3556mm, and 0.5334mm, respectively.

In the 10 state classification experiments of the rolling bearing under different loads, a sample set under a certain load (certain loads) is used as source domain data, a sample set under other loads is used as target domain data, and the classification performance of the width migration learning network is improved when the load change in the actual work of the rolling bearing is verified. The computer used for the experiment is a Windows10 operating system, the size of the memory is 16GB, and the model of the processor is Intel Core i 5-8265U.

And FFT is carried out on the vibration data of the rolling bearing at 4096 points of each sample to obtain a frequency domain amplitude sequence of 2048 points. Taking the load of 1hp as an example, fig. 6 shows a time domain signal diagram and a frequency domain amplitude signal diagram of the rolling bearing B07 state.

The sample set of rolling bearing state classification experiments under different loads is shown in table 1. Sample set 2_3 represents 2hp load data as source domain, 3hp load data as target domain, 2_13 represents 2hp load data as source domain, 1hp and 3hp load data as target domain; similarly, other sample sets are meant and so on.

TABLE 1 sample set constitution

In the experiment, GA parameters are specifically set as follows: the population size is 8, the cross probability is 0.95, the mutation probability is 0.05, the chromosome coding mode is binary coding, the next generation population selection mode is elite reservation tournament selection, and the iteration stop condition is set as: the size of the judgment threshold value of iteration stagnation is 1 multiplied by 10 ^-5 And finishing the optimization when the GA is judged to continuously iterate and stagnate for 1 generation.

On the premise that the proportion of the target domain training sample is 20%, an ablation experiment is adopted to respectively illustrate the improvement effect of the BLS network, the GA optimization effect and the model migration effect. In a comparison experiment for verifying BLS network improvement and model migration, 6 groups of sample sets listed in table 1 are respectively subjected to 5 repeated experiments, the average value of accuracy and training time is taken as the experiment result of each group, and the experiment results of the 6 groups of sample sets are averaged to obtain the final comparison result.

On the premise of using model migration and GA to optimize the network node structure, comparative experimental analysis was performed before and after BLS network improvement, and the results are shown in table 2. Compared with the BLS network, the average training time for improving the BLS network model building is slightly increased, but the obtained average test accuracy is higher by 1.4%, which shows that the improved BLS network maintains the advantage of fast classification of the BLS network and enhances the feature extraction capability of the BLS network.

TABLE 2 BLS network improvement and pre-and post-experimental results

To analyze the effect of GA optimization on the premise of model migration based on the improved BLS network, 5 sets of network parameters were randomly set, as shown in table 3.

Table 3 randomly set 5 sets of network parameters

The target domain test accuracy obtained according to the randomly set network parameters is compared with the experimental result of the GA optimization, and the comparison result is visualized, and the result is shown in FIG. 7. In FIG. 7, A represents the target domain test accuracy, N ₁ 、N ₂ 、N ₃ And N ₄ Representing a network parameter. As can be seen from fig. 7, the selection of the source domain network parameters has a large influence on the test accuracy of the target domain, and the diagnosis effect of the target domain network can be improved by migrating the appropriate network parameter combinations to the target domain. Taking a 2_3 sample set as an example, a 2hp load data training source domain network and a 3hp load data fine-tuning target domain network are adopted, the GA optimizes the parameters of the source domain network, and the test accuracy of the 3hp load fault in the target domain is higher, which shows that the mining of a pre-training model obtained by GA optimization on certain load knowledge of the source domain is helpful for the diagnosis of the faults of rolling bearings of other loads in the target domain, and the test accuracy of the faults of the rolling bearings under different loads can be improved.

And on the premise of using the improved BLS network and GA to optimize the network node structure, performing comparative experimental analysis before and after model migration. In the experiment without the model migration method, the source domain in table 1 is used as a training set, and the target domain is used as a test set, for example: 2_3 sample set, 2hp load data is used as training set, 3hp load data is used as testing set; model migration experiment the source domain samples in table 1 were as follows 2: 1, dividing a source domain training set and a source domain testing set, wherein 20% of total samples in a target domain are used for fine tuning the weight of a network output layer of the target domain, the rest samples are target domain testing sets, and the comparison experiment results of the two methods are shown in table 4.

As can be seen from Table 4, the average test accuracy after the model migration method is adopted is higher by 8.3%, and the training time difference of the two models is not much. Therefore, the model migration method is more suitable for solving the rolling bearing state diagnosis problem under different loads.

TABLE 4 Experimental results before and after model migration

In order to further verify the advantages of the method in solving the rolling bearing state diagnosis problem under different loads, the embodiment of the invention is used for improving the width model transfer learning and performing CNN, LSTM, BPNN and DBN based on the same sample set and transfer task ^[20] Reference [21 ]]Reference [22 ]]And (6) carrying out comparison.

(1) Setting relevant parameters of CNN, LSTM and BPNN

The hyper-parameter settings of CNN, LSTM and BPNN all depend on multiple manual experiments. The Epochs size is 100, the batch _ size is 32, the ReLU activation function is used between each layer of neurons, sorted using Softmax and optimized using Adam. Each network configuration is set as follows.

The CNN comprises 3 one-dimensional convolution blocks, the 1 st convolution block comprises 32 convolution kernels with the length of 1 × 3, the step length is 1, and the maximum pooling with the length of 1 × 5 and the step length of 2 is adopted; the 2 nd convolution block contains 64 1 × 3 convolution kernels, the step size is 1, and the maximum pooling with 1 × 5 and 2 step sizes is adopted; the 3 rd convolution block contains 64 1 × 3 convolution kernels, with a step size of 1; finally, a fully connected layer of 64 neurons. LSTM is a 2-layer network structure, each layer containing 800 neurons. The BPNN is a 6-layer network structure, specifically 2048-1024-512-256-128-64 network structure, 2048 indicates that the 1 st layer contains 2048 neurons, and similarly, the other numbers indicate the number of neurons corresponding to the number of layers.

(2) Comparative analysis of different methods

CNN、LSTM、BPNN、DBN ^[20] Reference [21 ]]And document [22 ]]The results of comparative experiments, which are respectively labeled as scheme one to scheme six, and 7 different methods are shown in fig. 8.

As can be seen from fig. 8, compared with the deep learning methods such as the first to fourth schemes, the average test accuracy of the method of the present invention is respectively higher by 19.4%, 17.8%, 19.7% and 3.0%, and the training time of the method of the present invention is less than that of the first to fourth schemes. The fifth scheme adopts a deep migration learning method, the sixth scheme adopts a width characteristic migration learning method, and compared with the fifth scheme and the sixth scheme, the test accuracy rate of the fifth scheme and the sixth scheme is almost the same, but from the training time, the training time of the method is 35.1s, and the training time of the fifth scheme and the sixth scheme is 1899.0s and 115.0s respectively, so that the training time of the method is less than that of the fifth scheme and that of the sixth scheme. In conclusion, compared with a deep learning network, a deep migration learning network and a width feature migration learning network, the method provided by the invention has obvious advantages in solving the problem of rolling bearing state diagnosis under different loads, and can obtain good test accuracy within a short training time.

Further, to verify the generalization of the proposed improved breadth model migration learning network, the MFPT database provided by the American mechanical Fault prevention technology Association was used ^[23] Experiments were performed. The method comprises the steps of processing original vibration data of the bearing by adopting the data preprocessing mode by using normal data loaded under the working condition of 270 pounds and bearing inner ring fault and outer ring fault data loaded under the working conditions of 50 pounds, 100 pounds, 150 pounds and 200 pounds, and respectively forming sample sets A, B, C and D by using data containing normal state, outer ring fault and inner ring fault under different loads. The migration experiment task composition of the MFPT database under different loads is shown in Table 5, wherein a task I represents that an A sample set serves as a source domain, a C sample set serves as a target domain, a task II represents that an A sample set serves as a source domain, two sample sets of C and D serve as target domains, and the meaning of other migration tasks is analogized.

TABLE 5 migration experiment task composition (MFPT database)

And constructing a source domain sample set and a target domain sample set according to the table 5, and performing fine adjustment on the target domain network output layer by taking 20% of the number of the target domain samples as training samples. The experiments were performed 5 times for each of the different migration tasks, and the average was taken as the experimental result for each migration task, as shown in fig. 9.

As can be seen from fig. 9, the average test accuracy of the method of the present invention in 6 groups of migration tasks is close to 100.0%, and the standard deviation of the test accuracy is small, the fluctuation range is small, and the standard deviation of a part of sample sets is 0. The training time in fig. 9 consists of two parts: the pre-training model establishing time and the target domain network fine-tuning time are compared with the average training time of 35.1s of the CWRU data set, and the average training time of the MFPT database does not exceed 40 s. The MFPT database experiment further verifies that the method can establish a classification model in a short time, and meanwhile, the fault diagnosis of the rolling bearing under different loads is realized with good classification accuracy, and the method has good generalization performance. To further verify the effect of the various part improvements of the proposed method, the ablation experiments described above were performed on the MFPT database, and the results are shown in fig. 10.

Scheme a in fig. 10 shows that the BLS network is not improved, model migration is performed based on the BLS network, and a GA is used to optimize the source domain network node structure; scheme b shows that the node structure of the source domain network is not optimized, and model migration is carried out based on the improved BLS network; and a scheme c shows that model migration is not carried out, a fault diagnosis model based on the improved BLS network is established, and the GA is adopted to optimize the network node structure. The experiment was performed according to the source domain and target domain sample sets in table 5, and specifically, each migration task of the scheme b was performed according to 5 sets of random network parameters in table 3.

As can be seen from FIG. 10, although the training time of the method of the present invention is longer than that of other schemes, the test accuracy is the highest, and the improvement mode of each part proposed by the present invention is helpful to improve the fault diagnosis accuracy of the rolling bearing under different loads. After the BLS network is improved, the test accuracy is improved by 3.6%, after the GA optimizes the network node structure, the test accuracy is improved by 1.1%, after the model is migrated, the test accuracy is improved by 1.1%, and the effectiveness of the method in diagnosing the faults of the rolling bearing under different loads in the aspect of diagnosing the accuracy is verified.

In order to further verify that the method can solve the problem of time consumption of deep learning network diagnosis, the MFPT database is compared and analyzed with the experimental results of CNN, LSTM and BPNN, and the relevant parameters of CNN, LSTM and BPNN are set as follows.

The hyper-parameter settings of CNN, LSTM and BPNN all depend on multiple manual experiments. The Epochs size is 200, the batch _ size is 64, the ReLU activation function is used between each layer of neurons, sorted using Softmax and optimized using Adam.

The CNN contains 4 one-dimensional convolution blocks, the 1 st convolution block contains 16 convolution kernels of 1 × 64 with a step size of 16, and maximum pooling with 1 × 5 and a step size of 2 is adopted; the 2 nd convolution block contains 32 convolution kernels of 1 × 3, the step size is 1, and the maximum pooling of 1 × 2 and 2 is adopted; the 3 rd convolution block contains 64 1 × 3 convolution kernels, the step size is 1, and the maximum pooling with 1 × 2 and 2 step sizes is adopted; the 4 th convolution block contains 128 convolution kernels of 1 × 3 with step size 1, and maximum pooling with step size 2 of 1 × 2 is used; finally followed by a fully connected layer of 30 neurons. LSTM is a 3-layer network structure, each layer containing 400 neurons. The BPNN is a 7-layer network structure, specifically, the network structure is 4096-2048-1024-512-256-64-128-64, 4096 indicates that the layer 1 contains 4096 neurons, and similarly, the other numbers indicate the number of the neurons corresponding to the layer number.

The comparison results of the method and CNN, LSTM and BPNN are shown in table 6, the experiment results in table 6 are tested according to the source domain and target domain sample sets in table 5, and as can be seen from table 6, compared with other deep learning methods, the method has the advantages of shorter training time, higher classification accuracy, lower standard deviation of average test accuracy and more stable experiment results when the method is used for solving the problem of fault diagnosis of the rolling bearing under different loads. The MFPT database further proves that the method has lower training time cost compared with a deep learning method, and can realize quick classification.

TABLE 6 comparison of the method of the present invention with deep learning network test results (MFPT database)

The invention provides a method for improving a BLS network, which establishes an enhanced node window in a cyclic mode, transversely expands a BLS network enhanced layer, introduces a Maxout activation function into the enhanced layer and improves the feature extraction capability of the BLS network. Compared with the BLS network, the improved BLS network improves the average classification accuracy of rolling bearing states under different loads by 2.5% on two data sets in the text; the improved BLS network is combined with the transfer learning to form an improved width model transfer learning network, after the transfer method is adopted, the average classification accuracy of the improved network on two data sets in the text is improved by 4.7%, and the problem that the source domain data and the target domain data of the rolling bearing under different loads are large in distribution difference can be effectively solved; based on the characteristic of few network parameters of the improved breadth migration learning network, the GA is used for optimizing a network node structure, and the generalization performance of the network can be improved by verifying and introducing the GA algorithm from the visual angle; compared with other methods, the method has the advantages that the training time is shorter, the classification model is quickly established on two experimental data sets within the average training time of 32.6s, the efficiency of establishing the model is 1.7-187.5 times that of other methods, and the fault diagnosis of the rolling bearing under different loads is realized with the average classification accuracy of 98.9%.

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.

The documents cited in the present invention are as follows:

[1]SOHAIB M,KIM J M.Fault diagnosis of rotary machine bearings under inconsistent working conditions[J].IEEE Transactions on Instrumentation and Measurement,2019,69(6):3334-3347.doi:10.1109/TIM.2019.2933342.

[2] the method comprises the following steps of Sunyao bell, Zhangetao and downstairs, self-adaptive blind extraction of rolling bearing fault signals [ J ] based on an equivariant self-adaptive source separation algorithm, 2020,42(10):2471-2477.doi:10.11999/JEJT190722.

[3] Wangyizhi, conservation, Zhangun, and the like, a rolling bearing state identification method [ J ] based on an ensemble empirical mode decomposition sensitive eigenmode function selection algorithm, electronics and informatics, 2014,36(03):595-600.doi:10.3724/SP.J.1146.2013.00434.

[4]WANG J,MO Z,ZHANG H,et al.A deep learning method for bearing fault diagnosis based on time-frequency image[J].IEEE Access,2019,7:42373-42383.doi:10.1109/ACCESS.2019.2907131.

[5] Wencholine, Lufina, deep learning-based failure diagnosis methods review [ J ] Electron and informatics, 2020,42(01):234-248.doi:10.11999/JEIT190715.

[6]GAO S,XU L,ZHANG Y,et al.Rolling bearing fault diagnosis based on intelligent optimized self-adaptive deep belief network[J].Measurement Science and Technology,2020,31(5):055009.doi:10.1088/1361-6501/ab50f0.

[7]KONG X,FU Y,WANG Q,et al.A High Generalizable Feature Extraction Method Using Ensemble Learning and Deep Auto-Encoders for Operational Reliability Assessment of Bearings[J].Neural Processing Letters,2020,51(1):383-406.doi:10.1088/1361-6501/ab50f0.

[8]WANG H,LIU Z,PENG D,et al.Understanding and learning discriminant features based on multi attention 1DCNN for wheelset bearing fault diagnosis[J].IEEE Transactions on Industrial Informatics,2019,16(9):5735-5745.doi:10.1088/1361-6501/ab50f0.

[9]ZHANG R,TAO H Y,WU L F,et al.Transfer learning with neural networks for bearing fault diagnosis in changing working conditions[J].IEEE Access,2017,5:14347-14357.doi:10.1109/ACCESS.2017.2720965.

[10] Beam cooling, Hangwenlong, Von Weber, etc. adaptive knowledge migration based on the criteria of consistency of classification errors [ J ]. electronic and informatics, 2019,41(11):2736-2743.doi:10.11999/JEIT181054.

[11] The rolling bearing life prediction method based on unsupervised depth model migration [ J ] is an automatic chemical report, 2021:1-12.doi:10.16383/j.aas.c200890.

[12] TL-LSTM based bearing fault acoustic emission signal identification research [ J ] Instrument and Meter report, 2019,40(05):51-59.doi:10.19650/j.cnki.cjsi.J1904766.

[13]CHEN C L P,LIU Z.Broad learning system:An effective and efficient incremental learning system without the need for deep architecture[J].IEEE transactions on neural networks and learning systems,2017,29(1):10-24.doi:10.1109/TNNLS.2017.2716952.

[14]FENG J,YAO Y,LU S,et al.Domain knowledge-based deep-broad learning framework for fault diagnosis[J].IEEE Transactions on Industrial Electronics,2020,68(4):3454-3464.doi:10.1109/TIE.2020.2982085.

[15]FU Y,CAO H,CHEN X.Adaptive broad learning system for high-efficiency fault diagnosis of rotating machinery[J].IEEE Transactions on Instrumentation and Measurement,2021,70:1-11.doi:10.1109/TIM.2021.3085940.

[16]YAN R Q,SHEN F,SUN C,et al.Knowledge Transfer for Rotary Machine Fault Diagnosis[J].IEEE Sensors Journal,2019:1-1.doi:10.1109/JSEN.2019.2949057.

[17] Zhang laugyang, Cheng Jun, et al bearing Intelligent Fault diagnosis based on lifting depth migration Autoencoder [ J ] mechanical engineering newspaper 2020,56

(9):84-90.doi:10.3901/JME.2020.09.084.

[18]YOSINSKI J,CLUNE J,BENGIO Y,et al.How transferable are features in deep neural networks？[C].Proceedings of the 27th International Conference on Neural Information Processing Systems,Montreal,Canada,2014:3320-3328.

[19]LOPARO K A.Bearing data center[OL].https://csegroups.case.edu/ bearingdatacenter/home.2019.9.

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

[21]KANG S,CHEN W,WANG Y,et al.Method of state identification of rolling bearings based on deep domain adaptation under varying loads[J].IET Science Measurement&Technology,2020,14(3):303-313.doi:10.1049/iet-smt.2019.0043.

[22]WANG Y,WANG C,KANG S,et al.Network-combined broad learning and transfer learning:a new intelligent fault diagnosis method for rolling bearings[J].Measurement Science and Technology,2020,31(11):115013.doi:10.1088/1361-6501/ab8fee.

[23]ERIC B.Bearing Fault Dataset[OL].https://mfpt.org/fault-data- sets.2019.9.

Claims (7)

1. A method for improving the rapid classification of rolling bearing states under different loads of width transfer learning is characterized by comprising the following steps:

firstly, acquiring a rolling bearing time domain vibration signal under a load A as source domain data, and acquiring a rolling bearing time domain vibration signal under a load B as target domain data;

secondly, preprocessing the source domain data and the target domain data;

inputting the preprocessed source domain data into a source domain pre-training model based on an improved width learning system for training to obtain source domain pre-training model parameters; the source domain pre-training model parameters comprise the number N of characteristic nodes in each characteristic node window in the width learning system network ₁ Number of feature node windows N ₂ The number N of enhanced nodes in each enhanced node window ₃ Number of enhanced node windows N ₄ And a feature layer weight parameter W _e And the enhancement layer weight parameter W _h ；

Transferring the source domain pre-training model parameters to a target domain network based on an improved width learning system to serve as initial network parameters; inputting the preprocessed target domain data into a target domain network based on an improved width learning system for training to obtain a rolling bearing state classification model;

and fifthly, inputting the preprocessed rolling bearing data to be classified into the rolling bearing state classification model to obtain a rolling bearing state classification result.

2. The method for rapidly classifying rolling bearing states under different loads according to claim 1, wherein in the first step, the load A and the load B are different and both comprise one or more load types.

3. The method for rapidly classifying the rolling bearing states under different loads by improving the width migration learning is characterized in that the rolling bearing states comprise a normal state and a fault state, and the fault state comprises multiple fault damage degrees of an inner ring fault, an outer ring fault and a rolling body fault.

4. The method for rapidly classifying rolling bearing states under different loads according to claim 3, wherein the preprocessing in the second step comprises performing fast Fourier transform on the source domain data and the target domain data to obtain a frequency domain amplitude sequence.

5. The method for rapidly classifying rolling bearing states under different loads according to claim 4, wherein the source domain pre-training model network node structure based on the improved width learning system is adaptively determined by combining a Genetic Algorithm (GA) in the third step, and the specific steps include:

step three, initializing GA: setting population size, cross probability, mutation probability, chromosome coding mode, next generation population selection mode and iteration stop conditions of the algorithm;

step three, defining a search space: n in source domain pre-training model parameters ₁ 、N ₂ 、N ₃ And N ₄ Forming a search space of the GA;

step three, obtaining the fitness value of individuals in the GA population: passing through feature layer weight parameter W by source domain training sample _e And linear activation function to obtain mapping characteristics, and passing the weight parameter W of the enhancement layer according to the mapping characteristics _h Obtaining an enhanced feature by the Maxout activation function, using the mapping feature and the enhanced feature as the input of a source domain pre-training model output layer based on an improved width learning system, and calculating the weight W of the source domain pre-training model output layer by using a ridge regression algorithm to obtain the classification accuracy;

step four, comparing fitness values of individuals in the population, and obtaining a next generation population according to GA initialization parameters;

step three, judging whether an iteration stopping condition is met, if so, obtaining a source domain pre-training model, and if not, repeating the step three to the step three;

and step three, inputting the source domain test sample into the source domain pre-training model, and verifying the source domain pre-training model.

6. The method for rapidly classifying rolling bearing states under different loads according to claim 5 and improving width migration learning, wherein in the third step, the weight parameter W of the enhancement layer is used according to mapping characteristics _h And the specific process of obtaining the enhanced features by the Maxout activating function comprises the following steps:

assume that each enhanced node window contains N ₃ An enhancement map, then the kth enhancement feature is represented as:

in the formula: z represents a mapping characteristic;

representing a kth weight parameter of the enhancement layer;

representing the k-th bias parameter of the enhancement layer; 1, 2, N ₃ ；

Definition H _n For the enhancement feature in an enhancement node window, adopting a Maxout activation function to obtain the enhancement feature, then

Enhancement layer N ₄ The enhanced features of each enhanced node window are expressed as:

7. the method for rapidly classifying rolling bearing states under different loads according to claim 6, wherein in the fourth step, the weight of the output layer of the target domain network is finely adjusted in the training process based on the target domain network of the improved width transfer learning system by combining with a ridge regression method.

Images