Research On Identification Method Of Rolling Bearing Life Stage Based On Transfer Learning

In recent years,wind power generation equipment,aeroengine,high-grade CNC machine tools and other large-scale mechanical equipment are developing towards the direction of precision and efficiency,and the safe and reliable operation of the equipment has become very important.As the core part of mechanical equipment,rolling bearing is widely used in rotating machinery.Its performance will directly affect the health of the whole equipment.It is not only convenient for experts to make maintenance strategy in advance,but also to avoid the direct economic loss caused by equipment downtime and damage.So it is more and more important to identify the life stage of rolling bearing.With the advent of the era of big data,traditional machine learning methods based on data-driven life stage identification of rolling bearing emerge one after another,but the traditional machine learning methods are limited by two assumptions:(1)training data and test data should meet independent and same distribution;(2)the number of training data should be enough.The traditional machine learning algorithm has failed in different working conditions.Migration learning relaxes the condition that training data and test data are independently and equally distributed,and uses existing knowledge to transfer "knowledge" in different but related domain problems,so as to solve the learning problem of target domain with little or no labeled sample data.In this paper,the main research object is the rolling bearing.The main contents of this paper are as follows: 1(1)In view of the problem that the distribution difference between training samples and test samples under different working conditions leads to the failure of effective identification in the life stage of rolling bearing,an improved identification method for the life stage of rolling bearing with balanced distribution adaptation is proposed.Under different conditions,the number of available training samples is limited,and the training data and test data do not meet the conditions of independent and identical distribution,which reduces the generalization ability of traditional machine learning life recognition model,and even makes the model not applicable.Therefore,a multi sample training set in the source domain is established by uniform random sampling without repetition.Under the function of balance factor μ,the weights of edge distribution and condition distribution of each training set are adjusted dynamically,and the pseudo tags in the target domain are iterated and optimized to minimize the difference between the two domains.Finally,the final recognition results of the target domain samples are obtained by identifying the consistency of multiple prediction tag vectors of the test samples.(2)In order to solve the problem that a few samples can not be effectively identified in the life stage identification of rolling bearing due to the limited imbalance of samples under different working conditions,a method of life stage identification of rolling bearing based on multiple classifiers integrated weighted balanced distribution is proposed.In order to better identify a small number of class samples,give different initial weights to the training samples,fully train a small number of class samples,at the same time,establish the weight matrix in the way of class prior probability approximate conditional probability,iteratively optimize the pseudo label,update the weight matrix,and finally combine the multi-classifier set strategy to set the appropriate base classifier into a strong classifier,and achieve the limited imbalance of samples Rolling bearing life stage identification.(3)In order to solve the problem of self-adaptive adjustment of cross domain feature alignment and distribution difference in the identification of rolling bearing life stage under complex conditions of multi-mode superposition,a method of identification of rolling bearing life stage under different conditions with subspace embedding feature distribution alignment is proposed.Under complex working conditions,there must be information redundancy and mutual coupling in the high-dimensional life state eigenvector composed of time-frequency characteristic parameters in the life stage of rolling bearing,and the different distribution eigenvectors in different working conditions are mostly located in different sub-spaces,which is not conducive to classification and recognition.Therefore,the second-order co-variance matrix is used to align the source domain features of different spatial distribution with the target domain features in the target domain subspace.The weights of edge distribution and conditional distribution of the two domains are estimated quantitatively,and the difference between them is adapted.Finally,a kernel function is constructed to build a classifier to identify the life stage of rolling bearing in the framework of structural risk minimization.At the end of the paper,the work of this paper is summarized,and the next research is prospected.