Research On Restoration Method Of Missing Signal Based On Parallel Factorization Model

Affected by the complex environment,the frequently used parts in large mechanical equipment will have the problem of incomplete collected data,resulting in the lack of signal.In order to solve this situation,this project is supported by the National Natural Science Foundation of China(Grant No.52075236,51675258),the Project supported by the Natural Science Foundation of Jiangxi Province,China(Grant No.20212ACB202005),the Project supported by the equipment Pre-Research Foundation of China(Grant No.6142003190210)and the Project supported by the Aviation Science Foundation(Grant No.201946030001),using the advantages of parallel factor analysis method,and other methods Combined,several novel and effective methods are proposed to solve the problem of incomplete signal acquisition in mechanical equipment due to objective or subjective reasons,and good recovery results have been achieved.The main contents of this paper include:The first chapter discusses the research background and significance of the subject,and expounds the research status of missing data recovery methods at home and abroad.In view of the shortcomings of the existing missing data recovery methods,this paper leads to the advantages of tensor analysis.Parallel factorization is one of the main methods of tensor analysis.Therefore,the principle,development status and application of parallel factor in mechanical fault diagnosis are discussed.On this basis,the main research contents and innovations of this paper are given.Chapter 2: The principle of variational Bayesian is discussed,and a new method is proposed to recover the missing signal by combining the parallel factorization algorithm and the variational Bayesian method.The simulation results show that the proposed method has better recovery effect than the traditional low-rank tensor completion algorithm under the same missing ratio,and the error with the original signal is smaller.Finally,the recovery ability of the proposed method is demonstrated by the actual collected bearing signals.From the experimental results,the proposed algorithm has a more accurate recovery effect,and the spectral characteristics are closer to the original signal.Chapter 3: Although the variational Bayesian parallel factorization algorithm can successfully recover the missing signal,in the process of operation,the complexity of the algorithm structure leads to a slow convergence speed,and the denoising effect of this algorithm is not good.Therefore,this chapter proposes a new algorithm that takes advantage of the smoothness constraint and introduces the smoothness constraint into the parallel factorization model.The variational Bayesian parallel factorization algorithm fixes the number of ranks,which can lead to overfitting.The method proposed in this chapter uses a new strategy to determine the rank of the model.Instead of fixing the value of the tensor rank,the value of the rank is gradually increased from the rank of 1 until the parallel factor model and the observed data are fully consistent.The proposed method is compared with the variational Bayesian parallel factorization algorithm,and the simulation results show that the restored result of the new method is closer to the characteristics of the original signal than the variational Bayesian parallel factorization method,and the error is smaller.Finally,the method is applied to the recovery of bearing missing signal,and the missing signal is effectively recovered,which verifies the feasibility of the method.Chapter 4: Standard local methods,such as smoothing or patching,do not achieve good performance in the face of very high percentages of missingness.Therefore,a new constraint model is considered,that is,the standard local method and global data structure method are adopted at the same time.The decomposition model is constructed as the sum of the outer products of the smoothed component vectors of the function,which are representated at the same time determined by the linearity combination of the smoothed basis functions.This method can solve the problem of low recovery accuracy when the missing ratio is very high.Through simulation analysis and compared with the smoothness constraint method,the results show that the new method has faster recovery speed,smaller error,and more accurate recovery effect.Finally,the advantages of high performance and low computational cost of the new method are further verified by actual working conditions.Chapter 5: The loss of signal in mechanical fault equipment is complete random loss.Complete random loss is divided into discontinuous random loss and continuous random loss.The previous chapters of the paper consider the way of discontinuous random loss.Considering the continuous random loss of signal,it is more in line with the situation of data loss in the actual industrial process.At the same time,in order to further compare the recovery ability of the proposed three methods in the face of different missing forms,this chapter applies the three methods proposed in this paper to recover the continuous random missing signal,and finally comes to the conclusion that the function smoothness constraint algorithm has the best recovery effect on the signal.The effect of variable Bayesian method is very poor in this missing form,and it can not recover the signal in this missing form.It can be concluded that under different missing forms,the function smoothness constraint method has a wider scope of application and better recovery effect.Chapter 6: Summarizes the content of the previous chapters of this paper,and makes a further prospect of the parallel factor model in the field of missing signal recovery.