Research On Key Technologies Of Bearing Life Prediction Based On Time Series Signal

With the increasing complexity,intelligence and refinement of modern mechanical equipment,the harsh and changeable working environment and the aging of equipment caused by long-time operation of equipment gradually reduce the remaining useful life(rul).It is particularly important to predict the remaining life according to the information and state of equipment or components in the actual situation.The overall process of residual life prediction algorithm design and visualization of reliability engineering and intelligent analysis results deserve further efforts and study.Bearings,which is the essential part of rotating equipment in the mechanical field,their failure will significantly reduce the availability and safety of the machine.At present,many research methods have emerged in the field of bearing life prediction,but because the life of rolling bearing is made by many reasons such as operating conditions,materials and processing technology,the service process of bearing has individual differences,which are also challenges in the research of remaining life prediction.At present,the remaining life prediction mainly focuses on point estimation,which needs further attention in measuring the uncertainty and reliability of algorithm model.In view of the above problems,this paper studies the remaining life prediction algorithm based on bearing vibration signal,the uncertainty measurement of algorithm model and the development of visual presentation of model analysis results.The main contents of this paper are as follows:Firstly,the data conditions and working conditions of the bearing signal are preliminarily analyzed.Using the characteristics of simple implementation and less calculation of wavelet threshold denoising,the bearing signal data obtained by the original sensor is denoised to weaken the influence of irregular noise on the next research of the data.In order to realize the compression of redundant data and increase the compactness and connection of time series data,the initial construction of input and output of time series data based on algorithm model is completed by means of data equidistant sampling and overlapping sliding window processing.Secondly,this paper analyzes the relationship and applicability between the current deep learning related methods and the remaining life prediction research of this subject,improves and optimizes them,and an attention coding and decoding algorithm—TCA-Seq2Seq,which is on the basis of the combination of time-domain characteristics and convolutional compression characteristics of vibration signals is proposed.Through the time-domain feature extraction and convolution feature compression of the original signal,the feature extraction efficiency of non-standard data such as bearing vibration signal is improved.By adding the attention mechanism to score the overall information provided by the encoder,the LSTM network integrating the attention score is used as the decoder to gradually predict the hidden layer eigenvalues representing the health indicator(HI),which improves the self-regulation ability of the model.Comparative experiments with other timing algorithms are carried out to test the rationality and superiority of this algorithm.Next,in the light of the issues that the present model situation of deep learning algorithm are mainly point estimation and lack of uncertainty measurement analysis of the model,a Bayesian reasoning method based on variational inference is proposed in this paper.Based on the data studied in this paper and the optimized model,the construction and optimization strategy of model uncertainty measurement are proposed.By means of cyclic sampling,the model after uncertainty measurement can iteratively predict and calculate the mean and variance of multiple rounds of results,and calculate the upper and lower boundaries of confidence intervals with certain confidence,so as to obtain a better measurement effect.At the same time of point estimation of prediction results,statistical analysis such as interval estimation is added to reduce the contingency of single prediction,so as to provide statistical basis for next decision-making suggestions in the field of remaining life prediction.Finally,on the basis of completing the training optimization of residual life prediction algorithm and the measurement of algorithm uncertainty,the visual display of the analysis results is analyzed and developed.This paper completes the construction of the database used to store the bearing vibration signal and the output information of the algorithm model.Through the C#script in Unity3D,the functions of external call of Python file,scene button binding,Unity3D data transmission and so on are developed,and the integrated process of visualization of intelligent analysis results of the model is realized.