Research On Methods Of Health Assessment And Fault Diagnosis Of Rotating Machinery Based On Deep Learning

Rotating machinery are widely used in production.If they fail,it is easy to cause safety accidents,so it is necessary to implement health management.However,at this stage,manual methods are difficult to efficiently and accurately perform health assessment and fault diagnosis on rotating machinery.Relying on the real engineering background,common problems in the implementation of health management for rotating machinery are summarized.Taking rotating machinery as the research object,this paper analyzes the effects from the perspectives of optimizing feature extraction,reducing the dependence on manual methods to improve efficiency,designing methods for automatically training health assessment models,and constructing transfer diagnosis models for variable workloads.The contents of health assessment and fault diagnosis have been studied.Aiming at the problems of traditional degradation assessment methods that rely on manual experience and complex signal processing methods,which can only be applied to specific scenarios,an adaptive health assessment model based on stacked denoising autoencoders and deep Qlearning network is constructed,and the training process is optimized.The training process is carried out in an end-to-end manner without human participation,which could the original data be directly and used for adaptive feature extraction and degradation status recognition,to obtain operating characteristic parameters representing the health state.Based on the Markov process,a health assessment model is constructed and health index is calculated.Then,the health evaluation curve is constructed to dynamically evaluate the degradation process by quantifying the deviation of health index from health status.Using rolling bearing dataset for experiments,the evaluation curve of health status assessment has good correlation,monotonicity and combination property,and better adaptability under different working conditions comparing with related methods.In fault diagnosis,a transfer model is built to solve the problem of significant decrease in diagnostic accuracy under machine load changing and external noise interference.Firstly,the maximum mean difference and distance variance are used comprehensively to measure difference between different distributions,and the joint distribution of source and target domain datasets is aligned on this basis.The entropy loss is used to improve the separability of features in shared subspace.A mathematical model with classification error,distribution difference,separability as optimization goals is established.Secondly,a dilated convolutional neural network with a zigzag expansion rate is designed for extracting multi-scale features and performing fault diagnosis.A two-stage training strategy is formulated.In the first stage,source domain dataset is used to train the diagnostic model.In the second stage,the model is transferred to target domain dataset.Finally,the average diagnosis accuracy of transfer diagnosis model for two bearing datasets can respectively reach 99.7% and 97.1% under different loads,and it has better performance than similar advanced methods.It can effectively deal with interference of different intensities of noise and maintain higher diagnosis accuracy.