Fault Diagnosis Of Mechanical Parts Based On Improved Autoencoder Models

In the context of intelligent manufacturing,it is becoming more and more important to realize accurate and timely fault diagnosis of key parts in the complex mechanical systems.But in the real industrial environment,the collected real data usually has the characteristics of large data,more noise,and few fault samples,which undoubtedly increases the difficulty of fault diagnosis.When traditional deep learning models face complex data,it is difficult to obtain ideal diagnosis results due to insufficient feature extraction capabilities and inability to deal with unbalanced small samples.Therefore,this paper designs two new deep autoencoder models for these problems.And they are used in the fault diagnosis of mechanical parts to improve the accuracy.Firstly,in order to solve the problem of insufficient feature extraction ability,this paper proposes a multi-channel transmission stacked pruning sparse denoising autoencoder.The new model proposes fully connected network architecture,non-optimal units pruning,multi-channel feature fusion and enhanced sparse expression,etc.,to change the traditional single-channel feature transmission method.At the same time,it further restricts the widening of the network and strengthens the sharing of characteristic information between layers.This model has achieved 99.94% classification accuracy in the fault diagnosis experiment of the bearing data set published by Case Western Reserve University,which verifies that the new model has stronger feature extraction and classification capabilities.Secondly,in order to deal with the problem of unbalanced small samples,this paper proposes a stacked reconstruction pruning autoencoder while preserving the advantages of multiple channels.The model is improved on two levels: At the data level,the improved over-sampling and window segmentation reconstruction algorithm are proposed to perform balanced expansion between classes and reconstruction expansion within each class;At the algorithm level,the shallow feature channel is proposed to ensure the full use of shallow information;then the support vector machine optimized by particle swarm optimization for small samples is used for top-level fine-tuning and classification.Afterwards,the bearing data set published by Paderborn University is used to verify that the model has a strong fault diagnosis and classification ability for unbalanced small sample data in different work scenarios.Then,in order to verify the fault diagnosising effectiveness of the two models in real industrial scenarios,this paper uses the real raw data collected in the case of the airconditioning expansion tube quality fault diagnosis project to design the fault diagnosis experiment of the large and small samples respectively.And the two models proposed in this paper have achieved 98.75% and 98.13% classification accuracy in two experiments respectively,which are better than common fault diagnosis models,which further demonstrates the practicality and superiority of the models.Finally,this paper is summarized,and the valuable research directions are prospected.