The Research On Mechanical Component Remaining Useful Life Prediction Method Using Deep Learning

Fatigue often occurs far below the material strength limit and yield limit.Fatigue failure is usually affected by factors such as stress concentration and surface roughness.Therefore,for predicting fatigue remaining useful life of mechanical parts,establishing accurate models to predict the remaining useful life(RUL)of mechanical components will bring great challenges.Deep learning provides powerful non-linear processing ability and shows great advantages in fatigue life prediction.This thesis will focus on the investigation and development of deep learning based methods to process acoustic emission(AE)signals and predict the fatigue RUL of the mechanical components.The main content of the thesis includes:(1)This thesis proposes a data preprocessing method for acoustic emission signals,which could make the deep learning model converge more easily and quickly with limited model capacity.In this thesis,a time window with variable-step is used as the input of the deep learning model.The model training data is augmented by moving the time window on acoustic emission data to solve the over-fitting problem.In the data augmentation process,a small expansion factor and a large expansion factor are used for labels with large amount of AE data.Therefore,the problem of learning from imbalanced data can be solved by this method to some extent.(2)This thesis uses a one-dimensional convolution kernel to process time-series AE signals.In general,more hidden layers are more difficult to train.This is normally affected by the activation function,the back propagation algorithm,and the initialization of model parameters.In this thesis,a residual block is used as the basic unit of the deep learning model.The depth and capacity of the deep learning model are increased by stacking the residual blocks.At the same time,the residual block is used to improve the model.In addition,the propagation algorithm of the model training process is improved by the residual block.The training difficulty caused by gradient vanishing is greatly reduced,and the robust performance of the model can still be maintained when the model depth is high.Faced with the enormous hardware requirements brought by data augmentation technology in the training process,this thesis uses the online learning method to alleviate the training difficulties caused by the shortage of computer memory.(3)This thesis introduces a deep learning model hyperparameter tuning method.Batch normalization is used to avoid the gradient explosion and gradients vanishing problem.The model hyperparameter values are determined by a random grid search method.This thesis uses max pooling to down-sample the data of each layer to reduce the data dimension.An advanced rectified linear unit function is used as activation function to mitigate the training difficulties caused by the gradient vanishing.By using the advanced adaptive momentum estimation optimization algorithm together with the residual block the model converge to the local minimum and saddle points during training can be avoided.Using a random grid search and a visualization module together to achieve code automatically find the network hyperparameter,and through a series of data preprocessing method to enhance the network ability of feature extraction,RUL prediction,through k-fold cross validation to evaluate the model.(4)To validate the deep learning based fatigue RUL prediction method developed,fatigue tests on metal specimens are performed and AE signals are collected.An acoustic emission data acquisition platform is constructed to collect the AE data during the fatigue tests.The accuracy and usability of the developed deep learning based method are validated by the collected AE data through k-fold cross validation.This thesis explains the important steps such as preprocessing of acoustic emission data,selection of a deep network structure,training process of the network,updating parameters,and the evaluation of network performance.According to the final prediction results,the accuracy of the method developed in this paper meets the requirements for the prediction of the fatigue remaining useful life of mechanical components.