Deep Learning Based Methods For Equipment Health Condition Assessment

The equipment presents a variety of morphological trends from the initial operation to the failure stage.The equipment basically has no obvious degradation trend in the early stage of the healthy condition,and shows a slower and stable degradation trend in the mid-term,but in the later stage,the closer the failure occurs,it will show an exponential rapid degradation.This leads to the imbalance of health and the multistage nature of degradation.In the health condition assessment and remaining useful life prediction,the deep learning model has the problems of class imbalance and large remaining useful life prediction errors near the time of failure.This thesis aims to solve the above two problems by building deep learning models separately.First,in the data preprocessing stage,the one-dimensional vibration signal of the bearing is converted into a two-dimensional tensor,so that convolutional neural networks can perform two-dimensional convolution on it,so as to better extract the depth features of the bearing,and realize the automatic extraction of the depth feature and end-to-end recognition and prediction.Aiming at the problem of class-imbalance,a health condition assessment model integrating auxiliary classifier generative adversarial network and convolutional neural networks is proposed.The auxiliary classifier generative adversarial network can learn the distribution and characteristics of samples well,and at the same time generate fitting samples consistent with the sample distribution to balance the minority classes.The convolutional neural networks mainly recognizes the health condition of the balanced data set.Meanwhile,in order to better achieve multi-classification and apply the results of health condition assessment to remaining useful life prediction,this thesis proposes to improve the auxiliary classifier generative adversarial network discriminator and the output layer of convolutional neural networks to the Softmax layer to realize the normalization of the various probability values of the recognition.This method was verified on the XJTU-XY data set and evaluated using utility indicators.The result reached 92.36%,which significantly improved the recognition accuracy of minority categories.Aiming at the multi-stage problem of degradation,a remaining useful life prediction model that integrates denoising convolutional autoencoder and multi-stage fully connected neural networks is proposed.denoising convolutional autoencoder is used to realize unsupervised learning,automatically extract bearing depth features,and establish different fully connected neural networks for each degradation stage.To better predict remaining useful life,this thesis further proposes to apply the result of the health condition assessment model as a weight to the multi-stage remaining useful life prediction.The verification result shows that the remaining useful life prediction error near the failure time is significantly reduced.