Research And Application Of Remaining Useful Life Prediction Of Components Based On Recurrent Neural Network Autoencoder

Components as an important part of system operation,the reasonable formulation of the maintenance time can improve operation efficiency and availability,reduce the operation and maintenance costs,as well as ensuring the safety of system performance.Therefore,the remaining useful life prediction technology for components has received extensive attention.In the field of railway transportation,realize the life prediction of heavy-duty freight car components,is the key to the transformation of the maintenance mode from "planned maintenance" to "condition maintenance".This dissertation focuses on the problem of life prediction for freight car wheels,and proposes a model for predicting the remaining useful life of components based on a recurrent neural network autoencoder.The model is carried out in two stages: the offline stage builds a health indicator library,and the online stage predicts the remaining useful life based on similarity.At the same time,the open datasets and the real wheel dataset are used to predict the remaining life of components,and the results are applied in the railway freight car condition repair system in trial operation.The main research contents of this dissertation are as follows:(1)Aiming at the two problems that appear in the current life prediction using the autoencoder model: the key features are not prominent and the decoding burden is heavy,a bidirectional long short-term memory network autoencoder(Bi LSTM-AS)prediction model is proposed.In the autoencoder of offline stage,a bidirectional long short-term memory network is used as the encoder-decoder structure,the attention mechanism and skip connection are introduced to highlight the key features affecting lifespan prediction.Then,based on the similarity matching principle,the prediction of the remaining useful life of the components in the online stage is carried out.And experimental verification is carried out on open datasets and real wheel dataset,the prediction accuracy has been improved.(2)Since the sensor monitoring data of components is affected by various operating conditions,the problem of incomplete extraction will occur during feature extraction.Therefore,the Bi LSTM-AS autoencoder is further optimized by introducing multi-scale fusion,to achieve the goal of better feature extraction under multiple operating conditions.In the experimental verification of the model on the open datasets and the real wheel dataset,the prediction results of the sub-dataset under multiple operating conditions have been significantly improved.(3)The proposed prediction model is applied to the railway freight car condition repair diagnosis and decision-making system.In the prediction module,according to the wheel condition monitoring data,the overhaul time and the remaining days of use are predicted,which can provide more scientific maintenance decision support for the condition repair of railway freight cars,and provide technical support for the maintenance work of operation and maintenance personnel.