Remaining Useful Life Prediction of Parts and Components based on BiLSTM Autoencoder with Attention Mechanism and Skip Connection

Abstract

Remaining useful life (RUL) prediction is one of the main tasks of prognostics and health management (PHM). This paper proposes a novel method for RUL prediction, Attention_mechanism and Skip_connection are introduced into bidirectional long short-term memory autoencoder (BiLSTM-AS). In the previous RUL prediction scheme based on BiRNN autoencoder, during the encoding process, in the process of embedding vector generation, the multi-dimensional sensor data of each timestep contributes equally, the embedding vector is only used in the first step of decoding. To overcome the above shortcomings, the BiLSTM-AS model is proposed. This model introduces attention mechanism based on the BiLSTM autoencoder, assigns weights to each timestep information and highlights the key timestep information contribution; to relieve the decoding burden of embedding vector, skip connection is introduced in the decoding process. This method evaluated on the datasets of C-MAPSS and compared with the latest prediction methods, the comparison shows that our results are competitive. Subsequently, experiments are carried out on the monitoring data of train wheelset experimental datasets, and the RUL prediction results are obtained, indicating that the model has good generalization capability.