FGTC Based Remaining Useful Life Prediction with Incomplete Bearing Dataset

Prognostics and health managemen is a vital technology for implementing the health management and predictive maintenance of rotating machinery. However, due to the lack of complete running data set of bearing, the practical application of life prediction is greatly restricted. Therefore, this paper proposes a feature generated time convolutional (FGTC) network framework, including coding layer time-convolutional network (TCN) and prediction layer bidirectional long short term memory attention mechanism (BiLSTM-AM), and decoding layer TCN network. Firstly, physical domain features are extracted from industrial equipment component sensor signals, and light gradient boosting machine is used to select more important features for fusion. The feature data is trend-decomposed using the complete ensemble empirical mode decomposition with adaptive noise to construct a performance degradation feature set. Degradation states and points are determined through reconstruction features, and degradation modes are classified using degradation speed. Secondly, different degradation mode feature data are input to the TCN for encoding and the BiLSTM-AM is trained on the hidden output. Finally, decode the prediction result to realize life prediction and feature data generation. The study uses bearings as an example and compares the proposed FGTC network with commonly used deep learning methods, with the experimental results confirming its superiority and feasibility.