Time-Varying Gaussian Encoder-Based Adaptive Sensor-Weighted Method for Turbofan Engine Remaining Useful Life Prediction

Remaining useful life (RUL) prediction, as an essential aspect of condition-based maintenance (CBM), has attracted substantial interest in industrial measurement. Recently, deep learning-based methods have achieved superior performance in turbofan engine RUL prediction. However, since the engine multisensor raw signals have noise and complex operation conditions, constructing representative features is challenging, which severely impacts the accuracy and generalization. Moreover, existing approaches tend to extract temporal dependencies and ignore identifying the contribution of different engine sensors. In this article, we focus on turbofan engine multisensor signals and propose a time-varying Gaussian encoder-based adaptive sensor-weighted (TGE-ASW) method to alleviate these problems. First, a time-varying Gaussian encoder (TGE) is built to enhance generalization and stabilize the training process of multisensor signals. Then, an adaptive sensor-weighted strategy is carried out to adaptive identify important sensors and weight signals conditioned on each sample. Finally, a convolutional neural network (CNN) is built to obtain high-level feature representation to predict the RUL. Experimental results on turbofan engine datasets demonstrate the superior performance over state-of-the-art methods and the effectiveness of processing and representing multisensor signals in industrial measurement.