An Entropy- and Attention-Based Feature Extraction and Selection Network for Multi-Target Coupling Scenarios

Sensor data analysis of machines' operating status plays essential roles in production and maintenance planning. Feature extraction and selection are critical procedures in sensor data analysis. However, existing methods for feature extraction and selection ignore the scenarios where multiple targets exist, and the obtained features cannot be fine-tuned to the specific target accordingly. A typical example of multiple targets is the remaining useful life (RUL) prediction under multiple failure modes of degraded machines. To address this issue, this paper develops an entropy- and attention-based feature extraction and selection network for multi-target coupling scenarios. Specifically, we proposed a multivariate symmetrical uncertainty-based filtering (MSUBF) algorithm to eliminate redundancy of originally extracted features by Tsfresh. Then, an attention-based joint learning network is proposed to automatically adapt features to different coupling targets. In a case study of degraded aircraft turbine engines with two coupling targets, i.e., failure mode classification and RUL prediction, the proposed method outperforms the benchmark methods that do not consider the interaction of coupling targets.