Abstract

Accurate  condition  monitoring  of  rotating  machinery requires integrating multi-sensor data to capture fault-related information  distributed  across  sensing  locations.  While attention-based  deep  learning  models  can  assess  sensor importance,  their  lack  of  transparency  limits  industrial adoption.  This  study  proposes  an  interpretable  sensor importance-based  multi-sensor  integration  framework combining  a  CNN-inspired  kernel  sharing  strategy,  a Transformer-based feature extraction module for local and global feature extraction, and a channel attention mechanism for  dynamic  sensor  weighting.  Attention  weights  in  the Transformer-based feature extraction module were analyzed in  the  frequency  domain to  reveal  spectral  components influencing sensor importance evaluation. Validation on a pump testbed with various speeds conditions shows superior fault diagnosis accuracy, robustness to unseen conditions, and clear alignment between high-weight sensors and known fault frequencies, supporting trustworthy AI-driven condition monitoring in practice.