Multisensor feature selector for fault diagnosis in industrial processes

To address the oversight of data feature properties and interactions in the fault diagnosis of multiple different sensors, we introduce a novel fault diagnosis method leveraging a multi-sensor feature selection mechanism. This method employs multiple parallel multi-view feature extraction modules to distill essential fault features from the data collected by different types of sensors and subsequently aligns these features within a unified information metric space. Within this space, a sensor source selector scrutinizes the features, pinpointing those with significant fault-relevant distinctions and calculating a weight matrix for the multi-view features based on these insights. The refined feature set is then forwarded to the fault pattern recognizer, which leverages these optimized features for the precise diagnosis of faults in industrial equipment. Experimental evidence from applying this method to a flash furnace system showcases a fault detection rate exceeding 99 %, markedly improving fault classification accuracy.