Abstract

In a multi-sensor system designed for fault diagnosis, the varying degrees of fault information present in each sensor's data are attributed to the diverse transmission pathways of the fault signal to individual sensors. Prior research has utilized feature analysis to merge multiple sensor’s data or simply concatenated data for input into deep learning model. However, existing approaches encounter difficulty in discerning the predominant fault information within each sensor's data, as the characteristics of all sensor data are abstracted during their propagation through neural network layers. To address this challenge, this study proposes a sensor fusion framework tailored for fault diagnosis in rotating machineries. Each sensor's data shares a common convolution kernel to maintain consistency based on the nature of defect signal in multi-sensor data. Additionally, a channel attention mechanism is implemented to weigh the significance of each sensor, rather than the channel itself. Experimental validation demonstrates that the proposed method appropriately integrates multi-sensor information in proportion to the fault information present in each sensor's data, surpassing the performance of existing multisensor fault diagnosis methods.