An approach for adaptive filtering with reinforcement learning for multi-sensor fusion in condition monitoring of gearboxes

Abstract

Condition monitoring of gearboxes is integral to maintaining floor safety, system stability, and inventory management. Capturing vibration response using sensors and subsequent response analysis is the standard procedure for gearbox fault detection. However, the sensors are susceptible to non-constant reliability due to the convolution of vibration responses from multiple sources, background noise interference, and transfer-path effect. The problem is multi-fold when ideal sensor attachment locations are unavailable due to spatial constraints of industrial floors. The response component reflective of the fault information must be enhanced for adequate fault severity estimations. The present study addresses this hurdle by proposing a multi-sensor framework with available sensor attachment locations for gearbox condition monitoring. Adaptive filtering is done in the framework with parameters optimised to enhance fault information. A proximal policy optimisation agent is trained with a reinforcement learning environment for parameter refinement. Further, fault severity estimation is achieved by a weighted fusion of spectral features reflective of the side-band excitation effect caused by gear fault. The proposed method is applied to datasets acquired from an in-house seeded fault test bed. The proposed method underscores superior performance compared to conventional single-sensor-based fault severity analysis and alternate fusion approaches.