A self-powered piezoelectric sensor with dual-branch CNN for timing belt health monitoring

This article introduces a novel self-powered piezoelectric sensor integrated with a convolutional neural network (CNN) for monitoring the health of timing belts during rotary operation. The sensor utilizes a two-point magnetic plucking mechanism, detecting torque-induced phase shifts between two magnets by measuring voltage variations. These phase shifts are also affected by wear on the timing belt, enabling the CNN to classify the belt's condition into three states: normal, transient, or fault, based on voltage data. However, relying solely on time-domain voltage signals makes it challenging to accurately assess the transient state. To address this, a dual-branch convolutional neural network model is proposed, designed to improve recognition accuracy even with limited data. Experimental results demonstrate that the dual-branch model significantly enhances feature extraction during the piezoelectric sensor's resonance, compared to traditional single-branch models. This leads to a more accurate distinction between transient and other states, achieving an average accuracy rate of 96%.