Multi-location fault detection with piezoelectric arrays and multi-task CNN learning

This article presents an innovative method for monitoring rotating instruments using piezoelectric array sensors and multi-task CNN (convolutional neural network) learning. The setup involves connecting piezoelectric patches, enabling spatial condition monitoring with a single voltage signal. The sensor array simultaneously tracks four bearing conditions and three gear conditions. However, a single-task CNN faces challenges, especially when trying to distinguish weak gear faults while simultaneously considering different bearing conditions. To address this, the article employs multi-task learning, simplifying the classification task by utilizing shared convolutional layers and two distinct fully connected layers dedicated to gear and bearing health states. This approach is also adaptable for extending fault detection to additional locations. Experiment demonstrates that multi-task learning achieved a 90% accuracy in identifying missing tooth defects in gears, outperforming the 78% accuracy of single-task learning. It confirms multi-task learning's effectiveness in detecting weak faults during multi-location defect monitoring using piezoelectric arrays.