Multi-scale CNN for Multi-sensor Feature Fusion in Helical Gear Fault Detection

Abstract

Fault detection and diagnosis of helical gears under high speed and heavy load conditions are rarely researched comparing with spur gears under light load and low speed conditions. It is a fact that the working conditions of helical gears are very complicated, thus multiple sensors mounted on its different locations can provide complementary information on fault detection and diagnosis. On this basis, a multi-scale multi-sensor feature fusion convolutional neural network (MSMFCNN) is derived, and it operates information fusion on both data level and feature level. MSMFCNN contains three parts, including a conventional one-dimensional CNN part, a multi-scale multi-sensor feature fusion part, and an output part. To better understand this network, theoretical foundation of MSMFCNN is given. Moreover, in order to demonstrate effectiveness of the proposed method, experiments are carried out on a parallel shaft gearbox test rig on which multiple acceleration sensors are mounted for data acquisition. The experimental results show that MSMFCNN can fully utilize multi-sensor information and get a high accuracy on helical gear fault detection and can also converge faster than standard CNN.