A FPGA accelerator-based WDCNN for rolling bearing diagnosis

Abstract

In industrial practice, deploying a neural network for the real-time diagnosis of rolling bearings is a challenging task. To address this problem, this paper proposes a method for deploying Deep Convolutional Neural Networks with Wide Firstlayer Kernels (WDCNN) on a field programmable gate array (FPGA). Vibration signals can be used to predict the bearing condition directly by the accelerator. Using high-level synthesis tools to optimize the accelerator, the prediction time of a sample (2,048 points) has been reduced to milliseconds. Our research results indicate that the accelerator can be used to accurately and rapidly predict the bearing state.