Online Knowledge Distillation-Based Multiscale Threshold Denoising Networks for Fault Diagnosis of Transmission Systems

Convolutional neural networks (CNNs) have developed rapidly in recent years, which has greatly promoted the advancement of intelligent fault diagnosis. Most currently available CNN-based diagnostic models are developed under the presumption that the acquired mechanical signals are invulnerable to noise. However, transmission systems usually operate under fluctuating conditions (e.g., variable speed and strong noise scenarios), making the fault-related pulse information in the mechanical signal easily swamped by noise. Therefore, it is challenging for these existing approaches to achieve satisfactory results in industrial scenarios. To deal with this problem, an online knowledge distillation-based multiscale threshold denoising network (OKD-MTDN) is developed in this research work. The main innovations and contributions of this research work include: 1) introducing a novel convolutional module, called the multiscale convolutional module (MCM), alongside a global attention module (GAM), for extracting a range of discriminative features generated from mechanical signals; 2) designing a multidilated threshold denoising module (MTDM) to expand the receptive field and filter out interference features; and 3) establishing an online knowledge distillation (OKD) algorithm to improve the generalization capability of OKD-MTDN. The HF-MS planetary gearbox dataset and the real-running high-speed rail dataset are utilized to verify the effectiveness of the proposed method. Experimental results show that the developed OKD-MTDN can achieve satisfactory results in various nonstationary scenarios.