Integrating the maximum mean discrepancy metric with time–frequency enhanced convolutional neural networks for fault diagnosis

Abstract

In the operation of rotating machinery, the condition monitoring of key components such as bearings and gears is very important. However, due to the complexity of the real environment and the diversity of equipment operating conditions, the collected mechanical vibration signals of bearings and gears often contain irrelevant noise, which brings challenges to signal processing. Traditional convolutional neural networks (CNNs), lacking denoising structures, have limited effectiveness in processing interfered vibration signals. In addressing these challenges, this paper introduces a novel approach, namely the Gramian-based CNN model and maximum mean square discrepancy (MMSD). For the model, we first propose a simple yet effective denoising strategy, namely Gramian noise reduction (GNR) based on the self-similarity of vibration signal cycles. Second, considering the limitations of traditional CNNs in handling interfered vibration signals, we develop a novel end-to-end GNR-based CNN model - Gramian Time–Frequency Enhanced Network (GTFE-Net), applied to bearing fault diagnosis. GTFE-Net consists of three parallel branches, each processing the original signal, the GNR-denosed signal, and spectral information. By integrating GNR into GTFE-Net, the network focuses more on feature extraction. Additionally, we introduce MMSD as a metric that comprehensively reflects differences in sample mean and variance information in the reproducing kernel Hilbert space, thereby enhancing domain confusion. The experimental results indicate that this method achieved a diagnostic accuracy of 77.19% for various transmission tasks in noisy environments. This finding provides reliable support for bearing fault diagnosis, demonstrating the method’s effectiveness and potential for application in complex environments.