Fault Diagnosis of Rolling Bearings Using Denoising Multi-Channel Mixture of CNN and Mamba-Enhanced Adaptive Self-Attention LSTM

Recent advancements in deep learning have significantly improved fault diagnosis methods. However, challenges such as insufficient feature extraction, limited long-range dependency modeling, and environmental noise continue to hinder their effectiveness. This paper presents a novel mixture of multi-view convolutional (MOM-Conv) layers integrating the Mixture of Experts (MOE) mechanism. This design effectively captures and fuses both local and contextual information, thereby enhancing feature extraction and representation. This proposed approach aims to improve prediction accuracy under varying noise conditions, particularly in rolling ball bearing systems characterized by noisy signals. Additionally, we propose the Mamba-enhanced adaptive self-attention long short-term memory (MASA-LSTM) model, which effectively captures both global and local dependencies in ultra-long time series data. This model addresses the limitations of traditional models in extracting long-range dependencies from such signals. The architecture also integrates a multi-step temporal state fusion mechanism to optimize information flow and incorporates adaptive parameter tuning, thereby improving dynamic adaptability within the LSTM framework. To further mitigate the impact of noise, we transform vibration signals into denoised multi-channel representations, enhancing model stability in noisy environments. Experimental results show that our proposed model outperforms existing state-of-the-art approaches on both the Paderborn and Case Western Reserve University bearing datasets, demonstrating remarkable robustness and effectiveness across various noise levels.