Fault diagnosis of rotating machinery using a signal processing technique and lightweight model based on mechanical structural characteristics

Intelligent Electromechanical Actuator (EMA) is becoming a key direction for future development, which increases the demands on the performance and complexity of fault diagnosis methods. Rotating machinery, the key rotating load-bearing components in EMA, directly impacts operational status of the entire machinery equipment. Operating in harsh environments and prone to noise, it requires fault diagnosis methods with robust noise resistance. Most existing fault diagnosis methods focus on Gaussian white noise, overlooking the impact on rotating machinery in real industrial scenarios. Additionally, they emphasize signal processing and neural networks, neglecting mechanical structure characteristics and lacking interpretability. To address these limitations, an interpretable fault diagnosis method of rotating machinery based on mechanical structural characteristics, called rolling element separation and channel-sharing, spatial-unidirectional convolutional network (RES-CSSUC), is proposed. This method integrates the rolling element separation (RES) signal processing technique and a lightweight model called CSSUCNet, which is mainly composed of channel-sharing, spatial-unidirectional convolutional layers. To assess the ability of the proposed method extracting features in noise conditions, Gaussian white noise and impact noise are respectively injected into the data. This method is demonstrated on Paderborn University bearing dataset and Planetary Roller Screw Mechanism (PRSM) dataset. Results show that RES-CSSUC achieves high diagnostic accuracy both in noise-free and noise conditions with fewer parameters, fewer FLOPs, and less memory usage, demonstrating strong robustness. It can reduce computational and storage costs, making it highly suitable for integration with mechanical equipment and facilitating the development of intelligent EMA.