Research on misalignment fault identification methods for coupling based on dynamic models and transfer component analysis

Abstract This study proposes an innovative diagnostic approach for coupling misalignment faults by integrating dynamic modeling with transfer learning. It tackles the challenge of limited fault samples in real-world diagnostic scenarios, a common obstacle to developing robust diagnostic models. A dynamic model of coupling misalignment is first established, generating diverse misalignment fault signals that are experimentally validated. Time- and frequency-domain features are then extracted from both simulated and experimental vibration signals. The Laplace Score (LS) method is utilized to select sensitive features, with simulated signal features serving as training samples and experimental signal features as testing samples. Transfer Component Analysis (TCA) is employed to reduce distributional differences between the two datasets. Subsequently, the Cuckoo Search (CS) algorithm optimizes a Support Vector Machine (SVM), yielding a robust model that effectively detects coupling misalignment faults. Experimental validation achieves 98.43% fault recognition accuracy, surpassing the baseline TCA-CS-SVM approach (90.31% without feature selection), thereby providing a novel solution for detecting coupling misalignment faults.