Investigation of Mechanical Fault Diagnosis Methods for On-Load Tap Changer Based on Texture Features of Curved-Surface Images Reconstructed in Phase Space

The main fault type in transformer on-load tap changer (OLTC) is mechanical failure. To address the issue of inadequate feature extraction for OLTC mechanical faults and increase diagnostic accuracy, this article proposes a fault diagnosis method for transformer OLTC based on the curved surfaces of vibration signals reconstructed in phase space. To improve feature extraction, the vibration signals from both normal and fault conditions are first decomposed using the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method, producing several intrinsic mode functions (IMFs). The IMFs undergo phase-space reconstruction to capture the nonlinear dynamic behavior of the signals. Texture features, including contrast, energy, homogeneity, and correlation, are then extracted from the XY–plane projection of the reconstructed phase-space surfaces using the gray-level co-occurrence matrix (GLCM). A Categorical boosting (CatBoost), model is employed for fault classification, achieving an overall accuracy of 95.8% on the test set and 96.9% in fivefold cross-validation, outperforming other classification models. Ablation experiments are conducted to analyze the contribution of each module, and comparative experiments with recent machine learning and deep learning methods further validate the effectiveness of the proposed approach. The results demonstrate that integrating vibration signal analysis with machine learning significantly enhances OLTC fault diagnosis, offering valuable reference data for equipment health monitoring and maintenance strategies.