Self-Attention Parallel Fusion Network for Wind Turbine Gearboxes Fault Diagnosis

The gearbox represents a critical component of the wind turbine and its proper functioning is fundamental to ensure high operational reliability. Efficient extraction of fault information from monitoring signals constitutes a crucial factor for successful fault diagnosis. This article presents a practical and effective self-attention parallel fusion network (SAPFN) for intelligent fault diagnosis of gearboxes. The proposed method leverages the Gramian angular field (GAF) matrix and continuous wavelet transform (CWT) to improve the accuracy of the diagnosis. Specifically, the GAF matrix is utilized to convert the original vibration signal from the Cartesian coordinate system to the polar coordinate system using trigonometric functions, which generates a matrix capturing the temporal correlation between different time points. The CWT technique is then employed to convert the vibration signals into time–frequency signals for further feature extraction. The SAPFN branch network is designed to extract fault features from both the GAF matrix and CWT signals, enabling a more comprehensive and accurate fault diagnosis. The SAPFN network employs a self-attention fusion layer that calculates the correlation matrix among features, enabling optimal allocation of weights to fault features and redundant features, thus enhancing the extraction of fault information. Moreover, the proposed SAPFN model is compared against several state-of-the-art methods recently published in the literature for benchmarking purposes. Finally, to validate the effectiveness of the proposed approach, comprehensive experimental tests are conducted under time-varying load and speed conditions.