Enhanced Gear Fault Net Using Squeeze Group and Depthwise Convolutions with Grad-CAM

Abstract Gearboxes provide a stable transmission ratio and are essential in wind turbines and circulation systems. Reducing downtime and maintenance costs prevents costly equipment failures and mitigates safety risks. Early gear fault identification is essential to achieving these goals. Conventional methods, such as wavelet transforms and Fast Fourier Transform (FFT), exhibit limitations in frequency resolution due to fixed time-frequency trade-offs, particularly under non-stationary conditions. In this study, the Constant Q Transform, combined with deep learning, outperformed traditional signal processing techniques in analyzing statistical features extracted from both the time domain and the frequency domain of vibration signals in gearboxes. Incorporating attention mechanisms, Inception blocks, grouped convolutions, Receptive Field Attention, and Squeeze- Aggregated Excitation in the proposed deep learning model enhances feature extraction and enables accurate classification. The proposed method achieves an accuracy of 96.03%, a precision of 97.16%, and a specificity of 97.28% in distinguishing between healthy and broken tooth gearbox conditions. The experimental results demonstrate that the proposed method outperforms existing state-of-the-art techniques. Additionally, Grad-CAM analysis is applied to validate the model’s decision-making process effectively, ensuring that the model focuses on relevant regions of the input data and enhancing the interpretability and reliability of the results.