Efficient Method for Gearbox Fault Diagnosis Under Variable Speed Conditions Using WVD Based Time Frequency Information of Vibration Signals and Deep Learning

The gearbox fault diagnosis under variable speed remains one of key challenges in industrial domain due to the nonlinear and time-varying nature of vibration measurements. In this paper, a highly-efficient hybrid method that combines the Wigner-Ville Distribution (WVD) based time-frequency representation and deep learning model to improve the accuracy of fault detection is suggested. WVD converts the signals of vibration to rich time-frequency representations in which complex fault-based patterns are recorded. We use an Attention-Based Convolutional Neural Network (Attention-CNN) to be able to extract useful features out of these representations, with _attention_ giving priority to the most pertinent areas. The features so extracted are further fed to an effective Random Forest (RF) model, and this effectiveness minimizes overfitting and maximizes the generalizability even across the changing speed settings. A comparative analysis of the proposed model is performed with vibration data of a publicly available gearbox fault dataset, which is outperforming the traditional techniques. The results show that the method is expecting 99.73 per cent accuracy, higher than the available methods of machine and deep learning. In addition to the above, a flexible measurement procedure is proposed to examine the fault diagnosis reliability under various loads to verify robustness in a real-life condition. The results support the fact that the suggested approach can enhance the gearbox condition monitoring and predictive maintenance approaches.