An Interpretation Method for Fault Diagnosis in Rolling Bearings: Integrating Whale Optimization Algorithm and SqueezeNet Model

The bearing is a commonly encountered element in mechanical systems, and the adoption of a precise and efficient methodology for diagnosing bearing faults can greatly enhance the operational safety and stability of the system. This study introduces an innovative methodology to tackle the problem of diagnosing faults in bearing vibration signal features that are weak and pose challenges. The method proposed in this study entails the utilization of an enhanced SqueezeNet network, which exhibits the ability to interpret features for the purpose of detecting bearing faults in images. The initial aim is to perform preprocessing on the original vibration signal data acquired from the bearing. The process entails converting the initial signal in the time domain into a two-dimensional representation known as a time-frequency map. In order to replicate the noise conditions found in the real working environment, a deliberate introduction of a noise signal is performed during this transformation, with a signal-to-noise ratio varying between 10 and 20. Following a comprehensive comparative analysis of different deep learning models, the SqueezeNet model has been selected due to its exceptional antinoise capabilities and superior accuracy. The progressive grid method is utilized for the purpose of optimization, aiming to analyze the fluctuations in network accuracy. The objective of this study is to examine the relationship between network accuracy and changes in hyperparameters. The SqueezeNet model, which has undergone fine-tuning, is subsequently employed in combination with diverse swarm optimization algorithms. Based on a comparative analysis of the loss function iteration curves, the whale algorithm has been selected as the preferred optimization algorithm. The aforementioned algorithm is subsequently utilized within the framework of resolving the fault detection issue. The experiment also demonstrates the model's feature extraction process by employing the t-SNE visualization method. Furthermore, the utilization of class activation mapping is employed to analyze the time-frequency diagrams pertaining to the four identified faults observed in the bearings. The objective of this approach is to improve the comprehensibility of the model in its classification task and offer a graphical depiction of the model's attention towards different features. The performance of the method is evaluated by utilizing the bearing dataset obtained from Case Western Reserve University. The findings indicate that the approach exhibits a considerable degree of precision, approaching nearly 100%, in the identification of bearing abnormalities. This validation affirms the method's robust potential and feasibility for practical applications, emphasizing its superiority in fault diagnosis and adaptability to diverse application scenarios.