GEARBOX FAULT DETECTION VIA PHYSICS-INFORMED PARALLEL DEEP LEARNING MODEL ARCHITECTURE

Ensuring the safety and reliability of mechanical system significantly hinges on accurate fault diagnosis of critical components comprising the system. Automating the diagnosis process can enhance efficiency and consistency of accuracy in the diagnosis stage. Deep learning methods tend to be playing increasingly indispensable roles in the establishment of the automated fault detection framework; however, they consistently require abundant data resources to leverage the uninterpretable features for fault detection tasks, which is posing major hindrances especially for situations where data collection for fault scenarios is expensive. Moreover, the physically unexplainable feature extraction process renders the nongenerative data augmentation preferrable by feeding the raw information repetitively without being informed of physics integrated in the data. The opaqueness of model greatly hinders the efficient exploitation of the raw data. To address these challenges, it is proposed in this research that the physicsinformed parallel model architecture can accommodate miscellaneous physically meaningful features in addition to the raw information. Specifically, the scalogram extracted by continuous wavelet transform is designated to be input together with raw data to the binary branches respectively in the parallel deep learning model architecture. By incorporating the extra interpretable features and information, our approach can completely leverage the data of restricted availability and render the model training physics-guided and understandable. The proposed algorithm is implemented on gear fault experimental data acquired from laboratory testbed to exhibit the physics- guided training process and the outstanding performances on varying size of data.