A novel temporal-spatial feature enhancement approach for cross-condition mechanical fault diagnosis with imbalanced data

In order to tackle the problems of precisely identifying fault modes and accomplishing cross-condition diagnosis using limited data under the background of imbalanced rolling bearing data, this paper comprehensively integrates transfer learning techniques and feature enhancement strategies. We present two feature enhancement modules: the neural ordinary differential equations-based temporal feature enhancement module (NODE-TFEM) and the path signature-based spatial feature enhancement module (PS-SFEM). Firstly, the residual-enhanced CNN (ResCNN) architecture is used to extract local features from vibration signals. Then the NODE-TFEM utilizes the local features as initial values, and then models the feature space as the systems of ordinary differential equations. By solving the systems, this module can capture the temporal evolution of ResCNN features. On the other hand, the PS-SFEM regards each vibration signal as a mathematical path. Through calculating its iterative integral, the module acquires a spatial feature representation of the entire vibration signal. This feature has a clear geometric meaning in mathematical terms, further enhancing the interpretability of the model. The validity of the proposed methodology has been meticulously verified on the Paderborn University bearing dataset, the Jiangnan University bearing dataset and the laboratory-acquired bearing dataset. These experiments comprehensively probed into the performance of the methodology, manifesting its potential to surmount the challenges in the domain of rolling-bearing fault diagnosis and furnishing robust evidence for its practical implementation in real-world mechanical maintenance scenarios.