Abstract:

To address the challenges of fault identification and decoupling in composite fault diagnosis of oilfield plunger-type water injection pumps, this study proposes a fault diagnosis algorithm based on adaptive time-frequency variational mode decomposition（VMD） and a multi-path fusion Transformer. First, an adaptive VMD algorithm incorporating time-frequency transformation preprocessing and adaptive mechanisms is designed to capture key fault features within composite fault sub-bands while dynamically adjusting the bandwidth and spectral center parameters. Then, a Transformer-based deep learning module with multi-path fusion capability is developed, which introduces both local path and global path attention mechanisms in a high-dimensional embedding space to enhance the ability to model frequency distribution discrepancies and improve composite fault feature recognition. Finally, experiments on a water injection pump dataset validate the proposed model. Under zero-sample training conditions for composite faults, the model achieves an average diagnostic accuracy improvement of 11.12% compared with mainstream composite fault diagnosis methods.

Key words:  , composite fault diagnosis, variational mode decomposition（VMD）, Transformer, zero-shot training