Abstract
Dynamometer card-based fault diagnosis confronts critical challenges including imbalanced data distributions, noise contamination, and constrained generalization capacity. This paper proposes a novel diagnostic framework integrating similarity-guided data augmentation with wavelet-optimized lightweight neural networks to address these limitations. The methodology employs kinematic feature recombination for minority class expansion, resolving data imbalance through physically consistent sample generation. Concurrently, vertical load disparities, temporal load variations, and auxiliary operational parameters are fused into standardized dynamometer card representations, enhancing feature discriminability without architectural modifications. We integrate Discrete Wavelet Transform (DWT) layers into stride-2 inverted residual modules ofMobileNet-V2, strategically embedding spectral noise suppression during feature downsampling while preserving diagnostically critical low-frequency patterns. Experimental validation confirms the framework's superior accuracy and noise robustness, achieving significant performance improvements over conventional approaches while maintaining computational efficiency. The proposed solution establishes a deployable framework for industrial fault diagnosis, balancing diagnostic precision with operational practicality in oilfield monitoring applications.