Hydraulic Regime Embedded Diffusion Generation: A Data Engineering Approach to Alleviate Class Imbalance in Pump Station Fault Diagnosis

Abstract

In this study, the topic of defect diagnosis for pump stations based on hydraulic monitoring data is addressed using data augmentation diffusion models. After encoding the operating head, flow demand, pump combination, and fault label using a condition encoder, a denoising network produces physically consistent temporal residuals. Unbalanced monitoring windows are less susceptible to sensor drift, have superior minority class recall, and have a higher macro-F1 score. The three primary components of the content are confidence-aware augmented training, conditional denoising generation, and hydraulic regime embedding. Even after being filtered by spectral, correlation, and classifier consistency tests, the generated samples remain similar to the real pump behavior and are not utilized to replace the field data. The enhanced cavitation recall rate increased to 91.6% from 78.2%, the minority-class recall standard deviation dropped by five standard deviations, and the controlled experiment's macro-F1 score increased by roughly 93.1% over the 86.4% in the original work. Consequently, in situations where there are few failure observations but a large amount of routine monitoring data, diffusion augmentation is a practical way to enhance diagnosis.