Domain correction for hydraulic internal pump leakage detection considering multiclass aberrant flow data

Abstract

Harsh working environment not only threatens the health of the hydraulic system but also the condition monitoring system. The latter problem will make data aberrant and disable lots of data-based fault detection methods. Inspired by the Fail-Safe principle, the multiclass aberrant data problem is investigated in this study from the perspective of transfer learning. Firstly, the Domain Correction, a variant of Domain Adaptation, is defined theoretically. Then, an indirect Domain Correction framework is proposed and applied to internal pump leakage detection with aberrant flow data. The Teacher-Student structure is the basis. Extra Correction Module is designed to better correct aberrant representation into normal. Layer-wise training and the Noisy Tune are performed to mitigate overfitting. The Self Correction Attention mechanism is presented to help the model focus on the well-measured parts of samples. The proposed method can improve the model's accuracy on the aberrant dataset from 47.1% to 95.0%, meanwhile, the accuracy on the well-measured dataset is guaranteed at 99.2%.