A Novel Nonlinear Mapping based Compatibility Method for Zero-Shot Classification in Intelligent Fault Diagnosis

Abstract

With the development of machine learning and deep learning techniques, intelligent fault diagnosis (IFD) have been widely used in industrial control systems to monitor and detect the possible system faults and ensure the safety of system operations. Due to the lack of the faults samples, supervised machine learning and deep learning techniques cannot achieve great efficiency in intelligent fault diagnosis for control system. Additionally, although some unsupervised machine learning and deep learning techniques have been used in intelligent fault diagnosis, the low efficiency on types detection and identification of unseen-faults is achieved. To address these issues, in this paper a novel nonliear mapping based compatibility method (NMC) is proposed to effectively detect and identify the types of unseen faults with zero-shot classification in intelligent fault diagnosis. In particular, In the proposed NMC method, the fault samples of control system (known as seen-fault samples) are collected with the Sliding-Window strategy. Then, the projection of the collected seen-fault samples in the hidden feature space is achieved by nonliear mapping. Finally, the compatibility of the seen-fault samples and their projections are obtained by the compatibility function, and the unseen-faults of control system (i.e., zero-shot fault samples) are detected and classified into the category with maximum compatibility to the unseen-faults. Via the evaluations with TEP database, the results show that the proposed NMC method can achieve great accuracy on category identification of unseen-faults in control system in comparison with existng zero-shot classification models (i.e., DAP, ALE and Feng) and achieve great robustness to noise. In addition, with few-shot or zero-shot fault samples, the proposed NMC method can also achieve better efficiency in comparison with supervised classification model.