Multi-Label Zero-Shot Learning for Industrial Fault Diagnosis
Jian Li, Haojin Tang, Dong Tang, Zhao Yang
Abstract
The scarcity of labeled fault samples in long-term running industrial systems has caused significant challenges for traditional data-driven fault diagnosis. To address this issue, a multi-label zero-shot learning (ML-ZSL) method for industrial fault diagnosis is proposed in this paper. Firstly, a multi-scale features extraction network based on one-dimensional convolutional neural network (1D-CNN) and convolutional block attention module (CBAM) is designed for exploiting the effective information from raw fault samples. Secondly, a multi-label learning approach based on binary cross-entropy with logits loss (BCEWithLogitsLoss) function is proposed to update the parameters of network, which is considered as a solution to the imbalanced problem between positive and negative samples in fault attributes. Finally, a semantic similarity metric space based on the Mahalanobis distance is constructed to better classify the unseen fault samples during the testing phase. Experimental results conducted on Tennessee-Eastman process (TEP) dataset demonstrate that our proposed ML-ZSL can achieve higher accuracy than the representative zero-shot learning methods in industrial fault diagnosis.