Abstract:
Traditional data-driven fault diagnosis methods rely on a large number of labeled fault samples, but the situation where there are no target fault samples for training is common in chemical processes. To address this issue, the idea of zero-shot learning (ZSL) is introduced and a fault diagnosis method based on attribute description based multi-unit self-attention mechanism (AMSM) is proposed. Firstly, the semantic autoencoder is pre-trained to extract semantic information containing fault attributes from the samples. Secondly, the self-attention mechanism adaptively assigns the attention weight of each unit to learn the correlation between units and dynamically adjusts the feature matrix to make the extracted sample features continuously match with the potential representation containing semantic information for attribute learning. Finally, fault diagnosis is achieved through determining the minimum Euclidean distance between the predicted attribute vector and the fault attribute label. To demonstrate the effectiveness of AMSM model, Tennessee Eastman (TE) process is divided into four datasets and compared with other zero-shot fault diagnosis methods and supervised learning methods. The ablation experiment is used to illustrate the importance of semantic encoder and self-attention mechanism on the model. The zero-shot fault diagnosis experiment based on TE process shows that AMSM can realize fault diagnosis without target fault samples.