Generalized zero-shot fault diagnosis based on fault similarity for hydrometallurgical process

Effective fault diagnosis techniques are important to ensure the normal operation of hydrometallurgical process. Generalized zero-shot fault diagnosis (GZSFD) technology can effectively diagnosis both seen and unseen faults in actual production, so as to improve production efficiency and reduce losses. However, traditional GZSFD methods have the domain shift problem (DSP) and over-rely on deep knowledge to establish the relationship between semantics/attributes and faults. This deep knowledge is difficult to acquire without sufficient understanding of the production process. In this study, a GZSFD method based on fault similarity (GZSFDFS) for hydrometallurgical process is proposed to overcome the limitations of traditional GZSFD. The core of GZSFDFS method is to build a fault similarity matrix (FSM) between seen and unseen faults using the superficial knowledge of whether state and operational variables will change as a fault occurs. Firstly, in order to extract representative features from the original data, a proper supervised learning method is used to establish feature extraction model, and the extracted features are used to establish a recognition model. Next, the prediction results for the test samples with respect to each seen fault can be obtained by utilizing the fault recognition model, and the appropriate threshold is selected to distinguish the unseen faults from the seen faults. For the unseen faults, the predicted results for the test samples with respect to each unseen fault are constructed based on the FSM. Then, reasonable discriminant rules are designed to determine the fault classes of the test samples. Finally, based on numerical examples and hydrometallurgical processes, the effectiveness and superiority of the proposed method are verified.