A safety evaluation method based on gating model & generalized zero-shot learning for industrial process

Safety evaluation refers to the comprehensive safety inspection and evaluation of system equipment. The aim is to identify potential security vulnerabilities and risks, and to provide early warning based on real-time unhealthy performance. However, most of the existing evaluation algorithms are based on seen performance data and lack information about unseen performance. This leads to poor generalization to unseen performance and prevent them from being accurately assessed. Therefore, for the case of online data containing both seen and unseen performance, this paper proposes a safety evaluation method based on gating model with generalized zero-shot learning. The feature extractor is constructed using historical data of seen performance. The attribute learner is trained using semantic vector information and driven to migrate the attribute knowledge to unseen performance. The domain shift is mitigated by the gating model to judge whether the samples belong to seen or unseen performance to improve the generalization ability of the model. The simulation results show that the evaluation accuracy of unseen performance improved from less than 5% to over 70%. The experimental results show an accuracy improvement of over 65% for unseen performance. This study centers on enhancing system security by evaluating the unseen performance that may have a significant impact or hazard on the system. Generalized zero-shot learning is creatively introduced into the safety evaluation to address the challenge of accurately assessing unseen performance. The proposed gating model effectively alleviates domain shift. This study provides a new perspective for safety evaluation to ensure security.