A new approach of integrating industry prior knowledge for HAZOP interaction

Abstract

Accidents often occur in the petrochemical industry, which have a negative impact on society and the environment. Learning Process Safety Knowledge (PSK) from accident cases is essential to prevent accidents and improve safety level. Hazard and Operability Analysis (HAZOP) is a popular hazard risk analysis method. Its report contains large-scale PSK, which can provide safety analysis and decision support for the industry. Subject to the characteristics of PSK, existing researches mine them in the form of sequence labeling. However, there are two intractable problems that cause the PSK mined by the model to be inaccurate. (1) PSK in HAZOP is domain specific, which is rare or even absent in general-domain texts. (2) The entity boundaries are ambiguous. Most domain-specific entities for HAZOP lack boundary characters. Inaccurate security knowledge is not acceptable from the perspective of process safety engineering. To solve the problems, we present a PSK mining architecture with External Lexicon Prior knowledge called EDPMA, EDPMA is prior knowledge-based multi-task HAZOP knowledge mining model. Specifically, EDPMA consists of prior knowledge constructor and sequence labeling model. The prior knowledge constructor expresses prior knowledge in the form of word embedding by three steps. For the sequence annotation model, we improve its embedding and decoding layers. The former incorporated the word vectors generated by the prior knowledge constructor, and the latter added the task of entity boundary prediction. We conduct multiple evaluation experiments on HAZOP datasets. The experimental results show that the accuracy, recall and F1-score of the EDPMA model are 92.92%, 91.85% and 92.38% respectively, which is better than the existing research. Our study represents a meaningful attempt to introduce prior knowledge in HAZOP knowledge mining and makes an important contribution to intelligence the field of process safety.