Leak Identification And Risk Assessment Of Chemical Process Based On Long And Short-term Memory Network

Due to dangerous raw materials and strict production conditions,safe production is a prerequisite for the development of the chemical industry.Leakage is one of the most common accidents in chemical plants.If environmental or human-induced leakage cannot be detected and identified in time,it will affect the normal operation of the process and even cause fire,explosion,and casualties.Among the existing leak identification methods,hardware methods are often disturbed by environmental factors and cannot be monitored in real time.Software methods are limited by the complexity of chemical process data,which leads to difficult modeling and high false alarm rates.In the leakage risk assessment methods,heuristic process analysis is often used to assess risks.This is not only difficult to capture potentially high-risk conditions in the process,but also requires a lot of time and labor costs.According to the current research status,this paper proposes a leak identification and risk assessment method based on long and short-term memory networks to ensure the safe operation of chemical processes.For the leakage identification model,a dynamic mechanism model is first built to simulate several leakage conditions and collect their process data.Second,the key variables in the process are selected and redundant variables are discarded through node weight filtering based on graph theory.Then,the industrial V&V idea is adopted to build four leak identification models(LSTM,GT-LSTM,CNN,and GT-CNN)based on long and short-term memory networks and convolutional neural networks.Finally,these four models are applied to the case of ammonia synthesis process with 8different leak conditions simulated to compare their performances.The results show that the F1 scores of GT-LSTM and GT-CNN are increased by 0.169 and 0.088 compared with LSTM and CNN respectively,which proves the importance of graph theory based variable selection.The F1 score of the GT-LSTM model reaches 0.971,0.063 higher than GT-CNN model,which proves that the GT-LSTM model is the most effective model for leak identification among these four models.For the risk assessment method,leakage data under a variety of abnormal conditions are first obtained based on the dynamic mechanism model to form a dynamic data set.Secondly,the optimal hyperparameters of the LSTM network are determined through orthogonal experiments on the dynamic data set.Then,based on data generated from the dynamic mechanism model,the LSTM network with optimal hyperparameters is used to learn and predict variables directly related to process risks under unknown conditions.Subsequently,the quantitative risk assessment and risk matrix are combined to determine the risk level of potential abnormal conditions.Finally,safety control schemes for high-risk conditions are designed and verified through mechanism model.The application results of 6 abnormal conditions in the ammonia synthesis process show that the failure of the D102 liquid level controller is a medium-high risk,and the failure of the D101 liquid level controller and the E105 cooling water failure are high risks.Among them,the two-hour radiant heat flux of the leakage caused by the E105 cooling water failure has the largest hazard range,reaching 31.3 meters,and the failure of the D101 liquid level controller shows more serious hazards in the early stage of leakage.Under the disturbance of different abnormal conditions,the system recoveries smooth operation under the control of the designed scheme,which proves the reliability of the proposed method.