(54g) Hybrid Mathematical Modeling and Deep Learning of Big Data for Optimal Sensor Placement and Tracking of Chemical Leaks

Leak incidents in chemical plants can easily escalate and spread to secondary and tertiary accidents, such as fire and explosions, and classified as one of major industrial accidents that may cause large personal injuries and material damage, if it not controlled reliably at the initial stage. A number of sensors are allocated near high-risk leak sources for earlier detection of leaks. Data reconciliation is performed to minimize nuisance alarms. In addition to data gathering, it is important to develop an active leak-source tracking system that promptly monitors the information and diagnoses about the leak source at the initial stage of the accident development – one of challenging inverse problems in safety. When the right source(s) of leak is diagnosed, it informs on-site response agents of the derived leak information, enabling systematic and efficient mitigation responses. Improving detectability and diagnostic performance of this kind of system contributes to the success rate of the whole response process and minimizes the consequence by the chemical leak and dispersion.

In this study, we propose an intelligent tracking system monitoring leaks spreading through a plant boundary and properly diagnosing the leak source in time. It learns the big data from dynamic simulations of representing probable leak scenarios, selected from risk analysis, and predicts the leak source by inputting a set of selected time-series sensor data to the deep neural networks. Training sets of data are obtained out of CFD simulations performed on the leak accident scenarios for actual chemical plant geometry and yearly distribution of weather conditions, and validated models are developed to predict leaking points by extracting concentration data from simulation results at virtual sensor locations.

Two types of popular network structure are studied to compare the best performance: deep feed-forward neural network (FNN) and recurrent neural network (RNN), reflecting on the time-series patterns of collected concentration data in real-time. The proposed tracking method based on deep neural networks not only shows high performance but also overcomes various problems of existing leak source tracking methods based on inverse vector tracking, such as local minimum, large demand time to track and essential use of sensitive mobile sensors, etc. Industrial case studies are discussed and it is expected that the proposed method can be widely used to replace the existing ad-hoc methods for leak source tracking.