Neural network-based source tracking of chemical leaks with obstacles

The leakage of hazardous gases poses a significant threat to public security and causes environmental damage. The effective and accurate source term estimation (STE) is necessary when a leakage accident occurs. However, most research generally assumes that no obstacles exist near the leak source, which is inappropriate in practical applications. To solve this problem, we propose two different frameworks to emphasize STE with obstacles based on artificial neural network (ANN) and convolutional neural network (CNN). Firstly, we build a CFD model to simulate the gas diffusion in obstacle scenarios and construct a benchmark dataset. Secondly, we define the structure of ANN by searching, then predict the concentration distribution of gas using the searched model, and optimize source term parameters by particle swarm optimization (PSO) with well-performed cost functions. Thirdly, we propose a one-step STE method based on CNN, which establishes a link between the concentration distribution and the location of obstacles. Finally, we propose a novel data processing method to process sensor data, which maps the concentration information into feature channels. The comprehensive experiments illustrate the performance and efficiency of the proposed methods.