Efficient Gas Source Tracking in Turbulent Environments Using a Probabilistic Vision Transformer

Turbulent environments pose significant risks to environmental safety and public health due to industrial gas leaks. Probability-Based gas tracking agents are considered a more effective strategy in turbulent conditions compared to reactive methods. However, the probabilistic methods that are highly adaptive to turbulence come with high computational costs, making it difficult to support large-scale or complex deep reinforcement learning methods. This study introduces the Probabilistic Vision Transformer, which uses probabilistic graph partitioning and self-attention mechanisms to model spatial probabilistic features for the agent. This approach overcomes the issues of dimensional explosion and high time complexity caused by the flattening of Bayesian probability graphs as input, enabling the solution of gas tracing problems in turbulent environments.