Proactive chemical leakage monitoring method using sparse sensor data based on deep learning

Ethylene leakage in enclosed petrochemical facilities poses serious safety risks, yet monitoring is often limited by sparse sensor deployment. Discrete measurements cannot characterise the global concentration distribution or its short-term evolution in geometrically complex indoor spaces. This study proposes a deep learning-based monitoring method that reconstructs and predicts two-dimensional concentration fields from sparse sensor data. Leakage monitoring is formulated as a spatiotemporal field inference problem, and a dual-module framework is developed, comprising a reconstruction module and a one-step prediction module. Both modules employ a customised LeakU-Net architecture that combines U-Net with depthwise separable convolutions and multi-scale feature extraction to efficiently capture dispersion features. High-fidelity Computational Fluid Dynamics (CFD) simulations are used to generate training and testing datasets, and the CFD data are validated through laboratory experiments. For safe validation while preserving similar dispersion characteristics to ethylene, nitrogen is used as the surrogate release gas (with an identical molecular weight to ethylene), and oxygen concentration is measured as a proxy indicator of dispersion. The proposed method achieves reconstruction MSE below 4.0 × 10^–4 and prediction MSE below 1.6 × 10^–5 across the evaluated scenarios. The integrated reconstruction–prediction pipeline runs in approximately 0.3 s, enabling near-real-time field estimation and short-term forecasting. The method provides an efficient and scalable solution for monitoring ethylene-related leakage scenarios under sparse sensing conditions.