Workface Methane Concentration Prediction Based on Transformer-KAN

Abstract

This paper proposes an underground methane concentration prediction system for working faces based on the Transformer-KAN, addressing the challenge of insufficient spatial correlation modeling in multi-sensor data. The core innovation lies in leveraging the self-attention mechanism of Transformer networks to dynamically compute the influence weights of sensors (atmospheric pressure, temperature, etc.) at different locations within the roadway on methane concentration at target points. This overcomes the limitations of traditional models reliant on fixed-distance decay models. Simultaneously, utilizing the differentiable spline basis functions of the KAN network, the complex nonlinear physical mapping relationship between atmospheric pressure gradients and methane concentration is constructed and explained. High-precision sensor networks are deployed in critical areas of coal mine tunnels. Data acquisition and preprocessing employ techniques like precise time synchronization and Kalman filtering. The system ultimately integrates the accelerated inference engine to achieve minute-level (target 200ms high-precision (target RMSE≤0.15%  prediction. Finally, real-time data from a mining face in a transport tunnel of a certain mine was selected to construct and validate on-site predictions, providing core technological support for the intelligent safety upgrade of coal mines and the "dual carbon" goals.