Multisensor Data-Fusion-Based Gas Hazard Prediction Using DSET and 1DCNN for Underground Longwall Coal Mine

Abstract

The combustible and noxious gases are among the prominent issues affecting the life and the mining operations of underground coal mines (UCMs). Commonly adopted hazard monitoring methodologies in UCM are fuzzy, rule-based systems, statistical methods, and other expert systems, but these models are not reliable for highly complex and nonlinear systems. The neural network’s ability to learn and create nonlinear relationships is beneficial to making hazard prediction models. Especially, convolutional neural networks (CNNs) auto feature extraction capabilities to make it more suitable for the task. But, UCM’s harsh and crucial environment may result in sensor malfunctioning and faults, giving rise to data uncertainty. Like other data-driven models, data uncertainty significantly affects CNN’s performance. This study involves designing an effective and reliable gas hazard monitoring system using a hybrid of Dempster–Shafer evidence theory (DSET)-based filter and one-dimensional CNN (1DCNN) classifier. The novelty of this study is the integration of DSET and 1DCNN to predict the UCM hazard more reliably, even in malfunctioning node scenarios. Inherent usage limitations on traditional communication techniques restrict the application of cloud-based machine learning (ML) methods and this study use novel edge implementation of the filter and the classifier using edge ML (EML) technology. The proposed model’s hazard classification accuracy is 99.6%, even in the faulty node scenarios, where the traditional approaches fail.