Smart and Fault-Tolerant Multisensor Fusion Model for UCM Methane Hazard Monitoring Based on Belief Divergence Backed DS Filter and Hybrid CNN-LSTM Classifier

The underground coal mine (UCM) dynamic and complex environment impose various hazards that significantly affect the mining production and safety of the personnel. Flammable and poisonous gases significantly contribute to many fatal accidents. This study proposes a real-time-based reliable gas hazard monitoring system using multi-sensor data fusion. A hybrid of CNN-LSTM-based deep neural network (HCLM) is developed to serve the purpose. Due to the challenging environment of the UCM, sensor malfunctioning is inevitable and severely affects the performance of HCLM. A novel front-end filter (FEF) is developed based on Damper Shafer's (DS) theory and belief divergence-based weighted credibility metric to overcome the drawback of HCLM. In the laboratory trial, it is observed that the hazard classification accuracy of HCLM for the faulty node scenarios is 85%. In contrast, the accuracy of the HCLM integrated with FEF is maintained at 98%, even for multiple faulty node cases. Another novelty of this study is the tinyML implementation of the proposed model. Due to UCM's inherent complexities and challenges, traditional wireless communications face operational difficulties. Hence, a cloud-based machine learning operation is not a feasible option in UCM. Hence, using the concept of tinyML, the proposed model is directly deployed on a microcontroller near the data sources, thereby reducing network latency and security issues.