Computer Vision-Based Fire and Explosion Risk Assessment Using Adaptive Dynamic Fault Tree Analysis

Abstract

Fire and explosion accidents are one of the most common reasons for industrial accidents that result in serious injuries, damage to equipment, pollution of environment, and loss of production. Traditional industrial safety systems employ smoke, heat, and gas sensors, and these sensors are capable of detecting any hazardous situation only when it has grown to a considerable extent. Advances in computer vision allow monitoring of industrial environment through an existing CCTV system, which opens new possibilities of early hazard detection and risk assessment. In this paper, we suggest the implementation of the computer vision solution to the problem of industrial fire and explosion risk assessment, which incorporates the use of Faster R-CNN for hazards detection, DeepSORT algorithm for tracking of multi-objects, Adaptive Dynamic Fault Tree Analysis (ADFTA) for probabilistic risk assessment, and Gradient-weighted Class Activation Mapping (Grad-CAM) for explainable artificial intelligence. Video from the industrial CCTV cameras is processed prior to detecting the following hazards: fire, smoke, combustible material, Gas Leakage Indicators, violation of Personal Protective Equipment (PPE). All these factors are provided to the ADFTA model to calculate the probability of fire and explosion and obtain a Fire and Explosion Risk Index (FERI), which is categorized into four levels: Low, Medium, High, and Critical. In addition, Grad- CAM enhances the transparency of the system through visualization of the areas of the image that are responsible for detecting a hazard. The proposed approach provides continuous monitoring and warning generation for industries under the environment of Industry 5.0.