An explainable machine learning based flashover prediction model using dimension-wise class activation map

This paper presents the development of an explainable machine learning based flashover prediction model, named xFlashNet. Synthetic temperature data from more than 17 000 fire cases are used for model development. The effect of missing data due to heat detectors to elevated temperature from the fire scene is also considered. xFlashNet utilizes multi-residual convolutional layers to effectively learn the indicative temperature features and dimension-wise class activation map (dCAM) to interpret the model decision. The proposed model is benchmarked against three current-state-of-the-art models. Results shows that the proposed model outperforms the benchmark models and it has an overall accuracy of about 92.9%. Based on dCAM, the model decision is analyzed. Depending on the location of the fire and the heat detector operating conditions, the proposed model shows the discriminative region of the temperature inputs which influence the model to make the decision. In addition, model testing against real fire data is conducted. It is believed that this present work contributes a step forward to bring trustworthy ML systems to fire safety applications and to enhance situational awareness for firefighting safety that can help reduce firefighter injuries and deaths.