EviFlash: Uncertainty-Aware Flashover Prediction Using Selective State Space Model

Flashover—the abrupt transition of a compartment fire to full development—demands early prediction with quantified confidence for risk-aware operations. We propose an uncertainty-aware evidential flashover predictor (EviFlash) that models multi-sensor temperature streams with a selective state space sequence model, enabling long-horizon inference with linear-time complexity suitable for streaming data. The network jointly predicts time-to-flashover and a binary flashover-within-horizon event, and represents uncertainty by combining heteroscedastic regression (aleatoric) with stochastic/deep-ensemble inference (epistemic), from which calibrated prediction intervals can be derived. Experiments on two synthetic multi-compartment datasets released by NIST (three- and six-compartment layouts) show that our method consistently reduces TTF mean absolute error and improves recall/F1 over strong recurrent and attention-based baselines, including Bi-LSTM/GRU variants and a recent graph-based flashover model. Ablation studies confirm that both the selective state space backbone and the uncertainty modeling contribute to these gains, and transfer tests across layouts indicate that the proposed model maintains competitive accuracy under changes in compartment geometry and sensor placement, suggesting that uncertainty-aware sequence modeling is a promising direction for early flashover warning in smart firefighting systems without relying on costly computational fluid dynamics simulations.