A Hybrid Sensor-Fusion and Graph-Based Routing Model for Explainable Multi-Agent Fire Evacuation

Fire evacuation guidance in multi-storey buildings must adapt to spreading fire and smoke, but adaptive machine-learning methods are hard to interpret and verify in regulated safety applications. This paper presents a deterministic alternative that extends the hierarchical building-graph model with sensor fusion with hysteresis, hard exclusion of hazardous nodes, risk-first route ranking, coordination of agents through virtual load, and a cellular-automaton movement model. Only one part of the model is fitted to data: a linear cost model with eight coefficients. It is used only to break ties between routes. All other parameters are set either by regulation or by engineering judgement. As a result, every recommendation can be traced back to sensor values and formulas. Thirteen simulated scenarios in four buildings, with up to 120 agents, were evaluated against four ablated variants with 20 Monte Carlo replicates each. Hard exclusion ensures that recommended routes contain no node classified as fire- or smoke-affected at planning time, which held in all scenarios for agents with an admissible route; this is a model-level property that depends on sensor classification and does not cover agents who do not evacuate. With a fixed coordination coefficient (λ = 0.35), coordination increased the number of exits used in eight scenarios, at a cost in maximum evacuation time in most scenarios where the difference was significant. In the two largest scenarios (120 agents), the standard deviation of the maximum evacuation time was 2.7 to 4.9 times lower than for a baseline with neither coordination nor calibrated weights, but not lower than for the variant without coordination, and the full system had the lowest variance in eight of thirteen scenarios. The only contacts with fire or smoke during movement (Scenario 12) resulted from sensor-classification latency. Code and data are provided for full reproduction.