Impact of Safety Signage Placement on Evacuation Behavior in Virtual Fire Scenarios Based on EDA Data

Evacuation efficiency during fire emergencies directly influences casualty rates. Traditional fire drills suffer from limited scenario realism, elevated safety risks, and single-modal data acquisition, making it difficult to objectively quantify participants’ physiological and behavioral responses under authentic fire conditions. To address these limitations, this paper employs a high-immersion, controllable virtual reality (VR) hotel environment to simulate fire evacuation while concurrently recording electrodermal activity (EDA). Five features are extracted from the EDA signal: PhasicData, PhasicDriver, Skin Conductance (SC), TonicData, and TonicDriver. Three variables are evaluated, signage height (1m, 0.5m, and 0m), spacing (5m and 10m), and presence of active fire, using a hybrid classification model that integrates an im-proved convolutional neural network (CNN), a Transformer-based sequence encoder, and a multi-layer spiking neural network. The CNN autonomously captures multi-scale local features to heighten sensitivity to stress-related signal changes, the Trans-former encoder models long-range temporal dependencies, and the spiking neural network simulates neural dynamics to accurately reconstruct transient stress spikes. Experimental results yield average classification accuracies of 93.36%, 97.71%, and 95.99% across the three scenarios, demonstrating the neuro-engineering value of combining VR-based simulations with EDA monitoring to optimize safety-sign placement.