Explainable Emergency Response Recommendation for Foundation Pit Collapse Accidents: An LLM-Enabled E/B/R Ontology-Driven Hybrid Reasoning Framework

Foundation pit collapse accidents require rapid identification of accident conditions and appropriate emergency measures. Accident investigation reports contain valuable response knowledge, but their unstructured format limits direct reuse. This study develops an LLM-enabled E/B/R ontology-driven hybrid reasoning framework for explainable emergency response recommendations. Accident knowledge is organized as Event (E), Bearing (B), and Response (R). The resulting ontology contains 51 subclasses, and 69 foundation pit collapse reports were retained from 1110 collected construction collapse reports to construct the case base. For a newly imported report, Qwen3.6-Plus extracts candidate E/B features within the ontology label space, while code validation and filtering remove invalid labels. The validated features are processed separately by rule-based reasoning (RBR) and case-based reasoning (CBR), and their responses are fused with source labels. On a 39-sample comparison set, Qwen3.6-Plus achieved a micro-F1 score of 0.821 for E/B extraction. Applications to three target cases showed that RBR provided rule-supported measures, whereas CBR contributed additional responses from related historical cases. A questionnaire assessment of the earlier prototype outputs, based on 19 valid responses, yielded an overall mean score of 5.97 on a seven-point scale. The framework provides a traceable means of reusing accident-report knowledge for construction emergency decision support while restricting the LLM to constrained extraction and result explanation.