FAST-AlertNet: Early warning fire-induced collapse of large-span steel truss structures

Abstract

Large-span steel trusses are often adopted as roof structures of public and industrial buildings. Besides significant property loss, the unexpected and sudden collapse of large-span steel trusses caused by fires also threatens the lives of firefighters and trapped people in fires. To prevent casualties to the greatest extent, real-time evaluation of the structural collapse state of the building at fire rescue scenes is critical. This paper presents a novel framework for early warning the collapse of large-span steel truss structures in fire based on proposing the FAST-AlertNet. The FAST-AlertNet can easily obtain real-time displacements based on easily-measured rotations and temperatures of the steel trusses during fire synchronously. Rotations and temperatures, which reflect uncertain structural parameters of steel trusses and fire scenarios, serve as inputs, while displacements are the outputs. A dynamic weighted loss function is employed in the FAST-AlertNet, yielding better performance reflecting the early-warning points than traditional loss functions. To address the high computational cost of obtaining real spatiotemporal patterns of temperature development inside the building in real fire scenes, transfer learning is utilized to transfer the interlinkage among displacements, temperatures, and rotations established by simplified parametric temperature curves, resulting in improved prediction accuracy in reality. A case study demonstrates the availability of the proposed framework, with the predicted remaining pre-collapse time of a typical steel truss satisfactorily matching the real value. Notably, the proposed framework provides a new methodology for early warning building collapse without requiring the difficultly-measured displacements in fire, which advances its feasibility for practical application.