Adaptive Multilevel Ensemble Intelligent Framework for Rapid Seismic Damage-State Classification of Mega-Sub Controlled Structural Systems Under Data-Scarce Conditions

Abstract

To address the difficulties in rapid postearthquake damage-state identification of complex high-rise structures, the high computational cost of nonlinear time-history analysis, and the limited availability of postearthquake samples, this study proposes an adaptive multilevel ensemble intelligent framework for mega frame structure (MFS) and mega-sub controlled structural system (MSCSS), aiming to achieve rapid and stable seismic damage assessment under data-scarce conditions. The proposed framework adopts regularized kernel extreme learning machine (RKELM) as the core classifier, where kernel mapping is used to characterize the complex nonlinear mapping relationship between ground motion intensity measures and structural damage states, and regularization is introduced to reduce the risk of overfitting under small-sample conditions. On this basis, a bidirectional beetle swarm adaptive optimization (BBSAO) algorithm is developed to achieve stable optimization of the kernel parameter and regularization coefficient of RKELM through bidirectional subpopulation cooperative search and adaptive step-size adjustment. Furthermore, a confusion-focused multiclass boosting (CFMBoost) mechanism is constructed to dynamically adjust sample weights according to class confidence and classification difficulty, enabling the base learners to focus more on easily confused and damage-sensitive states. The proposed framework is validated using the nonlinear responses and damage states of MFS and MSCSS under different ground motion inputs. The results show that the proposed framework achieves high accuracy and stability in small-sample damage-state classification, providing an efficient and reliable intelligent approach for rapid postearthquake damage assessment and emergency decision-making of complex high-rise structures.