An incremental broad ensemble learning framework for the identification of structural damage in conditions of data scarcity and non-stationarity

Abstract

The limited availability of structural damage samples in engineering practice, together with temporal variability in data acquisition, remains a key challenge for accurate intelligent structural damage detection. To overcome these obstacles, this study introduces an incremental broad ensemble learning framework for damage identification that utilizes drift sensing and node tracking, termed IBEL-DSNT. In particular, the concepts of dynamic weight allocation and incremental mechanisms for samples, broad learning systems, and feature nodes are first clarified to tackle the recognition challenges posed by insufficient samples and difficult-to-classify instances. This methodology capitalizes on the mathematical characteristics of pseudo-inverse computation to significantly improve the efficiency of model training. Following this, an adaptive node incremental mechanism, regulated by accuracy thresholds, is established to autonomously manage the scale of nodes and variations in precision, effectively alleviating concerns of excessive computational burden and severe overfitting that arise from unchecked node expansion during broad learning. In addition, a drift-detection-based node topology optimization strategy is developed to synchronize node evolution with model accuracy during incremental learning. Validation through three structural case studies demonstrates that the proposed approach attains a recognition accuracy exceeding 97 % with merely 150 training samples. In comparison to baseline methods characterized by simpler architectures, the suggested framework achieves a minimum of a 2 % enhancement in recognition accuracy by means of regulated increases in mapping nodes. With a moderate augmentation of both mapping and enhancement nodes, the rate of node expansion remains as low as 0.003 s per node, whilst improvements in recognition accuracy can surpass 20 %. Overall, the results confirm that the proposed method provides stable and accurate real-time damage identification, with strong adaptability to concept drift and robust performance under small-sample conditions.