Anomaly detection in structural health monitoring: Integrating a hybrid unsupervised learning network with isolation forest for vibration-based looseness identification in bolted connections

Bolted connections are widely used in engineering structures; however, bolt loosening remains a critical problem that can impair load transfer, compromise structural integrity, and, in severe cases, lead to catastrophic failure. This study proposes an unsupervised framework for vibration-based bolt-loosening detection by integrating a hybrid deep reconstruction network with the Isolation Forest algorithm. The proposed framework relies exclusively on intact-condition data for both network training and anomaly-model construction, which makes it suitable for practical real-world structural health monitoring scenarios. The hybrid network combines a multihead self-attention mechanism, convolutional neural network (CNN) layers, and a Convolutional Long ShortTerm Memory (ConvLSTM) layer to efficiently capture global temporal dependencies, local dynamic patterns, and sequential reconstruction characteristics in acceleration signals. Bolt loosening is identified through deviations between the original and reconstructed responses, quantified using three reconstruction-based damagesensitive indices: Root Mean Squared Error (RMSE), Jensen-Shannon Divergence (JSD), and Reconstructed-toOriginal Energy Ratio (ROER). These features are subsequently used as inputs to the Isolation Forest model for anomaly detection. The proposed method was validated on a full-scale laboratory steel frame benchmark and achieved prediction accuracies of 90.0% for tight connections and 93.3% for loose connections. The results demonstrate that the proposed framework provides an effective and practical solution for bolt-loosening detection by reducing the need for damaged-state data, lowering inspection costs, and enhancing the reliability of autonomous monitoring.