Deep Learning-Based Structural Health Monitoring for Bridges and High Rise Buildings Using Sensor Data

Abstract

Ageing civil infrastructure, growing traffic demand, and increasingly frequent extreme loading events have made continuous condition assessment of bridges and high-rise buildings an urgent engineering priority. Conventional visual inspection is episodic, labour-intensive, and often incapable of detecting hidden or early-stage deterioration. Structural Health Monitoring (SHM) supported by dense sensor networks offers a data-rich alternative, yet the sheer volume, noise content, and non-stationarity of field measurements limit the effectiveness of classical signal processing pipelines. This paper develops and evaluates a deep learning framework that transforms raw multi channel sensor streams into actionable damage diagnoses. A hybrid architecture that couples one-dimensional convolutional neural network (CNN) layers for automatic feature extraction with long short-term memory (LSTM) layers for temporal dependency modelling is proposed and benchmarked against support vector machines, multilayer perceptrons, standalone CNNs, and standalone LSTMs. The framework is exercised on two simulated case studies representative of field deployments: a three-span continuous girder bridge instrumented with accelerometers and strain gauges, and a forty-storey reinforced-concrete building instrumented with accelerometers and tilt sensors. Across four damage states, the hybrid model achieves 97.8% classification accuracy for the bridge dataset and 96.9% for the building dataset, outperforming all baselines while remaining robust to 10% additive measurement noise. The study also discusses data scarcity, environmental variability, model interpretability, and edge deployment as the principal barriers to practice, and outlines research directions including transfer learning, physics-informed networks, and digital-twin integration.