Intelligent Sensors for Real-Time Structural Health Monitoring Data Analytics

Thesis Abstract

The rapid evolution of sensor technologies and data analytics has enabled real-time structural health monitoring (SHM) systems to transition from passive data collection to proactive, decision-support platforms for civil infrastructure management. However, the deployment of intelligent sensors in diverse structural contexts faces challenges related to data heterogeneity, reliability under harsh environments, limited labeled failure data, and the need for interpretable analytics to support maintenance decisions. This study addresses these gaps by developing an integrated SHM data analytics framework that combines resilient sensing, edge computing, and advanced analytics to deliver timely, accurate condition assessments. The aim of the study is to develop and validate a scalable, real-time SHM data analytics architecture that (i) fuses multi-sensor streams from intelligent accelerometers, strain gauges, and fiber-optic sensors, (ii) detects anomalies and structural anomalies with high precision, and (iii) provides interpretable risk-based maintenance recommendations. Specific objectives include (1) design and deploy a heterogeneous sensor network on a representative transportation bridge and a high-rise modular building, (2) implement edge-enabled data processing using lightweight machine learning models to ensure low-latency monitoring, (3) develop a robust anomaly detection pipeline leveraging ensemble methods and domain-informed physics-based features, (4) evaluate model generalization under environmental variability using transfer learning and domain adaptation, and (5) formulate a decision-support framework that translates analytics into actionable maintenance interventions and safety alerts. Methodologically, the study adopts a mixed-methods approach supported by a pragmatic research design. The population comprises two real-world structures a 1.2-km reinforced concrete bridge and a 40-storey steel-frame building in a metropolitan region. A purposive sample of 120 sensor nodes (60 accelerometers, 40 strain gauges, 20 fiber-optic sensors) is deployed to collect continuous data over 18 months. Data collection instruments include high-resolution accelerometers (1 Hz–1 kHz sampling), distributed acoustic sensing for fiber-optic sensors, strain gauge arrays, GPS-based time synchronization, and environmental sensors measuring temperature, humidity, and wind. The analysis employs a hierarchical data fusion framework with edge computing for initial processing and cloud-based analytics for deeper inference. Statistical and machine learning techniques used comprise time-series forecasting (LSTM networks), unsupervised anomaly detection (Isolation Forest, One-Class SVM), supervised fault classification (Random Forest, Gradient Boosting), and physics-consistent feature engineering (modal analysis, frequency response functions). Model validation uses cross-validation across different structural states, with transfer learning to adapt models from one structure to another. The study also applies Bayesian updating to refine predictive confidence as new data are collected. Data quality assurance includes calibration protocols, sensor health checks, and redundancy schemes to mitigate data loss. Expected findings indicate that the integrated framework achieves improved detection latency (average < 2 seconds from anomaly onset), higher detection accuracy (precision > 92%, recall > 88%) across varying environmental conditions, and robust generalization demonstrated by successful transfer to the second structure with a drop in performance below 6 percentage points after domain adaptation. The analysis will reveal key feature drivers of structural degradation, such as modal damping changes and strain pattern shifts under temperature fluctuations, and quantify the added value of edge processing in reducing cloud latency and bandwidth demands. The study anticipates that the decision-support module will yield maintenance recommendations with quantified risk levels and prioritized intervention timelines. Contributions to knowledge include (i) a novel multi-sensor SHM data fusion architecture integrating edge and cloud analytics for real-time decision support, (ii) an interpretable hybrid analytics pipeline that blends data-driven models with physics-based features to enhance trustworthiness and adoption by practitioners, (iii) empirical evidence on transferability of SHM models across different structural typologies, and (iv) a comprehensive framework for translating analytic outputs into actionable maintenance strategies under uncertainty. The main conclusion underscores the feasibility and value of intelligent sensors combined with real-time analytics for proactive structural maintenance, while recommendations stress standardization of data schemas, sensor health monitoring procedures, and the development of interoperable dashboards for engineers and decision-makers.