Bayesian Online Real-Time Change Point Detection for Structural Monitoring

To detect structural anomalies in real time by identifying statistically significant shifts in noise event patterns associated with material cracking.

• Model structural behaviour as a time series

  Represent crack-related acoustic events as a probabilistic temporal process.

• Detect change points in real time

  Identify when the underlying distribution of events shifts significantly.

• Apply Bayesian inference

  Continuously update beliefs about system state as new data arrives.

• Handle noisy data

  Distinguish meaningful structural changes from background noise fluctuations.

• Enable early warning signals

  Flag abnormal patterns that may indicate structural degradation.

• Implemented Bayesian Online Change Point Detection (BOCPD) algorithm
• Modelled event counts using probabilistic distributions (e.g. Poisson/Gaussian)
• Computed run-length probabilities to track likelihood of regime changes
• Simulated structural noise data to validate detection performance
• Applied algorithm to real-world-inspired building crack event data
• Optimised for near real-time inference with incremental updates

• Translated a financial time-series algorithm to a structural health monitoring domain
• Detected subtle distribution shifts in noisy event data using Bayesian methods
• Enabled real-time anomaly detection rather than offline analysis
• Demonstrated how probabilistic modelling can outperform simple thresholding approaches
• Bridged theory (Bayesian inference) with a practical engineering application

• Demonstrated feasibility of real-time structural anomaly detection
• Provided a framework for early warning systems in infrastructure monitoring
• Reduced reliance on static thresholds for anomaly detection
• Showcased cross-domain application of advanced statistical methods

• Bayesian methods are powerful for real-time uncertainty-aware decision making
• Change point detection is highly transferable across domains (finance → engineering)
• Handling noise correctly is critical in real-world sensor data
• Online algorithms provide significant advantages over batch processing