Online change point methods for structural shock detection on resource-constrained edge devices: An analytical benchmarking study

Timely detection of structural shock events is critical for real-time health monitoring and system protection. This study evaluates five online detection methods for identifying shock-induced changes in drop-tower acceleration data: Bayesian Online Change Point Detection (BOCPD), Cumulative Summation (CUSUM), Expectation Maximization (EM)-based Gaussian mixture classification, Grey Systems Modeling (GM), and Isolation Forest. Each method is benchmarked against an offline detection baseline using precision, recall, F1 score, and detection delay. BOCPD achieved the best balance of accuracy and responsiveness (F1 score: 0.602; delay: 0.277 ms); Isolation Forest provided the second-highest F1 score (0.513), whereas GM achieved the shortest clean-signal detection delay (0.208 ms). The F1-score ranking is unchanged across four alternative offline reference definitions derived from the same acceleration signal. BOCPD, CUSUM, EM, and Isolation Forest retain their clean-signal performance under white Gaussian noise down to a signal-to-noise ratio of 0 dB, whereas the precision of GM degrades from 0.97 to 0.54. Parameter sweeps identify the reported BOCPD setting as its best tested configuration, but also show that cross-algorithm F1 rankings depend on the chosen parameters, particularly for GM and Isolation Forest. Repeating the evaluation on all five experimental runs, including a reference-light early-impact check, confirms the ordering BOCPD > Isolation Forest > CUSUM > EM on every run, whereas the rank of GM varies. Results highlight the effectiveness of lightweight online detection methods for shock detection and their algorithmic suitability for embedded structural health monitoring and control systems, pending validation on embedded hardware. Data and code are publicly available.