Machine Learning-Based Framework for Pre-Impact Same-Level Fall and Fall-from-Height Detection in Construction Sites Using a Single Wearable Inertial Measurement Unit.

Abstract

Same-level-falls (SLFs) and falls-from-height (FFHs) remain major causes of severe injuries and fatalities on construction sites. Researchers are actively developing fall-prevention systems requiring accurate SLF and FFH detection in construction settings prone to false positives. In this study, a machine learning-based approach was established for accurate identification of SLF, FFH, and non-fall events using a single waist-mounted inertial measurement unit (IMU). A total of 48 participants executed 39 non-fall activities, 10 types of SLFs, and 8 types of FFHs, with a dummy used for falls exceeding 0.5 m. A two-stage feature extraction yielded 168 descriptors per data window, and an ensemble SHAP-PFI method selected the 153 most informative variables. The weighted XGBoost classifier, optimized via Bayesian techniques, outperformed other current boosting algorithms. Using 5-fold cross-validation, it achieved an average macro F1-score of 0.901 and macro Matthews correlation coefficient of 0.869, with a latency of 1.51 × 10^-3 ms per window. Notably, the average lead times were 402 ms for SLFs and 640 ms for FFHs, surpassing the 130 ms inflation time required for wearable airbags. This pre-impact SLF and FFH detection approach delivers both rapid and precise detection, positioning it as a viable central component for wearable fall-prevention devices in fast-paced construction scenarios.