Alert Fatigue-Aware Risk Detection for Construction Safety

Introduction: Major failures in civil infrastructure typically develop through sequences of overlooked incidents that progressively accumulate into cascading risk. While existing monitoring systems identify individual threshold violations, they rarely model cumulative risk progression or account for operator alert fatigue. This study presents a proof-of-concept prototype combining machine learning anomaly detection with adaptive human-in-the-loop alert management for construction safety monitoring.

Methods: A web-based system was developed integrating an Isolation Forest model trained on 5,185 synthetic construction-site event sequences, generated across three risk profiles — normal, elevated, and critical — using five features per event: individual severity, cumulative risk, event type, site zone, and time step. Severity ranged from 0.1–0.3 for normal to 0.8–1.0 for critical sequences. Cascading risk was represented through cumulative severity accumulation across sequential events. Operator fatigue was modelled using predefined simulation rules across three response categories: fast (4-15 seconds), normal (15–60 seconds), and fatigued (above 60 seconds). Fatigued operators had a 30% alert ignorance probability applied per alert. No human participants were involved.

Results: The model produced anomaly rates of 10%, 20%, and 40% across normal, elevated, and critical scenarios respectively. All anomalies in the elevated-risk scenario occurred above the defined alert threshold. No alerts were triggered under normal conditions; elevated-risk and critical scenarios produced three and four alerts respectively. Alert operators averaged 3.7 seconds response time with 0% fatigue accumulation; normal operators averaged 8.4 seconds reaching 45% fatigue; fatigued operators averaged 34.2 seconds, ignored 66.7% of alerts, and reached maximum simulated fatigue. The adaptive mechanism escalated alert intensity when reduced responsiveness was detected.

Conclusions: This proof-of-concept demonstrates the feasibility of integrating cumulative risk modelling, Isolation Forest anomaly detection, and adaptive human-in-the-loop alert escalation within a construction safety monitoring framework. As evaluation relies entirely on synthetic data and simulated behaviour, findings represent an initial feasibility assessment. Future work should focus on validation using real incident datasets and physical sensor integration.