Fall-from-Height Detection Using Deep Learning Based on IMU Sensor Data for Accident Prevention at Construction...

Area of Science:

• Occupational safety
• Biomechanical engineering
• Machine learning applications

Background:

• Construction workers are at high risk for fall-from-height (FFH) accidents.
• Injury severity in FFH accidents correlates with peak acceleration.
• Predictive modeling using sensor data can enhance worksite safety.

Purpose of the Study:

• To develop a risk prediction model for fall-from-height accidents using IMU sensor data.
• To evaluate the effectiveness of various deep learning models for FFH risk assessment.
• To reduce fatal injuries through proactive accident prevention at construction sites.

Main Methods:

• Subjects performed non-fall (NF), low-hazard-fall (LF), and high-hazard-fall (HF) movements.
• Inertial Measurement Unit (IMU) sensors recorded three-axis acceleration and angular velocity at the T7 position.
• Deep learning models (1D-CNN, 2D-CNN, LSTM, Conv-LSTM) were trained and compared using Mean Absolute Error (MAE) and Mean Squared Error (MSE).

Main Results:

• Peak acceleration values were ≤4 g for general work and ≥9 g for FFHs.
• The Conv-LSTM model trained with MAE achieved the lowest error (MAE: 1.36 g).
• Classification based on predicted peak acceleration yielded a high accuracy of 97.6%.

Conclusions:

• The developed deep learning model effectively predicts fall-from-height risk levels.
• IMU sensor data combined with advanced algorithms can significantly improve construction site safety.
• This predictive approach holds potential for mitigating severe and fatal injuries in occupational settings.