Static and Dynamic Work Activity Classification from a Single Accelerometer: Implications for Ergonomic Assessment of Manual Handling Tasks

OCCUPATIONAL APPLICATIONS

A single wearable sensor (accelerometer) on the chest was employed to classify static and dynamic activities commonly observed in manual handling jobs. Utilizing only two features obtained from this sensor, 15 different simulated activities were classified with 93%–98% accuracy. The classification models developed here could be used to objectively quantify workers’ tasks through the course of their work shifts, thereby enabling more accurate and efficient ergonomic assessments while requiring using only a simple wearable device.

TECHNICAL ABSTRACT Background: The use of accelerometers for recognizing physical activity has increased substantially; however, there is a need to develop activity classification models that are effective for a large range of static and dynamic work activities commonly seen in manual handling. Purpose: This study aimed to develop an efficient classification model to recognize static and dynamic work activities from a single accelerometer attached to the chest. Methods: Accelerometer data were obtained during 15 simulated activities from 27 (13 males, 14 females) healthy adults. First-order differencing was employed to account for between-subject variability. Eighty-one datasets were created using nine different time windows and nine different data samples with randomly-selected start times. Medians of postural angles and the area-under-the-curve of transformed triaxial data were used as predictors in classification models (random forest and support vector machines). Results: In the random-forest models, the highest (98.2%) and lowest (93%) overall accuracies across the 15 activities were associated with time windows of 6 and 2 s, and data samples of 5000 and 1000, respectively. The findings also indicated that accuracy of the random-forest models improved with increasing time window durations and data samples. Finally, activity classification with the support vector machine model, corresponding to the random-forest model with the highest performance, achieved an accuracy of 95.5%. Conclusions: These results suggest that the use of only two features from a single accelerometer can help classify work activities, which in turn can improve real-time or remote ergonomic risk assessments of manual handling tasks. Future work is needed, though, to examine the generalizability of the current models to efficiently classify complex manual handling tasks in diverse worker groups.

Keywords:

• Area under the curve
• activity classification
• low back
• sensor
• machine learning

Disclosure statement

No potential conflict of interest was reported by the authors.