Two-class load weight discriminator in a lifting-up task using acceleration metrics of human body

Abstract

Low back disorder is a commonly observed worker injury in Japan. Statistical figures revealed that there were 600 to 900 workers who were absent from work more than four days every year because of the low back pain caused by handling heavy objects in the manufacturing industries. Use of mechanical lifters can be a solution but there are still many illshaped heavy objects which should be handled manually in the workplace. Wearable power assist devices can provide physical support to workers who are handling heavy loads in their daily work. The present paper proposes a two-class weight discriminator for lifting-up motion of a human worker, looking to use the output of the proposed discriminator in the control of the wearable power assist device in the future. The proposed discriminator mainly uses the magnitude of two dimensional acceleration spikes measured during a lift-up task using an accelerometer mounted on his/her shoulder, when he/she is trying to do a lift-up motion. We formulated and trained both the linear and the nonlinear support vector machines (SVMs) for the classification of the feature vectors, and evaluated the trained SVMs with independent evaluation dataset. Satisfactory discrimination accuracy has been observed both with the linear and the nonlinear SVMs which use the reaction acceleration feature values. We also evaluated the use of additional three dimensional accumulated body motion accelerations as supplemental feature vector elements. Higher dimensional SVMs were formulated and trained accordingly and the result of discrimination accuracy clarified both positive and negative aspects of high dimensional feature vector for the discrimination of two load weight classes in lift-up motions.