Combination of IMU and EMG for object mass estimation using machine learning and musculoskeletal modeling

• Main Author: Diaz, Claire
• Format: Others
• Language: English
• Published: KTH, Skolan för kemi, bioteknologi och hälsa (CBH) 2020
• Subjects: Musculoskeletal Disorders (MSD); Ergonomic assessment; Electromyography (EMG); Inertial Measurement Unit (IMU); Mass estimation; Machine Learning; Musculoskeletal modeling; Medical Engineering; Medicinteknik
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-289483

One of the causes of work-related Musculoskeletal Disorders (MSDs) is the manual handling of heavy objects. To reduce the risk of such injuries, workers are instructed to follow general guidelines on how to lift and carry objects depending on their mass. Current ergonomic assessments using wearable sensors can differentiate correct from incorrect body postures but are limited. Being able to estimate the mass of an object during ergonomic assessment would be a great improvement. This work investigates a combination of Inertial Measurement Units (IMUs) and Electromyography (EMG) sensors for offline estimation of an object’s mass for different movements. 10 participants performed 26 lifting and carrying trials with loads from 0 to 19 kg, while wearing a 17IMU motion capture system and EMG sensors on both biceps brachii and both erector spinae. Two methods were considered to estimate the carried mass: (1) supervised machine learning and (2) musculoskeletal modeling. First, the data was used to select features, train, and compare regression models. The lowest Mean Squared Error (MSE) for 10-fold cross-validation for lifting and carrying combined was 5.8113 for a Gaussian Process Regression (GPR) model with an exponential kernel function. Then, a MSE of 4.42 and a Mean Absolute Error (MAE) of 1.63 kg were obtained also with a GPR for Leave-One-Subject-Out Cross-Validation (LOSOCV) only for lifting and frontal carrying trials. For the same trials, the upper-extremity musculoskeletal model, scaled to each participant, estimated the mass with a MSE of 1.78 and a MAE of 0.95 kg. The study was restricted to lifting and frontal carrying, but the combination of the two technologies showed great potential for object mass estimation.