Selecting the appropriate input variables in a regression approach to estimate actively generated muscle moments around L5/S1 for exoskeleton control

Control strategies for trunk exoskeletons based on motion intent recognition: A review

BACKGROUND

Wearable trunk exoskeletons hold immense potential in fields such as healthcare and industry. Previous research has indicated that intention recognition control plays a crucial role in users' daily use of exoskeletons.

OBJECTIVE

This review aims to discuss the characteristics of intention recognition control schemes for intelligent trunk exoskeletons under different control objectives over the past decade.

METHODS

Considering the relatively late development of active trunk exoskeletons, we selected papers published in the last decade (2013 to 2023) from the Web of Science, PubMed, and IEEE Xplore databases. In total, 50 articles were selected and examined based on four control objectives.

RESULTS

In general, we found that researchers focus on trunk exoskeleton devices designed for assistance and motor augmentation, which rely more on body movement signals as a source for intention recognition.

CONCLUSION

Based on these results, we identify and discuss several promising research directions that may help to attain a widely accepted control methods, thereby advancing further development of trunk exoskeleton technology.

The effect of back muscle fatigue on EMG and kinematics based estimation of low-back loads and active moments during manual lifting tasks

This study investigated the effects of back muscle fatigue on the estimation of low-back loads and active low-back moments during lifting, using an EMG and kinematics based model calibrated with data from an unfatigued state. Fourteen participants performed lifting tasks in unfatigued and fatigued states. Fatigue was induced through semi-static forward bending. EMG, kinematics, and ground reaction forces were measured, and low-back loads were estimated using inverse dynamics and EMG-driven muscle model. A regression model was developed using data from a set of calibration lifts, and its accuracy was evaluated for unfatigued and fatigued lifts. During the fatigue-inducing task, the EMG amplitude increased by 2.8 %MVC, representing a 38% increase relative to the initial value. However, during the fatigued lifts, the peak EMG amplitude was found to be 1.6 %MVC higher than that observed during the unfatigued lifts, representing a mere 4% increase relative to the baseline unfatigued peak EMG amplitude. Kinematics and low-back load estimates remained unaffected. Regression model estimation errors remained unaffected for 5 kg lifts, but increased by no more than 5% of the peak active low-back moment for 15 kg lifts. We conclude that the regression-based estimation quality of active low-back moments can be maintained during periods of muscle fatigue, although errors may slightly increase for heavier loads.

Development of a real time estimation method of L5S1 moments in occupational lifting

Mechanical loading of the low-back is an important risk factor for the development of low-back pain. Real-time estimation of the L5S1 joint moment (M_L5S1) can give an insight to reduce mechanical loading. Model accuracy depends on sensor information, limiting the number of input variables to estimate M_L5S1 increases practical feasibility, but may decrease accuracy. This study aimed to find a model with a limited set of input variables without a large reduction in accuracy. We compared two approaches. The first was based on a simplified inverse dynamics model (SM) that requires a limited number of input variables (EMG/ground reaction forces, and orientations derived from an optoelectronic system (OMC)). Two variations were examined, to determine to what extent arm orientations were needed. The second approach was based on a regression model (RM) that uses the SMs as ground-truth. Two variations in terms of sensor use and calibration were examined. Test trials consisted of re-stacking a stack of 3 boxes. A high-end lab-based OMC-system was used as the gold standard (GS). Fifteen healthy participants, 9 males and 6 females (age 21-30) participated in this study. R², RMSE, and peak-difference with the GS M_L5S1 estimate were compared between models with a repeated-measures ANOVA. The SM including arm sensors performed similar or better than the regression models (r > 0.9 and RMSE < 15 % of average peak moment). However, from the perspective of practical feasibility and minimizing the required number of sensors during work, the best approach would be using one of the two regression model approaches.

A Simple Method to Optimally Select Upper-Limb Joint Angle Trajectories from Two Kinect Sensors during the Twisting Task for Posture Analysis

A trunk-twisting posture is strongly associated with physical discomfort. Measurement of joint kinematics to assess physical exposure to injuries is important. However, using a single Kinect sensor to track the upper-limb joint angle trajectories during twisting tasks in the workplace is challenging due to sensor view occlusions. This study provides and validates a simple method to optimally select the upper-limb joint angle data from two Kinect sensors at different viewing angles during the twisting task, so the errors of trajectory estimation can be improved. Twelve healthy participants performed a rightward twisting task. The tracking errors of the upper-limb joint angle trajectories of two Kinect sensors during the twisting task were estimated based on concurrent data collected using a conventional motion tracking system. The error values were applied to generate the error trendlines of two Kinect sensors using third-order polynomial regressions. The intersections between two error trendlines were used to define the optimal data selection points for data integration. The finding indicates that integrating the outputs from two Kinect sensor datasets using the proposed method can be more robust than using a single sensor for upper-limb joint angle trajectory estimations during the twisting task.

Optimizing Calibration Procedure to Train a Regression-Based Prediction Model of Actively Generated Lumbar Muscle Moments for Exoskeleton Control

The risk of low-back pain in manual material handling could potentially be reduced by back-support exoskeletons. Preferably, the level of exoskeleton support relates to the required muscular effort, and therefore should be proportional to the moment generated by trunk muscle activities. To this end, a regression-based prediction model of this moment could be implemented in exoskeleton control. Such a model must be calibrated to each user according to subject-specific musculoskeletal properties and lifting technique variability through several calibration tasks. Given that an extensive calibration limits the practical feasibility of implementing this approach in the workspace, we aimed to optimize the calibration for obtaining appropriate predictive accuracy during work-related tasks, i.e., symmetric lifting from the ground, box stacking, lifting from a shelf, and pulling/pushing. The root-mean-square error (RMSE) of prediction for the extensive calibration was 21.9 nm (9% of peak moment) and increased up to 35.0 nm for limited calibrations. The results suggest that a set of three optimally selected calibration trials suffice to approach the extensive calibration accuracy. An optimal calibration set should cover each extreme of the relevant lifting characteristics, i.e., mass lifted, lifting technique, and lifting velocity. The RMSEs for the optimal calibration sets were below 24.8 nm (10% of peak moment), and not substantially different than that of the extensive calibration.