Feasibility Study of Advanced Neural Networks Applied to sEMG-Based Force Estimation

Abstract

To find out the feasibility of different neural networks in sEMG-based force estimation, in this paper, three types of networks, namely convolutional neural network (CNN), long short-term memory (LSTM) network and their combination (C-LSTM) were applied to predict muscle force generated in static isometric elbow flexion across three different circumstances (multi-subject, subject-dependent and subject-independent). Eight healthy men were recruited for the experiments, and the results demonstrated that all the three models were applicable for force estimation, and LSTM and C-LSTM achieved better performances. Even under subject-independent situation, they maintained mean RMSE% of as low as 9.07 ± 1.29 and 8.67 ± 1.14. CNN turned out to be a worse choice, yielding a mean RMSE% of 12.13 ± 1.98. To our knowledge, this work was the first to employ CNN, LSTM and C-LSTM in sEMG-based force estimation, and the results not only prove the strength of the proposed networks, but also pointed out a potential way of achieving high accuracy in real-time, subject-independent force estimation.

Estimating Biomechanical Time-Series with Wearable Sensors: A Systematic Review of Machine Learning Techniques

“…In this review, we use the term feedback to refer to models which also use output and/or internal state variable feedback. For example, herein Elman networks [51], long-short term memory (LSTM) neural networks [52,53], and non-linear/linear autoregressive (with exogenous inputs) models [48,54] are all considered to have a feedback structure. In general, an exogenous input ( ) will be either the value of a sensor time-series at time , ( ), or a discrete feature which describes over some finite time interval.…”

“…Neural networks were the most popular regression model. 8.67%, 9.07%, and 12.13% respectively) for estimating contact forces at the distal forearm and was one of the few studies to use a leave-one-subject-out validation approach [53].…”

“…These techniques have been shown to improve estimation performance compared to other methods [35,45]. Most papers used the standard highpass filter, rectify, lowpass filter processing to estimate sEMG amplitudes and a broad range of lowpass filter cutoff frequencies were used [15,46,48,53,54,56,57,62,68,70,71,73,84,92,95]. In addition to enveloping techniques, some incorporate the fact that the observed sEMG is the superposition of pattern to exist [97].…”

Optimal Feature Selection for EMG-Based Finger Force Estimation Using LightGBM Model

“…To achieve accurate and efficient finger force estimation, the proper estimation model is to be first decided. Considering the requirement of generalisability due to large number of potential users, NN-based methods may not be the best choice as neural networks are usually hard to train and tune due to the subject-dependant cost for training and optimizing the models [14]. A LightGBM (Gradient Boosting Machine) model [15] under the general class of Gradient Boosting Decision Tree (GBDT) methods has been chosen for this estimation task with its wide generalisability, robustness against signal noises and less proneness to overfitting.…”

Force Estimation Based on sEMG using Wavelet Analysis and Neural Network

“…As for sEMG-force relationship, some researchers tried to build the model with computational algorithms such as the Hill model [17], polynomial fitting model [18], fast orthogonal search (FOS) [19], and parallel cascade identification (PCI) [20].Zhang et al studied the relationship between joint force muscle activation and joint force based on HD-sEMG and realized the recognition of joint force [21]. Xu et al used LSTM and CNN to achieve the evaluation of sEMG-based hand output force, but the force identification and prediction made by this method is the output effect of instantaneous force, and the required data samples and the requirements of the higher computing environment [22].…”