Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Effect of Muscle Fatigue on Surface Electromyography-Based Hand Grasp Force Estimation

Effect of Muscle Fatigue on Surface Electromyography-Based Hand Grasp Force Estimation Hindawi Applied Bionics and Biomechanics Volume 2021, Article ID 8817480, 12 pages https://doi.org/10.1155/2021/8817480 Research Article Effect of Muscle Fatigue on Surface Electromyography-Based Hand Grasp Force Estimation 1 2 2 2 2 3 Jinfeng Wang, Muye Pang , Peixuan Yu , Biwei Tang, Kui Xiang, and Zhaojie Ju Department of Information, Wuhan Huaxia University of Technology, 430223 Wuhan, China Intelligent System Research Institute, Wuhan University of Technology, 430070 Wuhan, China Intelligent System & Biomedical Robotics Group, University of Portsmouth, PO1 3HE Portsmouth, UK Correspondence should be addressed to Muye Pang; pangmuye@whut.edu.cn Received 19 May 2020; Revised 14 January 2021; Accepted 30 January 2021; Published 15 February 2021 Academic Editor: Dongming Gan Copyright © 2021 Jinfeng Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Surface electromyography- (sEMG-) based hand grasp force estimation plays an important role with a promising accuracy in a laboratory environment, yet hardly clinically applicable because of physiological changes and other factors. One of the critical factors is the muscle fatigue concomitant with daily activities which degrades the accuracy and reliability of force estimation from sEMG signals. Conventional qualitative measurements of muscle fatigue contribute to an improved force estimation model with limited progress. This paper proposes an easy-to-implement method to evaluate the muscle fatigue quantitatively and demonstrates that the proposed metrics can have a substantial impact on improving the performance of hand grasp force estimation. Specifically, the reduction in the maximal capacity to generate force is used as the metric of muscle fatigue in combination with a back-propagation neural network (BPNN) is adopted to build a sEMG-hand grasp force estimation model. Experiments are conducted in the three cases: (1) pooling training data from all muscle fatigue states with time-domain feature only, (2) employing frequency domain feature for expression of muscle fatigue information based on case 1, and 3) incorporating the quantitative metric of muscle fatigue value as an additional input for estimation model based on case 1. The results show that the degree of muscle fatigue and task intensity can be easily distinguished, and the additional input of muscle fatigue in BPNN greatly improves the performance of hand grasp force estimation, which is reflected by the 6.3797% increase in (coefficient of determination) value. 1. Introduction force simultaneously for a myoelectric hand [1]. Kim et al. obtained grasp force through upper limb forearm sEMG to control a teleoperation system in real-time [2]. Peternel Surface electromyography (sEMG) is the recording of myo- electric signals of muscle fiber contraction captured by elec- et al. proposed a muscle fatigue-based method for human- robot collaboration, by which the robot’s physical behaviour trodes attached on the surface skin. Due to this electrical manifestation, sEMG has the ability to represent the muscle can be adapted online to human motor fatigue [3]. It should be noticed that the effectiveness and robustness of these activation level and contains rich information of muscle applications are depended on the validation of the sEMG- force. This ability is widely applied in the accurate estimation of human joint moment which holds significant importance based force estimation which is highly affected by the proper- ties of sEMG signals. for robot control system design. Human hand grasp force estimation is one of the compelling applications among all The relationship between sEMG signals and muscle force is mostly extracted by either machine learning-based method of these implementations. The manipulability and dexterity or model-based method. Machine learning methods, such as of prosthetic hands, human-assisting devices, and telerobots are facilitated by grasp force estimation. Yamanoi et al. used artificial neural network [4] and support vector machines [5], enable the direct mapping from sEMG signals to desired sEMG signals to determine hand posture and estimate grip 2 Applied Bionics and Biomechanics itive difficulty brought by this property is seen in the esti- force estimation. The model-based method takes advantage of the musculoskeletal dynamics and incorporates the mation of muscle fatigue using solely frequency metric- human knowledge of physiology and motor functionality based sEMG signal representation. Thus, a more indicative metric is desired to function robustly during the lasting in an explicit way [6]. The disadvantages of the model- based method are that a general musculoskeletal system fatigue. The definition of fatigue as any reduction in the modeling for force estimation is still missing which is maximal capacity to generate force [12] allows the adop- attributed to the unknown properties, and the correspond- tion of the loss of maximal voluntary contraction (MVC) ing parameters are inherently difficult to identify. Machine to estimate muscle fatigue, where the degree of muscle fatigue is represented by the variant exerted force which learning-based methods mitigate the gap with a compro- mised yet acceptable interpretability. Among all the is relatively accurate to be measured by additional tangible modalities, sEMG together with classic regression models sensors. has been mostly investigated. Naeem et al. estimated joint Muscle fatigue has to be taken into account in order to force from EMG signals based on a back-propagation neu- acquire accurate grasp force from sEMG signals. However, so far nobody has been able to explain the relationship ral network (BPNN) [7]. Yang et al. compared different pattern regression methods to optimize the relationship between muscle fatigue and sEMG’s time-domain features. between sEMG signals and hand grasp force [8]. Zhang Even the conclusions of some studies are completely oppo- et al. used linear discriminant analysis (LDA) to realize site. In this paper, we propose an algorithm to quantitatively pattern recognition and artificial neural networks (ANN) estimate the degree of muscle fatigue and evaluate the results by three distinct methods. The substantial effect of muscle to establish the relationship between sEMG signals and fingertip force in each hand grasp modes [9]. fatigue on the performance of hand grasp force estimation Most current research is confined within the improve- is preliminarily demonstrated with experiments on 10 ment of accuracy and reliability for sEMG-based grasp force healthy subjects. As the muscle fatigue detection and grasp estimation through a single optimisation of regression algo- force estimation are improved, we believe that current applications such as presented in [1–3] will be benefited from rithms in a laboratory environment instead of a clinical scenario. And the practical factors in clinical settings such our proposed method. as fatigue, sweating, and electrode shift are normally ignored [10]. As one of the most critical factors, muscle fatigue influ- 2. Forearm Muscle Fatigue Evaluation ences the force estimation to a large extent in sEMG-based Based on the fact that muscle force will decline steadily dur- applications [11]. In daily activities, muscle fatigue leads to ing a sustained maximal contraction as shown in Figure 1, it failure of force generation to a required value at a normal is straight to adopt MFL as the index for evaluation of muscle muscle activation level [12]. When a muscle becomes fatigue. In this section, the definition of the proposed force- fatigued, the amplitude-related features of its sEMG signals based metric is given with an emphasis on the case of static are notably affected [13]. A typical example is that the root contraction for application. mean square (RMS) of sEMG increases when muscle fatigue happens. In grasp force estimation, RMS is the main feature 2.1. Maximum Force Loss (MFL). The proposed method to adopted for EMG-force regression. As a result, the perfor- estimate muscle fatigue depends on the measurement of mance of the pretrained force prediction model deteriorates, maximal voluntary contraction, which is performed by exert- which is attributed to the unstable RMS representation of ing maximum hand grasp force. To acquire reliable contrac- sEMG signals. It has been demonstrated that the variant of tion measurement, an easy-to-implement protocol is amplitude-based representation of sEMG-like multiscale designed in this paper. At the beginning of a measurement RMS (MRMS) gets almost doubled under fatigue condition session, the maximum force value is exerted by the subjects in a laboratory environment [14]. It is reasonable to incorpo- and recorded as MVC . After repetition of predefined types rate muscle fatigue in sEMG-based grasp force estimation of static contraction, the force value is recorded for multiple instead of solely depending on the plausible consistency of trials as MVC . The maximum hand grasp force, as shown sEMG signals. in Figure 2, will decrease over contraction tasks and reflects Frequency domain-based method is mostly explored to the remained muscle force capacity at the end of each trial. estimate muscle fatigue from sEMG signals by the analysis The termination of a session is determined by the failure to of mean frequency (MNF) or median frequency (MDF) accommodate the exertion of required force which indicates [15]. A general conclusion summarises the decreasing shift that the muscle is too fatigued to accomplish contraction of MNF or MDF along with the increase of muscle fatigue tasks. The required force value is recorded as MVC . MFL is [15]. Xie et al. applied MNF derived via Hibert-Huang finally defined as the following: transform to analyse fatigue sEMG signals [16]. Fernando et al. used the ratio of MNF to average rectified value MFL = MVC − MVC , ð1Þ (ARV) as the index of muscle fatigue and muscle fatigue i t is detected when MNF/ARV falls below a predetermined baseline [13]. Despite the promising results shown by the where MVC and MVC correspond to the initial and current i t transition between nonfatigue and fatigue status, the fre- MVC force. quency domain metrics exhibit without a determined To eliminate individual differences, the ratio of the vari- trend of shifting during singly-fatigue status [17]. An intu- ant maximum hand grasp force to the initial value is adopted Applied Bionics and Biomechanics 3 100 incorporate the influence of required force on muscle fatigue and MFL can be further redefined as MVC − MVC i t MFL = , ð3Þ MVC − MVC i f where MVC corresponds to the MVC force in the exhausted condition. The proposed muscle fatigue metric MFL can vary from 0 to 1 where 0 indicates the nonfatigue condition and 1 indicates the exhausted condition during static contraction tasks. Equation (3) is adopted together with the assumption of static contraction to estimate muscle fatigue in the follow- ing sections. 0 10 20 30 40 50 60 3. Hand Grasp Force Estimation Contraction time (s) In this preliminary study, back-propagation neural network Figure 1: MVC changes during a sustained maximal contraction (BPNN) is adopted to build sEMG based hand grasp force [12]. estimation model. 3.1. Experimental Protocol. Ten subjects (seven males and three females, mass 61:1±3 kg, height 1:70 ± 0:03 m, all right-handed) have been recruited in the experiment study. MVC MFL The subjects gave written informed consent before the exper- iment, and the study was approved by the ethics committee of Wuhan University of Technology. The experiment is con- MVC ducted with solely nondominant hands, i.e., the left hands, of our recruited subjects, where muscles are more prone to MVC fatigue during the measurement session [18]. The subjects are asked first to seat in a comfortable position with their forearm rest on the table. The sleeve with sEMG electrodes embedded is worn on the subject’s forearm with appropriate 150 fixation to avoid the electrode shifting. A hand-muscle devel- oper is held by the subjects’ nondominant hand. A pressure sensor is attached to the hand-muscle developer for the mea- surement of grasp force. With the forearm muscle initially at rest, the captured sEMG with an amplitude at 0 uV is secured prior to the measurement session. Then, the subject is asked 0 5 10 15 20 25 30 35 40 45 50 to hold the hand-muscle developer in the nondominant Contraction times hand, with the chair height subsequently adjusted to form an obtuse angle between the forearm equipped with sensor Figure 2: The measured maximum hand grasp force during a and the upper arm (shown in Figure 3). The sEMG signals measurement session. are easily interfered by cable movements or the surface elec- trodes relative movement caused by sleeve slipping ground the forearm. Thus, the subject is required to maintain his as the index of the degree of forearm muscle fatigue, defined posture as much as possible throughout the session to reduce as these artefacts. There are three sessions for one subject to perform: MVC − MVC i t named 50%, 60%, and 70% session. At the beginning of one MFL = : ð2Þ MVC i session, the subject is instructed by visual hints to conduct a 5-second hand grasps at MVC by exerting maximum hand 2.2. Case Study of Static Contraction. The definition given in grasp force with the hand-muscle developer, and the force is the previous section indicates the importance of required recorded as MVC . Then, a 10-minute rest is provided. After force in forearm muscle fatigue estimation. The greater the the break, the subject is asked to perform a hand grasp with a required force becomes, the less contraction time to maintain muscle contraction at x% MVC (x = 50, 60, 70, according to the required force level lasts and the muscle is easier to fall which session is performed) as steadily as possible for 10 sec- into fatigue state. In a case study where subjects perform onds. This grasp force is recorded as MVC . Then, a 5-second static contraction tasks by maintaining the required hand grasp at MVC is performed immediately, without a rest, and grasp force level as steadily as possible, it is necessary to the maximum hand grasp force is recorded as MVC . After Force (% control MVC) Hand grasp force (N) 4 Applied Bionics and Biomechanics Subject’s non-dominant hand ELONXI hardware box Electrode Pressure sensor Hand-muscle developer 16-channel electrode-embedded sleeve Figure 3: Experimental setup. that, another loop of 10-second x% MVC steady contraction signals and force measurements are captured and synchro- and 5-second MVC contraction is performed and repeated nized simultaneously during the experiment. multiple times continuously, without a break until MVC falls 3.3. Data Processing. The relation between sEMG and force below MVC . Then, one session is finished and a 30 minutes signals is extracted in an offline scheme. Two Sallen-Key fil- rest is given for the purpose of recovering from the muscle ters are employed to band-pass filter raw sEMG signals at a fatigue and preparing for the next session. The subject is pro- bandwidth between 20 Hz and 500 Hz. In addition, a notch vided with visual hints throughout the experiments to ensure filter with central cut-off frequency at 50 Hz (UK power line their adaption to the force variance. The entire experimental frequency) is used to remove the power line interference. procedure of one session is shown in Figure 4. The sEMG signals of each channel are segmented by the overlapped windowing technique [20] with a 300 ms window 3.2. Data Acquisition. In this study, three muscles closely and 100 ms window shift for feature extraction. In this study, related to hand grasp are selected to record the sEMG sig- RMS and MNF/ARV [21–23] are selected as sEMG features. nals, which are palmaris longus, flexor carpiulnaris, and Except for sEMG signal processing, the mean value of hand extensor digitorum. A 16-channel electrode-embedded grasp force data is adopted in each analysis window. sleeve (ELONXI, UK) is used to cover the aforementioned 3.4. Force Estimation Methods. BPNN is used to learn the forearm muscles to collect the sEMG signals where pal- association between sEMG signals and hand grasp forces. maris longus, flexor carpiulnaris, and extensor digitorum In order to evaluate the effect of muscle fatigue on hand grasp mainly correspond to 1-channel electrode, 3-channel elec- force exertion, we propose the following three methods and trode, and 5-channel electrode, respectively, as shown in compare them with locally acquired experiment data. Figure 5. The reference electrode is at the subject’s wrist. Time-domain feature driving machine learning-based Before wearing the electrode sleeve, the skin is cleaned method (TMLM, as shown in Figure 6): train the BPNN with by alcohol, and a 10-minute-rest is given after the elec- pooled training data from all muscle fatigue status together. trode attachment to improve the contact of the electrode The inputs of BPNN are three muscles’ sEMG feature RMS, with skin to reduce the resistance within [19]. sEMG sig- forming the feature vector nals are amplified by a factor of 5000 with linear range 20 Hz to 500 Hz and sampled at 1000 Hz. The FingerTPS ½ RMS , ð4Þ system (Pressure Profile Systems, Inc. (PPS), USA), origi- i×n nally utilised for capturing the tactile force on the finger pulp, is used to measure the hand grasp force in the where i is the channel and n is the number of window shift. experiment. Since the finger pulp is not the optimum pres- And the output is the measured hand grasp force. All the data sure point during hand grasp, a highly sensitive capacitive- acquired under three distinct hand grasp force levels are based pressure sensor is fixed to the appointed position on formed as the training/testing data for the BPNN. hand-muscle developer (shown in Figure 3). The sample Combined feature driving machine learning-based frequency is 100 Hz controlled by the PC clock. The sEMG method (CMLM, as shown in Figure 7): train the BPNN with Applied Bionics and Biomechanics 5 Start 10-second hand grasp at MVC 5-second hand grasp at MVC 5-second hand grasp at MVC Record MVC Record MVC 10 minutes rest MVC > MVC t f MVC = X % MVC 30-minute rest prepare for the next session Figure 4: Diagram of the experimental procedure of one session. Extensor digitorum (back) This method is identical with the above method in output Palm of non-dominant hand and selection of training/testing data. Fatigue feature driving machine learning method (FMLM, as shown in Figure 8): train the BPNN with estimated muscle fatigue value as an additional attribute. An additional input of the degree of muscle fatigue estimated by using (3) in combination with the RMS features is provided to the BPNN and expressed as ½ RMS ,MFL , ð6Þ i×n i×n Flexor carpiulnaris (front) Palmaris longus (front) where i is the channel and n is the number of window shift. Figure 5: Diagram of the experimental procedure. The output and selection method of training/testing data remain the same for all methods, as shown in the following combination of time domain and frequency domain features. three figures. MNF is often employed for the expression of muscle fatigue All methods adopt the BPNN architecture for force esti- information in sEMG-based force estimation. And Japanese mation, whose performance is dependent on the choice of researchers further proposed MNF/ARV, which has achieved network structure, training data, and testing data. The net- good results in muscle fatigue detection [13]. So, a combined work structure is adjusted by setting different number of feature vector is given as nodes from 2 to 20 in the hidden layer with the optimal results [24] provided by a three-layer BPNN. And a Log- Sigmoid function is selected as the transfer function in the ðÞ RMS ,ðÞ MNF /ARV i i i n n network. N 2 ð5Þ ∑ ðÞ F − F∧ 2 k k k=1 R =1 − , N Sx = , ð7Þ ðÞ −x ∑ F − F k=1 k k 1+ e where x is the input and e is the exponential function. More- where i is the channel and n is the number of window shift. over, a four-fold cross validation is adopted to avoid random F denotes the actual hand grasp force, F is the predicted classification of training data and test data from affecting the k k hand grasp force, F is the average of actual hand grasp force, prediction results, which helps ensure the reliability and and N is the number of testing data. stability of the model. 6 Applied Bionics and Biomechanics Output sEMG signals BPNN BPNN Training data parameters training Inputs BPNN RMS Data acquisition Data processing flexor pollicis longus preprocessing RMS Hand grasp force palmaris longus signal segmentation RMS carpiulnaris feature extraction Figure 6: Flowchart of grasp force estimation—TMLM. Output sEMG signals BPNN BPNN Training data parameters training Inputs BPNN RMS Data acquisition Data processing flexor pollicis longus preprocessing RMS Hand grasp force palmaris longus signal segmentation RMS carpiulnaris feature extraction MNF ARV MNF Combined feature ARV extraction MNF ARV Figure 7: Flowchart of grasp force estimation—CMLM. Output sEMG signals BPNN BPNN Training data parameters training Inputs BPNN RMS Data acquisition Data processing RMS flexor pollicis longus preprocessing Hand grasp force RMS palmaris longus signal segmentation carpiulnaris feature extraction MFL Maximum hand Muscle fatigue grasp force estimation Figure 8: Flowchart of grasp force estimation—FMLM. Applied Bionics and Biomechanics 7 sEMG signals of palmaris longus sEMG signals of flexor carpiulnaris sEMG signals of extensor digitorum 400 800 100 200 0 0 –100 –200 –100 –200 –400 –200 –300 –600 –300 –400 –800 –400 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5 5 5 ×10 ×10 ×10 Sampling points Sampling points Sampling points 50% MVC 60% MVC 70% MVC (a) RMS of palmaris longus RMS of flexor carpiulnaris RMS of extensor digitorum 110 90 80 160 60 50 40 80 10 20 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Window shift times Window shift times Window shift times 50% MVC 60% MVC 70% MVC (b) MNF/ARV of palmaris longus MNF/ARV of flexor carpiulnaris MNF/ARV of extensor digitorum 18 18 10 16 16 14 14 𝜇 𝜇 7 12 12 10 10 8 8 6 6 4 4 2 2 1 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Window shift times Window shift times Window shift times 50% MVC 60% MVC 70% MVC (c) Figure 9: Continued. RMS (V) Filtered skin surface voltage (V) MNV/ARV (Hz/V) MNV/ARV (Hz/V) Filtered skin surface voltage (V) RMS (V) MNV/ARV (Hz/V) RMS (V) Filtered skin surface voltage (V) 8 Applied Bionics and Biomechanics Actual hand grasp force of each sample Actual maximum hand grasp force of each sample Degree of muscle fatigue of each sample 36 50 1 0.9 0.8 0.7 40 0.6 30 0.5 0.4 0.3 30 0.2 0.1 24 25 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 Sample number Sample number Sample number 50% MVC 60% MVC 70% MVC (d) Figure 9: Data of static contraction tasks. (a) sEMG signals of three muscles. (b) RMS of three muscles. (c) MNF/ARV of three muscles (d) actual hand grasp force, maximum hand grasp force, and MFL. 1 1 1 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0 0 0 0 100 200 300 0 100 200 300 0 50 100 150 Time (s) Time (s) Time (s) 50% MVC 60% MVC 70% MVC i i i Figure 10: Muscle fatigue estimation results. Solid lines are linear fitting of the estimation. Table 1: Gradient of MFL. In this study, R is used to evaluate the estimation perfor- MVC 50% MVC 60% MVC 70% MVC mance of three methods, which can be expressed as f i i i Gradient 0.0077 0.0108 0.0532 N 2 ∑ F − F∧ ðÞ 2 i=1 i i 4. Results and Discussion R =1 − : ð8Þ ∑ F − F i=1 i i In this paper, an algorithm to quantitatively estimate the degree of muscle fatigue is introduced. And the effect of mus- The R can be comprehended as the percentage of the cle fatigue on hand grasp force estimation is evaluated by response variable variation that is explained by a linear conducting three distinct comparison methods. The experi- model [25] and ranges from 0 to 1. In general, the higher mental results of one subject are shown in Figure 9. They the R , the better the model fits the data. are sEMG signals, RMS, and MNF/ARV of three muscles in T-tests were used to verify differences in TMLM, CMLM, different levels of static contraction tasks, actual hand grasp and FMLM between different conditions. Differences among force, maximum hand grasp force, and MFL of each sample subjects are not considered in this paper, as muscle-level in different levels of static contraction tasks. These selected dynamic variation is commonly existed. p <0:05 is consid- features, seen in Figure 10, can basically reflect the force ered statistically significant for all tests. and muscle fatigue information. Hand grasp force (kg) MFL (100%) Maximum hand grasp force (kg) MFL (100%) MFL (100%) MFL (100%) Applied Bionics and Biomechanics 9 Table 2: R (mean ± sd) of predictions in TMLM. Number of nodes 2 3456 0:6530 ± 0:0314 0:7037 ± 0:0037 0:9093 ± 0:0205 0:6933 ± 0:0083 0:4682 ± 0:0949 R (mean ± sd) Number of nodes 7 8 9 10 11 0:8328 ± 0:0095 0:8201 ± 0:0013 0:7978 ± 0:0311 0:7738 ± 0:0183 0:8610 ± 0:0406 R (mean ± sd) Number of nodes 12 13 14 15 16 0:8259 ± 0:0007 0:7103 ± 0:0074 0:7561 ± 0:0273 0:6653 ± 0:0170 0:6531 ± 0:0713 R (mean ± sd) Number of nodes 17 18 19 20 0:7817 ± 0:0211 0:8261 ± 0:0021 0:8156 ± 0:0076 0:8033 ± 0:0145 R (mean ± sd) Table 3: R (mean ± sd) of predictions in CMLM. Number of nodes 2 3456 0:7321 ± 0:0599 0:8357 ± 0:0081 0:7897 ± 0:0276 0:9255 ± 0:0063 0:6594 ± 0:1223 R (mean ± sd) Number of nodes 7 8 9 10 11 0:7623 ± 0:0132 0:6229 ± 0:0818 0:6648 ± 0:0049 0:7149 ± 0:0563 0:7281 ± 0:0218 R (mean ± sd) Number of nodes 12 13 14 15 16 0:7579 ± 0:0357 0:7932 ± 0:0026 0:5954 ± 0:0214 0:4769 ± 0:1373 0:8362 ± 0:0035 R (mean ± sd) Number of nodes 17 18 19 20 0:7074 ± 0:0104 0:6305 ± 0:0201 0:5928 ± 0:0450 0:7207 ± 0:0422 R (mean ± sd) Table 4: R (mean ± sd) of predictions in FMLM. Number of nodes 2 3456 0:8158 ± 0:0096 0:8663 ± 0:0220 0:8425 ± 0:0055 0:9193 ± 0:0185 0:8356 ± 0:0149 R (mean ± sd) Number of nodes 7 8 9 10 11 0:8343 ± 0:0205 0:8795 ± 0:0016 0:8746 ± 0:0127 0:9572 ± 0:0030 0:8842 ± 0:0122 R (mean ± sd) Number of nodes 12 13 14 15 16 0:8312 ± 0:0474 0:8652 ± 0:0018 0:8892 ± 0:0017 0:7465 ± 0:1342 0:7358 ± 0:0545 R (mean ± sd) Number of nodes 17 18 19 20 0:8272 ± 0:0143 0:8100 ± 0:0063 0:7836 ± 0:0054 0:8472 ± 0:0010 R (mean ± sd) Table 5: Prediction results of three different methods. Method TMLM CMLM FMLM Number of nodes 4 5 10 0:8782 ± 0:0005 0:9065 ± 0:0011 0:9506 ± 0:0009 R (mean ± sd) 4.1. Results of Muscle Fatigue Estimation. The experiments levels of static contraction. The greater the required force last for 300 s, 240 s, and 50 s corresponding to 50% MVC becomes in static contraction, the faster MFL rises, shown (250 N), 60% MVC (300 N), and 70% MVC (350 N kg) in in Table 1, which implies that the task intensity can also be i i distinguished through the proposed metric. The feasibility static contraction tasks, respectively. Figure 10 shows the estimation results of muscle fatigue by the proposed method. of the proposed method is recognized in static contraction Dot arrays of different colors represent the estimation results tasks to estimate muscle fatigue quantitatively. at different force levels. Through linear fitting, it can be directly seen that MFL grows linearly with the increase of 4.2. Results of Hand Grasp Force Estimation. In order to weaken effect of network structure, initial weights, and bias contraction time, which is in accordance with Vøllestad’s assumption [12] of muscle fatigue’s variety law during a sus- values on the estimation performance, the neural network tained and steady contraction. In addition, the results show is retrained ten times at different numbers of nodes (from 2 that the gradient of time-varying MFL varies at different to 20) in the hidden layer. 10 Applied Bionics and Biomechanics Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison 2 2 2 2 R = 0.8818 R = 0.90022 R = 0.91813 R = 0.93697 36 36 36 40 34 34 32 32 30 30 32 28 28 26 26 24 22 24 24 0 1020304050 0 1020304050 0 1020304050 0 1020304050 Prediction samples Prediction samples Prediction samples Prediction samples Predicted value (50% MVC ) True value (50% MVC ) i i Predicted value (60% MVC ) True value (60% MVC ) Predicted value (70% MVC ) True value (70% MVC ) (a) Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison 2 2 2 2 R = 0.91705 R = 0.92491 R = 0.92531 R = 0.9347 36 36 36 36 34 34 34 32 32 32 30 30 30 28 28 28 26 26 26 24 24 24 22 22 24 22 0 1020304050 0 1020304050 0 1020304050 0 10 20 30 40 50 Prediction samples Prediction samples Prediction samples Prediction samples True value (50% MVC ) Predicted value (50% MVC ) i i True value (60% MVC ) Predicted value (60% MVC ) i i True value (70% MVC ) Predicted value (70% MVC ) i i (b) Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison 2 2 2 2 R = 0.95392 R = 0.95569 R = 0.95711 R = 0.96203 36 36 36 36 34 34 34 32 32 32 30 30 30 28 28 28 26 26 26 24 24 24 22 0 1020304050 0 1020304050 0 1020304050 0 10 20 30 40 50 Prediction samples Prediction samples Prediction samples Prediction samples True value (50% MVC ) Predicted value (50% MVC ) True value (60% MVC ) Predicted value (60% MVC ) True value (70% MVC ) Predicted value (70% MVC ) i i (c) Figure 11: Single four-fold cross-validation results of each method (a) TMLM. (b) CMLM. (c) FMLM. In TMLM, we pool training data from all muscle fatigue BPNN are shown in Table 4. The network structure of 10 states to make the network learn the differences among them nodes in the hidden layer brings the maximum mean R . alone. Predictions of BPNN are shown in Table 2. When the Its value is 0.9572. number of node is 4, the mean R is 0.9093, which is the Comparing the best prediction results of three different maximum. methods, as shown in Table 5 and Figure 11, it can be In CMLM, we employ MNF/ARV, one feature proposed indicated that the mean R obtained in TMLM is 0.9093. by Fernando’s team for the expression of muscle fatigue It just passes the baseline of applicability (0.9000), which information. Predictions of BPNN are shown in Table 3. implies the estimation performance of the BPNN model We set 5 nodes in the hidden layer, and the mean R of pre- in TMLM is not good enough and predicting model need diction results is 0.9255. to be readjusted. In CMLM, the mean R is 0.9255. This In FMLM, we incorporate the quantitative metric of mus- shows that employing MNF/ARV proposed by Fernando cle fatigue value as an additional input to explain the effect of et al. [13] in sEMG-based force estimation under fatigued muscle fatigue on hand grasp force estimation. Predictions of conditions is indeed feasible. But it is not an obviously Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Applied Bionics and Biomechanics 11 measure the subject’s current maximum grasp force. As a Table 6: R (mean ± sd) of predictions of all subjects under three methods. result, the force estimation in this work could only be proc- essed offline. So the future work is mainly to address how Method TMLM CMLM FMLM to estimate muscle fatigue online, that is, how to get MFL 0:9093 ± 0:0205 0:9255 ± 0:0063 0:9572 ± 0:0030 Subject 1 online. In fact, the results of this study have provided some potential and guiding ideas for the following work. Under 0:8910 ± 0:0092 0:9210 ± 0:0039 0:9548 ± 0:0009 Subject 2 static muscle contraction, the subject’s forearm muscle 0:8255 ± 0:0043 0:8517 ± 0:0064 0:8938 ± 0:0189 Subject 3 fatigue and muscle contraction time are approximately linear 0:8224 ± 0:0042 0:8464 ± 0:0049 0:8805 ± 0:0028 Subject 4 when maintaining a fixed level of hand grasp force. And this 0:8286 ± 0:0012 0:8571 ± 0:0014 0:9070 ± 0:0190 linear coefficient seems to have a nonlinear increasing rela- Subject 5 tionship with the target hand grasp force level. Therefore, a 0:9293 ± 0:0132 0:9411 ± 0:0209 0:9554 ± 0:0182 Subject 6 nonlinear estimation model of muscle fatigue could be more 0:9199 ± 0:0126 0:9262 ± 0:0050 0:9426 ± 0:0181 Subject 7 appropriated in this case, such as 0:8802 ± 0:0431 0:9359 ± 0:0088 0:9438 ± 0:0088 Subject 8 b×n%MVC 0:8153 ± 0:0268 0:8526 ± 0:0131 0:8753 ± 0:0267 Subject 9 MFL = a × e × t + c, ð9Þ 0:8350 ± 0:0097 0:8620 ± 0:0130 0:8983 ± 0:0121 Subject 10 where n%MVC is the target hand grasp force level. t is mus- cle contraction time. a, b, and c are model parameters. effective approach. Compared with the result in TMLM, R increases by 1.7816%. Data Availability As the main work of this study, the estimated muscle fatigue value is used directly as an additional input in FMLM. The EMG and force data used to support the findings of this The results show mean R can reach to 0.9572, which proves study are available from the corresponding author upon predicting model fits the data very well. It is a great improve- request. ment (5.2678%, above 5%, p <0:05) in estimation perfor- mance compared with CMLM. It is demonstrated that the Conflicts of Interest additional attribute is an applicable solution to the effect of muscle fatigue on sEMG-based hand grasp force estimation. The author(s) declare(s) that they have no conflicts of And MFL proposed in this paper is better than MNF/ARV interest. (p <0:05). For further explanation, the experimental results of all Acknowledgments subjects under different methods are presented as shown in Table 6. Statistics show that the mean R The authors would like to extend their gratitude to H. Liu, Q. values obtained under the three methods are 0.8656, 0.8919, and 0.9209. Gao and C. Li from Intelligent System & Biomedical Robotics Adopting MNF/ARV proposed in [12] for measure muscle Group, School of Computing, University of Portsmouth for fatigue could bring the 3.0383% growth in R in hand grasp assisting in the experimental process. This work was sup- ported in part by the National Natural Science Foundation force estimation. For comparison, using the MFL proposed of China under Grant No. 61603284, 61903286, and in this paper can increase R by 6.3797%. 52075530 and the AiBle project co-financed by the European The experimental results show that FMLM provides the Regional Development Fund. best estimation performance among the three methods. References 5. Conclusion [1] Y. Yamanoi, S. Morishita, R. Kato, and H. Yokoi, “Develop- In this paper, we propose an easy-to-implement method to ment of myoelectric hand that determines hand posture and quantitatively estimate muscle fatigue and evaluate the effect estimates grip force simultaneously,” Biomedical Signal Pro- of muscle fatigue on hand grasp force estimation. The exper- cessing and Control, vol. 38, pp. 312–321, 2017. iment results demonstrate that the incorporation of muscle [2] M. Kim, J. Lee, and K. Kim, “Tele-operation system with reli- fatigue metrics explicitly in the grasp force estimation has a able grasping force estimation to compensate for the time- substantial impact on the performance. When estimated varying sEMG feature,” in 2016 IEEE International Conference muscle fatigue value as an additional input in the machine on Robotics and Automation (ICRA), pp. 5561–5567, Stock- learning approach, the estimation accuracy improves to a holm, Sweden, 2016. large extent in FMLM. Because the neural network is trained [3] L. Peternel, N. Tsagarakis, D. Caldwell, and A. Ajoudani, offline and is computationally cheap, the proposed method “Robot adaptation to human physical fatigue in human–robot can be implemented in the current applications, such as in co-manipulation,” Autonomous Robots, vol. 42, no. 5, [1–3], as a calibration part to improve the effectiveness and pp. 1011–1021, 2018. robustness. At the same time, there are still some limitations [4] C. Choi, S. Kwon, W. Park, H. D. Lee, and J. Kim, “Real-time in this study. During the experiments mentioned above, the pinch force estimation by surface electromyography using an degree of the subject’s forearm muscle fatigue needs to be artificial neural network,” Medical Engineering & Physics, estimated at each moment, which requires to intermittently vol. 32, no. 5, pp. 429–436, 2010. 12 Applied Bionics and Biomechanics Biology Society, pp. 1449-1450, Amsterdam, Netherlands, [5] C. Castellini, E. Gruppioni, A. Davalli, and G. Sandini, “Fine detection of grasp force and posture by amputees via surface 1996. electromyography,” Journal of Physiology-Paris, vol. 103, [19] C. J. De Luca, “Use of the surface EMG signal for performance no. 3-5, pp. 255–262, 2009. evaluation of back muscles,” Muscle & Nerve, vol. 16, no. 2, [6] A. Erdemir, S. McLean, W. Herzog, and A. J. van den Bogert, pp. 210–216, 1993. “Model-based estimation of muscle forces exerted during [20] Y. Fang and H. Liu, “Robust sEMG electrodes configuration movements,” Clinical biomechanics, vol. 22, no. 2, pp. 131– for pattern recognition based prosthesis control,” in 2014 IEEE 154, 2007. International Conference on Systems, Man, and Cybernetics [7] U. J. Naeem, C. Xiong, and A. A. Abdullah, “EMG-muscle (SMC), pp. 2210–2215, San Diego, CA, USA, 2014. force estimation model based on back-propagation neural net- [21] N. A. Shrirao, N. P. Reddy, and D. R. Kosuri, “Neural network work,” in 2012 IEEE International Conference on Virtual Envi- committees for finger joint angle estimation from surface ronments Human-Computer Interfaces and Measurement EMG signals,” Biomedical Engineering Online, vol. 8, no. 1, Systems (VECIMS) Proceedings, pp. 222–227, Tianjin, China, p. 2, 2009. [22] S. Suryanarayanan, N. P. Reddy, and V. Gupta, “Artificial neu- [8] D. Yang, J. Zhao, Y. Gu, L. Jiang, and H. Liu, “Estimation of ral networks for estimation of joint angle from EMG signals,” hand grasp force based on forearm surface EMG,” in 2009 in Proceedings of 17th International Conference of the Engi- International Conference on Mechatronics and Automation, neering in Medicine and Biology Society, pp. 823-824, Mon- pp. 1795–1799, Changchun, China, 2009. treal, Quebec, Canada, 1995. [9] B. Zhang and S. Zhang, “Pattern-based grasping force estima- [23] Y. Koike and M. Kawato, “Estimation of dynamic joint torques tion from surface electromyography,” in 2017 International and trajectory formation from surface electromyography sig- Conference on Algorithms, Methodology, Models and Applica- nals using a neural network model,” Biological Cybernetics, tions in Emerging Technologies (ICAMMAET), pp. 1–5, Chen- vol. 73, no. 4, pp. 291–300, 1995. nai, India, 2017. [24] O. Fujita, “Statistical estimation of the number of hidden units [10] N. Jiang, S. Dosen, K.-R. Muller, and D. Farina, “Myoelectric for feedforward neural networks,” Neural Networks, vol. 11, control of artificial limbs—is there a need to change focus? no. 5, pp. 851–859, 1998. [in the spotlight],” IEEE Signal Processing Magazine, vol. 29, no. 5, 2012. [25] C. Hagquist and M. Stenbeck, “Goodness of fit in regression analysis – R2 and G2 reconsidered,” Quality and Quantity, [11] T. D. Lalitharatne, Y. Hayashi, K. Teramoto, and K. Kiguchi, vol. 32, no. 3, pp. 229–245, 1998. “Compensation of the effects of muscle fatigue on EMG- based control using fuzzy rules based scheme,” in 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 6949–6952, Osaka, Japan, 2013. [12] N. K. Vøllestad, “Measurement of human muscle fatigue,” Journal of Neuroscience Methods, vol. 74, no. 2, pp. 219–227, [13] J. B. Fernando, M. Yoshioka, and J. Ozawa, “Estimation of muscle fatigue by ratio of mean frequency to average rectified value from surface electromyography,” in 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 5303–5306, Orlando, FL, USA, 2016. [14] M. Navaneethakrishna and S. Ramakrishnan, “Multiscale fea- ture based analysis of surface EMG signals under fatigue and non-fatigue conditions,” in 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 4627–4630, Chicago, IL, USA, 2014. [15] Y. Masakado, Y. Noda, K. Hase, A. Kimura, and N. Chino, “Muscle fatigue assessed by frequency analysis of surface elec- tromyographic signals: topographical analysis in the same muscle,” The Japanese Journal of Rehabilitation Medicine, vol. 31, no. 6, pp. 409–414, 1994. [16] H. Xie and Z. Wang, “Mean frequency derived via Hilbert- Huang transform with application to fatigue EMG signal anal- ysis,” Computer Methods and Programs in Biomedicine, vol. 82, no. 2, pp. 114–120, 2006. [17] T. Öberg, L. Sandsjö, and R. Kadefors, “Subjective and objec- tive evaluation of shoulder muscle fatigue,” Ergonomics, vol. 37, no. 8, pp. 1323–1333, 1994. [18] H. Oka, “Estimation of muscle fatigue by using EMG and muscle stiffness,” in Proceedings of 18th Annual Interna- tional Conference of the IEEE Engineering in Medicine and http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Applied Bionics and Biomechanics Hindawi Publishing Corporation

Effect of Muscle Fatigue on Surface Electromyography-Based Hand Grasp Force Estimation

Loading next page...
 
/lp/hindawi-publishing-corporation/effect-of-muscle-fatigue-on-surface-electromyography-based-hand-grasp-MkQTZHUzLO
Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2021 Jinfeng Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
1176-2322
eISSN
1754-2103
DOI
10.1155/2021/8817480
Publisher site
See Article on Publisher Site

Abstract

Hindawi Applied Bionics and Biomechanics Volume 2021, Article ID 8817480, 12 pages https://doi.org/10.1155/2021/8817480 Research Article Effect of Muscle Fatigue on Surface Electromyography-Based Hand Grasp Force Estimation 1 2 2 2 2 3 Jinfeng Wang, Muye Pang , Peixuan Yu , Biwei Tang, Kui Xiang, and Zhaojie Ju Department of Information, Wuhan Huaxia University of Technology, 430223 Wuhan, China Intelligent System Research Institute, Wuhan University of Technology, 430070 Wuhan, China Intelligent System & Biomedical Robotics Group, University of Portsmouth, PO1 3HE Portsmouth, UK Correspondence should be addressed to Muye Pang; pangmuye@whut.edu.cn Received 19 May 2020; Revised 14 January 2021; Accepted 30 January 2021; Published 15 February 2021 Academic Editor: Dongming Gan Copyright © 2021 Jinfeng Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Surface electromyography- (sEMG-) based hand grasp force estimation plays an important role with a promising accuracy in a laboratory environment, yet hardly clinically applicable because of physiological changes and other factors. One of the critical factors is the muscle fatigue concomitant with daily activities which degrades the accuracy and reliability of force estimation from sEMG signals. Conventional qualitative measurements of muscle fatigue contribute to an improved force estimation model with limited progress. This paper proposes an easy-to-implement method to evaluate the muscle fatigue quantitatively and demonstrates that the proposed metrics can have a substantial impact on improving the performance of hand grasp force estimation. Specifically, the reduction in the maximal capacity to generate force is used as the metric of muscle fatigue in combination with a back-propagation neural network (BPNN) is adopted to build a sEMG-hand grasp force estimation model. Experiments are conducted in the three cases: (1) pooling training data from all muscle fatigue states with time-domain feature only, (2) employing frequency domain feature for expression of muscle fatigue information based on case 1, and 3) incorporating the quantitative metric of muscle fatigue value as an additional input for estimation model based on case 1. The results show that the degree of muscle fatigue and task intensity can be easily distinguished, and the additional input of muscle fatigue in BPNN greatly improves the performance of hand grasp force estimation, which is reflected by the 6.3797% increase in (coefficient of determination) value. 1. Introduction force simultaneously for a myoelectric hand [1]. Kim et al. obtained grasp force through upper limb forearm sEMG to control a teleoperation system in real-time [2]. Peternel Surface electromyography (sEMG) is the recording of myo- electric signals of muscle fiber contraction captured by elec- et al. proposed a muscle fatigue-based method for human- robot collaboration, by which the robot’s physical behaviour trodes attached on the surface skin. Due to this electrical manifestation, sEMG has the ability to represent the muscle can be adapted online to human motor fatigue [3]. It should be noticed that the effectiveness and robustness of these activation level and contains rich information of muscle applications are depended on the validation of the sEMG- force. This ability is widely applied in the accurate estimation of human joint moment which holds significant importance based force estimation which is highly affected by the proper- ties of sEMG signals. for robot control system design. Human hand grasp force estimation is one of the compelling applications among all The relationship between sEMG signals and muscle force is mostly extracted by either machine learning-based method of these implementations. The manipulability and dexterity or model-based method. Machine learning methods, such as of prosthetic hands, human-assisting devices, and telerobots are facilitated by grasp force estimation. Yamanoi et al. used artificial neural network [4] and support vector machines [5], enable the direct mapping from sEMG signals to desired sEMG signals to determine hand posture and estimate grip 2 Applied Bionics and Biomechanics itive difficulty brought by this property is seen in the esti- force estimation. The model-based method takes advantage of the musculoskeletal dynamics and incorporates the mation of muscle fatigue using solely frequency metric- human knowledge of physiology and motor functionality based sEMG signal representation. Thus, a more indicative metric is desired to function robustly during the lasting in an explicit way [6]. The disadvantages of the model- based method are that a general musculoskeletal system fatigue. The definition of fatigue as any reduction in the modeling for force estimation is still missing which is maximal capacity to generate force [12] allows the adop- attributed to the unknown properties, and the correspond- tion of the loss of maximal voluntary contraction (MVC) ing parameters are inherently difficult to identify. Machine to estimate muscle fatigue, where the degree of muscle fatigue is represented by the variant exerted force which learning-based methods mitigate the gap with a compro- mised yet acceptable interpretability. Among all the is relatively accurate to be measured by additional tangible modalities, sEMG together with classic regression models sensors. has been mostly investigated. Naeem et al. estimated joint Muscle fatigue has to be taken into account in order to force from EMG signals based on a back-propagation neu- acquire accurate grasp force from sEMG signals. However, so far nobody has been able to explain the relationship ral network (BPNN) [7]. Yang et al. compared different pattern regression methods to optimize the relationship between muscle fatigue and sEMG’s time-domain features. between sEMG signals and hand grasp force [8]. Zhang Even the conclusions of some studies are completely oppo- et al. used linear discriminant analysis (LDA) to realize site. In this paper, we propose an algorithm to quantitatively pattern recognition and artificial neural networks (ANN) estimate the degree of muscle fatigue and evaluate the results by three distinct methods. The substantial effect of muscle to establish the relationship between sEMG signals and fingertip force in each hand grasp modes [9]. fatigue on the performance of hand grasp force estimation Most current research is confined within the improve- is preliminarily demonstrated with experiments on 10 ment of accuracy and reliability for sEMG-based grasp force healthy subjects. As the muscle fatigue detection and grasp estimation through a single optimisation of regression algo- force estimation are improved, we believe that current applications such as presented in [1–3] will be benefited from rithms in a laboratory environment instead of a clinical scenario. And the practical factors in clinical settings such our proposed method. as fatigue, sweating, and electrode shift are normally ignored [10]. As one of the most critical factors, muscle fatigue influ- 2. Forearm Muscle Fatigue Evaluation ences the force estimation to a large extent in sEMG-based Based on the fact that muscle force will decline steadily dur- applications [11]. In daily activities, muscle fatigue leads to ing a sustained maximal contraction as shown in Figure 1, it failure of force generation to a required value at a normal is straight to adopt MFL as the index for evaluation of muscle muscle activation level [12]. When a muscle becomes fatigue. In this section, the definition of the proposed force- fatigued, the amplitude-related features of its sEMG signals based metric is given with an emphasis on the case of static are notably affected [13]. A typical example is that the root contraction for application. mean square (RMS) of sEMG increases when muscle fatigue happens. In grasp force estimation, RMS is the main feature 2.1. Maximum Force Loss (MFL). The proposed method to adopted for EMG-force regression. As a result, the perfor- estimate muscle fatigue depends on the measurement of mance of the pretrained force prediction model deteriorates, maximal voluntary contraction, which is performed by exert- which is attributed to the unstable RMS representation of ing maximum hand grasp force. To acquire reliable contrac- sEMG signals. It has been demonstrated that the variant of tion measurement, an easy-to-implement protocol is amplitude-based representation of sEMG-like multiscale designed in this paper. At the beginning of a measurement RMS (MRMS) gets almost doubled under fatigue condition session, the maximum force value is exerted by the subjects in a laboratory environment [14]. It is reasonable to incorpo- and recorded as MVC . After repetition of predefined types rate muscle fatigue in sEMG-based grasp force estimation of static contraction, the force value is recorded for multiple instead of solely depending on the plausible consistency of trials as MVC . The maximum hand grasp force, as shown sEMG signals. in Figure 2, will decrease over contraction tasks and reflects Frequency domain-based method is mostly explored to the remained muscle force capacity at the end of each trial. estimate muscle fatigue from sEMG signals by the analysis The termination of a session is determined by the failure to of mean frequency (MNF) or median frequency (MDF) accommodate the exertion of required force which indicates [15]. A general conclusion summarises the decreasing shift that the muscle is too fatigued to accomplish contraction of MNF or MDF along with the increase of muscle fatigue tasks. The required force value is recorded as MVC . MFL is [15]. Xie et al. applied MNF derived via Hibert-Huang finally defined as the following: transform to analyse fatigue sEMG signals [16]. Fernando et al. used the ratio of MNF to average rectified value MFL = MVC − MVC , ð1Þ (ARV) as the index of muscle fatigue and muscle fatigue i t is detected when MNF/ARV falls below a predetermined baseline [13]. Despite the promising results shown by the where MVC and MVC correspond to the initial and current i t transition between nonfatigue and fatigue status, the fre- MVC force. quency domain metrics exhibit without a determined To eliminate individual differences, the ratio of the vari- trend of shifting during singly-fatigue status [17]. An intu- ant maximum hand grasp force to the initial value is adopted Applied Bionics and Biomechanics 3 100 incorporate the influence of required force on muscle fatigue and MFL can be further redefined as MVC − MVC i t MFL = , ð3Þ MVC − MVC i f where MVC corresponds to the MVC force in the exhausted condition. The proposed muscle fatigue metric MFL can vary from 0 to 1 where 0 indicates the nonfatigue condition and 1 indicates the exhausted condition during static contraction tasks. Equation (3) is adopted together with the assumption of static contraction to estimate muscle fatigue in the follow- ing sections. 0 10 20 30 40 50 60 3. Hand Grasp Force Estimation Contraction time (s) In this preliminary study, back-propagation neural network Figure 1: MVC changes during a sustained maximal contraction (BPNN) is adopted to build sEMG based hand grasp force [12]. estimation model. 3.1. Experimental Protocol. Ten subjects (seven males and three females, mass 61:1±3 kg, height 1:70 ± 0:03 m, all right-handed) have been recruited in the experiment study. MVC MFL The subjects gave written informed consent before the exper- iment, and the study was approved by the ethics committee of Wuhan University of Technology. The experiment is con- MVC ducted with solely nondominant hands, i.e., the left hands, of our recruited subjects, where muscles are more prone to MVC fatigue during the measurement session [18]. The subjects are asked first to seat in a comfortable position with their forearm rest on the table. The sleeve with sEMG electrodes embedded is worn on the subject’s forearm with appropriate 150 fixation to avoid the electrode shifting. A hand-muscle devel- oper is held by the subjects’ nondominant hand. A pressure sensor is attached to the hand-muscle developer for the mea- surement of grasp force. With the forearm muscle initially at rest, the captured sEMG with an amplitude at 0 uV is secured prior to the measurement session. Then, the subject is asked 0 5 10 15 20 25 30 35 40 45 50 to hold the hand-muscle developer in the nondominant Contraction times hand, with the chair height subsequently adjusted to form an obtuse angle between the forearm equipped with sensor Figure 2: The measured maximum hand grasp force during a and the upper arm (shown in Figure 3). The sEMG signals measurement session. are easily interfered by cable movements or the surface elec- trodes relative movement caused by sleeve slipping ground the forearm. Thus, the subject is required to maintain his as the index of the degree of forearm muscle fatigue, defined posture as much as possible throughout the session to reduce as these artefacts. There are three sessions for one subject to perform: MVC − MVC i t named 50%, 60%, and 70% session. At the beginning of one MFL = : ð2Þ MVC i session, the subject is instructed by visual hints to conduct a 5-second hand grasps at MVC by exerting maximum hand 2.2. Case Study of Static Contraction. The definition given in grasp force with the hand-muscle developer, and the force is the previous section indicates the importance of required recorded as MVC . Then, a 10-minute rest is provided. After force in forearm muscle fatigue estimation. The greater the the break, the subject is asked to perform a hand grasp with a required force becomes, the less contraction time to maintain muscle contraction at x% MVC (x = 50, 60, 70, according to the required force level lasts and the muscle is easier to fall which session is performed) as steadily as possible for 10 sec- into fatigue state. In a case study where subjects perform onds. This grasp force is recorded as MVC . Then, a 5-second static contraction tasks by maintaining the required hand grasp at MVC is performed immediately, without a rest, and grasp force level as steadily as possible, it is necessary to the maximum hand grasp force is recorded as MVC . After Force (% control MVC) Hand grasp force (N) 4 Applied Bionics and Biomechanics Subject’s non-dominant hand ELONXI hardware box Electrode Pressure sensor Hand-muscle developer 16-channel electrode-embedded sleeve Figure 3: Experimental setup. that, another loop of 10-second x% MVC steady contraction signals and force measurements are captured and synchro- and 5-second MVC contraction is performed and repeated nized simultaneously during the experiment. multiple times continuously, without a break until MVC falls 3.3. Data Processing. The relation between sEMG and force below MVC . Then, one session is finished and a 30 minutes signals is extracted in an offline scheme. Two Sallen-Key fil- rest is given for the purpose of recovering from the muscle ters are employed to band-pass filter raw sEMG signals at a fatigue and preparing for the next session. The subject is pro- bandwidth between 20 Hz and 500 Hz. In addition, a notch vided with visual hints throughout the experiments to ensure filter with central cut-off frequency at 50 Hz (UK power line their adaption to the force variance. The entire experimental frequency) is used to remove the power line interference. procedure of one session is shown in Figure 4. The sEMG signals of each channel are segmented by the overlapped windowing technique [20] with a 300 ms window 3.2. Data Acquisition. In this study, three muscles closely and 100 ms window shift for feature extraction. In this study, related to hand grasp are selected to record the sEMG sig- RMS and MNF/ARV [21–23] are selected as sEMG features. nals, which are palmaris longus, flexor carpiulnaris, and Except for sEMG signal processing, the mean value of hand extensor digitorum. A 16-channel electrode-embedded grasp force data is adopted in each analysis window. sleeve (ELONXI, UK) is used to cover the aforementioned 3.4. Force Estimation Methods. BPNN is used to learn the forearm muscles to collect the sEMG signals where pal- association between sEMG signals and hand grasp forces. maris longus, flexor carpiulnaris, and extensor digitorum In order to evaluate the effect of muscle fatigue on hand grasp mainly correspond to 1-channel electrode, 3-channel elec- force exertion, we propose the following three methods and trode, and 5-channel electrode, respectively, as shown in compare them with locally acquired experiment data. Figure 5. The reference electrode is at the subject’s wrist. Time-domain feature driving machine learning-based Before wearing the electrode sleeve, the skin is cleaned method (TMLM, as shown in Figure 6): train the BPNN with by alcohol, and a 10-minute-rest is given after the elec- pooled training data from all muscle fatigue status together. trode attachment to improve the contact of the electrode The inputs of BPNN are three muscles’ sEMG feature RMS, with skin to reduce the resistance within [19]. sEMG sig- forming the feature vector nals are amplified by a factor of 5000 with linear range 20 Hz to 500 Hz and sampled at 1000 Hz. The FingerTPS ½ RMS , ð4Þ system (Pressure Profile Systems, Inc. (PPS), USA), origi- i×n nally utilised for capturing the tactile force on the finger pulp, is used to measure the hand grasp force in the where i is the channel and n is the number of window shift. experiment. Since the finger pulp is not the optimum pres- And the output is the measured hand grasp force. All the data sure point during hand grasp, a highly sensitive capacitive- acquired under three distinct hand grasp force levels are based pressure sensor is fixed to the appointed position on formed as the training/testing data for the BPNN. hand-muscle developer (shown in Figure 3). The sample Combined feature driving machine learning-based frequency is 100 Hz controlled by the PC clock. The sEMG method (CMLM, as shown in Figure 7): train the BPNN with Applied Bionics and Biomechanics 5 Start 10-second hand grasp at MVC 5-second hand grasp at MVC 5-second hand grasp at MVC Record MVC Record MVC 10 minutes rest MVC > MVC t f MVC = X % MVC 30-minute rest prepare for the next session Figure 4: Diagram of the experimental procedure of one session. Extensor digitorum (back) This method is identical with the above method in output Palm of non-dominant hand and selection of training/testing data. Fatigue feature driving machine learning method (FMLM, as shown in Figure 8): train the BPNN with estimated muscle fatigue value as an additional attribute. An additional input of the degree of muscle fatigue estimated by using (3) in combination with the RMS features is provided to the BPNN and expressed as ½ RMS ,MFL , ð6Þ i×n i×n Flexor carpiulnaris (front) Palmaris longus (front) where i is the channel and n is the number of window shift. Figure 5: Diagram of the experimental procedure. The output and selection method of training/testing data remain the same for all methods, as shown in the following combination of time domain and frequency domain features. three figures. MNF is often employed for the expression of muscle fatigue All methods adopt the BPNN architecture for force esti- information in sEMG-based force estimation. And Japanese mation, whose performance is dependent on the choice of researchers further proposed MNF/ARV, which has achieved network structure, training data, and testing data. The net- good results in muscle fatigue detection [13]. So, a combined work structure is adjusted by setting different number of feature vector is given as nodes from 2 to 20 in the hidden layer with the optimal results [24] provided by a three-layer BPNN. And a Log- Sigmoid function is selected as the transfer function in the ðÞ RMS ,ðÞ MNF /ARV i i i n n network. N 2 ð5Þ ∑ ðÞ F − F∧ 2 k k k=1 R =1 − , N Sx = , ð7Þ ðÞ −x ∑ F − F k=1 k k 1+ e where x is the input and e is the exponential function. More- where i is the channel and n is the number of window shift. over, a four-fold cross validation is adopted to avoid random F denotes the actual hand grasp force, F is the predicted classification of training data and test data from affecting the k k hand grasp force, F is the average of actual hand grasp force, prediction results, which helps ensure the reliability and and N is the number of testing data. stability of the model. 6 Applied Bionics and Biomechanics Output sEMG signals BPNN BPNN Training data parameters training Inputs BPNN RMS Data acquisition Data processing flexor pollicis longus preprocessing RMS Hand grasp force palmaris longus signal segmentation RMS carpiulnaris feature extraction Figure 6: Flowchart of grasp force estimation—TMLM. Output sEMG signals BPNN BPNN Training data parameters training Inputs BPNN RMS Data acquisition Data processing flexor pollicis longus preprocessing RMS Hand grasp force palmaris longus signal segmentation RMS carpiulnaris feature extraction MNF ARV MNF Combined feature ARV extraction MNF ARV Figure 7: Flowchart of grasp force estimation—CMLM. Output sEMG signals BPNN BPNN Training data parameters training Inputs BPNN RMS Data acquisition Data processing RMS flexor pollicis longus preprocessing Hand grasp force RMS palmaris longus signal segmentation carpiulnaris feature extraction MFL Maximum hand Muscle fatigue grasp force estimation Figure 8: Flowchart of grasp force estimation—FMLM. Applied Bionics and Biomechanics 7 sEMG signals of palmaris longus sEMG signals of flexor carpiulnaris sEMG signals of extensor digitorum 400 800 100 200 0 0 –100 –200 –100 –200 –400 –200 –300 –600 –300 –400 –800 –400 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5 5 5 ×10 ×10 ×10 Sampling points Sampling points Sampling points 50% MVC 60% MVC 70% MVC (a) RMS of palmaris longus RMS of flexor carpiulnaris RMS of extensor digitorum 110 90 80 160 60 50 40 80 10 20 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Window shift times Window shift times Window shift times 50% MVC 60% MVC 70% MVC (b) MNF/ARV of palmaris longus MNF/ARV of flexor carpiulnaris MNF/ARV of extensor digitorum 18 18 10 16 16 14 14 𝜇 𝜇 7 12 12 10 10 8 8 6 6 4 4 2 2 1 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Window shift times Window shift times Window shift times 50% MVC 60% MVC 70% MVC (c) Figure 9: Continued. RMS (V) Filtered skin surface voltage (V) MNV/ARV (Hz/V) MNV/ARV (Hz/V) Filtered skin surface voltage (V) RMS (V) MNV/ARV (Hz/V) RMS (V) Filtered skin surface voltage (V) 8 Applied Bionics and Biomechanics Actual hand grasp force of each sample Actual maximum hand grasp force of each sample Degree of muscle fatigue of each sample 36 50 1 0.9 0.8 0.7 40 0.6 30 0.5 0.4 0.3 30 0.2 0.1 24 25 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 Sample number Sample number Sample number 50% MVC 60% MVC 70% MVC (d) Figure 9: Data of static contraction tasks. (a) sEMG signals of three muscles. (b) RMS of three muscles. (c) MNF/ARV of three muscles (d) actual hand grasp force, maximum hand grasp force, and MFL. 1 1 1 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0 0 0 0 100 200 300 0 100 200 300 0 50 100 150 Time (s) Time (s) Time (s) 50% MVC 60% MVC 70% MVC i i i Figure 10: Muscle fatigue estimation results. Solid lines are linear fitting of the estimation. Table 1: Gradient of MFL. In this study, R is used to evaluate the estimation perfor- MVC 50% MVC 60% MVC 70% MVC mance of three methods, which can be expressed as f i i i Gradient 0.0077 0.0108 0.0532 N 2 ∑ F − F∧ ðÞ 2 i=1 i i 4. Results and Discussion R =1 − : ð8Þ ∑ F − F i=1 i i In this paper, an algorithm to quantitatively estimate the degree of muscle fatigue is introduced. And the effect of mus- The R can be comprehended as the percentage of the cle fatigue on hand grasp force estimation is evaluated by response variable variation that is explained by a linear conducting three distinct comparison methods. The experi- model [25] and ranges from 0 to 1. In general, the higher mental results of one subject are shown in Figure 9. They the R , the better the model fits the data. are sEMG signals, RMS, and MNF/ARV of three muscles in T-tests were used to verify differences in TMLM, CMLM, different levels of static contraction tasks, actual hand grasp and FMLM between different conditions. Differences among force, maximum hand grasp force, and MFL of each sample subjects are not considered in this paper, as muscle-level in different levels of static contraction tasks. These selected dynamic variation is commonly existed. p <0:05 is consid- features, seen in Figure 10, can basically reflect the force ered statistically significant for all tests. and muscle fatigue information. Hand grasp force (kg) MFL (100%) Maximum hand grasp force (kg) MFL (100%) MFL (100%) MFL (100%) Applied Bionics and Biomechanics 9 Table 2: R (mean ± sd) of predictions in TMLM. Number of nodes 2 3456 0:6530 ± 0:0314 0:7037 ± 0:0037 0:9093 ± 0:0205 0:6933 ± 0:0083 0:4682 ± 0:0949 R (mean ± sd) Number of nodes 7 8 9 10 11 0:8328 ± 0:0095 0:8201 ± 0:0013 0:7978 ± 0:0311 0:7738 ± 0:0183 0:8610 ± 0:0406 R (mean ± sd) Number of nodes 12 13 14 15 16 0:8259 ± 0:0007 0:7103 ± 0:0074 0:7561 ± 0:0273 0:6653 ± 0:0170 0:6531 ± 0:0713 R (mean ± sd) Number of nodes 17 18 19 20 0:7817 ± 0:0211 0:8261 ± 0:0021 0:8156 ± 0:0076 0:8033 ± 0:0145 R (mean ± sd) Table 3: R (mean ± sd) of predictions in CMLM. Number of nodes 2 3456 0:7321 ± 0:0599 0:8357 ± 0:0081 0:7897 ± 0:0276 0:9255 ± 0:0063 0:6594 ± 0:1223 R (mean ± sd) Number of nodes 7 8 9 10 11 0:7623 ± 0:0132 0:6229 ± 0:0818 0:6648 ± 0:0049 0:7149 ± 0:0563 0:7281 ± 0:0218 R (mean ± sd) Number of nodes 12 13 14 15 16 0:7579 ± 0:0357 0:7932 ± 0:0026 0:5954 ± 0:0214 0:4769 ± 0:1373 0:8362 ± 0:0035 R (mean ± sd) Number of nodes 17 18 19 20 0:7074 ± 0:0104 0:6305 ± 0:0201 0:5928 ± 0:0450 0:7207 ± 0:0422 R (mean ± sd) Table 4: R (mean ± sd) of predictions in FMLM. Number of nodes 2 3456 0:8158 ± 0:0096 0:8663 ± 0:0220 0:8425 ± 0:0055 0:9193 ± 0:0185 0:8356 ± 0:0149 R (mean ± sd) Number of nodes 7 8 9 10 11 0:8343 ± 0:0205 0:8795 ± 0:0016 0:8746 ± 0:0127 0:9572 ± 0:0030 0:8842 ± 0:0122 R (mean ± sd) Number of nodes 12 13 14 15 16 0:8312 ± 0:0474 0:8652 ± 0:0018 0:8892 ± 0:0017 0:7465 ± 0:1342 0:7358 ± 0:0545 R (mean ± sd) Number of nodes 17 18 19 20 0:8272 ± 0:0143 0:8100 ± 0:0063 0:7836 ± 0:0054 0:8472 ± 0:0010 R (mean ± sd) Table 5: Prediction results of three different methods. Method TMLM CMLM FMLM Number of nodes 4 5 10 0:8782 ± 0:0005 0:9065 ± 0:0011 0:9506 ± 0:0009 R (mean ± sd) 4.1. Results of Muscle Fatigue Estimation. The experiments levels of static contraction. The greater the required force last for 300 s, 240 s, and 50 s corresponding to 50% MVC becomes in static contraction, the faster MFL rises, shown (250 N), 60% MVC (300 N), and 70% MVC (350 N kg) in in Table 1, which implies that the task intensity can also be i i distinguished through the proposed metric. The feasibility static contraction tasks, respectively. Figure 10 shows the estimation results of muscle fatigue by the proposed method. of the proposed method is recognized in static contraction Dot arrays of different colors represent the estimation results tasks to estimate muscle fatigue quantitatively. at different force levels. Through linear fitting, it can be directly seen that MFL grows linearly with the increase of 4.2. Results of Hand Grasp Force Estimation. In order to weaken effect of network structure, initial weights, and bias contraction time, which is in accordance with Vøllestad’s assumption [12] of muscle fatigue’s variety law during a sus- values on the estimation performance, the neural network tained and steady contraction. In addition, the results show is retrained ten times at different numbers of nodes (from 2 that the gradient of time-varying MFL varies at different to 20) in the hidden layer. 10 Applied Bionics and Biomechanics Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison 2 2 2 2 R = 0.8818 R = 0.90022 R = 0.91813 R = 0.93697 36 36 36 40 34 34 32 32 30 30 32 28 28 26 26 24 22 24 24 0 1020304050 0 1020304050 0 1020304050 0 1020304050 Prediction samples Prediction samples Prediction samples Prediction samples Predicted value (50% MVC ) True value (50% MVC ) i i Predicted value (60% MVC ) True value (60% MVC ) Predicted value (70% MVC ) True value (70% MVC ) (a) Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison 2 2 2 2 R = 0.91705 R = 0.92491 R = 0.92531 R = 0.9347 36 36 36 36 34 34 34 32 32 32 30 30 30 28 28 28 26 26 26 24 24 24 22 22 24 22 0 1020304050 0 1020304050 0 1020304050 0 10 20 30 40 50 Prediction samples Prediction samples Prediction samples Prediction samples True value (50% MVC ) Predicted value (50% MVC ) i i True value (60% MVC ) Predicted value (60% MVC ) i i True value (70% MVC ) Predicted value (70% MVC ) i i (b) Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison Hand grasp force prediction result comparison 2 2 2 2 R = 0.95392 R = 0.95569 R = 0.95711 R = 0.96203 36 36 36 36 34 34 34 32 32 32 30 30 30 28 28 28 26 26 26 24 24 24 22 0 1020304050 0 1020304050 0 1020304050 0 10 20 30 40 50 Prediction samples Prediction samples Prediction samples Prediction samples True value (50% MVC ) Predicted value (50% MVC ) True value (60% MVC ) Predicted value (60% MVC ) True value (70% MVC ) Predicted value (70% MVC ) i i (c) Figure 11: Single four-fold cross-validation results of each method (a) TMLM. (b) CMLM. (c) FMLM. In TMLM, we pool training data from all muscle fatigue BPNN are shown in Table 4. The network structure of 10 states to make the network learn the differences among them nodes in the hidden layer brings the maximum mean R . alone. Predictions of BPNN are shown in Table 2. When the Its value is 0.9572. number of node is 4, the mean R is 0.9093, which is the Comparing the best prediction results of three different maximum. methods, as shown in Table 5 and Figure 11, it can be In CMLM, we employ MNF/ARV, one feature proposed indicated that the mean R obtained in TMLM is 0.9093. by Fernando’s team for the expression of muscle fatigue It just passes the baseline of applicability (0.9000), which information. Predictions of BPNN are shown in Table 3. implies the estimation performance of the BPNN model We set 5 nodes in the hidden layer, and the mean R of pre- in TMLM is not good enough and predicting model need diction results is 0.9255. to be readjusted. In CMLM, the mean R is 0.9255. This In FMLM, we incorporate the quantitative metric of mus- shows that employing MNF/ARV proposed by Fernando cle fatigue value as an additional input to explain the effect of et al. [13] in sEMG-based force estimation under fatigued muscle fatigue on hand grasp force estimation. Predictions of conditions is indeed feasible. But it is not an obviously Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Hand grasp force (kg) Applied Bionics and Biomechanics 11 measure the subject’s current maximum grasp force. As a Table 6: R (mean ± sd) of predictions of all subjects under three methods. result, the force estimation in this work could only be proc- essed offline. So the future work is mainly to address how Method TMLM CMLM FMLM to estimate muscle fatigue online, that is, how to get MFL 0:9093 ± 0:0205 0:9255 ± 0:0063 0:9572 ± 0:0030 Subject 1 online. In fact, the results of this study have provided some potential and guiding ideas for the following work. Under 0:8910 ± 0:0092 0:9210 ± 0:0039 0:9548 ± 0:0009 Subject 2 static muscle contraction, the subject’s forearm muscle 0:8255 ± 0:0043 0:8517 ± 0:0064 0:8938 ± 0:0189 Subject 3 fatigue and muscle contraction time are approximately linear 0:8224 ± 0:0042 0:8464 ± 0:0049 0:8805 ± 0:0028 Subject 4 when maintaining a fixed level of hand grasp force. And this 0:8286 ± 0:0012 0:8571 ± 0:0014 0:9070 ± 0:0190 linear coefficient seems to have a nonlinear increasing rela- Subject 5 tionship with the target hand grasp force level. Therefore, a 0:9293 ± 0:0132 0:9411 ± 0:0209 0:9554 ± 0:0182 Subject 6 nonlinear estimation model of muscle fatigue could be more 0:9199 ± 0:0126 0:9262 ± 0:0050 0:9426 ± 0:0181 Subject 7 appropriated in this case, such as 0:8802 ± 0:0431 0:9359 ± 0:0088 0:9438 ± 0:0088 Subject 8 b×n%MVC 0:8153 ± 0:0268 0:8526 ± 0:0131 0:8753 ± 0:0267 Subject 9 MFL = a × e × t + c, ð9Þ 0:8350 ± 0:0097 0:8620 ± 0:0130 0:8983 ± 0:0121 Subject 10 where n%MVC is the target hand grasp force level. t is mus- cle contraction time. a, b, and c are model parameters. effective approach. Compared with the result in TMLM, R increases by 1.7816%. Data Availability As the main work of this study, the estimated muscle fatigue value is used directly as an additional input in FMLM. The EMG and force data used to support the findings of this The results show mean R can reach to 0.9572, which proves study are available from the corresponding author upon predicting model fits the data very well. It is a great improve- request. ment (5.2678%, above 5%, p <0:05) in estimation perfor- mance compared with CMLM. It is demonstrated that the Conflicts of Interest additional attribute is an applicable solution to the effect of muscle fatigue on sEMG-based hand grasp force estimation. The author(s) declare(s) that they have no conflicts of And MFL proposed in this paper is better than MNF/ARV interest. (p <0:05). For further explanation, the experimental results of all Acknowledgments subjects under different methods are presented as shown in Table 6. Statistics show that the mean R The authors would like to extend their gratitude to H. Liu, Q. values obtained under the three methods are 0.8656, 0.8919, and 0.9209. Gao and C. Li from Intelligent System & Biomedical Robotics Adopting MNF/ARV proposed in [12] for measure muscle Group, School of Computing, University of Portsmouth for fatigue could bring the 3.0383% growth in R in hand grasp assisting in the experimental process. This work was sup- ported in part by the National Natural Science Foundation force estimation. For comparison, using the MFL proposed of China under Grant No. 61603284, 61903286, and in this paper can increase R by 6.3797%. 52075530 and the AiBle project co-financed by the European The experimental results show that FMLM provides the Regional Development Fund. best estimation performance among the three methods. References 5. Conclusion [1] Y. Yamanoi, S. Morishita, R. Kato, and H. Yokoi, “Develop- In this paper, we propose an easy-to-implement method to ment of myoelectric hand that determines hand posture and quantitatively estimate muscle fatigue and evaluate the effect estimates grip force simultaneously,” Biomedical Signal Pro- of muscle fatigue on hand grasp force estimation. The exper- cessing and Control, vol. 38, pp. 312–321, 2017. iment results demonstrate that the incorporation of muscle [2] M. Kim, J. Lee, and K. Kim, “Tele-operation system with reli- fatigue metrics explicitly in the grasp force estimation has a able grasping force estimation to compensate for the time- substantial impact on the performance. When estimated varying sEMG feature,” in 2016 IEEE International Conference muscle fatigue value as an additional input in the machine on Robotics and Automation (ICRA), pp. 5561–5567, Stock- learning approach, the estimation accuracy improves to a holm, Sweden, 2016. large extent in FMLM. Because the neural network is trained [3] L. Peternel, N. Tsagarakis, D. Caldwell, and A. Ajoudani, offline and is computationally cheap, the proposed method “Robot adaptation to human physical fatigue in human–robot can be implemented in the current applications, such as in co-manipulation,” Autonomous Robots, vol. 42, no. 5, [1–3], as a calibration part to improve the effectiveness and pp. 1011–1021, 2018. robustness. At the same time, there are still some limitations [4] C. Choi, S. Kwon, W. Park, H. D. Lee, and J. Kim, “Real-time in this study. During the experiments mentioned above, the pinch force estimation by surface electromyography using an degree of the subject’s forearm muscle fatigue needs to be artificial neural network,” Medical Engineering & Physics, estimated at each moment, which requires to intermittently vol. 32, no. 5, pp. 429–436, 2010. 12 Applied Bionics and Biomechanics Biology Society, pp. 1449-1450, Amsterdam, Netherlands, [5] C. Castellini, E. Gruppioni, A. Davalli, and G. Sandini, “Fine detection of grasp force and posture by amputees via surface 1996. electromyography,” Journal of Physiology-Paris, vol. 103, [19] C. J. De Luca, “Use of the surface EMG signal for performance no. 3-5, pp. 255–262, 2009. evaluation of back muscles,” Muscle & Nerve, vol. 16, no. 2, [6] A. Erdemir, S. McLean, W. Herzog, and A. J. van den Bogert, pp. 210–216, 1993. “Model-based estimation of muscle forces exerted during [20] Y. Fang and H. Liu, “Robust sEMG electrodes configuration movements,” Clinical biomechanics, vol. 22, no. 2, pp. 131– for pattern recognition based prosthesis control,” in 2014 IEEE 154, 2007. International Conference on Systems, Man, and Cybernetics [7] U. J. Naeem, C. Xiong, and A. A. Abdullah, “EMG-muscle (SMC), pp. 2210–2215, San Diego, CA, USA, 2014. force estimation model based on back-propagation neural net- [21] N. A. Shrirao, N. P. Reddy, and D. R. Kosuri, “Neural network work,” in 2012 IEEE International Conference on Virtual Envi- committees for finger joint angle estimation from surface ronments Human-Computer Interfaces and Measurement EMG signals,” Biomedical Engineering Online, vol. 8, no. 1, Systems (VECIMS) Proceedings, pp. 222–227, Tianjin, China, p. 2, 2009. [22] S. Suryanarayanan, N. P. Reddy, and V. Gupta, “Artificial neu- [8] D. Yang, J. Zhao, Y. Gu, L. Jiang, and H. Liu, “Estimation of ral networks for estimation of joint angle from EMG signals,” hand grasp force based on forearm surface EMG,” in 2009 in Proceedings of 17th International Conference of the Engi- International Conference on Mechatronics and Automation, neering in Medicine and Biology Society, pp. 823-824, Mon- pp. 1795–1799, Changchun, China, 2009. treal, Quebec, Canada, 1995. [9] B. Zhang and S. Zhang, “Pattern-based grasping force estima- [23] Y. Koike and M. Kawato, “Estimation of dynamic joint torques tion from surface electromyography,” in 2017 International and trajectory formation from surface electromyography sig- Conference on Algorithms, Methodology, Models and Applica- nals using a neural network model,” Biological Cybernetics, tions in Emerging Technologies (ICAMMAET), pp. 1–5, Chen- vol. 73, no. 4, pp. 291–300, 1995. nai, India, 2017. [24] O. Fujita, “Statistical estimation of the number of hidden units [10] N. Jiang, S. Dosen, K.-R. Muller, and D. Farina, “Myoelectric for feedforward neural networks,” Neural Networks, vol. 11, control of artificial limbs—is there a need to change focus? no. 5, pp. 851–859, 1998. [in the spotlight],” IEEE Signal Processing Magazine, vol. 29, no. 5, 2012. [25] C. Hagquist and M. Stenbeck, “Goodness of fit in regression analysis – R2 and G2 reconsidered,” Quality and Quantity, [11] T. D. Lalitharatne, Y. Hayashi, K. Teramoto, and K. Kiguchi, vol. 32, no. 3, pp. 229–245, 1998. “Compensation of the effects of muscle fatigue on EMG- based control using fuzzy rules based scheme,” in 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 6949–6952, Osaka, Japan, 2013. [12] N. K. Vøllestad, “Measurement of human muscle fatigue,” Journal of Neuroscience Methods, vol. 74, no. 2, pp. 219–227, [13] J. B. Fernando, M. Yoshioka, and J. Ozawa, “Estimation of muscle fatigue by ratio of mean frequency to average rectified value from surface electromyography,” in 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 5303–5306, Orlando, FL, USA, 2016. [14] M. Navaneethakrishna and S. Ramakrishnan, “Multiscale fea- ture based analysis of surface EMG signals under fatigue and non-fatigue conditions,” in 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 4627–4630, Chicago, IL, USA, 2014. [15] Y. Masakado, Y. Noda, K. Hase, A. Kimura, and N. Chino, “Muscle fatigue assessed by frequency analysis of surface elec- tromyographic signals: topographical analysis in the same muscle,” The Japanese Journal of Rehabilitation Medicine, vol. 31, no. 6, pp. 409–414, 1994. [16] H. Xie and Z. Wang, “Mean frequency derived via Hilbert- Huang transform with application to fatigue EMG signal anal- ysis,” Computer Methods and Programs in Biomedicine, vol. 82, no. 2, pp. 114–120, 2006. [17] T. Öberg, L. Sandsjö, and R. Kadefors, “Subjective and objec- tive evaluation of shoulder muscle fatigue,” Ergonomics, vol. 37, no. 8, pp. 1323–1333, 1994. [18] H. Oka, “Estimation of muscle fatigue by using EMG and muscle stiffness,” in Proceedings of 18th Annual Interna- tional Conference of the IEEE Engineering in Medicine and

Journal

Applied Bionics and BiomechanicsHindawi Publishing Corporation

Published: Feb 15, 2021

References