Control Of A Lower Limb Exoskeleton Robot By Upper Limb .

signal processed by the 50 Hz notch filter is shown in the Fig.4.The green curve is the original signal, and the black curve is thefiltered curve. It can be seen that the noise of the processedsignal is obviously reduced [14][15].C. Introduction of muscleThe human upper limb is composed of bone, joint andskeletal muscle. Bone and joint constitute the skeletonsupporting the whole body. These movements are thecompound movements of multiple degrees of freedomcoordinated by shoulder joint, elbow joint and wrist joint.When the upper limb is performing the corresponding action,each action is a single joint movement or a compoundmovement of multiple joints, which is dominated by differentmuscle groups, and the participation of each muscle group indifferent upper limb actions is also different. The main musclesinvolved in upper limb movement are pectoralis major, bicepsbrachii, triceps brachii, deltoid and brachioradialis. Theirfunctions in each movement mode are shown in Table 1. Theexperiment shows that biceps brachii and brachioradialisbrachii have higher accuracy in distinguishing arm throwingand arm lifting. sEMG signals of these two muscles arecollected as input signals of BP neural network [11].Fig. 4 Signal after frequency notch of biceps brachiiThen, since the effective signals of sEMG signal arebasically concentrated in 10-200Hz, Butterworth band-passfilter is used for further processing to remove the noise of otherbands. The processed muscle signal is shown in the Fig.5.Thered curve is the curve after pretreatment. Compared with theoriginal signal, the time-domain waveform of the pretreatedsEMG signal is smoother, and the signal energy is mainlyconcentrated in 10-200Hz.TABLE ITHE ROLE OF DIFFERENT MUSCLESMotion jointMotion modeCorresponding muscleAdductionPectoralis major,DeltoidAbductionDeltoid,TricepsFront swingPectoralis major,TricepsBack swingTriceps,SupraspinatusShoulder jointFlexionBiceps, Brachioradialis muscleElbow jointExtensionTricepsⅢ. PRETREATMENT AND FEATURE EXTRACTION OFSEMG SIGNALA.PretreatmentFig. 5 Signal processing of biceps brachii Butterworth band pass filterBecause the intensity of sEMG signal itself is very weak,it is easy to introduce other noises in the process of acquisition,such as power frequency interference, inherent noise ofacquisition equipment, and other biological signal noises suchas electrocardiogram signal. The introduction of a large amountof noise will seriously affect the accuracy of sEMG signalanalysis and motion control. Therefore, in addition tominimizing the acquisition error in the process of sEMGacquisition, it is necessary to further process the collectedsEMG signal [12][13].This paper preprocesses the sEMG signal according to thecommon forms and characteristics of noise interference,including band-pass filtering, power frequency removal andharmonic interference.Firstly, a notch filter is used to deal with the 50 Hzcommon frequency interference caused by the power supply.The principle of notch filter is band stop filter. The blockingfrequency is set to a small distance near the notch. Takingbiceps brachii as an example, the frequency domain of theB.Feature extractionsEMG has the characteristics of non-stationary signal, so itis difficult to obtain enough information from a single channelfor gesture recognition in this application scenario, so it isnecessary to collect data from multiple channels as recognitionsignals. If all the signals of the whole active segment are usedas input for recognition and extraction, it is a heavy workloadand difficult to achieve. Feature extraction can not onlycompress the dimension of feature space, but also distinguishthe differences of feature signals corresponding to differentgesture actions, and highlight their significance, so as toimprove the recognition rate of the classification system.Therefore, we need to use the feature extraction method toextract the characteristics of a group of signals for datadescription, so as to more effectively classify and identify,which is the main target of feature extraction [16].At present, the characteristics of sEMG signal can be1115

analyzed in time domain or frequency domain. Considering thatthe sEMG signal can reflect the muscle force information betterin time domain and has high real-time performance, this paperuses the time domain eigenvalue analysis of sEMG. In thispaper, four time-domain features with large discrimination areselected: absolute mean value, root mean square value, integralsEMG value and wavelength.Their expressions and physical meanings are as follows:The absolute mean represents the mean value of sEMGsignal in a certain period of time. The expression is as follows:1 （1）Fig. 7 Feature extraction results of brachioradialisRoot mean square value reflects the energy ofmyoelectric signal in a certain period of time. The expressionis as follows:1Ⅳ.ACTION CLASSIFICATIONIn this paper, BP neural network is selected to establish theangle prediction model. Its propagation direction is one-waypropagation, which belongs to multi-layer forward feedbacknetwork. Back propagation algorithm is used to train thenetwork. The layer of BP neural network can be divided intothree types, namely input layer, hidden layer and output layer.There is a complete connection between layers, but there is noconnection between neurons in each layer. A three-layer BPneural network can realize any mapping from n-dimension tom-dimension, so this paper selects three-layer BP neuralnetwork to build angle estimation model [17].The number of nodes in the input layer is determined bythe number of channels of sEMG signal. The experiment showsthat the absolute average value and wavelength of the twomuscles collected have the highest degree of discrimination forthe two arm movements set. Therefore, the absolute averagevalue and wavelength of each muscle are selected as the inputsignal of BP neural network, which has four input nodes. Thenumber of nodes in the output layer is determined by thenumber of actions to be classified. In this paper, it is initially setto classify the arm lifting and arm throwing actions, so thenumber of nodes in the output layer is 2. There are manychoices for the number of hidden layer units, but the choice ofthe number has a great impact on the network performance. Itsselection needs to be determined according to the problem to bestudied, the number of nodes in the input and output layer, thedesigner's experience and many experiments. Finally, thenumber of nodes selected in this paper is 8 [18].sEMG signal can accurately reflect the degree ofcontraction of related muscles, and then predict thecorresponding action through sEMG signal. Afterpreprocessing and feature extraction, the collected originalsignal is taken as the input data, and the corresponding twoactions are replaced by 0 and 1 as the output data to the BPneural network. After training, the BP neural network modelwhich can predict the joint angle can be obtained. The flowchart of joint angle prediction based on sEMG signal is shownin Fig.8.（2）The integral sEMG value reflects the intensity change ofsEMG signal with time.The expression is as follows:1 （3）The wavelength reflects the cumulative length of thewave in a certain period of time. The expression is as follows:1 1 （4）Finally, the feature extraction of sEMG data is carried outby sliding window method. The length of the window is500ms and the sliding distance is 50ms. The followingwaveform is obtained.Fig. 6 Feature extraction results of biceps brachii1116

that the absolute average value and wavelength of the twomuscles collected have the highest degree of differencebetween the two arm movements, so these two features areselected as the input signals of BP neural network. The outputsignal is set to 0 and 1, corresponding to normal walking andobstacle avoidance gait respectively. The classification resultsare shown in Fig.10.It can be seen that basically two kinds ofactions can be distinguished accurately. If the output result isset to be greater than 50%, it is regarded as the arm liftingaction, otherwise it is the arm throwing action. Therefore, evenif a few results are not very accurate, it will not affect thesubsequent control of the motor.Fig. 8 Control processIn this paper, two kinds of motion states are initially set,which are normal walking gait and high leg lifting gait,corresponding to normal arm swing and arm lifting. When theleft arm swings, the motors at the four joints cooperate witheach other to complete a gait movement followed by the left legafter the right leg moves forward; when the right arm swings,the left leg moves forward and the right leg follows; when thesame wearer lifts the arm, the wearer carries out a gaitmovement of high leg lifting across obstacles or up steps [19].In the experiment, sEMG signals of 10 healthy subjectsaged 20-50 were collected, including 5 males and 5 females.sEMG signals of their right upper limbs were collected tosimulate the clinical rehabilitation process of hemiplegicpatients. Each subject was in good health, full of rest, no musclefatigue and relaxed. Before collection, 75% alcohol was used towipe the surface of the muscle group to remove the dirt andenhance the conductivity. After waiting for the skin to drynaturally, the sensor was pasted on the right upper limb of thesubject according to the muscle group position selected above.During the collection, considering the influence oflong-time muscle fatigue on the experimental data, each subjectrepeated each action 10 times, with an interval of about 3seconds. In order to facilitate the extraction of active segments,2-3 seconds of idle time is reserved before and after eachaction, and the time to complete a complete action is about 7-8seconds. Fig.9 shows the signal acquisition site of two types ofactions.Fig. 10 The classification results of two movementsⅤ.EXPERIMENTS AND RESULTSFinally, the theoretical method is combined with theexperimental platform to carry out preliminary experimentalverification, and the knee joint elevation is used to replace thenormal walking gait with knee joint elevation of 30 degrees,and the knee joint is raised 60 degrees instead of the high leglifting gait. Because the risk of patients with lower limbdysfunction participating in the experiment, healthy young menwere selected as subjects. The experimental process is shown inFig.11.(a)Normal gait(b)High leg up gaitFig 9 Collection of subjects' sEMG signalAfter the sEMG signal acquisition experiment, 650 groupsof data were obtained by feature extraction, 500 groups of eachaction were used for classifier training, and the other 150groups were used for test experiment. The experiment shows(c) The motion curves of the two gaitsFig. 11 Switching experiment of two kinds of gait1117

Production Engineering”, Materials, Vol.13,No.24,pp.1267-1273,2020.[4] Xiong, DZ, Zhang, DH, Zhao, XG, Zhao, YW, “Deep Learning forEMG-based Human-Machine Interaction: A Review”, IEEE-CAA Journalof AutomaticaSinica, Vol.8, No.3, pp.512-533, 2021.[5] Shanmuganathan, V, et al,“R-CNN and wavelet feature extraction for handgesture recognition with EMG signals”, Neural Computing & Applications,Vol.32,No.21,pp.16723-16736,2020.[6] Mukhopadhyay, AK, et al,“An experimental study on upper limb positioninvariant EMG signal classification based on deep neural network”,Biomedical Signal Processing and Control, Vol.55,pp.378-385, 2020.[7] Chen, CK,et al,“Lower-Limb Electromyography Signal Analysis ofDistinct Muscle Fitness Norms under Graded Exercise Intensity”,Electronics, Vol.9,No.12,pp.2531-2537, 2020.[8] Zhuang, Y, Leng, Y, Zhou, J, Song, R, Li, L, Su, SVW, “VoluntaryControl of an Ankle Joint Exoskeleton by Able-Bodied Individuals andStroke Survivors Using EMG-Based Admittance Control Scheme”,IEEE Transactions on Biomedical Engineering, Vol.68, No.2,pp.695-705, 2021.[9] Feifei Qin, Han Zhao, Shengchao Zhen, Hao Sun, Yan Zhang，“LyapunovBased Robust Control for Tracking Control of Lower Limb RehabilitationRobot with Uncertainty”, International Journal of Control Automationand Systems, vol.18, pp. 76-84,2019.[10] Andre M. Barbosa Joa o Carlos M. Carvalho Roge rio S. Gonçalves,“Cable-driven lower limb rehabilitation robot”, Journal of the BrazilianSociety of Mechanical Sciences and Engineering, vol.40, pp.245-230,2018.[11] Chen, YM, Yang, ZL, Wen, YL, “A Soft Exoskeleton Glove for HandBilateral Training via Surface EMG”, Sensors, Vol.21, No.2,pp.672-679,2021.[12] Simao, M , Neto, P , Gibaru, O,“EMG-based online classification ofgestures with recurrent neural networks”, Pattern Recognition Letters,Vol.128,pp.45-51, 2019.[13] Abbaspour, S, et al,“Evaluation of surface EMG-based recognitionalgorithms for decoding hand movements”, Medical & BiologicalEngineering & Computing, Vol.58,No.3, pp.83-100, 2019.[14] Fajardo, JM, Gomez, O, Prieto, F, “EMG hand gesture classificationusing handcrafted and deep features”, Biomedical Signal Processing andControl, Vol.63, pp.167-173,2019.[15] Gregory, U, Ren, L,“Intent Prediction of Multi-axial Ankle Motion UsingLimited EMG Signals”, Frontiers in Bioengineering and Biotechnology,Vol.7,pp.563-569, 2019.[16] Saeed, B,et al,“Leveraging ANN and LDA Classifiers for CharacterizingDifferent Hand Movements Using EMG Signals”, Arabian Journal forScience and Engineering, Vol.46,No.2,pp.1761-1769,2020.[17] Nair,AS, et al. “Performance Analysis of Super Twisting Sliding ModeControllerbyADAMS-MATLABCo-simulationin Lower Extremity Exoskeleton”, International Journal of PrecisionEngineering and Manufacturing-Green Technology, vol.7, pp.743-754,2020.[18] Chen,C, et al. “Disturbance Observer-Based Patient-Cooperative Controlof a Lower Extremity Rehabilitation Exoskeleton”, International Journalof Precision Engineering and Manufacturing, vol.21, pp. 957-968,2020.[19] Poritz, Julia M P, et al, “User satisfaction with lower limb wearablerobotic exoskeletons”, Disability and rehabilitation. Auxiliary technology,vol.15, pp. 322-327,2020.The real-time data collected by sEMG sensor is importedinto the BP neural network model trained by Matlab. When theoutput of neural network is less than 0.5, it is considered thatthe normal gait should be performed at this time, and thecorresponding control signal is sent to the MCU throughUSART serial port. When the output is greater than 0.5, thehigh leg lifting gait is performed. After repeated experiments, itis found that the neural network can effectively classify the armswing and arm lift, and then control the motor to execute thecorresponding gait. However, due to the instability of thereal-time sEMG signal, there are a few cases of classificationdelay, that is, the whole arm lifting action is judged as armlifting after it is executed, and the real-time performance will beaffected to a certain extent. It is necessary to further optimizethe classification algorithm to improve the real-time control.Ⅵ.CONCLUSIONThis paper mainly designed a lower limb exoskeletonrobot based on the upper limb sEMG signal control, whichchanged the traditional control method, solved the problem thatthe lower limb sEMG signal strength of hemiplegic patientswas weak and affected the classification effect. And also solvedthe problem that the lower limb space was insufficient afterwearing the lower limb exoskeleton robot, which was not easyto collect the sEMG signal. The angle prediction model of BPneural network based on sEMG signal was established, and thecorresponding mode of upper limb movement and gaitmovement was designed. The real-time performance andaccuracy of motion prediction based on sEMG signal wereverified by experiments. Through this control method, thehuman-computer interaction ability was greatly improved, andthe rehabilitation enthusiasm of patients was increased.Although the project has achieved the expected goal andachieved certain research results, there is still a lot of work to befurther improved: on the one hand, it is necessary to carry outfurther experimental design and verification for patients withlower limb dysfunction who are really in the middle and latestage of rehabilitation; on the other hand, the reaction time ofgait switching and the classification accuracy of sEMG signalneed to be improved.ACKNOWLEDGMENTThis research is supported by National Natural ScienceFoundation of China (61703305) and Key Research ZDJC38500) and Innovative Cooperation Project Z00090).REFERENCES[1] Rajapriya, R, Rajeswari, K , Joshi, D, Thiruvengadam, SJ, “ForearmOrientation and Contraction Force Independent Method for EMG-BasedMyoelectric Prosthetic Hand”,IEEE Sensors Journal, Vol.21, No.5,pp.6623-6633, 2021.[2] Khera,P , Kumar,N , Ahuja,P,“MachineLearningbasedElectromyography Signal Classification with Optimized Feature Selectionfor Foot Movements”, Journal of Scientific & Industrial Research,Vol.79,No.11,pp.1011-1016,2020.[3] del Olmo, M, Domingo, R,“EMG Characterization and Processing in1118

Upper Limb sEMG Signal Abstract -In this paper, a lower limb exoskeleton robot based on upper limb sEMG signal controlledby designed for patients with lower limb functional injury in the middle and late stage of rehabilitation. It realized the patient's active and random control when wearing the lower limb exoskeleton for rehabilitation