Table 3 A comprehensive comparison with other sEMG motion intention algorithms
From: Myoelectric control algorithm for robot-assisted therapy: a hardware-in-the-loop simulation study
References | Robot | Sensors | Approximation | Implementation | Experiments | Calibration process | Type of Algorithm (black box, white box, grey box) | Results/errors | Advantages | Disadvantages |
|---|---|---|---|---|---|---|---|---|---|---|
Proposed MEC algorithm | Nukawa | Four pairs of electrodes in BF, VM, RF, and ST muscles. One ground electrode in C7 | A Butterworth filter with cut-off frequencies of 10 Hz and 500 Hz was used. Removing DC offset and full-wave rectification were also used. An LC of the four RMS envelopes is proposed, i.e., the features of the four channels were combined. The motion intention is a continuous value | Offline and online | 17 healthy subjects | MVC exercise for both flexion and extension muscles | Grey box | The algorithm detected the orientation of the intention 100% of the times for both extension and flexion exercises. The algorithm detected the intensity of the movement intention, in a comparable way to the MVC, in 94% and 59% of the cases during extension and flexion exercises, respectively. The robot can follow the desired trajectories in Cartesian space, imposed by the MEC algorithm. The position and speed error are small compared to the motion of each joint. The maximum position MAE is \(0.1^{\circ }\), \(6.3^{\circ }\), and \(0.3^{\circ }\) for the hip, knee, and ankle joints, respectively. Thus, the error is lower in the knee | It is a simple algorithm which requires a small amount of processor and no additional sensors. Moreover, the grey box model it uses a simple calibration process, using the well-known MVC tests. Also, it was tested on multiple subjects in comparison with other studies in its offline and online version | The LC proposed in this paper uses four sEMG channels instead of two |
[17] | None | An accelerometer placed on the forearm and sEMG electrodes on BB and TB | The raw signal was processed. Subsequently, the signals were rectified, and a signal normalization process was developed for each subject according to pre-recorded signals. The neural activity was calculated, and a Kalman Filter was used to predict motion | Real-time | 12 healthy subjects | Manual calibration | Black box | The model presented a high correlation with slow-motion trajectories (CC = 0.999–1). Moreover, the results showed high accuracy (97.4–98.6%) of prediction | It was validated with 12 healthy subjects and requires a simple manual calibration | It was tested on upper limb joints. Moreover, it requires additional sensors. The algorithm was tested in an offline fashion |
[18] | A virtual human model (VHM) | The sEMG signals were collected from the anterior deltoid, posterior deltoid, BB, and TB | The signals were rectified with RMS to obtain an amplitude envelope. Subsequently, a low pass filter was implemented, and the signals were normalized. The raw and pre-processed signals were the input of a three-layer back propagation neural network (BPNN) controller | Offline | Four healthy subjects | Offline training of the algorithm | Black box | The ANN performance in the estimation of the joint angle in each motion was computed using the mean squared error (MSE) method. The worst average for the MSE was 0.239 | It was tested using multiple joints and multiple movements | It requires to train a machine learning algorithm. Therefore, it requires a training data set, It was tested on upper limb joints. The algorithm was tested in an offline fashion |
[19] | HAL-3 | Two sensors near the flexor and extensor muscles | Signals were filtered and amplified, and the myoelectric activity was computed for both channels. Subsequently, the estimated muscle torque was computed as a linear combination of both, taking into account the equation of a straight line. Finally, a gain parameter was used to compute the torque for the actuator | Online | A healthy subject | A calibration process is necessary to obtain the conversion coefficients | Black box | The conversion coefficients depend on the sensor location and the operator’s physical condition | It uses a simple algorithm, and it was tested online in a commercial robotic exoskeleton | It requires a long calibration process, including additional sensors such as torque sensors |
[20] | A computer model of the index finger and wrist joints | Flexor digitorum superficialis (FDS) and flexor carpi ulnaris (FCU) | An RMS envelope was computed. Subsequently, a low-pass filter was used, and two different functions were used for the finger position | Online | 18 healthy subject | Simple calibration process of constants | Black box | The maximum errors obtained were \(3.77^{\circ }\). A direct relationship between the RMS and the motion of model was observed | It uses a simple algorithm, and it was tested in an online fashion. It requires a simple calibration process, and it was tested on 18 healthy subjects | It was tested on upper limb joints |
[21] | NEURO-exos | BB and TB | The EMG signals were processed obtaining a linear envelope (LE) through full-wave rectification. Both signals were conducted to a proportional controller to manipulate the flexion and extension of the exoskeleton | Online | Ten healthy subjects | Subjects selected the gains of the algorithm in a previous procedure | Black box | Subjects could fulfill the tasks during all trials, no matter the percentage of assistance, extra weight or movement pace | It requires a simple calibration exercise, it was tested in an online fashion, and it was tested on ten healthy subjects | It was tested on upper limb joints |
[23] | None | BB and TB | The MAV was computed. Subsequently, discriminant analysis and an SVM was used to classify the signals | Not reported | Three healthy subjects | Training of the algorithm | Black box | Classification accuracy for the discriminant analysis and the SVM was 96% and 99%, respectively | It was tested using multiple joints and multiple movements | It requires to train a machine learning algorithm. Therefore, it requires a training data set, it was tested on upper limb joints. The type of implementation is not reported |
[24] | None | BB and TB | A low-pass filter was used. Subsequently, two time-domain features were extracted and the signals were normalized. A linear state-space model was used to estimate joint motion | Offline | Two healthy subjects at two load levels | Offline training of the algorithm | Black box | The authors obtained a root-mean-square error ranging between 8.3 and 10.6%. Also, the prediction error of the average angle was around 10% | It overcomes subject-specific problems | It was tested in an offline fashion. It was tested on upper limb joints and just in two subjects |
[41] | iLeg | RF, VL, VM, BF and ST | Full wave rectification, low pass filter of 2 Hz of the sEMG signals which are inputs of a network Neural network. The angle and speed are also inputs to the neural network | Offline | One healthy subject | Training the neural network | Black box | The root-mean-square error is 0.67 N m for hip torque estimation and 0.37 N m for knee torque estimation | It can be used to perform a real-time coordinated active training with a rehabilitation robot. It was tested using multiple joints, hip, and knee | The proposed approach was tested with a circular-like trajectory. Requires additional sensors to measure angular position and speed |
[43] | Actuated leg-orthosis system | VL, RF and ST muscles | The sEMG signals were full wave rectified. Subsequently, a low pass filter was used. Finally, the processed signals were used as inputs of a Hill-type muscle model | Online | One healthy subject | The experimental torque was computed by employing the inverse dynamics | White box | From the experimental results the authors obtained a calibration accuracy with an RMSE ranging between 1.49 and 1.99 N m and the average R2 was 0.89 | The calibration process is subject-specific. The algorithm uses a white-box model, which makes it easier to understand. It was tested in an online fashion | The model requires knowing parameters such as the lengths of the muscles involved. The study only considers the knee joint |
[44] | None | Muscles of the quadriceps | Adaptive neural networks and fuzzy logic | Offline | One Healthy subject | Training the neural network and set the inference rules of the fuzzy logic | Black box | The performance of the algorithm after least square reached the desired torque level with a mean square error of 181.8 | This model uses different types of EMG-Torque profiles in one neural network. Many muscle activation profiles are used to estimate knee joint torque at different impedance levels that experiment the patient | It requires a training data set. Furthermore, it needs to set inference rules for the fuzzy logic |
[40] | None | VL | RMS of the sEMG signal. Subsequently, a particle swarm optimization (PSO) technique was used | Offline | One healthy subject | Training the algorithm | Black box | A Torque sum squared error ranging between 6148.26 and 25330.10. An average coefficient of determination \((R^{2})\) of 0.88 | The mathematical model for torque estimation is easy to implement since the equations are simple | The algorithm was tested on a single muscle and a single joint (knee) |
[42] | None | sEMG signals collected from VM, VL, Vastus intermedius and RF. Knee angle | The sEMG signals were rectified, a 6 Hz low pass filter was used, and the signals were normalized to be used as inputs to a Hill-type muscle model. The parameters of the model were optimized with quadratic minimums from a nominal torque signal and the torque signal estimated by the model | Offline | One healthy subject | Training the torque estimation algorithm | White box | The lowest error corresponds to the Sequence iii proposed by the authors and the cost function 1, also proposed by them 0.68% | The torque found after the identification of muscle parameters tendon can be used to detect the parameters of a model with reasonable accuracy | For the torque estimation, they used additional sensors such as an isokinetic dynamometer to measure torque and angular position. The latter is input to the muscle model |
[48] | HAL 3: four-link and three-joint | Extensor and the flexor of the knee and the hip | Method to assist motion through torque assistance corresponding to the operator’s intention | Online | One healthy subject | Manual calibration | Black box | With an assist ratio \(\hbox {Gr} = 0.6\), the result shows that EMG and the assist torque approach the constant values during walking. This result means that the myoelectricity is controlled by adjusting the assist torque | The algorithm is designed to assist movement and torque when walking | The algorithm was tested on a single healthy subject and uses additional floor reaction force sensors |
[45] | Leg exoskeleton | sEMG signals collected from RF, VL, and ST. Force and hall sensors | A dynamic human body model and the DFC of the actuator. In both approaches, a high-level control loop evaluates EMG signals and the current state of the human body and orthosis. The output is the desired motion expressed, as either the desired knee angle or torque | Online | One healthy subject | Isometric contractions of the knee flexor and extensor muscles without floor contact for the RF and ST are used for calibration | White box | The knee torque derived from the EMG signals is significantly lower compared to the trial without support | The algorithm used a function to obtain the force from the signal of sEMG to be used as input in the biomechanical model and thus be able to obtain the torque of the knee | The algorithm was designed to work only in the knee joint and was tested only to climb a step with two levels. The algorithm only was tested with a healthy subject |
[49] | KAFO | Left soleus (Sol), tibialis anterior (TA), VL and medial hamstrings (MH) | A physiologically-inspired controller to control artificial muscle forces using sEMG signals. Each artificial pneumatic muscle is controlled by a sEMG signal generated by a biological muscle, e.g, at the knee, they used VL to control the two artificial knee extensors and MH to control the two artificial knee flexors | Online | Three healthy male subjects | Simple tests were carried out to verify that the electrodes placement give appropriate signals for each muscle | White box | This robot produced a 22–33% of the peak knee flexor moment, a 15–33% of the peak extensor moment, a 42–46% of the peak plantar flexor moment, and a 83–129% of the peak dorsiflexor moment, all of this during regular walking | The algorithm includes an inspired control of the physiology of the knee and ankle. This algorithm controls every artificial pneumatic muscle with EMG signals | The algorithm was only tested in 3 healthy subjects. An additional component is needed to manage the pneumatic muscles |
[46] | Exoskeleton with 2-DOF, hip and knee | EMG signals from the biceps muscle and quadriceps muscle of the thigh. Also, the angle and the interaction forces are measured | They proposed a bidirectional human–machine interface including a neuro-fuzzy controller, based on EMG signals, and extended physiological proprioception (EPP) feedback system is developed by imitating the biological closed-loop control system of the human body | Online | A healthy male subject and a healthy male subject | sEMG signals and interaction forces were used to train the neuro-fuzzy network | Black box | The interaction force of the controller without the EMG feeding-forward item is more significant. The average value is 22.65 N after 100 tests, while the average value of the controller with EMG feeding-forward item is 12.46 N, which is 44.97% smaller than the previous one | The algorithm includes an extended physiological proprioception feedback system. It uses a neuro-fuzzy controller to decode human movement using sEMG signals that reflect the intention of the movement and the proprioception of angular feedback | The algorithm was only tested with two subjects. The system needs data previous to be used as a training sample to modify the parameters in the neuro-fuzzy network |
[47] | Exoskeleton system with 1-DOF joints: hip and knee | Information including sEMG, joint angle, and force are collected and analyzed in real time | Active-compliance control of the human–machine system is established based on real-time muscle force estimation and human–machine interactive force detection, while progressive treatment in accordance with stroke stage is realized by timely evaluation. EPP feedback system based on tactile stimuli is developed to help rebuild the closed-loop control system of the human body | Online | Three healthy male subjects | Not reported | White box | During the FE exercises, the interactive force remains from − 10 N to 10 N and the RMS value is 4.35 N, it indicates the exoskeleton joint can follow the movement of the human knee | In their rehabilitation system, an active coupling is mounted on a standing bed. It is designed to guarantee a comfortable and safe rehabilitation according to the structure and control requirements | The algorithm was tested on three healthy people. Besides, it uses an additional sensor for the extended physiological proprioception system, which generates greater data processing |