In this study, four healthy volunteers were recruited, aged 25–27 years old, in good physical condition and without basic diseases. There was no significant difference in the general data of the four healthy volunteers (p > 0.05). The experiment was conducted at Henan Xiangyu Medical Equipment Co., Ltd (Anyang City, Henan Province, China). All subjects signed informed consent.
3.1. Experimental System
The experimental interactive system of upper limb rehabilitation on-demand assist control strategy based on the subjective intention attenuation model includes upper limb rehabilitation exoskeleton robot body (ARE-II), sEMG signal acquisition sensor BITalino, control box, virtual scene, and so on. The experimental site is shown in
Figure 3. ARE-II has a total of eight degrees of freedom (DOF), which can realize the two-DOF movement of the patient’s scapula, the three-DOF movement of the shoulder joint, the one-DOF movement of the elbow joint, and the two-DOF movement of the wrist joint, and can realize the movement of any posture in the patient’s space. The designed upper limb rehabilitation robot is arranged with a three-dimensional force sensor at the upper and lower arm constraints, respectively, which can directly obtain the interaction force between the upper limb of the patient and the rehabilitation arm.
The upper limb rehabilitation exoskeleton robot (ARE-II) is applied to the rehabilitation training of patients with stroke or spinal cord injury. The left and right are symmetrically arranged. As shown in
Figure 4, the length of the mechanical arm is electrically adjustable to meet different sizes of human body. The
Z-axis of the three joints of the shoulder intersects at one point to form a ball joint. The
Z-axis of the three adjacent joints of joint 3, joint 4, and joint 5 is parallel. The shoulder of the robot is not three-axis perpendicular to each other. The angle between the three rotation axes of the shoulder joint is set as follows: the angle between joint 1 and joint 2 is 60°, the angle between joint 2 and joint 3 is 60°, and the angle between joint 1 and joint 3 is 90°. A fixed coordinate system
is established in the center of the human shoulder joint, and a local fixed coordinate system
is established in the center of the shoulder joint of the upper limb rehabilitation robot. In the initial state, the two coordinate systems coincide. The link coordinate system
is added to each joint of the upper limb rehabilitation robot mechanism, where the
,
, and
motion axes intersect at the axis of the shoulder joint of the robot arm, and the angle between the
and
coordinate axes is 90°.
and
correspond to the length of the upper and lower arms of the human upper limb. The three-dimensional force sensor for detecting the human–computer interaction force is located at the
S1 point on the big arm-connecting rod and the
S2 point on the small arm-connecting rod.
Since the two three-dimensional force sensors are in direct contact with the patient’s limbs, the collected data are the actual human–computer interaction force. In order to accurately calculate the expected assist force in the training task, it is necessary to separate the information of the patient’s body weight applied to the force sensor. In order to quickly solve the weight of different patients’ limbs, the human–computer interaction model is constrained to move in the sagittal plane, and the human–computer interaction model in the sagittal plane is established as shown in
Figure 5. In the figure,
and
are the centroid lengths of the affected limbs, respectively.
and
are the distance between the sensor and the joint length at the human–computer interaction constraint;
and
are the weight of each part of the affected limb; and
,
,
, and
are human–computer interaction forces, respectively.
and
are the joint angles of the manipulator, respectively. In order to facilitate the calculation, two angles
and
are defined in the sagittal human–computer interaction model, where
and
.
According to the principle of virtual work, the torque in each joint of the patient’s limb is balanced under the action of its own gravity and the support force of the mechanical arm. Through the analysis of the geometric relationship, the human–computer interaction force
at the elbow joint is only affected by the weight
of the patient’s forearm. The torque
at the elbow joint is obtained as follows:
Similarly, the human–computer interaction force at the shoulder joint is analyzed, and the shoulder joint torque
is obtained as follows:
The joint variable
in the above formula is obtained in real time through the joint angles of the robotic arm. The joint torque term
is obtained by transforming the human–computer interaction force data with the distance between each sensor and the joint force arm. In order to estimate the limb weight of different users as accurately as possible, the robotic arm drives the affected limb to perform random low-speed motion, during which the affected limb remains relaxed. A total of
k data are collected in the process, and a more accurate patient characteristic parameter item is obtained as follows:
The relationship between the equivalent interaction force at the end of the manipulator and the joint torque is established as follows:
In the following, the input parameter of the admittance control strategy of the upper limb rehabilitation robot is the human–computer interaction force–torque signal at the end of the rehabilitation robotic arm. Therefore, the active interaction force of each joint of the rehabilitation robotic arm is transferred to the end of the robotic arm.
The expected assist force of the training task is as follows:
The electromyographic signal acquisition sensor has five channels, and different electrode attachment points are set according to different experimental schemes. As shown in
Figure 6, CH1 was set on the upper end of the trapezius muscle, CH2 was set on the pectoralis major muscle, CH3 was set on the deltoid middle bundle, CH4 was set on the biceps brachii, and CH5 was set on the olecranon as the reference electrode. The subjective intention attenuation rate of the subjects was evaluated by measuring the sEMG signal level of the subjects in real time.
The construction of the virtual scene is completed on the Unity3D platform. The main purpose is to increase the fun and immersion of the subjects during training. At the same time, the real position of the robot’s end is displayed on the screen, allowing the subjects to grasp the training situation in real time. The object operated in the designed virtual scene is a mouse, and the patient operates the mouse to walk along an ideal trajectory within a safe area. The coordinates of the mouse and the end of the robotic arm are consistent, setting a task participation scoring mechanism; the closer the value is to the ideal trajectory, the higher the participation score. A 15 min training task was designed, in which participants had to resist the weight of their upper limbs and loads to complete the task. A training score of 60 or above was considered an effective experiment.
3.2. Experimental Scheme
In order to verify the effect of this method on the improvement of patients’ initiative, two comparative experiments were set up. In the first experiment, the traditional methods were used, namely, the simple passive training model (M1); the on-demand assist control method (M2) using only the force–position hybrid evaluation index
β; and the on-demand assist control method (M3) using the method in this paper. Four subjects participated in the experiment. During the experiment, in order to simulate the patient’s state, the subjects loaded a load of 10 pounds on the wrist. The experiment required the subjects’ upper limbs to perform circular trajectory motion in the sagittal plane. In the experiment, the subjects needed to try to “deceive” the robot to obtain higher assist force. The specific experimental scheme is shown in
Table 1.
The subjects controlled the upper limb to perform a circular motion trajectory on the sagittal plane. The specific parameters of the experiment were as follows: the center of the circular trajectory (x
0, z
0) = (400, −210) and the radius of motion r = 60 mm. At the same time, the admittance control model parameters were set as follows:
At the initial moment of the experiment, the subjective intention attenuation coefficient () and the limb training state force evaluation parameters were set. The moving time window of the subjective intention attenuation was set to 30 s, and the moving speed was 3 s. The moving time window of the training state force evaluation was set to 3 s, and the moving speed was 0.3 s.
During the experiment, the subject’s subjective scoring of the current fatigue state was recorded every 30 s. The fatigue state was divided into three grades. A higher score indicates that the subject experienced more fatigue in the 30 s time window, and the actual assist force of the robot was recorded in real time. In order to reduce the influence of accidental factors on the experimental results, the subjects were required to have enough rest time between the two experiments, and extensive simulation training was carried out before participating in the experiment. The experiment duration was set to 15 min to ensure the subjects entered the fatigue state.
3.3. Experimental Results
When the subjects participated in different comparative experiments, the human–robot interaction force between the subjects and the robot, the joint angle of the manipulator, and the surface sEMG signal of the subjects were continuously collected. In experiment 2, the end force data of subject C completing the three training tasks are shown in
Figure 7.
From the diagram, it can be seen that in the M1 training task, the subject completely relaxed the limb, and the mechanical arm drove the limb to complete the corresponding training task. In the M2 training task, the level of human–computer interaction remained low for a period of time at the beginning of the training. At 400 s, the subject was too weak to complete the task. At this time, the human–computer interaction force reached its first peak. At 420 s, the subject was challenged again and worked hard to complete the task goal, so the interaction force decreased rapidly. Then, at 500 s, the subject relied on the assist force provided by the robot and deceived the higher assist force by increasing the end force level and fluctuation frequency. In the M3 training task, the subjects constantly attempted to rely on their own strength to complete the training task, so the human–computer interaction force exhibited multiple peaks. From the original data of human–computer interaction force, it can be seen that only relying on the force position information in training to provide the assist force cannot avoid the patient’s dependence on the robot assist force. Therefore, it is necessary to extract the attenuation trend of patients’ subjective training intention from sEMG signals.
In order to ensure the accuracy of the subjective intention attenuation model and improve the calculation speed of the attenuation coefficient, it was necessary to select three to five features from the existing four-channel sEMG signals as the input of the attenuation model. Therefore, statistical analysis was conducted on these 16 features to obtain the correlation significance and correlation. The average
and
values of the data characteristic quantity compared with each other in the three fatigue states are given in
Table 2, where
is the test probability, and
is the influence of random error. When the feature
in the table, it means that the feature quantity has a significant difference between different fatigue degrees.
According to the data listed in the table, CH1 performs poorly. In the comparison of all features in the three fatigue states,
in only
MAV and
SSC, but the corresponding
, meaning that although there were significant differences between features, there were still a large number of overlapping areas in the feature data. In the comparison of all the characteristics of CH2 in the three fatigue states,
MAV,
WL, and
SSC features performed better, and their
p values were all less than 0.05, but in the tired/medium comparison,
, so only
SSC was selected as the input to avoid overfitting of the model. In the comparison of all the characteristics of CH3 in the three fatigue states, the
p-value and
F-value of
RMS and SSC were less than 0.05, which were significant and fatigue-related. As an input, the MAV and
RMS of CH4 had the same distinguishability. Therefore, the
SSC of CH2, the
RMS and
SSC of CH3, and the
MAV and
RMS of CH4 were selected as the input of the subjective intention attenuation model.
Figure 8a,b show the change curves of the five features extracted by the sEMG signal with time when subject D performs the M2 and M3 training tasks, respectively, intercepting the moment when the subjective intention attenuates the “fatigue” state. During the “fatigue” period, the five feature quantities have a significant improvement trend. At the same time, it can be seen that in the second half of the M2 training task, although the feature quantity also has a trend of improvement, the amplitude is not large, and the corresponding subjective evaluation of the subjects is also in a “medium” or “relaxed” state.
In order to verify the effectiveness of the active intention attenuation model, the feature quantities of the five training tasks completed by the four subjects were used as the input parameter
of the training sample set. At the same time, for subjective fatigue evaluation, data were recorded every 30 s in the training task and were used as the known category information
output of the training set, in which the “fatigue” state was set to “1”, the “medium” state was set to “0.6”, and the “easy” state set to “0.2”. The last experimental data were taken as the prediction sample input
and the real output value
Y of the prediction sample in the test set.
Figure 9 illustrates the actual and predicted values of the fatigue state of the subject
participating in the M3 training task. In this experiment, the difference between the predicted value and the true value of 80.57% of the data was within 0.02. If the tolerance range was extended to 0.05, the accuracy of the subjective intention attenuation model could reach 93.14%, and the average accuracy of the subjective intention attenuation recognition of the four subjects reached 87.62%. In order to further analyze the classification effect of QPSO-MLSSVM on volunteer participation status, the minimum MSE, mean MAE, and minimum RMS were used for evaluation, as shown in
Table 3.
Considering experiment 1, the results of the active verification experiment of subject A are shown in
Figure 10. Regarding experiment 2, the results of the fatigue identification and anti-dependence verification experiment of subject C are shown in
Figure 11.
Table 4 presents the two experimental results, and the data in the table are the average values of all subjects.
Subject A was selected to analyze the experimental data of three training tasks. In order to eliminate the influence of the assist force dependence phenomenon on the experimental results when subject A performed the M2 experiment, they were required to concentrate as much as possible and rely on their own strength to complete the task when performing the three training tasks.
The analysis chart shows that when the subject’s affected limb was unstable and in the fatigue state, the state parameter values increased; at this time, the average assist force of the M1 training task robot was 1.7335 N; the average assist force of the M2 training task robot was 0.97 N; the average assist force of the M3 training task robot was 0.6116 N; and the corresponding training scores were 98, 80.7, and 83.8, respectively. Compared with the traditional method, the average assist force of the overall training task was reduced by 64.72%. Compared with the assist force method based on force position information, the average assist force was reduced by 36.95%, indicating that the effective active torque of the affected limb greatly increased. At the same time, when the state of the affected limb was unstable, and the subject was in the fatigue state, the average assist force of the robot in this method was 1.0341 N, and the average assist force of the robot using M2 was 0.9614 N. Compared with the assist force, the two decreased slightly, but the decrease was not large, indicating that the method in this paper can provide sufficient assist force in the fatigue state.
In experiment 2, which was performed in order to verify the phenomenon of assist force dependence in the M2 training task, subject C had to try to “deceive” the robot to generate more assist force when performing this task. At the same time, in order to verify that the subjective intention attenuation model of M3 can prevent the phenomenon of assist force dependence, the M3 method was used to simulate and compare the data in the M2 training task.
Figure 11 shows the performance of subject C in trying to “deceive” the robot to provide more assistive force when performing the M3 training task.
The analysis of the chart shows that when subject C performed the M2 training task. Starting from 450 s, in order to “deceive” the robot to increase the auxiliary force, a deliberately intermittent force was generated to create the phenomenon of unstable motion. The robot increased the average auxiliary force to 1.2313 N. Using M3 to reanalyze the data, the average assist force of the corresponding robot did not increase but rather decreased to 0.6242 N, corresponding to the actual fatigue state of the subject. Therefore, the average assist force was reduced by 49.31% compared with the average assist force of M2. As shown in
Figure 11b, although subject C tried to “deceive” the robot several times, the assistive force did not improve.
Considering the results of subjects C/D participating in the M2 and M3 experiments, the average scores of the tasks and the proportion of subjective fatigue state of the subjects throughout the entire training period are shown in
Table 5. In order to compare the improvement effect of the on-demand assistive force control strategy on patient initiatives, subjects C/D also underwent five training tasks without the assistive admittance control strategy. By comparison, it can be seen that the average scores of M2 and M3 training tasks were 86.3 and 85.5, respectively, with little difference. However, the proportion of fatigue in M3 was much higher than that in M2. It can be seen that under the same task score, the subjective participation of patients in the M3 task was greater. Comparing the M1 task without assistance and the M3 task, it can be seen that the fatigue rate of M1 without assistance was slightly higher than that of M3, but the average score of M1 without assistance was much lower than that of M3. This is mainly because the subjects had to have more initiative to complete the M1 task without assistance training, and their physical strength decreased faster, which could not support the completion of the training task.