Abstract
Pneumatic muscle actuators (PMAs) own compliant characteristics and are suitable for use in rehabilitation equipment. This paper introduces a rehabilitation robot driven by PMAs devised in the Rehabilitation and Medical Robot Laboratory. Considering high nonlinearities inside PMAs, a single neuron tuned PID controller is carefully designed. Experimental setup is built up and trials are performed. Results demonstrate the proposed advanced PID algorithm can achieve better capacity in position tracking than the conventional PID controller.
1. Introduction
Features of physiological function recession in aging process include decreased limb flexibility, osteoporosis, muscle atrophy, and significant decline in loading capacity of bone tissue. The incidence of acute cardiovascular and cerebrovascular diseases and that of neurological diseases in the aged keep at a high level, and most of these patients have symptoms of hemiplegia [1, 2]. Quantity of patients with limb dyskinesia caused by other diseases, sports injury, and traffic accidents is increasing rapidly. For these patients, besides early surgical treatments and necessary medications, correct and scientific rehabilitation training plays an important role in restoration and improvement of limb motion functions [3]. Many patients suffer from muscle atrophy and lose limb mobility because of improper training methods and failure to get effective rehabilitation training. This unfortunately makes patients suffer tremendously and causes great burden to family and society [4, 5]. A lot of clinical practices have proved that, without scientific and sufficient rehabilitation training, many patients with impairment of limb motion functions cannot restore the walking capacity to normality and have to move in a typical asymmetric attitude. Thus, improvements in patients’ walking capacity and the ability to take care of themselves in lives make a lot of sense to both patients and society. However, dependence only on physiotherapists cannot meet requirements of patients’ rehabilitation because the amount of physiotherapists is small and one physiotherapist lacks the capacity of training several patients simultaneously. Another barrier is that different physiotherapists have different understandings about rehabilitation exercises and different evaluating indicators [6–10]. The above obstacles are harmful in achieving ideal training strength and effects. Besides, loss of necessary means of detection and control makes obstacles in meeting the needs of a large number of patients in modern society. Rehabilitation robots can help patients complete kinds of motion function rehabilitation training, such as arm restoring therapy and ankle rehabilitation [11–13]. Wu et al. designed a three-degrees-of-freedom lower rehabilitation robot involving hip, knee, and ankle joints and proposed an adaptive robust subcontroller for the robot to handle system uncertainties and disturbances from patients [14] Banala et al. developed a robot-assisted gait training algorithm and used a force-field controller to achieve more effective training [15]. Besides, other rehabilitation robots have been developed [16–29]. Fateh and Khoshdel presented a new voltage-based adaptive impedance force control for a lower limb rehabilitation robot, and gradient descent algorithm was adopted to tune impedance parameters to guarantee force controlling effect [30].
Most of the presented rehabilitation robots adopted motors as actuators, which leads the lack of compliance during recovery. In this research, a compliant ankle-rehabilitation robot driven by PMAs is presented and an advanced PID algorithm is devised to handle nonlinearities and disturbances inside the robot. This paper is arranged as follows. Section 2 introduces characteristics of the devised pneumatic muscle actuator-driven ankle rehabilitation robot. Section 3 proposes a single neuron adaptive PID controller for the robot; Section 4 establishes the experimental setup of the robot and performs the rehabilitation tests to validate the designed controller.
2. Structure of the Ankle-Rehabilitation Robot
2.1. Structure Description
The detailed structure of the ankle rehabilitation robot is addressed prudently and shown in Figure 1. Decomposition diagram of the ankle joint mechanism is displayed in Figure 2. One of the crucial issues to be handled is that placing angle and height of the training device must own the capacity of adjustment according to different patients and circumstances. Thus, the robot framework is made of aluminum profiles whose installing grooves are in unique T-shape (shown in Figure 3). By this way, arrangement angle and height of the robot can be adjusted arbitrarily in the framework. Besides, robot mounting panel has slots and holes. Slots allow the robot to move along and be fixed by bolts. Holes help the robot rotate around to modulate the arrangement angles and guarantee comfort during rehabilitation exercises. The panel framework is made up of aluminum profiles, whilst the robot is fabricated by light and high tensile aluminum. Another important issue is to guarantee safety during training exercises. This is realized by the limitation of the allowable angle range of the executive mechanism. Actually, angle limiting device (shown in Figure 4) is devised in the robot to avoid hurt on the patients’ ankles. The range of the executive mechanism in the robot is (−30°, 30°).
In traditional ankle rehabilitation robots, driving components are usually installed on one side of the ankle joint. This type of unilateral asymmetric structure (shown in Figure 5(a)) may exert uncertain force to joint, produces harmful torque on the robot joint in extra directions, and has detrimental influence on mechanical structure stability. Bilateral structure is adopted in the ankle joint design in order to preventing from producing harmful torques in extra directions. This bilateral structure keeps balance in different components of ankle rehabilitation robot and eliminates harmful torques created by drive power and forces from the feet of patients. Power supplying mechanism and transmission mechanism are both installed bilaterally about the ankle joint, which makes identical forces on the two sides of the structure. The symmetrical configuration can further guarantee the stability and reliability of mechanical transmission, make good use of space, and improve compact layout of the robot (shown in Figure 5(b)).
(a)
(b)
2.2. Prototype
The prototype of the robot (shown in Figure 6) adopts aluminum alloy to make sure of light weight and high strength. Aluminum alloy profile frame is used to regulate posture of the robot for various patients. A torque sensor is installed in one side of the output shaft of the transmission mechanism to measure driving torque or the human ankle, and an incremental encoder is fixed with the other side of the identical output shaft to record the real time rotation angle.
3. Single Neuron Tuned Adaptive PID Controller
3.1. Single Neuron Adaptive PID Algorithm
Proportional Integral Differential (i.e., PID) algorithm is widely used in various rehabilitation and medical robots because of concision, efficiency, and reliability. However, high nonlinearity and strong hysteresis of PMAs bring a tricky job in keeping high tracking capacity because of which conventional PID algorithm owns constant values of parameters P, I, and D. Considering single neuron strategy has the excellent capacity in regulating structural parameters, a single neuron tuned PID controller is devised, as shown in Figure 7. , , and are structural parameters of single neuron algorithm and represent P, I, and D of the PID controller, respectively. Values of , , and are online regulated according to some criterions, and this paper adopts a hybrid of supervisory delta learning rule and nonsupervisory Hebbian rule. A cost function is defined as the evaluation indicator of the tuning algorithm in the following equation:
The principle is to reduce by online tuning parameters , , and . Gradient descent algorithm is used as follows:
According to chain rule of differential equations, is calculated as follows:
Usually, is difficult to calculate because of complexity of the actual system. To simplify the calculation, it is replaced by . (i = 1, 2, 3) are inputs of neuron units , , and , respectively. In the single neuron PID controller, are assigned according to tracking errors of the system, i.e.,
The output is as follows:
3.2. Controller Design of the Rehabilitation Robot
The controller is devised using single neuron strategy tuned PID algorithm in Figure 7. Two electrical proportional valves with type of ITV1050-212N made by SMC Corporation are used in the setup to control the pressure inside the pair of PMAs. This type valve has an output range of pressure in (0.005, 0.9) MPa and linear correspondent controlling voltage in (0, 5) V. in Figure 8 is the initial input voltage of the first valve, i.e., the initial pressure inside the first PMA. Similarly, is the initial input voltage of the second valve, i.e., the initial pressure inside the second PMA. Values of and are tuned by the trial and error. is the transfer function of the proportional valve and has the following form:
Torque switch in the controlling block is used to guarantee the safety of patients by setting a range of torque. If the sampled torque value is larger than the upper limit Tmax, or less than lower limit Tmin, proportional valves will output zero to PMAs and the robot stops working. Tmax and Tmin are acquired by trial and error.
3.3. Stability Analysis of the Proposed Controller
Controlling parameters of single neuron PID algorithm vary along with the direction of negative gradient descent of , , and . Stability of the single neuron PID controller is analyzed by Lyapunov principle. Firstly, a Lyapunov function is defined as follows:
Variation of in the self-learning process of the single neuron model is expressed:
Similarly, variation of can be acquired as follows:
Consideringthenwhere , .
Thus,
Several conclusions can be acquired by Lyapunov principle:(a) is positive definite(b)If and only if , Δ is negative definite, which means stability of the system depends on the learning step(c)When k approaches infinity, approaches zero
Obviously, proper values of η make Δ < 0, which means the controlling system keeps stable.
4. Experimental Validations and Discussion
The entire robot system, including controlling component and air compressor, is shown in Figure 9. An AD/DA card of USB3102A type from Beijing Art Technology Development Co., Ltd., is adopted to control the robot and samples all status information from sensors. An air compressor is employed as the compressed air supplier. Main components in the robot are listed in Table 1.
(a)
(b)
To validate the effectiveness of the proposed single neuron tuned PID algorithm, several trials are conducted under different excitation. Experiments are performed on a male with weight of 77 kg and height of 176 mm in this study (shown in Figure 10). It is generally known that passive rehabilitation therapy is used for serious dyskinesia of limbs and must operate at a very low speed. A half sinusoidal wave excitation of 0.004 Hz with amplitude 50° is applied to the robot, and tracking responses and errors are plotted in Figures 11 and 12, respectively. Relative large errors occur at the zero position due to assembly errors of the robot and creep and high nonlinearities inside PMA. Figure 11(a) shows that the maximum tracking error of the PID controller is 5.905° at the time of 548.3 s, whilst that of single neuron tuned PID is 4.541°. Figure 11(b) shows an enlarged part for the portion A in Figure 11(a), from where the conclusion that single neuron tuned PID controller achieves smaller tracking errors than conventional PID controller can be drawn. Figure 12 shows tracking errors of passive rehabilitation therapy in Figure 11, which further proves the better capacity of single neuron regulated PID than classic PID. Figure 13 displays a half sinusoidal wave excitation of 0.002 Hz with amplitude 50° and different responses of the adaptive PID controller and classic PID controller. Figure 13(a) shows a maximum tracking error of 3.364° from PID controller responses and a maximum tracking error of 2.603°. Figure 13(b) shows an enlarged part for the portion in Figure 13(a), which further proves that the single neuron tuned PID controller achieves smaller tracking errors than the conventional PID controller. Figure 8 shows the corresponding tracking error performances in Figure 13, and comparison between response curves of the two controllers demonstrates better capacity of single neuron tuned PID algorithm in overcoming nonlinearities and disturbances of the robot.
(a)
(b)
(a)
(b)
5. Conclusions
Compliant ankle rehabilitation robot in this research is realized by employing pneumatic muscle actuators as the power source. A pair of PMAs is arranged in antagonistic form and provides driving torque through crank-slider mechanism. Prototype is manufactured and assembled. A single neuron tuned PID controller with torque safety switch is designed for the robot. Passive rehabilitation experiments are conducted and effectiveness of the adaptive controller is validated. Conclusion of experiments demonstrates that the devised controlling algorithm can improve accuracy of position tracking of the robot.
Data Availability
The datasets used or analyzed during the current study are available from the corresponding author on reasonable request.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Acknowledgments
This work was supported in part by the National Key R&D Program of China under Grants 2019YFC1511400, 2019YFC1511401, and 2019YFC1511402, Pre-research Union Fund of China Ministry of Education && PLA Equipment Development Department under Grant 6141A02033616, Fundamental Research Funds for the Central Universities under Grant B200202223, Strategic Rocket Innovation Fund of National Defense Science and Technology Industry under Grant Y18143, Natural Science Foundation of Jiangsu Province under Grant BK20170301, and Opening Fund of Jiangsu Key Laboratory of Special Robot Technology under Grant 2017JSJQR01.