Artificially induced joint movement control with musculoskeletal model-integrated iterative learning algorithm

Abstract

Neuromuscular electrical stimulation (NMES) delivers tiny electrical impulses to artificially induce muscle contraction, which can be used for the purpose of neurological rehabilitation or muscle strength enhancement. For neurological patients, NMES is not only able to artificially induce movements but also to improve physiological functions and proprioception. To make the induced movements functional, achieving accurate and persistent movement is important but difficult due to highly nonlinear time-variant properties of human musculoskeletal system. This paper proposes a musculoskeletal model-integrated iterative learning control (MMILC) strategy where the musculoskeletal model accelerates the self-learning of iterative learning control (ILC) to accomplish repetitive joint movements. Taking advantage of the feedforward information from the musculoskeletal dynamic model and the self-learning ability of ILC, fast response of the controller and accurate tracking performance can be achieved simultaneously. Both simulation and experiment research are conducted to evaluate the performance of the proposed MMILC in NMES-induced joint movement control. Comparing with traditional ILC, the proposed MMILC illustrates faster and better tracking performance in the control of typical knee joint movements. From the statistic results on six able-bodied subjects, the proposed MMILC can compensate the subject-specific differences in musculoskeletal characteristics responding to the artificial electrical stimulation and represent feasible, effective control of NMES to achieve different motor control purposes.

Introduction

Neuromuscular electrical stimulation (NMES) is an advanced technology which can be used for rehabilitation or training purpose. The persons who can benefit from NMES include patients with neurological diseases, functional disorders or defects [1] as well as able-bodied persons to relieve pain, enhance muscle strength and so on [2]. Reinnervation of the unimpaired nerves or muscles through NMES cannot only induce muscle contractions and corresponding movements but also improve physiological conditions, proprioception abilities and accelerate neural regeneration [3], [4]. The application of NMES in limb movement rehabilitation is limited, though it has been widely applied for heart beats and muscle exercise. One of the challenges is that the artificially induced movement cannot be persistently accurate as the motor task requires. Suitable stimulation strategy and accurate stimulation to the targeted nerves or muscles are therefore important to achieve the expected rehabilitation or training effects.

The existing commercial NMES systems can only provide predefined rehabilitation options with open-loop stimulation strategies. Once the induced movement deviates from the desired, not only the rehabilitation effect but also the safety of the subjects cannot be guaranteed as the patients lost their proprioception together with motor ability [5]. In the research of NMES, closed-loop stimulation may link the subject himself/herself and the stimulation by the human-machine interface (HMI) technology, such as brain-machine interface (BCI). In this case, the desired movement can be triggered from the decoding of the subject's movement intention [6], [7], [8], [9], [10]. Apart from the closed-loop between the user and the stimulation, the closed-loop between the stimulation and the corresponding motor response is also important and has been extensively researched [11], [12]. Such closed-loop focuses on resolving the deviation between the induced and the desired motor response. In fact, due to the time-variant and highly nonlinear properties between the electrical stimulation and muscle response [13], as well as individual differences in physical and physiological properties [14], the deviation from the desired response is often serious so that the motor response cannot contribute to the expected motor function and even make the user in danger (such as falling down if the lower limb joint cannot be actuated correctly and sufficiently) without necessary proprioception feedback in patients. That is also the main reason why the application of NMES for limb control is far behind that for pacemaker [15].

To address the performance of NMES for limb control, a lot of research were conducted to improve the accuracy and/or persistence of the joint movement in closed-loop. Wherein, musculoskeletal model-integrated control [16], [17], [18] allows to explain the internal principles of the movement rebuilt by NMES. In comparison, self-learning control such as neural network [19] and iterative learning control (ILC), is promising to have better tracking performance without the prior knowledge of the musculoskeletal system. To our knowledge, it is difficult to identify a musculoskeletal model to precisely describe the intrinsic mechanism of muscles, ligament and articular forces. The computation cost of neural network is too high to be applied for online tracking tasks. As for ILC algorithm, it was firstly proposed in 1984 and applied in time-variant and continuous systems [20] and has ever been applied to control the repetitive movement of industrial robot [21]. Its application was also extended to rehabilitation domain especially for some rhythmic movement control [22], [23], [24].

As we know, lower limb movements such as standing, walking and running are significant in endowing human outstanding locomotion ability [25]. Most lower limb movements represent rhythmic and periodic motor pattern under the regulation of spinal central pattern generators [26]. Due to the periodic characteristics, the control of NMES for lower limb motor recovery can be special comparing with voluntary movement control. In the existing research, ILC has been proposed to track repetitive movements with the assistance of NMES [27], [28], [29]. To achieve desired ankle movement for drop foot correction, ILC was used to induce 1–2 steps with NMES [30], [31]. ILC was also applied to control the electrical stimulation to the triceps to generate expected repetitive elbow movements for reaching tasks on unimpaired subjects [32], [33]. Although ILC presented its advantage in NMES control to generate repetitive joint movement, its performance in fast learning and accurate tracking can be further improved. It is especially important for NMES control where the reinnervated muscles are apt to be fatigued due to converse recruitment sequence with artificial electrical stimulation [34]. The balance between the learning efficiency and effectiveness will bring significant benefit to the users in effective movement regaining and sound experience as well.

Taking advantage of the prior information of the musculoskeletal systems and the self-learning ability of ILC, this paper proposes to integrate musculoskeletal model into the ILC algorithm to increase the self-learning ability of ILC. The musculoskeletal dynamic model contributes to accelerating the self-learning of ILC. In turn, the self-learning ability of ILC contributes to lowering the accuracy requirement of the musculoskeletal model. The proposed musculoskeletal model-integrated ILC (MMILC) finally guarantees both the accuracy and efficiency of the electrical stimulation control to achieve expected movement. The proposed control strategy is evaluated in simulation and experiment research for NMES-induced joint movement control.