Adaptive Control System for Functional Electrical Stimulation [Research Paper]

Adaptive Control System for Functional Electrical Stimulation [Research Paper]

Adaptive Closed-Loop Control System for Functional Electrical Stimulation (FES)

An Approach for Improving Mobility and Control in Neuromotor Prosthesis

David Ojika

Department of Electrical Engineering

California State University Los Angles





One of the major difficulties faced by those who are fitted with prosthetic devices is the enormous mental effort needed during the first stages of training which becomes even more dramatic with myographic prosthesis were muscle fatigue causes an additional burden to the user. A lot of the approaches to FES design have considered only a smaller part of the entire neuromuscular functioning of the body and tried replicate that function in isolation with the rest of the system. The general outcome of the approach is a system that does not fully satisfy the desire of the end-user. To correct this problem, and to efficiently replicate the function of a desired part of the body via electrical stimulation, a more pervasive approach away from the linear and traditional command and control signaling is proposed. This paper focuses on possible design improvements on existing FES systems by first breaking out its major components and then proposing entirely new solutions, or combining some already existing approaches to form what may be considered as an optimal hybrid solution. The areas of focus are pattern recognition for signal processing, sensing and feedback control, the main control unit and a supervisory unit.



Functional Electrical Stimulation is a technique that uses low levels of electrical current to stimulate nerves innervating extremities which have been impaired as a result of a head injury, spinal cord injury, stoke or other neurological disorders. FES is not a cure but can be used to restore function in people with disability. When used in muscular activities like walking or even in the cardiovascular aspect of breathing, it can be referred to as Neuromuscular Electrical Stimulation (NMES).


FES devices have been used commercially in the past to treat foot-drop problems in some patients with multiple sclerosis by stimulating the peroneal nerve during gait cycles.  Foot-drop is a condition caused by weakness or paralysis of the muscles involved in lifting part of the foot while walking. As a result of that, walking becomes a challenge and causes a person to either drag their foot or engage in irregular stepping pattern called gait cycle.


Generally, FES devices that are used to aid walking in disable persons are small, self-contained electronic devices which are attached to the leg just below the knee. They enable foot-lifting as a person walks, by electrically stimulating the nerve around the muscle associated with the stepping gait. During the swing phase, the device electronically stimulates the peroneal nerves, thereby activating the muscles that cause ankle dorsiflexion which may thus improve the person’s ability to walk.


Several of the approaches to FES design proposed by many authors in the past few years sought to address various challenges in the efficient use of FES devices especially in neuromuscular activities. Many FES devices, nonetheless, continue to exist as invasive or non-invasive, or as open-loop control systems, while newly proposed alternative designs still await long, exhaustive approval process from the FDA. A lot of the new design approaches considered a given and smaller part of the entire neuromuscular functioning of the body, and tried replicate that function in isolation with the rest of the system. The work in [1] demonstrated separately, the extraction of  Electromygraphic (EMG), Electronystagmogramic (ENG) and Electroencephalographic  (EEG) biopotentials  from the body for command and feedback purposes without a connection as to how these biopotentials relate to each other. Even though this technique might seem to be easier to design and implement, the general outcome is a system that does not fully satisfy the desire of the end-user.


In order to design a device that will yield a hundred percent satisfaction by the user whenever in use, the complete functioning of the neuromuscular system of the body should be looked at, having the interrelationship between the key components like the brain, nerves and muscles in mind. Particularly for patients with spinal cord injury, and whose mobility have been completely or partially impaired as a result of that, the linkage between the brain and the nerves is non-existent and will therefore not constitute a component in the design of such a system. To truly model this complex relationship, and to correctly replicate the function of a desired part of the body via electrical stimulation, a more pervasive approach away from the linear and traditional command and control signaling should be adopted.



This paper focuses on possible design improvements on existing FES systems by first, breaking out its major components, and then proposing entirely new solutions or combining some already existing approaches to form what may be considered as an optimal hybrid solution. The areas of focus are pattern recognition for signal processing, sensing and feedback control, the main system unit and a supervisory unit.

3.2 Pattern Recognition

Artificial Neural Networks or ANN have been proposed and used in many applications including hand-gesture recognition, face recognition, voice recognition and many more. ANNs may probably be the single most successful technology that have been used in the past two decades in a wide variety of applications in many areas. Its architecture offers improved performance and desirable properties due to its ability to model complex relationships between inputs and outputs or to find certain patterns in data. One major drawback however to this approach, especially in robotics, is the extensive diversity of training required by the system for a real-world application, but this problem can be mitigated with advanced hardware-software co-design processing techniques.  Artificial neural networks can also be completely autonomous and learn from input from outside and past experiences, or even self-teaching from sets of written rules unlike the traditional programming constructs where hard-written codes explicitly determine the operation of the system. The human body is composed of a complex interconnection of neurons which, with the help of the brain and spinal cord that form the central and peripheral nervous systems respectively, help to transmit biological signals to different parts of the body. It is this inherent nature of complex biological neural networks that inspired the development of a mathematical model called Artificial Neural Networks.


A prosthesis control based on neural networks for muscle EMG pattern classification was proposed in [2]. The system is based on real hardware and software for detecting and processing EMG signals. The outputs of the neural network were used to control the movements of a virtual prosthesis which mimicked what a real prosthesis would be doing. This strategy resulted in a 100% success rate when recognizing and classifying EMG signals initially collected from the muscles. Due to the relatively low computation requirements of ALN, they have been implemented in real-time in ALN based foot-drop correction system [6]. In this research, therefore, we propose an Artificial Logic Network model as the prime choice for processing the acquired potentials of interests from the body and for onward delivery of the output data to a central control unit. This approach will apply machine-learning techniques to sensory signals for accurate detection of gait events and muscle kinetics – under the supervision of brain signals – in order to accurately stimulate the muscle for desired joint angle, position and timing. The ALN implementation proposed is one based on FPGA that allows for a high performance neural processor as well as enabling a rapid reconfiguration of the cells in a node that make up the computational units of the neural processor. In a similar work in [8] where an adaptive logic network was implemented on an FPGA board, their approach enables the generation of hardware necessary to evaluate the very large combinatorial functions that results from an ALN.


3.3 Sensing and Feedback

Functional Electrical Stimulation is the most commonly used technology for improving motor function in individuals with spinal cord injury but only a few applications have successfully generated electrical stimulations that yield efficient functional outcomes. Much of the work in this field still remains within the confines of many research laboratories. Most FES devices in use today use an open-loop control system for the electrical stimulation of dysfunctional muscle systems. Even though a closed-loop control system would introduce a better stimulation pattern, it hasn’t seen much adaptation mostly due to the non-linear and time-varying effects of the human body muscle dynamics. The action of electrical stimulation to the muscles usually results in unpredictable reflexes that can generate strong perturbations. Some of the anomalies that have been studied to be a negative contributing parameter in the feedback control of artificial stimulation systems are muscle fatigue, muscle spasms and muscle retraining. All of these symptoms are visible in persons with SCI. To correct this problem, the FES control system should be able to detect and compensate for these undesirable inputs and exogenous disturbances. The research in [3] reported several closed-loop control systems that have been tested experimentally in the past for electrically induced knee extension against gravity. PID or Proportional integral Derivative, gain scheduling and sliding mode closed-loop control are some of the algorithms that all exhibited degradation in performance when these real-world nonlinear effects were included in the simulation.


Based on the foregoing, it is clear that the effective control of the muscle to carry out a desired function at the right timing and torque – after it has been electrically stimulated – lies in the compensation of the offset between the desired and actual muscle contraction at all times. If these muscles, or the afferent (sensory) nerves that they are directly attached to, can be correctly measured quantitatively at the right precision and timing, the error correction and feedback control would progressively reproduce electrical stimulations that would yield desirable muscular activities.


There are two broad sensor categories that can be used to measure muscular activity. Firstly, the external sensors and electrodes are non invasive and measure joint angles and muscle strain. Some of the sensors that have been used in this approach are gyroscopes, accelerometers, goniometers and position sensors. This approach would require external device attachments to parts of the human body and may not be convenient and esthetically appealing. Additionally the precision of these external sensors and their material property would introduce limitations to the accuracy of the measurement.


On the other hand, natural sensors which have reappeared in many research works for muscular feedback control are quite promising in terms of their ability to precisely characterize muscle reactions bio-electrically. Electroneurogram (ENG) is the resulting biopotential due to the electrical activity of the neurons, and EMG results from the electrical activity of the muscles cells. Both have been exploited as possible biopotentials for direct command and feedback signaling in functional electrical stimulation systems.   By harnessing the intrinsic characteristics of these biopotentials during muscle flexion and extension, the desired control output can be correctly achieved by a stimulator after pattern recognition and classification by the proposed ALN and onward processing by a digital controller.  The work in [4] reported that when muscle activity (recorded in the nerve cuff) and additional information about the gait cycle (possibly recorded*) with EMG were incorporated, functional detection ratios approaching 100% were achieved. Therefore, in conjunction with pattern recognition and classification processes by the proposed ALN, we propose a feedback control system incorporating a systematic combination of ENG and EMG. Both signals will serve as input to the main system control via the ALN.


3.4 Supervisory Unit

The desired joint orientation, position and gait timing required to achieve a muscular activity by an individual can effectively be controlled by the kinetics of the underlying muscle groups. There is no better way to judge a desired need than by the person who perceives its final results. For humans, this perception and judgment of achieved output is interpreted by the brain.  To this regard, we propose a supervisory logic, emanating from the brain, to give signals for particular muscular functions and to continuously supervise the output of the functions. There is no doubt however, that this man-machine interface would definitely require some training on how to interpret various brain biopotentials into certain machine functions. With our proposed ALN, EEG biopotentials recorded from the brain can be extracted and classified for use by a digital controller which is designed to generate multichannel electrical stimulations based on a certain class of input.


Numerous works have been done by researchers in the past few years in the field of BCI and EEG processing and for controlling external devices using biological signals from the brain. The work in [5] demonstrated a man-machine training system for development of EEG-based cursor control. The model required the subject to modulate their EEG signals voluntarily by using different thought patterns for different tasks. When electrodes are placed around the sensory motor cortex of the brain, the EEG signals recorded have a movement related potential (MRP) components in them which controls the movement of the desired part of the body. As with the system used to control the movement of a mouse on the screen, if these sensory motor components are correctly extracted and classified, they can be utilized as internal stimuli used to control the stimulation of the FES device when combined with the EEG and ENG (collected for feedback purposes). We propose this supervisory control of the FES stimulation by extracting and using beta signals that form part of the EEG acquired from the brain.  Research shows that Beta rhythm can be recorded over sensory motor cortex which is sensitive to physical and imaginary movements of the extremities.  [5] In order to provide a manual override to this almost autonomous control, we also propose the use of a user button to either enable or disable the entire FES system.


3.4 Main System Unit

The proposed control unit is one based on System-on-Chip (Soc) were signal preprocessing components including filters, ADC and multiplexers as well as RF front-end and complimentary microcontrollers are all integrated into one tiny Integrated Circuit. We also propose an implantable electrode array system attached to a multichannel stimulator within the Soc that is able to electrically stimulate multiple muscle cells based on input received by the ALN and after the stimulation parameter has been determined by the main controller.  An integrated system like this will not only miniaturize and cut down cost of the design, but will also minimize the power consumption of the system. In 2011, Xilinx unveiled the Zynq-7000 SoC family – the industry’s first Extensible Processing Platform (EPP) – supported by an extensive ecosystem of tools and IP providers. We propose this range of chip for a prototype implementation of the main system unit of the FES since they are targeted towards high performance computing and medical applications.




[1] Reconfigurable Processing for Low-Power Digital Signal Processors – David Ojika, Computer Systems Architecture Publication.

[2] Virtual Prosthesis Control Based on Neural Networks for EMG Pattern Classification

[3] A Comparison of Closed-Loop Control Algorithms for Regulating Electrically Stimulated Knee Movements in Individuals with Spinal Cord Injury
[4] Natural versus artificial sensors applied in peroneal nerve stimulation.

[5] Parallel Man–Machine Training in Development of EEG-Based Cursor Control

[6] Optimizing Machine Learning in Functional Electrical Stimulation in Assisted Foot-Drop Correction Based on Neural Networks: A Case Study
[7] Adaptive Filters, Simon Haykin, McMaster University
[8] Implementation of Adaptive Logic Networks on an FPGA board
[9] Medical RTOS

[10] Biopotential as Command and Feedback Signals in Functional Electrical Stimulation Systems

Share this post

Post Comment

Protected by WP Anti Spam