로봇 시스템의 근전도 기반 인간 동작 인식 및 연속 제어 방법

Abstract

This dissertation suggests electromyography (EMG)-based continual control and human motion recognition method for robotic systems such as a tel-operation device, a index finger prosthesis and a EMG-based electric wheelchair. The EMG signals are utilized as control input signals of prostheses considered muscle contraction dynamic and skeleton structure, and muscle robot interfaces to control robotic systems.

Since a raw EMG signal consists the summation of single motor unit action potential (SMUAP) generated from each muscle fiber innervated by the motor neurons, the EMG signal is very noisy. Therefore, the RMS operation is employed to extract the envelopes of the signal waveforms to reflect the moving average activities. The resulting signal of the RMS operation is utilized as the control signal for the robotic devices.

The EMG-based control method can be classified into two groups, according to whether a pattern of EMG signal is recognized or not. In the non-pattern recognition-based methods, the EMG signals are utilized as control input signals for the controllers and the motion extractors. However, the non-pattern recognition-based controller has a limitation of extending the multiple motion commands. In the pattern recognition-based methods, this EMG-based methods employ recognition techniques to extract the motion commands for controlling the robotic device for any users. However, the pattern recognition-based methods have limitations of the computational complexity and high cost compared to the non-pattern recognition methods.

To solve the problems, this dissertation suggests the EMG-based continual control methods which consists of a proportional controller to control the position, velocity, and torque of the robotic systems, and human motion recognition methods to switch controllers of the robotic systems. The suggested recognition methods can recognize the muscle continual commands during the voluntary muscle contractions within a short prior training time.

These subjects will be thoroughly discussed in this dissertation. Initially, The parameter of muscle contraction model extraction method is suggested using the human movement motion. Since the muscle skeleton is involved in other muscle skeleton, It makes a signal processing and a controller complex. Therefore, the EMG-based Continual Control method for movement extraction is suggested to consider the muscle skeleton.

Second, the non-pattern recognition-based methods have the main limitation to increase the number of control commands because of a muscle co-activation that a muscle and other muscle are activated coordinately when only one motion exercises. Also the number of muscles are not enough to generate the independent muscle commands. Even the users such as quadriplegia patients have three facial muscles or two shoulder muscles to generate the muscle commands. Therefore, the EMG-based continual control method for a muscle robot interface is suggested to solve the main limitation to increase the number of control commands in non-pattern recognition-based approaches.

Third, the pattern recognition-based methods employ recognition techniques to extract the motion commands to control the robotic systems for any users. As a result, the methods are able to adapt itself to physical and physiological conditions of any users. However, the pattern recognition-based methods require computational high cost compared to the non-pattern recognition methods. Therefore, the motion recognition method for EMG-based continual control is suggested to reduce the computational complexity and high cost.

As a result, The continual muscle commands are classified by using the suggested recognition methods during the voluntary muscle contractions within a short prior training time. The controllers of the robotic systems are switched by classifying the continual muscle commands. Additionally, The continual muscle commands can be utilized as input signals to control the position, velocity and torque of robotic systems.| 본 논문은 원격조정, 검지의수, 전동휠체어 등과 같은 로봇시스템을 위한 근전도기반 연속제어 및 동작인식 방법을 제안한다. 근전도 신호는 근육수축과 뼈구조가 고려된 의수나 로봇시스템을 조작하기 위한 근육로봇 인터페이스의 입력신호로 사용된다. 근전도 신호는 모터뉴런이 각각의 근섬유를 자극하여 생성된 단일 모터 유닛 활동 전위의 합으로 구성되므로, 매우 잡음이 크다. 그러므로 이동 평균이 반영된 신호파형형태의 포락선을 추출하는 제곱 평균방법을 사용한다. 제곱 평균을 취한 신호는 로봇 장치의 제어 신호로 사용된다. 근전도기반 제어방법은 근전도 신호의 패턴인식 유무에 따라서 크게 두 그룹으로 나눠진다. 비패턴 인식방법에서 근전도 신호는 제어기의 제어입력 신호나 모션추출을 위해서 사용된다. 하지만 비패턴 인식방법의 제어기는 모션명령 수를 증가시키는데 한계성을 가지고 있다. 패턴인식방법에서는 사용자에 상관없이 로봇장치를 제어하기 위한 모션명령을 추출하는 인식 기술을 사용하고 있다. 하지만 비패턴 인식방법은 연산복잡성과 비패턴 인식방법의 연산보다 큰 연산 부하를 가지고 있다. 이런 문제점을 해결하기 위해서 본 논문에서는 로봇 시스템의 위치, 속도, 토크 등을 제어 하기 위한 비례제어기로 구성된 근전도 기반의 연속제어 방법과 로봇 시스템의 제어기를 스위치 하기 위한 인간 모션 인식 방법을 제안하고 있다. 제안된 인식방법은 짧은 트레이닝 시간으로 자의적 근육 수축이 발생할 때 근육의 연속적인 명령을 인식 할 수 있다. 본 논문은 구체적으로 다음과 주제들로 구성되어있다. 첫째로 인간의 움직임 모션을 사용하여 근육수축모델의 파라미터 추출방법이 제안되었다. 근골격은 다른 근골격과 연결되어 있기 때문에 근전도 신호처리와 제어기를 복잡하게 만든다. 그러므로 근골격을 고려한 움직임 추출을 위한 근전도기반 연속제어 방법이 제안되었다. 둘째, 비패턴 인식 방법은 하나의 모션을 할 때, 여러 근육이 동시에 활성화되어 서로 상호작용을 하기 때문에 제어 명령 수의 늘리는데 한계를 가지고 있다. 또한 독립적으로 근육 명령을 생성 할 수 있는 근육의 수가 충분하지 않다. 심지어 사지마비 환자의 경우 얼굴에 3개의 근육이나 어깨의 2개의 근육만 사용 할 수 있다. 그러므로 비패턴 인식 방법을 사용할 때 제어 명령어 수 확장의 문제를 해결하기 위해서 근육과 로봇 사이의 인터페이스를 위한 근전도 기반의 연속제어 방법이 제안되었다. 셋째, 패턴인식 기반 방법은 어떤 사용자든지 로봇 시스템을 제어하도록 하기 위한 모션 명령을 추출하는 인식 방법을 사용하고 있다. 결과적으로, 이 방법은 사용자의 정신적 물리적 상태에 적응 할 수 있지만 이런 방법은 비패턴 인식 방법에 비해서 높은 연산 비용을 요구한다. 그러므로 근전도기반 연속 제어를 위한 연산 복잡도와 상대적으로 높은 연산 비용을 줄이는 모션인식 방법을 제안한다. 결과적으로 제안된 방법은 자의적인 근육 수축을 하는 동안 짧은 트레이닝 시간만으로도 연속적인 근육 명령을 분류가 가능하며, 연속적인 근육 명령을 식별하여 로봇시스템의 제어기를 스위치 한다. 부가적으로 이런 연속적인 명령은 로봇 시스템의 위치, 속도, 토크를 제어 하기 위한 입력 신호로 활용이 가능하다.; This dissertation suggests electromyography (EMG)-based continual control and human motion recognition method for robotic systems such as a tel-operation device, a index finger prosthesis and a EMG-based electric wheelchair. The EMG signals are utilized as control input signals of prostheses considered muscle contraction dynamic and skeleton structure, and muscle robot interfaces to control robotic systems.

Since a raw EMG signal consists the summation of single motor unit action potential (SMUAP) generated from each muscle fiber innervated by the motor neurons, the EMG signal is very noisy. Therefore, the RMS operation is employed to extract the envelopes of the signal waveforms to reflect the moving average activities. The resulting signal of the RMS operation is utilized as the control signal for the robotic devices.

The EMG-based control method can be classified into two groups, according to whether a pattern of EMG signal is recognized or not. In the non-pattern recognition-based methods, the EMG signals are utilized as control input signals for the controllers and the motion extractors. However, the non-pattern recognition-based controller has a limitation of extending the multiple motion commands. In the pattern recognition-based methods, this EMG-based methods employ recognition techniques to extract the motion commands for controlling the robotic device for any users. However, the pattern recognition-based methods have limitations of the computational complexity and high cost compared to the non-pattern recognition methods.

To solve the problems, this dissertation suggests the EMG-based continual control methods which consists of a proportional controller to control the position, velocity, and torque of the robotic systems, and human motion recognition methods to switch controllers of the robotic systems. The suggested recognition methods can recognize the muscle continual commands during the voluntary muscle contractions within a short prior training time.

These subjects will be thoroughly discussed in this dissertation. Initially, The parameter of muscle contraction model extraction method is suggested using the human movement motion. Since the muscle skeleton is involved in other muscle skeleton, It makes a signal processing and a controller complex. Therefore, the EMG-based Continual Control method for movement extraction is suggested to consider the muscle skeleton.

Second, the non-pattern recognition-based methods have the main limitation to increase the number of control commands because of a muscle co-activation that a muscle and other muscle are activated coordinately when only one motion exercises. Also the number of muscles are not enough to generate the independent muscle commands. Even the users such as quadriplegia patients have three facial muscles or two shoulder muscles to generate the muscle commands. Therefore, the EMG-based continual control method for a muscle robot interface is suggested to solve the main limitation to increase the number of control commands in non-pattern recognition-based approaches.

Third, the pattern recognition-based methods employ recognition techniques to extract the motion commands to control the robotic systems for any users. As a result, the methods are able to adapt itself to physical and physiological conditions of any users. However, the pattern recognition-based methods require computational high cost compared to the non-pattern recognition methods. Therefore, the motion recognition method for EMG-based continual control is suggested to reduce the computational complexity and high cost.

As a result, The continual muscle commands are classified by using the suggested recognition methods during the voluntary muscle contractions within a short prior training time. The controllers of the robotic systems are switched by classifying the continual muscle commands. Additionally, The continual muscle commands can be utilized as input signals to control the position, velocity and torque of robotic systems.