LEE- Sistem Dinamiği ve Kontrol Lisansüstü Programı

Bu topluluk için Kalıcı Uri

http://hdl.handle.net/11527/19459

Gözat

Yazar "Durmuş, İbrahim Hakkı" ile LEE- Sistem Dinamiği ve Kontrol Lisansüstü Programı'a göz atma
  • Öge
    Learning to walk on a human musculoskeletal system with powered prostheses
    (Lisansüstü Eğitim Enstitüsü, 2022-10-25) Durmuş, İbrahim Hakkı ; Yalçın, Hülya ; 503181621 ; System Dynamics and Control
    Generally, due to traumatic or vascular disorders, in some people, the lower leg is cut from below the knee and separated from the body. This procedure is called transtibial amputation. These people have difficulty in activities that require active use of leg muscles such as walking, running, climbing and getting support from the ground, and are generally unable to perform them, due to the loss of the integrity of one of their legs. Various prostheses have been designed so that transtibial amputees can perform walking, which is of great importance in daily activities. Among these prostheses, passive prostheses that do not produce active power and do not have a control loop, but are relatively more accessible for these reasons, are frequently used. Thus, individuals can use their shortened limbs to transfer loads to the floor with the help of passive prostheses and perform the walking action. However, in addition to the cost advantage and good accessibility, these passive prostheses cannot fully provide the functions of a healthy leg that benefits from the lost ankle joint and surrounding muscles. Among these functions, active force generation and variable shock damping are the most prominent ones. Active force generation is involved in the propulsion of the human body by pushing the body forward. Variable impact damping, on the other hand, serves to absorb the impact created by the instantaneous force on the ground at different paces and at different ground height changes. The loss of these important functions leads the person to complete these deficiencies in different ways. For example, passive prosthesis does not contribute actively to progress, and it must be lifted and carried forward with the help of other muscles in the body. Or, the amount of impact damping depends on the structural design and material properties of the passive prosthesis with constant rigidity and cannot be changed instantly. For this reason, people have to provide the impacts that occur against environmental variables and changing the walking tempo by changing their walking styles. It is seen that the hip workload increases in this type of walking. As a result of all these effects, the gait of the amputated people changes compared to before the amputation. Various muscles have to work harder, while various joints tend to angle differently from normal walking. As a result, the muscles and joints in the body work more and the comfort level decreases. In order to prevent these problems, it is aimed to add healthy ankle functions to newly designed prostheses. This type of prosthesis is called active prosthesis. Active prostheses aim to provide a gait closer to normal gait with various control strategies. Control strategies are made with predictive models created with previously collected data, and studies are also carried out on controllers with myoneural interfaces that act directly with the will of the person using it. In order for the designs that emerged through research to become the final product, they need to go through a number of design processes. Clinical testing and prototype productions are included in these design processes. Experiments in various human groups are required for clinical testing. Prototype productions are repeated with changes in design. These processes come back in time and cost, and from time to time they appear as an obstacle or limiting factor in the work. Reducing these processes is important in terms of increasing the target audience and scope of the studies to be carried out. Computer-aided design and simulation tools have long been used in various industries to accelerate design processes. Generally, these tools create three-dimensional or two-dimensional designs of artificial systems and enable them to be developed by evaluating them with simulations in terms of various design constraints. These simulation environments also include static and dynamic evaluations. However, these systems mostly consist of mechanical systems. It is difficult to model and simulate biological systems such as the human musculoskeletal system. Models created with these difficulties had to be created with a lower number of degrees of freedom, actuators and rigid parts than the real system for a long time, due to the high need for processing capacity to be used in the control phase. This situation has recently changed with the relative increase in accessible processing power, resulting in computer modelling of structures much more similar to the human musculoskeletal system. Models of the musculoskeletal system more similar to the degrees of freedom in humans can be manipulated with deep reinforcement learning controllers. These controllers can imitate the walking animations given to them to learn by using the muscles in the musculoskeletal system. Although it is still early to test behavioural movements such as long-term movement planning in a simulation environment, there is a great opportunity to test various artificial systems that will interact with the musculoskeletal system. Researchers working at Seoul National University and Seoul National University Hospital have shared the source code of musculoskeletal system and deep reinforcement learning controllers for use in further studies. In the thesis study, various characters with a deep reinforcement learning controller with the shared musculoskeletal system were created. The purpose of using these characters is to test active prosthetic controllers in interaction with a walking musculoskeletal system. The first of the characters represents a healthy person with the original musculoskeletal system, while the other three characters are transtibial amputees. One of the amputee characters uses a simple passive prosthesis, while the other two characters use two different active prostheses with their own controllers. Controllers of active prostheses make use of a prediction model. The prediction model was created with the data collected during the walking of the previously trained characters. In the first active prosthesis, the prediction model created by walking a healthy virtual character, while the prediction model created by walking a character with a passive prosthesis was used in the second active prosthesis. Walking performances were evaluated by running four different character simulations. The reward function, which is also used in deep reinforcement learning, was used as an evaluation criterion. Common limbs were evaluated for all four characters. Thus, it is ensured that the error used in the reward function is not low for characters with less limb number and no advantage is given due to calculation. When the reward values obtained in walking were examined, it was seen that the character with a healthy musculoskeletal structure exhibited the highest performance. Passive prosthetic character remained at the lowest performance value. Among the characters with active prosthetics, it was observed that the character using the passive prosthetic gait data in the prediction model provided better performance. An exemplary study was obtained for the use of virtual characters in design processes. At the same time, the muscle activations used by the characters during walking were recorded for further studies.