| dc.description.abstract | Even if they are not aware of it, healthy humans walk stably and comfortably. Amputee patients, on the other hand, have difficulty performing stable and comfortable walking due to their missing limbs. Powered prostheses have been developed and studied by researchers for decades as a treatment to restore their gaits as closely as possible to those of healthy people. Nonetheless, amputees continue to have several gait deficiencies as a result of a lack of interaction between the users and their prostheses. The user’s gait phase estimation is an important interacting factor. Because the user and the prosthesis should be treated as a coupled system, synchronized prosthesis control is required for stable walking and proper assistance. Researchers attempted to estimate user gait progression based on heel-strike, but this method is not adaptable to individual gait characteristics or speed change. As a result, the focus of this dissertation is on improving the adaptability of user gait phase estimation to individual gait traits and various walking speeds, allowing the powered prosthesis to provide a healthy human-like gait in a synchronized manner.
This dissertation makes three significant contributions to improving the estimation of the user
gait phase during ambulation: i) phase-shifting method; ii) piecewise phase variable; and iii) machine learning-based estimation. First, we propose a custom-built powered transfemoral prosthesis and its underlying control framework capable of mimicking human behavior for stable and synchronized walking with the user. The first chapter then presents a phase-shifting method to improve the user adaptability of gait phase estimation. A phase variable is a kinematic variable that can be used in prosthesis control to estimate the user’s gait phase. The thigh segment angle is particularly useful in calculating phase variables. Human data show that people have a cosine-like trend in their thigh segment movement while walking, which can be used to estimate their walking progression. However, a specific phase shift in the individual’s thigh profile was discovered, reducing the gait phase estimation accuracy in the terminal swing phase. By compensating for this phase shift, the linearity of the phase variable was increased, and heel-strike detection was significantly improved. Furthermore, a piecewise phase variable is proposed to achieve better adaptation to different walking speeds. At different walking speeds, people have different toe-off timings. To account for this variable toe-off timing, we estimate toe-off at each gait cycle and adjust the slope of the phase variable based on the estimated toe-off timing. This allows for a more natural roll-over while walking. Additionally, while walking with the prosthesis, the enhanced and timely push-off was achieved. In the previous chapter, we investigated a learning-based gait phase estimation method to improve robustness and speed adaptability. Traditionally, in their model training, researchers provide a linearly interpolated label based on the heel-strike. This, however, cannot account for variable toe-off timing while training the given model. As a result, we propose a new piecewise linear label as the ground truth to improve the speed adaptability of gait phase estimation. As a result of the proposed method, we were able to obtain highly accurate gait phase prediction and heel-strike detection. As a result, it appears that compensating for phase shifts in the human thigh profile could improve user adaptability. We could also improve speed adaptability by reflecting variable toe-off timing at different walking speeds. | |
