Intelligent Soft-sensor System to Add Haptic Perception to Underactuated Hand Prostheses
Nombre: LAURA VANESA DE ARCO BARRAZA
Fecha de publicación: 10/03/2023
Supervisor:
Nombre |
Papel |
|---|---|
| CAMILO ARTURO RODRIGUEZ DIAZ | Advisor * |
| MARIA JOSE PONTES | Co-advisor * |
Junta de examinadores:
| Nombre |
Papel |
|---|---|
| ARNALDO GOMES LEAL JÚNIOR | Internal Examiner * |
| CAMILO ARTURO RODRIGUEZ DIAZ | Advisor * |
| CARLOS ANDRES CIFUENTES GARCIA | External Examiner * |
| MARCELO MARTINS WERNECK | External Examiner * |
| MARIA JOSE PONTES | Co advisor * |
Sumario: Haptic exploration is the capability of the human hand to recognize objects and manipulate them without visual aid. This is achieved through kinesthetic and tactile perception, with kinesthetic perception sensing the body's movements and the forces applied to it. In recent years, there has been increasing research and development aimed at incorporating this ability into robotic hands. This has been made possible through sensors and machine learning algorithms, which enable robotic hands to interact with objects without human supervision. Since the research areas on robotic hands are looking for devices as similar as possible to the member, the work with soft robotics and the development of sensors that agree with that characteristics have been increased. The recent advances in the development of sensors with optical fiber have increased due to its well-known capabilities, such as low-cost, small size, low weight, robustness, biocompatibility, high sensitivity, and precision. This master dissertation aimed to develop an intelligent soft-sensor system to add haptic perception to underactuated hand prostheses. This sensor methodology monitors the physical interaction during grasping activities to detect the object type grasped by a data-driven approach. The haptic sensor system developed here was implemented in the upper-limb prosthesis PrHand, based on soft robotics actuation. The soft-sensor system resembles the kinesthetic perception of the human hand by implementing two sensing modalities, finger joint angles and fingertip contact force measurements implemented in the prosthetic fingers. The sensors are based on intensity variation with polymer optical fiber. For the angle sensor development, three fabrications were tested by axial rotating the sensors in four positions, and the fabrication way with the most similar response in the four rotations was chosen. The sensor used is the jacket remotion with cladding and core axial polish sensor. The sensors were located in the distal interphalangeal (DIP) joint of the prosthetic finger. The sensor characterization was made by performing six cycles of opening and closing per finger. The sensors presented a second-order polynomial response with R² higher than 92%. The contact force sensors were located at the fingertips to track the force made over the objects. Before anchoring the five sensors, each one was evaluated by making five cycles of compressing and decompressing. Almost all sensors presented a polynomial response with R² higher than 94 %; the fabrication process highly influenced the sensor behavior since in the drying process of the sensor, the fiber could have movements. For the Machine learning implementation, 24 objects of the Anthropomorphic Hand Assessment Protocol (AHAP) related to eight grasp types are used. Six machine learning algorithms are tested, four are supervised (Linear Regression, k-Nearest Neighbor, Supporter Vector Machine, and Decision tree), and two are unsupervised (K-Means Clustering and Hierarchical Clustering). To validate the algorithms was used the k-fold test with a k = 10, and the accuracy results for k-Nearest Neighbor (k-NN) was 98.5 ± 0.01 % and Decision Tree (DT) was 93.3 ± 0.2 %. The other four algorithms had a result lower than 30 %. One of the principal reasons related to the low accuracy of the unsupervised algorithms is the lack of clustering formation. It was concluded that just the k-NN and DT algorithms allow classified the grips types. One of the limitations of this study is to evaluate if the PMDS is the best option for this sensor because, after a while, the silicone starts to break, so the lifetime of the sensor is not as high as expected. In addition, for the characterization of the angle sensor as it was done on the fingers, it was difficult to ensure that for all trials, the fingers closed, in the same way always, causing some errors.
