Exo-Ada: A Boosting Model for Exoskeleton Angle Prediction

Petersen, Emil Thougaard ; Karlsson, Jonathan ; Thillemann, Palle

4. term

Software, Master

2020

2020-06-04

12

Eksoskeletter kan støtte både manuelt arbejde og rehabilitering. For at bevæge sig naturligt skal et eksoskelet kunne forudse, hvad brugeren har tænkt sig at gøre. En grundlæggende delopgave er at estimere albueleddets vinkel. I dette arbejde undersøger vi, hvordan man kan forudsige albuevinklen ud fra målinger fra et Force Myography (FMG) sensorarmbånd. Et FMG-armbånd registrerer trykændringer omkring underarmen forårsaget af muskelsammentrækninger. Traditionelt bliver en maskinlæringsmodel trænet på data fra én person og bruges kun til den samme person. Vi ønsker i stedet en løsning, der hurtigt kan tilpasses nye brugere og dermed kræver mindre træningsdata for hver ny person. Vi foreslår Exo-Ada, en ny boosting-tilgang designet til eksoskeletter, der gør det lettere at overføre en model til en ny bruger. Exo-Ada bygger på 2-Stage TrAdaBoost (en transfer learning-variant af boosting) og anvender en rectified Convolutional Neural Network (CNN), som gør brug af referencepunkter og dilaterede konvolutioner for bedre at fange mønstre i signalet. I tests på FMG-målinger fra flere personer overgår Exo-Ada flere baseline-metoder.

Exoskeletons can support both manual labor and rehabilitation. To move naturally, an exoskeleton must anticipate the user’s intended motion. A basic step in this process is estimating the elbow joint angle. This work addresses how to predict the elbow angle using measurements from a Force Myography (FMG) sensor armband. An FMG armband captures pressure changes around the forearm caused by muscle contractions. Traditionally, a machine learning model is trained on data from a single person and then used only for that same person. Our goal is to make this approach adapt quickly to new users, reducing the amount of training data required for each individual. We propose Exo-Ada, a new boosting approach tailored for exoskeletons that makes it easier to transfer a model to a new user. Exo-Ada builds on 2-Stage TrAdaBoost (a transfer learning variant of boosting) and uses a rectified Convolutional Neural Network (CNN) that incorporates reference points and dilated convolutions to better capture signal patterns. In tests on FMG measurements from multiple people, Exo-Ada outperforms several baseline methods.

[This abstract was generated with the help of AI]

Machine learning ; adaboost ; transfer learning ; regression ; cnn ; convolution ; reference points ; dilation ; exoskeleton ; joint angle ; session to session ; person to person ; mae ; exo-ada ; source ; target

Download
View record in AAU Student Projects