Deep learning for environmental state prediction and sensor fusion for intelligent wearable robots
Lead Research Organisation:
University of Bath
Department Name: Mechanical Engineering
Abstract
Current prosthetics are only capable of performing a limited range of activities. Multiple prostheses are therefore required to perform different activities, such as walking and running. Producing a device that can adapt to the users current requirement has the potential to improve the users quality of life. The research looks into improvements in prosthetics that can be made through intelligent adaptive controllers.
The work will aim to classify users activity, intent and efficiency informing an adaptive controllers response. Investigation into the roll machine learning can perform in setting controller gains based on an individuals requirements; with the aim of improving controller efficiency over time, and how dynamic changes to controller setup could be safely deployed. The system will be aim will be physically tested using small form factor electronics and embedded software to implement as a real time wearable device.
The work will aim to classify users activity, intent and efficiency informing an adaptive controllers response. Investigation into the roll machine learning can perform in setting controller gains based on an individuals requirements; with the aim of improving controller efficiency over time, and how dynamic changes to controller setup could be safely deployed. The system will be aim will be physically tested using small form factor electronics and embedded software to implement as a real time wearable device.
Organisations
People |
ORCID iD |
Pejman Iravani (Primary Supervisor) | |
Frederick SHERRATT (Student) |
Publications
Sherratt F
(2021)
Understanding LSTM Network Behaviour of IMU-Based Locomotion Mode Recognition for Applications in Prostheses and Wearables.
in Sensors (Basel, Switzerland)
Description | Analysis of a Long Short Term Neural Network trained for Locomotion Mode Recognition has identified that information about activity type is greatest in the early stance phase of gait. This is also the point at which most gait variations between individuals occurs in early stance. This may explain the poor generalization characteristics when novel subjects are presented to the model. This outcome demonstrates the need for the personalization of ML models to achieve acceptable accuracy. |
Exploitation Route | This provides motivation for investigation personalisation techniques for ML-based LMR models |
Sectors | Healthcare |
Title | Bath Natural Environment HAR Data Set |
Description | The data set contains recording from 5 9-axis IMU (MARG) sensors. Attached to ankles, hips and chest. The sensors were sampled at 100Hz. The experiment involved 22 subjects walking around natural environments wearing the five sensors. Though a BLE connections the sensors streamed data to an app on an android phone. The subjects labeled data in real time using buttons in the app. The data was collected in an unsupervised manner and shared with the researchers anonymously. The following activities were recorded; Walking, Ramp Ascent, Ramp Descent, Stair Ascent, Stair Descent, Stopped |
Type Of Material | Database/Collection of data |
Year Produced | 2020 |
Provided To Others? | Yes |
Impact | The publication "Understanding LSTM Network Behaviour of IMU-Based Locomotion Mode Recognition for Applications in Prostheses and Wearables" - doi 10.3390/s21041264 - was produced using this data set |
URL | https://zenodo.org/record/4390499 |