Development of Gait Classification System Using Six-Axis Inertial Sensors by CNN-LSTM

Restoration of walking function is the most important goal of rehabilitation. However, the evaluation of post-stroke rehabilitation lacks quantitative elements. In this study, we used an IMU to acquire the angles of three axes (rotation of x-axis, rotation of y-axis, and rotation of z-axis) of the k...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2024 Joint 13th International Conference on Soft Computing and Intelligent Systems and 25th International Symposium on Advanced Intelligent Systems (SCIS&ISIS) S. 1 - 5
Hauptverfasser: Ohno, Shotaro, Nagamune, Kouki
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 09.11.2024
Schlagworte:
Data models
Deep learning
Gait
IMU
Inertial sensors
Knee
Legged locomotion
Motion measurement
Pattern recognition
Predictive models
stroke
Time series analysis
Training
Training data
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Restoration of walking function is the most important goal of rehabilitation. However, the evaluation of post-stroke rehabilitation lacks quantitative elements. In this study, we used an IMU to acquire the angles of three axes (rotation of x-axis, rotation of y-axis, and rotation of z-axis) of the knee during walking motion as gait parameters. Using this data, machine learning analysis is performed. In this article, we discuss the gait parameters, the angular data of the three axes of the knee, and examine which of the three angular parameters increases or de-creases the generalization ability of the model. As a result, we identify the gait parameters to be used as a dataset. Time series data of the knee angle of one step in the gait measurement was used as the dataset. CNN - LSTM algorithm was used to construct the model. Five subjects were measured with the knee joint motion measurement device attached to both knees. A total of four types of gait movements, including normal gait and three types of gait movements that reproduce hemiplegia, were meas-ured for analysis. Out of one hundred thirty-six steps of data from five subjects, four subjects were used as training data and one unknown subject's data, which was not used during training, was applied to the test data. The results of the prediction of the estimated data show that rotation of z-axis was a parameter that reduced the generalization ability of the model, while rotation of y-axis was a parameter that increased the generalization ability of the model. Prediction results for the rotation of x-axis data alone showed poor accuracy, but when combined with the rotation of y-axis data, the generalization ability was improved.
DOI:10.1109/SCISISIS61014.2024.10760026