Please use this identifier to cite or link to this item: http://hdl.handle.net/10266/6889
Title: Muscle Force Prediction for Human Body Activities using Computational Intelligence Approaches
Authors: Bansal, Hunish Kumar
Supervisor: Rana, Prashant Singh
Kumar, Neeraj
Keywords: sEMG signal processing;anomaly detection;TimeGAN;LSTM;ARIMA;Prophet;iforest
Issue Date: 11-Oct-2024
Abstract: The analysis of surface electromyography (sEMG) signals plays a crucial role in understanding and quantifying muscle activation patterns, which has numerous applications in biomechanics, rehabilitation, and sports science. However, accurate modeling and prediction of sEMG time series data pose significant challenges due to the inherent complexities of physiological systems and the presence of inconsistencies and anomalies in the recorded data. This study presents an integrated machine learning framework that combines statistical anomaly detection techniques with generative adversarial networks (GANs) to enhance the predictive capabilities of sEMG-based muscle force forecasting models. The first part of the study focused on detecting knee abnormalities using sEMG data from a UCI Machine Learning Repository dataset. It included 22 adults performing walking, leg bending, and extending tasks. Data was collected from four muscles using a Biometrics Ltd. DataLog MWX8 and a goniometer. Preprocessing involved denoising, filtering, and normalization. The study noted a class imbalance due to longer task completion times for participants with knee abnormalities. To address this, three anomaly detection techniques (Isolation Forest, k-Nearest Neighbors, and Local Outlier Factor) were employed. Various machine learning classifiers, including LightGBM, XGBoost, Random Forest, Extra Trees, and Decision Tree, were trained and evaluated using multiple performance metrics. K-fold cross-validation assessed model robustness. The ensemble technique combining anomaly detection and machine learning classifiers showed significant improvements over previous studies. The LightGBM classifier, trained on data processed with the Isolation Forest technique, achieved 98.5% accuracy, surpassing previous best accuracies by 6-7%. Building upon the success of the anomaly detection and classification approach, the second part of the study focused on predictive modeling of muscle force using machine learning approaches like artificial neural networks (ANNs). To address challenges with time series data, the study proposed an integrated framework combining statistical outlier detection methods with generative adversarial networks (GANs). The framework integrates various anomaly detection techniques (Isolation Forest, KNearest Neighbor, One-Class Support Vector Machine, Histogram-Based Outlier Score, and Local Outlier Factor) to improve input data quality. TimeGAN architecture was used to synthesize realistic time-series data, addressing data scarcity and enabling more personalized predictive models. The framework was evaluated using a dataset of body movements and forces from 57 healthy individuals. Performance of prediction models like Long Short-Term Memory (LSTM) networks, Auto-Regressive Integrated Moving Average (ARIMA), and Prophet forecasting algorithm was compared with and without anomaly detection techniques. Results showed that integrating anomaly detection techniques significantly enhanced muscle force prediction performance. The Isolation Forest method combined with LSTM achieved a Pearson’s Correlation Coefficient of 0.95 and a coefficient of determination between 0.9 and 0.93, comparable to state-of-the-art approaches. This demonstrated the benefits of integrating statistical and generative AI techniques for time series analytics in muscle activation pattern modeling and prediction. The proposed integrated machine learning framework has several potential applications in the field of biomechanics, rehabilitation, and sports science. Accurate prediction of muscle activation patterns can facilitate the development of advanced prosthetic limbs and exoskeletons, enabling more natural and intuitive control systems. Additionally, the framework can support athletic training by providing real-time feedback and analysis of muscle activity, helping athletes optimize their performance and prevent injuries. Furthermore, the study contributes to the investigation of muscle physiology by enabling researchers to model and simulate various muscle activation scenarios, leading to a deeper understanding of the underlying mechanisms and potential interventions for musculoskeletal disorders.
URI: http://hdl.handle.net/10266/6889
Appears in Collections:Doctoral Theses@CSED

Files in This Item:
File Description SizeFormat 
PhD thesis_Hunish Bansal_CSED.pdfPhD Thesis2.28 MBAdobe PDFThumbnail
View/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.