A Method for Sensor-Based Activity Recognition in Missing Data Scenario.

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Sensor-based human activity recognition (HAR) is crucial for healthcare and smart-home applications.
• Existing HAR methods often fail with real-world data due to missing sensor information.
• This limitation hinders practical deployment in critical environments.

Purpose of the Study:

• To develop a robust sensor-based HAR method capable of handling randomly missing data without data imputation.
• To enhance the reliability and accuracy of HAR systems in the presence of data loss.
• To investigate the effectiveness of training machine learning models with simulated missing data.

Main Methods:

• Proposed a novel approach to train HAR models by explicitly introducing random missing data percentages into the training set.
• Utilized time-series statistical features for enhanced feature extraction.
• Evaluated Support Vector Machine (SVM) and Random Forest classifiers on synthetic and benchmark datasets (HASC, single chest-mounted accelerometer).

Main Results:

• Achieved a significant improvement in recognition accuracy from 80.8% to 97.5% on a synthetic dataset.
• Demonstrated improved performance across various missing data percentages, window sizes, and sliding widths on benchmark datasets.
• The proposed method effectively handles random data loss, enhancing HAR system robustness.

Conclusions:

• Training machine learning models with randomly missing data is a viable strategy to improve HAR performance in real-world conditions.
• The developed method offers a practical solution for deploying reliable HAR systems in healthcare and other domains.
• Further research can explore more complex missing data patterns and advanced feature engineering techniques.