A Machine Learning Framework for Balancing Training Sets of Sensor Sequential Data Streams.

Area of Science:

• Sensor data analysis
• Machine learning for time-series data
• Data augmentation techniques

Background:

• Smart technologies increasingly rely on sensor data and machine learning, making data imbalance a critical issue.
• Class-imbalanced datasets lead to poor performance and failure in smart technology applications.
• Sequential (time-series) sensor data presents unique challenges for imbalance resolution.

Purpose of the Study:

• To address the challenge of class-imbalanced sensor data in smart technologies.
• To develop and evaluate a data augmentation framework for balancing training datasets.
• To improve the reliability and accuracy of machine learning models trained on sensor data.

Main Methods:

• Development of an Unrolled Generative Adversarial Network (Unrolled GAN)-powered framework.
• Application of the framework to balance smartphone accelerometer and gyroscope sensor data for road surface monitoring.
• Validation using additional sensor data from an open data collection.

Main Results:

• Successful balancing of imbalanced sensor training data, particularly for smartphone accelerometer and gyroscope data.
• Demonstrated improvement in classification performance for heavily imbalanced datasets, with an F1 score increase from 0.69 to 0.72 (p<0.01) in a case study.
• Negligible performance improvement observed for slightly imbalanced or inadequate training sets.

Conclusions:

• The Unrolled GAN framework is effective for data augmentation in heavily imbalanced sensor datasets.
• The study highlights limitations related to data quality and computational efficiency, indicating areas for future research.
• Future work will focus on integrating data quality assessment and enhancing computational efficiency within the framework.