Deep Learning-Based Compressed Sensing for Mobile Sensor Data

Overview

As the capabilities of smart sensing and mobile technologies continue to evolve, storing diverse sensor data on smartphones and cloud servers becomes increasingly challenging. This project develops deep learning-based compressed sensing (CS) frameworks to efficiently compress and reconstruct smartphone sensor data, overcoming the limitations of traditional CS methods.

Approach

• Mob-ISTA-1DNet (CIKM 2024): An innovative CS framework that integrates deep learning with the iterative shrinkage-thresholding algorithm (ISTA) to adaptively compress and reconstruct smartphone sensor data. Designed to manage the complexities of sensor data including variability, dynamic changes, and different sampling rates
• Convolutional CS (SenSys 2022): A compressed sensing framework using convolutional neural networks (CNN) for compressing and reconstructing acceleration data, demonstrating dramatic improvements in reconstruction performance with minimal processing time

Results

• Mob-ISTA-1DNet enhances reconstruction accuracy and significantly reduces processing time, consistently outperforming traditional methods across various sensor types (accelerometer, gyroscope, barometer, etc.)
• Data collected from 30 volunteers over one month using a custom mobile application, validating the framework under real-world conditions
• Comparative analysis demonstrates superior performance over traditional compressed sensing methods in both reconstruction quality and processing efficiency

Significance

By combining deep learning with compressed sensing theory, this project enables practical and efficient compression of heterogeneous smartphone sensor data. The proposed frameworks address real-world challenges of mobile sensing data storage and transmission, making large-scale, continuous sensor data collection more feasible.