Physics-Enhanced PCA for Data Compression in Edge Devices

Qianyi Chen, Jiannong Cao, Yong Xia

Abstract: In smart cities, tremendous data are generated with edge devices continuously for scientific applications, such as structural health monitoring (SHM), leading to a high bandwidth burden to edge devices. Data compression is a typical technique for reducing data size and improving transmission efficiency. However, their operations for distinguishing redundant data might be unapplicable for a domain-specific scientific analysis and distort physical quantities of raw data (i.e., mode shapes of vibration data). Besides, they are too compute-intensive to be executed in resource-constraint edge devices. In this paper, we leverage physical knowledge to enhance a lightweight data compression method for edge devices in maintaining physical quantities. In particular, we propose physics-enhanced PCA for compressing data in a dynamic system — compressing the vibration data of a structure in the context of SHM. Physical knowledge is identified from a structure and guides the compression process to preserve the mode shape, an significant physical quantity for structures. We have formally analyzed the effectiveness of our approach and performed experiments to show that physical knowledge is essential for preserving mode shapes. Concretely, experiments in numerical and real-world structures show that physics-enhanced PCA can improve the accuracy by up to 56% compared with alternative baseline.