Latent Low-Dimensional Predictor Analytics for Engineering Applications

Description

Modern engineering and scientific systems often utilize numerous sensors to gather high-dimensional time series data for monitoring and operations. The success of DeepSeek for large language models highlights the effectiveness of low dimensional learning, particularly when computational resources and data volume are limited. This talk introduces a latent low dimensional dynamic predictor framework that concurrently achieves dimension reduction and optimal dynamic prediction. The dynamic latent variables, termed principal predictors, form low dimensional parsimonious predictor models for high-dimensional time series data. The solution process involves iterations to extract both dynamic and static subspaces. A maximum likelihood framework is employed to develop an iterative solution. The connection between principal predictors and DeepSeek low-dimensional approximation is explored. Examples from engineering and industrial manufacturing processes will be used to demonstrate the advantages of the proposed framework. This low-dimensional dynamic modeling approach has potential applications in prediction, control, and anomaly diagnosis.

Period	19 Aug 2025
Event title	2025 IEEE 21st International Conference on Automation Science and Engineering
Event type	Conference
Location	Los Angeles, United States, CaliforniaShow on map
Degree of Recognition	International