CRII: III: Motif-aware Contrastive Learning of Universal Time Series Representation

The University Of Texas Rio Grande Valley, Edinburg TX

Investigators

Abstract

Time series data (i.e., a collection of observations for a single subject at different time intervals) is one of the most prevalent data types in a wide range of research domains. For example, times series data are common in meteorology, medicine, and physics. However, the process of labeling time series data often demands a substantial amount of time. Typically, only domain experts have the capability to effectively label time series data. Consequently, data labels are often scarce in most real-world time series applications. To overcome this challenge, an approach called Contrastive Self-Supervised Learning (CSSL) has been developed in the research community. However, the effectiveness of CSSL framework is highly related to the definition of “semantic similarity” among time series data that represents the characteristics of the underlying system. In CSSL framework, the errors (whether they were wrongly considered similar or dissimilar samples) will likely generate meaningless or even misleading models that impact on the final artificial intelligence (AI)-driven models that utilize time series data. This project aims to address these challenges by conducting a systematic analysis of the CSSL framework for time series data and developing innovative algorithms to alleviate the identified limitations. Additionally, this project will specifically focus on smart manufacturing system related application where time series sensors are widely recognized for their energy efficiency but lack data labeling. Furthermore, the project has a detailed plan to integrate research outcomes into existing courses and develop new courses at the University of Texas Rio Grande Valley. Lastly, the project will promote the participation of Hispanic students in STEM research and participate in outreach events in the local community. The technical aims of the project are divided into two tasks corresponding to the two major steps in the CSSL framework: the view augmentation step, which aims at augmenting semantically correlated samples and the negative sample sampling step which aims at sampling semantically unrelated samples. The project aims to integrate the concept of time series motifs, a key primitive used to unveil underlying natural mechanisms in a wide variety of time series, into CSSL frameworks. Specifically, the project consists of two thrusts: 1) To address the challenge of difficulty to preserve semantic meaning in view augmentation operator, the project aims to design a motif-aware augmentation operator to use detected motifs to regularize the semantics of augmented data; and 2) To address the challenge of difficulty to sample high-quality negative samples, the project aims to design a motif-aware negative sampling to avoid low-quality negative samples induced by noises and redundancy existed in the time series data. The investigated theories and methodologies will deepen the understanding of the intrinsic working mechanism of the self-supervised time series representation and contribute to the research in a wide range of domain applications. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.