Theses and Dissertations
Date of Award
5-1-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
First Advisor
Yifeng Gao
Second Advisor
Li Zhang
Third Advisor
Haoteng Tang
Abstract
Time series data is perhaps one of the most broad data types that exist and is studied by diverse research fields. Recently, Self-Supervised Learning (SSL) training frameworks, the training frameworks to pre-train deep learning models without human annotations, have been proposed. Because human annotation for time series is typically associated with being costly, there is a growing interest in developing effective SSL for time series data. In SSL, the pre-trained model will often produce a time series embedding series summarized from the original time series to ensure temporal information is preserved. Although such representation can effectively capture the semantic information, the potential correlation between each time series sample can induce redundancy in the final representation.
To address this issue, this work first proposes a compression algorithm named SimCompress to compress the redundant encoder embedding series by removing similar consecutive embeddings. The proposed compression algorithm is easy to implement, and can be adapted to any SSL framework that produces an embedding series. Experiments show that the proposed process can possibly significantly decrease the amount of space required for data storage without reducing the performance accuracy on downstream tasks.
The impact of embedding redundancy is further examined in swarm robotic systems. Swarm robotics leverages groups of autonomous robots to perform complex tasks collaboratively. Recently, there has been growing interest in the Self-Supervised Learning framework (SSL) for social robots, yet very little research has been done on designing a SSL framework for foraging swarm robot systems. To address this, I propose a novel efficient self-supervised learning framework for swarm robotics. Specifically, I propose (1) a shared weight multi-instance based encoder–decoder structure for model pre-training, and (2) applying the SimCompress compression algorithm to reduce the space cost in the inference stage. Experiments show this system can match the performance of standard SSL frameworks while being more efficient. Finally, this work extends embedding compression beyond local redundancy removal. I propose an embedding series compression algorithm named CompressNet to compress the redundant encoder embedding series by summarizing similar consecutive embeddings into an individual embedding. The proposed compression algorithm is post-hoc and can be adapted to any SSL framework that produces an embedding series. Experiments show that the proposed process can possibly significantly decrease the amount of space required for data storage without loss of performance on downstream tasks.
Recommended Citation
Berry, B. (2026). Efficient Compression Framework for Time Series Self-Supervised Learning [Master's thesis, The University of Texas Rio Grande Valley]. ScholarWorks @ UTRGV. https://scholarworks.utrgv.edu/etd/1869
Comments
Copyright 2026 Brooklyn Berry. All Rights Reserved. https://proquest.com/docview/3371103985