Periodic Time Series Data Analysis by Novel Machine Learning Methodologies

Abstract

Period length extraction is considered a challenge in many research fields. To solve this problem, different methods have been proposed across many applications. For instance, supply chain management is an area that can greatly benefit from precise periodic information. In addition, periodic information on physiological data can provide insights into individuals’ health conditions, which is the motivation of this thesis. The difficulty of period length extraction involves the varying noise levels among working environments. A system that performs well in one environment may not be accurate in another. In this work, we explore two machine learning approaches, each of which attempts to solve the problem at a different noise level. The first algorithm, the period classification algorithm (PCA), utilizes historical labeled data as training material and classifies new instances. The PCA demonstrates robustness to both generated and natural noise. However, the training of the PCA is not economical if the data do not contain much noise. The second algorithm, the period detection algorithm (PDA), is used when the noise level is not very high. It does not require historical data, but rather detects the period length directly from the data stream. The PDA cannot tolerate as much noise as the PCA; however, it is more efficient and simpler to deploy. By investigating both algorithms on artificial and real-world datasets, we determined that they have advantages under different circumstances. In particular, the PDA outperforms the PCA when the system is noise-free, while it fails on real-world datasets, which usually contain a large amount of noise. In contrast, given that the training material is representative of test datasets, the PCA demonstrates high performance on both artificial and real-world datasets.