Diz 78 en
Pattern Discovery, Learning and Detection in Time Series
Author: Martin Ron
Machine learning tasks typically require large amount of data for training. This dissertation focuses on time series analysis, which is a frequent type of data collected in industry. The desired application is modeling and analysis of machines behavior. The behavior models require well-structured training data sets which are usually prepared manually. That is an expensive and exhausting task prone to errors. This complication limits the growth of applications of behavior models in industry, which motivated us to investigate the entire process of stochastic modeling of machines behavior to automate the deployment process as much as possible.
Our research was initiated by a task of modeling industrial robotic-manipulator behavior based on its power consumption, where we needed to segment a power-consumption time series by particular robotic operations. This analysis served as a support for research of optimal scheduling of robotic operations to reduce power consumption of robots. We expanded our target domain from the robotic power consumption to a general repetitive behavior observed in time series. Our findings are verified on the robotic use cases, but we keep our methods general enough to be applicable on wide range of industrial tasks, which was verified by successfully applying the methods on batch-heating ovens.
To reduce the manual interventions in the modeling process, we formulate the segmentation task as the motif discovery task. Motif discovery searches for repetitive behavior in time series and it has quadratic time complexity unless an approximation methods are used. One of our main contributions is a novel pre-filtering method that significantly reduces the time complexity of motif discovery to be linear for most real-life data sets, and ensures the high quality of the motifs. On the other end of the process, we propose a method of parameters-continuity preference which significantly reduces the size of the training data sets required for time-varying hidden Gauss-Markov models. These models provide great classification and inference performance, but there is rarely enough data to train them. Our contribution overcomes this issue by learning more information from less data. The entire work results in the Motif Discovery and Detection Framework which solves the whole task of machine behavior modeling.
The dissertation is structured according to the research progress. Each chapter contains the related work concerning the topic of the chapter, separate definition of the goals and discusses the findings from simulations and experiments. This elevates the modularity of our research, and the modularity of the Motif Discovery and Detection Framework, which promotes its adoption by practitioners.