States changes representation for time series
Full Text
Share
ENG- The exponential growth in wearable device usage has generated vast amounts of time series data, presenting opportunities for various applications like activity recognition and health monitoring. However, analyzing these datasets poses challenges due to their complexity and length. To address this, this doctoral thesis proposes the State Changes Representation for Time Series (SCRTS), a method aimed at extracting relevant information related to the dynamics of the time series while significantly reducing the dimensionality. Moreover, SCRTS is length-independent, enabling the application of this algorithm to frames (consecutive values of the variables in the time series related to a given class) of varying lengths while producing feature vectors of the same size. This aspect is crucial for classifications, ensuring uniformity in feature representations across different time series lengths.
The SCRTS algorithm is presented in two variants: the one-dimensional (1D-SCRTS) and the multidimensional (mD-SCRTS) approaches. In the 1D-SCRTS, each frame is represented by a sequence of states derived from vector magnitudes, which summarize the information of the interrelated variables at each time point of time series samples. In contrast, the mD-SCRTS considers individual variable values before discretization, allowing it to capture information related to all variable values independently.
The effectiveness of the SCRTS is demonstrated through activity classification experiments using three accelerometer datasets. Both the 1D-SCRTS and the mD-SCRTS exhibit outstanding dimensionality reduction capabilities while achieving considerable classification performance
L'accés als continguts d'aquesta tesi queda condicionat a l'acceptació de les condicions d'ús establertes per la següent llicència Creative Commons: http://creativecommons.org/licenses/by/4.0/