Time-Series Feature Extraction in Embedded Sensor Processing System

Sammanfattning: Embedded sensor-based systems mounted with tens or hundreds of sensors can collect enormous time-series data, while the data analysis on those time-series is commonly conducted on the remote server-side. With the development of microprocessors, there have been increasing demands to move the analysis process to the local embedded systems. In this thesis, the objective is to inves- tigate the possibility of the time-series feature extraction methods suitable for the embedded sensor processing systems.As the research problem raised from the objective, we have explored the traditional statistic methods and machine learning approaches on time-series data mining. To narrow down the research scope, the thesis focuses on the similarity search methods together with the clustering algorithms from the time-series feature extraction perspective. In the project, we have chosen and implemented two clustering algorithms, the K-means and the Self-Organizing Map (SOM), combined with two similarity search methods, the Euclidean dis- tance and the Dynamic Time Warping (DTW). The evaluation setup uses four public datasets with labels, and the Rand index (RI) to score the accuracy. We have tested the performance on accuracy and time consumption of the four combinations of the chosen algorithms on the embedded platform.The results show that the SOM with DTW can generally achieve better accuracy with a relatively longer inferring time than the other evaluated meth- ods. Quantitatively, the SOM with DTW can do clustering on one time-series sample of 300 data points for twelve classes in 40 ms using the ESP32 embed- ded microprocessor, with a 4 percentage of accuracy advantage than the fastest K-means with Euclidean distance in RI score. We can conclude that the SOM with DTW algorithm can be used to handle the time-series clustering tasks on the embedded sensor processing systems if the timing requirement is not so stringent.