Time-series signals are central to understand and identify the state of a dynamical system.
They are ubiquitous in many areas related to geosciences, climate, and structural health
monitoring. As a result, the theory and techniques for analyzing and modeling time-series
have vast applications in many different scientific disciplines. One of the key challenges
that the time-series data analysts face is that of information/data overload. Furthermore,
the sheer volume of the time-series data generated at the sensor node makes it difficult to
transport the data to centralized databases. These aspects pose an obstacle for data analysts
in detecting changes in the system response as early as possible. Instead, a workflow for
an efficient and automatic reduction of collected data at sensor nodes can enable timely
analyses and decrease event detection latency. Such a workflow can be useful for many
real-time monitoring and sensing applications. An attractive way to construct a
computationally efficient workflow for automated analysis of time-series data is through
machine learning. In this paper, we present a machine learning framework to construct
models to efficiently reduce the time-series data by means of feature extraction and feature
selection. In the first step of the framework, we apply a feature extraction and feature
filtering algorithm called “Feature Extraction based on Scalable Hypothesis (FRESH)” for
a given time-series data to extract comprehensive time-series signal features and then filter
the resulting features. In second step, we quantify the significance of each filtered feature
for predicting a set of labels/targets. Third, we construct a machine learning classifier,
which takes in important filtered features to classify the time-series signals. The proposed
framework is tested and validated against ultrasonic sensing datasets obtained from
multiphase flow loop experiments.
|