Model generation from sensor data : with frequent pattern mining

Abstract

This thesis examines the possibility of model generation from data sets. Automatic model generation from data sets is a subtopic of Model Intelligence. Model Intelligence aims to combine the benefits of Model Driven Engineering (MDE) with the benefits of machinelearning (ML) techniques. During a workshop about Model Intelligence a call for paper,namely "Model inferencers and automatic model generators from datas ets" [ MDE ]. In this work we propose a method to generate models with the help of machine learning from data sets and make a case study on a real-life data set coming from a sensor system of the (Plus-)Plus-Energy building of the Vienna University of Technology. Different definitions of"model" are considered to answer the question whether or not it is possible to generate models automatically from sensor data with the help of machine learning and data mining techniques.A frequent pattern recognition based method is introduced, which aims to generate a frequent pattern based model for sensors. To set boundaries for model generation, we discuss the basic concepts and introduce key techniques of metamodeling. We choose a suitable technique to define our metamodel, which then will serve as the basis of model generation. During defining the model and exploring different modeling techniques we mainly discuss methods and related concepts conforming to the Meta-Object Facility (MOF).After defining the model, we apply segmentation on the sensor-derived data set, and label different segments according to the Symbolic Aggregate approXimation (SAX) We apply frequent pattern mining on the segmented and labeled data sets. Based on the results of the patternmining, we carry out clustering to help us decide, whether the frequent patterns found in the sensor data are suitable to differentiate between different types of sensors, thus can serve as attributes in model definitions for the different types of sensors.Our results show that depending on what definition of "model" is used, it is possible to generate models either on the M1 or the M0 level of the M4 meta-modeling framework. Furthermore, aproposed model with Double Deep Instantiation (DDI) is feasible to model type definitionsfor the sensors. In conclusion, we show that generating type definitions for the sensors based on their frequent patterns is successful in most cases. We have found some instances with interesting caveats that call for further investigation.ix