Two main approaches of pattern recognition was applied to detect the operation states of electric appliances by examining the time series of their electricity consumption data. In similarity matching a distance function is utilizied to measure the dissimilarity of two time series, and a threshold to decide whether the two series match each other. In dynamic systems the same event may generate a different temporal alignment in different cases. Elastic distance measures may be applied to address this problem. In this work, dynamic time warping (DTW) was used as elastic measure.
The other approach was supervised machine learning with the utilization of classification algorithms. Two basic model classes were used: k-nearest neighbours classifiers (k-NN) and decision trees. The neares neighbours classifier is also distance based, assigns a class label to an unknown object by searching its k nearest objects and uses the most common label among their labels to predict a class membership. The most widely used distsance measure in k-NN is Euclidean distance, however, in this work it was combined also with DTW.
Decision trees form another non-parametric family of classifiers. It operates by splitting the parameter-space along a threshold applied to one of the parameters, thus generating a decision surface. The model is stored as a graph where each split corresponds to an internal node or the root node, and the leafs represent class memberships.
To be able to use these methods effectively the raw measurement data needs to be pre-processed or even the parameter-space has to be transformed into some more representative feature-space.
To support this process, a software framework was developed including several methods concerned with the individual phases. The framework was implemented in the Python programming language with the motivation of easy extendibility.
The pattern recognition task was resolved providing acceptable results. The methods also provided accountable performance regarding further real-time applicability.