Recognition of electronic devices by soft computing method

OData support
Molnár Károly
Department of Measurement and Information Systems

Today, the increase of computing capacity and the increasing complexity of problems

cause the steady spread of soft computing methods used in information systems. One

promising topic in the building automation field is the identification of electrical equipments

based on their power consumption parameters. Electrical devices identified this way can be

monitored,the given power consumptionhabits andpossiblemalfunction canbe discovered.

The thesis aims to achieve a working recognition system implemented in MATLAB

environment, with sufficient accuracy for identification of consumers. Future embedded

implementation should be taken account – in the design phase the reduced computing

capacity, and accuracy available in embedded systems, particularly.

Previous research results and application possibilities of the topic are reviewed, we touch

their suitability questions. The results of these papers served as a basis, we use some of

the methods used in these articles.

The difficulties of creating feature-vectors, particularly the problem of high-dimensional

data representation are described. We investigate two widely used solutions – they are the

principal component analysis (PCA) and linear discriminant analysis (LDA). We discuss

non-hierarchical clustering methods are discussed in those categories that are relevant to

resolv our problem - these are the k-means, kernel k-means and Gaussian Mixture Model


We overview classification techniques, focusing on their applicability to our problem –

especially the high computational requirements and recognition accuracy. The considered

classifiers are the neural network, the KNN method and some other distance-based classi-


The problems of detection efficiency, and also the suitable the analysis methods of these

techniques are presented. The presented techniques are the confusion matrix and ROC


The system is tested on real data samples in cross-validated manner. The implemented

identification algorythm works with sufficient precision, but we also present some recom-

mendations to increase the accuracy of the system.


Please sign in to download the files of this thesis.