Publikation

Framework for Analyzing Sounds of Home Environment for Device Recognition

Svilen Dimitrov

Mastersthesis DFKI / University of Saarland University of Saarland Saarbrücken 5/2014.

Abstrakt

Home environments are one of the subjects of study by Ambient Intelligent Systems for various purposes, including developments of elderly assistance systems and energy consumption optimization. Sensing the environment, via different sensors, is the first and crucial component of every Ambient Intelligent System. In this thesis we design and develop the Sound-based Device Recognition Framework to investigate the application of environmental sounds usage for touch-free audio-based device recognition in a home environment. For this purpose, we study the characteristics of the sounds dispersed by devices in a home environment. We use the acquired knowledge to implement different Sound Processing techniques for the extraction of a flexible set of features, which can be determined both manually and automatically. For the classification of gathered device acoustic fingerprints we use multiple optimized straightforward techniques of Supervised Learning as well as integrated established ones. Furthermore, we use a feedback from the user for creating an incremental learning system. After establishing a recognition basis for the recognition of fixed length sound buffers on demand, we implement a live recognition mode for real-time environment monitoring, providing runtime setup adjustments. These include changing the selected features, switching between Machine Learning algorithms, and recognition time interval choice, without interruption for modifications of the trained data. We then extend our work with the recognition of untrained simultaneously working known devices, utilizing Semi-supervised Learning. Finally, we create an automatic test utility to evaluate different aspects of the developed framework, including recognition rate performance for the different combinations of features and Machine Learning algorithms, as well as to study the reliability of the automatic mixing of trained data. Our evaluation shows satisfactory results in all tested aspects. Therefore we consider the development of our Sound-based Device Recognition Framework as complete and providing a solid base for further research.

Master-Thesis-Svilen-Dimitrov-2014.pdf (pdf, 8 MB)

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence