Abstract
Since time series signals arise always and everywhere in real life, time series signal processing and analysis have been attracted a lot of attention from researchers. Among various applications, localization and classification for time series signals are two important research fields. On the other hand, the rapid developing machine learning provides us a powerful tool to address the challenges in localization and classification. In this thesis, three research topics on machine learning based localization and classification for time series signals are demonstrated, including sound source localization, small unmanned aerial vehicle classification and localization, and acoustic scene classification. In the first study, a novel indoor sound source localization approach with probabilistic neural network is proposed. By proposing a generalized cross correlation based classification algorithm and a weighted location decision making method, sound sources can be localized in challenging environments with high reverberation and low signal-to-noise-ratio. In the second topic, a novel micro-Doppler signature based small unmanned aerial vehicle surveillance approach with long short-term memory neural network in low-frequency band is proposed. The surveillance consists of three stages, including detection, classification, and localization. In the third work, a novel acoustic scene classification approach based on data augmentation and ensemble learning with convolutional neural networks is proposed. The diversity issue and the unbalanced amount issue of the data from different acoustic scenes are addressed.