Sound Detection and Classification Algorithms based on Harmonic Characteristics

Abstract

Detection and classification of unmanned aerial vehicles (UAV) have been demanded to reduce many problems such as ground or midair collisions, privacy and security. It is difficult to detect UAV for radar since it is relatively small and flying at low altitude. However, sounds of engine and propeller can be measured easily by microphone. In this thesis, UAV sound detection and classification algorithms are proposed, which are based on harmonic spectral peaks of UAV. When two harmonic spectral peaks are mapped into a vector, it has a unique angle according to its harmonic orders. By pre-defining reference vectors based on known harmonic orders of target UAV and computing similarity of input vectors with the reference vectors, the UAV sound can be identified from non-harmonic sound. Inner product based algorithm and virtual array based algorithm are proposed to compute the similarities. Both algorithms show stable discrimination performance than Mel-frequency cepstral coefficient (MFCC). The under area of ROC curve was up to 0.975 by using the virtual array based algorithm. Harmonic order detection algorithm for unknown signal is also studied. Probabilities of input vectors to be certain reference vectors are computed for all possible cases and extreme learning machine (ELM) are used as a classifier. By concatenating the order detection algorithm and another ELM classifier, multiclass classification can be possible. It shows better performance than MFCC for four different classes of measurement sound data. Proposed algorithms can be usefully applied for detection and classification of other harmonic sounds as well as UAVs.