Mixed-signal distributed feature extraction for classification of wide-band acoustic signals on sensor networks.

摘要

This thesis proposes a distributed/collaborative scheme for the classification of wide-band acoustic events within the context wireless sensor networks (WSNs). The proposed method is characterized by a mixed-signal processing scheme for the extraction of feature vectors which reduces the processing, memory allocation, and power requirements of the sensor nodes. The mixed-signal scheme assumes that nodes have an analog frontend consisting of a microphone, a bandpass filter, a squaring element and an integrator. This chain of components produces an estimate of the energy over the frequency subband extracted by the filter. The WSN has the ability to form a distributed filter bank where a set of nodes self-organize to capture an acoustic signal (or event) across non-overlapping contiguous frequency subbands. The group of subbands resemble the frequency decomposition of a discrete wavelet packet transform (DWPT). The analog energy estimates are digitized (i.e., sampled) at a very low sampling rate (e.g., one sample per second), reducing the overall power, memory and processing requirements of the WSN. The energy measurements are relayed to a concentrator node where they are grouped to form a feature vector that serves as a signature for the acoustic event. The feature vectors can then be used with a statistical pattern classifier to identify the class to which the event belongs. We used a database of five acoustic classes: birds, explosions, cars, conversation and footsteps. The k-NN classifier, a back-propagation artificial neural network and a radial-basis function support vector machine (RBF-SVM) were evaluated against the dataset. An all-digital reference system was implemented using signals sampled at 44.1 KHz with a DWPT front end. Feature vectors were produced using subband energies and classified with the same pattern classifiers. Both systems were simulated in MATLAB using an eight-band filter bank. Simulations show that the proposed distributed mixed-signal solution performs as well as the all-digital system with 77.7% and 75.7% respectively for the RBF-SVM classifier.