Carvalho C., Martins C., Serralheiro A.

This paper describes a practical application using vector quantization (VQ) for detection of a moving train by means of the rail vibrations. The goal of this system is to remotely detect moving trains, in order to avoid labour accidents unfortunately common among railway workers. This system comprises both hardware and software subsystems. The hardware consists of a receiver having, as input, signals captured by an accelerometer in direct physical contact with the rail. The receiver circuit performs preliminary signal conditioning, namely, amplification, filtering and buffering. The latter subsystem (software) comprises a VQ process and the evaluation of distortion measures against pre-existing centroids. These represent in some way the spectra of two different classes of signals: “silence” (or absence of a moving train) and “moving train” (approaching or withdrawing). Prior to the detection, a training process must be carried out, so the two VQ centroids can accurately represent the two classes of signals. Therefore, training data must be collected and manually classified into the two previous classes. For both cases, centroids result from a clustering algorithm using reflection coefficients derived from a 14th-order Linear Predictive Coding (LPC) analysis. Incoming signals, either for training or classification purposes, are sampled at a rate of 16,000 samples per second and windowed into 300ms frames (window displacement is 100ms) and undergo an LPC analysis. The recognition or classification process is based on distortion measures to each one of the two centroids. In this work, two distortion measures are used and their performance compared: a modified Itakura-Saito distortion and the Euclidean vector distance. In the former, each vector has its energy normalized to unity while, for the latter, energy is used in the distance evaluation. Collected data was divided into training and classification corpora, respectively 80% and 20% and classification results are presented and discussed.