IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 52, NO. 6, DECEMBER 2003 1771 Classification of Audio Radar Signals Using Radial Basis Function Neural Networks Trent McConaghy, Member, IEEE, Henry Leung, Member, IEEE, Éloi Bossé, and Vinay Varadan Abstract—Radial basis function (RBF) neural networks are used to classify real-life audio radar signals that are collected by a ground surveillance radar mounted on a tank. Currently, a human operator is required to operate the radar system to discern among signals bouncing off tanks, vehicles, planes, and so on. The objective of this project is to investigate the possibility of using a neural network to perform this target recognition task, with the aim of reducing the number of personnel required in a tank. Different signal classification methods in the neural net literature are considered. The first method employs a linear autoregressive (AR) model to extract linear features of the audio data, and then perform classification on these features, i.e, the AR coefficients. AR coefficient estimations based on least squares and higher order statistics are considered in this study. The second approach uses nonlinear predictors to model the audio data and then classifies the signals according to the prediction errors. The real-life audio radar data set used here was collected by an AN/PPS-15 ground surveillance radar and consists of 13 different target classes, which include men marching, a man walking, airplanes, a man crawling, and boats, etc. It is found that each classification method has some classes which are difficult to classify. Overall, the AR feature extraction approach is most effective and has a correct classification rate of 88% for the training data and 67% for data not used for training. Index Terms—Audio signal, classification, neural net, radar, radial basis function. I. INTRODUCTION T HIS PAPER reports the result of a project supported by the Department of National Defence of Canada to inves- tigate the possibility of replacing a trained human operator by an artificial neural network in performing radar target recogni- tion. Radar is still an important sensor in current defence sys- tems for sensor surveillance, since radar is the only sensor which can effectively detect targets at long distances, in the dark and almost any weather conditions [1]. Currently, the identification function in a radar system is usually carried out by human op- erators who have had special training. However, the amount of incoming data is already far more than humans can handle, and Manuscript received October 1, 2001; revised July 27, 2003. T. McConaghy is with the Analog Design Automation, Inc., Ottawa, Ontario, ON K2P 2C2 Canada (e-mail: trent@analogsynthesis.com). H. Leung is with the Department of Electrical and Computer Engineering, University of Calgary, Calgary, Alberta, AB T2N 1N4 Canada (e-mail: leungh@enel.ucalgary.ca). E. Bossé is with the Decision Support Technology Section, Defence Research and Development Canada, Valcartier, Quebec, QC G3J 1X5 Canada (e-mail: eloi.bosse@drdc-rddc.gc.ca). V. Varadan is with the Department of Electrical Engineering, Columbia Uni- versity, New York, NY 10027 USA (e-mail: varadan@ee.columbia.edu). Digital Object Identifier 10.1109/TIM.2003.820450 the amount of available manpower is shrinking. As well, the re- duction of crew size in operating a surveillance system with no reduction in the surveillance performance is beneficial. Devel- opment of an automated system to carry out radar identification function is therefore highly desirable [2]–[4]. In this project, we consider radar signals collected by bat- tlefield radar systems that are installed in a tank. These signals are at audio frequencies, and a trained human operator is usu- ally required to stay inside the tank to recognize these audio radar signals. The data set that we use was collected by an AN/PPS-15 ground surveillance radar. In this type of radar, a microwave continuous wave signal of constant frequency is transmitted and is used to form a local oscillator for the received signal. The output signal can then be reduced to audio frequencies and represents the Doppler shift incurred because of target motion along the radar line of sight [5]. There are totally 13 classes of signals from various moving targets in this experiment, including trucks, tanks, men marching, a man walking, a man crawling, boats, small airplanes, and birds. Neural networks are investigated here for classifying the audio radar signals automatically due to its efficiency in processing audio signals [6], [7]. In particular, the radial basis function (RBF) neural network is used as the neural net classifier in this study [8]. The main reason is that a linear adaptive algorithm can be used in training the coefficients of the network. This makes the RBF net implementable for real-time application. For other neural networks such as recurrent neural networks [9]–[11], the training process may not be carried out in a real-time on-line fashion using current technologies [12]. Because audio signals are usually nonstationary, this on-line learning ability is preferred to allow the classifier being adaptive to the environment. In addition, an RBF is a universal approximator which has the capability of approximating a decision boundary of any shape. Based on our survey, we find three popular approaches for time series classification using neural networks. The main difference is the feature extraction process. The first approach applies a neural network classifier to the raw data directly without any separate feature extraction procedure. The input is basically the delayed values of the time series , i.e., for a network with input units. However, the raw data approach usually has a poor performance and we have the same observation for this audio radar classification problem. The detailed analysis of this approach is therefore omitted due to space limit. The second approach uses a model to represent the data, and applies a neural network classifier on the model parameters. Fig. 1 illustrates this approach. Here, 0018-9456/03$17.00 © 2003 IEEE