IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 3, JUNE 2001 749 Detection of Helicopters Using Neural Nets Sohail Akhtar, Moustafa Elshafei-Ahmed, Member, IEEE, and Mohammed Shahgir Ahmed, Senior Member, IEEE Abstract—Artificial neural networks (ANNs), in combination with parametric spectral representation techniques, are applied for the detection of helicopter sound. Training of the ANN detec- tors was based on simulated helicopter sound from four helicopters and a variety of nonhelicopter sounds. Coding techniques based on linear prediction coefficients (LPCs) have been applied to obtain spectral estimates of the acoustic signals. Other forms of the LPC parameters such as reflection coefficients, cepstrum coefficients, and line spectral pairs (LSPs) have also been used as feature vec- tors for the training and testing of the ANN detectors. We have also investigated the use of wavelet transform for signal de-noising prior to feature extraction. The performance of various feature extrac- tion techniques is evaluated in terms of their detection accuracy. Index Terms—Helicopter detection, linear prediction coefficient (LPC), line spectral pair (LSP), neural networks (NNs). I. INTRODUCTION H ELICOPTERS are highly mobile tactical weapon plat- forms. A helicopter can be used as an intruder’s trans- port, and as an escape vehicle after an intrusion has been com- mitted. The low-flying ability of such aircraft enables them to penetrate the defense system, undetected by conventional radar. Building a system of remote sensors to detect and track single and multiple very low-flying helicopters is an important defense problem. Detection of helicopters using their sound signatures has been a focus of a number of recent research works [1]–[8]. The conventional method for helicopter sound detection uses the ratio of the main and tail rotor frequencies and their har- monics as the key helicopter noise features. However, most of the helicopter sound detection studies have used simulated spec- trums based on only a fixed discrete spectrum from the rotors. In this paper, we propose an artifical neural network (ANN)-based helicopter sound detection system. The detection study of this paper is based on a helicopter sound simulator designed to pro- duce more realistic sound characteristics of the helicopter’s ro- tors, including the effect of aerodynamic vortex shedding, blade thickness noise, Doppler effect, as well as the atmospheric at- tenuation, and terrain effects [9]. Mori et al. [1], [2] investigated the use of a network of re- mote sound sensors to detect and track single and multiple very low-flying helicopters where information from different sensor nodes was exchanged to improve the resolution of the sound de- Manuscript received May 12, 1999; revised December 26, 2000. This work was supported by King Fahd University of Petroleum and Minerals. S. Akhtar and M. Elshafei-Ahmed are with the Department of Systems En- gineering, King Fahd University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia. M. S. Ahmed was with the Department of Systems Engineering, King Fahd University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia. He is now with E/E Engineering, DaimlerChrysler Corporation, Auburn Hills, MI 48326 USA. Publisher Item Identifier S 0018-9456(01)04387-X. tectors. Feder [3] studied the problem of estimation of the fun- damental frequency of the helicopter rotor in the presence of a wide band interfering signal, e.g., generated from a nearby jet engine. The aperiodic nature of the helicopter signal is modeled as a strictly periodic component with an interfering Gaussian noise with unknown spectrum. The parameters of the model are estimated using a likelihood function. Zhiping [4] applied cyclostationary estimation techniques for helicopter signal de- tection subject to time-varying Doppler shift and in the pres- ence of possibly nonstationary background noise. He also used a cyclic frequency smoothing method to have a better estimate of the average fundamental frequency of the time-variant Doppler shifted helicopter acoustic signal. Cabell et al. [5], [6] proposed two pattern classifiers: 1) a statistically based Bayes classifier and 2) an ANN classifier. Selected peaks of spectral amplitudes of the fundamental and first seven harmonics are fitted to least square regressions and used as features for the ANN classi- fier. The Bayes classifier identified 67% of the audio segments of three helicopters while the ANN classifier was correct 65% of the time. The results suggested that additional features cov- ering more about signal generation and propagation should be included in the ANN training to improve the performance. Elshafei and Ahmed [7] used a recorded helicopter sound and a number of nonhelicopter sounds to train their ANN detector. They utilized feature vectors based on the main spectral peaks and other components in the frequency band from 150 Hz to 350 Hz. The Goertzel algorithm (GFFT) was used for evaluating the spectrum. Feature vectors of length 30 were used as input vectors for the training and testing of the classifier. The ANN classifier with one hidden layer of neurons was used. The results obtained have shown 99.5% correct detection of helicopter over test samples. In this paper, different spectral parameters, such as linear pre- diction coefficients (LPCs), reflection coefficients (RCs), LPC cepstrum coefficients (CCs), and line spectral pairs (LSPs) are investigated. In the following section, we briefly give an overview of the structure of the ANN- based pattern classification procedure as used for helicopter sound detection. The third section out- lines the key features of the simulator used to generate the heli- copter acoustic signal. Extraction of the feature vectors using the LPC-based techniques is given in Section IV. The simulation re- sults are given in Section V. Finally, Section VI addresses some possible strategies for improving the detection performance in the presence of noise. II. DETECTION USING PATTERN CLASSIFICATION A pattern classification system aims to classify an object based on its previous knowledge of it. Such a system oper- ates in three phases: a training phase, a testing phase and a 0018–9456/01$10.00 © 2001 IEEE