Analyses and Tests of Three Signal Processing Methods for HelicopterIdentification

摘要

A helicopter in the battle field is a large threat to soldiers and armouredvehicles, it can suddenly appear behind hills, vegetation etc. A helicopter can also fly close to the ground and avoid detection by radar. Therefore it would be a great advantage still being able to recognize a helicopter, e.g. using its characteristic sound. Sound from helicopters and non-helicopters has been recorded with a microphone. Three signal processing methods for helicopter identification have been studied and compared. The physical model uses the characteristic frequencies of the helicopter and it was implemented as an algorithm. In the filter model, an AR (autoregressive) filter that extracts the characteristics of helicopter sound from white noise has been created. The neural network model accepts an arbitrary input signal and the output signal is 'helicopter/not helicopter'. The neural network uses the back-propagation training algorithm and an AR parametrization of the signals as pre-processing. The best classifier is the neural network model. It does not suffer from inflexibility like the physical model or the filter model. There is, though, need for further studies of better pre-processing methods.