Moving Target Indicator Using Neural Networks
2007
Here we, illustrate performance comparisons of the moving-target indicator using RBF, LP and BP neural network. In these simulations, we considered AR model P = 6. Number of samples in one sinusoid n = 13. The RBF neural network has N = 13 input nodes, K = 13 hidden nodes. It is trained by 25000 epochs. Sinusoidal signals corrupted by white Gaussian noise. Test set contains 100000 noise-only vectors and 25000 signal-plus-noise only vectors. The sampling period Tr = 1. Bc and Bs as the bandwidths of the noise spectrum and the target signal spectrum respectively. Here we considered Noise bandwidth spectrum as an threshold. The central frequency values of fc = 100 Hz and fs = 300Hz ,where fc and fs represent the central frequency values of noise and signal respectively. Signal-to-noise ratio varies using noise variance. For different bandwidths of the noise spectrum Bc and the target signal spectrum Bs, we estimated the corresponding frequency values of roots. Figure 2-(a), 2-(b) and 2-(c) show the relationship between the estimated frequency values of the roots and the SNR for LP, BP, and RBF respectively. Curves 1 and 3 represent the central frequency values of the noise and signal respectively. Curves 2 and 4 represent the estimated frequency values of the two roots. In case, of LP approach we are able to detect targets up to -25 dB and -60 dB SNR respectively as shown in Figure 2-(a) and 2-(b) while, in case of RBF neural network approach the signal can be found even if SNR = -100 dB as illustrated in Figure 2-(c). Figure 2-(d) show the comparison of LP, BP and RBF neural network performance. Table 1 show the performance comparasion between LP, BP And RBF MTI.