Radar Signal Detection By Multilayer Perceptron Trained With The Quadratic Error Cost Function

The main drawback of conventional radar signal detection techniques is thedifficulty of modeling for complex target, so that it’s hard to get a good performancewhen the target statistical properties vary from those assumed in the design.Performing the threshold detection process requires knowledge of the likelihoodfunction of the target probability distribution. However, machine learning can helpmake a good approximation to radar detectors without knowledge of the likelihoodfunctions. Multilayer perceptron is a kind of artificial neural networks, belonging tosupervised machine learning. Multilayer perceptron is a kind of nonlinear classifierwhich can deal with nonlinear separable problems. Multilayer perceptron based on theback-propagation can use specific learning algorithms to update itself parameters viaiteration, and achieve the purpose of self-optimization. So that it can be applied todifferent areas.We use the multilayer perceptron to approximate the optimum Neyman-Pearsondetector. The training process of multilayer perceptron can avoid the target propertymodeling and analysis of design parameters, and thus solve the problems mentionedabove. Nowadays multilayer perceptron neural networks have been widely used inimage processing, semantic recognition, diagnostic tests and other areas.The main considerations in designing multilayer perceptron are the selection ofthe cost function, activate functions and iterative algorithm. In order to minimize thecost function, multilayer perceptron design process needs to determine the structureand calculates its synaptic parameters. Depending on the specific issue, thereplacement of cost function can sometimes get better results. Different kind of costfunctions have their own characteristics, the quadratic error cost function with fastconvergence properties, more in line with radar signal detection system timelinessrequirements.In the thesis, multilayer perceptron trained to minimize the quadratic error costfunction is used to approximate the optimum Neyman-Pearson detector. First, radarechoes are generated according to the characteristics of the radar echoes. Then, thestandard radar detection system is modeled to generate three kinds of signals from different stages of processing in the system. Finally, the signals are applied into themultilayer perceptron to see which kind of signal suited best. Our simulation resultsshow that the modulated signal can suit the multilayer perceptron best. The multilayerperceptron with quadratic error cost function has a good approximation toNeyman-Pearson.