Sea Clutter Target Recognition Based on Modular Neural Network

Sea clutter is a highly nonlinear signal; it has a non-stationary nature both in time and space. Many scholars found that sea clutter data are chaotic. This indicates that the system has some internal rules, but it is difficult to make an analytic formula. The BP neural network with self-organizing fuzzy spatial mapping ability of artificial happens to be a powerful tool for such kind of problem. But it is not enough to get a better result for sea clutter target recognition use an only predict method. So we use modular integrated neural network method in this experiment. The results show that the modular integrated neural network can improve the recognition performance. Introduction of Sea Clutter Traditional Sea Clutter Analysis Based on the Mathematical Modeling. Sea clutter usually refers to the ocean surface radar backscatter echoes, is a very complex dynamical phenomena [1]. Sea clutter is a highly nonlinear signal; it has a non-stationary nature both in time and space. We can use mathematical modeling approach to analyze the sea clutter ultimate purpose is to improve the target detection performance. The classical statistical modeling approach is taking the sea clutter as a stationary random process; establishing a distribution model of the process, then performing target detection based on the maximum likelihood guidelines. There are several commonly used sea clutter distribution model, Log-Normal distribution [2], Weibull distribution [3], K distribution [4], K+Rayleigh distribution [5], etc. But classical detection methods based on statistical models generally require relatively high SCR when it is able to better target detection. In high sea state conditions single statistical distribution models cannot fully described sea clutter characteristics; generally it is difficult to obtain a satisfactory result. Sea ripple can be seen as corresponding to different roughness of surface, so some scholars have proposed the theory with fractal sea clutter nonlinear analysis and modeling [6]. Fractal characteristics of sea clutter has been described [7], multifractal and extended Fractal are often used to describe the sea-clutter [8]. The Chaotic Characteristics of Sea Clutter. Many scholars found that sea clutter data are chaotic when they are analyzing the sea clutter data [9]. Sea clutter is a very complex dynamic system; it is also a highly non-linear time-varying system; it is affected by the sea conditions, weather changes, etc. The chaos characteristics of sea clutter means that there is some kind of internal law. It is showed as correlation between short sequences in time series. Simon Haykin proposed a state space model as follows. )) ( ( ) 1 ( n x f n x   (1) ) ( )) ( ( ) ( n w n x h n y   (2) In which ) (n x is d-dimensional state vector in discrete Time, ) ( f is a nonlinear vector function, ) (n y is the only detectable value of the system, ) ( h is a nonlinear value function, ) (n w is additive noise. This indicates that the system has some internal rules, but it is difficult to make a analytic formula. The BP neural network with self-organizing fuzzy spatial mapping ability of artificial happens to be a powerful tool for such kind of problem. Using a BP neural network to predict a short time chaotic signal trained to build its unknown internal power system. But the chaotic 6th International Conference on Electronic, Mechanical, Information and Management (EMIM 2016) © 2016. The authors Published by Atlantis Press 1230 properties of the sea clutter will reduce in high sea conditions. At the same time, some scholars questioned the chaotic dynamics of sea clutter [10]. They suggested that sea clutter with positive and negative Lyapunov exponent, but this is not a chaotic criterion. Some scholars point out sea clutter does not have the power rate sensitivity characteristics which is the typical feature of the chaotic signal. Thus, It is not enough to get a better result for sea clutter target recognition use an only predict method. It is very susceptible to interference effects of sea conditions and noise. So we use modular integrated neural network method in this experiment. Two separate modules are the predictive neural network module and the feature classification neural network module. The Experiment Procedure and Results Analysis The Predictive Neural Network Module. We can build a multi-level prediction model as follows. Figure 1. Multi-level prediction model When there is only sea clutter signal, it should meet this prediction model, the energy of prediction error is small. When the target superposed on a background of sea clutter it should break this prediction model, the energy of prediction error is bigger. By comparison error energy, target detection can be performed under sea clutter. The sea clutter data used in the experiment is IPIX radar data of Mcmaster University in Canada.We take No. 17 IPIX data for training. The main target appears at the gate 09, the secondary target appears between the gate 08 and gate 11. Frist, construct a series of single-step sequence prediction single hidden layer BP neural network [11]. Then take every 10 data points as a packet to forecast the next data point; use the front 1000 data points for each range gate to make the training set. Finally, use the same network structure and parameters for training and testing. The training condition is the convergence error reaches a given threshold or the training times reached. Now, we compared several network convergence curves and the energy of prediction error.