A Multilayer Artificial Neural Network for Target Identification Using Radar Information

Transcription

1 Available online at A Multilayer Artificial Neural Network for Target Identification Using Radar Information James Rodrigeres 1, Joy Fundil 1, International Hellenic University, School of Economics & Business Administration, Thessaloniki, Greece Abstract In this paper a novel application for identification and recognition of flying objects is introduced. The procedure uses the back propagation learning algorithm for multilayer perceptron s and employs the incoming radar information as input for target recognition. The proposed algorithm can ideally be extended for use in an online learning process as the incoming radar information for known objects are being collected. The possibility of using fuzzy logic for identification is also investigated. Keywords: Back Propagation Learning Algorithm, Radar, Flying Object Introduction The purpose of radar target identification is in fact recognition of the type of flying object and classifying the information obtained from radar. The available radar information usually include, the line of sight, the distance, the velocity vector, and the radar cross-section (RCS) value. Flying objects of interest are various types of ships, airplanes, helicopters, and missiles. Although, many other applications such as pollution detections, weather forecasting and even sonar detection can be considered. For instance a recurrent neural network was used by T.Ziemke [1995]

2 for the detection of oil spills from Doppler radar imagery. Also, a generalized regression neural network scheme for target orientation using back-scattered data was presented by Kabiri et al. [1999]. Much of the work on artificial neural network (ANN) application is focused on situations in which an ANN is used in combination with a calculus based technique for solving a given problem. But, neural networks are capable of classifying, identifying and decision making without having to deal with calculus of the problem, as is the case in natural neural networks. It is apparent that future works in this area should focus on development of robust autonomous self-learning ANN which are capable of mimicking natural neural networks. Radar cross-section is an important parameter in target identification applications. RCS is the ratio of the electromagnetic energy that upon incident with an object in space is radiated back with the same period. The RCS of an object is usually a function of many parameters, such as: size, distance, orientation, configuration, and the transmitting medium. The ways neural networks operate are quite simple and their power is mainly due to their collective performance. The behaviors of biological neurons are quite complex and nonlinear but, a simple mathematical model can e designed based on their response to the sum of the incoming inputs. Fuzzy logic was introduced in 1960's by Professor L. Zadeh, and have found many applications in control engineering ever since. Combination of Fuzzy logic and neural networks have created the new subject of fuzzy-neural networks. In this paper a neural network is designed and taught for target identification using radar tracking information fuzzy-neural networks may as well be considered for this purpose. The classification categories for this identification can simply be: 'large', 'medium', and 'small', or it can include more detail parameters such as the type the

3 size and even the brand of flying or sailing targets. Of course, as the expected classification becomes more detailed, the more complex a neural network is required. Also, the teaching algorithm should have the diversity and the learning rate so that during a learning step other features are not destroyed. In the first part of this paper, after a general introduction, the concepts of artificial neural networks are discussed. Next, a brief discussion of fuzzy logic for identification is given. After that, the RCS analysis and computation using theoretical and experimental methods are presented. Following that, simulation results are discussed and finally, the conclusions are presented. Artificial neural networks Artificial neural networks are usually constructed from a collection of neuron like elements for performing a specific task and the range of possible applications for them are as large as natural beings capabilities. These applications include classification and decision making problems in the areas of communication, robotics, control, pattern-recognition, vision, and many yet unexplored area. Figure 1. A multilayer neural network with typical learning algorithm

4 Figure 1 shows a feed forward multi-layer neural network with three layers and a number of neurons in each layer. The first layer is usually called the input layer and must be the same size as the number of inputs. The last layer is called the output layer and must have the same size as the desired output. The middle layer are called the hidden layer and they should be sized properly according to the complexity of the required task. Generally the number of hidden layers and the number of neurons in them can be considered as optimization parameters. It has been shown that a network of interconnected neurons with n layer and m neurons in each layer is capable of classifying or representing any set or relation, and, the ability of a neural network for performing a given task is only limited by the network's complexity and size. One of the most popular algorithms used for teaching the network is "the least square back propagation method". This method is a simplified calculus base optimization procedure based on the steepest decent algorithm. dx dt ( x, t) E( x) (1) x Where, ( x, t) is a symmetric positive definite matrix which is called the learning matrix, and in general is dependent on the time and the vector x. the procedure finds the needed corrections for synoptic weights of all the neuron in the network in hyper plane direction of the gradient matrix such that the performance index E(x) is minimized. The performance index is usually defined as the difference between the actual and the desired response. The relations for such optimization problem are available in most neural network literature, [3]. Fuzzy Logic for Identification Fuzzy logic applications in control systems have been considered for practical development of expert systems ever since it was first introduced in the 1960's. in

5 general fuzzy logic provides the means for dealing with decision making problems using linguistic descriptions similar to the way human intelligent system works. When the set A is crisp or standard then any subset MA is either a number of A or not, i.e. there is no choice in between. However, for fuzzy sets a parameter may have a degree of membership associated with it. Figure. Fuzzy sets for describing radar information The main reason fuzzy sets are introduced here is because of the uncertainties that exist in many aspects of radar information. These uncertainties are mainly due to imprecision and sensor noise well as environmental noise which are present during scanning. Fuzzy logic and neural network can work together in two predominant forms. In the first form, fuzzy variables are designed for use in a multilayer network. Therefore, NN trainings are based on fuzzy parameters, i.e., neural networks are used to improve the performance of fuzzy logic. In the second form, the process of fuzzy logic reasoning is replied through the use of an artificial neural network, i.e., fuzzy logic and reasoning are built into artificial neural structure in order to achieve the information-based logic.

6 These logic can be used for aiding the proposed neural network for better classification and identification of the incoming target data. For example the RCS values could be categorized using linguistic parameters Large, Medium, and Small and the target direction (TD) could be categorized using region assignment, as shown in figure. Fuzzy decision making is based on solid rules such as: If X is A and Y is B then Z C1 is Since these rules use fuzzy variables, each decision will have a degree of truth associated with it, which in general depends on the degree of membership of the involved variables. The final crisp decision is made by difuzzification of the results in accordance with the degree of trueness of each crisp rule based decision. A multilayer neural network can be designed through the use of connectionist structure for replication of this fuzzy radar information-based algorithm. The implementation and simulation of this procedure will be presented in a future paper. The RCS Analysis and Computation Usually analytical computation of RCS for complex objects is difficult but, for simple objects, such as spheres, cylinders, and rectangular plates, this can be obtained using known theories. Estimation of the RCS for more complex objects can be accomplished using the superposition principal. RCS has the dimension of cross sectional area and is usually measured as compared to that for some ideal reference object. An experimental relation for RCS computation from reference [1] is: RCS P L (4 ) L 1 Lmr Lmt (4 ) rt r L p () Gr 0 Pr Gt Where,

7 Pt, is the transmitter power. Gt, is the gain of the transmitter antenna. Gr, is the gain of the receiver antenna in the direction of the target. rt, is the distance between the target and the transmitter antenna. r, is the distance between the target and the receiver antenna. Lp, is the numerical coefficient for computation of the polarization losses Lt, is the numerical coefficient for computation of transmitter losses Lr, is the numerical coefficient for computation of the receiver losses Lmr, Lmt, are the numerical coefficient for computation of the losses in the receiver and the transmitter medium Theoretically the RCS value is defined as: RCS Power density of the signal coming to the receiver. Powerd density of the transmiti ng wave toward the target. or RCS Er (3) Lim 4 R E t in which, Et, is the intensity of the reflected field in the receiver location and, Er, is the intensity of the transmitted radiation toward the target.

8 Table 1. Radar cross-section of some simple objects Geometry type Size dependence Frequency dependence Formula Flat plate L 4 F 4 a b RCS cylinder L 3 F 1 Sphere L F 0 RCS RCS ab a Right dihedral corner reflector L 4 F 8 a b RCS Detection success highly depends on the power of the processing techniques to extract target information from the background environment signals. A target amount of effort has been given to stochastic modeling and application of new techniques such as neural networks to overcome these problems. The target reflection is usually compared with the reflection of an ideally reflective flat object which radiates back the wave in a isentropic process. The RCS of simple objects can theoretically be computed using the Maxwell wave equation by considering the boundary conditions and the surface material. This method is quite complicated and time consuming. A simpler method which is frequently used is the high frequency approximation and the geometric theory of diffraction. The RCS of some simple objects and their dependence to the direction and frequency are presented in table 1.

9 It should be noted that RCS of most objects are extremely direction and frequency dependent. Although, the principal of superposition applies only for linear and independent systems, yet, relatively good estimates for the RCS of complex objects may be obtained by breaking them in to simple parts. A word of caution in this regard is the interaction of radiated wave reflected from the surrounding surfaces, which is usually not considered in the superposition process. The physics behind a high frequency radar based integrated maritime surveillance system is discussed in Sevgi et al. (004). Simulation Results Radar target identification is a complicated and erroneous task which have traditionally been conducted by a human operators using past experiences, radio communication devices, optical instruments and radar information. Artificial neural networks are quite well structured for one-to-one mapping of any shape of input pattern into a related desired output. The power of neural networks in human like reasoning and decision making have made them suitable for finding the solution for numerous problems in the area of pattern recognition and machine vision. Here, a novel approach for target identification using radar information is presented. The inputs to the network are the target direction (DT), and the radar cross-section (RCS). The output of the network is chosen to be an index binary number which is used to identify the target. Figure 3 shows the RCS values of three typical radar targets as a function of direction. Because, the indicated objects are geometrically symmetric with respect to the XZ plane only the RCS values for angles between 0 to 180 are sketched. The values for directions between 180 to 360 would be a mirror image of the sketch.

10 A three layer neural network with ten neurons in each hidden layer was used for this identification problem. The network was thought for 14 epochs, by using the least square error minimization and back propagation method. Figure 4 shows the error values during the learning process. Although, the RCS values of different objects, as shown in figure 3, are extremely direction dependent, the learning algorithm was able to coverage to the desired solution with virtually zero error. The simulation results are summarized in Figure 5. As indicated in this Figure, the network after learning was able to identify each target according to their respective RCS and direction. The interesting point in this problem is that the network learning can be an on-line and an ongoing process during radar tracking. In other words, the network can be thought during operation according to various incoming signals. In this simulation, it is assumed that the RCS numerical values are distance corrected using equation I, i.e., the RCS values used are all adjusted to correspond to a given reference distance between the object and the radar. However, in a more realistic situation the distance could also be used as an input to the neural network. This increase the size of the network and reduces the radar information preprocessing, before they are used in the identifying neural network. Figure 3. The RCS values for three different targets

11 Figure 4. Output index desired and actual values for the simulation Conclusions Target identification using radar information is a difficult task, which is usually done by human operators. Efficient algorithms for target identification can simplify the job of human operators considerably. Artificial neural networks computing power have proved to be very efficient in decision-making and classifier systems. A neural network was designed and simulated for target identification using radar information. The network was able to distinguish different targets using distance adjustment RCS values and the direction information. More realistic scenarios can readily follow using the same procedure. A fuzzy neural network is also proposed for this target identification. The network is designed for replication of the fuzzy information-based radar target classification. The implementation and simulation of this procedure will be presented in a future paper. References

12 1. Bachman, Christain G., (198), "radar Targets, Heath and Company.. Brookner, Eli, (1988), "Aspect of Modern Radars", Artech House. 3. Cichocki, Andrzej and Rolf Unbehauen, (1993), "Neural Networks for optimization and Signal processing", John Wiley & Sons. 4. Kabiri, Ali, Nima Sarshar and Kasra Barkeshli, (001), "A Generalized regression Neural Network (GRNN) Scheme For Robust Estimation of Target Orientation Using Bach-Scattered Data", Department of Electrical Engineering, Sharif University of Technology. 5. Sevgi, Levent and Anthony M. Ponsford, (004), "An HF Radar Based Integrated Maritime Surveillance System", Raytheon System Canada Ld. Integrated System Division. 6. Ziemke, T., (1995), "A Recurent Neural Network for the Detection of Oil Spills from Doppler Radar Imagery", Department of Computer Science, University of Skovde, S5418 Skovde, Sweden. 7. Shaw, T.S., "Fuzzy Control of Industrial Systems, Theory and Applications", Kluwer Academic Publishers, 1998.