A novel Method for Radar Pulse Tracking using Neural Networks

Transcription

1 A novel Method for Radar Pulse Tracking using Neural Networks WOOK HYEON SHIN, WON DON LEE Department of Computer Science Chungnam National University Yusung-ku, Taejon, KOREA Abstract: - Within restricted response time, in order to track pulse train of Radar and predict the next pulse time, we propose the Neural network with Dynamic structure. This Dynamic structure consists of basic structures. And it is a variable structure in that the basic structures are added to the previous structure as input signal is received more. This structure is applicable to the case that has noise pulses, missing pulses and requires rapid response time. In this study, we apply the proposed neural network to predict a radar pulses with a nonlinear characteristics and test it for 4 level stagger, patterned signal with sinusoidal type. Key-Words: - Pulse, Tracking, PRI, Dynamic structure, Neural Network Introduction Multi-layered feed-forward neural networks have been developed and used as predictor of stock prices and weather [],[4]. There are, however, many restrictions when we apply it to extremely rapid real time series problem (e.g. radar signal processing), because it requires many processing time. In radar processing problem input data is changed every hundreds of microsecond to tens of millisecond, and it also has often missing pulse and noisy data. In this complex radar signal environment, in order to handle the special signal only, we have to track the pulse train and predict the next TOA(Time Of Arrival). Also, tracking for frequency is needed in the frequency domain. Many modern radars use complex pulse repetition interval(pri) modulations like stagger, jitter and pattern type signal against the many intentional interference methods[2],[3]. So the pulse tracking system which is applied in EW(Electronic Warfare) must track many types of PRI modulation and the patterned frequency signals, even in the complex multi-emitter condition -- missing and spurious pulses. In this paper, we propose a dynamic neural network structure that can track the patterned signal even in the noise environment within a tolerable error range. This structure uses Error Back Propagation(EBP) basically. Since we have to get a precise prediction and a fast response time, we output a tracked data by using a simple structure for small amount of received signal. As the number of received signals is increased, we expand size of the structure and make its accuracy high. The structure is expanded until the number of input node exceeds one cycle of patterned signals. To test this structure, we use a sinusoidal type signal and 4 level stagger of the patterned signal. 2 Radar signal environment and characteristics Radar uses pulse trains to acknowledge information of targets. This pulse train is transmitted with some time interval that is called PRI. Many radar systems change the PRI for the purpose of preventing the tracking of a ECM equipment(tracker) and processes the signal effectively[2]. Received radar signal from an antenna is described with periodic function equation as follows: X + t + = Xt + F( ti) et () Where, X t+ is time of arrival (TOA), X t is previous time, F(t i ) is function that changes PRI pattern, e t is sum of a radio wave transmission error and its receiving error in the antenna. Difference of two consecutive TOA, X t+ - X t is PRI and its range is usually several tens of microseconds to several hundreds of microseconds. Since e t is small relatively, PRI is varied according to the function F(t i ). If F(t i ) is constant then it is called stable PRI, if it has a pattern, it is called a sine type PRI or a triangle type

2 PRI, for instance. Also, in frequency domain a pattern is applied like PRI and it is described as equation (2). Freq + t + = Freq + F( ti) et (2) Freq t+ is next frequency, Freq is base value, F(t i ) is function that changes a frequency pattern, and e t is sum of error. In equation (), (2), function F(t i ) has a similar pattern and characteristics. So, the tracking algorithm for PRI and frequency can be used commonly. As equation (), (2), the radar pulse has missing pulses in the consecutive pulse trains besides noise e t. This radar pulse problem makes hard to track the signal, but tracker has to follow the pulse train and predict next TOA. In this work we suppose that the tracker receive a brief information(pattern type, range of values) about PRI and frequency from ES(Electronic Support) part as in other tracking systems. 3.2 Dynamic structure As in case of tracking radar pulse trains, although it has some error, for the fast response, we proposed a dynamic structure that varies its structure according to the lapse of time(fig.). The neural network structure we propose in this paper has a basic structure using EBP learning algorithm. Dynamic structure means that structure of neural network is variedaccording to the dimensionof the input data. For example, Fig. (a) is a basic structure of this dynamic structure and it has 20 input nodes, 40 hidden nodes and one output node. We can change the structure according to the requirement of the tracker. The number of input nodes in the input layer, for instance, could be determined by the first response time that is requested by a tracking system. The combined structure of these two basic structures is (b), while (c) and (d) are the ones with 3 and 5, respectively. 3 Neural Network structure 3. Simple 3-layer approach In radar pulse tracking we can use a simple 3-layer neural network[5]. As described in (3), objective function is defined as the sum of the square of the difference between the output( y k ) and desired output ( yd k ). (a) 20(basic structure) (b) 40 2 T p = ( ydk yk ) (3) 2 In time domain, a tracking system that tracks consecutive input signals needs an accurate prediction and a fast response. We use simple 3 layer neural network for tracking system in which its input node is PRI and its output node is a predicted PRI. By experiment, the number of input nodes in this network must be much larger than that of PRI of one cycle for a stable tracking. For example, if a period of pattern is 50milli-seconds and PRI is 500micro -seconds, the number of pulse becomes 00. So the number of input nodes is set to 00. If it is fully connected network, it will operate after all input nodes are assigned. It means that the network will generate output data(prediction) after the lapse of 50ms. (c) 60 (d) 00 Fig. A structure is changed according to the requirement of the tracker. If it is a patterned signal which has a periodical, the input node number of the last combined structure must be lager than the number of pulses in a cycle of the pattern. If it is so, the finally structured neural network can learn the information which is included within a whole cycle, and it can predict a next value(pri or frequency) accurately.

3 The number of input nodes in the basic structure is related to the minimum response time required in the application. This paper takes 20 input nodes as an example, which means we output a predicted result after receiving 20 pulses. If the number of received signals increases with the lapse of time, another 20-cell is appended to the previous structure. Then, a neural network structure having 40 input nodes is made. After this time, the predict data is made by using this new structure. It continues until the number of pulses corresponding to one cycle is received. A summary of the algorithm is: ) Define the minimum response time of tracking system, RT min 2) Design the basic structure of dynamic structure neural network BN input = RT min /PRI (BN input : input node number of Basic structure) 3) Determine the number of maximum input node, MN input CPRI MN input (CPRI : cycle time of pattern/pri) 4) Determine structure number to append to previous structure, SN SN= MN input / BN input 4 Experiments and Discussion The dynamic structure suggested in this paper can be applied to the prediction of a periodical and non-stationary signal system. This experiment uses the signals that got from the experimental set-up in unechoic chamber. The experiment signals are obtained by measuring the signals received through the antenna with special measuring equipment. Radar simulator makes the radar signal data and three antennas transmit the signal. At the other side one antenna receives that RF signal and converts it to digital data with A/D converter. Filter & Emitter identifier filters the noise. Tracker system tracks the three radar signals and predicts a next pulse event. Table. Experimental radar features No PRI sine No. emitter of Table has a sinusoidal pattern type in frequency modulation. In case of the No. signal we suppose that the minimum response time of tracking system, RT min is 9.6milliseconds. Then, BN input (input node number of basic structure) is 9.6ms/480µs=20. We designed a basic structure with 20 input nodes, 40 hidden nodes and one output node. The number of one cycle pulse(cpri) is about 42(20ms/480µs). So, the maximum input node number, MN input is decided to be 42. SN= 42 / 20 =3. Thus, we design three networks. One is a basic structure(20 input nodes), another is a two combined structure(40 input nodes) and the other is a three combined structure(60 input nodes). Each network is learned with the sample data. When the learned network is practiced in real data, we have to count the incoming PRI data. If incoming data is less than 20, then neural network does not produce output and when incoming data is between 20 and 40 then it produces output using basic structures. And then a second basic structure is appended to the first basic structure (e.g. two combined structure) (+/-00,20ms) 2 stable sine 400 (+/-00,30ms) No.-Freq-20 fix 480 Stagger (4 level) 650, , 672 jitter 560 Output data.005 Radar Simulator Filtering & Emitter Identification Tracker System 5 Fig.2 Experimental set-up Fig.3. Tracked result by a basic structure

4 The tracked results by the structure is made of 20 input nodes(basic structure) are shown in fig.3. From the time when the number of pulses received consecutively are 40 and more, structure (b) in Fig. is used, and its tracking results are shown in Fig.4. Similarly, Fig.5 shows the results when the number of PRI is 60 and over No No.-Freq Fig.4. Tracking result by 2 basic structures (40 input nodes) From the test result, as we see in Fig.3, the first response time of the tracker is 20x480us and its maximum error is 25.9, but in Fig.4 its maximum error is 6.6, and in fig.5 it is 2.9. These show that the tracker gradually becomes more accurate as the structure gets increased. No.2 in Table is the one whose frequency is fixed and the PRI is 4 level stagger. So we design neural networks: one for frequency, another for PRI No.-Freq-60 Output data Fig.6. Tracking result Stable Fig.7 shows the predicted results of signals when pulses are missing on the way and signals are not input into the neural network. When signals are not received, then the final PRI is used as input of the neural network. PRI No.2-stagger Fig.7. Tracking result-4 level stagger Fig.8 is the tracking results of sine type frequency(no. 3) with noise No.3-Freq Fig.5. Tracking result by 3 basic structures (60 input nodes) Fig.6 shows the tracking results of stable type frequency with noise Fig.8. Tracking result Sine type

5 5 Conclusion EBP algorithm has a memory effect in supervised learning environment and has a characteristic of learning repetitive signals effectively. In this paper we propose a neural network having a dynamic structure, which consists of basic strucuturess with EBP algorithm. This structure is good to predict periodical time series signals with noise. Here we apply the dynamic structure to the signals with the following environment (a) with noise which gets larger or smaller than the basic PRI (b) with missing pulses (c) with prediction response time required in several pulses We test our dynamic structure in radar tracking system having more than one of the above environments. This radar tracking system predicts the period of it s target well. For more study we are to work on algorithm with a shorter response time with a dedicated hardware. References: [] C. Gent and C. Sheppard, A general purpose neural network architecture for time series prediction, in Proc. IEE ICANN 92, 992, pp [2] Richard G. Wiley, Electronic Intelligence : The Analysis of Radar Signals, 2nd ed., pp , Artech House, 993. [3] Gregory P. Noone, A Neural Approach to Automatic Pulse Repetition Interval Modulation Recognition, IDC99 Proceedings, 999, pp [4] McClelland, Rumelhart, and the PDP Research Group, PARALLEL DISTRIBUTED PROCESSING Vol. 2, The MIT Press, 986. [5] Simon Haykin, NEURAL NETWORKS : A comprehensive foundation, 2nd ed., pp , Prentice Hall, 999.