Elsevier

Neural Networks

Volume 14, Issue 3, 1 April 2001, Pages 325-344
Neural Networks

Contributed article
A What-and-Where fusion neural network for recognition and tracking of multiple radar emitters

https://doi.org/10.1016/S0893-6080(01)00019-3Get rights and content

Abstract

A neural network recognition and tracking system is proposed for classification of radar pulses in autonomous Electronic Support Measure systems. Radar type information is considered with position-specific information from active emitters in a scene. Type-specific parameters of the input pulse stream are fed to a neural network classifier trained on samples of data collected in the field. Meanwhile, a clustering algorithm is used to separate pulses from different emitters according to position-specific parameters of the input pulse stream. Classifier responses corresponding to different emitters are separated into tracks, or trajectories, one per active emitter, allowing for more accurate identification of radar types based on multiple views of emitter data along each emitter trajectory. Such a What-and-Where fusion strategy is motivated by a similar subdivision of labor in the brain.

The fuzzy ARTMAP neural network is used to classify streams of pulses according to radar type using their functional parameters. Simulation results obtained with a radar pulse data set indicate that fuzzy ARTMAP compares favorably to several other approaches when performance is measured in terms of accuracy and computational complexity. Incorporation into fuzzy ARTMAP of negative match tracking (from ARTMAP-IC) facilitated convergence during training with this data set. Other modifications improved classification of data that include missing input pattern components and missing training classes. Fuzzy ARTMAP was combined with a bank of Kalman filters to group pulses transmitted from different emitters based on their position-specific parameters, and with a module to accumulate evidence from fuzzy ARTMAP responses corresponding to the track defined for each emitter. Simulation results demonstrate that the system provides a high level of performance on complex, incomplete and overlapping radar data.

Introduction

Radar Electronic Support Measures (ESM) involve the search for, interception, location, analysis and identification of radiated electromagnetic energy for military purposes. ESM hereby provide valuable information for real-time situation awareness, threat detection, threat avoidance, and for timely deployment of counter-measures (Brown & Thurbon, 1998; Davies and Hollands, 1982, Grant and Collins, 1982, Schleher, 1986, Schleher, 1999, Sciortino, 1997, Tsui, 1986, Wiley, 1993).

A critical function of radar ESM is the real-time identification of the radar type associated with each pulse train that is intercepted. Current approaches typically involve sorting incoming radar pulses into individual pulse trains, then comparing the pulse train characterizations with a library of parametric descriptions, which yields a list of likely radar types. This task is challenging because of: increases in environment density (e.g. pulse Doppler radars that transmit hundreds of thousands of pulses per second); dynamically changing environments; multiplication and dispersion of the modes for military radars; agility in parameters such as pulse repetition interval, radio frequency and scan; unknown and reserve modes for which no ESM library entry exists; overlaps between the parameters of different radar types in the ESM library; and noise and propagation effects that lead to erroneous or incomplete signal characterization. These aspects of the problem place severe stress on current ESM systems.

In this paper, an alternative approach is examined. A new recognition system combines diverse sources of information in order to predict the most likely radar type for each intercepted pulse. Type-specific parameters of the input pulse stream are used to classify pulses according to radar type, while environment-specific parameters are used to separate pulses corresponding to active emitters. Such separation allows the system to accumulate the classifier's responses for each emitter, and therefore to predict an emitter's identity based on one or multiple responses.

A key component of the new recognition system is a neural network classifier that is trained to determine the types of radar emitters present in the environment. The system learns autonomously, directly from data collected in the field, to identify pulse parametric ranges corresponding to specific radar types. Aside from avoiding some of the pulse sorting, training on data from the actual environment to approximate an unknown mapping function may deliver greater predictive accuracy. Furthermore, the need for by-hand construction of an emitter library is obviated.

From an ESM standpoint, training a system directly on radar data is a radical departure from current practice. At present, data are collected, analyzed, combined with prior information, and distilled into ESM libraries off-line by skilled analysts. New libraries, containing explicit radar type descriptions, are disseminated to the field as needed. One inconvenience of the conventional approach is that it is very complex, time-consuming, and does not allow for rapid modifications of ESM libraries upon discovery of new radar modes in the field. Using a neural network able to learn incrementally offers a framework for refining familiar, or adding unfamiliar, radar type descriptions on the fly.

In a particular realization of the recognition system, fuzzy ARTMAP (Carpenter et al., 1992, Carpenter et al., 1991b) is considered for neural network classification of pulses from their type-specific parameters, whereas nearest-neighbor matching with a bank of Kalman filters (Bar-Shalom and Li, 1993, Blackman, 1986) is considered for separation of pulses from their environment-specific parameters. The features of the system include: (1) by virtue of the fast-learning capabilities of ARTMAP neural networks (Carpenter et al., 1991a, Carpenter et al., 1991b), new information from familiar or unfamiliar radar type classes can be learned incrementally without retraining on the whole data set; (2) classification decisions can be made on the basis of single pulses or, for greater accuracy, on the basis of streams of pulses that have been determined to come from a given emitter. This determination is performed either by a time-of-arrival (TOA) deinterleaver, or, when TOA deinterleaving is not practical, by a Kalman filter that tracks the bearing and amplitude of the pulses. The system is thus an example of a neural system combining temporal—When—and positional—Where—information with featural—What—information to arrive at its decision. It is well-known that the mammalian brain also divides What and Where computations into separate, but mutually interacting, cortical processing streams. Our What-and-Where model shows how this strategy can generate higher accuracy in identifying radar emitters; (3) the ‘familiarity discrimination’ extension of fuzzy ARTMAP, called ARTMAP-FD (Carpenter et al., 1991a, Carpenter et al., 1991b), allows the system not only to detect pulses from unfamiliar radar type classes (not presented during training), but also to determine the threshold for rejection based on all of the training data, without the need for holding back a portion for a validation set. This ability to determine the reject threshold on-line makes possible on-line learning of pulses from unfamiliar classes (LUC) (Granger, Rubin, Grossberg & Lavoie, 2000); (4) new extensions to fuzzy ARTMAP permit both training and testing on data with missing components, and the use of unlabeled training data (Granger et al., 2000).

Conventional approaches to, and challenges of, radar type identification in radar ESM systems are reviewed in the next section. A system-level overview of our novel neural network recognition system is provided in Section 3. A radar pulse data set used for proof-of-concept simulations is presented in Section 4. The three main components that form a specific implementation of the recognition system are described in 5 An ARTMAP neural network for classification, 6 Pattern clustering, 7 Sequential evidence accumulation. In Section 5, the fuzzy ARTMAP neural network is applied to the classification of pulses according to radar type from functional, type-specific parameters. Then, aspects of this network for dealing with incomplete radar data are proposed and tested. In Section 6, a module for clustering incoming pulses by emitter based on environment-specific parameters is described. In Section 7, a module that accumulates evidence from fuzzy ARTMAP responses corresponding to the tracked emitters is proposed. Finally, these three components are connected, and global simulation results using this particular realization of the entire recognition system are presented and discussed.

Section snippets

Overview

The basic functionality of current radar ESM approaches can be decomposed into three tasks: reception of radar signals; grouping of pulses according to emitter; and identification of corresponding radar types.

Radar signals are passively intercepted by the receiver portion of the ESM system. In typical theaters of operation, intercepted signals are a mixture of electromagnetic pulses transmitted from, typically, several sources. Simultaneous illumination by these sources causes overlap and

Adaptive learning and ESM

In this section, a new approach is described for radar type recognition. When collection platforms are brought into a theater of operations prior to military interventions, data from radars of interest can be collected and analyzed. Collection platforms include in-theater tactical aircraft and ships, unmanned aerial vehicles, and stand-off assets such as electronic warfare (EW) aircraft. Data collected prior to, or during, the conflict are analyzed either on-line (e.g. on a ship) or off-line

Radar pulse data

The data set used for the computer simulation contains approximately 100,000 consecutive radar pulses gathered over 16 s by the Defense Research Establishment Ottawa during a field trial. After the trial, an ESM analyst manually separated trains of pulses coming from different emitters. Each pulse was then tagged with two labels: a radar type number and a mode number. Since ESM trials are complex and never totally controlled, not all pulses could be tagged and a sizable residue was obtained.

An ARTMAP neural network for classification

An enhanced ARTMAP neural network is used to classify incoming radar pulses according to radar type from parameters in the What data stream. ARTMAP refers to a family of neural network architectures capable of fast, stable, on-line, unsupervised or supervised, incremental learning, classification, and prediction (Carpenter et al., 1992, Chandra et al., 1988). ARTMAP networks have several attractive features for applications such as electronic support measures (ESM). Because they can perform

Pattern clustering

The objective of pattern clustering in the What-and-Where recognition architecture shown in Fig. 2. is to group patterns from the Where data stream into tracks. Impinging signals contain information about emitter status, which may change in time. Desirable features for such on-line clustering include the ability to initialize new tracks whenever new emitters are detected, to adjust tracks in response to emitter maneuvers, and to delete tracks as emitters leave or stop transmitting.

Several

Sequential evidence accumulation

Sequential evidence exploits Where information by combining the responses of pattern clustering with neural network classification. In short, the classifier's responses are accumulated according to track, thus offering predictions from multiple views of an emitter.

Conclusions

A novel What-and-Where architecture has been proposed for recognition and tracking of radar emitters for Electronic Support Measures (ESM). This architecture combines a neural network classifier, an on-line clustering algorithm, and an evidence accumulation module. Once trained on samples of data gathered in the field of operation, the neural network classifier can predict the radar type intercepted pulses based on their What parameters. Meanwhile, the clustering algorithm separates these

Acknowledgements

This research was supported in part by the Defense Advanced Research Projects Agency and the Office of Naval Research ONR N00014-95-1-0409 (S.G. and M.A.R.), the National Science Foundation (NSF IRI-97-20333 (S.G.), the Natural Sciences and Engineering Research Council of Canada (E.G.), and the Office of Naval Research ONR N00014-95-1-0657 (S.G.).

References (59)

  • M.R. Anderberg

    Cluster analysis for applications

    (1973)
  • J.A. Anderson et al.

    Radar signal categorization using a neural network

    Proc. IEEE

    (1990)
  • Y. Bar-Shalom et al.

    Estimation and tracking: principles, techniques and software

    (1993)
  • C. Bishop

    Neural networks for pattern recognition

    (1995)
  • S.S. Blackman

    Multiple-target tracking with radar applications

    (1986)
  • G. Bradski et al.

    STORE working memory networks for storage and recall of arbitrary temporal sequences

    Biological Cybernetics

    (1994)
  • J.P.R. Browne et al.

    Electronic warfare

    (1998)
  • G.A. Carpenter et al.

    Fuzzy ARTMAP: a neural network architecture for incremental supervised learning of analog multidimensional maps

    IEEE Trans. on Neural Networks

    (1992)
  • G.A. Carpenter et al.

    ART-EMAP: a neural network architecture for object recognition by evidence accumulation

    IEEE Trans. on Neural Networks

    (1995)
  • Carpenter, G. A., Rubin, M. A., & Streilein, W. W. (1997a). ARTMAP-FD: familiarity discrimination applied to radar...
  • G.A. Carpenter et al.

    Threshold determination for ARTMAP-FD familiarity discrimination

  • Chandra, V., Jyotishi, B. K., & Bajpai, R. C. (1988). Some new algorithms for ESM data processing. Proc. 20th...
  • S. Chen et al.

    Orthogonal least squares learning algorithm for radial basis function networks

    IEEE Trans. on Neural Networks

    (1991)
  • T.M. Cover et al.

    Nearest neighbor patterns classification

    IEEE Trans. on Information Theory

    (1967)
  • C.L. Davies et al.

    Automatic processing for ESM

    Proc. IEE

    (1982)
  • A. Dermiriz et al.

    Semi-supervised clustering using genetic algorithms

  • R.C. Dubes et al.

    Algorithms for clustering data

    (1988)
  • R.O. Duda et al.

    Pattern classification and scene analysis

    (1973)
  • K. Fukunaga

    Introduction to statistical pattern recognition

    (1990)
  • Cited by (0)

    View full text