Analysis of the effects of bearings-only sensors on the performance of the neural extended Kalman filter tracking system

The neural extended Kalman filter (NEKF) has proven to be a quality maneuver target tracking system when the sensors provide a fully observable measurement, such as a radar's range-bearing measurement or a position report. As with any state estimation technique, the NEKF requires observability in order to estimate the target track states. Observability is needed as well to train the weights of the neural network, since the neural network training paradigm is coupled to the target states. Passive sensor systems, such as electronic surveillance measures and passive sonar arrays, provide an angle-only measurement. Such bearings-only measurements make the tracking system an unobservable system. For a Kalman filter estimator, this will result in the eigenvalues of the error covariance matrix to grow without bound. For the NEKF, since both the target state and the weights of the neural network are affected by the lack of observability, the results could be more pronounced. In this paper, the application of the NEKF in bearings-only tracking problems is analyzed to determine the effects on performance. The analyzed cases look at a single sensor platform in four important scenarios: a stationary platform and straight-line target, a stationary platform and a maneuvering target, a maneuvering platform and a straight-line target, and a maneuvering platform and a maneuvering target.