Synthesis of a Maneuvering-Target Angular Tracking System Based on an Extended Kalman Filter and a Regime-Balanced Multilayer Perceptron Neural Network under Dynamic-Model Mismatch

Trinh Thi Minh¹, Vu Quang Luong^2*, Nguyen Huy Hoang³, Pham Dao Minh Duc³

Abstract

This paper synthesizes a maneuvering-target angular tracking system using an extended Kalman filter combined with a regime-balanced multilayer perceptron neural network. Angular tracking accuracy degrades when a maneuvering target departs from the fixed dynamic model embedded in an extended Kalman filter (EKF), even when the measurement noise remains nominal. This paper proposes a regime-balanced multilayer-perceptron residual-compensated EKF (RB-MLP-RC-EKF) for piecewise accelerating and reversing targets under Singer-model mismatch. A compact MLP estimates only the unmodeled forcing in the transverse-acceleration channel from a causal window of innovation and estimated-state features. Training windows are balanced across mild, positive-acceleration, reversal, and recovery regimes, while the learned forcing is bounded and low-pass filtered before being injected into the EKF prediction. In the primary simulation, the proposed method reduces angular tracking RMSE by 50.8%, 45.1%, and 6.35% relative to EKF, adaptive-Q EKF, and conventionally trained MLP-RC-EKF, respectively. Across 40 Monte Carlo trials, the corresponding median reductions are 47.0%, 41.2%, and 5.66%, while median LOS-rate estimation RMSE decreases by 49.0% relative to EKF. The results support localized residual learning as a lightweight approach to dynamic-model mismatch.

Keywords:

Angular tracking; dynamic-model mismatch; extended Kalman filter; maneuvering target; multilayer perceptron; residual compensation; Singer model