CRAPA - A measurement-preprocessing algorithm based on correntropy and attention mechanism

In complex environments, non-Gaussian noise leads to the accumulation of state estimation deviations in traditional Kalman filter (KF), severely degrading the robustness of target tracking systems. Although existing methods optimize filtering performance through noise statistical modeling or robust loss functions, they are still limited by issues such as fixed parameters, model mismatch, and high computational complexity. This paper proposes a measurement preprocessing algorithm based on correntropy and residual attention mechanism (CRAPA). By calculating dynamic weights through the Gaussian kernel correntropy of historical residuals and current residuals within a sliding window, it adaptively identifies and suppresses abnormal measurements. CRAPA adopts modular architecture, combining feedforward suppression mechanisms to block noise propagation. While retaining the optimality of the KF framework, it achieves efficient abnormal detection with relatively low complexity. Simulation experiments show that under Gaussian-mixture noise and Gaussian impulse hybrid noise scenarios, CRAPA significantly reduces the mean square error of KF. Its dynamic window mechanism responds faster than noise statistical modeling algorithms, and its tracking accuracy is comparable to that of Huber-KF, with stronger parameter adaptability. This validates the robustness and tracking performance of CRAPA in non-Gaussian noise, especially in impulse noise environments, providing theoretical support for subsequent extensions to maneuvering target scenarios and deep learning integration.