Abstract:To satisfy the modern track traffic's demand for maintaining high precision and continuity of navigation under complex environmental conditions, and to address the issue of positioning drift caused by data outages in nonlinear dynamic integrated navigation systems, this paper proposes a novel adaptive estimation method for observation errors in nonlinear dynamic integrated navigation systems. This method is based on a Probabilistic Time Series Transformer model, aiming to resolve the aforementioned issues. By introducing self-learning capabilities through the Probabilistic Time Series Transformer, the method adaptively adjusts the impact of state prediction and observation information outages on the dynamic navigation system. The Probabilistic Time Series Transformer is composed of a dual-loop system of a generative model and an inference model, combined with LSTM network to tackle the challenges of multivariate time series modeling. The integrated navigation system based on the Probabilistic Time Series Transformer optimizes the error compensation mechanism by establishing a relationship between the current Kalman filter gain and the optimal estimation error, thereby improving the accuracy and stability of the nonlinear navigation system. Experimental results demonstrate that the proposed method not only effectively controls the impact of GNSS outages on the nonlinear navigation system but also accurately estimates and compensates for observation model system errors. The average positioning error in various complex scenarios is less than 10m. The suppression of positioning drift in the observation model is better than that of other filtering methods.