Deep learning-driven multi-sensor observation fusion filtering method

Multi-sensor fusion estimation is of great significance in the field of information fusion. However, conventional estimation methods often suffer from performance limitations due to strong reliance on model information and fixed weight assignment, making it difficult to achieve unified representation of heterogeneous observations and dynamic evaluation of sensor reliability. To address the model mismatch problem inherent in traditional model-driven multi-sensor observation fusion filtering, this paper proposes a deep learning-based observation fusion filtering algorithm that integrates model-driven observation fusion filtering theory with a data-driven self-attention neural network. The algorithm consists of an observation feature extraction module, a dynamic weight assignment module, and a noise parameter estimation network. By encoding the temporal features of multi-source heterogeneous observation sequences, a unified representation across sensors is realized; the attention mechanism is leveraged to dynamically learn the fusion weight of each sensor; meanwhile, the system noise covariances are estimated in a data-driven manner, and the final state estimate is obtained through a weighted observation fusion framework. Simulation experiments demonstrate that the proposed algorithm outperforms existing data-driven methods, such as KalmanNet and IFNet, as well as covariance intersection fusion and arithmetic average fusion algorithms in terms of estimation precision. Furthermore, the filtering time consumption analysis indicates that the algorithm satisfies real-time requirements.