A Novel Data Fusion Method of Redundant Gyroscope Systems With Nonuniform Accuracy

In the scenario of sensor performance degradation, the predictive management of the inertial navigation system needs to maintain reliable accuracy. However, traditional methods are difficult to effectively handle the problem of data fusion of gyroscope with heterogeneous accuracy and dynamic degradation. This paper proposes a new data fusion method for redundant gyroscope systems based on reinforcement learning. This method adopts the soft actor-critic (SAC) reinforcement learning algorithm, breaking through the limitations of traditional models that rely on pre-defined noise models or system models. It achieves adaptive real-time optimization of fusion weights through dynamic interaction with the environment. Based on the principle of information non-increase, a three-stage data fusion model is constructed to precisely match the entire life cycle state of the gyroscope from mild degradation, moderate degradation to severe degradation. To solve the problem of objective determination of stage boundaries, a intersection point detection network is designed. Through convolution feature extraction, bidirectional GRU temporal modeling and multi-branch fusion architecture with difference perception, the critical transition point of sensor stability mutation is automatically identified. Experiments are verified on the inertial instrument control platform. The results show that the accuracy loss in the mild degradation stage is only 7.67%. The fusion accuracy in the severe degradation stage is improved by more than 16% compared to the weighted least squares method (WLS) and the fruit fly optimization weighted least squares method (FFO-WLS), and the single-step inference time is as low as 0.82ms. It has the characteristics of high accuracy, strong adaptability and efficient execution, providing active maintenance support for predictive management.