Performance Analysis of Typical Data Fusion Algorithms for Inertial Measurement Arrays

This paper investigates data fusion for multi-MEMS gyroscope arrays by comparing four methods: numerical averaging, weighted least squares, direct estimation Kalman filtering, and indirect estimation Kalman filtering. The state-estimation characteristics and observability of the two Kalman-filter models are also analyzed. The performance of the four methods is evaluated through controlled simulations, static experiments, dynamic turntable experiments, and array-size analysis. The simulation and static experimental results show that when the IMUs exhibit similar Allan bias-instability characteristics, the four methods yield relatively similar results in terms of bias instability. When the Allan bias-instability characteristics of the IMUs differ, numerical averaging provides poorer performance, whereas the other three methods yield comparable results in terms of bias instability. When different error components are present and exhibit conflicting trends, fixed weighting based on a single statistical indicator may lead to weight mismatch. In the dynamic experiment, the three-axis gyroscope and three-axis accelerometer measurements are separately processed using the same fusion method, and the resulting six-axis fused data are used as inputs to the PSINS inertial navigation system, with the final horizontal position drift adopted as a system-level performance metric. Under the tested dynamic conditions, the direct estimation Kalman filter achieves the smallest final horizontal position drift, followed by the indirect estimation Kalman filter, while both outperform numerical averaging and fixed-weight fusion. The array-size analysis shows that the performance gain gradually diminishes as the number of IMUs increases. The geometric knee point is located at approximately 21 IMUs, and the main performance-transition range is approximately 21–30 IMUs. Under the simulation and experimental conditions considered in this study, this range provides a favorable trade-off among fusion performance, computational burden, and array complexity.