An Efficient Ensemble Mamba Architecture for IMU Data Processing in the FAST Feed Cabin

The high-precision pose measurement of the Five-hundred-meter Aperture Spherical radio Telescope (FAST) feed cabin is crucial for its scientific output. However, it is often limited by the computational intensity and parameter sensitivity of the Precise Strapdown Inertial Navigation System (PSINS). To address this, we propose an ensemble architecture that achieves a superior balance between accuracy and efficiency for real-time Inertial Measurement Unit (IMU) data processing. This architecture is based on a lightweight Gradient Boosting Decision Tree (GBDT) model, combined with a multi-scale pyramid Mamba and the original Mamba. The GBDT classifies IMU data into high-speed and low-speed states. The pyramid Mamba, with its strong capacity for capturing detailed features, processes the high-speed data to enhance computational accuracy, while the original Mamba rapidly processes the low-speed data. In the model improvement analysis, evaluation on the 133 GB FAST dataset demonstrated that our ensemble model achieved progress compared to the original Mamba, with RMSE and MAE reduced by over 30% and 66% respectively. In the effectiveness analysis, although the traditional PSINS method showed only an 8% gap in RMSE, our approach achieved a remarkable 57% improvement in processing efficiency. In a comparative study of deep learning models, the Ensemble Mamba achieved a superior RMSE of 0.003222m, outperforming both LSTM and Transformer, while its inference time of 10.1569 ms meets the real-time requirements. Furthermore, a generalization capability analysis on a vehicle dataset demonstrated its robust cross-domain performance, yielding the lowest RMSE of 0.044m and confirming its universal modeling capability for sequences with state transitions. This work demonstrates the potential of selective state space models in high-frequency, long-sequence scientific data processing.