基於 Transformer 的自主載具定位系統之故障檢測與排除

Abstract

本研究延續本實驗室前期的錯誤偵測與排除架構，使用雙塔結構的 Transformer 對多感測器的數據進行資訊整合與特徵學習，改善傳統方法在多重故障辨識上的不足。在訓練模型的過程中，本研究分析並模擬感測器的錯誤訊號，結合開源資料集進行自監督學習。此外，根據道路設計規範與車輛自主定位需求，文中明確界定定位失效的判斷標準，用來評估模型在定位精度與系統安全性的效益。實驗結果顯示，本模型在單一感測器故障情境下可排除 80 % 以上的失效案例；在多感測器同時故障的情境下，亦能成功偵測超過 70 % 的錯誤訊號，展現出良好的穩定性與偵測能力。此外，本模型可於無顯示卡加速的環境下運行，適用於資源受限的車載系統部署。

This study builds upon our laboratory’s previous fault detection and exclusion (FDE) framework by employing a dual-tower Transformer architecture to perform information integration and feature learning from multi-sensor data, addressing the limitations of traditional methods in identifying multiple simultaneous faults. During model training, common sensor faults are analyzed and simulated, and self-supervised learning is conducted using an open-source dataset. In addition, based on road design standards and the requirements of autonomous vehicle localization, this study defines clear criteria for localization failure, which serve as the basis for evaluating the model’s effectiveness in improving localization accuracy and system safety. Experimental results show that our model can excluding over 80 % of localization failures caused by single-sensor faults. In scenarios involving multiple simultaneous sensor failures, the model successfully detects more than 70 % of faulty signals, demonstrating strong stability and detection capability. Moreover, the model operates efficiently without GPU acceleration, making it suitable for deployment in resource-constrained in-vehicle systems.