Real-time Prediction of UAV Pilot Operational Errors and Adaptive Threshold Optimization under Temporal Models

Operational errors by UAV pilots have become a critical bottleneck constraining industry safety. Existing prediction methods mostly adopt unified models while neglecting individual pilot capability differences. This study proposes a real-time prediction framework integrating pilot adaptability features with operational temporal data. A dual-channel network architecture separately extracts temporal and adaptability features, utilizing cross-modal attention for deep fusion. An adaptive threshold algorithm dynamically adjusts warning parameters based on pilot capability levels. Experiments on 665 pilots demonstrate 89.7% accuracy, 0.928 AUC, and 3.2-second average warning time. Complex reaction, control function, and motion stability are identified as key predictive indicators, contributing 28.3%, 25.6%, and 19.5% respectively. The adaptive threshold strategy reduces false negative rates for low-adaptability pilots by 50.4% and false positive rates for high-adaptability pilots by 64.6%. Leave-one-pilot-out validation confirms generalization capability (86.3% accuracy). These findings support transforming UAV safety management from post-incident analysis to proactive prevention and from unified standards to personalized adaptation.无人机（UAV）飞行员操作失误已成为制约行业安全的关键瓶颈。现有预测方法多采用统一模型，却忽略了个体飞行员的能力差异。本研究提出一种融合飞行员适应性特征与操作时序数据的实时预测框架。该框架采用双通道网络架构分别提取时序特征与适应性特征，并利用跨模态注意力机制实现深度融合。此外，研究设计了自适应阈值算法，可基于飞行员能力水平动态调整预警参数。针对665名飞行员的实验结果表明，该框架准确率达89.7%，曲线下面积（AUC）为0.928，平均预警时间为3.2秒。研究识别出复杂反应能力、操控功能与运动稳定性为核心预测指标，其特征贡献占比分别为28.3%、25.6%与19.5%。自适应阈值策略可使低适应性飞行员的假阴性率降低50.4%，高适应性飞行员的假阳性率降低64.6%。留一飞行员交叉验证（Leave-one-pilot-out validation）结果证实模型具备良好泛化能力，其准确率达86.3%。上述研究成果推动无人机安全管理从事后事故分析向事前主动预防转型，从统一标准化管理向个性化适配管理演进。