基于雷达航迹序列的无人机/飞鸟类型识别是空中安全监管的关键。在实际应用中，随着航迹数据的不断到来，需要准确且快速地实现无人机/飞鸟分类。本文提出“多特征快速综合、多似然序贯决策、多因子长时精分”的短-中-长多尺度动态分类机制。在多特征快速综合中，按照相同的物理含义将输入航迹向量划分为位置类（代表目标态势占位）、速度类（代表目标态势变化）、辐射类（代表目标材质结构），分别导入短时多头一维卷积神经网络并采用通道注意力机制进行多类别特征综合，从而实时度量目标属性置信；在多似然序贯决策中，统计目标属性置信的似然分布，设计具有多级化的长短时置信似然决策逻辑，从而在更长时间跨度上实现目标属性的综合推理；在多因子长时精分中，提出速率/航向角变化、速率/航向角趋势等多因子度量，进而采用随机森林对难分样本进行长时多特征精确分类。本文算法在实际雷达航迹数据中分类准确率、虚警率和漏检率三项指标均优于现有算法，验证了所提算法的有效性。

The identification of unmanned aerial vehicles (UAVs) and flying birds based on radar track sequences is crucial for air safety supervision. In practical applications, with the continuous arrival of track data, it is essential to achieve accurate and rapid classification of UAVs/flying birds. A short-medium-long multi-scale dynamic classification mechanism characterized by “multi-feature rapid synthesis, multi-likelihood sequential decision, and multi-factor long-term precise classification” is proposed. In multi-feature rapid synthesis, the input track vector is categorized based on the same physical meaning: position-related features (representing the target’s situation occupancy), velocity-related features (representing target’s situation changes), and radiation-related features (representing target’s material structure). These features are then fed into a short-term multi-head one-dimensional convolutional neural network (1D-CNN) and synthesized using a channel attention mechanism, enabling real-time measurement of target attribute confidence. In multi-likelihood sequential decision-making, the likelihood distribution of target attribute confidence is statistically analyzed, and a multi-level decision logic incorporating both short-term and long-term confi-dence likelihood is designed to achieve comprehensive inference of target attributes over a longer time span. In multi-factor long-term precise classification, multiple factors measurements including velocity/heading angle changes and velocity/head angle trends are proposed and the random forest algorithm is then employed to accurately classify hard-to-distinguish samples with multiple features over a long period of time. The proposed algorithm outperforms the existing algorithms in terms of classification accuracy, false positive rate, and false negative rate in real radar track data, verifying the effectiveness of the proposed algorithm.