基于雷达测量的用于炮位侦察的Transformer网络

Transformer network for fire locating from radar measurements

依赖雷达测量数据的炮位侦察在遇到炮弹低射角时面临极大的挑战。雷达观测数据弧段短、测量误差大,且具有数据批量小、非线性、不完整等特点,炮位外推困难。广泛用于自然语言处理领域的Transformer网络具有长距离依赖、全自注意力机制等特点,在长距离序列建模方面具有较大优势。该文提出了一种时间戳编码的方法,首次应用于Transformer网络来表征空气动力目标的飞行轨迹,并外推炮位位置。同时建立了大规模雷达侦测仿真数据集用于网络训练,并与传统炮位侦察算法,如卡尔曼滤波类算法、长-短周期记忆网络等进行了对照实验。结果表明:Transformer网络在预测炮位时收敛性能好,圆概率误差指标优于其他方法。

The fire locating which relies on radar measurement data is facing great challenge when the launch angles of cannonball are too low.Much barriers,for instance short observable segment from radar,large measurement error on radar elevation,small samples,incomplete segment and high non-linearity between measurement and fire location,make it more difficult.The Transformer network,which has been widely used in the field of natural language processing(NLP),has been improved by a timestamp encoding method to capture the flight trajectory of aerodynamic targets and estimate the fire location in this paper.Benefit from its long-distance dependence and self-attention mechanism,it works better on fire locating from radar measurements.A large-scale radar measurement dataset for cannonball flying has been created as performance benchmark for training,validation and testing.Compared with the classical fire locating methods,such as Kalman filter-like methods,long short-term memory(LSTM)based methods,the proposed Transformer based method is superior to the formers on the circular error probability(CEP)metric.