The extremely low flux of ultra-high energy cosmic rays (UHECR) makes their direct observation by orbital experiments practically impossible. For this reason all current and planned UHECR experiments detect cosmic rays indirectly by observing the extensive air showers (EAS) initiated by cosmic ray particles in the atmosphere. The world largest statistics of the ultra-high energy EAS events is recorded by the networks of surface stations. In this paper we consider a novel approach for reconstruction of the arrival direction of the primary particle based on the deep convolutional neural network. The latter is using raw time-resolved signals of the set of the adjacent trigger stations as an input. The Telescope Array (TA) Surface Detector (SD) is an array of 507 stations, each containing two layers plastic scintillator with an area of 3 m². The training of the model is performed with the Monte-Carlo dataset. It is shown that within the Monte-Carlo simulations, the new approach yields better resolution than the traditional reconstruction method based on the fitting of the EAS front. The details of the network architecture and its optimization for this particular task are discussed.

超高能宇宙射线 (UHECR) 的极低通量使得通过轨道实验对其进行直接观测几乎是不可能的。出于这个原因，所有当前和计划中的 UHECR 实验都是通过观察由大气中的宇宙射线粒子引发的广泛的空气簇射 (EAS) 来间接探测宇宙射线的。世界上最大的超高能 EAS 事件统计数据由地面站网络记录。在本文中，我们考虑了一种基于深度卷积神经网络重建初级粒子到达方向的新方法。后者使用一组相邻触发站的原始时间分辨信号作为输入。望远镜阵列 (TA) 表面探测器 (SD) 是由 507 个站点组成的阵列，每个站点包含面积为 3 m 的两层塑料闪烁体² . 模型的训练是使用 Monte-Carlo 数据集进行的。结果表明，在蒙特卡罗模拟中，新方法比基于 EAS 前沿拟合的传统重建方法产生更好的分辨率。讨论了网络架构的细节及其针对此特定任务的优化。