Unsupervised machine learning methods present a promising approach for detecting fragments produced from meteors and bolides as distinct signatures within Doppler weather radar data. A method combining principal component analysis (PCA), t-distributed statistical neighbor embedding (t-SNE), and data point pruning based on the nearest neighbor algorithm is introduced as a process to detect outlier meteor signatures from terrestrial weather signatures using the national NOAA WSR-88D Doppler radar network. This method is applied against unlabeled data from four weather radar sites during two bolide events: the KFWS radar for the Ash Creek bolide and the KDAX, KRGX, and KBBX radars for the Sutter’s Mill bolide. The combined algorithm results in an accuracy rate of 99.7% and can classify the data in <8 min for a 121,000 return sized data set. However, the classifier’s recall and precision rates remained low due to difficulties in correctly classifying true-positive meteorite fall events. This method enables the expedited detection of materials from bolides and meteors that fall within the national radar network, leading to the faster confirmation of meteorite fall events and subsequent dispatch of recovery teams.

中文翻译：

---

使用人工智能/机器学习在多普勒天气雷达数据中检测 Bolide 碎片

无监督机器学习方法提供了一种很有前景的方法，可以检测流星和火流星产生的碎片作为多普勒天气雷达数据中的不同特征。介绍了一种结合主成分分析 (PCA)、t 分布统计邻域嵌入 (t-SNE) 和基于最近邻算法的数据点剪枝的方法，作为使用国家 NOAA 从陆地天气特征中检测异常流星特征的过程。 WSR-88D 多普勒雷达网络。该方法适用于两次火炬事件期间来自四个天气雷达站点的未标记数据：用于 Ash Creek 火炬的 KFWS 雷达和用于 Sutter's Mill 火炬的 KDAX、KRGX 和 KBBX 雷达。组合算法的准确率为 99.7%，并且可以在 <8 分钟内对返回大小为 121,000 的数据集的数据进行分类。然而，由于难以正确分类真正的陨石坠落事件，分类器的召回率和准确率仍然很低。这种方法可以加快检测落入国家雷达网络内的火流星和流星的材料，从而更快地确认陨石坠落事件并随后派遣救援队。