The thermal protection system (TPS) for spacecraft is easily damaged by various thermal and mechanical loads, which adversely affects the thermal protection performance of the system. TPS damages diagnosis is one of the most complex and challenging problems for spacecraft structural reliability. The embedded distributed optical fiber sensor can directly reflect the physical field distribution of the structure and then evaluate its health status. One of the challenges of this technique is to accurately extract signal characteristics and reconstruct damage state information. A quantile random forest and self-organizing map (SOM) neural-network-based two-step damage diagnosis framework for thermal protection systems is investigated in this Paper. In this Paper, the combination of physical interpretation and data driving is used to analyze the strain anomaly of the TPS specimen and obtain the damage diagnosis results, including the location, influence area, and categories of damage. First, the abnormal distribution of strain values caused by different types of damage is studied by numerical simulation. Then, the outliers of the experimental strain distribution data are detected by using quantile random forest. A signal features vector is extracted from each signal segment, and the SOM neural network is trained to classify damage. Using the trained model to classify the test set, the accuracy of damage classification is 100%. The experiment result shows great potential for the proposed approach is not only possible to detect the presence of damage, but it is also able to locate and estimate the extent of the damage, which helps one make an appropriate decision on the diagnosis.

航天器的热保护系统（TPS）容易受到各种热负荷和机械负荷的损害，这会对系统的热保护性能产生不利影响。TPS损伤诊断是航天器结构可靠性最复杂，最具挑战性的问题之一。嵌入式分布式光纤传感器可以直接反映结构的物理场分布，然后评估其健康状况。该技术的挑战之一是准确地提取信号特征并重建损伤状态信息。本文研究了基于分位数随机森林和自组织图（SOM）神经网络的热防护系统两步损伤诊断框架。在本文中，物理解释和数据驱动相结合，用于分析TPS试样的应变异常，并获得损伤诊断结果，包括位置，影响区域和损伤类别。首先，通过数值模拟研究了由不同类型的损伤引起的应变值的异常分布。然后，通过使用分位数随机森林来检测实验应变分布数据的异常值。从每个信号段中提取信号特征向量，并对SOM神经网络进行训练以对损伤进行分类。使用训练有素的模型对测试集进行分类，损坏分类的准确性为100％。实验结果表明，该方法不仅有可能发现损坏的存在，而且具有很大的潜力，