Fourteen selected parameters of the reactor response were used to train a multi-connected Support Vector Machines (SVM) model. With proper optimization, the SVM model demonstrated accurate prediction for hot-leg break LOCA, cold-leg break LOCA, and SGTR, and their break sizes. The prediction accuracy is higher in the early phase (<~15 s) of the accidents than the later phase of the accident, implying that accident-discerning information is contained in the early response of the accident. Such characteristics of machine-learning aided accident characterization are complementary to the typical man-made decisions, which often require a considerable progression of accidents before one can tell the accident characteristics (i.e., location and size). The presented study also investigated the robustness of the model by quantifying the prediction accuracy in case some sensors (reactor water level and pressure) of the instrument system output wrong signals during the accidents. It was found that the model is not equally built upon all used parameters, demonstrating biased dependency on specific parameters. Hence, malfunction of some sensors (i.e., reactor water level) noticeably increases prediction errors while wrong signals of other sensors (i.e., reactor pressure) have a limited impact on the prediction accuracy.

反应器响应的 14 个选定参数用于训练多连接支持向量机 (SVM) 模型。通过适当的优化，SVM 模型展示了对热腿断裂 LOCA、冷腿断裂 LOCA 和 SGTR 及其断裂尺寸的准确预测。事故早期（<~15 s）的预测精度高于事故后期，表明事故早期响应中包含事故识别信息。机器学习辅助事故表征的这些特征与典型的人为决策相辅相成，通常需要大量的事故进展才能判断事故特征（即位置和规模）。本研究还通过量化预测精度来研究模型的稳健性，以防万一仪表系统的某些传感器（反应堆水位和压力）在事故期间输出错误信号。发现该模型并非平等地建立在所有使用的参数上，这表明对特定参数的依赖性有偏差。因此，一些传感器的故障（即反应堆水位）显着增加了预测误差，而其他传感器的错误信号（即反应堆压力）对预测精度的影响有限。