Abstract Live-line maintenance is a high risk activity. Hence, lineworkers require effective and safe training. Virtual Reality Training Systems (VRTS) provide an affordable and safe alternative for training in such high risk environments. However, their effectiveness relies mainly on having meaningful activities for supporting learning and on their ability to detect untrained students. This study builds a student model based on Learning Analytics (LA), using data collected from 1399 students that used a VRTS for the maintenance training of lineworkers in 329 courses carried out from 2008 to 2016. By employing several classifiers, the model allows discriminating between trained and untrained students in different maneuvers using three minimum evaluation proficiency scores. Using the best classifier, a Feature Importance Analysis is carried out to understand the impact of the variables regarding the trainees’ final performances. The model also involves the exploration of the trainees’ trace data through a visualization tool to pose non-observable behavioral variables related to displayed errors. The results show that the model can discriminate between trained and untrained students, the Random Forest algorithm standing out. The feature importance analysis revealed that the most relevant features regarding the trainees’ final performance were profile and course variables along with specific maneuver steps. Finally, using the visual tool, and with human expert aid, several error patterns in trace data associated with misconceptions and confusion were identified. In the light of these, LA enables disassembling the data jigsaw quandary from VRTS to enhance the human-in-the-loop evaluation.

中文翻译：

---

虚拟现实培训系统中学生建模的学习分析：Lineworkers 案例

摘要 带电维护是一项高风险的活动。因此，线路工人需要有效和安全的培训。虚拟现实培训系统 (VRTS) 为在此类高风险环境中进行培训提供了一种经济实惠且安全的替代方案。但是，它们的有效性主要取决于开展有意义的活动来支持学习以及发现未受过培训的学生的能力。本研究建立了一个基于学习分析 (LA) 的学生模型，使用从 2008 年至 2016 年开展的 329 门课程中使用 VRTS 对线路工人进行维护培训的 1399 名学生收集的数据。通过使用多个分类器，该模型允许区分使用三个最低评估熟练度分数，对训练有素和未经训练的学生进行不同的操作。使用最好的分类器，进行特征重要性分析以了解变量对受训者最终表现的影响。该模型还涉及通过可视化工具探索受训者的跟踪数据，以提出与显示错误相关的不可观察的行为变量。结果表明，该模型可以区分受过训练和未受过训练的学生，其中随机森林算法脱颖而出。特征重要性分析表明，与受训者最终表现最相关的特征是轮廓和路线变量以及特定的机动步骤。最后，使用可视化工具，并在人类专家的帮助下，确定了跟踪数据中与误解和混淆相关的几种错误模式。鉴于这些，