The passive optical network (PON) is widely used in optical fiber communication thanks to its low cost and low resource consumption. However, the passiveness brings about a critical problem that it requires manual work to identify the topology structure, which is costly and prone to bringing noise to the topology logs. In this paper, we provide a base solution firstly introducing neural networks for such problems, and based on that solution we propose a complete methodology (PT-Predictor) for predicting PON topology through representation learning on its optical power data. Specifically, we design useful model ensembles (GCE-Scorer) to extract the features of optical power with noise-tolerant training techniques integrated. We further implement a data-based aggregation algorithm (MaxMeanVoter) and a novel Transformer-based voter (TransVoter) to predict the topology. Compared with previous model-free methods, PT-Predictor is able to improve prediction accuracy by 23.1% in scenarios where data provided by telecom operators is sufficient, and by 14.8% in scenarios where data is temporarily insufficient. Besides, we identify a class of scenarios where PON topology does not follow a strict tree structure, and thus topology prediction cannot be effectively performed by relying on optical power data alone, which will be studied in our future work.

中文翻译：

---

电信运营商无源光网络拓扑预测中的时间序列表示学习

无源光网络(PON)由于其低成本和低资源消耗而被广泛用于光纤通信。然而，这种被动性带来了一个关键问题，即需要人工识别拓扑结构，成本高且容易给拓扑日志带来噪声。在本文中，我们首先针对此类问题提供了一个基本解决方案，并基于该解决方案提出了一种完整的方法（PT-Predictor），用于通过光功率数据的表示学习来预测 PON 拓扑结构。具体来说，我们设计了有用的模型集成（GCE-Scorer）来提取光功率的特征，并集成了抗噪训练技术。我们进一步实施了一种基于数据的聚合算法 (MaxMeanVoter) 和一种新颖的基于 Transformer 的投票器 (TransVoter) 来预测拓扑。与以往的无模型方法相比，PT-Predictor在电信运营商提供的数据充足的场景下，预测准确率提升了23.1%，在数据暂时不足的场景下，预测准确率提升了14.8%。此外，我们确定了一类 PON 拓扑不遵循严格树结构的场景，因此仅依靠光功率数据无法有效地进行拓扑预测，这将在我们未来的工作中进行研究。8% 在数据暂时不足的场景。此外，我们确定了一类 PON 拓扑不遵循严格树结构的场景，因此仅依靠光功率数据无法有效地进行拓扑预测，这将在我们未来的工作中进行研究。8% 在数据暂时不足的场景。此外，我们确定了一类 PON 拓扑不遵循严格树结构的场景，因此仅依靠光功率数据无法有效地进行拓扑预测，这将在我们未来的工作中进行研究。