Water levels in rivers are measured by various devices installed mostly in remote locations along the rivers, and the collected data are then transmitted via telemetry systems to a data centre for further analysis and utilisation, including producing early warnings for risk situations. So, the data quality is essential. However, the devices in the telemetry station may malfunction and cause errors in the data, which can result in false alarms or missed true alarms. Finding these errors requires experienced humans with specialised knowledge, which is very time-consuming and also inconsistent. Thus, there is a need to develop an automated approach. In this paper, we firstly investigated the applicability of Deep Reinforcement Learning (DRL). The testing results show that whilst they are more accurate than some other machine learning models, particularly in identifying unknown anomalies, they lacked consistency. Therefore, we proposed an ensemble approach that combines DRL models to improve consistency and also accuracy. Compared with other models, including Multilayer Perceptrons (MLP) and Long Short-Term Memory (LSTM), our ensemble models are not only more accurate in most cases, but more importantly, more reliable.

中文翻译：

---

用于检测遥测水位数据异常的深度强化学习集成

河流中的水位由主要安装在沿河偏远地区的各种设备测量，然后通过遥测系统将收集的数据传输到数据中心以进行进一步分析和利用，包括对风险情况进行预警。因此，数据质量至关重要。但是，遥测站中的设备可能会发生故障并导致数据错误，从而导致误报或漏报。发现这些错误需要有专业知识的有经验的人，这非常耗时且不一致。因此，需要开发一种自动化方法。在本文中，我们首先研究了深度强化学习（DRL）的适用性。测试结果表明，虽然它们比其他一些机器学习模型更准确，特别是在识别未知异常方面，它们缺乏一致性。因此，我们提出了一种结合 DRL 模型以提高一致性和准确性的集成方法。与其他模型相比，包括多层感知器 (MLP) 和长短期记忆 (LSTM)，我们的集成模型不仅在大多数情况下更准确，而且更重要的是，更可靠。