Internet of Things (IoT) is a rising topic and finds the way into various domains, changing the way we interact with technology and gather data. For instance, in the healthcare sector, IoT-enabled devices such as wearable fitness trackers and remote patient monitoring systems have transformed the landscape by providing real-time health data, enhancing patient care, and enabling proactive interventions.

While networks of IoT devices represent complex structures, they produce significant amounts of time series data, which may be prone to anomalies, unexpected patterns or deviations from the norm within the data. Anomalies in IoT data can cause potential issues, security threats, or operational irregularities, making anomaly detection a critical aspect of ensuring the reliability and security of IoT ecosystems. Nevertheless, the inherent resource constraints of IoT devices raise concerns regarding the precision of their emitted observations. This impreciseness in measurements can be characterized as noise.

In statistics or machine learning, this noise is referred to as data (aleatoric) uncertainty. In addition, the model itself introduces another layer of uncertainty, known as model (epistemic) uncertainty, which becomes particularly interesting in anomaly detection scenarios where differing anomalies from normal patterns requires a nuanced understanding of the model's uncertainty.
Addressing both types of uncertainties is necessary, as neglecting either could overlook a significant aspect of overall uncertainty and impair robustness of anomaly detection systems.

Current approaches to uncertainty in time series anomaly detection mostly focus on quantifying epistemic uncertainty and ignore data dependent uncertainty. However, consideration of noise in data is important as it may have the potential to lead to more robust detection of anomalies and a better capability of distinguishing between real anomalies and anomalous patterns provoked by noise.

In this work, we propose an extended version of a LSTM Autoencoder for anomaly detection which is able to quantify both aleatoric and epistemic uncertainty in time series separately.
We show that implemented mechanisms are effectively recognising different levels of noise. Further, we investigate the impact of anomalies on uncertainty. Finally, we evaluate effectiveness of the approach by evaluating it on real-world datasets.