Anomalies in time series data are a considerable concern, as they can indicate potential risks or issues. For instance, in the case of earthquakes, anomalies can manifest as irregular seismic signals that show sudden spikes or drops in data. Similarly, in the financial domain, anomalies were evident in events such as the Wall Street Crash of 1929. In engineering, anomalies can be observed in signals with spikes, which may indicate ultrasound reflections off walls or individuals. These examples all raise the question of how to identify anomalies in new signals within a set of normal data.
The core issue lies in determining whether a new signal is anomalous within a dataset of normal signals. This problem is distinct from the task of detecting anomalies within a single signal, which is a well-studied challenge. When faced with a new signal, the goal is to assess whether it differs significantly from the signals considered “normal” in the dataset. This distinction underscores the importance of developing effective methods for detecting anomalies in time series data to mitigate potential risks and ensure accurate analysis and decision-making.