Anomaly Detection via Re-encoding in Autoencoder-Based Compression for Time Series Monitoring Applications

Monitoring systems generate and transmit large volumes of data to processing facilities capable of performing multiple tasks. To reduce transmission and storage costs, data is often compressed, with autoencoders (AEs) emerging as a promising neural network-based approach. This work considers a scenario where the receiver is responsible for both reconstruction and anomaly detection. We propose a novel anomaly detector that operates on the receiver side, approximating the standard anomaly score of conventional AE-based detectors. The proposed approach requires no fine-tuning, as the compression process itself ensures strong detection performance. Moreover, its performance can be further enhanced through a common regularization technique. We validate our method through experiments on two distinct time series datasets: ECG signals and acceleration data.