MMD-SCD

This project is part of a CIFRE PhD thesis aimed at developing anomaly detection algorithms for time series.

The repository contains the implementation and comparison of four CUSUM-based algorithms. Our main contribution — the HKCUSUM algorithm (Histogram-based KCUSUM) — significantly reduces the false alarm rate, whilst maintaining a detection delay comparable to that of state-of-the-art methods, such as KCUSUM (Kernel-based CUSUM) and TSCUSUM (Two-sided CUSUM).

This study has led to a paper accepted at the EUSIPCO conference, to take place in september 2026. The publication of this repository aims to guarantee the reproducibility of experiments and promote research in the field of anomaly detection.