MAAT: Mamba Adaptive Anomaly Transformer

Authors: Abdellah Zakaria Sellam*, Ilyes Benaissa, Abdelmalik Taleb-Ahmed, Luigi Patrono, Cosimo Distante

Paper

You can check our paper now in Engineering Applications of Artificial Intelligence journal in this link: https://doi.org/10.1016/j.engappai.2025.111685

Abstract

Anomaly detection in time series poses a critical challenge in industrial moni- toring, environmental sensing, and infrastructure reliability, where accurately distinguishing anomalies from complex temporal patterns remains an open problem. While existing methods, such as the Anomaly Transformer lever- aging multi-layer association discrepancy between prior and series distribu- tions and DCdetector employing dual-attention contrastive learning, have advanced the field, critical limitations persist. These include sensitivity to short-term context windows, computational inefficiency, and degraded performance under noisy and non-stationary real-world conditions. To ad- dress these challenges, we present MAAT (Mamba Adaptive Anomaly Trans- former), an enhanced architecture that refines association discrepancy model- ing and reconstruction quality for more robust anomaly detection. Our work introduces two key contributions to the existing Anomaly transformer ar- chitecture: Sparse Attention, which computes association discrepancy more efficiently by selectively focusing on the most relevant time steps. This re- duces computational redundancy while effectively capturing long-range de- pendencies critical for discerning subtle anomalies. A Mamba-Selective State Space Model (Mamba-SSM) is also integrated into the reconstruction mod- ule. A skip connection bridges the original reconstruction and the Mamba- SSM output, while a Gated Attention mechanism adaptively fuses features from both pathways. This design balances fidelity and contextual enhancement dynamically, improving anomaly localization and overall detection performance. Extensive experiments on benchmark datasets demonstrate that MAAT significantly outperforms prior methods, achieving superior anomaly distinguishability and generalization across diverse time series applications. By addressing the limitations of existing approaches, MAAT sets a new stan- dard for unsupervised time series anomaly detection in real-world scenarios.

Keywords: MAAT, Transformer, Association Discrepancy, Gated Attention, Mamba-SSM, Sparse Attention, Anomaly Detection, Unsupervised Learning

Architecture

Results

Get Start

1. Create new virtual environement with Python 3.10 .
2. Clone the repo.
3. Install the requirements using: pip install -r requirements.txt.
4. Download data from this Link
5. Train and evaluate. using the scripts in ./scripts folder:

bash ./scripts/SMD.sh
bash ./scripts/MSL.sh
bash ./scripts/SMAP.sh
bash ./scripts/PSM.sh
bash ./scripts/SWAT.sh
bash ./scripts/NIPS_TS_Swan.sh
bash ./scripts/NIPS_TS_Water.sh

Citation

If you find this repo useful, please cite our paper.

@article{SELLAM2025111685,
title = {Mamba Adaptive Anomaly Transformer with association discrepancy for time series},
journal = {Engineering Applications of Artificial Intelligence},
volume = {160},
pages = {111685},
year = {2025},
issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2025.111685},
url = {https://www.sciencedirect.com/science/article/pii/S0952197625016872},
author = {Abdellah Zakaria Sellam and Ilyes Benaissa and Abdelmalik Taleb-Ahmed and Luigi Patrono and Cosimo Distante},
keywords = {Transformer, Association discrepancy, Gated attention, Mamba state space model, Sparse attention, Anomaly detection, Unsupervised learning},
}

Contact

If you have any question or want to use the code, please contact z.abdellah.sellam@gmail.com, ilyesbenaissa7429@gmail.com, abdelmalik.taleb-ahmed@uphf.fr, cosimo.distante@cnr.it.