Paper Title
xLSTM for Time Series Surveillance: Detecting Anomalies in Multivariate Data

Abstract
Detecting and understanding irregular behaviors across multiple variables in time-dependent data is essential for maintaining the stability and robustness of contemporary industrial operations. However, this process is complicated by several factors, including the scarcity of annotated anomaly instances, the inherently unstable nature of real-world signals, and the demand for immediate detection with minimal delay. While transformer-based frameworks like TranAD have made progress in tackling these challenges, they continue to encounter hurdles related to their ability to scale across large datasets, adjust to a wide range of temporal behaviors, and remain robust when dealing with noisy or unpredictable inputs. In this study, we present xLSTM—a novel architecture for anomaly detection and root cause analysis—that consistently surpasses TranAD in various evaluation metrics. Departing from TranAD’s transformer-based approach, xLSTM leverages an advanced LSTM design integrated with cross-layer attention, allowing it to efficiently model both transient variations and extended temporal patterns. xLSTM integrates a hybrid feature fusion mechanism that combines temporal, statistical, and frequency-domain insights for superior anomaly characterization. Additionally, it incorporates lightweight dynamic gating and adaptive memory modules, enabling faster inference and lower computational overhead. Fig. 1: Depiction of the xLSTM architecture family, which augments traditional LSTM by introducing specialized memory cell types—sLSTM featuring novel memory fusion and mLSTM incorporating structured matrix memory. These components are composed into xLSTM blocks to enhance the modeling of sequential data. xLSTM also features improved robustness through noise-aware training and context-sensitive anomaly scoring, which enhances accuracy in real-world noisy data. Leveraging advanced meta-learning strategies, it requires fewer training samples while achieving higher generalization across datasets and domains. Empirical evaluations on six publicly available datasets demonstrate that xLSTM surpasses state-of-the-art methods, including TranAD, in both anomaly detection and diagnosis. xLSTM achieves up to 25% higher F1 scores, reduces training time by over 99%, and offers 2× faster inference speeds, making it highly suitable for real-time industrial applications. Keywords - Component, Formatting, Style, Styling, Insert