Abstract
Anomaly detection for time series data has been proven useful in both academia and industry and can be categorized into global and local methods based on the range of data used to determine anomalies. To the best of our knowledge, there are no methods that unify global and local methods together. In this study, we propose an approach that unifies global and local anomaly detection to achieve superior performance. To make this unification possible, we discretize time series data and partition the discrete data into global and local anomaly detection parts based on a concept of local factor. We design a rule-based global method using non-redundant association rules with tolerances to handle the difficulty of the latency phenomenon and the excessive number of rules in time series data. We develop a Transformer-based local method with Locality Sensitive Hashing (LSH) Attention to better fit discrete data. Finally, we conduct an extensive experimental study to verify the effectiveness and efficiency of our approach.
Introduction
Anomaly detection in multivariate time series is a critical task for maintaining the health of complex industrial systems. Traditionally, detection methods fall into two categories: Global Methods, which identify violations of long-term patterns across the entire dataset, and Local Methods, which focus on deviations from recent trends in a data point's immediate neighborhood.
While both approaches are valuable, real-world anomalies often exhibit a mix of global and local characteristics. Furthermore, state-of-the-art deep learning models often suffer from high computational complexity ($O(n^2)$) and sensitivity to the "latency phenomenon"—where different sensors respond to the same event at slightly different times, leading to false positives. This paper introduces a novel framework that bridges this gap by unifying both detection paradigms into a single, efficient pipeline.
The Unified Framework
The proposed approach follows a multi-stage process involving data discretization, partitioning, and parallel detection engines. By separating the data into global and local components, the system can apply the most appropriate algorithm to each type of pattern.
Methodology
1. Data Discretization
Continuous sensor data is first converted into discrete "attribute-value pairs." The authors utilize three distinct strategies—trend-based (clustering segments), cluster-based (k-means), and EM-based (Gaussian mixture models)—to ensure that the discretization process captures the essential physical properties of the monitored system. This step acts as a noise filter and clarifies the boundaries between normal and abnormal states.
2. Data Partitioning via "Local Factor"
The discretized data is partitioned using a Local Factor (LF), which is calculated based on the frequency (support) of specific items. High-frequency items, representing common system states, are funneled into the Global Detection Part (GDP). Rare items, which are more likely to represent localized shifts, are sent to the Local Detection Part (LDP).
3. Global Detection: Association Rules (AR)
For the GDP, the framework employs a rule-based engine that mines non-redundant association rules. To combat the latency phenomenon, the authors introduce tolerances ($tol$) to these rules. This allows the model to ignore transient violations caused by sensor lag, significantly improving precision and reducing the number of rules required by orders of magnitude.
4. Local Detection: AT-LSH
The LDP utilizes an improved version of the Anomaly Transformer. By replacing standard Self-Attention with Locality Sensitive Hashing (LSH) Attention, the model reduces computational complexity from $O(n^2)$ to $O(n \log n)$. This change also provides greater stability when processing the sparse, "long-tailed" distributions typical of discretized local data.
Experimental Evaluation
The researchers evaluated their approach on three high-dimensional, real-world industrial control system (ICS) benchmarks: SWaT (Secure Water Treatment), WADI (Water Distribution), and HAI (HIL-based Augmented ICS). These datasets contain hundreds of sensors and actuators monitoring complex physical processes.
Key Results
The unified approach demonstrated significant improvements over 10 representative baselines, including USAD and the original Anomaly Transformer:
- Accuracy: F1 scores increased by an average of 4.7% on SWaT, 11.9% on WADI, and 7.0% on HAI compared to previous state-of-the-art models.
- Efficiency: Total running time was reduced by 39.5% to 60.3% across the datasets due to the efficiency of LSH attention and the partitioned architecture.
- Rule Optimization: The global engine used between 110x and 462x fewer rules than standard association mining by focusing only on non-redundant patterns.
Conclusion
This research demonstrates that global and local anomaly detection are complementary rather than competing approaches. By unifying them through a discretization and partitioning strategy, the proposed framework achieves higher accuracy and better computational efficiency than single-paradigm models. Future work will explore more automated discretization methods to reduce the need for manual threshold tuning and investigate the application of this framework to broader data mining tasks.
Summary updated