Abstract:For modern industrial applications, accurately detecting and diagnosing anomalies in multivariate time series data is essential. Despite such need, most state-of-the-art methods often prioritize detection performance over model interpretability. Addressing this gap, we introduce MIXAD (Memory-Induced Explainable Time Series Anomaly Detection), a model designed for interpretable anomaly detection. MIXAD leverages a memory network alongside spatiotemporal processing units to understand the intricate dynamics and topological structures inherent in sensor relationships. We also introduce a novel anomaly scoring method that detects significant shifts in memory activation patterns during anomalies. Our approach not only ensures decent detection performance but also outperforms state-of-the-art baselines by 34.30% and 34.51% in interpretability metrics.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the lack of anomaly detection and interpretability in multivariate time - series data. Specifically: 1. **Accurately detect and diagnose anomalies in multivariate time - series data**: In modern industrial applications, accurately detecting and diagnosing anomalies in multivariate time - series data is crucial. However, most of the existing state - of - the - art methods tend to focus more on detection performance and neglect the interpretability of the model. 2. **Improve the interpretability of the model**: Although existing methods perform well in detection performance, they are deficient in explaining the root causes of anomalies. Especially in environments such as intelligent manufacturing, it is very important to identify which sensors or interactions between sensors cause anomalies. Therefore, a method that can provide clear explanations while maintaining high detection performance is required. To solve the above problems, the paper introduces MIXAD (Memory - Induced Explainable Time Series Anomaly Detection), a model designed specifically for interpretable anomaly detection. The main contributions of MIXAD include: - **Introduce a memory module for the first time in multivariate time - series anomaly detection**: By storing the complex dynamics of data extracted from spatio - temporal feature extractors, MIXAD can model the intra - metric and inter - metric dependencies simultaneously, thus providing an in - depth understanding of the nature of the data. - **Propose a novel anomaly scoring mechanism based on memory activation pattern analysis**: This method not only combines the accuracy of anomaly detection but also explains the root causes of anomalies by analyzing the differences in memory activation patterns during normal and abnormal periods. Unlike existing methods that mainly rely on prediction or reconstruction errors, MIXAD determines which features cause anomalies and how they are affected through a more complex scoring method. - **Significantly improve the interpretability and detection performance of the model**: Experimental results show that MIXAD improves the interpretability indicators by 34.30% and 34.51% respectively compared to the existing state - of - the - art baseline methods, and is also competitive in detection performance. In summary, MIXAD aims to significantly improve the interpretability of multivariate time - series anomaly detection while maintaining high detection performance by introducing a memory module and a novel anomaly scoring mechanism.

MIXAD: Memory-Induced Explainable Time Series Anomaly Detection

PatchAD: A Lightweight Patch-based MLP-Mixer for Time Series Anomaly Detection

A Multi-scale Patch Mixer Network for Time Series Anomaly Detection

Segmentation-Based Adversarial Denoising Auto-Encoder for Anomaly Detection in Multivariate Time Series Data

Multilevel Anomaly Detection for Mixed Data

HybridAD: A Hybrid Model-Driven Anomaly Detection Approach for Multivariate Time Series

A robust multi-scale feature extraction framework with dual memory module for multivariate time series anomaly detection

A novel spatio-temporal memory network for video anomaly detection

Memory-augmented Adversarial Autoencoders for Multivariate Time-series Anomaly Detection with Deep Reconstruction and Prediction

MEMTO: Memory-guided Transformer for Multivariate Time Series Anomaly Detection

MixMamba: Time Series Modeling with Adaptive Expertise

TSMAE: A Novel Anomaly Detection Approach for Internet of Things Time Series Data Using Memory-Augmented Autoencoder

SimAD: A Simple Dissimilarity-based Approach for Time Series Anomaly Detection

Adaptive Memory Broad Learning System for Unsupervised Time Series Anomaly Detection

AcME-AD: Accelerated Model Explanations for Anomaly Detection

Multivariate Time Series Anomaly Detection and Interpretation Using Hierarchical Inter-Metric and Temporal Embedding

DyMix: Dynamic Frequency Mixup Scheduler based Unsupervised Domain Adaptation for Enhancing Alzheimer's Disease Identification

MGAD: Mutual Information and Graph Embedding Based Anomaly Detection in Multivariate Time Series

Embarrassingly Simple MixUp for Time-series

Explainable Time Series Anomaly Detection using Masked Latent Generative Modeling

Memory-Augmented Spatial-Temporal Consistency Network for Video Anomaly Detection.