MSDG: Multi-Scale Dynamic Graph Neural Network for Industrial Time Series Anomaly Detection.

Area of Science:

• Industrial IoT and Sensor Networks
• Machine Learning for Time Series Analysis
• Graph Neural Networks for Anomaly Detection

Background:

• Industrial plants generate vast amounts of real-time sensor data with inherent time series and spatial correlations.
• Existing Graph Neural Network (GNN) models often fail to simultaneously capture dynamic spatial-temporal dependencies in sensor data.
• A need exists for advanced anomaly detection methods capable of handling complex, multi-scale correlations in industrial operational data.

Purpose of the Study:

• To develop a novel Multi-Scale Dynamic Graph Neural Network (MSDG) for robust anomaly detection in industrial sensor data.
• To address the limitations of existing models in capturing simultaneous dynamic correlations across time and space.
• To improve the accuracy and reliability of anomaly detection in critical industrial environments.

Main Methods:

• A multi-scale sliding window mechanism was employed to process sensor data at various temporal scales.
• A dynamic graph neural network architecture was designed to model intricate spatial-temporal dependencies within multivariate sensor data.
• Anomaly detection was performed by reconstructing sensor data sequences and analyzing reconstruction errors.

Main Results:

• The proposed MSDG model demonstrated superior performance in anomaly detection across three real-world public datasets.
• The multi-scale approach effectively captured both short-term and long-term dependencies in sensor data.
• The dynamic graph construction successfully modeled the evolving spatial-temporal relationships between sensors.

Conclusions:

• The MSDG model offers a significant advancement in anomaly detection for industrial sensor data by effectively handling dynamic spatial-temporal correlations.
• The proposed method provides a more comprehensive understanding of operational data compared to traditional approaches.
• MSDG is a promising technique for enhancing the safety and efficiency of industrial operations through reliable anomaly identification.