Lightweight and Fast Time-Series Anomaly Detection via Point-Level and Sequence-Level Reconstruction Discrepancy

Area of Science:

• Artificial Intelligence
• Machine Learning
• Industrial IoT

Background:

• Unsupervised time-series anomaly detection (TSAD) is crucial for industrial safety.
• Industry 4.0 necessitates TSAD deployment on resource-limited edge devices.
• Existing TSAD models are often too computationally expensive for edge applications.

Purpose of the Study:

• To propose a lightweight and fast unsupervised TSAD model (LFTSAD) suitable for edge deployment.
• To address the limitations of existing deep learning models in terms of accuracy, timeliness, and resource consumption.
• To enable real-time anomaly detection on edge devices for enhanced production safety.

Main Methods:

• Developed LFTSAD, an unsupervised lightweight TSAD model using a contrastive architecture with two two-layer multilayer perceptron networks (MLPs).
• Implemented a dual-branch reconstruction network to generate point-level and sequence-level reconstruction discrepancies.
• Designed a novel anomaly scoring scheme combining both reconstruction discrepancies for improved detection accuracy.

Main Results:

• LFTSAD demonstrates 3-10 times improvement in timeliness compared to state-of-the-art (SOTA) models.
• The model consumes only half the resources of comparable deep learning models.
• Achieved accuracy comparable to or superior to existing deep SOTA TSAD models.

Conclusions:

• LFTSAD is the first lightweight All-MLP-based TSAD model designed for resource-limited edge devices.
• The proposed model effectively balances accuracy, timeliness, and low consumption for edge TSAD.
• LFTSAD offers a viable solution for real-time industrial anomaly detection and safety assurance.