Learning in Two-Scales Through LSTM-GPT2 Fusion Network: A Hybrid Approach for Time Series Anomaly Detection.

Area of Science:

• Machine Learning
• Industrial Monitoring
• Time Series Analysis

Background:

• Anomaly detection (AD) in industrial multivariate time series data (MTS) is critical for machinery health monitoring.
• Industrial MTS often contain noise due to system complexity and environmental variations.
• Existing AD methods often overlook the latent space, focusing solely on the original data space.

Purpose of the Study:

• To propose a novel deep learning neural network, LGFN, for anomaly detection in MTS.
• To effectively detect anomalies in both the original input space and the latent space of MTS.
• To improve the accuracy and efficiency of anomaly detection in industrial machinery.

Main Methods:

• Developed a multi-scale feature extraction and data reconstruction deep learning neural network (LGFN).
• Designed LGFN to analyze anomalies in both the original input space and the learned latent space.
• Conducted comparative experiments against five established AD methods on five public MTS datasets.

Main Results:

• The proposed LGFN method achieved state-of-the-art or comparable performance across multiple datasets.
• LGFN demonstrated superior space efficiency compared to a GPT-2-based AD method.
• Ablation studies confirmed the essential contribution of each module within the LGFN architecture.

Conclusions:

• LGFN offers a robust and efficient solution for anomaly detection in noisy industrial MTS.
• Analyzing both original and latent spaces significantly enhances AD performance.
• The developed method provides a valuable tool for damage estimation and monitoring in critical industrial machinery.