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

Anomaly detection (AD) in multivariate time series data (MTS) collected by industrial sensors is a crucial undertaking for the damage estimation and damage monitoring of machinery like rocket engines, wind turbine blades, and aircraft turbines. Due to the complex structure of industrial systems and the varying working environments, the collected MTS often contain a significant amount of noise. Current AD studies mostly depend on extracting features from data to obtain the information associated with various working states, and they attempt to detect the abnormal states in the space of the original data. Nevertheless, the latent space, which includes the most essential knowledge learned by the network, is often overlooked. In this paper, a multi-scale feature extraction and data reconstruction deep learning neural network, designated as LGFN, is proposed. It is specifically designed to detect anomalies in MTS in both the original input space and the latent space. In the experimental section, a comparison is made between the proposed AD process and five well-acknowledged AD methods on five public MTS datasets. The outcomes demonstrate that the proposed method attains state-of-the-art or comparable performance. The memory usage experiment illustrates the space efficiency of LGFN in comparison to another AD method based on GPT-2. The ablation studies emphasise the indispensable role of each module in the proposed AD process.