An Unsupervised Deep Learning Approach for Health Assessment of Distributed Systems

Different from binary classification approaches used in fault detection and anomaly detection, health assessment technology calculates a health score by leveraging the current operational state of a system. This quantifies the system's health status and enhances convenience in system maintenance. In recent years, the assessment of health status in distributed systems has emerged as a prominent research focus. However, distributed systems possess complex topology information and generate large amounts of metric data. Additionally, distributed systems usually function normally, which makes it challenging to obtain enough abnormal data for training models. Traditional expert-based methods and machine learning techniques struggle to address these challenges. Therefore, this paper proposes an unsupervised learning model called GraphLSTM-VAE, based on graph neural networks (GNN). This model utilizes graph neural networks to extract spatial features from topology information and long short-term memory networks (LSTM) to extract temporal features from metric data. The unsupervised training of the model employs the structure of a variational auto-encoder (VAE). Through the comparison and analysis of three sets of experiments, this paper demonstrates the feasibility of the proposed model for evaluating the health status of distributed systems and shows that the model outperforms some advanced algorithms.