Unsupervised learning enables instance-level microservice fault detection using traces and resource metrics

Microservice systems exhibit complex and time-varying dependencies, while elastic scaling and load balancing further obscure fault propagation paths and diversify anomaly patterns. These characteristics make instance-level fault detection challenging, especially when labeled fault data are unavailable. This paper presents an unsupervised likelihood-based framework for instance-level microservice fault detection from distributed traces and resource metrics. Each request trace is encoded as a Multidimensional Feature Trace (MFT) in a compact coordinate (COO) format, which preserves call structure and invocation order while fusing span latency with instance-level resource KPIs such as CPU and memory usage. To reduce distribution mixing caused by heterogeneous trace structures, we perform edit-distance based prototype clustering and learn a pattern-specific normal baseline for each cluster. For baseline modeling, we employ a BiLSTM-VAE to capture sequential dependencies in MFTs and further improve posterior expressiveness using a RealNVP flow. During online detection, trace log-likelihoods are evaluated against cluster-specific normal baselines using KDE-based tail-probability hypothesis testing. Experiments on the AIOps2020 and TrainTicket datasets achieve F1-scores of 0.979 and 0.985, respectively, outperforming several strong baselines under the adopted evaluation protocol. Ablation results further show that multidimensional feature fusion, pattern-wise baseline separation, and posterior flow modeling each contribute to the overall detection performance.