Cost-Sensitive Mamba Sequence Modeling for Fault Detection in Cloud-Native Microservice Systems

Abstract

While microservice architecture improves system scalability and iteration efficiency, it also introduces more complex temporal dependencies and cross-service coupling behaviors, resulting in anomalies exhibiting characteristics such as low frequency, long tails, varied forms, and strong signal-noise ratios. Against this backdrop, extreme class imbalances and asymmetric costs for false positives and false negatives further exacerbate the practical difficulty of anomaly detection. This paper proposes a cost-sensitive Mamba sequence anomaly detection method for rare microservice faults. The method encompasses unified multivariate monitoring data alignment and cleaning, sliding window sample construction, long sequence representation learning based on a state-space mechanism, anomaly probability modeling and threshold alarm decision-making, and a cost-sensitive optimization strategy that explicitly injects business costs into the loss function. This method aggregates key temporal information through gating state updates, enhancing its ability to characterize long-range dependencies and dynamic fluctuations. Simultaneously, it adjusts the optimization intensity of different false positives with cost weights, making the model more aligned with risk control needs under imbalanced data conditions. Comparative evaluation based on an open-source microservice anomaly detection dataset shows that the proposed method has a competitive advantage in detection quality and probability calibration consistency, and maintains a more reasonable trade-off between alarm intensity and risk identification, making it suitable for online monitoring and intelligent operation and maintenance scenarios.