A Multi-Scale Deep Learning and Uncertainty Estimation Framework for Comprehensive Anomaly Detection in Cloud Environments

Abstract

This study proposes a multi-scale deep learning-based detection method to address the complexity, dynamics, and diversity challenges of anomaly detection in cloud service systems. By introducing multi-scale feature extraction and cross-scale fusion mechanisms, the method effectively characterizes system behavior evolution across different temporal granularities, enabling the capture of both short-term burst anomalies and long-term structural anomalies to improve detection comprehensiveness and accuracy. In terms of model architecture, hierarchical feature modeling and context-aware mechanisms are employed to achieve a deep representation of semantic associations and temporal dependencies among multidimensional metrics. In addition, an uncertainty estimation module is introduced to calibrate boundary samples and low-confidence predictions, which effectively reduces false positives and false negatives and enhances system stability and robustness in highly dynamic environments. The method is also systematically evaluated under various environmental factors, including hyperparameter variations, resource interference, sampling granularity, and data distribution drift. Experimental results show that it outperforms existing methods on multiple key metrics and demonstrates strong adaptability and discriminative power. Overall, the proposed multi-scale detection framework provides reliable technical support for intelligent operation, automated anomaly management, and complex service state monitoring in cloud computing systems, offering an effective solution for ensuring stability in large-scale distributed environments.