Abstract: The microservice architecture is increasingly adopted as the preferred approach for enterprises to build complex application systems due to its flexibility, scalability, and high cohesion with low coupling. However, massive microservice monitoring indicators are generated with significant variations in quantity and granularity across different system layers, resulting in high computational costs being incurred during data processing and anomaly detection. Moreover, existing research primarily relies on key performance indicators (KPIs) while neglecting potential correlations between cross-layer monitoring metrics, which limits the comprehensiveness and precision of system state perception and anomaly detection.To address these challenges, we propose Metric Correlation-Based Dual Coding - Microservice Anomaly Detection (CDC-MAD), a dual-encoder model that integrates raw and structurally correlated monitoring features to improve anomaly detection under complex microservice environments. Firstly, the dynamic time warping (DTW) algorithm is employed to measure inter-indicator correlations, based on which a K-nearest neighbor network structure is constructed. Subsequently, hierarchical community clustering is implemented to identify critical indicators sensitive to abnormal states, thereby building a streamlined yet effective correlated feature space that reduces computational overhead. Then, one-dimensional convolutional neural networks (1D-CNN) are utilized to extract features and unify dimensions from raw data. Dual-path encoders are designed to fuse features from both raw and correlated spaces, enabling enhanced capture of abnormal characteristics. Finally, the reparameterization technique of variational autoencoders (VAE) is applied to strengthen feature representation robustness. Experimental results demonstrate that compared to similar anomaly detection methods such as VAE, OmniAnomaly, and SDFVAE, the proposed approach achieves average precision improvements of 2.99%, 39.06%, and 1.92% respectively. Comprehensive experiments across multiple datasets validate the effectiveness of the CDC-MAD method.