In customs supervision scenarios, import and export cargo data exhibits significant non-uniform distribution characteristics due to variations in cargo types and seasonal fluctuations. Traditional anomaly detection methods, which rely on uniform distribution assumptions, face performance limitations. To address this critical issue, this paper proposes the HADN framework, which enhances anomaly detection accuracy for non-uniform time series through density-adaptive preprocessing and multimodal feature modeling. The core innovations of this study include: (1) A dynamic density discrimination mechanism based on variance analysis, enabling real-time partitioning of sparse and dense subsequences; (2) A dual-path feature processing architecture combining Fourier decomposition with resampling (for periodic dense data) and context-aware imputation (for sparse data with semantic mutations); (3) A local-global representation contrastive decision module that optimizes anomaly criteria through multidimensional feature alignment. Existing methods primarily designed for uniform distribution scenarios struggle to address feature drift caused by density heterogeneity. Experimental results demonstrate that HADN achieves a 1.62% improvement in F1-score compared to state-of-the-art baselines across five benchmark datasets, validating its technical superiority in non-uniform distribution scenarios.