CovDiff-PaAno keeps PaAno's encoder and rewrites only the scoring: multivariate anomaly AUC-ROC rises from 0.7612 to 0.7843

The improvement can sit entirely in the inference-stage scoring rule, without redesigning or retraining the encoder. CovDiff-PaAno keeps PaAno's one-dimensional convolutional patch encoder and its training objective, and changes only how query patches are compared with normal memories, adding views built from temporal differences. PaAno scores raw patch embeddings by cosine nearest-neighbor distance to normal memories, a rule that measures angular disagreement only, ignores the scale and correlation of embedding dimensions, and does not directly reflect changes in slope or curvature. On TSB-AD-M, with identical data, training boundaries, encoder, optimization settings, random seeds and evaluation pipeline, the ten-seed mean raised AUC-ROC from 0.7612 for reproduced PaAno to 0.7843 and Range-F1 from 0.4043 to 0.4384; on TSB-AD-U, AUC-ROC went from 0.8532 to 0.8626 and Range-F1 from 0.4822 to 0.4924.

The three views share one frozen encoder but each holds its own normal reference distribution, so raw values, slope changes and curvature changes are each scored against their own normal memory. Single-view cosine scoring supplies one normal reference distribution, and reusing it for difference views would face mismatched embedding distributions and incompatible branch score scales. Each view builds a coreset by mini-batch k-means over ℓ2-normalized candidate embeddings, estimates a covariance regularized with ϵ=10⁻⁵ and computes Mahalanobis distance through its pseudo-inverse, averages the k=3 nearest memories, and fuses with fixed weights α=0.3 and β=0.15; across all 200 TSB-AD-M series with seeds 2000, 2021 and 2022, the shared frozen encoder stayed within 0.005 of the best view-specific encoder on every metric.

The controlled comparison is against reproduced PaAno; published results for other methods serve as context rather than locally rerun head-to-head evidence. Evaluation covers 200 multivariate series in TSB-AD-M and 870 univariate series in TSB-AD-U, and the encoder-sharing experiment is limited to TSB-AD-M with three seeds.

The component ablation and fusion-weight sensitivity experiment each use a single seed, so small differences between configurations remain diagnostic rather than statistically conclusive. Empirical-CDF calibration substantially reduced every reported metric, while z-score, robust z-score and min–max calibration changed the default fusion by at most 0.0024 in absolute value and their paired bootstrap intervals included zero. Full-covariance estimation may become unstable when the memory bank is small relative to the embedding dimension, and the boundary-alignment rule has not been compared with padding, forward alignment or boundary exclusion.