Interval width fuzzy aggregation for streaming process monitoring on the Tennessee Eastman benchmark

Abstract

This study evaluates a transparent interval-width fuzzy anomaly score for streaming process monitoring. The reported IT2-DRA configuration converts standardized variables into Gaussian normality memberships with widths in [0.8, 1.2], takes complementary anomaly memberships, averages their interval endpoints with fixed normalized weights, and smooths the result. The width band is a design assumption; the resulting score is neither an estimated accident probability nor a calibrated measure of epistemic confidence. On 21 classic Tennessee Eastman (TE) fault streams, the reported mean area under the receiver operating characteristic curve (AUC) was 0.8099 ± 0.1667 and the capped detection-delay mean was 251.4 samples. Hotelling T², Multivariate Exponentially Weighted Moving Average (MEWMA), and a shallow autoencoder achieved mean AUCs of 0.9183, 0.9117, and 0.8758 and shorter mean delays. The reported Friedman AUC test gave p = 6.83e−06; Holm-adjusted paired comparisons favored these three baselines. The type-1 counterpart had a statistically detectable but small mean AUC advantage of 0.0006. The false-alarm comparison did not reject equality (p = 0.2483), which does not establish equivalence. A per-variable contribution bound explains why weighted averaging can dilute localized deviations, while joint dependence is not represented explicitly. The findings delimit this interval-midpoint construction on the stated benchmark and do not assess general rule-based interval type-2 inference.