Reliability-Constrained Multimodal Anomaly Detection for Streaming Non-Stationary Time Series Data Analysis
Year: 2026
Type: article
Abstract: . Streaming decision support systems frequently face two coupled operational constraints: multimodal input streams are non-stationary and periodically exhibit modality degradation (missing values, additive noise, gain or scale shifts), while alarm frequency must be limited by an explicit false alarm rate (FAR) budget. Naive multimodal residual scoring then inflates anomaly scores during degraded segments, pushing FAR-constrained thresholds upward and reducing detector sensitivity to genuine anomaly events. The aim is to design and empirically validate a residual scoring policy that retains anomaly sensitivity under alternating modality degradation at a fixed FAR budget. The research tasks are to formalise the streaming multimodal anomaly detection problem under a FAR budget in a causal setting, to design an online modality reliability estimator from cheap causal degradation cues (missingness and feature-energy inflation), to construct a reliability-gated rule for selecting the source of the residual evidence, and to validate the approach on a controlled stream and on the UCI Air Quality dataset. The paper proposes RC-AD, a reliability-constrained residual scoring policy that combines standardised residual scores from modality-specific and early-fusion multimodal predictors, online modality reliability weights estimated from short causal windows using the missingness and feature-energy inflation, and a winner-takes-all rule: in clean windows the early-fusion residual score is used, otherwise the residual score of the currently more reliable modality. RC-AD is a control policy on top of an arbitrary forecasting backbone, which makes it lightweight, auditable, and easy to integrate into existing monitoring pipelines. The empirical study uses a causal protocol with ten fixed seeds and 95% confidence intervals: on a controlled benchmark with alternating modality degradation and injected event anomalies, RC-AD improves Recall@FAR from 0.103 to 0.335 at FAR 0.05 and from 0.182 to 0.426 at FAR 0.10, outperforming naive multimodal, equal-weight fusion and single-modality baselines; a demonstration on UCI Air Quality at FAR 0.01 confirms the same trend against the multimodal baselines. The scientific novelty consists in formulating a multimodal anomaly detection policy with an explicit online reliability constraint that, for the first time, combines gating of early-fusion evidence by causal degradation cues with selection of the most reliable single modality at a prescribed FAR budget. The practical significance is supported by reproducible Recall@FAR improvements on both controlled and real sensor data and by the lightness and interpretability of the method, which suits auditable decision support systems Keywords: machine learning, data analysis, information systems, decision support systems, multimodal time series, non-stationary time series, streaming anomaly detection, data quality degradation, modality reliability estimation, false alarm rate budget.
Source: ELECTRICAL AND COMPUTER SYSTEMS
Authors: Illia Uzun, Mykhaylo Lobachev
Institution: Odesа Polytechnic National University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7162933869