Multi-AXIS: Explaining Multivariate Time-Series Anomalies with Detector-Derived Contextual Evidence

Time-series anomaly detection (TSAD) typically returns scores or labels, whereas practical analysis also requires explanations. Existing language-model interfaces combine raw observations, statistical or textual summaries, detector outputs, or continuous time-series representations, but generally do not connect ”what was observed” at a channel–time location to ”why it was judged anomalous.” We propose Multi-AXIS, which constructs two complementary forms of evidence. Contextual Evidence fuses a frozen detector's position-wise contextual representations with anomaly-discrimination signals at the same locations. Observable Evidence combines a global visual reference with local target-window values. Their channel–time alignment lets the model check contextual judgments against concrete observations. A lightweight interface further constructs channel-level and question-relevant event-level representations and supplies both forms of evidence to a frozen vision-language model (VLM). We build approximately 70,000 multivariate time-series question–answer pairs. Experiments on synthetic and real-world data show that Multi-AXIS produces more accurate, complete, and evidence-based anomaly explanations while achieving competitive overall performance.