Industrial-MultiSensor-AnomalyTS

Industrial equipment produces continuous sensor data, requiring systems to decide if recent history warrants an alert. This study evaluates whether domain prompting and native reasoning improve this decision in an anomaly-triage context. Two local models—NVIDIA Nemotron 3 Nano 4B and Gemma 4 e4b—were tested under 2 local-model conditions on 168 labelled 60-minute windows from a synthetic physics-informed dataset. A GPT-5.6 Sol provider-default reference was used for comparison. The evaluation showed opposite Nemotron/Gemma reasoning results. Reasoning improved Nemotron's accuracy but reduced Gemma's, while causing a latency increase for both. This highlighted a Gemma false-alert trade-off (catching most positive windows but generating excessive false positives) and a Nemotron conservative trade-off. No local effect met the predefined 20-percentage-point threshold. We make no deployment/universal-ranking claim; rather, reasoning is a model-specific setting balancing missed anomalies, false alerts, latency, and cost. ## Why anomaly triage matters Industrial monitoring yields raw data requiring actionable decisions. The critical first step is anomaly triage. Missing genuine anomalies delays intervention, while excessive false alerts overwhelm staff. While large language models are proposed for this analysis, evaluating them requires acknowledging binary-task limitations. Models received one hour of telemetry to return a simple alert or no alert. They were not asked to diagnose faults, predict future failures, estimate remaining useful life, or explain their reasoning.