LLMs Clean Faulty Sensor Data: 80% Error Reduction in Industrial Predictions

Industrial AI predictions often fail not because the forecasting model is weak, but because the input data is silently corrupted—biased, delayed, or derived from shared sensors. Traditional fixes rely on numerical correlations, alarms, or explicit process equations, which aren't always available. Enter LLM-Guided Measurement Credibility Correction (MCC), proposed by Youcheng Zong, Runda Jia, and Dakuo He. MCC taps into process documents—manuals, logs, schematics—and uses an LLM to extract measurement semantics: what each sensor actually means and how it relates to others. These semantics are then fed into a lightweight numerical model (just 0.5–2.0k parameters) that independently verifies each measurement's credibility before the predictor sees it. If a conflict is detected—say, a flow rate derived from the same pressure sensor as another variable—MCC corrects it on the fly at 0.089 ms per step.

The results speak volumes. Across multiple complex industrial forecasting and soft-sensing tasks, +MCC achieved a 30.7% average reduction in mean absolute error on real test protocols and a staggering 80.3% reduction when inputs were deliberately corrupted. This pre-inference credibility layer is model-agnostic and lightweight enough to integrate into existing pipelines without retraining the main predictor. By turning static documentation into dynamic guardrails, MCC addresses a critical blind spot: data quality degradation that traditional approaches miss. For engineers running predictive maintenance, quality control, or process optimization, this means fewer false alarms and more reliable outputs—especially in settings where sensor drift or replacement is common.