Towards trustworthy ai monitoring of roll-to-roll electrode coating: an evidence-centred digital measurement twin with uncertainty-aware sensor fusion

This is the authors' abstract. We don't add key points for this paper.

Industrial AI monitoring increasingly transforms high frequency plant data into process state estimates and quality predictions, yet many systems still treat sensor readings as exact, stable and equally reliable. This limits trustworthiness because plant measurements are affected by calibration status, drift, missingness, environmental exposure, maintenance events and traceability gaps. We present an evidence-centred digital measurement twin for roll-to-roll electrode coating and drying that embeds baseline measurement uncertainty and dynamic evidence adjustment, metrological traceability, confidence scoring, data quality flags and FAIR metadata directly into an evidence-centred process monitoring and quality decision workflow. The framework was demonstrated using the study specific MDT-R2R-90 dataset, comprising 90 days, 25,920 timestamped records, 42 measurement channels, 270 batch or roll quality records and eight documented faults, drift or maintenance events. Each measurement channel was linked to an uncertainty budget and timestamp resolved evidence adjusted dispersion field, enabling transparent evidence weighted state estimation of moisture and thickness with propagated evidence. The twin also generated confidence, traceability, anomaly risk and quality risk outputs to qualify each prediction. Process disturbances and measurement chain degradations were analysed separately, and slowly logged vibration, acoustic, vision and NIR variables were interpreted as upstream processed features rather than raw dynamic signals. Batch level evidence linked monitoring outputs to 127 releases, 102 engineering reviews and 41 hold decisions. Quality risk scores separated pass, marginal and fail outcomes at mean level, with scores of 11.1, 86.0 and 99.2, respectively, while borderline records were retained for audit. In leakage controlled internal fivefold decision level validation, evidence-aware outputs improved macro-F1 from 0.701 to 0.829 and reduced expected calibration error from 0.151 to 0.047. These results demonstrate the internal feasibility and consistency of the proposed approach for one anonymised plant dataset; external validation across plants, products and operating regimes remains required.