Fuzzy Fault Detection and Band-Conditioned Data Quality Assessment for Sensor Streams: A Continuous Glucose Monitoring Case Study

Abstract

Sensor faults can propagate beyond raw measurements and distort the aggregates used to interpret a monitored phenomenon. We propose a domain-agnostic, five-layer fuzzy pipeline that characterizes the reliability of a sensor stream jointly over time and value bands and carries it through to every derived aggregate: (L1) fault tagging of each sample against five fuzzy failure primitives; (L2) fuzzy windowing; (L3) band reliability; (L4) routed repair with propagated reconstruction uncertainty; and (L5) veracity aggregation with linguistic protoforms. An end-to-end acquisition architecture is also specified. The pipeline was instantiated for continuous glucose monitoring (CGM) with clinical knowledge of glucose dynamics and applied to a 14-month recording of an adult with type 1 diabetes (FreeStyle Libre 3, 5-min cadence), of which 294 days contained at least 120 readings. Clean reference days remained essentially unchanged (mean sample quality 0.98–0.99). On the most corrupted day, the estimated time below range fell from 16% to 3%, with a veracity of only 0.14. Over the recording, 38% of the days were too unreliable to phenotype and 5% changed phenotype. Joint time- and band-conditioned reliability assessment thus prevents corrupted sensor data from producing falsely precise clinical conclusions.