A Two-Layer Signal Validation and Temporal Analysis Framework for Environmental Sensor Fault Detection in a Small-Scale Vertical Farm

Reliable environmental sensing is essential for stable monitoring and control in vertical farms; however, high data-acquisition completeness does not ensure that measurements are valid, temporally consistent, or operationally meaningful. This study proposed an interpretable two-layer framework that distinguished direct signal-quality errors from in-range temporal and operating-context abnormalities in a Wi-Fi-based small-scale vertical farm. Nine sensor nodes were installed across nine cultivation beds under empty-bed and continuous LED-on conditions. Air temperature, relative humidity, CO2 concentration, and light intensity were measured at 30 s intervals for 10 days and transmitted through MQTT, Node-RED, and InfluxDB on a Raspberry Pi local server. Layer 1 screened timestamp discontinuities, invalid or null packets, and physically invalid measurements. Layer 2 evaluated only valid records using moving-window variability, rate of change, peer-node residuals, and LED-on context consistency, and a study-specific F-code taxonomy converted detected abnormalities into interpretable operating-status labels. Detection thresholds were determined from a separate two-day period immediately preceding the 10-day evaluation and were not adjusted using the evaluation records. Of 259,200 expected records, 258,910 were received (99.888% completeness). Layer 1 identified 290 missing timestamp positions, 25 invalid packets, and 40 out-of-range records. The framework achieved 99.68% accuracy, 98.49% precision, 98.89% recall, 98.69% binary F1 score, and 84.69% macro-averaged F-code F1 score. Combining signal validation with temporal and context analysis detected abnormalities that static range validation could not identify.