Data-Driven Calibration and Reference-Free Drift Detection for Multi-Modal Sensor Systems

Low-cost multi-modal sensors make high-resolution exposomic and urban-health monitoring affordable, but their data are only as useful as they are trustworthy. Measurement bias, cross-sensitivity, temporal drift and domain shift between devices and sites erode reliability, and reference instruments are rarely available for every sensor. We present a two-part reliability workflow for this realistic setting, in which reference data exist for only a subset of sensors, and evaluate it on five custom multi-modal devices deployed at two Brussels sites (9 to 68 days of data per device). For calibration, we benchmark Multiple Linear Regression, Elastic Net, Random Forest, Gradient Boosting and a Domain Adversarial Neural Network (DANN) for O3, NO2, PM2.5 and PM10 on the source device and on other devices at the same site and at the other site using MAE, RMSE, R2, bias, device-balanced aggregation and block-bootstrap intervals. Tree ensembles achieve the lowest source-device errors for most pollutants, yet for PM10, they lose their advantage over the uncalibrated sensor when transferred, a caution for any deployment that calibrates on one unit. DANN attains the highest mean Calibration Transfer Ratio in the transfer settings, although with the available device pairs, the differences are not statistically significant. For modalities without a reference, a cross-sensor residual monitor combined with statistical detectors reliably flags injected gradual drift in sensors that can be reconstructed from the others (UV bands, PM1, temperature, humidity and CO), while Lux, pressure and VOC remain harder. The monitor detects cross-sensor inconsistency rather than isolating the faulty sensor, and the results reflect an injected fault model, not real sensor ageing. Together, these results give a transparent, reproducible baseline and clear practical guidance for building trustworthy low-cost sensor networks.