Low-cost particulate matter (PM) sensor networks can only deliver indicative-quality measurements if their calibration scales across devices, locations, and time. De Vito and co-workers showed that a single 'global multi-unit' machine-learning calibration, trained on aggregated data from a few co-located IoT multisensor devices, matches per-device (ad-hoc) calibration performance for PM2.5 in one geographic area (Portici, Italy) across winter and summer campaigns. It is open whether such a zero-transfer-sample global calibration law remains valid as deployment diversity grows — across locations with different aerosol composition, source mix, and microclimate, and over long-term sensor aging and seasonal regime shifts — and which detection, re-calibration/update, and model-execution (on-board vs edge) strategies are needed to maintain it.

Background

Low-cost air quality monitoring systems (LCAQMS) use optical particulate matter sensors — light-scattering devices that count particles and convert counts to mass concentrations for size fractions such as PM2.5 (particles below 2.5 μm aerodynamic diameter) and PM10 — and are cheap enough for dense urban networks. Their raw readings are biased by environmental interference (especially relative-humidity-driven hygroscopic growth of particles, which changes light scattering), by the aerosol's composition and size distribution, and by unit-to-unit fabrication variance. The standard remedy is field calibration: co-locating each device next to a regulatory reference-grade analyzer for weeks and fitting a regression or machine-learning correction (using raw sensor channels, humidity, and temperature as inputs), evaluated by the coefficient of determination (R2R^2) and mean absolute error (MAE). This per-device ('ad-hoc') calibration gives the best accuracy but is unscalable: every unit needs its own co-location, and 'concept drift' — statistical change between training and operating conditions caused by seasonal emission and weather shifts or device relocation — plus progressive sensor aging or poisoning (degradation of the sensing element over months to years of operation) forces frequent repetition.

One escape route is a global (also called general or multi-unit) calibration: a single model, trained on data from a small set of co-located units of the same device type, applied 'zero-transfer-sample' to all other units. Precedents include the general calibration model for low-cost gas sensors of Malings and colleagues, and the United States-wide PurpleAir PM2.5 correction of Barkjohn, Gantt, and Clements — a simple humidity-dependent equation fitted once and applied across the continental US, showing that a single correction can work at national scale for one device type. Conversely, Nowack and colleagues showed that machine-learning calibrations of low-cost NO2 and PM10 sensors transfer poorly between monitoring sites, so cross-location generalization of data-driven calibrations is not automatic.

The source paper (De Vito et al., IEEE Transactions on Instrumentation and Measurement, 2024) trains global calibrations on aggregated data from kk co-located ENEA MONICA multisensor devices (a microcontroller-based IoT air quality unit with multiple gas and PM sensors) at an Italian environmental-agency (ARPAC) reference station in Portici, near Naples, during three campaigns (winter 2020/21, summer 2021, winter 2021/22), and applies them to unseen MONICA units. Its central result is that the global multi-unit calibration matches ad-hoc per-unit calibration performance on PM2.5, both short-term and in winter-to-summer long-term transfer, potentially enabling deployments of hundreds of devices with one shared, computationally light calibration executable on board or at the edge. However, the evaluation covers a single geographic area and a limited seasonal span. The authors report that in long-term transfer a few models exhibit negative R2R^2 (worse than predicting a constant), a variance shared with ad-hoc models and attributed to summer conditions with low concentrations near the sensors' detection limit; they state that changing environmental and emission conditions, and aging or poisoning effects, will degrade performance over time, that the limited variance in a short-term single-location training set cannot confer generalization to very different input distributions, and that on-board execution complicates model updates (firmware updates) relative to edge/cloud execution, which instead costs raw-data transmission. Their explicit future work is exploration of the methodology 'in increased diversity scenarios with respect to location and environmental conditions along with exploitation of deep learning algorithms for improved performance and generalization.'

Problem Statement

Establish the validity domain of a single zero-transfer-sample global multi-unit calibration law for low-cost particulate matter (PM2.5) multisensor devices — one machine-learning model trained on aggregated data from a small set of co-located units and applied to all other units — as deployment diversity grows, and determine how that validity is maintained in operation. Concretely, with respect to the methodology validated by De Vito et al. in a single geographic area (Portici, Italy) over winter and summer campaigns: (1) characterize how the calibration's accuracy transfers to locations with differing aerosol composition, emission source mix, and microclimate — in particular whether a law derived in one origin environment can be applied zero-sample elsewhere, and what training-data diversity supports cross-location validity; (2) quantify how progressive sensor aging and poisoning degrade the stability and validity duration of a single global calibration law, and determine the model-update cadence and mechanism needed to keep the network within indicative-measurement accuracy over years of operation; (3) develop and validate systematic strategies to detect and remediate performance failures under extreme seasonal or deployment-regime shifts, such as the observed episodes of negative R2R^2 under very different humidity and concentration distributions; and (4) quantify the operational tradeoff between on-board model updating (firmware-update logistics) and edge/cloud centralized updating (data-transfer volume and energy cost) for deployed networks operating under such a global calibration. These four components jointly define a single objective — determining when one shared calibration law suffices for a heterogeneous, long-lived PM sensor network and how its accuracy is sustained — and a complete resolution must address all four, with no single component independently sufficient. Determining this validity domain includes establishing the error metric and threshold applied to held-out units under which a calibration law counts as valid, and the environments, deployment durations, and device conditions over which that validity holds.

References

  1. Saverio De Vito, Gerardo D'Elia, Sergio Ferlito, Girolamo Di Francia, Miloš D. Davidović, Duška Kleut, Danka Stojanović, Milena Jovaševic-Stojanović, 'A Global Multiunit Calibration as a Method for Large-Scale IoT Particulate Matter Monitoring Systems Deployments', IEEE Transactions on Instrumentation and Measurement 73, 1–16 (2024), DOI: 10.1109/TIM.2023.3331428, preprint arXiv:2310.18118, https://arxiv.org/abs/2310.18118
  2. Gerardo D'Elia, Matteo Ferro, Paolo Sommella, Sergio Ferlito, Saverio De Vito, Girolamo Di Francia, 'Concept Drift Mitigation in Low-Cost Air Quality Monitoring Networks', Sensors 24(9):2786 (2024), DOI: 10.3390/s24092786, https://doi.org/10.3390/s24092786
  3. Petar Zhivkov, Stefka Fidanova, 'Machine Learning Calibration Transfer for Low-Cost Air Quality Sensors: Distance-Based Uncertainty Quantification in a Hybrid Urban Monitoring Network', Atmosphere 17(4):335 (2026), DOI: 10.3390/atmos17040335, https://doi.org/10.3390/atmos17040335
  4. Saverio De Vito, Gerardo D'Elia, Sergio Ferlito, Elena Esposito, Gabriele Piantadosi, Girolamo Di Francia, 'Remote Calibration strategies for Low Cost Air Quality Multisensors: a performance comparison', Proceedings of ISOEN 2024, DOI: 10.1109/ISOEN61239.2024.10556107, https://doi.org/10.1109/ISOEN61239.2024.10556107
  5. Saverio De Vito, Antonio Del Giudice, Gerardo D'Elia, Elena Esposito, Grazia Fattoruso, Sergio Ferlito, Fabrizio Formisano, Giuseppe Loffredo, Ettore Massera, Paolo D'Auria, Girolamo Di Francia, 'Future Low-Cost Urban Air Quality Monitoring Networks: Insights from the EU's AirHeritage Project', Atmosphere 15(11):1351 (2024), DOI: 10.3390/atmos15111351, https://doi.org/10.3390/atmos15111351
  6. Karoline K. Barkjohn, Brett Gantt, Andrea L. Clements, 'Development and application of a United States-wide correction for PM2.5 data collected with the PurpleAir sensor', Atmospheric Measurement Techniques 14, 4617–4637 (2021), DOI: 10.5194/amt-14-4617-2021, https://doi.org/10.5194/amt-14-4617-2021
  7. Peer Nowack, Lev Konstantinovskiy, Hannah Gardiner, John Cant, 'Machine learning calibration of low-cost NO2 and PM10 sensors: non-linear algorithms and their impact on site transferability', Atmospheric Measurement Techniques 14, 5637–5655 (2021), DOI: 10.5194/amt-14-5637-2021, https://doi.org/10.5194/amt-14-5637-2021
  8. Carl Malings, Rebecca Tanzer, Aliaksei Hauryliuk, Sriniwasa P. N. Kumar, Naomi Zimmerman, Levent B. Kara, Albert A. Presto, R. Subramanian, 'Development of a general calibration model and long-term performance evaluation of low-cost sensors for air pollutant gas monitoring', Atmospheric Measurement Techniques 12, 903–920 (2019), DOI: 10.5194/amt-12-903-2019, https://doi.org/10.5194/amt-12-903-2019
  9. Priyanka deSouza, Karoline Barkjohn, Andrea Clements, Jenny Lee, Ralph Kahn, Ben Crawford, Patrick Kinney, 'An analysis of degradation in low-cost particulate matter sensors', Environmental Science: Atmospheres 3, 521–536 (2023), DOI: 10.1039/d2ea00142j, https://doi.org/10.1039/d2ea00142j
  10. Priscilla Adong, Engineer Bainomugisha, 'Transferability of Machine Learning-Based Calibration Models for Low-Cost PM2.5 Sensors Across Heterogeneous African Cities', Meteorological Applications 33(4) (2026), DOI: 10.1002/met.70228, https://doi.org/10.1002/met.70228
  11. Jingzhuo Zhang, Li Bai, Na Li, Yu Wang, Yibing Lv, Yaolong Shi, Chao Yang, Chi Xu, 'Low-Cost Particulate Matter Mass Sensors: Review of the Status, Challenges, and Opportunities for Single-Instrument and Network Calibration', ACS Sensors 10(5), 3207–3221 (2025), DOI: 10.1021/acssensors.4c03293, https://doi.org/10.1021/acssensors.4c03293