Source-linked AI summary
Low-Cost Sensor Calibration for Indoor Air Quality Monitoring: A Dataset, Evaluation Scenarios, and a Lightweight Model
Jinyong Yun, Seokho Ahn, Hyungjin Kim, Sungbok Shin, Young-Duk Seo
TL;DR
Strict pairwise calibration requires a reference sensor at each deployment location, a costly fit for indoor monitoring across spatially and temporally heterogeneous spaces. The paper builds a six-month, five-location dataset and four deployment scenarios, then proposes a lightweight temporal calibration model. It reports strong performance across the scenarios at low edge-inference cost, while noting limits from its single-building data and event annotation resolution.
Problem
Strict pairwise calibration requires a co-located reference sensor at every deployment location, while practical monitoring needs calibration across locations with few references and robustness to spatial and rapid environmental changes.
Source: §1 · p. 1Method
The paper constructs a six-month, five-location indoor dataset with event metadata, defines four deployment scenarios, and proposes a lightweight temporal calibration model.
Source: §1 · pp. 1–2Results
The model ranked first or second across deployment scenarios with low edge-inference cost, and the conclusion reports strong performance across all four scenarios.
Source: §8.2 · p. 7, §9 · p. 7Takeaways & Limitations
The benchmark and model address calibration across locations, time periods, and event conditions while maintaining low edge-inference cost.
Source: §8.2 · p. 7, §9 · p. 7The dataset comes from five locations within one building, and 30-minute event annotations support event-conditioned periods rather than precisely aligned event responses.
Source: §8.3 · p. 7
Abstract
from arXiv · showhide
Low-cost sensors enable scalable indoor air quality monitoring but require calibration because of nonlinear distortions, noise, and temporal drift. The conventional strict pairwise calibration setting requires a co-located reference sensor at each deployment location and does not account for spatial and temporal heterogeneity. To address these limitations, we introduce a six-month dataset comprising multivariate indoor air-quality measurements from low-cost and reference sensors with contextual metadata collected at five locations. Using this dataset, we define four evaluation scenarios. The reference-efficient and location-transfer scenarios evaluate spatial generalization, whereas the long-term drift and event-conditioned scenarios assess robustness to gradual and abrupt distribution shifts. Based on these scenarios, we derive design requirements and propose a lightweight temporal model that combines input-window compression with residual temporal and feature fusion. Experiments show strong calibration performance across all four scenarios with low edge-inference cost.
1 Introduction
Indoor air quality varies across locations and over time, while dense deployment of high-accuracy reference sensors is costly. The paper addresses this tension with a multi-location dataset, deployment-oriented evaluation scenarios, and a lightweight calibration model.
1 Introduction
- Spatial heterogeneity and short-term events can make a small sensor set miss local changes and the overall air-quality distribution.Dense high-accuracy sensing is therefore desirable, but costly to deploy throughout an indoor space.Source: §1 · p. 1
- Strict pairwise calibration requires a co-located reference sensor at every low-cost sensor location, making broad indoor monitoring prohibitively expensive.The paper seeks calibration across locations using only a few reference sensors, robust to rapid environmental changes and spatial heterogeneity.Source: §1 · p. 1
- The benchmark combines a six-month, five-location dataset, four scenarios for limited references and spatial or temporal shifts, and a lightweight model.The dataset pairs low-cost and reference measurements with event metadata, including door and window states.Source: §1 · pp. 1–2
2 Related Work
The related work reviews low-cost sensor calibration methods and datasets, identifying separate gaps in indoor dataset coverage and in jointly addressing deployment challenges.
2 Related Work
- The related work surveys prior low-cost sensor calibration methods and the datasets used in calibration studies.This framing spans both sensor calibration datasets and models.Source: §2 · p. 2
2.1 Sensor Calibration Datasets
- None of the compared datasets combines synchronized measurements across multiple indoor locations with long-term coverage and event metadata.The proposed dataset pairs low-cost and reference measurements at five indoor locations over six months, with door, window, and occupancy metadata.Source: §2.1 · p. 2
2.2 Sensor Calibration Models
- Existing methods address deployment challenges separately; the proposed model combines temporal compression and residual fusion for limited references, unseen locations, distribution shifts, and low edge-inference cost.Prior examples include temporal and long-term outdoor calibration, efficient edge deployment, and unsupervised calibration without paired references.Source: §2.2 · p. 2
3 Indoor Sensor Calibration Scenarios
The four scenarios relax two assumptions of strict pairwise calibration: a reference at every deployment location and stable calibration relationships over time. They separate spatial generalization from gradual drift and event-driven shifts.
3 Indoor Sensor Calibration Scenarios
- S1 reference-efficient calibration uses limited reference measurements across locations observed during training, while S2 transfers calibration to a target location excluded from training.For S2, target references are reserved for evaluation; it therefore requires stronger spatial generalization than S1.Source: §3 · p. 2
- Figure 1 compares standard calibration with the four deployment scenarios by their relaxed assumptions, calibration regimes, and world settings.The scenarios are reference-efficient, location-transfer, long-term drift, and event-condition.Source: §3 · p. 2, Figure 1 · p. 3
- S3 long-term drift tests a model trained initially and evaluated later, after gradual sensor-specific response changes within a closed-world setting.S1–S2 relax reference coverage; S3–S4 relax temporal stability.Source: §3 · pp. 2–3
- S4 event-condition evaluates abrupt, externally induced distribution shifts, whose magnitudes and delays vary by location.Door or window changes are examples; this is an open-world setting requiring handling of temporal and location-specific responses.Source: §3 · p. 3
4 Indoor Sensor Calibration Dataset
The dataset pairs low-cost and reference sensor measurements across five indoor locations, adds contextual metadata, and supports tests of spatial and temporal deployment shifts.
4.1 Data Collection
- The six-month dataset collects synchronized, paired low-cost and reference measurements across five locations in one university laboratory.Five environmental variables are recorded at one-second intervals, and the processed dataset contains 6,102,195 paired observations across 25 tasks.Source: §4 · p. 3, §4.1 · p. 3
4.2 Metadata
- Door and window states, occupant counts, timestamps, and business-hour status provide contextual metadata for event-conditioned evaluation and further analysis.Door and window states are recorded every 30 minutes; the benchmark uses those variables to identify event-conditioned periods.Source: §4.2 · p. 3
4.3 Preprocessing
- After sanity filtering and one-minute resampling, channel means differ across locations, indicating placement- and environment-dependent calibration relationships.Before imputation, valid-value ratios ranged from 82.9% to 89.2% across locations and channels.Source: §4.3 · pp. 3–4, Table 2 · p. 3
4.4 Support for Deployment Scenarios
- Aligned measurements at all five locations support controlled comparisons of sparse-reference S1 and held-out-location S2 calibration.Location-specific channel statistics, particularly at the South node near a window and the North node near an entrance, reflect spatial heterogeneity.Source: §4.4 · p. 4
- Continuous six-month records support tracking calibration over time for S3, while node placements and repeated door and window changes provide event episodes for S4.The event episodes support evaluation of abrupt, delayed, and location-dependent sensor responses.Source: §4.4 · p. 4
5 Model for Indoor Sensor Calibration
The model is designed to meet five requirements spanning limited reference supervision, unseen-location transfer, temporal shifts, and edge efficiency. It uses temporal compression and residual temporal-feature fusion to support calibration across these conditions.
5.1 Design Requirements
- The five requirements cover sharing limited reference supervision, transfer to unseen locations, gradual drift, abrupt event responses, and lightweight edge inference.These correspond to R1–R5 across the four evaluation scenarios and resource-constrained deployment.Source: §5.1 · p. 4
5.2 Model Design
- An attention-free MLP combines input-window temporal compression with residual temporal-feature fusion to address R1–R4 while keeping inference and memory costs low for R5.The design operates in a compact temporal space for edge deployment.Source: §5.2 · p. 4
- The model compresses an L-step input window into R temporal representations, with R ≪ L, to reduce later computation and favor transferable temporal patterns.It omits location-specific information and estimates the reference measurement at the window’s final time step.Source: §5.2 · pp. 4–5
- Temporal fusion captures trends, delayed responses, and rapid changes; feature fusion learns nonlinear dependencies, while residual connections preserve the sensor trajectory during correction.Batch normalization stabilizes intermediate representations, and the temporal input supports modeling event direction and duration when the transition and response fall within the window.Source: §5.2 · p. 5
- Initial compression and small hidden dimensions reduce parameter count and inference cost, supporting real-time calibration on resource-constrained edge devices.The output head aggregates latent temporal and feature dimensions into a calibrated estimate for the final time step.Source: §5.2 · p. 5
6 Experimental Settings
The experiments assess calibration across standard conditions and four scenarios, including limited references, unseen locations, long-term drift, and event-conditioned shifts. They compare accuracy and computational cost using defined protocols, baselines, and implementation settings.
6.1 Evaluation Objectives
- The evaluation tests whether the dataset and four scenarios expose challenges missed by strict pairwise evaluation, and whether calibration remains strong under spatial and temporal shifts at low computational cost.Tables 3–7 report calibration results, while Figure 3 evaluates edge latency and model size.Source: §6.1 · p. 5
6.2 Dataset Usage and Evaluation Protocol
- S1 trains with reference supervision at one anchor and evaluates non-anchor locations; S2 transfers each location-trained model to the other four without adaptation.S1 and S2 split each location chronologically into 70% training, 15% validation, and 15% test data.Source: §6.2 · p. 5, Table 4 · p. 6, Table 5 · p. 6
- S3 measures relative RMSE increase from the first to last of four deployment periods; S4 evaluates event-period CO₂ and temperature RMSE at window- and door-related locations.S3 uses chronological 50%/10%/40% training, validation, and deployment splits and averages results across five locations.Source: §6.2 · p. 6, Table 6 · p. 6, Table 7 · p. 6
6.3 Baselines
- Baselines include attention-based, linear/MLP-based, and calibration-specific methods; Veli is excluded because it is unsupervised rather than part of the supervised benchmark.The compared calibration-specific methods are SenDaL and TESLA.Source: §6.3 · p. 6
6.4 Evaluation Metrics
- Accuracy is measured by RMSE in each variable’s physical unit, while efficiency uses per-sample inference time and peak inference memory.Inference time is averaged over repeated runs after graph warm-up; peak memory is measured relative to the pre-inference baseline.Source: §6.4 · p. 6
6.5 Implementation Details
- Models share a TensorFlow 2.14 training protocol and processor-based efficiency setting; embedded evaluation converts trained models to TensorFlow Lite and deploys them on an Arduino Nano 33 BLE Sense.The common training protocol is adopted from Ahn et al. (2025) and follows the iTransformer configuration.Source: §6.5 · p. 6
7 Experimental Results
The proposed model performs strongly under co-located reference supervision and across deployment-oriented calibration scenarios. Its results also show that rankings vary by reference coverage, location, time, and event condition, while edge inference remains efficient.
7.1 Overall Calibration Performance
- The proposed model achieved the lowest RMSE for all three outdoor variables and four of five indoor variables, ranking second for indoor temperature under co-located reference supervision.Source: §7.1 · p. 6
7.6 R5. Efficiency and On-device Performance
- On Arduino Nano 33 BLE Sense, the proposed model was faster than TESLA at every tested input window and faster than DLinear for long windows, with a FlatBuffer size between theirs.Source: Figure 3 · p. 7, §7.6 · p. 7
7.2 R1. Reference-Efficient Calibration
- Under reference-efficient calibration, the proposed model had the lowest RMSE for CO, CO2, humidity, and PM1, while model rankings differed between strict pairwise and limited-reference settings.Source: §7.2 · p. 7
7.3 R2. Location-Transfer Calibration
- For unseen-location transfer, the proposed model's RMSE was higher than in S1 for all five variables; its S1-to-S2 increase ranged from 1.3% for PM1 to 61.2% for CO.Source: §7.3 · p. 7
7.4 R3. Long-Term Calibration Drift
- Across six months, the proposed model had the lowest relative RMSE increase for all five variables, but its RMSE still rose by 24.6–57.2%.Source: §7.4 · p. 7
7.5 R4. Event-Conditioned Calibration
- The best model varied across event conditions: the proposed model ranked first or second in all four cases, while TESLA and SenDaL each led one condition.Source: §7.5 · p. 7
8 Discussion and Limitations
Model rankings changed across deployment conditions, which the authors interpret as distinct calibration capabilities that aggregate evaluation can hide. The dataset's single-building scope and event-metadata timing constrain what its evaluations represent.
8.1 Dataset and Scenario
- Rankings changed with reference coverage, target-location exposure, deployment time, and event conditions, revealing distinct calibration capabilities that standard aggregate evaluation can hide.Source: §8.1 · p. 7
8.2 Model and Deployment
- The model ranked first or second across deployment scenarios with low edge-inference cost, but its 24.6–57.2% S3 degradation motivates periodic recalibration or online adaptation.Source: §8.2 · p. 7
8.3 Single-Environment Scope and Data Fidelity
- The dataset covers five locations in one building, and fixed-interval metadata supports event-conditioned periods rather than precisely aligned event responses.Source: §8.3 · p. 7
9 Conclusion
The paper presents a deployment-oriented benchmark for indoor low-cost sensor calibration. It combines a six-month, five-location dataset and four scenarios with a lightweight temporal model, reporting strong performance across scenarios at low edge-inference cost.
9 Conclusion
- The benchmark uses a six-month dataset from five locations and four scenarios covering limited reference availability, unseen locations, long-term drift, and event-conditioned shifts.Source: §9 · p. 7
- The authors derive design requirements from those scenarios and propose a lightweight temporal model for robust and efficient calibration.Source: §9 · p. 7
- Experiments show strong performance across the four scenarios with low edge-inference cost; future work will extend the dataset and explore online adaptation for long-term drift.Source: §9 · p. 7