Research map / Time series / Long-term forecasting
Real-Time Calibration Model for Low-Cost Sensor in Fine-Grained Time Series
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TESLA, a small Transformer for calibrating low-cost fine-grained sensors that uses logarithmic binning of tokens, local+global embeddings, and a linear feature-wise aggregator to get Transformer-level accuracy at near-linear-model speed on microcontrollers.
Problem
Low-cost, low-tech sensors give inaccurate fine-grained (minute-level) readings; existing calibration models either lack accuracy (linear), or are too computationally heavy for IoT hardware (Transformers), and time-series forecasters do not transfer directly to single-sensor calibration.
Approach
Reduce the sequence to about log2 N tokens via logarithmic binning (recent tokens kept fine, older ones aggregated with learnable weights), add a global embedding for single univariate series, and replace the token-wise feed-forward network with a feature-wise linear aggregator, yielding O(log^2 N) attention cost.
Key results
- TESLA achieves the best average RMSE for PM10 (14.75 vs 15.46 for PatchTST and 15.65 for iTransformer) across Antwerp, Oslo and Zagreb with N=360.
- Ablation: log-scaled binning with local+global embedding and linear aggregator gives an average gain of +6.88% over the full Transformer baseline in RMSE.
- On High-distribution case-study windows, TESLA improves over iTransformer by 34.86%, and averages 3.81% improvement in the case study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TESLA instead of PatchTST | Sensor calibration · SensEURCity PM10 average, N=360 | RMSE: 15.46 -> 14.75 | Helped |
| Use TESLA instead of iTransformer | Sensor calibration · SensEURCity PM10 average, N=360 | MAE: 8.91 -> 8.57 | Helped |
| Use TESLA instead of Informer | Sensor calibration · SensEURCity PM2.5 average, N=360 | RMSE: 6.31 -> 6.16 | Helped |
| Use TESLA instead of PatchTST | Sensor calibration · SensEURCity PM2.5 average, N=360 | MAE: 3.50 -> 3.23 | Helped |
| Use TESLA instead of Transformer | Sensor calibration · SensEURCity PM1 average, N=360 | RMSE: 3.42 -> 3.19 | Helped |
| Use TESLA instead of Transformer | Sensor calibration · SensEURCity PM1 average, N=360 | MAE: 1.83 -> 1.71 | Helped |
| Replace uniform binning with logarithmic binning (ablation) | TESLA · SensEURCity, N=360 | Avg. gain (RMSE): -1.20% -> -0.45% | Helped |
| Add global embedding (ablation) | TESLA with logarithmic binning and feed-forward aggregation · SensEURCity, N=360 | Avg. gain (RMSE): -0.45% -> +1.90% | Helped |
| Replace feed-forward aggregation with feature-wise linear aggregation (ablation) | TESLA with logarithmic binning and local+global embedding · SensEURCity, N=360 | Avg. gain (RMSE): +1.90% -> +6.88% | Helped |
Surprising findings
- Newly crafted linear models (Linear, DLinear, NLinear) underperformed Transformer-based models on calibration, contrary to their strong showing in forecasting.
- Minimal gap between RMSE and MAE, attributed to a 'mean trap' where limited variation makes these metrics insensitive to calibration differences.
Limitations
- Evaluated only on PM sensor data from SensEURCity; other sensor types such as temperature and humidity are untested.
- Reference sensor precision may be insufficient, and the static train/test setup requires recalibration or drift correction.
- Relies on a single high-cost reference sensor for labels.
Experimental setup
Datasets: SensEURCity
Benchmarks: SensEURCity
Metrics: RMSE MAE FLOPS memory footprint parameter count inference speed on Arduino Nano 33 BLE Sense Flatbuffer size
Baselines: Linear NLinear DLinear Transformer Informer PatchTST iTransformer
Builds on: PatchTST iTransformer DLinear Informer
Closest papers by meaning
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.83)
- SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention (ICML 2024 · similarity 0.82)
- Testing Calibration in Nearly-Linear Time (NeurIPS 2024 · similarity 0.81)
- Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention (NeurIPS 2026 · similarity 0.81)
- Hierarchical State Space Models for Continuous Sequence-to-Sequence Modeling (ICML 2024 · similarity 0.81)
- A Scalable Multi-Task Model for Virtual Sensors (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.