Real-time Calibration Model for Low-cost Sensor in Fine-grained Time Series

Abstract

Precise measurements from sensors are crucial, but the data are usually collected from low-cost, low-tech systems, which are often inaccurate. Thus, they require further calibrations. To that end, we first identify three requirements for effective calibration under practical low-tech sensor conditions. Based on the requirements, we develop a model called TESLA, Transformer for effective sensor calibration utilizing logarithmic-binned attention. TESLA uses a high performance deep learning model, Transformers, to calibrate and capture non-linear components. At its core, it employs logarithmic binning, to minimize attention complexity. TESLA achieves fast and consistent real-time calibration, even with longer sequences and finer-grained time series in hardware-constrained systems. Experiments show that TESLA outperforms existing novel deep learning and newly crafted linear models in accuracy, calibration speed, and energy efficiency.

Environment Requirements

• See requirements.txt

File Directory Structure

• dataset: Contains experimental datasets

  • ./finedust_concentration_alldrop_firstcol_dataset/{region}_{sensor}_FirstCol.csv
  • Please download via: https://drive.google.com/file/d/1sQpUlwFw2H4dyyXfvxcMKpTQhli8rhgj/view?usp=sharing

• lbt: Contains the experimental package

  • ./lbt

• result: Stores TESLA weights and experimental results during training

  • ./{save_path}/{region}_{pm}_{window_size}.h5 - TESLA best weights for the validation dataset
  • ./{save_path}/{region}_{pm}_{window_size}.log - Loss history during training
  • ./{save_path}/{region}_{pm}_{window_size}.png - Plot for the test dataset using the best weights

Command Examples

• Training & Evaluation Process
  Evaluation scalar results are displayed in the console, and plots are saved to {save_path}.

  python main.py --save_path ./model --pm 10 --window_size 360 --gpu 0 --seed 2024
  

• Only Evaluation Process With model weights loaded from {save_path}.

  python main.py --save_path ./model --pm 10 --window_size 360 --gpu 0 --seed 2024
  

Command-line Arguments

• --save_path Default: "./model"
  Description: Directory path to save the model and results.

• --pm
  Default: 10 Description: Select from (1, 2.5, 10). Parameter for calibration.

• --window_size
  Default: 360
  Description: Window size for data processing. Recommended values are (15, 60, 360, 720, 1440).

• --max_sensor_scale Default: 300
  Description: Maximum sensor scale value.

• --epoch Default: 10
  Description: Number of training epochs.

• --batch_size Default: 32
  Description: Batch size for training.

• --learning_rate Default: 1e-4
  Description: Learning rate for the optimizer.

• --shuffle_sensor_size Default: 4
  Description: Number of shuffle sizes between sensors during training.

• --gpu Default: None Description: GPU device number for training. (e.g., 0). If not specified, CPU will be used.

• --seed Default: 2024 Description: Random seed for reproducibility.

Citation

If you would like to cite this paper, please make a copy of the following text. Thank you🤣

@article{ahn2025tesla,
      title={Real-Time Calibration Model for Low-Cost Sensor in Fine-Grained Time Series},
      author={Ahn, Seokho and Kim, Hyungjin and Shin, Sungbok and Seo, Young-Duk},
      journal={Proceedings of the AAAI Conference on Artificial Intelligence},
      year={2025},
      month={Apr.},
      pages={3-11}
      volume={39},
      number={1},
      url={https://ojs.aaai.org/index.php/AAAI/article/view/31974},
      DOI={10.1609/aaai.v39i1.31974},
}