CaliFormer: Leveraging Unlabeled Measurements to Calibrate Sensors with Self-supervised Learning

Abstract

Accurate calibration of low-cost sensors is critical for improving their potential in environmental monitoring. State-of-the-art (SOTA) methods based on supervised learning commonly calibrate sensor measurements with ground truth from the immediate past or future. However, these techniques rely heavily on labeled data which is challenging to obtain in real-world scenarios. Thus, this paper introduces CaliFormer, a novel representation learning model using self-supervised learning to extract time- and spatial-invariant knowledge from unlabeled measurements. Moreover, we propose pre-training enhancements and model architecture modifications to help train CaliFormer. We then fine-tune the calibration model with the learned representations, which is supervised by limited labeled data. Finally, we comprehensively evaluate our calibration model with a dataset collected by low-cost sensors. Results show that our model outperforms other SOTA calibration methods significantly.