New AI Model Fixes Broken Sensors Faster Than Current Methods
Researchers have developed a machine learning architecture that automatically recalibrates faulty sensors with greater accuracy than existing approaches. The technique could reduce costs and downtime across industries relying on sensor networks—from air quality monitoring to manufacturing—by eliminating manual calibration work.
Originaltitel: Latent-Space Controlled VAE (LaC-VAE): An ML Architecture for Sensor Calibration
<p>In this paper we introduce Latent-space Controlled Variational Autoencoder (LaC-VAE), a novel, theoretically-grounded machine learning architecture that reframes sensor calibration as a joint representation learning and prediction problem. Our core architectural innovation is the explicit constraint of the latent space of the LaC-VAE to directly represent the calibrated prediction-output of the model. This forces the encoder network to act as the calibrator while the decoder acts as a sensor simulator. We benchmark this novel calibration model against five established models, including CVAE, Regression-VAE, and traditional methods (RFR, SVR) and LSTM. The models were validated on two real-world datasets. The first comprising of air quality data from the Seoul Metropolitan and the second, a novel dataset, from a Cape Town Home. Our results demonstrate that the LaC-VAE achieves superior predictive performance, recording the lowest Mean Absolute Error (MAE) and highest R2 values across the primary comparative analyses. Additionally, the LaC-VAE consistently exhibits the lowest Kullback-Leibler and Jensen-Shannon divergence metrics indicating a strong ability to match the calibration data statistically. Lastly, by enabling us to split sensing into a simulator and a calibrator part, LaC-VAE would also facilitate in the development of digital twins of processes. We further analyse the drift performance of all the models using a Cumulative Sum (CUSUM) analysis.</p>