Variational Autoencoder for Calibration: A New Approach

Travis Barrett, Amit Kumar Mishra, Joyce Mwangama · 2025

Read Variational Autoencoder for Calibration: A New Approach by Travis Barrett, Amit Kumar Mishra, Joyce Mwangama (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-00475.txt.

Introduction

Travis Barrett, Amit Kumar Mishra, and Joyce Mwangama propose a new use for variational autoencoders: calibrating sensors. An autoencoder compresses data into a compact latent code and rebuilds it; their twist trains that latent space to serve directly as the calibration output. One model then does two jobs at once, reconstructing raw readings and emitting corrected values.

A proof of concept on an existing multi-sensor gas dataset shows the idea works. Outputs from both heads, the calibration output and the reconstruction, stayed statistically similar to their respective ground truths. The authors discuss how future testing should proceed and how the approach could expand. If it holds up, calibration could ride along inside the same network that already processes sensor streams.

Catalog record

CREATORTravis Barrett, Amit Kumar Mishra, Joyce Mwangama
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-00475

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Variational Autoencoder for Calibration: A New Approach Thanks: This work has been supported with funding from Sentech Soc Ltd, South Africa. Travis Barrett, Amit Kumar Mishra, Joyce Mwangama Abstract In this paper we present a new implementation of a Variational Autoencoder (VAE) for the calibration of sensors. We propose that the VAE can be used to calibrate sensor data by training the latent space as a calibration output. We discuss this new approach and show a proof-of-concept using an existing multi-sensor gas dataset. We show the performance of the proposed calibration VAE and found that it was capable of performing as calibration model while performing as an autoencoder simultaneously. Additionally, these models have shown that they are capable of creating statistically similar outputs from both the calibration output as well as the reconstruction output to their respective truth data. We then discuss the methods of future testing and planned expansion of this work. I Introduction The use of low-cost sensors, or any sensors for that matter, requires a consistent management of sensor drift and calibration to achieve accurate measurements. This is particularly important when using low-cost sensors as they suffer from a range of challenges depending on the different sensor technologies. For air quality sensors, this is especially common as these devices are influenced by various factors . Cross-sensitivities, environmental factors and sensor degradation cause these devices to drift and become inaccurate over time [ 1 , 2 ] . This makes calibration critically important for optimal sensor performance. As stated by Liang and Daniels [ 3 ] , failure to invest in calibration may leave large uncertainties in retrieving reliable low-cost sensor data that further hinders its broader applications. Machine Learning (ML) as a method for calibration has been widely discussed in open literature. Implementations of ML have been used to calibrate data in weather measurement stations for low-cost temperature sensors [ 4 ] , for low-cost particulate matter sensors [ 5 , 6 ] , and many other implementation such as increasing the speed of calibration of Metal Oxide Semiconductor (MOS) sensors using transfer learning as shown by Robin et al. [ 7 ] . In the work done by Mu and Chen [ 8 ] , they show a calibration model building method based on semi-supervised learning is presented for in situ measurements of process components based on near-infrared (NIR) spectra where there are insufficient labeled samples. This use of a Conditional Variational Autoencoder (CVAE) allows for the generation of NIR spectra to increase the size of training data and when combined with the S2-LN calibration model, achieve a higher labeling accuracy than the other compared methods. Variational Autoencoders have been shown to function as calibration models as well as showing promise for drift detection. In the work done by Hossain et al. [ 9 ] , a Variational Autoencoder (VAE) is presented for performing drift compensation and error detection. This implementation allows for processing data in parallel to enabling the Variational Autoencoder to compare the data to the distribution and correct the output accordingly. We also see the VAE used in the process of developing soft sensors, as seen in the works of Dai et al. [ 10 ] and Shen et al. [ 11 ] . These works make use of the ability of the VAE to deal with non-linearity. Mihailescu [ 12 ] showed the use of a weakly supervised VAE for domain adaptation. It is clear that the VAE model has promising properties for calibration, error detection and domain adaptation, shown in current literature. In this paper we present a new use for the Variational Autoencoder (VAE) architecture as a method for sensor calibration. We propose that the VAE can be used to calibrate sensor data by training the latent space as a calibration output. This allows for the facilitation of calibration of sensor data while retaining the autoencoding functionality. In this paper we show a proof of concept for this approach to sensor calibration with the use of the dataset presented in De Vito et al. [ 13 ] . As shown in our previous work [ 14 ] , Machine Learning (ML) could be a viable solution to sensor calibration. The intention of this work with new data and new model types is to find different approaches to this problem of sensor calibration with machine learning. Our contributions can be summarized as follows: • Present new VAE for calibration. • Use the latent space of the VAE as a calibration output. • Demonstrate the model on an existing dataset. The remainder of the paper is organized in the following way. We will discuss the dataset used for the experimentation proposed in this paper. We will then describe the VAE and the proposed VAE for calibration. Following this we will present the methodology of experimentation used in this body of work. The results of the experimentation set out in the preceding section will then be shown. A discussion into the results will be presented and finally conclusions will be drawn about the work presented. With this conclusion, improvements to this work and future work will be presented. II Dataset This section describes the dataset used for the work in this paper. It gives the background information of the dataset and then describes the content of the dataset as described by the creators. II-A Information Created by De Vito et al. [ 13 ] , the dataset contains the responses of a multi-sensor gas sensing system deployed in an Italian city. This dataset includes these responses captured in the form of hourly averages as well as the recorded gas concentrations from a reference analyzer. This dataset has been used by many publications around sensor calibration and low-cost sensor performance. Some of these papers include further work by De Vito et al. [ 15 ] as well as the work by Yu et al. [ 16 ] . As this is a extensively used dataset it was a natural choice for our testing. II-B Collected Data The dataset is described as having 9358 instances of hourly averaged responses from an array of 5 metal oxide chemical sensors (MOS). These readings were taken in the field in a polluted area of an Italian city at road-level. The data was collected with ground truth hourly averaged concentrations for CO, Non-methane Hydrocarbons, Benzene, Total Nitrogen Oxides (NOx) and Nitrogen Dioxide (NO 2 ) and were provided by a co-located reference certified analyzer. This data recording took place from March 2004 to February 2005. Evidence of cross-sensitivities as well as both concept and sensor drift are present as described in the work done by De Vito et al. [ 13 ] . III Variational Autoencoder This section discusses the Variational Autoencoder as well as the proposed calibration from latent space Variational Autoencoder. III-A Variational Autoencoder The Variational Autoencoder is an autoencoder machine learning model. It is characterized as having two distinct networks in the model [ 17 ] . The first is the encoder network which takes the input data and maps it to a lower-dimensional latent space. The second being the decoder, which maps the lower-dimensional latent space back to the input data. The latent space of

Read the complete text