Enhancing sensor systems through sensor drift identification and compensation using Jensen-Shannon divergence and CTGAN

Abstract

Many modern gas detection systems use machine learning (ML) to achieve consistent accuracy, even in the face of challenges like sensor drift and class imbalance. Sensor drift refers to the gradual decline in a sensor’s performance over time, which makes it difficult for ML models to correctly identify gas types. Additionally, many gas detection datasets are imbalanced, meaning that some gas types are underrepresented, which affects the performance of the models. To address these challenges, two important techniques are used, in this work, namely, sensor drift detection and data augmentation. However, current methods often struggle to handle the complexity of gradual or non-linear drift along with the imbalanced datasets. Moreover, traditional data augmentation techniques are often not suitable for tabular data, which is commonly used in sensor-based systems. This thesis introduces a novel approach that uses Jensen-Shannon (JS) divergence to detect and measure sensor drift, and Conditional Tabular Generative Adversarial Networks (CTGAN) to generate synthetic data that improves dataset balance and compensates for class imbalance. JS divergence allows us to precisely identify sensor drift by comparing the probability distributions of sensor data over time, helping us better understand how drift affects classification accuracy. CTGAN is used to create high-quality synthetic tabular data, ensuring better representation of minority gas classes. Our research focuses on a gas sensor dataset collected over 36 months, which includes multiple batches of sensor readings. We use JS divergence to detect and measure drift, and CTGAN to generate synthetic data that addresses both sensor drift and class imbalance. This combination of JS divergence and CTGAN is applied for the first time to improve the accuracy of gas detection systems. By evaluating ML models such as SVM and MLP, we demonstrate significant improvements in classification accuracy, with gains of up to 20% in some batches. Our findings highlight the effectiveness of this approach in addressing both sensor drift and data imbalance, contributing to a better understanding of how to improve gas detection systems. In summary, this thesis makes important contributions to the fields of sensor drift detection and data augmentation in gas detection systems. It presents a new methodology for quantifying drift using JS divergence and addresses class imbalance with CTGAN, enhancing the accuracy and reliability of ML models in sensor-based applications.

Date	18 Apr 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Andy Shih (Supervisor) & Kuljeet Kaur (Co-supervisor)