Scholarlink

<p>Early detection of chemical gases plays a critical role in public health, environmental safety and industrial monitoring. The need for faster and more accurate gas detection methods is increasing, especially in sensitive industrial applications, and machine learning (ML) algorithms are increasingly utilized by researchers to improve gas sensor performance by enabling earlier prediction of hazardous concentrations. This study presents an application-driven hybrid intelligent framework that first leverages a beta-regularized conditional variational autoencoder (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\beta \)</EquationSource> </InlineEquation>-cVAE) to enrich the limited time-series exposure dataset obtained through ammonia (NH<sub>3</sub>) gas exposure in an experimental setup. The augmented data are subsequently used to train a Gated Recurrent Unit network (GRU) to predict the maximum gas concentration using only the early sensor readings, thereby eliminating the need to wait for full gas exposure. The proposed hybrid method achieves accurate early prediction of the peak concentration across various concentration levels, enabling a significant improvement of sensor response time. Building upon a previously proposed generative-predictive framework for hydrogen sensing, this study adapts and extends the methodology to NH<sub>3</sub> detection using a <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\beta \)</EquationSource> </InlineEquation>-regularized cVAE-GRU architecture under limited data conditions. By modeling temporal patterns in sensor response data, the framework is potentially adaptable to other time-series sensing applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generative augmentation and early time-series prediction of gas concentrations: a Beta-Conditional Variational Autoencoder–Gated Recurrent Unit approach

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Halil Ibrahim Turan,
  • Mohand A. Djeziri,
  • Lisa Weber,
  • Virginie Martini,
  • Marc Bendahan,
  • Harun Taha Hayvaci

摘要

Early detection of chemical gases plays a critical role in public health, environmental safety and industrial monitoring. The need for faster and more accurate gas detection methods is increasing, especially in sensitive industrial applications, and machine learning (ML) algorithms are increasingly utilized by researchers to improve gas sensor performance by enabling earlier prediction of hazardous concentrations. This study presents an application-driven hybrid intelligent framework that first leverages a beta-regularized conditional variational autoencoder ( \(\beta \) -cVAE) to enrich the limited time-series exposure dataset obtained through ammonia (NH3) gas exposure in an experimental setup. The augmented data are subsequently used to train a Gated Recurrent Unit network (GRU) to predict the maximum gas concentration using only the early sensor readings, thereby eliminating the need to wait for full gas exposure. The proposed hybrid method achieves accurate early prediction of the peak concentration across various concentration levels, enabling a significant improvement of sensor response time. Building upon a previously proposed generative-predictive framework for hydrogen sensing, this study adapts and extends the methodology to NH3 detection using a \(\beta \) -regularized cVAE-GRU architecture under limited data conditions. By modeling temporal patterns in sensor response data, the framework is potentially adaptable to other time-series sensing applications.