Scholarlink

<p>The deployment of gas sensors in industrial and mining environments is greatly affected by sensor drift, which leads to false gas concentration estimation and unreliable safety assessment. Traditional models trained under ideal conditions often fail to generalize this drift, thus highlighting the need for adaptive learning models that can work under limited or absent target-domain labels. This paper proposes a drift-aware gas sensing framework that integrates feature-level and instance-level domain adaptation to mitigate the impact of sensor drift, thereby providing reliable health-oriented decision support with a predicted Modified Health Quality Index value. The proposed model utilizes CORrelation ALignment to mitigate covariate shift, and TrAdaBoost.R2 is employed to perform instance reweighting and suppress misleading source-domain instances. The framework is evaluated using a publicly available gas-sensor-array-drift dataset. The baseline regression models attain high values of predictive accuracy, but they suffer greatly during drift conditions, revealing negative <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> values in the latter batches. The joint approach of CORAL and TrAdaBoost.R2 brings the RMSE error down by 44.9% and improves the <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> metric of ethanol concentration measurements from 0.21 to 0.78. Though the regression model performance is affected during the severe drift conditions, the classification model built using the MHQI approach attains an accuracy of 84% on the drifted batches, indicating the model is well-equipped with the concept of hazards regardless of the inaccuracy of the concentration estimation during severe conditions.</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

Drift-Aware Gas Concentration Estimation and Health Risk Assessment in Mining Environments Using Feature and Instance-Level Domain Adaptation

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Aparna Singh,
  • Surabhi Solanki,
  • Sachin Kumar,
  • B. Vasumathi,
  • Arindam Biswas,
  • Vaibhav Srivastav,
  • Pulakesh Roy,
  • Surbhi Sharma

摘要

The deployment of gas sensors in industrial and mining environments is greatly affected by sensor drift, which leads to false gas concentration estimation and unreliable safety assessment. Traditional models trained under ideal conditions often fail to generalize this drift, thus highlighting the need for adaptive learning models that can work under limited or absent target-domain labels. This paper proposes a drift-aware gas sensing framework that integrates feature-level and instance-level domain adaptation to mitigate the impact of sensor drift, thereby providing reliable health-oriented decision support with a predicted Modified Health Quality Index value. The proposed model utilizes CORrelation ALignment to mitigate covariate shift, and TrAdaBoost.R2 is employed to perform instance reweighting and suppress misleading source-domain instances. The framework is evaluated using a publicly available gas-sensor-array-drift dataset. The baseline regression models attain high values of predictive accuracy, but they suffer greatly during drift conditions, revealing negative \(R^2\) R 2 values in the latter batches. The joint approach of CORAL and TrAdaBoost.R2 brings the RMSE error down by 44.9% and improves the \(R^2\) R 2 metric of ethanol concentration measurements from 0.21 to 0.78. Though the regression model performance is affected during the severe drift conditions, the classification model built using the MHQI approach attains an accuracy of 84% on the drifted batches, indicating the model is well-equipped with the concept of hazards regardless of the inaccuracy of the concentration estimation during severe conditions.