A Suppression Method of Concentration Background Noise by Transductive Transfer Learning for a Metal Oxide Semiconductor-Based Electronic Nose

Huixiang Liu, Qing Li, Zhiyong Li, Yu Gu2020 · DOI

TL;DR

This paper introduces a novel method, MICF-IFLD, to tackle the significant challenge of sensor drift in electronic noses (e-noses). By combining a Maximum Independence of Concentration Features (MICF) approach to reduce concentration-related noise and Iterative Fisher Linear Discriminant (IFLD) to address sensor self-drift, the method achieves improved classification accuracy on datasets affected by these issues.

What Problem Does This Paper Address?

Electronic noses (e-noses), particularly those based on Metal Oxide Semiconductor (MOS) sensors, are valuable tools for identifying odors across various fields like medical diagnosis, food engineering, and environmental monitoring. However, a major hurdle in their practical application is sensor drift. This drift, caused by factors like sensor aging, poisoning, or environmental changes, alters the sensor's response over time, rendering previously trained classification models inaccurate. This problem is exacerbated by the fact that the concentration of the target substance is often unknown, making it difficult to distinguish between drift caused by concentration variations and inherent sensor drift. Existing methods for drift compensation often require reference gases or labeled target data, which are not always feasible. This paper aims to address the challenge of classifying gases using e-noses in the presence of both concentration-dependent noise and long-term sensor drift, particularly in a transductive transfer learning setting where target domain data is unlabeled.

Methods

The proposed method, MICF-IFLD, combines two key techniques:

  • Maximum Independence of Concentration Features (MICF): This component aims to reduce the influence of concentration variations on the sensor's signal. It works by learning a latent subspace where the extracted features are maximally independent from the concentration information. This is achieved by utilizing the Hilbert-Schmidt Independence Criterion (HSIC) to minimize the dependence between the projected features and the concentration features. The goal is to create a feature representation that is less sensitive to varying concentrations of the same gas.
  • Iterative Fisher Linear Discriminant (IFLD): This component addresses the sensor's self-drift over time. Since the target domain data is unlabeled, a direct application of standard Fisher Linear Discriminant (FLD) is not possible. IFLD uses a pseudo-labeling strategy, similar to the Expectation-Maximization (EM) algorithm. It first trains a classifier on the labeled source domain data. Then, it uses this classifier to predict labels for the unlabeled target domain data (pseudo-labels). FLD is then applied to this combined dataset, iteratively refining the pseudo-labels and the transformation matrix. This process aims to minimize within-class scatter and maximize between-class scatter, effectively reducing the impact of long-term sensor drift and improving class separability.

The MICF-IFLD algorithm first applies MICF to transform the data into a concentration-independent subspace and then uses IFLD on this transformed data for classification.

Datasets

The paper evaluates the proposed MICF-IFLD method on two datasets:

  1. Authors' Laboratory Dataset: This dataset was collected using a PEN3 E-nose (Airsense Analytics GmbH) over 45 days, measuring ethanol and n-propanol solutions at various concentrations (10% to 90%). This dataset was used to demonstrate the effectiveness of MICF in reducing concentration-dependent variations.
  2. Publicly Available Gas Sensor Drift Dataset: This dataset, created by Vergara et al., is a well-known benchmark for evaluating drift compensation methods. It comprises 13,910 records from 16 MOS sensors over three years, measuring six different gases at various concentrations. The dataset is split into batches based on acquisition time, with Batch 1 used as the labeled source domain and Batches 2-10 as unlabeled target domains. This dataset was crucial for evaluating the overall performance of MICF-IFLD in handling both concentration and long-term drift.

Key Insights and Ideas

  • Concentration as a major interference: The research strongly highlights that variations in gas concentration significantly impact the feature space of MOS sensor arrays, often more so than the inherent differences between distinct gases. This underscores the need to explicitly address concentration effects.
  • MICF for concentration independence: The MICF method effectively creates a feature subspace where the influence of concentration is minimized, leading to more consistent distributions for the same gas across different concentrations. This is a crucial step before tackling long-term drift.
  • IFLD for transductive drift compensation: The IFLD approach provides a practical way to apply FLD in a transductive setting (unlabeled target data) by iteratively refining pseudo-labels. This is key to handling sensor drift over extended periods.
  • Synergistic effect of MICF-IFLD: Combining MICF and IFLD proves more effective than either method alone. MICF handles immediate concentration variations, while IFLD addresses the slower, long-term drift, leading to a more robust classification system.
  • Practical implications for real-world e-noses: The paper demonstrates that the proposed method can improve classification accuracy and potentially reduce the recalibration frequency of e-noses used in long-term monitoring applications.

Key Results

  • Laboratory Dataset: The MICF method improved classification accuracy from 73.81% to 87.62% when compared to using original features with a BPNN classifier, demonstrating its effectiveness in mitigating concentration-induced variations.
  • Gas Sensor Drift Dataset: The MICF-IFLD method achieved the highest average classification accuracy of 76.17% among the compared methods, outperforming mainstream machine learning algorithms (RF, XGBoost, SVM, BPNN) and other transfer learning methods (TCA, Comgfk-ml, SMIDA).
  • Performance Degradation: While all methods showed performance degradation over time due to increasing drift, MICF-IFLD exhibited a slower degradation rate and higher overall accuracy, suggesting improved robustness.
  • Sensor Importance: For the Chinese liquor classification experiment, the fifth MOS sensor (MOS5) was identified as having the largest contribution to distinguishing between different types of liquors.

Limitations and Future Work

  • Fundamental Drift Resolution: The authors acknowledge that while MICF-IFLD improves performance and reduces recalibration rates, it cannot fundamentally resolve sensor drift. The drift still exists and causes performance degradation over time.
  • Computational Complexity: While the method is described as simple and effective, the iterative nature of IFLD might introduce computational overhead, especially with very large datasets or frequent retraining.
  • Future Work: The paper suggests that further research could explore more advanced domain adaptation techniques or sensor fusion strategies to more effectively combat sensor drift. Investigating the theoretical limits of drift compensation and developing methods that can adapt to more complex and dynamic drift patterns would also be valuable.

Summary updated

BibTeX
@misc{liu2020suppression,
  author = {Huixiang Liu and Qing Li and Zhiyong Li and Yu Gu},
  title = {{A Suppression Method of Concentration Background Noise by Transductive Transfer Learning for a Metal Oxide Semiconductor-Based Electronic Nose}},
  year = {2020},
  doi = {10.3390/s20071913},
}
Figure