Enhancing the Discrimination Ability of a Gas Sensor Array Based on a Novel Feature Selection and Fusion Framework

Changjian Deng, Kun Lv, Debo Shi, Bo Yang, Song Yu, Zhiyi He, Jia YanSensors 2018 · OpenReview

TL;DR

This paper introduces a new framework for gas sensor arrays (electronic noses) that improves odor identification by intelligently selecting the most informative features and then fusing them. This approach achieves higher recognition rates and is more robust to sensor drift compared to conventional methods.

What Problem Does This Paper Address?

Electronic noses (E-noses) are powerful tools for detecting and identifying gases, mimicking the human olfactory system. However, they face several challenges:

  • High Dimensionality and Redundancy: Raw sensor data can be very high-dimensional, with many features being redundant or irrelevant, which can degrade classification accuracy and increase computational cost.
  • Sensor Drift: Over time, sensors can change their response characteristics (drift), leading to decreased performance and unreliable identification.
  • Limited Information per Feature: A single sensor's response might not fully capture the complex characteristics of an odor, necessitating the combination of information from multiple sensors and features.

The paper aims to address these issues by developing a method that can effectively select the most discriminative features and then fuse them in a way that maximizes classification performance while mitigating the impact of sensor drift.

Methods

The core of this work is a novel feature selection and fusion framework.

  1. Feature Selection:
    • Separability and Dissimilarity Indices: The authors propose a method that uses two key metrics to evaluate features:
  2. Separability Index (SI): Measures how well a feature can distinguish between different gas classes. A higher SI means better class separation.
  3. Dissimilarity Index (DI): Measures the correlation between features. A higher DI indicates less shared information, meaning the features are more diverse and less redundant.
  4. KNN-based Optimization: The K-Nearest Neighbor (KNN) classifier is used in conjunction with these indices to determine the optimal subset of features. The algorithm iteratively adds features, evaluating the combined SI and DI, and also considers the classification accuracy achieved by KNN with the selected subset.
  5. Feature Fusion:
    • Classification Dominance Strategy: Instead of simply concatenating features, this method assigns weights to different feature types based on their individual classification accuracy. Features that perform better individually receive higher weights.
  6. Weighted Fusion: The weighted features are then fused to create a final decision, aiming to leverage the strengths of each feature type and suppress the weaknesses of less informative ones.

Datasets

The proposed framework was evaluated on two publicly available datasets:

  • Dataset I: This dataset involves four types of wound bacteria (uninfected, Staphylococcus aureus, Pseudomonas aeruginosa, and Escherichia coli). It utilizes a sensor array with metal oxide and electrochemical sensors. Seven types of features were extracted from the sensor responses.
  • Dataset II: This is a larger dataset containing 13910 measurements from an E-nose system with 16 gas sensors, collected over 36 months, exhibiting significant long-term drift. It involves six pure gaseous substances (ethanol, ethylene, ammonia, acetaldehyde, acetone, and toluene) at various concentration levels. Eight types of features were extracted, including those derived from exponential moving averages to capture drift.

Key Insights and Ideas

  • Feature Importance is Not Uniform: Not all sensor features contribute equally to odor identification. Some are highly discriminative, while others are noisy or redundant.
  • Separability and Dissimilarity are Complementary: Relying solely on separability might select correlated features, while relying only on dissimilarity might miss important discriminative information. Combining both provides a more robust selection.
  • Intelligent Fusion Outperforms Simple Concatenation: Directly combining all features (conventional fusion) is less effective than a weighted fusion strategy that prioritizes features with higher individual classification power.
  • Feature Selection is Crucial for Drift Mitigation: By selecting only the most relevant and stable features, the proposed method can significantly reduce the negative impact of sensor drift on classification accuracy.

Key Results

  • Dataset I: The proposed feature selection and fusion framework achieved a classification accuracy of 97.5% with k=1 and the correlation distance metric. This is significantly higher than conventional methods.
  • Dataset II: The framework achieved a classification accuracy of 80.11% for the dataset with significant drift. This demonstrates its effectiveness in handling challenging real-world conditions.
  • The paper shows that the proposed method outperforms both using all features without selection and a conventional feature fusion approach.

Limitations and Future Work

  • Computational Cost of Feature Selection: While the proposed method is more efficient than pure wrapper methods, the feature selection process itself still involves significant computation.
  • Generalizability: While tested on two distinct datasets, further validation on a wider range of gas mixtures and sensor types would be beneficial.
  • Real-time Application: The paper focuses on offline analysis. Adapting the framework for real-time E-nose applications might require further optimization.
  • Future Work: The authors suggest that the framework can be further improved by exploring more advanced feature selection and fusion techniques, and by applying it to more complex gas mixtures and real-world scenarios.

Summary updated

Abstract

In this paper, a novel feature selection and fusion framework is proposed to enhance the discrimination ability of gas sensor arrays for odor identification. Firstly, we put forward an efficient feature selection method based on the separability and the dissimilarity to determine the feature selection order for each type of feature when increasing the dimension of selected feature subsets. Secondly, the K-nearest neighbor (KNN) classifier is applied to determine the dimensions of the optimal feature subsets for different types of features. Finally, in the process of establishing features fusion, we come up with a classification dominance feature fusion strategy which conducts an effective basic feature. Experimental results on two datasets show that the recognition rates of Database I and Database II achieve 97.5% and 80.11%, respectively, when k = 1 for KNN classifier and the distance metric is correlation distance (COR), which demonstrates the superiority of the proposed feature selection and fusion framework in representing signal features. The novel feature selection method proposed in this paper can effectively select feature subsets that are conducive to the classification, while the feature fusion framework can fuse various features which describe the different characteristics of sensor signals, for enhancing the discrimination ability of gas sensors and, to a certain extent, suppressing drift effect.
BibTeX
@inproceedings{deng2018enhancing,
  author = {Changjian Deng and Kun Lv and Debo Shi and Bo Yang and Song Yu and Zhiyi He and Jia Yan},
  title = {{Enhancing the Discrimination Ability of a Gas Sensor Array Based on a Novel Feature Selection and Fusion Framework}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=0G7CrMs_Ht},
}
Figure