On the performance evaluation of hybrid and mono-class sensor arrays in selective detection of VOCs: A comparative study

Ankur Srivastava, Vinayak P. DravidSensors and Actuators B Chemical 2006 · DOI

TL;DR

This paper tackles the challenge of selectively identifying volatile organic compounds (VOCs) using non-selective chemical sensors. The authors demonstrate that a hybrid approach, combining different sensor types (hybrid class) with the same sensor type operated at varying temperatures (mono-class), significantly improves detection accuracy and quantification compared to using either approach alone. This work offers a promising path towards more robust and reliable electronic nose systems for various applications.

What Problem Does This Paper Address?

The accurate and selective detection of Volatile Organic Compounds (VOCs) is a critical task in fields ranging from environmental monitoring and industrial safety to medical diagnostics. However, many chemical sensors suffer from cross-sensitivity, meaning they respond to multiple chemicals, making it difficult to pinpoint a specific VOC. Traditional "electronic nose" (e-nose) systems address this by using an array of non-selective sensors and employing pattern recognition algorithms to decipher the unique response patterns of different VOCs.

The challenge lies in creating sensor arrays that are both selective and practical. Building large arrays of diverse sensor materials (hybrid class sensors) can be complex and expensive, especially for integrated systems. This paper explores alternative strategies to enhance selectivity, specifically investigating the effectiveness of using an array of identical sensors operated at different temperatures (mono-class sensors) and, crucially, a combination of both hybrid and mono-class sensors.

Methods

The researchers employed a systematic approach to evaluate different sensor array configurations for VOC detection. Their methodology involved:

  • Sensor Selection: Four types of Taguchi sensors (TGS821, TGS826, TGS2611, TGS2620) were chosen for their known responsiveness to various VOCs.
  • Experimental Setup: A custom-built automated test-bed was used, featuring a gas delivery system with precise flow control, a gas chamber, and a data acquisition system.
  • Temperature Control: For mono-class sensor experiments, the sensors were operated at three different heater voltages, corresponding to temperatures between 200-500 °C, to generate varied response patterns.
  • VOC Exposure: Saturated vapors of five target VOCs (ethanol, ammonia, acetone, toluene, and chloroform) were generated and introduced to the sensors at controlled concentrations (10, 20, and 30 sccm) mixed with nitrogen carrier gas.
  • Data Analysis:
    • Sensitivity Calculation: Sensor sensitivity was defined as the ratio of sensor resistance in nitrogen ambient to that in the test vapor.
  • Principal Component Analysis (PCA): PCA was used to reduce the dimensionality of the data, identify redundant features (sensors), and visualize the separation of different VOCs and their concentrations.
  • Artificial Neural Networks (ANNs): Backpropagation trained feed-forward neural networks were employed for both identification (classification) and quantification (predicting concentration) of the VOCs. Different ANN architectures were tested using raw sensor data, principal components, and selected sensor features as input.
  • Case Studies: Three distinct scenarios were evaluated:
  • Hybrid Class Sensors: An array of different sensor types operated at a fixed temperature.
  • Mono-Class Sensors: An array of the same sensor type operated at different temperatures.
  • Hybrid and Mono-Class Sensors: A combined array where both hybrid and mono-class sensors were operated at different temperatures.

Datasets

The study utilized experimental data generated from exposing the selected gas sensors to five specific VOCs (ethanol, ammonia, acetone, toluene, and chloroform) at three different concentrations (10, 20, and 30 sccm). This generated a dataset for training and evaluating the pattern recognition algorithms. The paper does not introduce a new benchmark dataset but rather uses the experimental data collected for this specific study.

Key Insights and Ideas

  • Temperature is a Powerful Feature: Operating a single type of sensor at different temperatures can generate distinct response patterns, effectively mimicking the selectivity achieved by using an array of different sensor materials. This is a key insight for simplifying sensor array design.
  • Hybridization Boosts Selectivity: Combining different sensor types (hybrid) with the same sensor type operated at varying temperatures (mono-class) leads to significantly improved classification accuracy. This suggests that exploiting both material diversity and operational diversity is crucial.
  • PCA for Feature Selection: PCA proved invaluable for understanding data redundancy and selecting the most informative sensor inputs for the neural network models, leading to more efficient and potentially more accurate classification.
  • Ambiguity Resolution: The study successfully demonstrated how seemingly ambiguous responses from individual sensors can be resolved by considering the collective response of the array, especially when incorporating temperature variations.

Key Results

The study achieved impressive results, particularly with the hybrid and mono-class sensor combination:

  • Identification (Classification):
    • Hybrid class sensors alone achieved 87.71% classification accuracy.
  • Mono-class sensors (TGS2620 at different temperatures) achieved 100% classification accuracy.
  • The combined hybrid and mono-class array also achieved 100% classification accuracy, with easier training and fewer learning cycles compared to the mono-class only approach.
  • Quantification:
    • For the combined array, predicted concentrations of VOCs were found to be within 10% of the target concentrations for most input feature sets.
  • R-values (correlation between predicted and target concentrations) for the combined array were as high as 0.9990, indicating excellent quantification performance. This was significantly better than hybrid-only or mono-class only approaches, which showed lower R-values.

Limitations and Future Work

  • Integrated Array Challenges: While the hybrid and mono-class approach simplifies selectivity, the paper acknowledges that achieving different operating temperature zones on a single integrated substrate remains a significant technological challenge. This would require specialized heater geometries and materials with high thermal conductivity.
  • Limited VOC Set: The study focused on a specific set of five VOCs. Future work could explore the performance of this hybrid approach with a broader range of VOCs and more complex mixtures.
  • Sensor Drift and Stability: The paper does not explicitly discuss sensor drift or long-term stability, which are critical considerations for real-world applications. Further investigation into the robustness of the proposed method over time would be beneficial.
  • Computational Cost: While the combined approach improved training efficiency for ANNs, the computational resources required for training and inference with larger, more complex arrays might still be a factor.

The authors suggest that the demonstrated ability to achieve high selectivity and quantification using a combination of sensor types and temperature variations opens avenues for developing more sophisticated and practical electronic nose systems.

Summary updated

BibTeX
@inproceedings{srivastava2006performance,
  author = {Ankur Srivastava and Vinayak P. Dravid},
  title = {{On the performance evaluation of hybrid and mono-class sensor arrays in selective detection of VOCs: A comparative study}},
  booktitle = {Sensors and Actuators B Chemical},
  year = {2006},
  doi = {10.1016/j.snb.2005.11.034},
}
Figure