Linear and Non-Linear Modelling Methods for a Gas Sensor Array Developed for Process Control Applications

Riadh Lakhmi, M. Fischer, Quentin Darves-Blanc, Rouba Alrammouz, Mathilde Rieu, Jean-Paul ViricelleSensors 2024 · DOI

TL;DR

This paper presents a multi-sensor array and evaluates linear (PLS) and non-linear (ANN) models for detecting and quantifying H₂, CO, and CH₄ in gas mixtures relevant to "Power to X" applications. While both models show promise, ANN excels at predicting methane due to its non-linear sensor response, and both models' performance degrades with sensor aging, highlighting the need for robust calibration and future aging compensation strategies.

What Problem Does This Paper Address?

The "Power to X" (PtX) sector, which focuses on converting renewable energy into storable forms like hydrogen and synthetic fuels, relies heavily on processes involving gases such as H₂, CO, CH₄, and CO₂. Effective monitoring and control of these processes are crucial for safety and efficiency. However, commercially available gas sensors often suffer from a lack of selectivity, meaning they respond to multiple gases, making it difficult to accurately determine the concentration of individual components in a mixture. This paper addresses this challenge by developing and evaluating a multi-sensor platform and sophisticated modeling techniques to overcome sensor non-selectivity and enable accurate gas mixture analysis for PtX applications.

Methods

The researchers employed a multi-sensor platform composed of five commercial sensors with diverse transduction principles (electrochemical, catalytic, infrared, and metal-oxide semiconductor) to capture richer, less correlated information. They then developed and compared three modeling approaches:

  • MLR-OLS (Multi Linear Regression - Ordinary Least Squares): A basic linear regression technique. This was quickly disqualified due to poor performance during training.
  • MLR-PLS (Multi Linear Regression - Partial Least Squares): A linear multivariate regression technique that handles multicollinearity among predictors, making it suitable for sensor array data.
  • ANN (Artificial Neural Network): A non-linear modeling approach, explored with varying numbers of neurons and hidden layers, to capture potentially complex relationships between sensor signals and gas concentrations.

The models were trained using controlled gas sequences and then validated on independent datasets. A key aspect was also evaluating the models' performance after a year of sensor aging.

Datasets

The study utilized custom-generated gas mixtures for training and validation. These mixtures involved single analytes (H₂, CO, CH₄) and binary combinations thereof, with concentrations controlled by a gas bench setup. The data was collected over time, and transient parts of the sensor responses were removed to focus on steady-state behavior for model building. The paper also includes data from a year-long aging study of the sensor platform, providing a unique dataset for evaluating model robustness over time.

Key Insights and Ideas

  • Sensor Fusion is Key: Combining sensors with different transduction principles is essential for overcoming the inherent non-selectivity of individual sensors and providing a more comprehensive picture of the gas mixture.
  • ANNs Shine for Non-Linear Responses: The ANN model demonstrated superior performance in predicting CH₄ concentrations. This is attributed to the logarithmic response of the MOX sensor to CH₄, a non-linear relationship that ANNs are better equipped to model compared to linear methods like PLS.
  • PLS is a Solid Linear Baseline: MLR-PLS proved to be a capable linear model, offering good predictions for H₂ and CO, where sensor responses are generally more linear.
  • Sensor Aging is a Significant Challenge: After one year of operation, both PLS and ANN models experienced a significant deterioration in prediction quality. PLS models suffered from concentration offsets, while ANN models showed a drop in sensitivity. This highlights the critical need for ongoing calibration and potentially adaptive modeling strategies.
  • CO₂ is Easily Discriminated: The NDIR sensor for CO₂ was found to be highly selective, meaning CO₂ detection did not require complex multivariate analysis, simplifying its integration into the system.

Key Results

  • Initial Performance: Both MLR-PLS and ANN models achieved good concentration predictions for H₂, CO, and CH₄ on fresh sensor data.
  • CH₄ Advantage: ANN outperformed PLS in CH₄ prediction, with a validation RMSE of 497 ppm compared to 755 ppm for PLS.
  • Aging Impact:
    • ANN predictions on aged sensors showed high underestimation of H₂ and CH₄ concentrations.
  • PLS predictions on aged sensors exhibited significant concentration offsets (positive for H₂/CO, negative for CH₄).
  • After offset compensation, PLS performance on aged sensors improved considerably, even surpassing ANN in some aspects, though still not matching the unaged platform.
  • RMSE Values: Table 3 shows training RMSEs for MLR-PLS (H₂: 66 ppm, CO: 35 ppm, CH₄: 656 ppm) and ANN (H₂: 103 ppm, CO: 34 ppm, CH₄: 671 ppm). Table 4 shows validation RMSEs for MLR-PLS (H₂: 197 ppm, CO: 26 ppm, CH₄: 755 ppm) and ANN (H₂: 207 ppm, CO: 19 ppm, CH₄: 497 ppm).

Limitations and Future Work

  • Sensor Aging: The most significant limitation identified is the performance degradation of the models due to sensor aging. The authors suggest future work to investigate sensor signals at different aging stages and develop an "aging model" as a second layer to compensate for these effects.
  • Complex Mixtures: The current study focused on binary mixtures. Future work aims to extend this to more complex gas compositions, such as ternary and quaternary mixtures, which will likely require new models and potentially additional sensors to enrich the information captured by the array.
  • Offset Removal: While post-treatment algorithms were explored, significant offsets in aged PLS predictions were difficult to fully remove without periodic zero calibration. This suggests that robust calibration strategies are crucial for long-term deployment.

Summary updated

Abstract

New process developments linked to Power to X (energy storage or energy conversion to another form of energy) require tools to perform process monitoring. The main gases involved in these types of processes are H<sub>2</sub>, CO, CH<sub>4</sub>, and CO<sub>2</sub>. Because of the non-selectivity of the sensors, a multi-sensor matrix has been built in this work based on commercial sensors having very different transduction principles, and, therefore, providing richer information. To treat the data provided by the sensor array and extract gas mixture composition (nature and concentration), linear (<i>Multi Linear Regression-Ordinary Least Square</i> "MLR-OLS" and <i>Multi Linear Regression-Partial Least Square</i> "MLR-PLS") and non-linear (<i>Artificial Neural Network</i> "ANN") models have been built. The MLR-OLS model was disqualified during the training phase since it did not show good results even in the training phase, which could not lead to effective predictions during the validation phase. Then, the performances of MLR-PLS and ANN were evaluated with validation data. Good concentration predictions were obtained in both cases for all the involved analytes. However, in the case of methane, better prediction performances were obtained with ANN, which is consistent with the fact that the MOX sensor's response to CH<sub>4</sub> is logarithmic, whereas only linear sensor responses were obtained for the other analytes. Finally, prediction tests performed on one-year aged sensor platforms revealed that PLS model predictions on aged platforms mainly suffered from concentration offsets and that ANN predictions mainly suffered from a drop of sensitivity.
BibTeX
@inproceedings{lakhmi2024linear,
  author = {Riadh Lakhmi and M. Fischer and Quentin Darves-Blanc and Rouba Alrammouz and Mathilde Rieu and Jean-Paul Viricelle},
  title = {{Linear and Non-Linear Modelling Methods for a Gas Sensor Array Developed for Process Control Applications}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24113499},
}
Figure