Deep Learning Framework With Essential Pre-Processing Techniques for Improving Mixed-Gas Concentration Prediction

Multiple gas detection in mixed-gas environments is a challenging issue in many engineering industries because some of the gases can raise defect rates and reduce production efficiency. For chemo-resistive gas sensors, a precise estimation can be challenging because of the measurement variance and non-linear nature of the gas sensors, especially in a low concentration environment. A simple application of the deep learning models, however, does not yield sufficiently accurate predictions of the concentrations of multiple gases in gas mixtures; thus, it is essential to develop basic strategies for enhancing the accuracy in all possible ways. In this study, we develop a deep learning framework for achieving high accuracy of gas concentration prediction by studying the essential pre-processing techniques, learning task design, and architecture design. For the pre-processing, we study several aspects of processing time-series sensor data and identify the key techniques for complementing deep learning models' limitations. We utilize the mixed-gas nature for the learning task design and show that multi-task learning can generate a synergistic effect. Additionally, we show that a further improvement is possible by considering on-off classification as a part of the hybrid learning task. Concerning architecture design, we investigate Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) models after applying the identified pre-processing techniques. CNN outperformed other models in a joint analysis with the learning task. The effectiveness of our framework is confirmed with the UCI gas mixture dataset acquired using a chemical detection platform where 16 chemical sensors are exposed to ethylene, CO, and methane gases. Using the dataset, we study the basic techniques that can be effective to mixed-gas prediction. For the UCI dataset, our deep learning framework achieves a significant improvement in estimation accuracy when compared to the previous studies.