Progress in Research on VOC Molecule Recognition by Semiconductor Sensors
Hongyu Liu,Gang Meng,Zanhong Deng,Meng Li,Junqing Chang,Tiantian Dai,Xiaodong Fang
DOI: https://doi.org/10.3866/pku.whxb202008018
2020-01-01
Acta Physico-Chimica Sinica
Abstract:Metal oxide semiconductor (MOS) gas sensors have been widely used in military and scientific research, as well as various industries; this is because of the unique advantages of MOS gas sensors including their small size, low power consumption, high sensitivity, and good silicon chip compatibility. However, the poor selectivity of MOS sensors has restricted their potential application in the Internet of Things (IoT) era. In this paper, progress in the research addressing the selectivity issues of MOS sensors is reviewed, and three strategies for selective MOS sensors, and performance improvements of MOS, e-nose, and thermal modulation, are introduced. Research on the performance improvements of MOS-sensitive materials provides an important guarantee for fast and accurate identification of trace gas molecules. The e-nose system adopts an array of sensors with distinct surface chemical properties; more "features" of volatile organic compound (VOC) molecules can be extracted by enlarging the number of sensor arrays, providing a "many-to-one" or "many-to-many" approach to discriminate VOC gas molecules via pattern recognition/machine learning algorithms. For thermal modulation technology, the working temperature of the sensor is intentionally swept during one measurement cycle, and the dynamic response signals of the sensor to different VOC gases under a given temperature mode are tested. Combined with signal processing and pattern recognition/machine learning, the "one-to-many" recognition of VOC gas molecules is realized by a single MOS sensor. Principal component analysis (PCA), linear discriminant analysis (LDA), and neural network (NN) pattern recognition/machine learning algorithms are compared in this review. Among them, the LDA algorithm based on supervised learning can be used as a signal dimension reduction or pattern recognition method. It is mainly applicable to the gas identification and classification of small datasets of VOC gas molecules. LDA is superior to PCA (based on unsupervised learning) in identifying and classifying VOC gas molecules. Compared with the LDA algorithm, an artificial neural network (ANN) based on the back-propagation algorithm, as a highly robust machine learning classification model, has the potential to process large datasets and realize the classification and identification of multiple kinds of VOC gases. Finally, the deep learning algorithm of convolutional neural networks (CNNs), with the performance of data dimension reduction, feature extraction, and robust identification, is expected to be applied in the field of VOC gas identification. Based on the performance improvement of MOS, a combination of multiple modulation methods and array technology, as well as the latest developments of deep learning algorithms in the artificial intelligence (AI) field, will greatly enhance the VOC molecular recognition capability of nonselective MOS sensors.