Co-training neural network-based infrared sensor array for natural gas monitoring

Abstract

High-precision infrared gas sensor arrays require high calibration costs due to the defect of cross-sensitivity of infrared sensors in gas mixtures. In this paper, we build an infrared sensor array for natural gas monitoring and propose to optimize the performance with low-cost unlabeled data through the co-training MLPNN (Multi-layer perceptual neural network). We use two MLPNNs with different data views, activation functions, and structures as base learners to facilitate co-learning. The base learner picks out high-confidence unlabeled gas samples at its own data view and then exchanges them for mutual improvement. We build a data acquisition platform in the laboratory to evaluate the performance of various algorithms. The experimental results show that the proposed method has the highest prediction accuracy for the gas concentrations compared with related algorithms. Co-training MLPNN can effectively exploit unlabeled gas data, and its error is reduced by 7–19%. In addition, the method has stable performance and high calibration cost benefits. The work shows a paradigm for the concentration monitoring of mixed gas via infrared sensors and gives a low-cost implementation method. Not only natural gas but also other gas mixtures can be measured with similar schemes.