Non-linear carbon dioxide determination using infrared gas sensors and neural networks with Bayesian regularization

Carbon dioxide gas concentration determination using infrared gas sensors combined with Bayesian regularizing neural networks is presented in this work. Infrared sensor with a measuring range of 0–5% was used to measure carbon dioxide gas concentration within the range 0–15000 ppm. Neural networks were employed to fulfill the non-linear output of the sensor. The Bayesian strategy was used to regularize the training of the back propagation neural network with a Levenberg–Marquardt (LM) algorithm. By Bayesian regularization (BR), the design of the network was adaptively achieved according to the complexity of the application. Levenberg–Marquardt algorithm under Bayesian regularization has better generalization capability, and is more stable than the classical method. The results showed that the Bayesian regulating neural network was a powerful tool for dealing with the infrared gas sensor which has a large non-linear measuring range and provide precise determination of carbon dioxide gas concentration. In this example, the optimal architecture of the network was one neuron in the input and output layer and two neurons in the hidden layer. The network model gave a relationship coefficient of 0.9996 between targets and outputs. The prediction recoveries were within 99.9–100.0%.