Self-adaptive temperature and humidity compensation based on improved deep BP neural network for NO2 detection in complex environment.

The accuracy and reliability of gas sensor are directly by temperature and humidity. In this study, an improved deep Back Propagation (BP) neural network was designed to lower the impact of environmental factors on NO 2 gas sensor based on PbS nanoparticles sensitive film. The contradiction between high performance and low time complexity was usually faced by current compensation methods of gas sensor. Moreover, poor self-learning ability always resulted in more recognition errors. To solve the above problems, a 14-layer deep BP neural network model was constructed after hyperparameter searching. Stochastic Gradient Descent (SGD) algorithm with Mini-batch algorithm was adopted to well balance the model performance and the training time complexity, resulting in 76.68% performance improvement and nearly 6 times training time reduction after 1000 epochs, respectively. Softplus activation function was combined with Adam optimizer to further improve the model performance with a good recognition accuracy (1.37% relative error, corresponding to 0.0087 Mean Square Error (MSE)). The self-learning and self-adaptability of the improved deep BP neural network made it an excellent compensation method for the gas sensor applied in complex environments. • An improved deep BP neural network was designed to lower the environmental factors' impact on PbS based NO 2 gas sensor. • SGD algorithm with mini-batch was chosen to balance the model performance and the training time complexity. • Softplus activation function combined with Adam optimizer were selected to obtain accurate prediction of NO 2 concentrations. • Self-adaptation and self-learning capabilities of the model helped to achieve good recognition accuracy. [ABSTRACT FROM AUTHOR]