Research On Mixed Gas Identification And Concentration Detection Algorithm Based On Sensor Array

In human daily life and industrial production,accurate and rapid detection and identification of toxic and harmful gases in the environment play an early warning and preventive role for avoiding the occurrence of dangerous gas accidents,and is of great significance to human health and safety,social harmony and stability.Compared with traditional chemical analysis instruments,machine olfaction has gained broad application prospects with the advantages of rapid detection,simplicity and low cost.Facing difficulties of cross-sensitivity and long response time,accuracy and efficiency of gas identification using metal oxide sensors need to be improved.This research focuses on the data-driven multivariate mixed gas identification technology,including investigation of mixed gas concentration prediction algorithm based on deep learning,and that of gas classification and concentration prediction algorithm based on multi-task learning.The goal is to accurately identify the composition of the gas and predict its corresponding concentration without achieving a steady-state response of a sensor.Specific works include:1.A mixed gas concentration prediction algorithm,based on convolutional neural network and long short-term memory neural network,is proposed.According to the characteristics of long sequences of sensor data,the structure and size of rectangular convolution kernels were optimized for simultaneous feature extraction,and a one-dimensional convolutional network combined with LSTM for efficient feature learning was constructed.The relationship between the prediction effect of the proposed method and the length of intercepted time series data is discussed.Upon comparing with the existing methods LSTM,SVR and RC,we conducted comparative experiments using UCI public dataset.The results showed the algorithm proposed in this paper has the best gas concentration prediction,shorter response time and faster convergence speed,and the prediction effect of the model is the best when the 20s response segment is intercepted.The concentration prediction accuracy R~2 of carbon monoxide and ethylene mixture gas reached 0.99.Relative to the maximum concentration value,the prediction error of ethylene is 8%,and that of carbon monoxide is 2%.Compared with the LSTM network,the R~2 is increased by 13%,and compared with the PSO+SVR network,the R~2 is increased by 14%.2.This study investigated the parameter sharing characteristics of multi-task learning,and proposed three neural network algorithms based on different multi-task learning frameworks,namely hard parameter sharing,soft parameter sharing and Cross-stitch parameter sharing,combined with Bayesian optimization.The algorithm tunes the hyperparameters of the network.A comparative experiment among the three algorithms was conducted using the UCI public dataset,and their performance was evaluated.The experimental results show that the three algorithms based on the multi-task learning framework can achieve the classification accuracy of Acc.In the context of specific tasks,the multi-task learning framework with soft parameter sharing is better than hard parameter sharing,with higher accuracy and better generalization.