Study On Cross-Sensitivity Suppression Method Of Sensor Array Detection For Mixture Gas Dissolved In Transformer Oil

Gas sensing technology is the key to the on-line monitoring of dissolved characteristic gases in transformer oil.The gas sensing array technology that realizes the automatic resolution and detection of the mixed gas is the development direction of multi-component mixed gas detection.Cross-sensitivity is a technical bottleneck for gas sensing array detection.A gas sensor array combined with intelligent algorithms is an effective means to solve this problem.This paper designs a gas sensor array for three typical transformer fault gases:hydrogen?H₂?,carbon monoxide?CO?and acetylene?C₂H₂?.A database of samples is created using data from a large number of gas sensitivity tests.Combined with various intelligent algorithms,a qualitative identification model of mixed gas based on deep belief network and a quantitative analysis model of mixed gas based on optimization algorithm are established.This paper effectively suppresses the cross-sensitivity of the mixed gas and realizes the intelligent resolution and detection of the mixed gas.The main work of the paper is:Eight TGS sensors are selected for H₂,CO,and C₂H₂,and assembled into a gas sensor array.Design gas distribution,testing,collection and other experimental processes.Build a gas sensing array testing platform.Experimental studies of the gas sensor characteristics of the sensor array are conducted for detection sensitivity,repeatability,and selectivity.Through a large number of experimental results,gas sensitive experimental sample databaseof three characteristic gases are formed.Using BP neural network and DBN network to establish qualitative recognition models of mixture gas.Combined with the pre-processed gas sensitivity experiment sample database,identify and analyze the two models.The results show that compared with the single BP neural network model,the qualitative recognition rate of the mixed gas is increased by 5 percentage points with the classifier DBN network.It can effectively reduce the cross-sensitivity,and the average accuracy rate reaches 94.11%.Using SVR?Support Vector Regression?to establish quantitative analysis models of H₂,CO,and C₂H₂.Combined with the pre-processed gas sensitivity experiment sample database,calculate the model.The result shows that the quantitative model and prediction effect of H₂ are good,followed by C₂H₂,and CO is not ideal.GA?Genetic Algorithm?and SVR are combined to form a genetic support vector regression machine.PSO?particle swarm algorithm?and SVR are combined to form a particle swarm support vector regression machine.The optimized quantitative analysis models solve the problem that traditional parameter optimization methods are easily trapped in local minima.The optimized H₂,CO and C₂H₂ gas quantitative analysis models have better reliability?MSE<0.1?and high prediction accuracy?squared correlation coefficient SCC>0.96?.