Research On Explosive Classification And Recognition Based On Electronic Nose

At present,the world as a whole is relatively peaceful,but local conflicts and terrorist attacks occur from time to time.And new energy has developed rapidly in various fields around the world in recent years.However,due to its flammable and explosive properties,such as hydrogen energy,its safety problems in the process of production,storage and use cannot be ignored.Therefore,it is necessary to study the detection and classification of explosives.In this thesis,machine olfaction is applied to the detection and classification of explosives.An array of eight sensors and a virtual array of a single sensor are constructed,which can realized the detection and identification of three nitro explosives（octogen（HMX）,pentaerythritol tetranitrate（PETN）,Trinitrotoluene（TNT））and explosive gas hydrogen（H₂）,respectively.The specific research contents of this thesis are as follows:（1）The array containing 8 sensors was prepared by using organic carrier to mix tin oxide（SnO₂）and tungsten oxide（WO₃）with different mass ratios by ball milling,then calcining and annealing at high temperature.The morphology and phase of the sensors were characterized,which proved that SnO₂ and WO₃ were successfully composited,and the porous structure of its surface was conducive to target gas entering the sensitive material and combining with the active sites.The array was welded with the substrate in the test cavity by the gold wire ball welding,and the resistance data of the array was collected by the data acquisition card.Three kinds of explosives are decomposed into target gas by ultraviolet radiation,which reacts with sensor array,and detection limit of explosives is as low as 500 ng.In order to realize the rapid detection and classification recognition of explosives,this thesis constructs a CNN-LSTM network model based on convolution neural network（CNN）and short-term memory neural network（LSTM）.The data of the first 5 s,10 s and 15 s after the sensor array starts to respond to target gas are substituted into the CNN-LSTM model for learning,respectively.The recognition accuracy is 72.6%,95.6%and 97.7%,respectively.Compared with traditional pattern recognition algorithm,the effect of constructed CNN-LSTM model is significantly better than that of traditional pattern recognition algorithm.（2）A layer of SnO₂ film was deposited on the interdigital electrode by magnetron sputtering method,and then a layer of platinum（Pt）was sprayed on the surface of the SnO₂ film by an ion sputtering instrument,and a SnO₂/Pt sensitive layer was obtained by annealing.The gas sensitivity（resistance and impedance）of the sensor to H₂,acetaldehyde,acetone and ethanol was tested,respectively.The results show that the resistance response of the sensor to the four gases has large response values,good repeatability and linearity（10-50 ppm）.However,the response values of the four gases are similar in some concentration ranges,the accuracy of identifying different gases by resistance response values is very low（55.5%）.Therefore,impedance spectrum analysis technology was used to build a virtual sensor array to improve the recognition accuracy in this thesis.By analyzing its impedance gas sensing performance,it is found that under the selected operating frequency,the linearity of its impedance imaginary part response values（all greater than 0.99）is significantly higher than that in resistance.In addition,the equivalent circuit of the sensor,the Nyquist circle of the sensor in four gases,and the relationship between the real part and the imaginary part and the operating frequency are also analyzed.Finally,the impedance response spectra at 10 driving frequencies were selected by impedance analysis,and the"fingerprint spectra"of four gases were obtained.Combined with linear discriminant analysis（LDA）algorithm,the recognition accuracy of the four gases is improved to 94.4%.