The Study And Application Of Pattern Recognition In The Classification Of Gas Samples

The olfaction of mammal can be replaced by the electronic nose (EN) system. The odor can be recognized by the artificial system. Some dangerous tasks which were previously carried out by human beings can now be executed by EN. The EN could also be used in the evaluation of the quality of food, beverage, and alcohol. Another application field is anti terrorism in which EN could be employed to detect the potential bomb attack. Much attention is being paid on the EN and it becomes one of the braches of the bionics.The paper focused on the pattern recognition (PARC) in EN. The commercial available sensors were assembled to establish the testing platform. PARC was carried out after acquisitions of the data. The primary work in the thesis is as followed:1. 5 different model commercial available gas sensors provided by Figaro Corp. were assembled as an array. The homemade circuit converted the conductance of the gas sensors into the voltage signal.2. Combined with PCI-6251 data acquisition card, a data acquisition procedure, which was developed in LabVIEW, was deployed to collect the variation of the voltage signal of the reference resistors with which the gas sensors were in series. The data were saved as data file whose extension name is lvm and could be edited by Excel for further processing.3. The vapors of the acetone and the formalhyde samples (Group A) were tested for 150 times respectively in the gas sample acquisition section. 300 pieces of data in total were obtained in the section. Besides, the solutions of methanol and erguotou (an alcoholic product in domestic market, Group B) were confected, and 100 vapor samples were gathered.4. The interface called Exlink between Excel and Matlab was utilized to develop VBA code in order to reduce the dimensions of the data. The reduction of the data dimensions could bring the benefit of greatly decreasing the calculation. Meanwhile, the 'dimension disaster' caused by the high dimensioned data could be avoided. 25 feature points were extracted out of 1,500 data points to represent one gas sample in each experiment. The feature points were then passed on to PARC for sorting.5. The thesis generally demonstrated the principles of three main categories of the PARC. Linear Discriminant Analysis (LDA), Cluster Analysis (CA) and Backpropagation Artificial Neural Network (BP-ANN) were picked out and the corresponding principles and calculation steps were fully studied. The three PARC routines were applied in the classification of the samples from Group A and Group B. The influence of the different training functions to BP performance and classification results were discussed. Finally, the conclusion was figured out that the three PARC routines could successfully classify the samples from Group A, but they could not distinguish the samples from Group B.6. The feasibility, accuracy and calculation resource costs of three PARC routines were analyzed and compared. The conclusion came out: l)If the dimensions of the data samples were small, the LDA and CA may fast classify the samples due to the simplicity of the algorithms, and they were easy to implemente in programming languages, such as C. When the dimensions increase, the calculation might be augmented because of the inversion of the matrix in LDA and the complexity of the similarity functions of CA. The increased calculation might lead to the fact that LDA and CA were no longer the fast and easy algorithms. BP has highly non-linearity. Although the classification boundaries in any shape could be approached by BP in theory, the calculation was of great complexity and difficult to implemented in programming languages. CA was concluded as the most suitable approach for the experiment in the thesis.The main features of the thesis are as followed:1. The self-designed data acquisition software has a friendly user interface. The voltage signal in pace with the concentration of the gas sample could be read out in real-time. The curves of one sample could be clearly observed after importing the data into Excel.2. The half-open gas chamber could gather the dynamic information of gas sample. The structure of the chamber increased the information for PARC as well as the speed of the experiments. 3. The whole process of the feature data extraction (FDE) was automatically done by program, required little participation of the operators. The FDE code was module-designed and divided into two parts: front-end and back-end. In front end, Excel was utilized as the data display and the user interface for data input. All the calculation was processed in Matlab in the back-end. The code in the back-end also followed the rule of module design. The code which completed complicated tasks was encapsulated as a standalone function.4. The most important feature was that the thesis proposed an idea that the dividing point in LDA could be optimized based on genetic algorithms (GAs).