System Construction And Algorithm Research In Human Wound Infection Detection Based On Electronic Nose Technology

Since the types and concentrations of metabolites of different wound pathogen aredifferent, electronic nose system can discriminate the types of wound infection pathogenby detecting the wound headspace gases. With the characteristics of being non-invasive,rapid, efficient, convenient, electronic nose has become a new and attractive diagnosticmethod. Applying the electronic nose in wound infection detection, an electronic nosehardware system for wound infection detection is built, and the methods of analyzing,processing and pattern recognition of electronic nose signal are studied in this thesis.The main work includes:①The common wound infection pathogens of human and SD (Sprague-Dawley)male rats are used as experimental subjects. According to the pathogen metabolites andthe sensitivity of gas sensors, an electronic nose hardware experimental system forhuman wound infection pathogens detection is designed and constructed. The systemconsists of an array of gas sensors, data acquisition system and flow control system.Based on the hardware experimental system, experiments of detecting seven types ofcommon human wound infection pathogens culture medium and four types of SD ratsinfected with different pathogens were performed. The experimental results show thatthe array of gas sensors has significant response on the headspace gases of seven typesof common pathogens culture medium and four types of SD rats’ wounds, and theresponse patterns are different. These data provide a basis for the subsequent analysis,processing and pattern classification of electronic nose signal.②Feature extraction is a key point of the electronic nose signal patternclassification. Performances of features have serious impact on the effect of thesubsequent pattern classification. Traditional feature extraction methods of electronicnose signal extract the steady-state response (the maximum of response) as a feature,without taking into account the additional information on the entire sensor responsecurves. Aiming at the problem that the maximum feature has not enough information forclassification and the classification results of wound infection detection are not good,the impact of common pretreatment methods and different feature extraction methods ofelectronic nose signal on the effect of subsequent pattern recognition are researched.First, the effects of differencing, fractional differencing, relative differencing, logarithmdifferential and normalization to the maximum feature are compared and studied. Then the impact of ten kinds of dynamic response feature, based on the original responsecurve, curve fitting, and transform domain, on the accuracy of the electronic nosepattern recognition algorithms, is analyzed. Experimental results show that, the featureextraction methods based on Fourier transform and wavelet transform etc. can greatlyimprove the discrimination ability of the electronic nose system for wound infectiondetection.③For wound infection detection electronic nose pattern recognition is essentiallya high-dimensional, small sample, non-equidistant, non-linear separable patternrecognition system, classifiers based on empirical risk minimization are ineffective. Thesupport vector machine which is based on the structural risk minimization is applied inwound infection detection electronic nose pattern recognition. The basic principle ofstatistical learning theory is analyzed, the advantages of support vector machines indealing with high-dimension, small sample and non-linear separable data are proven,and the impact of the support vector machine model parameters on generalizationperformance are explored. A grid search method and the genetic algorithm are used tooptimize the classifier parameters, respectively. Experimental results show that,compared with neural networks, support vector machine classifier has a huge advantagenot only in the recognition rate but also in the time consumption in terms of woundinfection detection electronic nose pattern classification.④Aiming at the problem that support vector machine model parameters and thedifferences of importance of each sensor will seriously affect the recognization effect,the particle swarm optimization algorithm is introduced to the optimization of supportvector machine model parameters and sensor array. A new electronic nose signalprocessing method that combines synchronization optimization of support vectormachine model parameters and sensor array based on particle swarm optimizationalgorithm with feature extraction based on wavelet transform is proposed to classify SDrats wound infection detection data. The wavelet approximation coefficients of sensorresponses are extracted as feature, and the features’ importance factor is found out byparticle swarm optimization algorithm, and weight the features by the importance factor,and the synchronization optimization of support vector machine model parameters andsensor array is realized. The results show that, compared with the neural network andmaximum feature, this method has the potential advantages in classification rate andtime consumption in SD rats wound infection data recognition.⑤Further, aiming at the problem that there are strong background interferences such as rat self-body odor and sample environment odor in SD rats in wound infectiondetection experiment, the impact of background interference on the recognition effect isanalyzed. A reference vector-based independent component analysis is proposed toremove the background interference and the identification of wound infection detectioncombined with artificial neural network is achieved. According to the independentcomponent analysis (ICA) model with noise, background interference signal is regardedas an independent component of the source signals. ICA is used to extract independentcomponents and discriminate background interference component and the useful signalsby correlation between independent components and one reference vector. Experimentalresults show that the method effectively remove strong background interference, andenhance discrimination ability of wound infection detection for electronic nose.