Research Of Intelligent Data Processing Algorithms For Detection Of Wound Infection Based On Electronic Nose

Since the types of metabolites of different wound pathogen are different, electronicnose (E-nose) system can be used to discriminate the types of wound infection pathogenby detecting wound headspace gases. With the characteristics of being non-invasive,rapid, efficient and convenient, E-nose has become a new and attractive diagnostic tool.The intelligent algorithms system of E-nose used in the field of wound infectionincludes the acquirement of original response of sensors, data processing and patternrecognition. An ideal data processing technique can extract more helpful informationwhich ensures that the classification accuracy of E-nose reaches the desired level, butthe traditional data processing methods used in E-nose are principal component analysis(PCA), fisher discriminant analysis (FDA) which are known as linear analysistechniques. Those novel techniques which have been successfully applied in other fieldssuch as image analysis, object identification have hardly been used to deal with theproblems of E-nose.The purpose of this thesis is to research and find a novel intelligent data processingtechnique which can improve the classification accuracy of E-nose in predicting theclass information of different wound infection. The main research content of theintelligent data processing technique in this thesis is the kernel principal componentanalysis, and the research includes the preparation of its input data, parameteroptimization, performance improvement of the kernel function and the dimensionalityreduction of the output data. All the above research is to find an effective way whichcan improve the classification accuracy of E-nose in wound infection detection. Themain research work and the achievements of this thesis are shown as follows:①SD (Sprague-Dawley) male rats are used as experimental subjects, and allexperiments are based on the common wound infection pathogens of human. Accordingto the pathogen metabolites and the sensitivity of gas sensors, an electronic nosehardware experimental system for human wound infection pathogens detection isdesigned and constructed. The system consists of an array of gas sensors, dataacquisition system and flow control system. Based on the hardware experimentalsystem, experiments of detecting four types of SD rats infected with different pathogenswere performed. Experimental results show that the array of gas sensors has significantresponse on the headspace gases of four types of SD rats’ wounds, and the response patterns are different. These data provides a basis for the subsequent data processingand pattern classification of E-nose.②The constructing of the original feature matrix is the first and important step ofdata processing for E-nose. Wavelet coefficients contain a wealth of information, andthey have been frequently used to represent the original response in the field of patternrecognition. When wavelet coefficient is used as the feature of response of sensorsdetecting wound infection, the problem is that there are more than one coefficient whichwe are interested. It’s easy to understand, different coefficients will play different rolesduring the representation of the original response of sensors, the loss or ignoring of onecoefficient will lead to loss of key information. A novel weighted summation model isproposed to solve this problem, and each wavelet coefficient will be weighted accordingto its role in letting the E-nose make a correct judgment. The weighted coefficients ofthis model are optimized by an enhanced quantum delta particle swarm optimization(EQPSO). Experimental results prove that the original feature matrix contains morehelpful information when weighted summation model is used to deal with the waveletcoefficients, and the classification accuracy of E-nose is higher than other featureextraction methods.The body odor of rats becomes the background interference during the woundinfection sampling experiments of E-nose. Spatial correlation method is used to removethe background inference, during which the wavelet coefficients of the mixed signal(containing the odor sampled from the wound and the body odor of the rats) and thebody odor of the rats are analyzed by the spatial correlation method to remove thebackground from the response of E-nose. Experimental results show that theclassification accuracy of E-nose has been improved obviously when the backgroundinference is processed by spatial correlation method.③There are some different parts for a whole intelligent algorithms system ofE-nose such as feature extraction, dimension reduction and pattern recognition et al.,and there are some parameters which are from different parts needing to be set. Thesetting of these parameters can influence the performance of the intelligent algorithmsystem of E-nose. A synchronous optimization model of E-nose is proposed to find thevalue of different parameters of E-nose, and optimization goal of the model is uniform.Meanwhile, an enhanced quantum-behaved particle swarm optimization (EQPSO) isproposed as the optimizing technique of the optimization model, and the ability ofEQPSO in searching the global optimization is improved, and this has been proved by experimental results of the optimization of testing functions. The experimental result ofsynchronous optimization model of E-nose is also ideal when EQPSO is adopted as theoptimization method.④The classification accuracy of E-nose in detecting the wound infection is notvery ideal when the original feature matrix is used as the input data of classifier. A newfeature must be found to represent the odor pattern of the wound infection, to improvethe classification accuracy of E-nose. Previous work has proved that this original featurematrix is uneven and nonlinear, so the data processing method which will be used toprocess the original feature matrix should have the ability of dealing with nonlinearproblem. Kernel principal component analysis is a classical nonlinear data processingmethod, but its performance can be easily influenced by its kernel function. Consideringthe actual application of E-nose, a novel weighted kernel principal component analysiscombined with multiple kernel learning and KPCA is proposed, the mapping ability ofthe kernel function is improved, and the information which is mapped to thehigh-dimension is abundant. The detailed mathematical derivation of the proposedmethods has been given. Besides, the pre-processing of the input data, admissibilityanalysis of the novel weighted kernel function and parameters setting are also studied inthis thesis. Experimental results prove that the performance of the proposed method isthe best among all different methods which are used to deal with the original featurematrix.⑤When KPCA is used to process the original feature matrix, the data is mappedfrom the input space to a high-dimension space. More computation will be needed byclassifier if the output data of KPCA is put into it directly. A novel manifold learningmethod, supervised locality preserving projections (SLPP), is used to reduce thedimension of the KPCA output. SLPP shares many properties of the classical manifoldlearning, namely, it can transmit as much nonlinear information from the input space tothe low-dimension space as possible. What’s more, it can provide the explicit mappingexpression, which can let it deal with the new sample which is acquired during theapplication of E-nose. Experimental results prove that the performance of SLPP is betterthan other methods on the premise of not reducing the classification accuracy of E-nose.