Research On Feature Extraction And Feature Selection Algorithm For Wound Infection Odor Detection

Diabetic foot ulcers,as a typical chronic wound,are the main cause of disability and death in diabetic people.At present,the incidence of diabetic foot ulcers has shown a high trend.The resulting chronic wound infection has become a serious threat to human health and has caused serious burden on social.Compared with traditional wound detection methods,odor detection technology has the advantages of rapid detection,convenient use and non-invasiveness,and is very suitable for daily screening of wound infections.The research of wound infection detection feature engineering is essential,because it has clear practical significance for the task of wound infection odor detection.This thesis takes the dual-channel odor sensing system to detect wound infection as the subject background.Based on the reliable rat wound infection experiment,the key technologies in the intelligent algorithm based on odor detection are researched.The detection system is composed of high-Field Asymmetric Ion Mobility Spectrometry(FAIMS)detection unit and electronic nose(e-nose)detection unit working in parallel.Aiming at the feature extraction problem of FAIMS unit and the feature selection problem of electronic nose unit,in-depth research on them was conducted,respectively.The main research contents are as follows.(1)Using FAIMS ion peak spectral line feature extraction method to identify wound infection odorCurrent FAIMS feature extraction methods follow traditional methods for analysis,such as image feature extraction methods and one-dimensional signal compression methods,some results have been achieved,while there is a lack of algorithmic research on the characteristics of FAIMS data.In this thesis,based on the characteristics of FAIMS datas,research on feature extraction methods based on ion peaks was carried out.For the first time,this method systematically defines the effective ion peak,then finds the shape of each effective ion peak by judging the type of peak and the position of the peak foot and regions.Finally,five contour features of each ion peak were extracted.Experimental results show that the overall characteristics of ion peak line features are superior,and they have information missing for other types of features,which can complement each other.The performance of the ion peak line feature method combined with image features for feature selection is significantly better than that of image features alone,and the recognition rate difference is more than 10% ～ 20%.(2)An electronic nose feature selection method based on local LDA and FGE is proposedFor current research on feature selection algorithms of the electronic nose,researchers mostly use the methods of evaluating single features,and then select several high-scoring features as the final feature set.These methods did not consider the interaction between features,which is easy to cause unstable performance of the selected feature subset.The LDA(Linear discriminant analysis)eigenvalue evaluation method can evaluate the feature group,but the evaluation is inaccurate for small samples of complex data.This thesis presents a new feature evaluation method based on local LDA(Linear discriminant analysis),combined with the feature group evolution algorithm FGE(Feature group evolution)to finally achieve the selection of the optimal feature subset.Local LDA feature evaluation method uses clustering to find multiple clusters in the sample distribution,and uses each cluster in each class as the basic unit to obtain the LDA eigenvalue score,finally,through a series of weighted summation,the evaluation score of the feature data is obtained.Local LDA combined with feature group evolution FGE can select a feature set with strong discriminating performance and stable performance.Experimental results show that the local LDA is more accurate than the LDA method in feature evaluation.Compared with the method of LLDAFGE and LDAFGE,the highest accuracy index of LLDAFGE is 4.12% higher than that of LDAFGE,and the average accuracy index is 2.86% higher.Compared with the 7 feature selection methods as a whole,the performance of the proposed method ranks among the top three in all results,proving that the method is stable and effective.