Research On Classification Of Ecological Environment Sounds Based On Ensemble Subspace

With the advent of big data and artificial intelligence,machine learning has been further promoted.In the field of machine learning,ensemble learning,as a hot research topic,has not only received extensive attention,but also has been widely applied to various fields.The recognition of eco-environmental sounds is of great significance.It not only enhances our understanding of the surrounding ecological environment,but also better protects the ecological environment and species diversity.How to apply ensemble learning to the field of eco-environmental recognition to improve the classification recognition rate is a problem worthy of further study.In this paper,two sets of ecological environment sound datasets EnAudio1 and EnAudio2 are studied,and seven different feature selection methods,Correlation,GainRatio,InfoGain,OneR,ReliefF,SymmetricalUncert,csFs,stability evaluation methods and separation measurement methods are studied.The experimental analysis is carried out by means of Matlab,Weka and other softwares.The content and conclusions mainly have the following points:(1)Feature selection of eco-environmental sound data.Firstly,the MFCC,CELP,?MFCC and ??MFCC feature values are extracted by a series of transformations such as preprocessing of the sound signal.Then,the feature set is extracted according to different proportions by different feature selection methods,and each method is classified in the decision tree.The classification accuracy rate on the device was compared and analyzed.Then,the two groups of data sets were compared with the stability evaluation method on the decision tree classifier.The results show that the csFs method and the SU method have better stability on the eco-environmental sound data set.In addition,InfoGain and SU have the best results in the separation metric experiment with decision tree and naive Bayes as the classifier.(2)Ensemble classification based on subspace.By introducing the subspace,the ensemble of feature subspace is introduced,and the method of partitioning feature subspace is studied.The independent feature subspace division method,random subspace division method and weight subdivision method are compared.Experiments,and compared with the results of Bagging algorithm,the results show that the ensemble based on feature subspace can significantly improve the classification and recognition rate of ecological environment sound data,and the ensemble effect of independent subspace is better than the weight sub-stack ensemble method.And random subspace ensemble method.(3)Multi-models feature selection ensemble environmental sound classification.The differences and diversity of different feature selection methods are fully considered.The different feature subsets selected by different feature selection methods are integrated,and the decision tree is used as the base classifier.The experimental results show that the ensemble feature selection The classification accuracy rate is not necessarily higher than the single feature selection method,but in most cases,the classification accuracy of ensemble feature selection is significantly improved,and the ensemble method MmEnFs1 combined with the csFs method is nearly 7% higher than MmEnFs2.