| Recently,with the continuous expansion of human cognitive boundaries,the demand for specific behavior recognition in special places such as Elderly safety monitoring,Campus bullying warning,and Perimeter security identification become more complicated.However,mainstream behavior recognition technologies are generally limited by many aspects such as equipment installation,cumbersome identification and huge amount of data.Owing to the low cost,convenient laying,stable physical properties and distributed monitoring,Phase Sensitive-Optical Time Domain Reflectometer(φ-OTDR)has been widely used in behavioral recognition.This paper through experiments including data collection,feature extraction,feature selection and pattern recognition,creatively solves the problems of small feature extraction range,single feature selection,biased evaluation index and insufficient recognition accuracy in the field of optical fiber sensor recognition.The specific research contents are as follows:(1)Based on the investigation and analysis of the research status and technical weaknesses of behavior recognition methods,this paper proposes methods based onφ-OTDR system which are feasible schemes for specific behavior recognition research in special places.(2)Based on the principle of φ-OTDR system,this paper focuses on the experimental platform building and 4 kinds of behavior modes including No abnormal environment,Walking,Falling and Falling during walking.After the experiment,286 groups of No abnormal environment signals,839 groups of Walking signals,408 groups of Falling signals and 414 groups of Falling during walking signals were obtained,which reserves as data set.(3)The Time domain,Frequency domain and Time-frequency domain fusion feature extraction method has been proposed.Based on the feature extraction principle,141 types of feature datas are extracted.And based on SVM,Random Forest and MLP learning models,the fitting performance of the Time domain,Frequency domain and Time-frequency domain fusion feature data set was analyzed by ROC curves,which were compared with that of single analysis domain model such as Time domain.The results show that the AUC area is up to 0.9899,and the small range of feature extraction in the field of optical fiber sensing was solved.Besides,the proposed method exhibits excellent model fitting performance on different learning models.(4)The Multi-feature Comprehensive Scoring Mechanism has been proposed,which uses Spearman correlation coefficient,Distance correlation coefficient,Ridge regression,Accuracy,Processing time and Sensitivity as the base feature selection models.Based on 5 kinds of learning models such as SVM and Random Forest,the differences between MFCSM and classic recognition methods such as feature selection and deep learning in Accuracy,Processing time and Sensitivity were analyzed.The results show the Accuracy of MFCSM could reach 94.08%,the Processing time can be as low as 187 milliseconds each time,and the Sensitivity can reach 95.87%.The proposed method solves the problems of biased evaluation indicators,single feature selection methods,and insufficient recognition accuracy,and its performance in engineering indicators is superior to traditional feature selection methods and classic deep learning methods.(5)Realizing the application expansion and performance revalidation of the methods proposed in this paper.Facing the Perimeter security field,according to the original proposed methods,a total of 295 groups of 4 types of behavior patterns including Climbing,Shaking,Knocking and Static were extracted.The results show that,the highest Accuracy can reach 98.45%,the lowest Processing time is 5.35 milliseconds each time,and the highest Sensitivity can reach 98.43%,which verified the universality and generalization ability of the proposed algorithm in different scenarios.The proposed methods provide novel processing idea and a solid application foundation for research in this field. |
PDF Full Text Request