Research On Structural Damage Detection Algorithms Based On Deep Bidirectional Recurrent Neural Network

Structural vibration data are high dimensional timing data from the data type,but timing data are not independently and uniformly distributed.In the detection of abnormal points of structural vibration data,the algorithm used does not make full use of the characteristics of timing data.In the aspect of feature extraction,the result of principal component extraction is not interpretable.Finally,the traditional machine learning algorithm is used for damage identification.Firstly,it is inferior to deep learning in recognition accuracy.Secondly,some models are prone to overfitting.In this paper,based on the information of dynamic and dynamic response of structure acquired by health monitoring sensors,the research on data-driven structural damage identification method is a hot spot in the interdisciplinary field of computer science and civil engineering.Based on information theory,machine learning,and the depth of the neural network method based on artificial intelligence theory,such as temporal dependencies for structural health monitoring data of high dimension characteristic,dynamic threshold is proposed based on fusion filtering type of exception handling,monitoring data of feature selection and structural damage identification technique based on bidirectional circular neural network system and its key algorithms,for structural health monitoring data analysis field of research to provide new research ideas and methods,this article main research content is as follows:(1)the data preprocessing technology of structural health monitoring based on filter feature selection with information gain rate as screening index is analyzed in detail.This paper designs the data pre-processing technology of structural health monitoring which is aimed at the problem of too high dimension of collected data.This paper presents a set of data feature extraction process,which uses the information gain rate as the evaluation index of filter feature selection to carry on the feature selection.And put forward two sets of solutions for the information entropy can not be used in continuous data:1)Referring to the improvement of C4.5's ID3 algorithm,the continuous data are discretized and separated,and then the gain rate of information is calculated by using the discretized data;2)Using the idea of GDA for reference,the continuous distribution modeling is used.However,in the process of modeling,the monitoring data do not necessarily belong to Gao Si distribution,so the non-parametric hypothesis test should be added to verify the correctness of the distribution hypothesis before establishing the distribution,and then the parameter estimation of the continuous distribution should be made.Finally,the information gain rate can be obtained by calculating the continuous information entropy using the continuous distribution.(2)Research on abnormal point processing algorithm of structure perception data based on dynamic thresholdResearch on abnormal point processing algorithm of structure perception data based on dynamic threshold.Combined with the theoretical firmness of statistics,the dynamic threshold generation with the central limit theorem as the theoretical core is designed to determine the outliers.After the outliers are determined,the outliers are filled by Lagrange interpolation.Finally,a solution to the runge phenomenon in interpolation is given(3)Research on the key technology of bridge structure damage identification based on Bi-Directional LSTM.In view of the time series correlation characteristics of bridge monitoring data,combining with the superiority of Bi-Directional LSTM in time series data fitting,a Bi-Directional LSTM model based on structural monitoring data is constructed.At the same time,the design idea of the model is analyzed deeply.Related contents such as training optimization and construction and realization;In order to verify the reliability of the method,Bookshelf and Spring Beam are used to validate the data set,and the proposed Bi-Directional LSTM-based bridge damage identification is more accurate.(4)In order to verify the structural damage identification scheme proposed in this paper based on anomaly detection,feature selection and damage identification,Bookshelf and Spring Beam were used as verification data sets,and obfuscating matrix,ROC and AUC were used as evaluation indexes for comparative experiments.(1)through the and One-Class SVM and SVM algorithm,and verify the monitoring data of feature selection based on information gain rate is better than the other two models(2)through the selection of and PC A and variance comparison,abnormal points detection based on dynamic threshold value was verified in structural damage identification is better than the other two models(3)Bi-Directional LSTM in bridge structure damage identification effect is good,and the typical SVM,CART,and within DNN LSTM comparison,discovered the Bi-Directional LSTM at best.