Research On Anomaly Detection And Recovery Of Bridge Health Monitoring Data Based On Deep Learning

The accuracy of security evaluation of bridge structures is closely related to the data collected by sensors.However,due to the quality defect of the sensor itself,the long period of operation and the severe environment,the sensor fault frequently occurs,and the bridge monitoring system produces anomaly data.If unreliable monitoring data are used to evaluate the security of bridge,it will lead to the omission and false alarm of the bridge monitoring system.How to accurately detect and update the anomaly data has become the focus in the field of brigde health monitoring system.Above all,this paper proposes a deep learning-based data anomaly detection and recovery method.The specific research work is as follows:(1)The variation of deflection and strain of classical bridge under moving load and temperature load is analyzed,and the response function of bridge monitoring system is constructed on this basis.According to the composition of the response function,each link of the bridge monitoring system is analyzed,and six common anomaly data types,characteristics and causes of anomaly data are summarized.(2)In order to solve the problem that the monitoring data and its characteristic information of the bridge are difficult to be obtained,the methods of extraction and visualization are used for preprocessing.As for the extraction of monitoring data,first of all,MATLAB software was used to write codes to load the monitoring data file,then the data were named according to the date and sensor number,and the monitoring data were classified and stored according to the type.Finally,continuous wavelet transform is used to visualize the time-frequency domain information of monitoring data,and the visualized image can be used for feature extraction and training of convolutional neural network.(3)Aiming at the problem of bridge monitoring data anomaly detection,a suitable convolutional neural network model is designed and established.After importing training samples,the optimal model was trained and saved for data anomaly detection,and the detection accuracy of the model reached 98.07%.By summarizing and analyzing the detection results,the quality of monitoring data can be evaluated,which can reflect the health status of the bridge sensors.In addition,three methods are used to repair the bridge monitoring distortion data,namely,the missing data repair method based on long-short term memory network,the outlier data repair method based on 3-sigma criterion and the trend data repair method based on linear regression.