Research On Identification Methods For Monitoring Data Anomaly And Structural Load And Damage Based On Deep Learning

The application research of deep learning technology in the process of structural health monitoring is of great significance.In this thesis,deep learning in the structural health monitoring data anomaly detection,damage identification under unknown excitation,and unknown concentrated load time history reconstruction are researched.When the sensor is exposed to harsh environmental conditions,a variety of anomalies caused by sensor failure or damage will affect the measurement data of the SHM system.Faced with the massive data generated by continuous monitoring of the structure,manual removal of anomalies will lead to low efficiency.Therefore,it is of great significance to use more advanced deep learning methods for anomaly detection.After anomaly detection of monitoring data and elimination of anomalies,the more important step is to perform damage detection and condition assessment of the structure.For damage detection of structures under seismic excitation,different seismic excitations will inevitably lead to different structural response data.In practical applications,the future seismic excitation characteristics cannot be predicted by the pre-trained convolutional neural network.Therefore,it is very necessary to study the automatic detection of structural component damage under unknown seismic excitation.A very important part of structural health monitoring is to identify external loads.Accurate identification of the actual stress of the structure is the basis for decision-makers to conduct a structural safety assessment and decision-making.For concentrated load identification,most machine learning methods use shallow neural networks in machine learning,and cannot directly learn the end-to-end functional relationship between response and load.Moreover,the feature extraction of structural response also relies on manual experience,and most of the concentrated load forms are impulse loads.Therefore,the use of deep learning methods to more automatically identify the unknown concentrated load time history is a problem worthy of study.The first chapter of this thesis summarizes the current research status of data anomaly identification,damage identification,and load identification in the process of structural monitoring,and clarifies the current research deficiencies and the main work of this thesis.The second chapter of this thesis proposes a data anomaly detection method based on vibration signals and using a one-dimensional convolutional neural network.First,the anomaly detection problem is modeled as a time series classification problem.The original time series undergoes data preprocessing and data augmentation,including data expansion and down-sampling to construct new samples.For a small number of samples in the data set,two methods are used for data expansion,and samples with the same label are added without increasing the original samples.The down-sampling method of extracting the maximum and minimum values can effectively reduce the dimensionality of the input samples while retaining the characteristics of the data to the greatest extent.Adding the hyperparameter tuning of the class weights makes the convolutional neural network more effective in dealing with unbalanced training sets.The effectiveness of the proposed method is proved by anomaly detection of acceleration data on a long-span bridge.The results show that the proposed method can automatically detect multiple data anomalies with high accuracy.The third chapter of this thesis proposes a structural damage identification method based on structural response transmissibility function and wavelet packet energy,using a convolutional neural network under unknown seismic excitation.The transmissibility function(TF)of the structure response is used to eliminate the influence of unknown seismic excitation.Besides,the inverse Fourier transform and wavelet packet transform of TF are used to reduce the influence of frequency band and extract damage-sensitive features.Generate structural response data under environmental excitation,and use undamaged data and a small amount of damaged data to construct a training sample set to train a convolutional neural network.Then,the response of the structure under the unknown seismic excitation is processed accordingly,and the trained network model is used to locate the damage.The fourth chapter of this thesis proposes a method of reconstructing unknown concentrated loads based on structural response signals and using a recurrent neural network.The Recurrent Neural Network architecture is mainly composed of three LSTM layers,which are trained through a large number of dynamic responses and white noise loads to learn the complex inverse mapping between structural input and output.And verify the proposed method through the numerical simulation of the beam bridge.The results show that this method can reconstruct the time history for several types of loads even when the time history of the concentrated load is unknown.For more load types,the applicable scope and limitations of this method need to be studied.