Research On Prediction Method Of Bridge Dynamic Monitoring Signal Based On Machine Learning

In structural health monitoring system(SHM),due to various factors such as sensors,data loss and damage will inevitably occur,and data loss will easily lead to unnecessary loss of bridge structure.This paper aims to reconstruct missing data in SHM.Research on missing data prediction method based on SAM-LSTM model.The contents are as follows:(1)Firstly,the long-term and short-term memory network(LSTM)model is used to predict the wavelength and acceleration data in healthy state.Based on this model,three kinds of hyperparameter prediction experiments were carried out.The LSTM model is optimally designed for the prediction experiment of different number of iterations,the prediction experiment of different hidden layer nodes,and the prediction experiment of different learning rates.By comparing the maximum absolute percentage error(MAX),mean absolute error(MAE),root mean square error(RMSE)and other evaluation indicators,it is determined that the number of model iterations is 1000,the learning rate is0.01,and the number of hidden layer nodes is 32,and the prediction accuracy of the model is high.The vibration excitation test is carried out on I-beam.The vibration signals measured by fiber Bragg grating(FBG)sensor and piezoelectric acceleration sensor are used as training and test data,and the machine learning algorithm is used for prediction and analysis.(2)The self-attention mechanism method(SAM)and the LSTM model are combined to establish an improved SAM-LSTM model.Combining the advantages of the two methods,the SAM layer is added after the LSTM network layer.SAM assigns different weights of data to extract effective feature information.Based on the FBG sensor and piezoelectric acceleration sensor,the measured dynamic response data of the structure under healthy state are processed and used as SAM-LSTM input.The prediction results of the LSTM,RNN,ARIMA algorithm and the SAM-LSTM algorithm were compared,and the MAX,MAE,RMSE and other indicators were compared,and the difference between the predicted value and the measured value of the sequence signal by different algorithms was evaluated.The differences between the predicted and measured values of sequence signals by different algorithms are evaluated.It was verified that the improved SAM-LSTM had better prediction effect and the smallest error.Take acceleration prediction in healthy state as an example,the MAE value of SAM-LSTM is 74.5%lower than ARIMA model,62.9%lower than RNN model and 59.4%lower than LSTM model.The test results show that SAM-LSTM has high accuracy in predicting wavelength and acceleration data under healthy conditions.(3)When a sensor fails,data cannot be obtained correctly at the failed sensor.In this case,there is an urgent need for some way to reconstruct the missing sensor data.Taking the numbering direction of the sensors as the sequence direction,the time series of sensor fault locations is reconstructed using the monitoring data between adjacent sensors.From a practical engineering point of view,lateral signal prediction is performed based on SAM-LSTM.16 FBG sensors and 6 piezoelectric accelerometers are used to measure the dynamic response data of the structure in the healthy state,and the data is processed as the input of the model SAM-LSTM.Compared with other model evaluation indicators and prediction results,the experiment shows that SAM-LSTM can effectively predict the sequence data between different sensors,and the prediction accuracy is higher than LSTM,ARIMA,RNN.(4)Explore the recovery performance of SAM-LSTM model for missing wavelength and acceleration data under different damage conditions in practical engineering.Wavelength and acceleration under different damage conditions are used as model inputs.SAM-LSTM predicts damage information,and compares it with the prediction results of LSTM,ARIMA and RNN through evaluation indicators.And compared with the prediction results of wavelength and acceleration in healthy state.It shows that the SAM-LSTM method can effectively predict the wavelength and acceleration data after damage.It is of great significance to reconstruct missing damage data in practical engineering.