Structural Vibration Signal Recovery And Damage Identification Considering Sensor Faults Based On Deep Learning

Sensors are a key component of the structural health monitoring(SHM)systems,which may experience performance degradation or even failure during service.The error signal collected by the fault sensor will affect the accuracy of the evaluation result of structural health status.Therefore,the sensor signal recovery and unsupervised structural damage identification under sensor failure are studied,in order to effectively diagnose the structural state under the condition that the sensor fails and no damage data can be used for training.The main research contents and achievements are as follows:(1)A fault diagnosis method of acceleration sensor based on convolutional neural network(CNN)is proposed.A sensor fault diagnosis model based on convolutional neural network was established by taking the vibration acceleration time history data of the structure as the data set,and the sensor fault type and fault location were determined.The feasibility of the proposed method is verified by IASC-ASCE SHM Benchmark phase I numerical model and a suspension bridge,and the influence of different noise levels on the sensor fault diagnosis results is discussed.(2)A fault sensor data recovery method based on deep convolution generation adversarial network(DCGAN)is proposed.According to the sensor fault diagnosis result,the data set is processed accordingly.The signal of the faulty sensor is taken as the input data of the discriminant network,and the signal recovery model based on DCGAN is used to recover the fault sensor signal through the correlation training between the latent features of the residual health sensor signal and the fault sensor signal.The feasibility and reliability of the proposed method are verified by IASC-ASCE SHM Benchmark phase I numerical model and a suspension bridge,and the effects of different fault sensors and different noise levels on the signal recovery results is discussed.(3)An unsupervised structural damage identification method based on the improved GAN(IGAN)is proposed.The structure of the generated network is encoder-decoder-encoder.The encoder reduces the dimensionality of the input vibration response to a lower-dimensional vector,and then the decoder decodes the vector and reconstructs the generated sample.Finally,the resulting sample is encoded by additional encoders,which output it into a representation of its latent features.Minimizing the distance of latent features after encoding twice during the training process learns the basic features hidden behind complex data.During the test phase,unknown state data is entered and the difference between the latent features of two different codes is used to evaluate the structural state.The feasibility and reliability of the proposed method were verified by IASC-ASCE SHM Benchmark phase I numerical model,IABMAS BHM Benchmark numerical model and aluminum frame structure in laboratory.The effects of different noise levels and sensor positions on damage identification results were discussed.