Optical Fiber Monitoring And Damage Identification Method Based On Deep Neural Network

Under the influence of various factors,bridge structures can suffer from different types of damage,increasingly eroding structural performance and leaving the potential for future structural failure.Timely detection of damage and formulation of reasonable measures can effectively reduce the probability of the extreme events.Currently,the main direction of research on damage detection methods is that the measured structural response data are processed and analyzed to obtain information about structural damage.And then the judgments about the structural condition are made.It is usually called data-driven method.However,a huge amount of data is accumulated under the long-term monitoring of bridges.And the general data-driven approach cannot effectively utilize such Massive data.Deep convolutional neural network has been proven to be very powerful in the field of image recognition to handle big data.In addition,conventional sensors such as accelerometers have many limitations for monitoring large structures,such as susceptibility to electromagnetic interference,the need to deploy dense sensors,etc.To address the above issues,this paper proposes a damage detection method combining deep learning and fiber optic sensors.In the meantime,the research studies the differences in the features learned by DCNN from different data and the damage detection performance.And it verifies the applicability of vibration-based transfer learning in damage detection.The main research content is as follows:1.The DCNN model is designed based on data type and number of sensors,including the selection of network depth,convolutional kernel size and number,etc.The network is then built using a deep learning framework based on the Python language.Simultaneously it is programmed to visualize the learned convolution kernels to better understand the operation mechanism of the models.2.The acceleration and wavelength data collected from the steel beam vibration experiment are used to build three data sets for the training of the constructed DCNN.The damage detection is then divided into a classification task and a regression task.Each task is utilized to identify the structural damage state,locate and quantify the damage respectively.Finally,the generalization performance of the training models using the test set data is validated and compared.The results of the study indicate that the model trained with the fusion data of acceleration and wavelength shows better performance in both tasks and each model learns different convolutional kernels.3.Feature parameters of the trained model in the steel beam damage detection task are transferred to the task of reinforced concrete beams.And the performance is compared with the model that is not loaded pre-trained model parameters.The test results show that transfer learning can greatly reduce the training time of the model and maintain a good prediction performance.The feature visualization results demonstrate that the model that has undergone transfer learning has approximately the same feature parameters as the pre-trained model.