Deep Learning Structural Damage Identification Considering Optimal Sensor Placement

Structural Health Monitoring(SHM)is an important means to ensure the safety operation and healthy service of structures,which mainly includes three aspects: monitoring,diagnosis,and evaluation.Reasonable sensor placement is the important prerequisite to effectively monitor the structure and collect high-quality data.Structural Damage Identification(SDI)is the core part of SHM system,which directly determines the effectiveness and scientificity of structure status evaluation.Traditional damage identification methods are greatly influenced by expert experience and domain knowledge.In recent years,deep learning SDI technology has become a new emerging branch in SHM field,which has attracted scholars’ attention.This thesis focuses on the two main topics of sensor optimization and deep learning SDI in SHM,specifically,the work mainly includes:(1)A new individual initialization mechanism and two individual evolution strategies are proposed to meet the practical requirements of structural sensor optimization based on intelligent optimization algorithms.Considering the uncertainty of the number of sensors and the dynamic changing characteristics of their proportion in the candidate measuring points of the structure,an individual initialization method combining sensor coverage density is proposed.To address the defects of information singularity and certain blindness in the evolution direction during individual evolution in the artificial bee colony algorithm,a matching and preservation strategy for bee colony individual evolution mechanism is proposed.Combining the advantages of the simplicity and efficient iteration of the beetle antennae search algorithm,a beetle-swarm evolution competitive algorithm with a swallowing elimination strategy is developed.The validity and superiority of the proposed methods are verified based on the modal assurance criterion of a certain bridge structure,and two indicators are used to evaluate the optimization results of two types of optimization algorithms,obtaining a better sensor optimization placement solution with better performance.(2)A multi-objective optimization criterion for sensor placement optimization is constructed based on the diversity requirement of measurement point data in the process of structural damage identification.A multi-objective function expression is given that considers both the orthogonality of the structural modal shape vector and the completeness of the structural damage information by integrating the modal assurance criterion and Fisher information matrix.Taking tower steel structure and a lattice tower benchmark structure as examples,the effectiveness of the constructed multi-objective optimization criterion is discussed,and the sensor placement results under single/multi-objective optimization criteria are compared using two evaluation indicators,providing an effective basis for sensor optimization for deep learning SDI data collection.(3)To directly use structural acceleration response signals for single-domain data-driven deep learning SDI,a three-dimensional data construction method and a deep learning SDI model with three-dimensional data processing capability are proposed.To expand the feature information of the structural response data,a “major and subsidiary” data construction method is designed combining empirical mode decomposition and intrinsic mode function information.At the same time,a convolutional neural network model with three-dimensional data processing capability(3-Dimensional Signal Convolutional Neural Networks,3DS-CNN),whose network structure may be removal flexibly,is constructed.Using the acceleration data corresponding to the two types of structural sensor measurement results in(2)to construct a data set,the SDI performance of the “major and subsidiary” data construction method and 3DS-CNN model is systematically explored.Furthermore,the impact of sensor selection on SDI is discussed,providing an effective data set construction method and deep learning architecture model for cross-domain data-based SDI.(4)Considering the practical differences between structural numerical simulation data and actual measurement data,a transfer learning SDI network model based on cross-domain data feature Fusion(3DS-CNNF)is proposed based on the core idea of transfer learning and the designed 3DS-CNN framework.The 3DS-CNNF is trained with structural numerical simulation data to preliminarily learn structural damage features.The shallow network parameters of the network are frozen,and the end layer is replaced with a new classification one.The end layer parameters of 3DS-CNNF are fine-tuned using structural actual measurement data,gradually learning the common features of cross-domain data.The effectiveness of the proposed method is verified using cross-domain data of a tower steel structure,demonstrating that the optimized 3DS-CNNF model under the cross-domain data feature fusion strategy can overcome the problem of single application scenarios of 3DS-CNN model relying only on single-domain data training,and improve the SDI accuracy and generalization ability of the network model.The methods provide an effective approach and beneficial attempt for deep learning-based SDI under cross-domain data.