Machine Learning-Assisted Timber Structural Damage Detection Based On Smart Sensing

Timber structures are the treasure of China’s architectural culture.Because of its green,low-carbon,light weight,and high strength,it has gradually been the prospective structural form under the "double carbon" national strategy.However,in long-term service,timber structures are inevitably degraded due to the coupling effects of environmental variations,creep effect under long-term load,and sudden disasters,resulting in accumulated structural damage.In extreme cases,the above damages may even lead to catastrophic accidents such as structural collapse.Thus,it is of great significance to identify structural damages which would affect the service performance of timber structures to ensure structural service safety.The elastic wave-based non-destructive testing(NDT)technology shows increasing attention to identifying wood structure damage and evaluating service performance.Compared with the traditional micro-drilling resistance method and structural modal analysis method,the elastic wave-based detection method can realize the active identification of local damage in specific areas according to the propagation properties and signal characteristics of the elastic wave at different structural damage boundaries.However,the existing non-destructive testing of timber structures based on elastic waves mainly relies on the manual inspection with instruments and equipment,which efficiency and accuracy are highly affected by operators.Thus,this research proposes a novel timber damage detection method by integrating piezoelectric-derived stress wave sensing and machine learning-assisted signal interpreting,aiming to achieve intelligent monitoring and accurate identification of typical timber structural damages that would affect structural performance,such as moisture content,cracks,internal void,bolt looseness,and stiffness degradation.The research has improved the detection efficiency,identification accuracy,and high-efficiency processing capacity of massively data from existing nondestructive testing methods and has particular scientific significance and application value for ensuring the service safety of timber structures in China.The main contents involved are as follows:The propagation law and frequency dispersion characteristics of stress waves in various anisotropic wood were studied.Firstly,based on the basic principle of elastic wave propagation and the assumption of wood material properties,the dispersion equation of elastic wave in anisotropic wood is deduced through substituting geometric equation,constitutive equation,and stress boundary conditions into the movement equilibrium equation.Then,the theoretical solution of simple components is extended to the numerical simulation of complex members.Finally,the simulation principle and parameter selection of elastic wave propagation are clarified.A detection method for wood moisture content based on piezoelectric sensing and the na(?)ve Bayesian algorithm is proposed.Firstly,the dynamic simulation of elastic wave propagation is conducted to investigate the propagation characteristics of stress waves in wood components under different moisture contents.Then,the feasibility of using piezoelectric materials as actuators and sensors to identify the moisture content of timber structures is validated through experiments,and the correlation mechanism between the feature of stress wave signals and moisture content is analyzed.Finally,based on stress wave signals obtained in tests,the eigenvalues in the time domain and frequency domain are extracted as training features of the machine learning algorithm,and a na(?)ve Bayesian classification model is established to identify moisture content.A crack identification method of timber structures based on piezoelectric sensing and convolution neural networks is proposed.Firstly,the stress wave propagation of the numerical model under different crack depths is analyzed through elastic wave dynamic simulation,and then the piezoelectric sensing method is used to identify timber structural cracks through tests.The short-time Fourier transformation is conducted based on the time-series stress wave signals obtained in tests.The spectrum containing time-frequency domain damage information is used as the input of convolution neural networks for training and feature learning.An identification method of internal holes in timber structures based on the acoustic vibration method and convolution neural networks is proposed.Firstly,the vibration equation of particles on the surface of the timber column with internal holes is deduced,and the relationship between hole diameter and vibration frequency is established.Secondly,the multi-physical field coupling simulation is carried out based on the numerical model,and the correlation between the vibration frequency of particles in the airfield and particles at the surface of the timber column is analyzed.Then,the experimental verification of the proposed method is conducted by considering the influence of multivariable factors.Mel cepstrum transform is applied to sound signals collected in the experiment,and the spectrum is used as the input of deep convolution neural networks for training and feature learning.Finally,the robustness of the proposed method to environmental and operational factors is verified.A detection method of looseness of bolted joints in timber structures based on the piezoelectrical impedance method is proposed.Firstly,based on the self-excitation sensing characteristics of piezoelectric materials,the correlation equation between bolt preload and piezoelectric impedance of sensor-timber structure system is deduced.Based on the correlation equation,the influence of preload on electromechanical impedance is studied through parameter analysis.Considering the requirement of longterm monitoring for the durability of piezoelectric sensors,a smart washer that can integrate with the timber bolted joint is designed.Then the influence of environmental temperature and humidity on the performance of the smart washers was verified by tests.Finally,the effectiveness of using the electromechanical impedance method to identify the looseness of timber bolted joints is verified through experiments.An identification method of stiffness degradation of timber structure joints based on piezoelectric sensing and machine learning algorithms is proposed.Considering the important role of joints in energy dissipation of timber structures under horizontal earthquake actions,taking the mechanical behavior stage of joints under monotonic loading and low cyclic loading as the identification object,the piezoelectric sensing method is used to percept damage.Then,extract the eigenvalues of stress wave signals collected in tests in the time and frequency domain as the input of machine learning algorithms.Based on the machine learning classification algorithms,the identification method of stiffness degradation of wood structure joints is established,and the identification performance of various machine learning algorithms is compared.