Data-driven Mrthod For Sound Source Locating And Damage Identification Of Composite Plates

Composite panels are widely used in various fields due to their various excellent properties.However,the continuous accumulation of damage during the service of materials will lead to a significant reduction in structural strength and seriously affect safety.In this context,the research on the structural health monitoring(Structure Health Monitoring,SHM)method and non-destructive testing(Non-destructive Testing,NDT)technology of composite structures is of great significance,an excellent damage identification method for composite material panels can provide strong support for staff to formulate maintenance strategies in a timely manner.However,due to the limitations of the material properties,the traditional damage identification methods for composite plates often need to rely on more complex physical models,which is difficult to balance accuracy and efficiency.In recent years,artificial intelligence methods that have been emerging continuously provide new ideas for SHM technology,and the powerful nonlinear mapping ability of neural networks can effectively make up for the shortcomings of traditional methods.Therefore,this thesis considers the combination of machine learning algorithm and non-destructive testing technology,and applies a sound source localization and damage identification method for composite material panels based on BP neural network.A metal aluminum plate and two carbon fiber reinforced composite material plates with different lamination directions are used as the experimental objects,and the position of the sound source on the surface of the plate is found by using the BP neural network.The neural network can be trained using only the dataset generated from the center point positions of each region to predict the location of the sound source inside the plate.The main research work is as follows:Firstly,follow the principle of simple to complex,and use BP neural network to conduct sound source localization experiments on isotropic metal aluminum plates with relatively simple material properties.Establish databases in two different environments,namely the COMSOL numerical simulation platform and the corresponding piezoelectric transducer(PZT)excitation-scanning laser Doppler vibrometer(SLDV)receiving physical experiment platform.Experimental results show that the trained model can map the arrival time of the first wave packet in the signal to the region where the sound source is located in both environments.In particular,the prediction accuracy in the numerical simulation environment is extremely high,and the area where the sound source is located can be 100% accurately located whether it is near the center of the area or near the border of the area.Due to the influence of various environmental factors in the physical experiment,the accuracy rate is slightly reduced,but the area where the sound source is located can still be located.Next,the proposed method is applied to composite panels with more complex material properties.The COMSOL numerical simulation platform is used to establish two composite material plate models with different material properties and simulate the propagation of five-peak narrowband modulated sinusoidal signals in them.The SLDV was used to build a corresponding non-contact experimental platform,and two composite laminates with different lamination directions were scanned to obtain initial time domain signals.After establishing the database respectively,input it into the BP neural network for training and testing.The results show that: under the numerical simulation environment,the prediction accuracy rate of the two plates can reach 100%whether it is near the sound source or at the boundary of the area,and the area where the sound source is located can be accurately located.In the physical experiment,the accuracy rate is slightly worse.In the orthotropic laminate,it can reach 100% near the sound source,and the boundary decreases to 96.3%.The same 100% accuracy near the sound source in the unidirectional laminate was reduced to 88.9% at the border.Finally,discuss whether the proposed method can predict the location of impact sources in composite plate structures without relying on material properties.Four piezoelectric sensor and an oscilloscope are used to build an experimental platform for impulse response testing,and the impact source is generated by manually releasing the steel ball.A training data set is generated in the center of each area,and the area to which the impact sources in other locations in the area belong is predicted,and the robustness of the model is tested at the same time.The results show that the trained neural network model can accurately predict the area where the impact source is located,whether in isotropic aluminum panels or composite panels with different ply directions.At the same time,the prediction result has strong robustness,changing the diameter and drop height of the falling steel ball within a certain range does not affect the accuracy of the prediction.The method of expanding the training set is used to improve the prediction error of the impact source location at the boundary of each region.After adding samples at the same distance of 60 mm from the center as the training set,the performance of the model was further improved,and the prediction accuracy also reached 100%.After refining the detection area to 16,the performance of the model did not decline significantly.After further refinement to 25 regions,the prediction accuracy and efficiency of the model have decreased to a certain extent,but acceptable prediction accuracy can still be obtained.