Time-frequency Analysis Of Cement Concrete Pavement Broadband Array Audio Signals And CNN Of Pavement Void Classification Research And Application

Cement concrete pavement(concrete pavement)has the advantages of strong bearing capacity,high temperature resistance,long service life,etc.It is widely used in highway system construction.However,with the increasing traffic flow,the concrete pavement is repeatedly rolled by daily vehicles,coupled with natural factors,which directly affect the strength and bearing capacity of the pavement slab,changing its internal structure,thus shortening its service life.One of the most important damage phenomena is pavement void.The consequences of void not only cause cracks or broken slabs on pavement slabs,but also affect the safety of traffic and transportation.Therefore,it is of great significance to study the classification of void detection for concrete pavement.Firstly,this paper studies the detection methods of void at home and abroad,aiming at the traditional acoustic vibration detection method,it proposes to use microphone array to collect audio signals.Secondly,An improved CSM method based on the traditional coherent signal subspace(CSM)method is given,i.e.the differential space smoothing CSM method is used to estimate the DOA of the broadband array audio signal of the pavement void position.Then the signals are processed by adaptive beamforming and Wigner-Ville time-frequency analysis to obtain directional audio signals and time-frequency characteristic maps.Finally,the obtained time-frequency diagram is taken as the basic empty detection data,and classified and compared with the results by different Convolutional Neural Network structural models.In the network structure model,LeNet-5 and VGGNet-16 are studied firstly,and then two different improvement methods are given according to the advantages and disadvantages of VGGNet-16.One is to add convolution stone with the size of after the fourth convolution layer module,splice with the output result features of the fifth convolution layer module,and then enter the full connection layer.The other is to remove the first two full connection layers and add global average pooling.The main purpose of the two improved methods is to ensure the integrity of image features,reduce computational complexity and reduce overfitting.Through processing and analyzing the collected audio data,the results show that the CSM method based on differential spatial smoothing is more accurate than the traditional broadband signal source localization method in the process of DOA estimation.In addition,in the process of Convolutional Neural Network classification,the classification accuracies of LeNet-5 and VGGNet-16 adopted in this paper are 92.2% and 96.5% respectively.However,the classification accuracies of the two improved VGGNet-16 structures are 97.9% and 98.1% respectively.In the later stage,the robustness of the network model is analyzed by adding different intensity Gaussian white noise to the adaptive beamformed audio signal.The results show that the lower the signal-to-noise ratio of the signal,the lower the classification accuracy.For different signal-to-noise ratios,the network model has different degrees of decline,and the anti-noise ability is also different.