Research On Lamb Wave Modal Decomposition And Crack Damage Monitoring Technology Of Rail Weld Based On GAN

As of 2020,my country's national railway operating mileage will reach 146,300 kilometers,of which 38,000 kilometers are high-speed railways,which play an important role in ground passenger and freight transportation.In the railway track transportation network,the fatigue loss of rail welds due to the force and pressure generated by the wheel-rail relationship when the load train passes through often produces small cracks that are more difficult to detect.Small cracks develop into large cracks through more fatigue loss.Cracks bring the risk of rail breakage.Based on this article,I did some research:· Starting from the goal of detecting rail weld cracks,using the basic theory of Lamb wave propagation in elastic media,the propagation characteristics of Lamb waves under anisotropic material of rail welds are explored,and the slow propagation process of Lamb waves is analyzed.Velocity curve,phase velocity dispersion curve and group velocity dispersion curve;· Explored the theoretical basis needed to build the Lamb wave modal decomposition and crack detection model based on the generative countermeasure network,and theoretically verified the feasibility of the underlying mathematical principles and model of the generative countermeasure network to converge in the target task of this article;· With the help of the calculated Lamb wave dispersion curve and the excitation frequency of the Lamb wave in the data acquisition experiment,the sensor layout plan in the data acquisition process is designed.Perform feature engineering analysis on the collected real damage data,and use the analysis results to further design the damage data collection plan under the simulation state,and expand the sample distribution of the overall data set;· The time-domain and frequency-domain characteristics of the simulation data set are explored separately,and the data preprocessing scheme is designed to maximize the difference of the crack damage data characteristics by using the time-frequency characteristics analysis results of the complete data set.Preprocess the overall data set according to the plan,and use the preprocessed data set as the training data and test data of the model;· In the model building stage,according to the characteristics of the data obtained from the analysis,this paper designs a two-stage model according to the special structure of the adversarial generation network and the transfer learning theory in the model compression theory.And according to the different tasks performed by the model and the difference of the model structure,the loss function and the network optimizer of the two-stage model are designed separately.At the same time,use the results of the design to train the network to obtain a two-stage model with different characteristics;· According to the different tasks performed by the model,the evaluation indicators of the model are designed,and the indicators of the two-stage model based on different network structures are evaluated.The calculation method of the model decline rate index is designed to calculate and analyze the decline rate of the second-stage model.At the same time,the model cost-effective index calculation method is designed to calculate and evaluate the cost-effectiveness of the twostage model,and select the most cost-effective model.The article also compares the two-stage model with advanced deep learning algorithms and traditional digital signal processing algorithms on different tasks.Visually compare the results of the model's task execution with the results in the real data training set.The experimental analysis results show that the two types of models have good evaluation indicators for the tasks corresponding to the real data sets collected in the experiment.In the main tasks,the various task indicators of the model are close to the effect of the advanced model algorithm,indicating that the model can be used in most tasks.Complete tasks well and efficiently.At the same time,the comparison of the visualization results of the model shows that the model has a strong task execution ability under real conditions,and is highly sensitive to rail weld cracks,showing an effect close to the current optimal deep learning model or digital signal processing algorithm.