Research On Metal Fatigue Crack Detection Method Based On RNN Eddy Current Pulse Thermography

In the process of long-term complex operation of high-speed rail vehicles,the axle,wheelset,bogie and other key structural parts bear large loads,and are prone to fatigue cracks,impact damage,weld cracking and other defects.If it is not found in time at the early stage,the crack would be expanded gradually and consequently the whole equipment would be damaged,resulting in the loss of human and material resources,and even the existence of significant security risks.Therefore,it is a very important task to check the health of the key structural parts of the train regularly to ensure the safety of the equipment,which has considerable feasibility study potential.Eddy current pulse thermal imaging is a new non-contact nondestructive detection method for surface defects.It uses a device that generates a high-frequency pulsed electric field to generate eddy current on the surface of the tested specimen,and then detects the defect by measuring the heat change generated by the eddy current through a thermal imager.Analysis of the variation of heat image can ascertain information such as the type,a position and size of defects.and other information,which can be effectively applied to the detection of metal cracks.However,this technique usually relies on human experience to extract features for crack detection and recognition,so it is difficult to achieve automation and intelligence.To overcome these problems,this thesis presents an eddy current pulse thermal imaging metal fatigue crack inspection and recognition method using recurrent neural network(RNN),which combines the features of eddy current pulse thermal imaging technology and recurrent neural network,in order to achieve automatic and intelligent detection and recognition of metal crack defects.In this paper,the metal plate was taken as the research object to analyze the temperature distribution characteristics of the specimen when it was excited by eddy current pulse.In order to make the heat generation effect of crack region better,the control variable method was used to study the influence law of the excitation intensity,excitation distance and excitation time on the heat generation of crack.The temperature distribution of the surface of the tested metal specimen was obtained by infrared thermal imager.Input the collected infrared thermal images to the PC terminal for real-time data processing to obtain eddy current pulse thermal images and make data sets.Firstly,the HOG-LSTM model is devised according to the recurrent neural network(RNN)and applied to the classification of whether the eddy current pulse thermal image contains cracks.Then,according to the obtained data sets of eddy current pulse thermal images with different sizes,the designed Bi-LSTM model is used to enhance the time sequence information in the feature vector,and the eddy current pulse thermal images with different sizes of cracks are extracted and classified.The proposed results demonstrated that the Bi-LSTM model was effectively detecting metal crack defects and accurately identifying crack sizes.Relative to more conventional computer learning patterns and other deep learning patterns,the algorithm has greater discrimination accuracy,and the recognition and classification accuracy of existing crack image data sets reaches 100%.It can effectively detect and identify metal fatigue cracks,and has a wide range of application prospects for automatic detection of metal fatigue cracks.