Rail Defect Identification And Classification Method Based On Recurrent Neural Network

With the great development of the current railway infrastructure,China’s railway maintenance task is increasingly heavy.In order to fully ensure the safety of the train,it is necessary to timely detect the safety of railway-related equipment,especially the bearing body of the train rail.At present,the detection of rail defect still relies on manual means,which leads to the low efficiency of rail defect detection.To improve the efficiency of rail flaw detection and enhance the accuracy of defected data detection,it has become an urgent technical problem to realize the accurate identification and positioning of rail defect in rail flaw detection data through computer and other technical means.This paper collects the rail flaw detection data by using ultrasonic flaw detection equipment,and identifies the rail defect type in the rail flaw detection data by studying the relevant data processing methods,and help the flaw detection staff to improve the flaw detection efficiency.The main contents of this paper are as follows:(1)The data of rail ultrasonic flaw detection obtained by flaw detection equipment are cleaned and denoised.According to the encryption method and data format of different flaw detection equipment,the rail ultrasonic B-scan data is decrypted and analyzed,the original information is recovered,and the output is a digital sequence in time sequence according to the corresponding probe format.(2)According to the distribution characteristics of rail defect data in different channels,the detection modules of the railhead,rail waist,and rail bottom are determined to simplify the classification and identification of rail defected data.The independent sample size of each type of injury was determined.According to the damage spectrum of the B-scan image,the actual collected sequence data was segmented,and the positive and negative samples were distinguished by the color marking method.The standard data set was made in the above way.Based on this work,a typical defected database is established,and 260 thousand typical rail defects of all types are obtained,and 255 thousand data sets of rail defect have been produced.(3)According to the characteristics of different types of flaw detection data of railhead,rail waist,and rail bottom,the recurrent neural network models with long short-term memory unit(LSTM)as neurons are established under Tensor Flow architecture.The neuron structure,input layer,hidden layer,output layer,and other related parameters are determined.All data sets are sent to the LSTM model for training.After repeated parameter adjustment,the final result is obtained,and the optimal parameters are determined.Based on the existing data scale of the database,the accuracy of the railhead,rail waist,and rail bottom is 89%,96%,and 95%respectively.(4)The optimized model is encapsulated by the algorithm,and the appropriate threshold is set.A rail ultrasonic flaw detection data recognition and detection system based on a recurrent neural network is designed to identify and locate the rail defect in the flaw detection data and verify the reliability of the method.Finally,it is applied to the flaw detection system of railway rail flaw detection vehicle to help the flaw detection personnel accurately judge the type and location of rail defected.At present,the data recognition and detection system of rail ultrasonic flaw detection based on the recurrent neural network has been applied in the flaw detection workshop of Jinan public works depot.Both the accuracy rate and the false alarm rate are better than the existing system of flaw detection vehicle and manual judgment;At the same time,it can eliminate 100% of nonecho data and more than 90% of non-suspected echo data;the system significantly improves the efficiency of flaw detection data playback and re-inspection,and in the practical application of flaw detection,it can help flaw detection operators reduce the workload of flaw detection data playback by more than 80%,which greatly improves the work efficiency of flaw detection,timely and accurately locate the rail defect location,and greatly reduces the working pressure and labor intensity of flaw detection personnel.