Research On Rail Ultrasonic Flaw Detection Method Based On Machine Learning

As an important part of the train track,steel rail will inevitably have various injuries to its internal structure after long-term service.Damage detection of in-service rails and timely elimination of structural diseases can reduce the safety hazards of train operation.At this stage,China’s railway department adopts the operation method of "skylight repair" in terms of rail flaw detection,and uses relevant equipment to carry out ultrasonic flaw detection on the rail during the skylight period;After completing the on-site flaw detection operation,and then deliver the flaw detection data to the data playback personnel,the manual operation method is used to further analyze the flaw detection data,at this stage,railway employees need to maintain a high degree of concentration,carefully watch the B display map of the flaw detection data,find out the damage information of the inspected rail,and the re-inspection of the flaw detection data often takes a lot of time.At present,the contradiction between heavy flaw detection tasks and insufficient efficiency of manual injury re-inspection is becoming increasingly acute,and it is urgent to improve the playback efficiency of flaw detection data.In view of this problem,this paper studies the ultrasonic flaw detection data of rails,builds corresponding damage classification models for different parts of rails based on relevant machine learning algorithms,and writes a damage detection software to assist railway departments to complete flaw detection data playback work more efficiently.The main research contents of this paper are:(1)Preprocessing of ultrasonic testing data for steel rails.By means of data pruning,a large number of B-display patterns without echoes are removed from the original data,completing the cleaning work of the flaw detection data.And decrypt the flaw detection data with different encryption methods,converting it into a table file format containing information about each detection channel.(2)Production of rail flaw detection sample dataset.Based on the detection principle of the ultrasonic probe equipped in the rail flaw detector,the rail detection parts corresponding to the data in different detection channels are clarified,and the corresponding independent flaw detection samples are extracted from the rail head,rail waist and rail bottom of the rail.At the same time,the normal sample data and the damage sample data are distinguished by the color marking method,and the flaw detection data of different parts of the rail is made.(3)Research on the imbalance of injury data sets at the data level.Aiming at the imbalance of sample categories in the dataset,the data enhancement method based on the conditional tabular generative adversarial network(CTGAN)is studied,and the generative model is used to synthesize the sample data that approximates the true distribution of rail damage,which expands the number of minority samples in the unbalanced dataset,and balances the positive and negative category ratio of the dataset to a certain extent.(4)A rail damage classification model was constructed,and the category imbalance problem was studied at the algorithm level.Aiming at the damage distributed in different positions in the rail,a flaw detection data classification model of rail head,rail waist and rail bottom is constructed based on the extreme gradient boosting algorithm(XGBoost).At the same time,the Focal Loss function is introduced into the XGBoost model,which further improves the classification accuracy of the model.Finally,under the balanced test set,the accuracy of the three injury classification models constructed reached 90.1%,97.6% and 94.7%,respectively.(5)Develop rail damage detection software.Based on the improved XGBoost rail damage classification model,a rail damage detection software has been developed,achieving the goal of efficiently analyzing flaw detection data using computers.This is beneficial for flaw detection data playback personnel to more efficiently complete damage re inspection work,thereby further improving the damage detection rate and damage determination accuracy.At present,the rail damage detection software developed in this paper has been applied to the Jinan Engineering Section.Compared with the traditional manual injury re-inspection work,the use of this software to detect and analyze the flaw detection data shows that the detection efficiency of rail damage is higher,and the detection rate of damage is also improved.