Rail Defect Detection Based On Memory-Linked Metric Learning Anomaly Detection

With the development of railway,high speed and intensive rolling will significantly increase the probability of various fatigue damage and they lay a hidden danger for major safety accidents.Therefore,it is vital to detect defects of the rail.ECPT is a mature method for defect detection which can detect surface and subsurface defects simultaneously.The commonly used algorithms mostly trained with the defect samples whereas this data is rare and cannot cover all types of defects.Whereas anomaly detection frameworks trained with large amount of accessible background samples,and is able to detect various types of defects.However,most anomaly detect algorithms work in unsupervised mode,and can not find the weak anomaly areas accurately.We proposed metric learning module to modify the memory linked auto-encoder and apply it to detect rail defects with ECPT system.The proposed algorithm models the normal samples and add some defect samples to supervise the training process firstly.Then,the trained model defect the anomaly areas according to the learned normal patterns.The embedded metric learning module help the model learn robust representation of the background patterns and enlarge the distance between defect patterns and background patterns in the feature space.The main contribution of the work is as following:· An ECPT based rail defect detection system was established to test the quality of track information collection under different conditions.After the system parameters were optimized,thermal images of the excited part of the rail with multiple types of defects were collected,and the effective data were made into anomaly detection datasets.· An embedded metric learning module is proposed to improve the memory linked autoencoder.The modification optimize the performance of weak defect detection.The proposed model work in semi-supervised mode and can learn representative memory items to describe the background patterns.Experimental results show that the metric learning module can improve the accuracy of the model and ensure the generalization of the model.· Based on the above models,multi-memory module is proposed to improve the adaptability of the model.The proposed module enables the model adapt to various kinds of backgrounds.Experimental results show that the multi-memory module can improve the adaptability of the model and supress the performance descend caused by catastrophic forgetting.