基于卷积神经网络与注意力机制的铁路基床翻浆病害识别研究OA

[Objective]Due to the concealed nature of mud pumping diseases in ballastless track subgrade,the detection and identification performance of track inspection vehicles and ground-penetrating radar systems is currently not significant.[Methods]Based on the measured vibration response data of subgrade mud pumping,a simple and efficient method for identifying mud pumping diseases in ballastless track subgrade,which combined an attention mechanism with a one-dimensional convolutional neural network,was proposed.[Results]The convolutional neural network incorporating the attention mechanism showed an average accuracy improvement of 30.4%,21.33%,31.25%,and 20%on the track slab acceleration validation set compared to traditional convolutional neural network and machine learning methods such as random forest,support vector machine,and gradient boosting tree.To identify the most suitable dataset for the model to recognize subgrade mud pumping diseases and enhance the overall performance of practical identification,six datasets—acceleration,dynamic displacement,and velocity of the base plate and the track slab—were utilized as model inputs.Additionally,the changes in accuracy and loss value with training epochs were analyzed.[Conclusion]The results show that the fitting process of the base plate dynamic displacement dataset is more stable,and the average accuracy on the validation set reaches 98.75%.Therefore,it is recommended to collect the base plate dynamic displacement dataset during on-site testing and utilize it as the training dataset for identifying subgrade mud pumping diseases to achieve better performance.