Researches On S700K Switch Machine Fault Diagnosis Method Based On Ensemble Deep Learning

As an important part of railway signal infrastructure,turnout plays an important role in controlling train running direction,maintaining passenger safety,and improving train speed.S700 K AC electric switch machine is mainly used as the switch equipment to control the action of turnouts in railway field.The normal working state of switch machine plays a key role in successfully fininshing starting,unlocking,switching,locking and other processes of turnouts.At present,the fault diagnosis of switch equipment is mostly realized by manually judging the length,peak value and trend of action current or action power curve of switches.However,this method has some disadvantages,such as poor diagnosis effect,low efficiency,heavy workload,lack of real-time,etc.It is gradually difficult to meet the requirements of railway industry operation and maintenance.In order to make up for the defects of traditional fault diagnosis methods,this dissertation uses an ensemble deep learning method to diagnose faults of S700 K AC electric switch machines.The specific research contents are as follows:Firstly,the relevant data of S700 K switch machines is divided into normal and common fault types,which are preprocessed to generate the switch machine action current curves,which is used as the dataset for neural network training in deep learning.Secondly,several Convolution Neural Network(CNN)models are established,and the dataset is input into the models for the fault diagnosis of S700 K switch machines.The influence of convolution layers,convolution kernel size and other parameters on the model performance is explored,and the models are comprehensively evaluated and selected according to the accuracy,precision,recall,and other related indexes.Thirdly,several Deep Residual Network(Res Net)models as Res Net-18 and Dense Connected Convolution Network(Dense Net)models as Dense Net-121 are established to diagnose faults of S700 K switch machines.The model diagnosis effect is visually evaluated by confusion matrix and ROC curve,and compared with CNN models to judge the significance of their optimization effect.Fourthly,Ensemble Learning(EL)is used to ensemble the three neural network models,including homogeneous ensembles,heterogeneous ensembles,and hybrid ensembles,to compare the optimization effect of various ensemble methods on the fault diagnosis accuracy of these models.Experimental results show that Res Net and Dense Net achieves better diagnostic effects than CNN,and the highest fault diagnosis accuracy reaches 97.95%.Simultaneously,the fault diagnosis effect of each model is improved in varying degrees after EL.Among them,hybrid ensembles have the best promotion effect,1.31%,and the highest accuracy is 98.63%.It demonstrates the feasibility and superiority of the fault diagnosis method.