Full Cycle State Evaluation Of S700K Switch Machine Based On Convolutional Neural Network And Fuzzy Clustering

The S700 K electric switch machine is an indispensable outdoor signal device in high-speed railway transportation in our country.Its role is mainly to convert,lock the turnout and timely feedback on the position and working status of the turnout.At the present stage,the condition diagnosis of S700 K rutting machine in China basically relies on manual experience and regular troubleshooting,testing and maintenance.However,with the increase in high-speed railway lines and the increase in total mileage,this method can no longer meet the current fault monitoring.and it is of great significance for the operation safety of high-speed rail to effectively identify the working state of S700 K switch machine,namely normal,sub-health,fault and serious fault.is of great significance to the operation safety of high-speed rail,especially the discovery of sub-health state,will greatly improve the maintenance efficiency,the accident will be excluded in the embryonic stage.In order to solve the above problems,we propose a new algorithm based on convolutional neural networks and fuzzy clustering to diagnose the running state of the S700 K switch machine.The research content of this thesis mainly includes:First,the research status of fault diagnosis and state evaluation is analyzed,and use the switch machine power curve rail switch machine can represent the relationship between the performance of the state.Taking S700 K switch machine as the research object,aiming at fault diagnosis and full-cycle state assessment of switch machine,a feature extraction method based on convolutional neural network is proposed by taking advantage of the ability of convolutional neural network to extract tiny features.A lightweight convolutional neural network was used to remove the softmax classification layer and retain only the first 11 feature extraction layers.The first 11 layers of Squeezenet were used as feature extractor to extract 1000 features of power curve of S700 K switch machine,and the standard model library was established.For the whole-cycle state assessment of the switch machine,a feature extraction method of switch machine based on an improved residual network was proposed.Res Net network was selected to design a dimension adaptive global mean pooling layer(GAP)instead of the full-connection layer(FC).A few samples were used to train GAP parameters,and the power curve image of switch machine was input into the network,and in the GAP layer output characteristics of 512 data,establish the switch machine health status evaluation characteristic vector library.Secondly,in the aspect of fault diagnosis,fuzzy model recognition was used to replace the classification layer of Squeezenet.The fuzzy model recognition algorithm was used to calculate the grid progress between the standard model library and the samples to be tested,and the classification was achieved according to the principle of proximity.In a state of full cycle evaluation,a global average pooling layer is used instead of Res Net’s fully connected layer,and the GAP layer parameters are trained using a small amount of data.Secondly,fuzzy clustering is used to replace the classification layer of Resnet network,and feature vectors are output in GAP layer to construct the feature vector matrix between the health state feature vector library and the samples to be tested.The dynamic clustering diagram is obtained by fuzzy clustering,to realize the status assessment of the switch machine.Meanwhile,because of the universality of the method,the bearing open data set of Case Western Reserve University is used for experimental verification.The results show that the improved convolutional neural network method proposed in this thesis is not only applicable to the status diagnosis of the switch machine but also applicable to the fault diagnosis of bearings.The number of training samples can be appropriately increased or decreased according to the curve complexity of the detection target to improve the feature extraction ability of the convolutional layer and increase the classification accuracy.