Rotor Shaft Trajectory Pattern Recognition Method Based On Convolutional Neural Network

Rotor is one of the key components in rotating machinery and equipment,this is because the running state of the rotor reflects the running state of rotating equipment,once the operation of the rotor failure,will lead to unstable running rotating machinery downtime,even,in turn,bring immeasurable loss,due to uncontrollable factors such as work environment,the influence of the rotor will appear during the operation of different problems.Therefore,it is of great engineering significance and academic value to carry out fault diagnosis of rotor state of rotating machinery equipment for preventing accidents and maintaining the performance of equipment running.Axis path as an important way of detecting rotor,its development trend in intelligent diagnosis as the core,the traditional way is to manually extract the feature classifier combination,the axis trajectory classification,and combined with prior knowledge,with the axis of track types corresponding equipment running status,thus infer the fault,with the development of the computer,the axis path deep learning was applied to the identification of hot gradually,however,deep learning applied to the axis of trajectory identification model of network difficult training and face less data difficult to deploy to the hardware problem,in view of this the paper puts forward the data to enhance the preliminary training of network application in the axis trajectory identification,And the light weight network is applied to the problem of axis trajectory identification.On this basis,some research problems are carried out for the identification of axle center trajectory.The main contents and innovations of this paper are as follows:(1)To solve the overfitting problem that is prone to occur when deep convolutional neural network is applied under small data,this paper proposes two major deep learning-based solutions.The first method is the enhancement of the axis trajectory data.Based on the obtained manual classification of the axis trajectory database,data enhancement is carried out before input to the network to expand the size of the data set,so that the model can obtain sufficient trainable data.In order to solve the problem that data enhancement after more than 5-7 times of data enhancement has little effect on the network,a pre-training network model is proposed to identify the axis trajectory.The priori knowledge accumulated by deep neural network in image classification tasks on other large data machines is used as feature extractor to recognize the axis trajectory.The pre-training model is fine-tuned to make it more suitable for the current axis trajectory classification task.The proposed method has advantages in the accuracy of small sample axis trajectory identification,which is verified by the simulation database and the unbalanced axis trajectory data obtained from experiments.(2)In order to solve the deployment problem of the axis trajectory recognition model based on deep convolutional neural network,this paper proposes the method of applying the lightweight model to the axis trajectory recognition.In this method,Mobile Net is applied for pre-axis trajectory recognition task based on deep separable convolution.The network depth is deeper,but the number of parameters is only one-ninth of the convolutional neural network based on the same depth of ordinary convolution.For the problem that the network is too deep and the training on small data is easy to overfit,the pre-training lightweight network is combined with the axis trajectory recognition task,which not only realizes the application of the deep network lightweight model in the axis trajectory recognition,but also avoids the overfit caused by the small sample training deep convolutional neural network.The experiment is carried out on the simulation axis trajectory database and the experimental axis trajectory database.Compared with one ninth of the same deep network and one fourth of the same type network,the neural network-based axis trajectory recognition model has achieved parameter lightening at the cost of 4% accuracy.