Research On Fine Grained Recognition Algorithm For Axis Orbit Of Rotary Machine Based On Deep Learning

With the development of China's manufacturing industry,fault diagnosis and monitoring of large rotary machine is coming increasingly important,while axis orbit is a powerful signal that can reflect the running state of rotary machine.However,traditional method to identify axis orbit of rotary machine usually chooses the pattern recognition method of "Feature extraction + Classifier",which has not been able to effectively identified the severity of rotary machine fault,due to the factors such as the lack of ability of feature to describe faults and the lack of recognition performance of classifier.Based on all above,the research on fine-grained pattern recognition of the axis orbit is carried out by using deep learning.On the basis of summarizing the research status of the fault mechanism of the axis orbit of rotary machine,the purify of the axis orbit and the automatic identification of the axis orbit,the main problems existing in axis orbit identification are analyzed and the related axis orbit databases are established.Firstly,the axis orbit database named “A” containing different faults has been established.Besides,five indicators based on slenderness,degree of bending,width ratio between the rings,angular span ratio,and minimum radius of curvature have been put forward to match different severity of faults.So that the fine-grained graphics database of axis orbit,named “B”,which contain 15 kinds of graphics is established.After that,the convolution neural network(CNN)for axis orbit recognition is constructed based on the CNN model LeNet-5.And the most suitable network structure for fine-grained axis orbit recognition is obtained by experiments on adjusting the structural parameters of the network.Firstly,on the basis of the original LeNet-5 model,the number of linear combinations between the convolution layers is increased to improve its capability of feature extraction.Then,full connected layer is reduced to streamline its network structure and a softmax layer is added to optimize its classification ability.Finally,by using database “B”,the structure parameters of this CNN network are optimized from several aspects,including the number of convolution layers,the number of convolution layers' channels,the number of layers and nodes of the full connected layers,and the use of Relu layer and Dropout layer,etc.In order to test the effectiveness of this algorithm,the optimized CNN network,some existing main algorithm of axis orbit recognition and the related feature descriptors are tested in the simulated database and actual measured axis orbit database respectively.The tests in simulated databases “A” and “B” show that the improved axis orbit recognition algorithm based on LeNet-5 has the highest recognition rate in all databases.In the fine-grained graphics library of axis orbit “B”,its accuracy rate is up to 97.63%.And the real-time performance is also the best.The recognition speed of single sample in database “B” is 0.0427 milliseconds.In addition,the rotor test rig,“STS1000” on-line vibration monitoring and analysis system was built to simulate different degrees of rotor unbalance faults and to obtain the fine-grained figure of the unbalanced fault.The experiment on measured fine-grained axis orbit of unbalance fault shows that the algorithm proposed in this paper also has the highest accuracy,up to 97.14%,and only 0.128 milliseconds for single sample recognition.The experimental results show that this algorithm is fast,accurate and robust,and can meet the actual requirements of the fine-grained recognition of axis orbit.