The Research On Fault Diagnosis Method Of Motorized Spindle Bearing Based On Convolutional Neural Network

As the core component of high-grade NC machine tool,motorized spindle has been used widely in varieties of sophisticated and special manufacturing areas.Its performance and operation state affect directly machining accuracy and product quality of the equipment.The rolling bearing is the main supporting form of the motorized spindle.Its life must be greatly influenced with the complicated operating conditions of motorized spindle.Once the bearing failure,it will cause the motorized spindle and even the whole production line to be paralyzed.In order to ensure the safe and reliable operation,it is essential to diagnose faults of the motorized spindle.With the development of deep learning,the concept about self-learning characteristics from the data itself provides a new research direction for fault diagnosis.Convolutional neural network as a method of deep learning with the strongest generalization ability,has a broad development prospect in the field of fault diagnosis.In this thesis,taking the motorized spindle bearing fault as the research object,the improved convolutional neural network is applied to fault diagnosis of the motorized spindle bearing.Input samples of this method are provided by time frequency analysis method which can transform vibration signal into images.The main work are summarized as follows:(1)Combined with theoretical knowledges of the motorized spindle bearing,its fault frequency characteristics were analyzed.Mild and heavy faults of bearing outer ring were simulated.The experimental platform of vibration signal collection in this thesis were described.Meanwhile,common processing methods of vibration signal were introduced.(2)The structure and training methods of convolutional neural network were researched.This thesis illustrated how the Tensorflow framework implemented convolutional neural network by using Python to write some important functions.At the same time,the excellent classification performance of convolutional neural network was verified by simulation signal of mild and heavy faults of bearing outer ring.With the short-time fourier transform,continuous wavelet transform and S transform,effects of training iterations on the classification accuracy were analyzed.(3)In order to distinguish fault types of motorized spindle bearing,a fault diagnosis method of motorized spindle bearing based on improved LeNet-5 convolutional neural network was designed.By comparing several time-frequency analysis methods,it was concluded that time-frequency images of vibration signal after S transform were the best.These time-frequency images were used for subsequent training and testing.At the same time,the LeNet-5 model was optimized and improved in many aspects.On this basis,a fault diagnosis method of motorized spindle bearing based on bagging and blocking convolutional neural network was proposed.The stability ability of convolutional neural network was further improved by fine-tuning training samples and determining fault types with voting principle.The experiment showed that the improved method can effectively distinguish the normal state and the three fault states of motorized spindle bearing.Compared with the other methods,classification accuracy of the improved method was improved and its training process was more stable.(4)Because of the insufficient learning of shallow convolutional neural network,it was difficult to distinguish faults with higher similarity.Therefore,a fault diagnosis method of motorized spindle bearing based on improved VGG-16 convolutional neural Fault signal which was collected by motorized spindle bearings with nine different working conditnetwork was designed to distinguish nine kinds of fault types with higher similarity.ions was transformed into input samples by S transform.The VGG-16 model was improved by using the Batch Normalization method.The results showed that the BN method not only performed better than the Dropout method,but also had the best performance when placed in the full connection layer.The improved method can not only make up for the deficiency of the superficial neural network,but also ensure the training process fast and steady,and improve the fault diagnosis rate.The fault diagnosis method of motorized spindle bearing based on convolutional neural network can identify effectively the fault classification of motorized spindle bearing and improve the accuracy of fault diagnosis.