Subjects: Mechanical Engineering >> Other Disciplines of Mechanical Engineering Subjects: Computer Science >> Computer Application Technology submitted time 2017-12-26
Abstract: There are some shortcomings in the existing methods of fault diagnosis of planetary gearbox: First, the traditional methods are complex and can not effectively diagnose the planetary gearbox faults. Second, the methods based on convolution neural network mostly diagnose gearbox faults and rarely are used to diagnose planetary gearbox. In order to effectively diagnose complex faults and variable working conditions, fault tree structure, working condition parallel structure and multi-SoftMax convolution neural network are proposed for the first time. Fault tree structure can handle a variety of complex faults and see the diagnosis effect of each node. The parallel structure can handle variable conditions, and predict speed and load. A series of experiments are carried out using the vibration data of ours laboratory planetary gearbox, which indicated that the method can accurately diagnose the complex faults and variable working conditions of the planetary gearbox, and the accuracy is 97%. It is verified that the multi-SoftMax convolution neural network has strong generalization ability, and the advantages of the fault tree structure.
Peer Review Status:Awaiting Review
Subjects: Mechanical Engineering >> Other Disciplines of Mechanical Engineering Subjects: Computer Science >> Computer Application Technology submitted time 2017-12-26
Abstract: The existing methods of bearing diagnosis have some disadvantages: The conventional method has complex mathematical calculation and poor diagnosis effect. It generally only diagnoses the fault location and irrespective of the load and the fault size. The existing convolutional neural network method use the traditional convolution neural network. A network can only output a property and can not simultaneously diagnose multiple properties. In order to simultaneously diagnose the fault location, fault size and load, for the first time put forward a multi-attributes convolution neural network (MACNN) and applied to the bearing fault diagnosis. The multi-attribute convolution neural network is trained using one-dimensional vibration signal training . The advantages lies in overcoming the shortcomings of the traditional method: the diagnosis result of any combination of the fault attributes can be obtained, the network parameters are less, the method is simple, the generalization ability is strong and the accuracy rate is high. A series of tests have been carried out using the bearing data of Case Western Reserve University. The results show that the proposed method can accurately diagnose several properties of bearing faults with high accuracy and good generalization ability.
Peer Review Status:Awaiting Review
Subjects: Mechanical Engineering >> Other Disciplines of Mechanical Engineering Subjects: Computer Science >> Computer Application Technology submitted time 2017-08-29
Abstract: The traditional methods of qualitative diagnosis of bearing require complex and difficult knowledge of mathematics and deep domain knowledge. The method based on deep belief network overcomes the shortcomings of traditional methods but the amount of network parameters are huge thus the netword is difficult to train. Convolution neural network time-frequency image method uses wavelet packet transform to obtain time-frequency image. As the strong feature learning ability and generalization ability of convolution neural network a qualitative fault diagnosis method of bearing based on convolution neural network is proposed. Convolution neural network is trained directly on one-dimensional vibration signal. The advantages of this method are to overcome the shortcomings of the traditional methods the amount of parameters are much less and training is effective compared to the deep belief network don’t need use wavelet packet transform to obtain time-frequency image. A series of comprehensive tests are carried out by using the data of Case Western Reserve University and our own laboratory. They show that the method can diagnose the bearing fault accurately and the accuracy is higher than that of all other methods. For the first time the convolution neural netword trained on Case Western Reserve University’s data accurately diagnoses the fault type of our own laboratory’s bearing which indicates that the method can be used in practical.
Peer Review Status:Awaiting Review