Research On Fault Diagnosis Method Of Non-stationary Rolling Bearing

With the continuous improvement of the level of national defense and military modernization,the level of automation and intelligence of military equipment is increasing day by day.The use of mechatronics and mobile equipment such as radar and communication equipment has greatly improved the military's combat capability.At the same time,certain problems also occur.Sometimes the equipment sometimes fails during the automatic or semi-automatic operation.Some failures often seriously affect the operation of the equipment and even cause the equipment to be unusable.Rolling bearings are common components in the transportation,deployment,scanning and other processes of motorized equipment,and are one of the weak links of motorized equipment.Therefore,the health monitoring and fault diagnosis of rolling bearings are very important for the healthy and safe operation of the entire equipment and fulfilling the mission of the combat unit.Significance.In this paper,combined with the rolling bearing failure,it will cause periodic shock vibration.The collection and analysis of vibration signals can effectively analyze the actual bearing state.For the bearing working environment,the noise is large,the vibration signal is non-stationary and nonlinear,and the signal is submerged in noise.The traditional signal processing method is used to analyze the non-stationary signal characteristics of the bearing,and it is found that the signal processing method is difficult to distinguish the fault type directly in a noisy environment.It is necessary to use further complex algorithms to calculate according to the signal characteristics.Satisfy.Further combined with the development of machine learning,the bearing fault diagnosis method based on the classic convolutional neural network is studied,but the convolutional neural network shows the problem of insufficient utilization of the signal time-frequency domain space and position relationship in the diagnosis,so this paper proposes a decentralized Attention capsule network model.The main research of the thesis has the following aspects.First,combined with traditional signal processing methods,the characteristics of non-stationary state signals are explored by means of time-domain,frequency-domain,and time-frequency domain analysis.Using the vibration signals of the same faulty bearing under different speeds and load conditions to analyze the time domain index,it is concluded that the time domain index of the bearing under non-stationary conditions does not have strict statistical law characteristics;the theory and envelope spectrum are calculated by the fault characteristic frequency The analysis method further verifies that the vibration signal under non-stationary state does not have obvious regularity in the frequency domain;using time-frequency domain analysis to further explore the signal characteristics,it is concluded that the non-stationary signal of the bearing can show on the joint distribution map of time-frequency domain.There is a fixed proportional relationship between frequency conversion and fault characteristic frequency.Secondly,according to the characteristics of the fixed frequency relationship curve between the frequency and the characteristic frequency in the time-frequency domain,the domain transformation diagnosis method based on the time-frequency domain conversion order domain is studied.A method for searching the maximum value of the module in a small range for a specific working condition is proposed,and the instantaneous speed is extracted in the time-frequency domain.According to the extracted instantaneous speed information,the method based on frequency domain coordinate reconstruction and time-frequency domain transformation order domain is studied to obtain fault diagnosis data with obvious order.Finally,after the transformation of the order domain,there are still problems such as fuzzy recognition and how to realize the intelligent diagnosis method.The intelligent recognition algorithm based on the traditional convolutional neural network is studied.The data in the order domain is reconstructed using coordinates,and deep learning is compared.The four classic models of Alex Net,VGG16,Googlenet,and Res Net50 in the development process analyzed the advantages and disadvantages of the model,and further proposed a method of distracting capsule network.The accuracy of model diagnosis has been greatly improved.The method in this paper has certain reference value for the study of fault diagnosis methods based on machine learning,and has certain practical significance for improving the intelligent monitoring and maintenance of equipment.