Research On Diagnosis Method Of Deep Groove Ball Bearing Based On Domain Adversarial Network

Deep groove ball bearings play an important role in modern rotating machinery and equipment,and their damage is closely related to whether the equipment can run smoothly.Therefore,monitoring the running status of deep groove ball bearings and diagnosing the location and severity of their faults have become an important part in improving the stability of mechanical equipment.In recent years,intelligent diagnosis methods based on deep learning have been widely used in fault diagnosis.However,it is difficult to meet the conditions of sufficient training samples and complete fault state labels,and the problem of insufficient generalization ability when working conditions change also restrict its development.In order to solve these two problems,this paper proposes a deep transfer diagnosis model based on domain adversarial network.It is trained by the simulation dataset obtained from the dynamic model and the collected experimental dataset,and then the task of diagnosing the bearing fault category and severity can be completed.The dynamic model takes the deep groove ball bearing as the simulation object,it is based on the Hertz elastic contact theory,and considers six degrees of freedom in horizontal and vertical directions of the bearing inner ring,outer ring and bearing seat.The fault defects form of bearings are local single points and composite rectangular defects on the inner and outer raceways of different scales.By analyzing the geometric relationship and movement process when the rolling elements passing through the defect area,the corresponding displacement excitation function can be obtained,and then the contact force and damping force between the rolling elements and the inner and outer raceways can be calculated.Finally,the differential equations of the entire bearing vibration system can be obtained.The simulation dataset with sufficient samples and rich fault types can be obtained by solving the dynamic model,and then compare it with the experimental dataset collected by sensors in the time domain,frequency domain,time-frequency domain,and complexity index to verify their similarity,that provide a solid data foundation for the subsequent intelligent diagnosis work.The structure of the established transfer diagnosis model consists of three parts,which are the feature extractor which based on the ResNet network combined with the attention mechanism,the domain discriminator based on the global discriminator and subdomain discriminator,and the label classifier which aim to distinguish bearing fault states.In the feature extractor structure,the feature extraction process is optimized from the two perspectives of channel attention and convolution kernel attention.In the label classifier structure,the mixed loss function is used to improve the classification ability of the diagnostic model.And in the domain discriminator structure,the marginal distribution and conditional distribution is adapted at same time to improve model’s transfer performance.Then,an adversarial relationship between the domain discriminator and the feature extractor is formed,which can improve the parameter training speed and feature extraction ability of the built diagnosis model.The obtained simulation data and experiment data are transformed into two-dimensional time-frequency images after preprocessing,and then input them into the diagnosis model can complete the work of parameters training.Finally,the collected experiment data was used to verify the performance of the diagnosis models,and extending the models to variable working conditions.In comparison with the existing diagnosis models,the superiority of the proposed diagnosis method can be proved.