Fault Detection System For Motor And Reducer Of Welding Robot In Automobile Production Line

With the development of industry 4.0,welding robots support many high-end industries.The use frequency of welding robots in factories is generally increased.In order to avoid sudden failure and delay in production,it is extremely necessary to monitor its status and detect its failure.At present,fault detection based on large data is the main way to identify equipment faults.However,inadequate fault data labels can not be obtained,resulting in serious data imbalance,which affects the judgment of fault type.In this paper,a data acquisition method based on dynamic simulation is proposed to expand the data set,because there are some differences between simulation data and real data,and then a transfer learning method is proposed to realize fault diagnosis of main components of the robot.In this paper,according to the operating state and vibration characteristics of RV Reducer,the translational torsional dynamic model is constructed by using the centralized mass method,and the motor rotor dynamic model is constructed by Jeffcott rotor system.For the fault,the time-varying meshing stiffness of RV Reducer Gear and cycloid gear pitting fault is solved by using the energy method,and the differential equation is solved by using the fourth-order Runge Kutta method to obtain the fault characteristics,and then the test-bed is used to collect the control data,Verify the effectiveness of the model and provide data basis for migration learning.For feature extraction,a time-frequency processing method of variable window short-time Fourier(FSST)is adopted,and compared with short-time Fourier(STFT)and continuous wavelet transform(CWT)in terms of data richness and clarity.It is found that while increasing the amount of data information,it inevitably leads to the decline of readability.Through the verification of the above time-frequency extraction method by neural network,it is concluded that the improvement of the amount of information will get better performance in model training.The two-dimensional data and one-dimensional data of simulation faults are classified by using DenseNeural Network(DenseNet)and Deep Forest(DF).At the same time,the comparison between random forest and traditional classification algorithms proves that deep network has natural advantages for representation learning.Maximum Mean Difference(MMD)is introduced to constrain the distance between source domain and target domain,so as to realize the fitting and migration from simulation data to real data.Data filtering is realized through DF to further improve the accuracy of the model.From the perspective of B/S server,the platform functions are designed,the springboot framework is used for rapid development,the MVC architecture is used to organize the platform content,and the functions are subdivided into specific steps according to the idea of Test Driven Development(TDD),so as to realize rapid development.At the same time,the code is easy to test,expand and maintain.