Research On Industrial Robot Operation Condition Monitoring Based On Transfer Learning

As an indispensable part of automated production and intelligent manufacturing,industrial robots are widely used in various fields due to their high efficiency,security and high flexibility.And undoubtedly,prognostics health management,especially in the field of condition monitoring,are significantly important.Ho wever,there are limited research on industrial robots in this field.The article mainly focus on the following parts: Firstly,analysing industrial robots regarding its basic structure,and also conducting kinematics analysis of industrial robots,the structure and possible failure mode of RV reducer.Then,acquiring the faulty data through the experimental platform.Furthermore,establishing a fault identification model based on small sample,and building a transfer fault identification model under variou s working loads of robots.Moreover,integrating the transfer fault identification model into the system,which is Labview-based industrial robot status monitoring system.The methodology applied is offline training and online deploying.The article contai ns following parts:(1)By analyzing the structure and kinematics of industrial robots,and common failure modes of RV reducers,aiming at small samples and highly detection accuracy,a method based on transfer support vector machine is introduced to build the fault detection model under small sample condition.Besides,the experimental data is acquired by low-cost attitude sensors.This method contains two parts: data dimension reduction through feature extraction of faulty data;and then input the TSVM mo del for fault recognition in the case of small samples.When the feature dimension of the model is 15 after dimensionality reduction,the average recognition accuracy of the six sample proportion is the highest,reaching 98.15%.By comparing with other alg orithms,it is shown that the recognition accuracy and stability of TSVM are better than the comparison algorithm.(2)Due to the limited nonlinear fitting ability of the shallow structure,like TSVM model,a Wasserstein distance-guided domain adaptation algorithm using deep learning structure is proposed.Then,in catering to different loads of industrial robots,the cross-domain fault recognition model is built by using vibration signals of industrial robots.Firstly,the preprocessing of data that convert the time-series signals into time-frequency signals through STFT.After the time-frequency transformation,inputting the data into the WDGDA model for training.The average recognition accuracy of the model under six transfer tasks is 95.13%.Compared with the fault recognition models established by other transfer algorithms,the recognition accuracy of the WDGDA model is much better.(3)Aiming to embed the model into the Labview-based industrial robot state monitoring system,the two-dimensional input model should be changed into the one-dimensional input model,as well as maintaining good recognition accuracy.Then,through the built-in python interface of labview,integrating the converted model into this system.As a result,the offline training and online deploying mode of intelligent system have been completed.