Research On Health Condition Assessment Method Of RV Reducer Based On Ordinal Classification And Cost-sensitive Learning

The RV reducer is a high-torque transmission device and is widely used in industrial robots.It has become the most core part of the joints of industrial robots,and the reliability of the RV reducer is crucial.If the operating state cannot be accurately judged in actual production,it will bring great disadvantages to the maintenance and repair of the RV reducer,and it will also have a bad influence on the production environment.Therefore,it is very important to study the health assessment of RV reducer.At present,the research on performance status recognition and reliability monitoring of RV reducer is relatively lacking at home and abroad.The main research is to analyze the failure mechanism of RV reducer through fault tree and discriminate the fault type based on acoustic emission.Data-driven methods based on vibration signals are modeled using machine learning,making it easier to implement online status assessment.This thesis analyzes the health status of RV reducer on the performance degradation test of RV reducer through the vibration data of RV reducer from the actual project.And compared to the performance test of the RV reducer based on the platform,the data-driven method based on the vibration signal is more convenient and effective.The main work of this thesis is divided into two parts:First of all,through the RV reducer vibration test platform,the vibration signal is collected.Data sets are formed at different speeds,and then the collected data is subjected to wavelet threshold denoising.In order to obtain and utilize more comprehensive signal information,multi-domain feature parameters are extracted,including time domain,frequency domain,time-frequency domain,and entropy.The obtained feature set contains irrelevant redundant features that need to be selected using the ReliefF algorithm.Later,several commonly used dimensionality reduction algorithms,including Principal Component Analysis(PCA),Linear Discriminate Analysis(LDA),Isometric Mapping(ISOMAP),and Local Linear Embedding(LLE),three-dimensional visualization of them,and comparison of their effects on the standard support vector machine,as the basis for the selection of dimensionality reduction methods.Second,establish a classification model that considers ordinal classification and cost-sensitivity.The RV reducer vibration data used in this thesis is divided into four states with performance trends,namely "healthy","good","general" and "poor".Conventional classification algorithms do not consider the sorting of response variables,but this order information can contribute to the classification.At the same time,it needs to consider that different categories of misclassification costs are different.Therefore,in light of the actual background knowledge of the problem,ordinal classification and cost-sensitive learning were deeply studied,and experiments were compared using different algorithms.The results show that the ordered classification can use the inherent order information,and the cost-sensitive learning can improve the recall rate of the "poor" class while maintaining the accuracy of the model.Finally,an ordinal classification method based on cost-sensitive SVM is proposed.Comparison experiments of various algorithms show that the performance is optimal.