Acoustic Signal Based Grinding Belt Wear Condition Monitoring

With excellent mechanical properties and corrosion resistance,nickelbased superalloy can maintain high strength and fatigue resistance even in high temperature and extremely loaded working environment.Robotic abrasive belt grinding has many advantages,such as strong material removal ability,stable finishing quality and good flexibility,therefore,it is frequently applied for processing nickel-based superalloy.In grinding of superalloy,wear can be a serious problem which will lead to a serious decline in processing efficiency and forming quality,therefore,effective belt condition monitoring is a key element of intelligent robotic belt grinding system.This thesis paper presents an acoustic signal based tool condition monitoring(TCM)system for belt grinding of nickel-based superalloy.In the test process,a constant-force grinding mode is adopted and the acoustic signal from every grinding cycle is recorded.Acoustic signal is analyzed in time domain and frequency domain to extract the tool condition related features.Using a six-layers wavelet disposition method,the acoustic signal is divided into seven parts according to different frequency distributions and the signal components in D6(312.5~625Hz),D5(625~1250Hz),D2(5000~10000Hz),and D1(10000~20000Hz)are selected as the effective components which are used for further belt condition recognition.The wear process can be distinguished into initial wear period,the steady state period and accelerated wear period.By mapping the acoustic features of grinding sound and conditions of grinding belts,a tool condition classifier based on back propagation neural network(BPNN)and a tool condition classifier based on random forest(RF)algorithm are trained and applied in prediction of grinding belt conditions.The prediction accuracy of NN model is 90% while the accuracy of RF model is 94%.In order to effectively improve the utilization rate of grinding belt,if the belt wear stage is in the later stage of accelerated wear,this thesis introduces a multiple linear regression(MLR)model to predict the grinding ability of belt during this phase.The trained MLR model was used to predict the test samples.The average error of the predicted results was 5.3%.This work proposes an acoustic signal based detection method that combines a RF classifier and a MLR model to detect different wear periods and evaluate the remaining grinding ability for robotic belt grinding of nickel-based superalloy.