Research On Prediction Method Of Cutting Tool Life For Dicing Saw

As one of the most important pieces of equipment in the semiconductor manufacturing process,dicing saw is mainly used to cut wafers into chips of a certain size.With the continuous optimization of chip manufacturing processes and the shrinking of chip sizes,the requirements for cutting quality on the dicing saw are becoming increasingly high.As a critical component in the cutting process,the cutting tool directly impacts the cutting quality.To improve the cutting quality and manufacturing efficiency of the dicing saw,it is necessary to predict the remaining useful life(RUL)of the cutting tool.Predicting the RUL helps production managers replace the tool in a timely manner,which can prevent production downtime and quality issues caused by tool wear.Both production efficiency and chip quality will also be ensured.The main jobs include the following aspects:(1)For DISCO’s ZH-05 series tools,the cutting datas recorded by ADT Israel’s 8230 fully automatic dual-axis dicing saw were selected for tool life prediction study.Six features were selected as inputs to the tool prediction model,including tool exposure,tool thickness,cutting speed,feed speed,spindle speed,depth of cutting.And these six features are used as the input to the tool prediction model.(2)In the RUL of the dicing saw tool prediction experiment,in order to improve the problems of traditional BP neural network prediction models,such as low prediction accuracy and the tendency to fall into local optima,a new method is proposed in this paper.The proposed method leverages the global optimization capabilities and adaptive features of both particle swarm algorithms(PSO)and genetic algorithms(GA)to enhance the performance of the prediction model.(3)For the problem of the dicing saw tool life prdediction,three tool life prediction models were established: namely BP model,PSO-BP model,GA-BP model.Experiments are conducted using cutting data collected from an 8230 fully automatic two-axis dicing saw to predict scribing machine tool life.Results indicate that the proposed PSO-BP model and GABP models outperform the traditional BP model.The PSO-BP prediction model proposed had higher accuracy than the BP prediction model by 2.7% and 1.9% based on 15 and 50 test samples,respectively.Similarly,the GA-BP prediction model had higher accuracy than the BP prediction model by 4% and 3.5% based on 15 and 50 test samples,respectively.These findings demonstrate that PSO and GA,by leveraging their global optimization capabilities and selfadaptive features,are able to overcome the limitations of the BP neural network model.Overall,the proposed prediction methods offer significant potential for practical application in dicing saw tool life prediction.