Abstract
表面粗度是工件品質的一項重要指標。然而基於經濟上的考量，一般工廠對於工件表面粗度之檢驗多採抽樣方式。為了提高整體之加工品質，有必要找尋一些方法來量測或監測每一工件之表面粗度。線上量測表面粗度的方法可分為直接量測及間接量測，直接量測通常使用光學方法直接量取工件表面波紋，其缺點為光學訊號容易受切屑及切削液的干擾。而間接量測則藉由一些較不受外界干擾之感測器來估計表面粗度值。基於上述之考量，本研究擬採用類神經網路來處理切削力訊號進而達到間接量測車削表面粗度之目的。除此之外，吾人將比較傳統歸分析方法及類神經路在有頂心及無頂心車削時之表面粗度估測能力 。研究結果得知，在有頂心車削時振動量很小，使用迴歸模式即可約略估計出工件之表面粗度。然而，在無頂心車削時振動量變大，此時類神經網路比迴歸模式較能建立削力訊號與表面粗度之關係。It is well know that surface finish of machined parts plays animportant role in their functioning; it's a significant indexof quality management. However, due to economic consideration,the current mehtods for determining surface roughness ofmachined parts usually depend on statistical techniques. Inohter words, only a few samples from a group of machined partsare measured. In order to promote the global quality ofmachined parts, it is of interest to search for methods tomeasure or monitor the surface roughness of each machined part.In-process measurement of surface roughness can be classifiedas direct and indirect techniques. Direct techniques employoptical methods to measure surface profile directly. The directin-proces techniques are rather of limited use and unacceptablebecause the optical signal may be distorted by chip generationor cutting fluid during machining process. While indirecttechniques estimate surface roughness by means of ohterinformation from some sensors which have better resistance tothe surrounding noises. Due to the consideration mentionedabove, an indirect surface roughness monitoring system will bedeveloped in this thesis. Via neural network, the signalsdetected from dynamometer are handled to estimated surfaceroughness of workpiece turned with or without support of dead-center. Besides, the estimating results via neural network andregression method are compared. The results indicate thatregression model can deal with the case when workpiece issupported with dead-center. While, vibration phenomena issevere when workpiece is not supported with dead-center, andneural network have the better ability to simulate therelationship between cutting force signals and surfaceroughness than regression model does.