【正文】
2構(gòu)造模型與效果的分解23構(gòu)造模型xijk=μ+ai+bj+ck+(ab)ij+(ac)ik+(bc)jk+(abc)ijk+eijkΣai=Σbj=Σck=0,Σ(ab)ij=Σ(bc)jk=Σ(ac)ik=0,Σ(abc)ijk=0效果的分解,則有如下之情形:l A因子之效果差=4|(a1a2)|=|((1)+c+b+bc)(a+ab+ac+abc)|=|(a1)(b+1)(c+1)|l B因子之效果差=|(a+1)(b1)(c+1)|l C因子之效果差|(a+1)(b+1)(c1)|l AB因子之效果差=|(a1)(b1)(c+1)|=|(abab+1)(c+1)|=|(abc+ab+c+1)(ac+bc+a+b)|l AC因子之效果差=|(a1)(b+1)(c1)|l BC因子之效果差=|(a+1)(b1)(c1)|l ABC因子之效果差=|(a1)(b1)(c1)|【例1】求AB因子之效果差A(yù)B因子之效果差=|(abc+ab+c+1)(ac+bc+a+b)|= |(2+4+3+2)(5+8+3+3)|=8公式:因子偏差平方和=(因子效果差)2/總實(shí)驗(yàn)數(shù) 在2N型之場(chǎng)合,因子之效果差可以2者之差來(lái)表示(亦可用偏差平方和來(lái)表示),但在3N型時(shí),要表示3者之差異,通常以偏差平方和來(lái)表示因此 SAB=82/8=8直交與交絡(luò)因子之效果可以分離出來(lái),稱為直交,若無(wú)法分離出來(lái)稱為交絡(luò)直交表1. 利用上述效果的分解之方法,我們可將其以+符號(hào)列成下表NOABC代號(hào)DATAABABCACBCABC111112+++++++2112c3+++3121b5+++4122bc8+++5211a3+++6212ac3+++7221ab4+++8222abc2+++每行符號(hào)DATA之和即為因子之效果差1812=6偏差平方和=(因子效果差)2/總實(shí)驗(yàn)數(shù)36/82. 上表雖可按照效果的分解之方法求得,實(shí)際可依以下作法完成l 列出A,B,C之符號(hào),即A為1時(shí)為+,A為2時(shí)為-,以此類推l AB行可依A行若為”+”B行若為”+”,則AB行得”+”, A行若為”-”B行若為”+”,則AB行得”-”,等之同號(hào)為+異號(hào)為-之原則計(jì)算3. 我們亦可將其以1,2符號(hào)列成下表,其作法同上NO.ABC代號(hào)ABABCACBCABC1111(1)11111112112c11122223121b12211224122bc12222115211a21212126212ac21221217221ab22112218222abc22121124. 因?yàn)槊啃卸贾苯唬魺o(wú)交互作用,則可增加一因子,依此方式檢討,若全無(wú)交互作用,則23型,可配置7個(gè)兩水準(zhǔn)因子,如下表所示,田口即根據(jù)此方式,建立及推廣其直交表。CResults after lst disassembly/ reassembly(H2):-35。 Variable Search)步驟1:粗估(Ballpark)1. 列出(找出)可能會(huì)影響的零組件(分好與壞的PART)或可能會(huì)影響的變數(shù)(分出高及低的水準(zhǔn))2. 確認(rèn)這些要因中,會(huì)包括有影響變異之大要因作法:將最好的組合,稱為Good(High)與最差的組合,稱為Bad(Low) 各3次實(shí)驗(yàn),共6次,實(shí)驗(yàn)採(cǎi)6次隨機(jī)試驗(yàn)解析:GOODBAD實(shí)驗(yàn)X1X2X3Y1Y2Y3(1)中位(2)全距R1R2(3)(4) 參考Lord39。s test for two independent samples. In this test the sample ranges R1,R2 replace s1, s2. This is a quick test, no more robust under nonnormality than the t test, and even more vulnerable to erroneous sample extreme values. Table A 7(ii) applies to two independent samples of equal size. The mean of the two ranges, w = (R1 + R2)/2, replaces the w of the paired test and 2—1takes the place of D. The test of significance is applied to the numbers of worms found in two samples of 5 rats, one sample treated previously by a wormkiller. See table . We have 2—1 = and w = (219 + 147)/2 = 183. From this, tR = (2—1)/ = , which is beyond the 1% point, , shown in table A 7(ii) for n = 5. TABLE NUMBER OF WORMS PER RATTreatedUntreated12337814327519241240265259286Means, Ranges, R219147To find 95% confidence limits for the reduction in number of worms per rat due to the treatment, we use the formula (2—1)—tR 2-1 (2—1) + tR ()(183 2- + ()(183)60 2-1284l The confidence interval is wide, owing both to the small sample sizes and the high variability from rat to rat. Student39。(6) 判斷所選擇因子中有影響的大要因存在,可進(jìn)行步驟2(7)如果 判斷所選擇的因子中無(wú)影響大要因存在,回到步驟1例:HOUREMETERAn hour meter , built by an electronics pany , had a 2025 percent defect rate because several of the units could not meet the customer39。為差異的要因分析提供強(qiáng)列的線索。同時(shí),它也是失敗故障分析的有力工具。討論:請(qǐng)舉出在LCD之製程中,空間之變異有哪些。l 全品之內(nèi)相同各件之間的差異,譬如一片晶圓上數(shù)百粒晶體之間品質(zhì)出入很大。經(jīng)過恰當(dāng)對(duì)策後,空間面要因所產(chǎn)生的品質(zhì)變異可望消除大半。這種隨機(jī)性要因是會(huì)再度出現(xiàn)的,所以它們有反覆性。變異要因檢討解析例某家瓷磚製造商磁磚褙紙之褙紙黏度品質(zhì)不易控制,搜集數(shù)據(jù)如下表(1)橫條之內(nèi)(每條5片瓷磚)(2)橫條之間(3)時(shí)間,另外,將以上數(shù)據(jù)繪製成 multivari charts(包括每條中最高黏度每時(shí)段平均黏度、每條平均黏度),如圖 ( 問題) 1那一方面的變因有最大的變異? 2你可以找到什麼端倪? 包括非隨機(jī)的趨勢(shì)。如同第2步驟,.觀察且記錄此對(duì)差異,4. 重複此搜尋步驟,第三,第四,第五,和第六對(duì),直到觀察的差異顯現(xiàn)出有重覆的模式。SHAININ使用Lord test 之步驟步驟BetterCurrent(1)實(shí)驗(yàn)Bamp。CResults after 2nd disassembly/ reassembly(H3):-37。步驟終止。例:HOUREMETER(Component serach)An hour meter , built by an electronics pany , had a 2025 percent defect rate because several of the units could not meet the customer39。2. 之分配,n[()/s]2為趨近分配,此處值如下表所示,從表中可知,d2=(1+1/4n),當(dāng)組數(shù)夠大時(shí),=d23. 當(dāng)組數(shù)k=1,即為一般所稱之R,在不致誤解下,有時(shí)亦表示成,同理有時(shí)以d2表示。外側(cè)也和E兩板的defects 最多各達(dá)56及54,而中間B、C和D三板各只有18及10個(gè)defects ?!挂虼?,我們宜研究比對(duì)「良板」與「劣板」。(對(duì)策)(Corrective Action)過量的心偏表示整組機(jī)板的夾具不太適當(dāng),而且預(yù)熱區(qū)溫度過高。F)及unsoldered connections孔位鑽穿小孔。F480450G銲錫時(shí)間secH助銲劑沫高度STAGE 1 BALLPARK在波銲機(jī)臺(tái),以全高水準(zhǔn)十組和全低水準(zhǔn)十組進(jìn)行實(shí)驗(yàn),如此隨機(jī)反覆三回。因此,我們宜將它們從參數(shù)清單中剔除。(3)於製程我們宜將助銲劑密度,輸送帶速度和輸送帶坡度等,都設(shè)定在高水準(zhǔn)。因此,上述的改善成果得以確認(rèn)。(2)如果只願(yuàn)意接受2OPPm的銲錫defect rate。 and down to less than l0 ppm! A systematic road map was followed:1. Multivari study2. Paired parisons3. B 4. Variables search5. B 6. Full factorials7. Scatter plots8. Positrol9. Process certification This road map is not necessarily a rigid one to fo1low in all circumstances. In fact, no two chronic problems are exactly alike. The number and sequencing of the DOE tools do vary from problem to problem. Yet the elegance of these Shainin techniques and their bined power can make any chronic problem disappear.SHAININ解決問題基本想法1. 特性值與工程規(guī)格計(jì)量值CPK (計(jì)數(shù)值PPM) 評(píng)價(jià)過程能力2. 現(xiàn)況分析:尋找差異處l 查檢表設(shè)計(jì)?時(shí)間層?空間層?重覆層l 多元表分析: multivarti chart找出重要層l Components Search由零組件不良所構(gòu)成原因,找出重要影響之零組件(項(xiàng)目)l Paired Comparisons從良品與不良品對(duì)照找出故障模式3. 原因分析:可能要因與真因驗(yàn)證l Variables Search由過程中所採(cǎi)用之條件(參數(shù)變數(shù))因不好之組合而構(gòu)成,找出重要影響之變數(shù)(項(xiàng)目)l Full Factorials檢討重要項(xiàng)目之主效果與組合效果l B vs .C –驗(yàn)證最佳與現(xiàn)有之差異l Realistic Tolerance Parallelogram (scatter plots)重要變數(shù)公差訂定4. 對(duì)策計(jì)畫與