附錄A：資料篩選

hlperng

Annex A (informative) Data screening 附錄A：（參考）資料篩選

There are a number of general requirements that apply to all the statistical tools referred to in this standard.  These requirements are mainly concerned with the design and control of the testing.
有一些適用於在本文件中所提及的所有統計工具的一般要求，這些要求主要著重在試驗的設計和控制。

There are several areas that the test engineer should explore before commencing any testing, for example:
- purpose of the test;
- test equipment available;
- design of the test;
- setting up of the test;
- control of the test;
- collection of the data from the test.
任何試驗開始之前，試驗工程師有幾個方面應該探討，例如：
– 試驗目的；
– 可用的試驗設備；
– 試驗設計；
– 試驗規劃；
– 試驗控制；
– 試驗資料收集。

A carefully planned test increases the chance that the data can be analyzed using referenced statistical standards.
經過精心規劃的試驗，會增加引用的統計標準可以分析資料的機會。

However, before applying any of the statistical tools, the data should be screened.
然而，在應用統計工具前，資料應該進行篩選。

Particular attention should be paid to any unusual patterns in the data.  Unusual patterns may suggest that there is a mixture of different populations.  For example, simple exploratory data analysis could be conducted. Such analysis could take the form of simple plots of time against failure number, bar charts, pie charts, tabulations and statistical techniques such as Analysis of Variance (ANOVA).  ANOVA may give a more formal indication of the existence of significant population differences.  Unusual patterns in the data may suggest test equipment problems or data collection problems or several different failure modes.
應該特別注意的是資料中的任何不尋常態樣，不尋常的資料態樣可能代表著有不同的群體混合在一起。例如，可以進行簡單的探索式資料分析，這種分析可以採用簡單的繪圖形式，例如時間對失效次數的關係圖、直條圖、圓形圖、表格，以及統計技術，例如變異數分析(ANOVA)。變異數分析分析可能會得到更加正式的指標，表明群體存在顯著的差異。資料中不尋常的態樣，可能是測試設備問題，或者資料收集的問題，或是有幾個不同的失效模式。

In the case of failed items, the engineer should investigate, if possible, the the failure modes or failure mechanisms before applying the statistical techniques. Mixing data from different failure modes may give erroneous results.  At best, only average data for the actual mixture of failure modes may be estimated, note that the mixture may change with test condition (for example Arrhenius equation).   
對於失效的物品，如果可以的話，工程人員在應用統計技術之前應該先分析其失效模式或失效機理。來自不同的失效模式的混合資料，可能會得到錯誤的結果。實際混合多種失效模式的混合資料，充其量只能估算其平均值，要注意的是混合資料可能會隨著試驗條件而變（例如阿倫尼爾斯方程式）。


A.1 失效模式分類 (Classification of Failure Modes)

Figure A.1 shows a bar chart for the number of failures of a particular type of item grouped by failure mode.  By classifying the data into failure modes,the engineer has a better knowledge of where particular problem areas lie.  Statistical modeling of the data therefore be carried out on the data for each category. This is in fact an example of a Pareto diagram, i.e., where a large proportion of the faults are due to a small number of causes.
圖A.1的長條圖為一特定類型項目分組的失效模式的失效次數數量，透過失效模式資料的分類，工程人員對於特定問題發生的部位可以有更好的知識。因此，應該分別針對每一類別的資料建立其統計模型。這其實是一個柏拉圖的例子，亦即故障有很大的比例是因為少數的原因所造成的。


圖 A.1 ：柏拉圖


A.2 故障分類 (Classification of Faults)

Figure A.2 shows a pie chart of fault categories for a repaired item.  Each fault has been categorized so that future occurrences of these faults can be avoided through changes in design, manufacturing process, component suppliers, etc.  This is particularly useful in development programmes as the number of design, component and external faults should decrease, over a period of time, as the design becomes more stable and the manufacturing processes have been set up and verified.  This type of data screening is often used when monitoring reliability growth programmes.  
圖A.2顯示已修復物品的故障分類圓形圖，每一個故障都已經分類，透過變更設計、製造過程、零件供應商等，可以避免未來再發生這些故障。這點對研發方案特別有用的，因為經過一段時間後，設計會變得更加穩定，而且製造過程也會完成建置和查證，設計、元件和外部故障的數目應該會減少。監視可靠度成長計畫時，通常使用這種型式的資料篩選。


圖A.2：故障分類範例

A.3 不同失效模式混合 (Mixtures of Different Failure Modes)
Figure A.3 shows a plot of operating times to failure for non-repaired items of a particular type.  Notice that there appears to be three separate sets of data.  For example, for an electronics product, the first set could be due to early life failures such as design problems, incorrect component rating, process problems, dry joints or mechanical damage, etc.  The middle set could be due to random failures or perhaps leaking packages.  The third set could be associated with wear-out mechanisms such as solder joint fatigue.  By plotting this data, the engineer can observe these different stages and therefore further investigate the cause of these failures.  This analysis should be done before attempting to model the data for compliance, comparison or estimation.  
圖A.3為某種特定類型不修復物品的失效發生操作時間，請注意此圖似乎有三個獨立的資料集合，以電子產品為例，第一組可能是因為早期失效所得到，如設計問題、不正確的零件額定等級、製程問題、乾燥接頭或機械損壞等；中間這組資料可能是因為隨機失效或封裝洩漏所造成的；第三組資料可能與磨耗機理有關，如焊點疲勞。透過這些資料的繪製，工程人員可以觀察到這些不同的階段，進而進一步調查這些失效的原因。這種分析應該在符合性、推定或比較資料分析之前完成。


圖A.3：不同失效模式混合範例

A.4 不同群體混合 (Mixture of different population)
If it is suspected that some items within an operational system have a manufacturing defect, what could be done to investigate the extent of the problem in terms of expected life?  A number of items of the same type could be tested to investigate the expected life.  The test would have to be able to induce the same failure mechanism as that observed in the field of operation.  Suppose there are 60 items and three different batches to be tested.  The resultant data would firstly have to be screened for a mixture of differing populations.  A check as to whether there is any significant difference between batches would therefore have to be carried out using exploratory techniques and other statistical technique, for example, ANOVA.
如果懷疑在一個操作系統中有製造瑕疵時，要研究這個問題與期望壽命的關係，有甚麼事可以做？有一些相同類型的物品可以利用試驗來調查其期望壽命，該試驗會有能力誘導出與實際操作時相同的失效機理。假設有60件分成三個不同的批次進行試驗，所產生的資料將首先進行篩選是否為不同群體的混合，檢查確定是否有任何批次之間的顯著差異，因此必須使用探索技術和統計技術等來進行，例如，變異數分析(ANOVA)。

Figure A.4 shows the times to failure for non-repaired items divided into batches.  
圖A.4 顯示不修復物品分成批次的失效發生時間。

It appears that batch C items have behaved differently from the others and that perhaps there are two separate populations, formal statistical techniques and failure analysis may confirm this hypothesis.  Further analysis of this data using tools described in this standard would only be meaningful if data from batches A and B were analyzed separately from batch C data.
批次C表現得不同於其他批次，顯示也許有兩個獨立的群體，正式的統計技術和失效分析或許可以確認此一假設。如果批次A和B的資料分析與批次C的資料分開分析，那麼使用本標準中描述的工具來進一步分析這些資料才有意義。


圖A.4：不同群體混合範例

A.5 故障診斷 (Fault Diagnosis)

Analysis of data of a particular non-repaired item indicated a larger number of fault incidents than expected, resulting in a higher estimated failure rate.  Closer inspection of the data showed that the times to failure were as shown in Figure A.5.  
某種不修復物品的資料分析顯示其故障事件比預期還大，因此得到較高的失效率推定值，經過仔細的觀察資料結果顯示其失效發生時間如圖A.5所示。

Further investigation of the population showed that there were five items of this type on each printed circuit board.  After questioning the diagnostic engineer it was found the only one failure had occurred but the engineer thought that it was best to replace all five items.  Thus if the estimated failure rate has been accepted without further investigation it would have implied a shorter expected life and consequently too many spares may have been ordered.  
對群體做進一步研究結果顯示，在每片印刷電路板上有5個這一類型的物品，詢問過診斷工程師後才發現只發生一次失效，但是工程師卻認為最好將5個物品全部換掉。因此，如果接受失效率推定值結果而沒有進一步研究，這意味著會有比較短的期望壽命，其後果就是訂購太多的備份件。


圖A.5：診斷問題範例

A.6 試驗裝備故障 (Test Equipment Faults)

Figure A.6 shows operating time to failure for a single repaired item.  The data appears to have a large number of early failures followed by an increasing trend in the times between failures.  
圖A.6為單一已修復物品的操作失效發生時間，資料顯示有大量的早期失效次數，接著是失效間隔時間有增加的趨勢。

The reasons for this clustering effect should be investigated before model fiting.  When complex simulation equipment is used for testing purposes the initial set-up of the equipment is crucial.  The early failures shown in Figure A.6 could be the result of errorneous parameters set in the simulation.  
在模型嵌合前應該先研究資料群聚效應的道理，當試驗使用複雜的模擬設備時，其初始設定非常重要，在圖A.6的早期失效有可能就是在模擬時因為參數設定錯誤所造成的結果。

When selecting a distribution or process for modeling data, the user should firstly use engineering judgement and knowledge of the items under test.  Simple exploratory data analysis techniques and more formal statistical techniques should, together with failure analysis confirm or modify engineering judgement.  
當選擇一個分布或過程建立資料模型時，分析者首先應該利用工程判斷及對受試物品的知識，接著應該以簡單的探索式資料分析(exploratory data analylsis)技術和更正式的統計技術，結合失效分析，來證實或修改工程判斷。


圖A.6：資料分群範例