外文翻譯----石油質(zhì)量在線監(jiān)控系統(tǒng)的設(shè)計(jì)與測(cè)試

1、<p>　　畢業(yè)設(shè)計(jì)(論文)外文資料翻譯</p><p>　　題 目: 石油質(zhì)量在線監(jiān)控系統(tǒng)的設(shè)計(jì)與測(cè)試 </p><p>　　院系名稱： 專業(yè)班級(jí)： </p><p>　　學(xué)生姓名： 學(xué) 號(hào)： </p>&

2、lt;p>　　指導(dǎo)教師： 教師職稱： </p><p>　　起止日期： 09.2.23—4.30地 點(diǎn): 中心實(shí)驗(yàn)樓 </p><p>　　附 件： 1.外文資料翻譯譯文；2.外文原文。 </p><p>　　附件1：外文資料翻譯譯文</p><p&

3、gt;　　石油質(zhì)量在線監(jiān)控系統(tǒng)的設(shè)計(jì)與測(cè)試</p><p><b>　　摘要</b></p><p>　　本文總結(jié)了作者對(duì)石油質(zhì)量傳感器的應(yīng)用與測(cè)試。當(dāng)機(jī)器處于高速運(yùn)轉(zhuǎn)狀態(tài)時(shí)，燃油保持正常的流動(dòng)性是至關(guān)重要的。因此在本文中，作者描述了石油感應(yīng)原理，而且在實(shí)踐中驗(yàn)證了石油質(zhì)量傳感器自動(dòng)檢測(cè)石油質(zhì)量的能力，并詳細(xì)而精確地對(duì)石油污染物進(jìn)行了分類。該傳感器能進(jìn)行流體檢測(cè)，能根

4、據(jù)化學(xué)規(guī)律對(duì)污染物進(jìn)行分類，并能具體地估計(jì)污染物變化趨勢(shì)，實(shí)施更高層次的通信協(xié)議，從而最終檢測(cè)出石油的質(zhì)量或污染程度。正如作者測(cè)試時(shí)所做的一樣，石油質(zhì)量監(jiān)控系統(tǒng)的傳感器設(shè)計(jì)成一個(gè)能夠直接插入油管的塞子。作者還提出了對(duì)實(shí)驗(yàn)方法、壓力反饋潤(rùn)滑系統(tǒng)和預(yù)防污染的總體看法以及測(cè)試結(jié)果。同時(shí)作者對(duì)石油質(zhì)量檢測(cè)系統(tǒng)進(jìn)行了為期6個(gè)月測(cè)試，并在石油的變質(zhì)界限下驗(yàn)證了多個(gè)污染物的分類。</p><p>　　關(guān)鍵詞：潤(rùn)滑油污染，潤(rùn)滑油

5、分析技術(shù)，狀態(tài)監(jiān)測(cè)和內(nèi)燃機(jī)油，設(shè)備監(jiān)控</p><p><b>　　導(dǎo)言</b></p><p>　　狀態(tài)檢修（CBM）是一個(gè)新興的維修理念，其采用積極的監(jiān)測(cè)，以確定系統(tǒng)的組成部分以及運(yùn)行是否正常，使根據(jù)診斷和預(yù)測(cè)剩余使用壽命去維修成為可能。CBM減少了生命周期維護(hù)成本，改善了系統(tǒng)的安全性，并增加了運(yùn)行的可預(yù)見性。液壓機(jī)油、齒輪油、潤(rùn)滑劑和其他正在使用的液態(tài)物在使用中

6、變質(zhì)，是機(jī)器運(yùn)行失常的共同原因之一，因此潤(rùn)滑劑的質(zhì)量監(jiān)測(cè)是對(duì)CBM系統(tǒng)比較理想的補(bǔ)充。按照傳統(tǒng)方法，潤(rùn)滑油的質(zhì)量監(jiān)測(cè)可以通過定期取樣、實(shí)驗(yàn)室檢測(cè)或使用現(xiàn)場(chǎng)測(cè)試裝置分析完成。然而，這些分析方法比較費(fèi)時(shí)、成本過高，誤差較大，而且有較長(zhǎng)的滯后時(shí)間。其誤差可能的來源主要包括取樣位置、容器交叉污染以及測(cè)試方法的準(zhǔn)確性。一個(gè)自動(dòng)化、現(xiàn)場(chǎng)油質(zhì)量監(jiān)測(cè)系統(tǒng)能提供連續(xù)、實(shí)時(shí)的潤(rùn)滑劑變質(zhì)信息，讓機(jī)器避免不必要的系統(tǒng)磨損，使維護(hù)時(shí)間間隔最佳化，并能盡早處理設(shè)備

7、問題。這些技術(shù)通過減少設(shè)備的停機(jī)時(shí)間，以及降低運(yùn)營(yíng)成本使利益最大化。</p><p>　　一些方法能實(shí)現(xiàn)潤(rùn)滑劑的實(shí)時(shí)監(jiān)測(cè)，但大部分方法主要針對(duì)三個(gè)主要的類型：質(zhì)量、殘?jiān)驮亍Ｆ渲杏幸环N方法是使用元素的感應(yīng)原理檢測(cè)污染物，然而，這一技術(shù)在檢測(cè)粒徑在500μm以下的粒子時(shí)受到限制，這是由于粒子在不足5微米至20微米會(huì)造成的60％的發(fā)動(dòng)機(jī)磨損。這種方法對(duì)燃料和水這樣的污染也是不敏感的，而這些污染普遍存在于柴油機(jī)系統(tǒng)中

8、并嚴(yán)重地降低潤(rùn)滑劑的工作性能。由于存在大量處于失效模式的石油，使利用感應(yīng)技術(shù)測(cè)量石油導(dǎo)電率的技術(shù)受到限制，這是因?yàn)橥茰y(cè)出油質(zhì)量的參數(shù)只有一個(gè)。其他不同于油的老化機(jī)理的檢測(cè)技術(shù)已經(jīng)發(fā)展起來，但無法檢測(cè)其他主要失效模式。還有一些檢測(cè)系統(tǒng)使用多個(gè)復(fù)雜的傳感器而且需要針對(duì)現(xiàn)有的油循環(huán)和診斷控制系統(tǒng)做一些有意義的修正，因此直接阻礙了這些方法在商業(yè)上的應(yīng)用。</p><p>　　本文中所描述的智能型石油傳感器（SOS）能提供

9、在線、即時(shí)、低成本的石油質(zhì)量即時(shí)分析。本文的目的是論證SOS技術(shù)能克服上述以光譜學(xué)為基礎(chǔ)的寬頻阻抗分析技術(shù)中所具有的缺點(diǎn)。</p><p><b>　　測(cè)量基礎(chǔ) </b></p><p>　　智能型石油質(zhì)量檢測(cè)系統(tǒng)采用了正在申請(qǐng)專利的、低功率的寬頻阻抗測(cè)量技術(shù)以及使用多傳感器融合技術(shù)和模型分析軟件包，旨在預(yù)測(cè)流體質(zhì)量的變化。電氣化學(xué)的阻抗光譜學(xué)（EIS）方法，是將一個(gè)

10、復(fù)雜的交流信號(hào)加入寬頻光譜系統(tǒng)從而測(cè)量系統(tǒng)的響應(yīng)，以確定石油的質(zhì)量。系統(tǒng)的阻抗是由所加的激勵(lì)信號(hào)和響應(yīng)信號(hào)之間的不同點(diǎn)決定的。通過掃描寬范圍的頻譜，傳感器獲得一個(gè)能比較好的反映石油實(shí)際阻抗的測(cè)量數(shù)據(jù)?？破澛宸蛩够热擞昧艘粋€(gè)相似的方法模擬細(xì)胞的電氣化學(xué)阻抗。通過發(fā)射一個(gè)寬頻信號(hào)，而非一個(gè)單信號(hào)，石油檢測(cè)系統(tǒng)在少于30秒內(nèi)能完成對(duì)石油的檢測(cè)。相比之下，傳統(tǒng)的EIS測(cè)量法需要50分鐘以上才能完成石油質(zhì)量檢測(cè)，因此這種在線檢測(cè)方法無法處理可能

11、在這段時(shí)間內(nèi)發(fā)生工藝參數(shù)變化的過程，另外，有害污染物的出現(xiàn)和石油基本原料以及添加劑的變質(zhì)通常也會(huì)引起油質(zhì)的變化。這些變化會(huì)影響油的介電性能、導(dǎo)電率、阻抗、電容和其他主要性能。SOS檢測(cè)系統(tǒng)使用對(duì)電化學(xué)的模擬來表現(xiàn)潤(rùn)滑劑的阻抗響應(yīng)，也使用基于模型的參數(shù)估計(jì)方法來鑒別樣品中所含的污染物。圖1顯示了在污染物不斷變化的條件下的基本電化學(xué)阻抗測(cè)量電路。</p><p>　　圖1：智能型石油傳感器的原理圖</p>

12、<p>　　智能型石油質(zhì)量傳感器的設(shè)計(jì) </p><p>　　智能型石油傳感器（圖2）是一個(gè)獨(dú)立的單元，包括敏感元件、信號(hào)調(diào)理、數(shù)據(jù)處理和通信模塊。電化學(xué)阻抗頻譜測(cè)量元件由二個(gè)均衡分離的同中心圓筒組成，傳感器采用頭形幾何形狀，以最大限度地提高樣品表面面積，同時(shí)盡量減少阻抗，以利于流體流動(dòng)。SOS在信號(hào)條件和數(shù)據(jù)處理方面有一些特殊的功能，比如動(dòng)態(tài)結(jié)構(gòu)增益和濾波器的選擇，檢測(cè)信號(hào)的發(fā)生，高速數(shù)據(jù)的獲取，

13、數(shù)據(jù)分析，以及外部通信等。</p><p>　　圖2：智能型石油傳感器的實(shí)物圖</p><p>　　目前傳感器的設(shè)計(jì)支持RS-232通信和控制器局域網(wǎng)（CAN）通信。通過這些接口，傳感器能對(duì)流體測(cè)量、傳感器狀態(tài)、配置信息以及軟硬件更新信息進(jìn)行通信。包含CAN通信協(xié)議的支持能簡(jiǎn)化并整合現(xiàn)有的檢測(cè)控制系統(tǒng)。</p><p>　　在進(jìn)行EIS測(cè)量之后，傳感器進(jìn)行特征抽出和

14、分類，并運(yùn)行一系列相關(guān)的運(yùn)算法則，從而在收集的阻抗數(shù)據(jù)中篩選出流體的質(zhì)量信息。這一個(gè)步驟包括特征抽出和分類兩個(gè)過程。特征抽出是被用來估計(jì)電化學(xué)阻抗模型參數(shù)的方法，比如大阻抗和表面特性，由此產(chǎn)生的功能是提供一個(gè)阻抗測(cè)量和實(shí)際的流體特性變化之間的聯(lián)系。該傳感器的嵌入式處理器使用線性最小二乘擬合算法進(jìn)行特征抽出，這種方法便于在嵌入式系統(tǒng)中實(shí)現(xiàn)，而不會(huì)犧牲系統(tǒng)的性能和準(zhǔn)確性。 </p><p>　　作者利用他們的經(jīng)驗(yàn)在多

15、傳感器數(shù)據(jù)融合、分類、數(shù)據(jù)挖掘等方面建立一個(gè)最符合應(yīng)用要求的實(shí)時(shí)、原位、獨(dú)立的傳感器分類器。作者嘗試用幾個(gè)基于線性判別分析、貝葉斯概率模型、強(qiáng)大的故障檢測(cè)與隔離方法以及神經(jīng)網(wǎng)絡(luò)的分類器體系結(jié)構(gòu)進(jìn)行試驗(yàn)。為了評(píng)價(jià)分類方法，作者考慮了數(shù)據(jù)集、所需的訓(xùn)練樣本大小、分類準(zhǔn)確性和執(zhí)行時(shí)間之間兼容性。貝葉斯概率分類器最適合這一特定應(yīng)用的要求，這是因?yàn)樗苓_(dá)到高精確度并能達(dá)到合理的效果。貝葉斯分類器適用于功能獨(dú)立的針對(duì)概率分布的假設(shè)。這個(gè)模型的特點(diǎn)是

16、采用了多元的、正常的概率密度函數(shù)并計(jì)算概率的每個(gè)特征向量。因此，根據(jù)貝葉斯定理，一個(gè)特征向量X ＝x和C ＝i可能屬于同一類，即：</p><p><b>　?。?）</b></p><p>　?。ㄗⅲ涸谶@些變量中，大寫字母表示變量，而小寫字母表示當(dāng)前值或觀察值）</p><p>　　貝葉斯分類器使用了由特征向量賦予的后驗(yàn)概率作為判別式，它能使

17、假設(shè)簡(jiǎn)化，其功能是獨(dú)立的。因此，分類器使用了判別函數(shù)：</p><p><b>　?。?）</b></p><p>　　分類器的體系結(jié)構(gòu)由燃料，水和煤煙確定，而石油包括三個(gè)層次，以此來利用從傳感器得到的實(shí)時(shí)數(shù)據(jù)和歷史分類。第一層和第二層貝葉斯分類器通過一個(gè)交叉耦合的體系結(jié)構(gòu)連接。這些交叉耦合的分類器由于搜索空間而發(fā)生振蕩，因?yàn)樗鼈冃枰褜け舜碎g可以接受的分類。這樣一個(gè)

18、簡(jiǎn)單的架構(gòu)降低了系統(tǒng)設(shè)計(jì)的復(fù)雜性，同時(shí)提高了分類的穩(wěn)定性和精確度。第三層更新了從歷史趨勢(shì)信息中預(yù)測(cè)的分類，因此，系統(tǒng)不會(huì)對(duì)污染物尖峰和異常測(cè)量反應(yīng)過度，從而防止了在污染物種類和水平中快速切換。</p><p><b>　　分類精度</b></p><p>　　分類器的性能往往被認(rèn)為是使用了一個(gè)混合矩陣。混合矩陣包含有關(guān)實(shí)際情況的信息（估計(jì)），并預(yù)測(cè)所產(chǎn)生的分類方法?；?/p>

19、合矩陣是一個(gè)n×n大小的包含條件概率的隨機(jī)矩陣，而其每一個(gè)元素定義了其預(yù)測(cè)概率。因此，分類的精確度由以下方程給出：</p><p><b>　　（3）</b></p><p>　　這種估計(jì)說明分類器有嚴(yán)重的誤差，這一點(diǎn)在該系統(tǒng)中是特別重要的?？紤]到誤差的嚴(yán)重性，作者設(shè)計(jì)了一個(gè)低成本的精度估計(jì)方法，即從實(shí)際情況中反映距離對(duì)誤差的影響。作者還使用了硬度系數(shù)作為指

20、標(biāo)，以評(píng)估分類器的性能。硬度系數(shù)糾正了分類器的估計(jì)值與真實(shí)值之間的誤差程度，并預(yù)測(cè)了可能發(fā)生意外的情況。在一個(gè)多級(jí)分類器中，式（4）給出了貝葉斯定理：</p><p><b>　　（4）</b></p><p>　　在該式中，N是指總數(shù)，和分別為混合矩陣中行和列的數(shù)值。 </p><p><b>　　實(shí)驗(yàn)評(píng)估</b><

21、;/p><p>　　作者設(shè)計(jì)了一個(gè)的測(cè)試平臺(tái)，該測(cè)試平臺(tái)包含潤(rùn)滑油在線流量的典型環(huán)境條件，用于測(cè)試SOS在一個(gè)復(fù)雜的環(huán)境中準(zhǔn)確地檢測(cè)劣質(zhì)油的能力。測(cè)試平臺(tái)能模擬真實(shí)環(huán)境的壓力、溫度和流量情況，并提供了一種手段來執(zhí)行污染物的測(cè)試。SOS在測(cè)試平臺(tái)上在線測(cè)試油質(zhì)，以探測(cè)和跟蹤MIL-PRF 9000H型潤(rùn)滑劑中的水體污染，燃料稀釋情況和煙塵污染。在所有的測(cè)試中，傳感器頭部流速均為1.1米/秒（25 psi），并在172k

22、Pa（25磅），51.6℃（125°F）的環(huán)境條件下進(jìn)行。 </p><p>　　在第一階段的測(cè)試中，作者以單調(diào)遞增的方式在每種類型的污染物的變質(zhì)界限下使系統(tǒng)受到污染。作者定義變質(zhì)界限為：煙塵的質(zhì)量比為1％、燃料的體積比為5％、水的體積比為0.2％。在2-3小時(shí)內(nèi)，把每種污染物緩慢加到測(cè)試平臺(tái)中并使污染物在平臺(tái)內(nèi)均勻分布，而且6個(gè)獨(dú)立測(cè)試的污染物是各不相同的。這種辦法表明，除了單一污染物的線性影響外，多

23、種污染物的共同作用和污染物之間的相互作用以及石油添加劑引起的非線性效應(yīng)也會(huì)對(duì)檢測(cè)結(jié)果造成影響。通過在幾個(gè)區(qū)域中的搜索空間（多污染物，低于變質(zhì)界限）中采集數(shù)據(jù)可以更容易地識(shí)別出以前未曾出現(xiàn)過的分類。 </p><p><b>　　優(yōu)化策略</b></p><p>　　作者分析了從上述檢測(cè)方法中收集的數(shù)據(jù)以檢驗(yàn)分類器的性能和決定最佳分類策略，該分類策略確定了健康狀態(tài)時(shí)分類

24、器可以檢測(cè)到的最大變動(dòng)。分類器的高分辨率可以檢測(cè)在健康狀況時(shí)的細(xì)微變化，這對(duì)用戶是非常有利的。然而，分類精度卻與分類策略成反比，因此，有必要權(quán)衡策略和準(zhǔn)確性。 </p><p>　　為了確定最佳的分析策略，作者將污染范圍劃分為10個(gè)級(jí)別和并對(duì)幾個(gè)預(yù)定的水平分類數(shù)據(jù)集的分類器性能進(jìn)行了評(píng)估，但減少了分類準(zhǔn)確性，而且導(dǎo)致策略的增加。雖然分類性能非常高（90％以上的準(zhǔn)確度），但是，策略的缺失會(huì)使用戶的預(yù)見能力大大減弱。

25、由于只有健康、警告和危急三個(gè)狀態(tài)，用戶無法預(yù)測(cè)何時(shí)污染程度將達(dá)到變質(zhì)界限。另外，每個(gè)污染物的8-10級(jí)分類在準(zhǔn)確性上有一個(gè)顯著減少，而這使得用戶能預(yù)測(cè)傳感器輸出。由于這些原因，作者為石油潤(rùn)滑油的多污染物分類選擇了每個(gè)污染物五級(jí)分類。 </p><p><b>　　分類的精度評(píng)估 </b></p><p>　　作者根據(jù)潤(rùn)滑油系統(tǒng)測(cè)試平臺(tái)上收集的數(shù)據(jù)評(píng)估三個(gè)層次的貝葉斯分

26、類器的性能，評(píng)估了采用交叉驗(yàn)證方法的分類器，并使用70％的數(shù)據(jù)進(jìn)行分類，使用剩下的30％用于檢測(cè)。圖3顯示分類器的混合矩陣，在這個(gè)矩陣，行表示系統(tǒng)的實(shí)際分類（基于估計(jì)過的污染物），列表示分類的結(jié)果。對(duì)于一個(gè)理想的分類器，其混合矩陣應(yīng)該有取決于對(duì)角線的所有數(shù)值。錯(cuò)誤分類表現(xiàn)為數(shù)值對(duì)混合矩陣對(duì)角線的偏離。正如圖3所示，分類器能非常準(zhǔn)確地預(yù)測(cè)大多數(shù)分類。例如，SOS對(duì)石油的分類（行1）有95％的準(zhǔn)確性。然而，對(duì)高污染燃料（行10），分類器只取

27、得61％的準(zhǔn)確性。這主要是由于系統(tǒng)燃料污染敏感程度的限制，而且這是當(dāng)前系統(tǒng)中所固有的，因此，高水平的其他類型的污染物分類會(huì)取代燃油污染物分類。</p><p>　　圖3：分類器的混合矩陣</p><p>　　SOS的實(shí)驗(yàn)室測(cè)試結(jié)果</p><p>　　作者進(jìn)行了傳感器的驗(yàn)證測(cè)試，他們?cè)谑芪廴镜臏y(cè)試平臺(tái)與已知或未知的燃料、水和煙塵污染物中，定期對(duì)測(cè)試平臺(tái)上的樣品油進(jìn)行

28、抽樣檢查，并將樣品進(jìn)行了實(shí)驗(yàn)室分析。測(cè)試需要2天時(shí)間，每一天開始時(shí)必須在測(cè)試平臺(tái)進(jìn)行排水、清潔然后注滿。由此產(chǎn)生的實(shí)驗(yàn)室分析報(bào)告證實(shí)，智能石油質(zhì)量檢測(cè)系統(tǒng)有檢測(cè)污染物和對(duì)多污染物同時(shí)進(jìn)行檢測(cè)的能力。 </p><p>　　獨(dú)立分析實(shí)驗(yàn)室根據(jù)美國(guó)材料試驗(yàn)學(xué)會(huì)（ASTM）D3524進(jìn)行氣相色譜分析，以衡量燃料稀釋的程度。圖4表明，實(shí)驗(yàn)室測(cè)量的實(shí)際的燃料水平在用高度準(zhǔn)確性和一致性的傳感器預(yù)測(cè)的燃料稀釋范圍之間跳躍。注：

29、樣品標(biāo)簽“A”指第1天的測(cè)試，“B”指第二天的測(cè)試。實(shí)際超出傳感器預(yù)測(cè)范圍的污染水平，遠(yuǎn)不止一級(jí)。 SOS的取得了非常好的實(shí)驗(yàn)結(jié)果，也有較好的誤差率（美國(guó)材料試驗(yàn)學(xué)會(huì)標(biāo)準(zhǔn)規(guī)格的2％）。 </p><p>　　圖4：使用氣相色譜分析進(jìn)行燃料稀釋水平論證</p><p>　　兩個(gè)獨(dú)立的實(shí)驗(yàn)室使用庫侖卡爾.菲舍爾測(cè)試（美國(guó)ASTM D6304）對(duì)水濃度進(jìn)行分析。圖5顯示了實(shí)驗(yàn)室檢測(cè)的實(shí)際水污物水

30、平和SOS的分類范圍。該圖突出了可能發(fā)生在分析實(shí)驗(yàn)室之間的變化。根據(jù)實(shí)驗(yàn)室的報(bào)告，SOS的分類效果較好，但是實(shí)驗(yàn)室環(huán)境的不一致導(dǎo)致結(jié)論的測(cè)量精度受到限制。實(shí)驗(yàn)室之間的差異最有可能是由于校正鋅烷基二硫代磷酸（ZDDP）時(shí)不適當(dāng)?shù)臏y(cè)量所造成的干擾。因?yàn)镾OS依賴于實(shí)驗(yàn)室分析，因此，分析方法和實(shí)驗(yàn)室的選擇對(duì)傳感器的整體精度是至關(guān)重要的。 </p><p>　　圖5: 使用卡爾.菲舍爾滴定進(jìn)行水污染論證</p>

31、;<p>　　作者還使用了兩個(gè)實(shí)驗(yàn)室估計(jì)石油樣本中的煙塵水平（一個(gè)使用威爾克斯煙塵表，另一個(gè)使用FTIR分析），并且對(duì)SOS的預(yù)測(cè)結(jié)果進(jìn)行比較。值得注意的是，這兩個(gè)實(shí)驗(yàn)室實(shí)際報(bào)告的煙塵水平比實(shí)際值要低。 </p><p><b>　　今后的工作 </b></p><p>　　該傳感器證實(shí)了在測(cè)試平臺(tái)環(huán)境中檢測(cè)多種污染物的能力，作者所解決的幾個(gè)問題，將提高

32、該傳感器在商業(yè)上的成功。傳感器一些改進(jìn)，如擴(kuò)大檢測(cè)頻率范圍，改善溫度補(bǔ)償算法，提高了傳感器對(duì)特殊污染物的靈敏度，這些方法將提高分類器的精確度和分辨率。系統(tǒng)還可以做進(jìn)一步改善，如減少電路規(guī)模，改善溫度的限制，并降低傳感器的尺寸大小，這樣，傳感器可以獲得更廣泛的應(yīng)用。 </p><p>　　在應(yīng)用中，作者已經(jīng)將石油感應(yīng)能力擴(kuò)展到其最初的工作中。傳感器所表現(xiàn)的性能非常好，能在變速箱系統(tǒng)中監(jiān)測(cè)水含量和潤(rùn)滑油質(zhì)量。在柴油機(jī)

33、和變速箱中應(yīng)用傳感器的仿真測(cè)試將用于進(jìn)一步驗(yàn)證傳感器的功能，并檢測(cè)出其潛在的局限性。 </p><p><b>　　結(jié)論 </b></p><p>　　本文中作者介紹了智能石油質(zhì)量檢測(cè)技術(shù)，采用寬帶光譜方法，電化學(xué)技術(shù)，以及先進(jìn)的多傳感器數(shù)據(jù)融合方法，從而設(shè)計(jì)出實(shí)時(shí)、內(nèi)置的石油質(zhì)量分析裝置。從6個(gè)月的連續(xù)測(cè)試和數(shù)據(jù)分析顯示出的結(jié)果表明，該分類器具有靈活性、魯棒性以及準(zhǔn)

34、確性。試驗(yàn)中還確定了需要改進(jìn)的地方，以改善分類器的精確度和分辨率。在實(shí)驗(yàn)室中對(duì)石油樣品的分析證實(shí)，智能石油質(zhì)量檢測(cè)裝置能夠探測(cè)和跟蹤的石油中的燃料、水和煙塵污染水平。進(jìn)一步的試驗(yàn)表明，SOS能適應(yīng)其他更廣范圍的污染。 </p><p>　　作者還提供了額外的安裝和測(cè)試應(yīng)用方法，從而使石油質(zhì)量傳感器可以得到進(jìn)一步的開發(fā)和驗(yàn)證。這些評(píng)估，以及在專門的實(shí)驗(yàn)室進(jìn)行的地面實(shí)況數(shù)據(jù)收集，將提供一種手段去發(fā)展的石油質(zhì)量檢測(cè)技術(shù)

35、，并證明其有能力在這種環(huán)境中追蹤和預(yù)測(cè)石油污染。從根本上說，SOS技術(shù)提供了一個(gè)關(guān)鍵方法去實(shí)現(xiàn)有效的流體系統(tǒng)監(jiān)測(cè)，從而延長(zhǎng)機(jī)器的壽命，盡量減少對(duì)環(huán)境的影響，并降低生命周期成本。 </p><p><b>　　致謝 </b></p><p>　　這一努力的部分工作，得到了海軍研究辦公室（美國(guó)海軍研究局）的支持。作者要感謝伊格納西奧.佩雷斯博士（美國(guó)海軍研究局）以及肯.斯

36、凱德爾，詹姆斯.蘇瓦松和維基.拉里莫爾（NSWC ）的投入和支持。附件2：外文原文</p><p>　　EXPERIENCES AND TESTING OF AN AUTONOMOUS ON-LINE OIL QUALITY MONITOR FOR DIESEL ENGINES</p><p><b>　　Abstract </b></p><p&

37、gt;　　The paper summarizes the author’s application and testing of an oil quality sensor for diesel engine applications. Maintaining healthy fluid systems is critical to keeping machinery in a high readiness state. The au

38、thors describe the oil sensing principles and recent experiences proving the sensor’s ability to autonomously assess oil quality and classify specific diesel oil contaminants. The sensor includes both analog and digital

39、electronics enabling the sensor to perform fluid interrogations, </p><p>　　Key Words: Lubricant Contamination, Lubricant Analytical Techniques, Condition Monitoring, Internal Combustion Engine and Oils, Equi

40、pment Monitoring. </p><p>　　Introduction </p><p>　　Condition-based Maintenance (CBM), an emerging maintenance philosophy, employs active monitoring to determine the health of a component or syst

41、em and enables maintenance based upon the diagnosis and predicted remaining useful life. CBM provides the potential for reduced life cycle maintenance costs, improved safety, and increased operational readiness. Because

42、contamination and in-service degradation of hydraulic fluids, gear oils, lubricants, and other in-service fluids are among the most commo</p><p>　　Several methods exist for real-time condition monitoring of

43、lubricants; most of these methods target one of three main categories: quality, debris, or elemental techniques. Several technologies employ inductive transducer elements to detect particle contamination. However, this t

44、echnology shows limited promise for particle sizes below 500μm, which is insufficient considering that particles between 5 and 20 microns cause 60% of all engine wear. This method is also insensitive to contaminations su

45、</p><p>　　The Smart Oil SensorTM (SOS) described in this paper provides a real-time analysis of in-service fluids that is online, in-situ, real-time, and inexpensive. The goal of this paper is to demonstrate

46、 that the SOS technology overcomes the aforementioned drawbacks inherent in other oil analysis techniques by implementing a novel, broadband impedance spectroscopy based approach. The paper also presents the results of e

47、xtensive testing. </p><p>　　Basis of Measurement </p><p>　　The Smart Oil SensorTM system employs a patent-pending, low-powered, broadband impedance measurement coupled with multi-sensor fusion a

48、nd a model-based analysis package designed to be capable of predicting fluid quality and degradation for a range of fluid systems.</p><p>　　The electrochemical impedance spectroscopy (EIS) approach involves

49、injecting a complex alternating current signal into a system over a wide frequency spectrum and measuring the system’s response to determine oil quality. The impedance of the system is determined by analyzing the differe

50、nces between the injection (excitation) and response signals. By scanning across a wide-range of frequencies, the sensor obtains a measurement that is rich with information and better reflects the actual impedance</p&

51、gt;<p>　　Figure 1: Smart Oil Sensor Principle</p><p>　　Smart Oil SensorTM Design </p><p>　　The SOS (Figure 2) is a stand-alone unit that includes sensing element, signal conditioning, dat

52、a processing, and communications elements in an integrated package. The EIS measurement element is comprised of two concentric cylinders of uniform separation. The authors selected the sensor head geometry to maximize sa

53、mple surface area while minimizing impedance to fluid flow. The SOS houses signal conditioning and data processing electronics capable of functions such as dynamically reconfigurable g</p><p>　　Figure 2: The

54、 Smart Oil Sensor?</p><p>　　The current sensor design supports both RS-232 and Controller Area Network (CAN) communications. Through these interfaces, the sensor can communicate fluid measurements and sensor

55、 status or receive configuration and firmware updates. Including CAN protocol support simplifies integration into existing diagnostic and control systems. </p><p>　　Feature Extraction and Classification Afte

56、r performing an EIS measurement, the sensor executes a series of algorithms to extract fluid quality information from the gathered impedance data. This process includes feature extraction and classification processes. Fe

57、ature extraction is the method used to estimate electrochemical impedance model parameters, such as bulk-resistivity and interfacial properties. The resulting features provide the link between an impedance measurement an

58、d actual fluid pro</p><p>　　The authors leveraged their experiences in multi-sensor data fusion, classification, and data mining to build a classifier that best meets the application requirements for a real-

59、time, in-situ, stand-alone sensor. The authors experimented with several classifier architectures based on linear discriminant analysis, Bayesian probabilistic models, robust fault detection and isolation, and neural net

60、works. To evaluate the classifier methodologies, the authors considered compatibility with the data s</p><p><b>　　（1）</b></p><p>　　A Bayes classifier uses the class posterior probabi

61、lities given by the feature vector as discriminant. A Na.ve Bayes classifier makes the simplifying assumption that the features are independent, given the class. Hence, the classifier uses the discriminant function: <

62、/p><p><b>　　（2）</b></p><p>　　The classifier architecture selected for the identification of fuel, water, and soot within diesel oil consist of three tiers that utilize real-time data fr

63、om the sensors and historical classifications. The first and second tier Bayesian classifiers are connected through a cross-coupled architecture. These cross-coupled classifiers oscillate through the search space as they

64、 hunt for a mutually acceptable classification. Such a simple architecture decreases the complexity of the system design whi</p><p>　　Methods of Determining Classifier Accuracy </p><p>　　Classif

65、ier performance is often interpreted using a confusion matrix. A confusion matrix contains information about the actual (estimated) and predicted classifications generated by a classification system. The confusion matrix

66、 is a stochastic, n ×n sized matrix of conditional probabilities, where each of its elements ( pij ) defines its probability of predicting a class i given an example of an actual class j. Hence, the accuracy of the

67、classifier is given by the equation:</p><p><b>　　（3）</b></p><p>　　Such an estimate does nottake into account the error severity, which is particularly important in this application.

68、To account for error severity, the authors devised a cost-sensitive accuracy estimate that weighs each misclassification with a weighted coefficient ( wij ) that reflects the distance of the misclassification from the ac

69、tual class. </p><p>　　The authors also used the Kappa coefficient as a metric to evaluate the classifier’s performance. The kappa corrects the degree of agreement between the classifier’s predictions and rea

70、lity by considering the proportion of predictions that might occur by chance. In a multi-class classifier, Eq. (4) gives the Fleiss’ kappa：</p><p><b>　?。?）</b></p><p>　　where N is th

71、e total number of instances and xi. and xi are the column and row counts, respectively, of the confusion matrix.</p><p>　　Experimental Evaluation</p><p>　　The authors designed a lubrication test

72、 stand that recreates typical in-line flow environmental conditions to test the SOS’s ability to take accurate readings of contaminated oil in a dynamic environment. The test stand, replicates real-world pressure, temper

73、ature, and flow scenarios, and provides a means to perform seeded fault contamination testing. The SOS was tested inline on the test stand to detect and track water contamination, fuel dilution, and soot contamination in

74、 MIL-PRF 9000H lubric</p><p>　　In the first phase of testing, the authors contaminated the system in a monotonically increasing manner up to the condemning limit of each type of contaminant. The test sponsor

75、 defined condemning limits as 1% soot by mass, 5% fuel by volume, and 2000 ppm (0.2%) water by volume. Slow addition of each contaminant to the test stand over a 2-3 hour period allowed for uniform distribution of the co

76、ntaminant within the test stand. six separate test sequences were performed in which the order of contam</p><p>　　Determination of Optimum Resolution</p><p>　　The authors analyzed the data colle

77、cted from the aforementioned test plan to verify the performance of the classifier and determine the optimal classification resolution. The resolution of the classifier determines the maximum change in health state that

78、the classifier can detect. A classifier with a high resolution can detect fine changes in the health state and is most beneficial to the user. However, classification accuracy is inversely proportional to classifier reso

79、lution. Thus, there is an</p><p>　　To determine the optimum resolution, the authors divided the contamination range into 10 levels of resolution (N) and evaluated the performance of the classifier on several

80、 data sets for these predetermined levels. While classifier performance is very high (above 90% accuracy), for three classes per contaminant, the lack of resolution provides a user little prognostic capability. With only

81、 ‘healthy’, ‘warning’, and ‘critical’states, a user cannot predict when contamination levels will reach a co</p><p>　　Evaluation of Classifier Accuracy </p><p>　　The authors evaluated the perfor

82、mance of the three-tiered Bayesian classifier using data collected on the scaled lube system test stand. The authors evaluated the classifiers using a cross validation approach, which trains the classifier on 70% of the

83、data and uses the remaining 30% for testing. Figure 3 shows a confusion matrix for the classifier. In a confusion matrix, the rows indicate the actual classification of the system (based on estimated contamination) and t

84、he columns indicate the clas</p><p>　　Figure 3: Confusion Matrix Showing Results with Correct and Incorrect Classifications</p><p>　　Lab-based Verification of SOS Results </p><p>　　

85、The project sponsors conducted a sensor verification test in which they contaminated the test stand with known as well as unknown levels of fuel, water, and soot contaminantion, sampled oil from the test stand at regular

86、 intervals, and dispatched the samples to a laboratory for analysis. Testing ran over a 2-day period and the test stand was drained, cleaned and refilled at the start of each day. The resulting laboratory analysis report

87、s confirmed the ability of the Smart Oil SensorTM to detect </p><p>　　The independent analysis lab performed gas chromatography according to ASTM D3524 to measure fuel dilution. Figure 4 shows that the actua

88、l fuel levels measured by the lab fall between the fuel dilution bounds predicted by the sensor with a high level of accuracy and consistency. Note: Samples labeled as set ‘A’ denote day 1 of testing and set ‘B’ denote d

89、ay two. The actual contamination levels that fall outside of the predicted bounds were never off by more than one class. SOS results trend very</p><p>　　Figure 4: Fuel Dilution Level Verification using Gas C

90、hromatography</p><p>　　Two independent laboratories performed water concentration analyses using a coulometric Karl-Fischer test (ASTM D6304). Figure 5 shows the actual water contaminant level detected by th

91、e labs and the bounds of the SOS classification. The plot highlights the variability that can occur between analysis labs. The SOS classifications trend well with both lab reports; however, conclusions on measurement acc

92、uracy are limited by lab inconsistency. The discrepancy between the labs is most likely due to i</p><p>　　The authors also employed two laboratories to estimate the soot levels in the oil samples (one using

93、Wilkes Soot meter and the other using FTIR analysis) and compared these to the SOS predicted results. Both labs actually reported significantly lower soot levels than were actually added. </p><p>　　Figure 5:

94、 Water Contamination Verification using Karl Fischer Titration</p><p>　　Future Work </p><p>　　While the sensor demonstrated the capability of identifying and trending multiple contaminants in a

95、test stand environment, the authors are addressing several issues that will improve the commercial success of the sensor. Improvements such as expanding the interrogation frequency range, improving temperature compensati

96、on algorithms, and increasing the sensor’s sensitivity to particular contaminants will enhance classifier accuracy and resolution. Further improvements such as reducing electronics</p><p>　　The authors have

97、already extended their initial work by employing the oil sensing capability in the applications shown in Table 1. Moving beyond diesel lubes, the sensor performs extremely well while monitoring water content and lubrican

98、t quality within gearbox systems. ‘Realworld’ application testing of the sensor in diesel and gearbox applications will be used to further verify the sensor’s capabilities and identify potential limitations. </p>

99、<p>　　Conclusion </p><p>　　In this paper, the authors describe the Smart Oil SensorTM technology that employs broadband spectroscopy approaches, electrochemical techniques, and advanced multi-sensor dat

100、a fusion methods to present a near real-time, inline oil analysis device. The results from the 6 months of continuous testing and data analysis demonstrate the flexibility and robustness of the classifier and highlight t