1294089 九、發明說明： 【發明所屬之技術領域】 熱管的概念首先由R. S. Gaugler於1944年所提出，但當時並未實用化， 直到 1963 年由 G_M· Grover 申請的「Evaporation-Condensation Heat Transfer1294089 IX. Description of the invention: [Technical field to which the invention pertains] The concept of heat pipe was first proposed by R. S. Gaugler in 1944, but it was not put into practical use until 1963. Evaporation-Condensation Heat Transfer was applied by G_M. Grover.
Device」專利並首度提到熱管(HeatPipe)—詞。熱管基本上可分為毛細構 管與熱虹吸式熱管兩種，兩者都是利用工作流體在相變化時所具有的潛熱來輸 送大量熱量的機構，由於它在微小的溫差下操作就可以有很大的熱傳量，因此 有熱之超導體的美稱。 熱管最初的發展乃應用於太空技術，為因應太空中無重力場之環境上，工 作流體返回加熱部無法依賴重力的作用，故必須利用毛細作用將液體送回，整 個過承中亦不需從外部提供額外的動力，且熱管中央為中空的狀態，整體的質 量非常輕，這是熱管的第一特色，也是應用於太空技術的理由之一，初期的研 究也都是以太空船為中心，而在1966年，美國RCA公司首先將熱管商品化， 從此熱管才廣泛地應用於地面。 熱管的應用||圍非常廣，包括人造衛星、鍋爐的廢熱回收、電器及電子零 件的冷卻、馬達的冷卻、太陽能及地熱的有效利用等。 本發明「智慧型理論評估熱管滲透率之方法與裝置」，係以類神經網路系 統為依據’估算出不同的燒結時間與燒結溫度設計對整體熱管滲透率之影響， 作精準有效的分析判斷以及快速的找出估熱管滲透率設計之最佳化的方法。 1294089 【先前技術】 燒結體内部具有不規則形狀之連續性孔隙，在有水壓力作用之環境下，流 體能在此孔隙内流動，滲透率實驗的目的，乃藉由此實驗瞭解鋼粉在各種不同 的燒結時間與燒結溫度下，金屬燒結之滲透率變化情況，用以測定水在燒結體 孔隙内流動之難易程度；至於滲透率的實驗步驟為： 丨·使用游標卡尺量測燒結試樣的長度及其直徑，並計算燒結試樣之斷面積； 2.將燒結後之試片，在周圍包覆防漏膠帶，以避免水由周圍滲透，包覆好後 | 置入流管中； 3·將乾淨的水由供水裝置將水引入滲透管内，調整供水量略多於滲流量，直 到滲流管内保持固定之水位； 4·由刻度尺讀記皿器水面滲流管溢水口間之水頭距離，並用溫度計量測當時 滲流水之水溫並計計錄之； 5·虽水連續流經燒結體試體一段時間而達穩定流後，在按下碼錶開始計時之 同時’立即將量筒或燒杯置放於導水口下，以便收集流經土樣試體之滲流 水；當流量與時間成線性比例時，可判斷其達穩定流； 6·當已收集到足夠的滲流水後，再按下停錶並同時移開量筒或燒杯並稱得收 集水重； 7·計算水力傳導係數，並計算滲透率繪成圖表。 目前使用技術之缺點·· 1·在進行人工檢測滲透率時，容易因人為疏失而造成錯誤； 2·實驗設備易損耗而導致實驗誤差； .1294089 3·鋼粉試片的製作以及實驗設備保養維修昂貴； ' 《由於目前使用之技術需較多的工作人員進行人工檢測與實驗，因而造成人 力資源的浪費及消耗； 5·目前使用之技術層面在整個工作流程上較無連續性與自動化，容易在工作 程序上發生錯誤及脫序的情況； 6·此發明完全由電縣代替人工完成所需要的工作，因而減少大量人力資源 的消耗； 7·此發明由模酬始絲神經網路推算完成，完全由電腦進行，因而在整個 的工作流程上，具備完善的連續性與自動化；同時在工作上也較具有智慧 型的發展空間。 【發明内容】 建立起―套智慧型理論評估鮮滲透率之方法與裝置，係以類神 ^網路為錄，將其模擬酬之結果供_神_路祕鱗及學習後推算 出不_渗^較計對熱管之影響；_本發微善目紐術之優點如下： 1.由於此估算方法與裝置之研發經冑與製造成本低符合高經濟低成本之效 益； 2·此發明方法與裝置凡全由電腦模擬計算來完成，既快速且較精準，因此不 需投入大量之研發時間與縮短工作時效，使其能夠在最短的時間内獲得最 高之效率；具有時效工作性能，減少因時間因素改變所造成之誤差； 3.由於此發明驗㈣職擬_不會發錢端·失而造成錯誤； 《此發明所使狀設備乡為電觀模擬軟體，@峨備不易損耗； 5·此發明所使用之设備多為電腦及模擬倾，因而維修成本低； 1294089 6.此發明完全由電腦來代替人工完成所需要—作㈣減少大量人力 的消耗； 資源 7·此發明由模擬開始至類神經網路推算完成，完 工作流程上，具備完善的連續性與自動化； =，打’ _在整個的 發展空間。 、在作上也較具有智慧型的 【實施方式】 本發明智慧型理論評估熱管滲透率之方 二大科，[科實財如__路_^’其實财私要分為 • 出估熱管滲透率設計條件，第二部分實施方^厂據’如第5圖所示，估算 網路架構之輸入層進行主轴轉換與降 么以K_L exP肋―方法對類神經 網路運算時間、 弟圖所不，使其大幅減少類神經 第一實施例·· -種㈣雜絲滲透較村 顧 含右整發彻細 置如第5圖所示，其方法流程包 S有學I過程、檢測過程與再學 :一 出節點分別對應於爾紗果，ra L 略嗎入向量& ’而《個輸 路輪出值與縣軸暖㈣Y崎蜂嶋_傳遞神經網 結值調整蝴圭化，以獲取最精破，、化作為目標函數，將神經網路各節點鍵 神經網路每，域得倒傳遞 之步驟包含有： 第6圖所示’以下就神經網路學習過程 步驟一 ··設定初始鍵結值& 魏小亂數，·及沿水平轴之飽和函數之偏差(_认均 步驟二:當停止條件尚未到達 神、遂網路輸出值與訓蜂樣本實際輪出差 1294089 則、於容許值，作麵三到麵十，關結絲料程； v驟二·對每一個訓雜本，作步驟四到步驟九； 步驟四：將輸入向量(熱管渗透率設計參數之屬性向量)撕到隱藏層的 每一個節點作全連結運算； 步驟五:計算隱藏層之輸出值〜，〜=/(:^_〜， i 其中作用函數定義為/(θ— 1 ; 步驟六：計算每-個輸出節點八，^^以一❸； ►步驟七.計算每一個輸出節點之回傳誤差^ — ， (八，hl，2，···，㈣及鍵結修正量^=吨、，辑=吨，其中學習率 〇<α<1，實際輸出值& ; 步驟八.言十算隱藏層每-節點之鍵結修正值、=吵，及回傳誤 ㈣，'hjJ:.ia，…，ρ、; /驟九·更_每個鍵結值〜(_)，〜(_风(_)，4(臟），其中 ί jkinew)-^jk{old)^y uiJ(new)^uiJ(old) + AuiJ 9 · 以峰物•岣，咖幻，㈣+，； 步驟十：測試停止條件是否為真； 檢測過程’如第7 ®所示，'其步驟含括有， 翔-：輸入待測熱管渗透率設計參數樣本之屬性向量； H ‘神、_路評估其熱管滲透較計，騎階後之錄作為神經網路 輸入向量’並計算神經網路各輸出節點輸出值； #驟=·檢測是否有誤判樣本（其中誤判樣本代表神經網路實際評估誤差 9 1294089 大於*許值時之樣本）’如私，將糊樣本雜齡於學習樣本 資料庫以利再學習資料之取得； 步驟四：檢測是否結束’如果否回步驟一，如果是則停止檢測過程； 再學習過程，如第8圖所示，其步驟包含有， 步驟一：將所有誤判樣本加入於學習樣本資料庫； 步驟二：重新精上述學f過程，使神經網路各_權討重新進行調整 達至最佳化’以有效對麵讀目之糊縣不會再發生誤判， 而提升本發明之估算精確度； 藉由上述諸流程之組合，以其熱管滲透率設計之”項屬性向量作為倒傳 知種輸出結果，依序進行學f過程'檢測過程與再學f過程，其中學習過程 中’藉由神經網路利用訓練樣本已知之輸入值與輸出值(即學習樣本資料庫 中，訓練樣本屬性向量與其相對應之輸出結果）調整各節點權重，使神經網路 輸出值與樣本實際輸紐之誤差最小作為目標函數，將各節點麟值調整至最 佳化，以提升神經網路估算精度，學習過程結束後並固定各節點權重，以利檢 測過程之估算；檢測過程中將等待檢測樣本屬性向量作為輸入向量，經由神經 網路進行熱管_率設狀評估，職聰如果有誤繼本，卿誤判樣本資 料儲存於學習樣本資料庫以利再學習資料之取得;再學習過程是經由誤判樣本 加入於學習樣本紐庫，使神輯路織各節點雜，⑽至後續物過程對 於相似或雷同之前述翻樣林再產生翻情形，俾提升智慧型評倾管渗透 率開口之方法與裝置之估算精度。 1294089 本發明智慧型評估熱管滲透率之方法與裝置，其中該智慧型評估熱管滲透 率開口設計之裝置包含有個人電腦内設有學習樣本資料庫與熱管滲透率設計 之方法，其中該學習樣本資料庫儲存有訓練樣本之熱管渗透率設計的屬性向 量，該屬性向量是經由電腦模擬分析後所獲取，每一筆訓練資料包含3〇項數 據，前30項(或前"維)數據為熱管滲透率設計之屬裡向量，依序為&到χ3〇， 其中Χι為第一點溫度，X2為第二點溫度，X3為第三點溫度為第四點溫 度，Xs為第五點溫度，Χό為第六點溫度，X7為第七點溫度，&為第八點溫度， _ 為為第九點溫度，χι〇為第十點溫度，Χιι為第十一點溫度，X12為第十二點溫 度，X!3為第十三點溫度，XH為第十四點溫度，X1S為第十五點溫度，又16為 對應第一點溫度的時間，·Χ17為對應第二點溫度的時間，乂18為對應第三點溫 度的時間，Χ19為對應第四點溫度的時間，χ2〇為對應第五點溫度的時間，χ2ι 為對應第六點溫度的時間，χ22為對應第七點溫度的時間，χ23為對應第八點 溫度的時間，X24為對應第九點溫度的時間，X2S為對應第十點溫度的時間， X26為對應第十一點溫度的時間，义27為對應第十二點溫度的時間，χ28為對應 第十二點溫度的時間’义29為對應第十四點溫度的時間，^:知為對應第十五點 溫度的時間’第31項(或/!+/項)是神經網路輸出節點Y以熱管滲透率，再依 據現有已知標準方法所獲取之評估標準值； 如上所述之智慧型熱管滲透率之方法與裝置，其中該檢測過程步驟二之神 經網路評估熱管滲透率設計驟說明如下： 步驟一：將學習過程將中所訓練完成之神經網路鍵結值固定、，％，$片； 步驟二：對每一個測試樣本，作步驟三至步驟五； 11 1294089 步驟三:將，入向量(熱管渗透率設計之屬性向量)撕到隱藏層的每一個 節點作全連結運算； 步驟四·計算隱藏層之輸出值;^ ’其中伽函數定義為 步驟五·計算每—個輸出層之輸出值λ，。 j 第二實施例： 士一種智慧型理論評估熱管滲透率之方法與裝置，如第！圖所示，其方法 含有學程、細雜與再”聰，其巾學習與程之 步驟含括有， “ 從予S樣本資料庫，藉由全部訓練樣本前π項熱管渗透率的屬性向 量（訓練樣本之前，行數據），計算相關矩陣Λ (auto_論 輪X)，R = x调㈣，其h是熱管滲透率之屬性向量，♦) 是類另J /出現之機率讀函數，Ε{χχ}是在第，個麵的屬性向量 期望值，w為熱管滲透率之屬性向量數目； 步驟， « 將特徵向#正規化(nGnnalize)，設起始; 求取降階矩陣Φ ’如第！圖所示，降階矩陣φ由相關矩陣及之最大 ‘個特徵值所對應之讀特徵向量所組成m· 步騾四•·主軸轉換與降階，如第 4 示z圖所不，由％ ，將輸入向量由易之” 维降階至义之m維，其中％是主軸向量； 步驟五：棚雜梅__#，㈣3 _鮮以⑽為 12 1294089 輸入向量，由倒傳遞神經網路對訓練樣本進行學習評估訓練樣本 熱管滲透率設計條件，如果估算誤差小於容許值，則回步綠二， 輸入向量再降一階重複上述步驟，直到找出可以估算所需之最低 (或最少)主軸維度為止，如果估算誤差大於容許值，則估算縮需之 最少主轴維度為w = m + l並停止學習過程； 檢測過程之步驟如下， 步驟一：輸入待測之熱管滲透率設計條件之屬性向量； 步驟二：將熱管滲透率設計條件之屬性向量進行主軸轉換與降階，藉&七 expansion方法將熱管滲透率設計條件之屬性向量進行主軸轉換與 降階，其中主轴降階之維度，是由學習與降階過程中維持熱管滲 透率設計條件估算精度所需最少之主軸向查數； 步驟三：神細路職鮮滲透率辑，紐階後之主⑽騎經網路輪 入向量，並計算神經網路各輸出節點之輸出值，以最大輸出值之 輸出節點相對應於所代表之熱管滲透率設計條件，作為檢測樣本 之熱管滲透率設計； 步驟四·檢測是否有誤判樣本，如果有，將誤判樣本魏儲存於學習樣本 資料庫以利再學習資料之取得； 步驟五·檢測是騎束，如果雜步驟_，如果是赚止檢測過程； 再學習過程，如第4圖所示，其步驟含括有，. '驟·將所有誤判樣本加人於學習樣本資料庫； 竭- ·鱗進行上述學雜_抛，使神賴路各節點權重可重新進 13 1294089 行調整達至最佳化，以有效對近似或雷同之誤判樣本不會再發生 誤判’而提升本發明之估算精確度； 藉由上述諸流程之組合，以不同的燒結溫度與時間w項屬性向量作為倒 傳遞神經網路的《個輸入向量〜，=1，2，···，λ，$個輸出節點九，=1，2，".^分別對 應於S種輸出向量，因此Υ以滲透率，然後再依序進行學習與降階過程、檢測 過程與再學I過程，其中學習與降階過程中，藉由κ丄方法將熱管 滲透率設計參數之屬性向量轉換至正交主轴上，避免屬性向量彼此干擾，並求 取維持估算精度所需最少之峰向量數，崎低神經網路估算之複雜度，同時 神經網路利用訓練樣本已知之輸入值與輸出值(即學習樣本資料庫巾，訓練樣 本屬性向量與其相對應之滲透率設計條件）調整各節點權重，使神經網路輸出 與樣本實轉丨值之縣最小作為目標减，將各軸鍵結侧整至最佳 化’以提升神經網路估算精度，學習與降階過程結束後並固定各節點權重，以 利檢測過程德算；檢_程情等碰纖本屬軸量_ Μ哪 方_行主_換與降階，並將降階後之主軸作為輸入向量經由神經網路進行 滲透率射#估，評估過程如果有誤判樣本，則將誤判樣本資料儲存於學習樣謂 2料庫關再學轉料之取得;再學f過程觸誤判樣本加人於學習樣本 貝料庫使K L expansion重新調整主軸方位與神經網路調整各節點權重，以 達至後續檢_轉於她讀狀前麟雛林再產生糊_，俾提升 9慧型評絲管較較雜絲置德雜度。 、如上所述之智慧型熱管滲透率之方法與裝置，其中該智慧型評估熱管滲 透率之裝置包含有個人電腦⑽辦習樣本資料庫_型熱管滲透率之方 1294089 法’該學習樣本資料庫儲存有訓練樣本的設計參數之屬性向量，該設計之屬性 向量是經由模擬分析後所獲取，每一筆訓練資料包含31項數據，前30項(或 則”維)數據為熱管滲透率設計參數之屬性向量依序為Xl到χ3〇，其中Χι為第 一點溫度，X2為第二點溫度，χ3為第三點溫度，為第四點溫度，Xs為第五 點溫度，X6為第六點溫度，叉7為第七點溫度，X3為第八點溫度，叉9為第九點 溫度’ χ10為第十點溫度，Xll為第十一點溫度，Xi2為第十二點溫度，Xi3為 第十二點溫度，為第十四點溫度，Xis為第十玉點溫度，Xi6為對應第一點 籲溫度的時間，χ17為對應第二點溫度的時間，Xi8為對應第三點溫度的時間， 為對應第四點溫度的時間，X2〇為對應第五點溫度的時間，X21為對應第六 點溫度的_，X22鱗應第七點溫朗賴，為對絲八點溫度的時間， X24為對應第九點溫度的時間，Xls為對應第十點溫度的時間，知為對應第十 -點溫度的時間，X”為對應第十二點溫度的時間，知為對應第十三點溫度 的時間，X29為對應第十四點溫度的時間，义❶為對應第十五點溫度的時間， 第31項(或㈣項)疋神經網路輪出節點¥以滲透率，再依據現有已知鮮方 •法所獲取之評估標準值。 如上所述之智慧型評估熱管渗透率之方法與裝置，其中該學習與降階過 程步驟五之_祕評雜雜轉之料餅，是赠狀錄向量作為倒 傳遞神經網路的輸入向量，而M固輪出節點分別對躺種輸出向量，藉由倒 傳遞神經網路輸出值與樣本實際輪出狀誤差最小化作為目標函數，將神經網 路各節點U調整至最佳化，簡取最精確之輸出結果，並獲得倒傳遞神經 網路每-個節點之最佳鍵結值，町就神經網路學f步驟·如下： 15 Ϊ294089 步驟―：設定初始鍵結值w及沿水平轴之飽和函數之偏差㈣θ肩 為很小亂數； s’田停止條件尚未到達時，神經網路輸出值與訓練樣本實際輪岐· 許值’作麵三到轉十，酬絲學習過程； 步驟三：對每一個訓練樣本，作步驟四到步驟九； 步驟四·將輸入向量(主輛向量傳送到隱藏層的每一個節點作全連結運 算； α 其中作用函數定義為·; · 步釋六：計算每一個輸出節點八，； 步驟七：計算每-個輪出節點之回傳誤差， ·0^ = 1，2，···4及鍵結修正量^=吨、，化=吨，其中學習率 0<以1，實際輪出值。 鲁步驟八：計算隱藏層每—節點之鍵結修正值Η=泌;〜，△—;及回傳誤The Device patent first mentions the HeatPipe-word. The heat pipe can be basically divided into a capillary structure and a thermosyphon heat pipe, both of which are mechanisms for utilizing the latent heat of the working fluid in the phase change to transport a large amount of heat, since it can be operated under a slight temperature difference Great heat transfer, so it has the reputation of hot superconductor. The initial development of heat pipes was applied to space technology. In the environment where there is no gravity field in space, the working fluid returning to the heating part cannot rely on gravity. Therefore, it is necessary to use capillary action to return the liquid. The entire overrun does not need to be external. Providing additional power, and the central part of the heat pipe is hollow, the overall quality is very light, which is the first feature of the heat pipe and one of the reasons for the application of space technology. The initial research is also centered on the spacecraft. In 1966, the US RCA company first commercialized the heat pipe, and since then the heat pipe has been widely used on the ground. The application of heat pipes|| is very wide, including satellites, waste heat recovery of boilers, cooling of electrical and electronic parts, cooling of motors, efficient use of solar energy and geothermal energy. The invention relates to a method and a device for evaluating the permeability of a heat pipe by an intelligent theory, which is based on a neural network system to estimate the influence of different sintering time and sintering temperature design on the overall heat pipe permeability, and accurately and effectively analyze and judge And quickly find ways to optimize the design of heat pipe permeability. 1294089 [Prior Art] The continuous pores with irregular shapes inside the sintered body, in which the fluid can flow in the pore under the action of water pressure, the purpose of the permeability experiment is to learn the steel powder in various experiments. The variation of the permeability of the metal during the different sintering time and sintering temperature is used to determine the ease of water flow in the pores of the sintered body; the experimental steps for the permeability are: 丨 The length of the sintered sample is measured using a vernier caliper And its diameter, and calculate the broken area of the sintered sample; 2. After the sintered test piece, wrap the leak-proof tape around to avoid the water from penetrating from the surrounding, after being wrapped, placed in the flow tube; The clean water is introduced into the permeation pipe by the water supply device, and the water supply is adjusted slightly more than the seepage flow until the water level in the seepage pipe is maintained; 4. The water head distance between the overflow pipe of the water surface seepage pipe is read by the scale and used Temperature measurement measures the water temperature of the seepage water at the time and counts it; 5. Although the water continuously flows through the sintered body test body for a period of time and reaches a steady flow, start timing by pressing the code table. Immediately place the measuring cylinder or beaker under the water guiding port to collect the seepage water flowing through the soil sample; when the flow rate is linearly proportional to the time, it can be judged to reach a steady flow; 6. When enough has been collected After seeping the water, press the stopwatch and simultaneously remove the measuring cylinder or beaker and weigh the collected water; 7. Calculate the hydraulic conductivity and calculate the permeability as a graph. Disadvantages of the current use of technology ··· When performing manual detection of permeability, it is easy to cause errors due to human error; 2. Experimental equipment is easy to wear and cause experimental error; .1294089 3·Preparation of steel powder test piece and maintenance of experimental equipment Expenditure is expensive; 'Because the technology currently used requires more staff to carry out manual testing and experimentation, resulting in waste and consumption of human resources; 5. The technical level currently used is less continuous and automated throughout the workflow. It is easy to make mistakes and out-of-order in the working procedure; 6. This invention completely replaces the labor required by the electricity county to complete the required work, thus reducing the consumption of a large amount of human resources; Completed, completely computer-based, so in the entire workflow, with complete continuity and automation; at the same time, it also has a smarter development space in the work. [Summary of the Invention] Establishing a method and device for assessing the fresh penetration rate of the wisdom theory, the system is based on the class of the gods, and the results of the simulation are used for the _ God _ road secret scales and learning after calculation The effect of osmosis on the heat pipe is as follows: 1. The advantages of this method are as follows: 1. The development and cost of this estimation method and device are low in cost and high in cost and low cost; And the device is all done by computer simulation, which is fast and accurate, so it does not require a lot of research and development time and shortened work time, so that it can get the highest efficiency in the shortest time; The error caused by the change of time factor; 3. Because of the invention (4) job _ will not send money, the error will result in the wrong; "The invention of the invention is the electrical simulation software, @峨备 is not easy to wear; 5 ·The equipment used in this invention is mostly computer and analog tilt, so the maintenance cost is low; 1294089 6. This invention is completely replaced by computer instead of manual completion - (4) to reduce the consumption of a large amount of manpower; Analog neural networks begin to calculate complete, the complete workflow, with perfect continuity and automation; =, hit '_ in the whole development. In terms of work, it is also more intelligent. [Embodiment] The intelligent theory of the invention evaluates the heat pipe penetration rate of the two major subjects, [Ke Shicai as __路_^' is actually divided into: • Permeability design conditions, the second part of the implementation of the factory ^ according to the figure, as shown in Figure 5, the input layer of the network architecture is estimated to perform spindle transformation and reduction with K_L exP ribs - method for neural network operation time, brother map No, it greatly reduces the nerve-like first embodiment. · - (4) hybrid wire penetration is better than that of the village. The right hair is finely arranged as shown in Figure 5, and its method flow package S has a learning I process and a detection process. And re-learning: a node corresponding to the yarn, ra L slightly into the vector & 'and a "transit value and the county axis warm (four) Y-snake _ _ transfer neural network value adjustment butterfly, In order to obtain the most sophisticated, and as the objective function, the steps of the neural network of each node of the neural network are reversed. The steps of the domain are as follows: Figure 6 below, the following steps are taken in the neural network learning process. Set the initial key value & Wei small random number, · and the saturation function along the horizontal axis Poor (_ acknowledgment step 2: When the stop condition has not yet reached God, 遂 network output value and training bee sample actual wheel difference 1294089, in the allowable value, face three to face ten, close the wire course; v second • For each training, proceed to Step 4 to Step 9; Step 4: Tear the input vector (the attribute vector of the heat pipe permeability design parameter) to each node of the hidden layer for the full-join operation; Step 5: Calculate the hidden layer The output value is ~, ~=/(:^_~, i where the action function is defined as /(θ-1; Step 6: Calculate each output node eight, ^^ to one point; ►Step seven. Calculate each one The return error of the output node ^ - , (eight, hl, 2, ···, (four) and the key correction amount ^ = tons, series = tons, where the learning rate 〇 < α < 1, the actual output value & Step 8. Words are hidden in the hidden layer per-node bond correction value, = noisy, and back error (four), 'hjJ:.ia, ..., ρ,; / IX · more _ each key value ~ (_), ~(_风(_), 4(dirty), where ί jkinew)-^jk{old)^y uiJ(new)^uiJ(old) + AuiJ 9 · to peaks • 岣, 咖幻, (d) +,; Step 10: Test stop Whether the condition is true; the detection process is as shown in the 7th ®, 'its steps include, Xiang-: input the attribute vector of the sample design parameter of the heat pipe permeability to be tested; H 'God, _ road to evaluate its heat pipe penetration ratio After the riding step is recorded as the neural network input vector' and the output values of the output nodes of the neural network are calculated; #STEP=·Detect whether there is a misjudgment sample (where the misjudged sample represents the actual evaluation error of the neural network 9 1294089 is greater than * value Sample of time) 'If you are private, you will use the sample of the sample to learn the sample database to facilitate the acquisition of the data. Step 4: Check if it ends. If you don't go back to step 1, if it is, stop the detection process; then learn the process, As shown in Figure 8, the steps include: Step 1: Add all the misjudgment samples to the learning sample database; Step 2: Re-finish the above process f, so that the neural network can be re-adjusted to the most Jiahua's effective comparison of the past will not lead to misjudgment, but improve the estimation accuracy of the present invention; by the combination of the above processes, the design of the heat pipe permeability The vector is used as the inverted output of the seed, and the process of the process of detecting and re-learning f, in which the input and output values known by the training samples are used by the neural network (ie, the learning sample database) In the training sample attribute vector and its corresponding output result) adjust the weight of each node, so that the error between the neural network output value and the sample actual input is minimized as the objective function, and the node values of each node are adjusted to optimize to improve the nerve Network estimation accuracy, after the end of the learning process and fixed the weight of each node, in order to facilitate the estimation of the detection process; in the detection process will wait for the detection sample attribute vector as an input vector, through the neural network for heat pipe _ rate design evaluation, In error, the sample data is stored in the learning sample database to facilitate the acquisition of the learning materials. The learning process is added to the learning sample vacuum through the misjudgment sample, so that the gods can woven the various nodes, (10) to the follow-up The process is to reproduce the similar situation or similar to the aforementioned reversal forest, and to improve the wisdom of the inclined pipe permeability opening method and The accuracy of the estimation set. 1294089 The method and device for intelligently evaluating the heat pipe permeability of the present invention, wherein the smart device for evaluating the heat pipe permeability opening design comprises a method for designing a learning sample database and a heat pipe permeability rate in a personal computer, wherein the learning sample data The library stores the attribute vector of the heat pipe permeability design of the training sample. The attribute vector is obtained after computer simulation analysis. Each training data contains 3 items of data, and the first 30 items (or the former " dimensional) data are heat pipe penetration. The vector of the rate design belongs to & to χ3〇, where Χι is the first point temperature, X2 is the second point temperature, X3 is the third point temperature is the fourth point temperature, and Xs is the fifth point temperature, Χό is the sixth temperature, X7 is the seventh temperature, & is the eighth temperature, _ is the ninth temperature, χι〇 is the tenth temperature, Χι is the tenth temperature, and X12 is the tenth. Two-point temperature, X!3 is the thirteenth temperature, XH is the fourteenth temperature, X1S is the fifteenth temperature, and 16 is the time corresponding to the first temperature. ·Χ17 is the temperature corresponding to the second point. Time, 乂18 is the corresponding number The time of the point temperature, Χ19 is the time corresponding to the temperature of the fourth point, χ2〇 is the time corresponding to the temperature of the fifth point, χ2ι is the time corresponding to the temperature of the sixth point, χ22 is the time corresponding to the temperature of the seventh point, χ23 is the corresponding time At 8 o'clock temperature, X24 is the time corresponding to the ninth temperature, X2S is the time corresponding to the tenth temperature, X26 is the time corresponding to the eleventh temperature, and the meaning 27 is the time corresponding to the twelfth temperature, χ28 In order to correspond to the temperature of the twelfth point, '29 is the time corresponding to the temperature of the fourteenth point, ^: the time corresponding to the temperature of the fifteenth point' is the 31st item (or /!+/ item) is the neural network. The output node Y is based on the heat pipe permeability, and is based on the evaluation standard value obtained by the existing known standard method; the smart heat pipe permeability method and device as described above, wherein the neural network of the second step of the detection process evaluates the heat pipe permeability The design steps are as follows: Step 1: The neural network key value of the training completed in the learning process is fixed, %, $ piece; Step 2: For each test sample, step 3 to step 5; 11 1294089 three: The input vector (the attribute vector of the heat pipe permeability design) is torn to each node of the hidden layer for the full-join operation; Step 4: Calculate the output value of the hidden layer; ^ 'where the gamma function is defined as step 5. Calculate each output The output value of the layer is λ,. j Second Embodiment: A method and device for evaluating the heat pipe penetration rate by a smart theory, such as the first! As shown in the figure, the method contains the course, the fine and the "Cong Cong, the steps of the learning and the process of the towel are included," from the S sample database, the attribute vector of the π heat pipe permeability before all the training samples (Before training the sample, the row data), calculate the correlation matrix Λ (auto_ argument X), R = x (4), where h is the attribute vector of the heat pipe permeability, ♦) is the probability J/occurrence read function, Ε{χχ} is the expected value of the attribute vector in the first and second faces, and w is the number of attribute vectors of the heat pipe permeability; step, « normalizes the feature to #nnnize, sets the start; finds the reduced order matrix Φ ' The first! As shown in the figure, the reduced order matrix φ is composed of the correlation matrix and the read eigenvector corresponding to the largest 'eigenvalues'. m· Step •4·· Spindle transformation and reduction, as shown in the fourth graph, the graph is not , the input vector is reduced from the dimension of the dimension to the m dimension of the meaning, where % is the principal axis; Step 5: shed mei __#, (4) 3 _ fresh to (10) is 12 1294089 input vector, by inverse neural network The training sample is used to evaluate the design conditions of the heat pipe permeability of the training sample. If the estimation error is less than the allowable value, then return to green two, and the input vector is further reduced by one step to repeat the above steps until the minimum (or minimum) that can be estimated is found. Up to the spindle dimension, if the estimated error is greater than the allowable value, the minimum spindle dimension of the estimated shrinkage is w = m + l and the learning process is stopped; the steps of the detection process are as follows, Step 1: Input the properties of the heat pipe permeability design conditions to be tested Vector; Step 2: Perform the spindle transformation and reduction of the attribute vector of the heat pipe permeability design condition, and perform the spindle transformation by using the attribute method of the heat pipe permeability design condition by the & seven expansion method. The order of reduction, in which the dimension of the main axis is reduced, is the minimum number of main axial inspections required to maintain the accuracy of the design conditions of the heat pipe permeability during the learning and reduction process; Step 3: The fine penetration rate of the Shenjing Road, after the New Zealand The main (10) rides through the network to enter the vector, and calculates the output value of each output node of the neural network. The output node corresponding to the maximum output value corresponds to the heat pipe permeability design condition represented by the heat pipe permeability design of the test sample. Step 4 · Detect whether there is a misjudgment sample. If there is, the sample will be stored in the learning sample database to facilitate the re-learning of the data. Step 5: Detection is riding, if the step _, if it is the profit detection process The re-learning process, as shown in Figure 4, includes steps in it. 'Sc. Add all misjudged samples to the learning sample database; exhaust--scales to carry out the above-mentioned learning _ throwing, so that God Lai Road The node weights can be re-adjusted to 13 1294089 row adjustments to optimize, so that the approximation or similar misjudgment samples will no longer be misjudged' to improve the estimation accuracy of the present invention; The combination of different sintering temperatures and time w attribute vectors as the "input vector ~, 1, 2, ···, λ, $ output nodes 九,=1,2," ;.^ corresponds to the S kinds of output vectors, respectively, so the permeability and then the learning and reduction process, the detection process and the re-learning I process, in which the learning and reduction process, by the κ丄 method The attribute vector of the heat pipe permeability design parameter is converted to the orthogonal main axis to avoid the attribute vectors from interfering with each other, and to obtain the minimum number of peak vectors required to maintain the estimation accuracy, the complexity of the estimation of the neural network, and the use of the neural network. The known input values and output values of the training samples (ie, the learning sample data library, the training sample attribute vector and the corresponding permeability design conditions) adjust the weights of each node, so that the neural network output and the sample real-time depreciation are the smallest The target is reduced, and the axis bonding side is optimized to 'improve the neural network estimation accuracy. After the learning and reduction steps are completed, the weights of each node are fixed to facilitate the detection process. The fiber is the axis quantity _ Μ 方 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ The sample data is stored in the learning sample 2 and the re-learning of the material is obtained; the re-learning process is added to the learning sample, and the KL expansion re-adjusts the spindle orientation and the neural network to adjust the weight of each node. After reaching the follow-up inspection _ turn to her reading before the Lin Linlin and then produce paste _, 俾 9 9 慧 慧 慧 评 评 评 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 The method and device for intelligent heat pipe permeability as described above, wherein the intelligent device for evaluating the heat pipe permeability includes a personal computer (10) sample sample database _ type heat pipe permeability rate 1294089 method 'this sample database The attribute vector of the design parameter of the training sample is stored. The attribute vector of the design is obtained through simulation analysis. Each training data contains 31 items of data, and the first 30 items (or "dimensional" dimension data are heat pipe permeability design parameters. The attribute vector is sequentially from X1 to χ3〇, where Χι is the first point temperature, X2 is the second point temperature, χ3 is the third point temperature, the fourth point temperature, Xs is the fifth point temperature, and X6 is the sixth point. Temperature, fork 7 is the seventh temperature, X3 is the eighth temperature, fork 9 is the ninth temperature ' χ 10 is the tenth temperature, X11 is the tenth temperature, Xi2 is the twelfth temperature, Xi3 is The twelfth point temperature is the fourteenth point temperature, Xis is the tenth jade point temperature, Xi6 is the time corresponding to the first point temperature, χ17 is the time corresponding to the second point temperature, and Xi8 is the time corresponding to the third point temperature. Time, for the fourth point temperature The time, X2〇 is the time corresponding to the temperature of the fifth point, X21 is the _ corresponding to the temperature of the sixth point, and the X22 scale should be the seventh point of temperature, which is the time of the eight-point temperature of the wire, and X24 is the temperature corresponding to the ninth point. Time, Xls is the time corresponding to the tenth temperature, known as the time corresponding to the tenth-point temperature, X" is the time corresponding to the temperature of the twelfth point, and is known as the time corresponding to the thirteenth temperature, X29 is corresponding The time of the fourteenth temperature, the time is the time corresponding to the temperature of the fifteenth point, the third item (or (four) item) is the permeability of the neural network from the node, and then according to the existing known fresh method The evaluation criteria value obtained. The intelligent method and device for evaluating the heat pipe permeability as described above, wherein the learning and reducing step process step 5 is the input vector of the back-transfer neural network. The M-solid-out node respectively outputs the vector to the lying seed, and the neural network output node and the actual round-out error are minimized as the objective function, and the neural network nodes U are adjusted to be optimized. The most accurate output, and obtain the optimal key value for each node of the inverted neural network, the following steps: 15 Ϊ294089 Step-: Set the initial key value w and along the horizontal axis The deviation of the saturation function (4) θ shoulder is a small random number; when the s' field stop condition has not yet arrived, the neural network output value and the training sample actual rim·value value face three to ten, the reward learning process; Three: For each training sample, do step four to step nine; Step four · Transfer the input vector (the main vector is transmitted to each node of the hidden layer for full-join operation; α where the action function is defined as · · · Step 6 : Calculate each output node eight, Step 7: Calculate the backhaul error of each round-out node, · 0^ = 1, 2, ··· 4 and the key correction amount ^= ton, ization = ton, where Learning rate 0 < to 1, the actual round value. Lu step eight: calculate the hidden layer per-node bond correction value Η = secretion; ~, △ -; and back error
W ^^；=(Σ^)Α,(1-Λ,) . (hjJ = i,2,...,p) ; φ 步驟九:更新每一個鍵結值％和)，〜(·)Α(_)，士㈣，其中 〜(卿)=)+△〜， ^k(new) = ek(〇ld)^A0k » 0j(new)^0J(〇ld) + A0J ； 步驟十：測試停止條件是否為真； 如上所述之智慧型評估熱管滲透率設計之方法與裝置，其中該檢測過程 •步驟三之神經網路評估熱管滲透率設計步驟說明如下： ， 16 1294089 步驟一：將學習與降階過程中所訓練完成之神經網路鍵結值固定 步驟二··對每一個測試樣本，作步驟三至步驟五； 步驟三：將輸入向量(主轴向量)JC,·傳送到隱藏層的每一個節點作全連結運 算； 步驟四：計算隱藏層之輸出值沁，士），其中作用函數定義為 步驟五：計算每一個輸出層之輸出值λ，。W ^^;=(Σ^)Α,(1-Λ,) . (hjJ = i,2,...,p) ; φ Step 9: Update each key value % and ), ~(·) Α(_),士(四), where ~(卿)=)+△~, ^k(new) = ek(〇ld)^A0k » 0j(new)^0J(〇ld) + A0J ; Step 10: Test Whether the stopping condition is true; the method and device for intelligently evaluating the heat pipe permeability design as described above, wherein the neural network evaluation heat pipe permeability design steps of the detection process and the third step are as follows: , 16 1294089 Step 1: Will learn The neural network key value fixed in the process of the reduction step is fixed. Step 2· For each test sample, perform step 3 to step 5; Step 3: Transfer the input vector (spindle vector) JC, · to the hidden layer Each node performs a full join operation; Step 4: Calculate the output value of the hidden layer 沁, 士), where the action function is defined as step five: calculate the output value λ of each output layer.
II
17 1294089 【圖式簡單說明】 第1圖K-LExpansion結合神經網路之智慧型評估熱管滲透率架構圖 _ 第2圖智慧型評估熱管滲透率之特徵選擇與降階流程圖 第3圖K-LExpansion結合神經網路之智慧型評估熱管滲透率之學習架構 流程圖 ' 第4圖K-LExpansion結合神經網路之智慧型評估熱管滲透率之再學習架構 流程圖 第5圖倒傳遞神經網路架構 第6圖智慧型評估熱管滲透率之神經網路訓練過程流程圖 第7圖智慧型評估熱管滲透率之神經網路測試過程流程圖 第8圖智慧型評估熱管滲透率之神經網路再學習過程流程圖 參 【主要元件符號說明】 I為第一點溫度 X2為第二點溫度 X3為第三點溫度 X4為第四點溫度 X5為第五點溫度 X6為第六點溫度 x7為第七點溫度 Xg為第八點溫度 Χίο為第九點溫度 xu為第十點溫度 Xl2為第十一點溫度 Xl2為第十二點溫度 \13為第十三點溫度 x14為第十四點溫度 Xl5為第十五點溫度 Y為熱管滲透率17 1294089 [Simple description of the diagram] Figure 1 K-LExpansion combined with the intelligent evaluation of the heat pipe permeability structure of the neural network _ Figure 2 Intelligent evaluation of the heat pipe permeability characteristics selection and reduction of the flow chart Figure 3 K- LExpansion combined with neural network intelligent evaluation of heat pipe permeability learning architecture flow chart 'Figure 4 K-LExpansion combined with neural network intelligent evaluation of heat pipe permeability re-learning architecture flow chart Figure 5 inverted transmission neural network architecture Figure 6: Flow chart of neural network training process for intelligent evaluation of heat pipe permeability. Figure 7 Flow chart of neural network test process for intelligent evaluation of heat pipe permeability. Figure 8: Neural network re-learning process for intelligent evaluation of heat pipe penetration rate Flow chart reference [main component symbol description] I is the first point temperature X2 is the second point temperature X3 is the third point temperature X4 is the fourth point temperature X5 is the fifth point temperature X6 is the sixth point temperature x7 is the seventh point Temperature Xg is the eighth point temperature Χίο is the ninth point temperature xu is the tenth point temperature Xl2 is the eleventh point temperature Xl2 is the twelfth point temperature\13 is the thirteenth point temperature x14 is the fourteenth point temperature Xl5 is The fifteenth point temperature Y is the heat pipe permeability
x16為對應第一點溫度的時間 Χπ為對應第二點溫度的時間 Xl8為對應第三點溫度的時周 Xl9為對應第四點溫度的時間 Χ20為對應第五點溫度的時間 χ2ι為柯應第六點i度的時間 X22為對應第七點溫度的時間 X23為對應第八點溫度的時間 X24為對應第九點溫度的時間 X25為對應第十點溫度的時間 X26為對應第十一點溫度的時間 X27為對應第十二點溫度的時間 X28為對應第十三點溫度的時間 X29為對應第十四點溫度的時間 X30為對應第十五點溫度的時間 18X16 is the time corresponding to the first point temperature Χπ is the time corresponding to the second point temperature Xl8 is the time corresponding to the third point temperature X11 is the time corresponding to the fourth point temperature Χ20 is the time corresponding to the fifth point temperature χ2ι is Ke Ying The time X22 of the sixth point i is the time corresponding to the temperature of the seventh point X23 is the time corresponding to the temperature of the eighth point X24 is the time corresponding to the temperature of the ninth point X25 is the time corresponding to the temperature of the tenth point X26 is corresponding to the eleventh point The time X27 of the temperature is the time corresponding to the temperature of the twelfth point X28 is the time corresponding to the temperature of the thirteenth point X29 is the time corresponding to the temperature of the fourteenth point X30 is the time 18 corresponding to the temperature of the fifteenth point