201245997 六、發明說明: 【發明所屬之技術領域】 ㈣ίΓ明是有關於—種發電機效能·方法及裝置,且 有’—種風力發電機之轉子系統的效能評估方法 【先前技術】 ,力,錢是近幾年來發展蓬勃的新興雜產業。截 、、先计王球裝機各量每年增長超過如%,年安裝量 ^ 彳心千瓦發電量將占全球總量的12%。以全球乾 淨^源的彳又資趨勢與發展來看’大型風力發電仍是目前增 長取快的選擇方案。囿於氣候風場難醇確預測與不穩定 的It A之下’發展—套監控系統來預先評估目前發電效能 ,可能發生故障,藉以穩定風力發電發電品f的需求一直 是此一產業之重點發展項目。 風力發電機為一種透過轉子將風能轉換成機械能,再 經過齒輪系統升速,最後透過發電機使機械能轉變為電能 的一,裝置。影響風力發電機發電效能的因素非常多,例 如風場的穩定與否、機械能轉換過程中的能量損耗等,但 其中最主要的因素還是取決於風力發電機如何在環境中擷 取最多的風能,並將風能有效的轉換成機械能,也就是所 謂的轉子效率。以氣體動力學的角度來看,可藉由氣動轉 矩、氣動功率簡單模擬轉子效率: 201245997201245997 VI. Description of the invention: [Technical field to which the invention pertains] (4) Γ 是 是 是 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机 发电机Money is a booming new hybrid industry in recent years. The total amount of the installed capacity of the ball is increased by more than 100% per year. The annual installation volume ^ 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千 千Looking at the trend and development of the world's clean and dry sources, large-scale wind power generation is still the current choice for growth. In the climate wind field, it is difficult to predict and unstable under the It A 'development-set monitoring system to pre-evaluate the current power generation efficiency, possible failure, so that the demand for stable wind power generation products has always been the focus of this industry. Development project. A wind turbine is a device that converts wind energy into mechanical energy through a rotor, then accelerates through a gear system, and finally converts mechanical energy into electrical energy through a generator. There are many factors affecting the power generation efficiency of wind turbines, such as the stability of wind farms, energy loss during mechanical energy conversion, etc., but the most important factor depends on how wind turbines draw the most wind in the environment. Energy and efficient conversion of wind energy into mechanical energy, also known as rotor efficiency. From a gas dynamic point of view, the rotor efficiency can be easily simulated by pneumatic torque and aerodynamic power: 201245997
Jy<<cpix,epitch) ^~pnRlcMiepUch) p areo * rot 2A1Jy<<cpix,epitch) ^~pnRlcMiepUch) p areo * rot 2A1
CO 的氣動等風機轉子 為尖速比。cU6> . . Ab θ ^ 一係數是風機的重要特徵/以篁功率係數,此 風能與轉子_功率p解歧频提供之理論 響到其係數理論值的高接影 能的能力。 逆阳〜響轉子於裱境中擷取風 電機二力率公式可·此時風力發 矩r所ίί 再透過外加感測器掘取轉子氣動轉 矩7;0<所得出的轉子效率 做有效的監測。但=== 知χ二钇月b里功率係數〜认^僅能透過實驗插值得 算出^理論可求得確切之值,故無法由此一公式直接計 軋功率,另一方面因成本與技術上等現實面考量, 二亦無去針對氣轉矩U做監測,也因此事實上並無 侍知轉子效率實際值。 難,t了避免直接量測氣動轉矩與計算其理論值的困 歸!$知技術均只針對發電機發電效率做監測’其方法可 =為以下幾類:例如建模法、額外的感測器、透過經驗 進行錯誤彳貞測等。圖1是習知發電機電力監控的示意 4 201245997 圖。請參照圖1,習知的電力監控系統Η)是利用_輪箱 η與發電機13轉換後的電力進行監控。 電 葉片㈣―轉速〜等』 由轉子系統11產生氣動轉矩^,經由 給發錢13。評料置14則會根 據^電機u所測量_平均功率魏ρ_,生成發電 效率指標。此情祕驗供發蚊#域,但糾法針對 :子^的效能做監控’在發電效率不如預期的情況之 下,亦無法進—步得知Μ部何種關鍵零件出了問題。 【發明内容】 有鑑於此,本發明提出一種風力發電機的效能評估方 法及糸統,可評估轉子效能。 /本發明提出-種風力發電機的效能評估方法。此方法 係擷取風力發電機的鋒㈣,Μ每 包括風力發電機_子祕的多㈣徵參數。 ^ 述^運轉㈣映射至非線性形態的特徵空間,然後 個單分類模型贿映射躺運轉㈣在特 =最後’在操取到新進運轉資料時,將此二 至特徵空間’並_上述求得的單分類模型計算映 錢職贿㈣分佈的她度储,用為 汗估轉子纟統之贱的效魅標。 巧 ιΐΓ緒出—觀力發賴的效能評m 1包括 貝㈣取單元、資料映射單元、狀態分佈建立單元;= 5 201245997 度度量單元。其中,資料擷取/ 的多筆運轉資料,每—筆運轉係用以擷取風力發電機 系統的多個特徵參數。資料映射二包括風力發電機的轉子 射至非線性形態的特徵空間。壯=兀係用以將運轉資料映 分類模型描述映射後的 1轉° ^分=建立單元係利用單 佈。相似度度量單元細龍”巾的狀態分 與狀態分佈的相似度指標,“作進運轉資料 七:;====狀 —為壤本I明之上述特徵和優點能更明顯易僅,下文特 舉實施例’並配合所附圖式作詳細說明如下。 【實施方式】 為了避免直接量測氣動轉矩與計算其理論值的困 難’本案採用單類別分類的概念,透過監測轉子系統的其 他變數以建立其狀態分佈基準,藉此只需利用現行系統中 可擷取到資料與正常狀態數據來訓練模型,即可在不需要 外加感測器與花費大量時間蒐集異常狀況數據資料的情形 下’有效的建立轉子效能之評斷指標。透過此一評斷指標 就能在整機異常狀況下看出是否是因轉子所造成 ,進而可 6 201245997 推論出其他關鍵元件是否正常。 以風力風電機為例,圖2是依照本發明一實施例所繪 示之轉子效能評估的示意圖。請參照圖2,本實施例的^ =能評估系統20除了將風速仏、葉片仰角^ι、轉子 轉速吟。,專二個變數輪入轉子系統21,由轉子系統2i產生 氣動轉矩,並經由齒輪箱22提供給發電機23外,還 將這三個變數構成一筆運轉資料輸入評估裝置24。接著, 在正常運轉下取得足夠的數據,由評估裝置2 4訓練出發電 機23的正常運轉模型,並建立一轉子效能指標以表示&子 系統21的效能。此模型是以非線性的型態在多維空間中分 布,其建模流程與技術將詳述於後。最後,此轉子效能^ 標便可在風力發電機運轉時用來判別當前的數據 受。如被接受,代表轉子效率正常;如被拒絕康= 效率處於異常狀態,此時,風力發電機也就處於異常狀熊。 亦即,錯誤被偵測出。 ^ 圖3是依照本發明一實施例所繪示之風力發電機的效 能評估裝置的方塊圖。圖4是依照本發明一實施例所繪>示 之風力發電機的效能評估方法的流程圖❶請同時參照圖3 及圖4’本實施例的效能評估裝置3〇例如是配置^力發 電機(未繪示)内或是與風力發電機連接的具運算能力^ 電子裝置,其包括資料擷取單元31、資料映射單元32、狀 態分佈建立單元33及相似度度量單元34。以下即搭配效 能評估衫3G +的各項元件㈣本㈣之風力發電機的 效成评估方法的詳細步驟: 201245997 首先,由資料擷取單元31 轉資料(步驟S402)。其中,—二々風力發電機的多筆運 電機的轉子系統的多個^^徵表:^轉資料包括風力發 統運轉時的風速、葉片角度及轉子^參數則包括轉子系 詳言之’本實關研究的對象 以其控⑽統所記錄之實際運轉資 ^^風料電機, y. - . Λ- _ 貝料為依據’進行轉子效 m :料中有三個特徵值,分別為風速、 ==度以及轉子轉速,其例如是每10分鐘記錄一次。假 =發電機的啟動風速為每秒4公尺(4m/s),關機風速 ‘、、母;25公尺(25m/s) ’風力發電機處於滿餘態時轉速 上限約為每分鐘轉速(Rev〇luti〇n Per Minme,RpM) 16 7, 發電機的最低轉子轉速為12RPM。藉上述條件,可預設標 準以對資料擷取單元31所擷取的資料進行正常狀態資料 選取,以去除不正常或離散的運轉資料,其步驟可分為: (1)將每一筆運轉資料中的風速與一個風速區間比 較,以去除風速在風速區間之外的運轉資料。例如,可選 出風速在風速區間[4,25]m/s的運轉資料,而去除在此風速 區間之外的運轉資料; * (2)將每一筆運轉資料中的轉子轉速與最低轉子轉速 比較,以去除轉子轉速低於最低轉子轉速的運轉資料。例 如,可選出尚於轉子最低轉速12 RPM以上的運轉資料, 而去除其他低於该最低轉子轉速的運轉資料;以及 (3)去除較為離散或曾經有出現警訊記錄的運轉資 料。 8 201245997 31所=心it㈣料映射單元32將資料擷取單元 =二==態的,空間(步驟 實::係將運轉資料映射至非:上本 =r正常運鳴的模型,作為後續二 對於這些映射後的運轉資料,狀態分佈建立單元33 = 莫型來描述映射後的運轉資料在特徵空 S4。6)。詳言之,本實施例係依據 運轉,的特色㈣以支持向量f料插述(s卿。λ vector defnPti〇n,SVDD)模型或核主成份分析(Kernel pr腦Pal component analysis ’ KpCA)模型等單分類 來對轉子基準狀態分佈進行描述,藉此可用崎正常狀能 與異常狀態的資料進行分類,上述兩 ^ 態分佈的資料例如是預先儲存在資料儲存單==大 中’以作為後續評估新進運轉資料是否正常的依據。 一 ^ 34 _上述狀態分佈建立 二ϋ:的單刀類模型,—算新進運轉資料與狀態分 =相,度指標,而用以作為評估轉子㈣之效能的效能 ϋ二=S408) °其中’上述的新進運轉資料例如是由 ==:’並肖料映射單™特 相似度度量單元%所計算的相似度指標可在風力發 201245997 電機運轉時,用來評估當前的轉子效率,並進而判別當前 轉子的狀態是否正常。若此相似度指標落在預設區間内, 即代表轉子效率正常;若非在預設區間内,則代表轉子效 率處於異常狀態’此時效能評估裝置30即可發出警告以通 知相關人員進行處理。 、本發明採用單類別分類的概念,有效建立複雜迴轉機 械内轉子正常運轉的基準狀態分佈,而藉由將新監測到的 運轉^料與此基準狀態分佈進行相似度度量,可達到效能 評斷與故障偵_目的。上述的單分賴型包括支持向量 資料描述(SVDD)模型及核主成份分析(KpCA)模型, 以下即分別舉一實施例詳細說明。 在使用支持向夏資料描述模型作為單分類模型的實 佈建立單元33會求取對於映射後的運轉資 的狀態分佈。詳言之,支二: ';l )之結果可解釋為特徵空間上對於正常資 料具有某種最佳包覆1 、 灿能Μ㈣h 嫌其表面即為絲分類正常 述成下列的最佳化問題, I復了以f田 最小化+ ; 限制於 ~R +ξί> ξ(>〇, V/ --- Jl — VI, — 1 ...]VI ° 其中,’)為資料Xi映射至特徵 及為超球體的半徑,C_權重(Pe讀y 201245997 鬆弛變數(slack variable),Μ為訓練資料的資料筆數。 藉著拉格朗日乘數法(Lagrange multiplier )可得其對偶問 題(Dual problem )如下: /=1 i=ly=l 限制於=卜 i=l 其中 為拉氏函數(Lagrangefunction),尺(χ,,χ))為 一事先定義的核函數,%為拉氏乘數(Lagrange multipliers)。以此實施例而言,核函數足(w)選用高斯 核函數(Gauss kernel function) ’與其參數(parameter) σ的關係如下: ^(xI?Xy) = exp( χ. — Xj / 2(j2) ° 接著’再透過卡羅需-庫恩-塔克條件 (Karush-Kuhn-Tucker Conditions,KKT conditions)與懲 罰權重C可求得超球體半徑及與球心β : r2=k^)~1^a^ ; α ⑷。 其中,Φ〇,·)為第/筆訓練資料映射至特徵空 果。 、。 μ對於新監測之運轉資料文,狀態分佈建立單元^ S3轉資料映射到特徵空間後的特徵向量超 球體的球心β的距離,然後再計算此距離與超球體的半徑 11 201245997 π的比值,以做為相似度指標#幻,其公式如下: d⑻= ||Φ(《) —α||/Λ。 ▲上述的相似度指標除了可作為評估轉子系統效 能的效能指標外’當此相似度指標响超過了預設數值 時’例如峨小還可判定轉子系統的運#出現 提供故障之警示。 需說明的是,對於上述使用支持向量資料描述模型建 立轉子基準狀態分佈的綠,本發明雜供賴性的調整 機制,以求得對於正常狀態資料的最佳包覆。詳言之圖 5是依照本發明—實_所繪示的湘支持向量資料描述 模,建立轉基準狀態*佈的方法流程圖。請參照圖5, 本實施例的方法步驟如下: 首先,設定支持向量資料描述模型的核參數及懲罰權 重(步驟S5G2)。接著’對於映射至特徵空間的運轉資料, 利用所β又疋的核參數計算每兩筆運轉資料的核函數值(步 驟S504)。然後,再根據卡羅需_庫恩-塔克條件及懲罰權 重,利用所計算的核函數值求得超球體的半徑及球心( 驟 S506)。 Ϊ算出超球體的半徑及球心之後,即利用此超球體計 异運算資料落在超球體外部的比例(步驟S5〇8),並判斷 此比例是否大於預設比例(步驟S510),據以判別是否此 超球體的大小符合需求。詳言之,狀態分佈建立單元33 例如會計算映射至特徵m的運轉資料落在超球體之外的 比例並與預a又目標拒絕比例(Target)比較, 12 201245997 此預設比例例如是1%、3%或5%。其中,若所計算的比例 大於預設比例,則狀態分佈建立單元33會回到步驟S502, 重新設定核參數及懲罰權重,而重新計算超球體的半徑及 球心;反之,若所計算的比例不大於預設比例’則狀態分 佈建立單元33即可利用所求得的超球體的半徑及球心,計 算映射後的新進運轉資料與狀態分佈的相似度指標(步驟 S512),藉以做為評估轉子系統效能的依據。 另一方面,在使用核主成份分析(KPCA)模型的實 施例中,狀態分佈建立單元33會利用核主成份分析模型求 取對於映射後的運轉資料滿足一資料散佈最大原則 (maximal amount of variance)的多維子空間,以描述狀 態分佈。詳言之,核主成份分析為非線性的降維分析技巧, 透過重建錯誤的計算,藉以達成單分類的功效。 基本士,主成份分析可以看成解矩陣特徵值的問題, 例如,其結果為特徵空間上中的—個q維本徵子 空間(dgenspacO。此子㈣的找尋必 大原則,也就是找尋特徵空間中共 二 ,度。根據推導,特徵向量; 資料Φ(χ)的線性組合: 勹”J映射後運轉 y=Za^(x.) 〇 13 201245997 運轉下的運轉 可定義下列的 其中,Χί為訓練資料(即轉子系統正常 資料)、《,為權重值。針對映射後資料岭), 核矩陣(kernel matrix )尤. ^=(Φω·Φ(χ7.)) = 6χρφ._χ|/2σ2)〇 藉此原問題可轉換成另一個等價的特徵值問題: Μλ〇ί = Κα 〇 作其中’上述核矩陣尤解出來的特徵们由大到小排列 可得),其所對應到的特徵向量分別為 aW3,··./。對於新監測之運算資料文,可藉由此運轉 資料;e在維度g之多維子空間上的重建錯誤值柳,以做為 該相似度指標,此重建錯誤值〆幻的定義如下: /?(幻=凡(幻—ΛΟ2。 其中 上述的相似度指#抑除了可作為評估轉子系統效 能的效能指標外,當此相似度指標々(句超過了閥值時,例 如⑽)>;i,還可判定轉子系統的運算出現異f,進而提供 故障之警示。 " 需說明的是,對於上述使用核主成份分析模型建立轉 201245997 子基準_分佈㈣法’本發㈣提彳 ,,以求得對於正常狀態資料的最佳描述。詳言:。周二機 是依照本發明—實施例麟㈣核主成份分析模j 立轉子基準狀,¾、分佈的方法流程圖。請參關太 例的方法步驟如下: 本貫知 首先,設定核主成份分析模型的核參數及多維子空 的維度(步驟_2)。接著,對於映射至特徵空間的=轉 資料’所設定的核參數計算每兩筆運轉賴的核函數 值,並聚集成為核矩陣(步驟S604)。然後,解出此核矩 陣的多個特徵值及各個特徵值對應的特徵向量 S606)。 卜 在解出核矩陣的特徵值之後,接著則計算映射至特徵 空間的各個運轉資料在該維度之多維子空間上的重建錯誤 值(步驟S608 )。根據預設目標拒絕比例(Target咏⑼⑽ rate)(例如是1%、3%或5%)與重建錯誤值計算出對應 閥值,根據此一閥值計算測試資料的錯誤拒絕率(False Rejection Rate ’ FRR)與錯誤接受率(False AcceptanceThe rotor of the CO such as aerodynamics is a tip speed ratio. cU6> . . Ab θ ^ A coefficient is an important feature of the fan / the power factor of the ,, the wind energy and the rotor _ power p decomposed frequency provide the theoretical ability to respond to the high theoretical power of the coefficient. The reverse yang ~ ring rotor draws the wind power two-force rate formula in the dilemma. At this time, the wind power is r ίί and then the rotor aerodynamic torque is extracted through the external sensor 7; 0; the resulting rotor efficiency is effective Monitoring. However, === knowing the power factor of the second month b, the recognition can only be calculated through the experimental interpolation. The theory can be used to determine the exact value, so the power cannot be directly calculated by this formula. On the other hand, the cost and technology Considering the real-life considerations, the second does not have to monitor the gas torque U, so there is no actual value of the rotor efficiency. Difficult, t avoid the difficulty of directly measuring the aerodynamic torque and calculating its theoretical value! The knowing technology only monitors the generator power generation efficiency. The method can be as follows: modeling methods, additional sensors, error guessing through experience, etc. Figure 1 is a schematic diagram of a conventional generator power monitor 4 201245997. Referring to Fig. 1, the conventional power monitoring system (监控) is monitored by using the power converted by the _ wheel box η and the generator 13. The electric blade (four) - the rotational speed ~, etc., generates a pneumatic torque ^ from the rotor system 11, and sends money 13 through. The evaluation set 14 will generate the power generation efficiency index based on the average power Wei ρ_ measured by ^ motor u. This secret test for the mosquitoes # domain, but the correction method for: the performance of the child ^ to monitor 'in the case of power generation efficiency is not as expected, can not enter - know what key parts of the crotch problems. SUMMARY OF THE INVENTION In view of the above, the present invention provides a method and system for evaluating the performance of a wind power generator, which can evaluate rotor performance. / The present invention proposes a method for evaluating the effectiveness of a wind power generator. This method draws the wind turbine's front (four), and each of the wind turbines includes multiple (four) signs. ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ The single-category model calculates the distribution of the money bribe (four) distribution, and uses it as a measure of the effectiveness of the rotor.巧 ΐΓ ΐΓ — —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— —— Among them, the data capture / multiple operational data, each pen operation is used to capture a number of characteristic parameters of the wind turbine system. The data map 2 includes the feature space of the wind turbine rotor to the nonlinear shape. Zhuang = 兀 is used to map the operational data into a classification model description of 1 turn ° ^ ^ = the establishment of the unit system using a single cloth. Similarity measure of the state of the similarity measure unit and the state distribution of the thin dragon towel, "for the operation data seven:; ==== shape - the above characteristics and advantages of the soil I can be more obvious and easy, only the following The embodiment will be described in detail with reference to the accompanying drawings. [Embodiment] In order to avoid the difficulty of directly measuring the aerodynamic torque and calculating its theoretical value, the concept of single-category classification is adopted in this case. By monitoring other variables of the rotor system to establish a state distribution reference, it is only necessary to utilize the current system. The data can be retrieved by training the data with the normal state data, so that the evaluation index of the rotor performance can be effectively established without the need for an external sensor and a large amount of time to collect abnormal condition data. Through this judgment indicator, it can be seen whether the rotor is caused by the abnormal condition of the whole machine, and then it can be inferred that other key components are normal. Taking a wind power motor as an example, FIG. 2 is a schematic diagram of rotor performance evaluation according to an embodiment of the invention. Referring to Fig. 2, the ^= energy evaluation system 20 of the present embodiment has a wind speed 仏, a blade elevation angle, and a rotor speed 吟. The two variables are wheeled into the rotor system 21, the aerodynamic torque is generated by the rotor system 2i, and supplied to the generator 23 via the gearbox 22, and these three variables constitute an operational data input evaluation device 24. Next, sufficient data is obtained under normal operation, the normal operation model of the generator 23 is trained by the evaluation device 24, and a rotor performance index is established to indicate the performance of the & subsystem 21. This model is distributed in a multidimensional space in a nonlinear form, and its modeling process and techniques will be detailed later. Finally, this rotor performance can be used to determine the current data reception while the wind turbine is running. If accepted, it means that the rotor efficiency is normal; if it is rejected, the efficiency is in an abnormal state, at this time, the wind turbine is also in an abnormal shape. That is, the error is detected. FIG. 3 is a block diagram of an apparatus for evaluating the performance of a wind power generator according to an embodiment of the invention. 4 is a flow chart showing a method for evaluating the performance of a wind power generator according to an embodiment of the present invention. Referring to FIG. 3 and FIG. 4 together, the performance evaluation device 3 of the present embodiment is configured, for example, by The computer (not shown) or the computing device connected to the wind power generator, the electronic device, includes a data capturing unit 31, a data mapping unit 32, a state distribution establishing unit 33, and a similarity measuring unit 34. The following is the detailed steps of the evaluation method of the wind turbine generator of the (3) part (4) of the performance evaluation shirt 3G +: 201245997 First, the data extraction unit 31 transfers the data (step S402). Among them, the multiple rotors of the multi-motor motor of the Eryi wind turbine have the following conditions: The wind data, the blade angle and the rotor parameters of the wind turbine system include the details of the rotor system. The object of this actual research is based on the actual operation of the control (10) system, ^^ wind power motor, y. - . Λ- _ shell material based on 'rotation efficiency m: material has three characteristic values, respectively, wind speed , == degrees and rotor speed, which is recorded, for example, every 10 minutes. False = the starting wind speed of the generator is 4 meters (4m / s) per second, the wind speed of the shutdown ', mother; 25 meters (25m / s) 'The upper limit of the speed of the wind turbine is about every minute when the wind turbine is in the full state (Rev〇luti〇n Per Minme, RpM) 16 7. The minimum rotor speed of the generator is 12 RPM. By the above conditions, the standard can be preset to select the normal state data of the data captured by the data acquisition unit 31 to remove abnormal or discrete operational data, and the steps can be divided into: (1) Each operation data is divided into: The wind speed in the middle is compared with a wind speed interval to remove the operating data of the wind speed outside the wind speed interval. For example, the operating data of the wind speed in the wind speed range [4, 25] m/s can be selected, and the operating data outside the wind speed interval can be removed; * (2) Comparing the rotor speed in each running data with the lowest rotor speed To remove the operating data of the rotor speed below the minimum rotor speed. For example, operating data that is still above the minimum rotor speed of 12 RPM can be selected to remove other operating data below the minimum rotor speed; and (3) to remove operating data that is more discrete or has a warning record. 8 201245997 31 = heart it (four) material mapping unit 32 will be the data extraction unit = two = = state, space (steps are:: the system will map the operational data to non: the current = r normal operation model, as the follow-up two For these mapped operational data, the state distribution establishing unit 33 = Mo type to describe the mapped operational data in the feature space S4. 6). In detail, this embodiment is based on the characteristics of operation, (4) with support vector f material insertion (sqing. λ vector defnPti〇n, SVDD) model or kernel principal component analysis (Kernel pr brain Pal component analysis 'KpCA) model The single-class classification is used to describe the rotor reference state distribution, so that the data of the normal state and the abnormal state can be classified. For example, the data of the above two states are stored in advance in the data storage list ==大中' as a follow-up. The basis for assessing whether new operational data is normal. A ^ 34 _ the above state distribution establishes a two-knife model: the calculation of the new operational data and the state of the sub-phase, the degree index, and is used as a measure of the effectiveness of the rotor (four) performance ϋ = S408) ° where The new operational data is, for example, the similarity index calculated by the ==:' and the singular mapping unit TM similarity metric unit % can be used to evaluate the current rotor efficiency when the wind power 201245997 motor is running, and then determine the current Whether the state of the rotor is normal. If the similarity index falls within the preset interval, it means that the rotor efficiency is normal; if it is not within the preset interval, it means that the rotor efficiency is in an abnormal state'. At this time, the performance evaluation device 30 can issue a warning to notify the relevant personnel for processing. The invention adopts the concept of single category classification, effectively establishes the reference state distribution of the normal operation of the inner rotor of the complex rotary machine, and achieves the performance judgment by measuring the similarity between the newly monitored operation material and the reference state distribution. Fault detection_purpose. The above-mentioned single-distribution type includes a support vector data description (SVDD) model and a kernel principal component analysis (KpCA) model, which are respectively described in detail below. The actual distribution establishing unit 33 using the support summer data description model as the single classification model obtains a state distribution for the mapped operational assets. In detail, the result of branch 2: ';l) can be explained as the optimization of the normal data for the normal data in the feature space 1 , Cannon Μ (4) h, the surface is the normal classification of the silk classification , I is restored to minimize the field +; is limited to ~R +ξί>ξ(>〇, V/ --- Jl — VI, — 1 ...] VI ° where, ') is mapped to the data Xi The feature is the radius of the hypersphere, C_weight (Pe reads y 201245997 slack variable, Μ is the number of data of the training data. The dual problem can be obtained by Lagrange multiplier (Dual problem ) is as follows: /=1 i=ly=l is limited to =Bu i=l where is Lagrange function, ruler (χ,,χ)) is a predefined kernel function, % is Lagrangian Multiplier (Lagrange multipliers). In this embodiment, the kernel function foot (w) uses the Gauss kernel function 'the relationship with its parameter σ as follows: ^(xI?Xy) = exp( χ. — Xj / 2(j2 ° ° Then, through the Karush-Kuhn-Tucker Conditions (KKT conditions) and the penalty weight C, the radius of the supersphere and the center of the sphere β: r2=k^)~1 can be obtained. ^a^ ; α (4). Among them, Φ〇,·) is mapped to the feature empty material for the pen/train data. ,. μ For the newly monitored operational data, the state distribution establishing unit ^ S3 transfers the distance of the data to the spherical center β of the feature vector hypersphere after the feature space, and then calculates the ratio of the distance to the radius of the hypersphere 11 201245997 π, As the similarity index #幻, its formula is as follows: d(8)= ||Φ(") -α||/Λ. ▲The above similarity index can be used as a performance indicator to evaluate the effectiveness of the rotor system. When the similarity indicator exceeds the preset value, for example, the rotor system can be determined to provide a warning of failure. It should be noted that, for the above-mentioned green using the support vector data description model to establish the rotor reference state distribution, the adjustment mechanism of the hybrid supply of the present invention is obtained to obtain the optimal coating for the normal state data. In detail, FIG. 5 is a flow chart of a method for establishing a transition reference state* cloth according to the description of the Hunan support vector data description module according to the present invention. Referring to FIG. 5, the method steps of this embodiment are as follows: First, the kernel parameters of the support vector data description model and the penalty weights are set (step S5G2). Next, for the operation data mapped to the feature space, the kernel function value of each of the two pieces of operation data is calculated using the kernel parameter of β and 疋 (step S504). Then, according to the Carol-Tucker condition and the penalty weight, the radius of the supersphere and the center of the sphere are obtained by using the calculated kernel function value (S506). After calculating the radius of the hypersphere and the center of the sphere, the ratio of the hypersphere calculation data to the outside of the supersphere is used (step S5〇8), and it is determined whether the ratio is greater than a preset ratio (step S510). Determine whether the size of this hypersphere meets the demand. In detail, the state distribution establishing unit 33 calculates, for example, the ratio of the operation data mapped to the feature m to the outside of the hypersphere and compares it with the pre-a target rejection ratio (Target), 12 201245997, the preset ratio is, for example, 1%. , 3% or 5%. If the calculated ratio is greater than the preset ratio, the state distribution establishing unit 33 returns to step S502, resets the kernel parameter and the penalty weight, and recalculates the radius and the center of the supersphere; otherwise, if the calculated ratio The state distribution establishing unit 33 can calculate the similarity index of the mapped new running data and the state distribution by using the obtained radius and the center of the supersphere (step S512), thereby using the evaluation. The basis for the effectiveness of the rotor system. On the other hand, in the embodiment using the kernel principal component analysis (KPCA) model, the state distribution establishing unit 33 uses the kernel principal component analysis model to obtain a maximum amount of variance for the mapped operational data. Multidimensional subspace to describe the state distribution. In detail, the kernel principal component analysis is a nonlinear dimensionality reduction analysis technique, which is used to reconstruct the erroneous calculations to achieve the effect of single classification. The basic component analysis can be regarded as the problem of solving the eigenvalues of the matrix. For example, the result is a q-dimensional eigenspace space in the feature space (dgenspacO. The principle of finding the sub-fourth is to find the feature. A total of two degrees in space. According to the derivation, the eigenvector; the linear combination of the data Φ(χ): 勹"J mapping operation y=Za^(x.) 〇13 201245997 The operation under operation can define the following, Χί Training data (ie, rotor system normal data), ", is the weight value. For the mapped data ridge", the kernel matrix (kernel matrix) is especially. ^=(Φω·Φ(χ7.)) = 6χρφ._χ|/2σ2) 〇The original problem can be converted into another equivalent eigenvalue problem: Μλ〇ί = Κα 〇 其中 ' ' ' ' ' ' ' ' ' 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤 尤The vectors are respectively aW3,··./. For the newly monitored operation data, the data can be manipulated by this; e. The reconstruction error value on the multidimensional subspace of dimension g is used as the similarity index, and this reconstruction is performed. The definition of the error value is as follows: /? (幻幻凡凡(幻- ΛΟ 2. The above similarity refers to # 抑 can be used as a performance index to evaluate the effectiveness of the rotor system, when the similarity index 々 (sentence exceeds the threshold, for example (10)) >; i, can also determine the rotor system The operation of the operation f is different, and thus provides a warning of failure. " It should be noted that for the above-mentioned use of the nuclear principal component analysis model to establish the transfer of 201245997 sub-base _ distribution (four) method 'this hair (four), in order to obtain a normal state The best description of the data.Details: The Tuesday machine is a flow chart of the method according to the present invention - the embodiment of the core (four) nuclear principal component analysis module, the vertical rotor reference, 3⁄4, distribution method. Firstly, the kernel parameters of the kernel principal component analysis model and the dimensions of the multidimensional subspace are set (step_2). Then, for the kernel parameters set to the =transfer data of the feature space, every two operations are calculated. The kernel function values are aggregated into a kernel matrix (step S604). Then, a plurality of eigenvalues of the kernel matrix and eigenvectors corresponding to the respective eigenvalues S606) are solved. And then calculating a reconstruction error value of each operation data mapped to the feature space on the multi-dimensional subspace of the dimension (step S608). According to the preset target rejection ratio (Target咏(9)(10) rate) (for example, 1%, 3% or 5%) Calculate the corresponding threshold with the reconstructed error value, and calculate the False Rejection Rate 'FRR and False Acceptance of the test data based on this threshold.
Rate ’ FAR)的比例(步驟S610),並與一預設比例值比 較,以判斷此比例是否大於預設比例值(步驟S612)。其 中,若所計算的比例大於該預設比例值,即回到步驟 S602,重新設定核參數及多維子空間的維度,並重新計算 核主成份分析模型的特徵值及各個特徵值對應的特徵向 量;反之,若所計算的比例不大於預設比例值,則利用所 求得的特徵值及各個特徵值對應的特徵向量,計算映射後 15 201245997 的新進運轉資料與狀態分佈的相似度指標(步驟s6i4)。 综上所述,本發明之風力發電機的效能評估方法及裝 置透過環境與狀態資料建立轉子狀態之基準狀態分佈,^ 由與基準分佈之相似度度量達到效能評斷與故障偵測。其 中二透過相似度度量轉化之可量化的效能指標,可指明轉 子系統的健康程度,降低監測成本,並簡化傳統上異常偵 測所需另外佈建感測器之流程。 ' 雖然本發明已以實施例揭露如上,然其並非用以限定 本發明’任何所屬技術領域中具有通常知識者,在不脫離 本發明之精神和範圍内,當可作些許之更動與潤飾,故本 發明之保護範圍當視後附之申請專利範圍所界定者為準。 【圖式簡單說明】 圖1是習知發電機電力監控的示意圖。 圖2是依照本發明一實施例所繪示之轉子效能評估的 示意圖。 圖3是依照本發明一實施例所繪示之風力發電機的六文 能評估裝置的方塊圖。 、> 圖4是依照本發明一實施例所緣示之風力發電機的六文 能評估方法的流程圖。 > 圖5是依照本發明一實施例所繪示的利用支持向量資 料描述模型建立轉子基準狀態分佈的方法流程圖。 圖6是依照本發明一實施例所繪示的利用核主成份八 析模型建立轉子基準狀態分佈的方法流程圖。 刀 16 201245997 【主要元件符號說明】 ίο:電力監控系統 11、 21 :轉子系統 12、 22 :齒輪箱 13、 23 :發電機 14、 24 :評估裝置 20 :效能評估系統 30 :效能評估裝置 31 :資料擷取單元 32 :資料映射單元 33 :狀態分佈建立單元 34 :相似度度量單元 S402〜S408 :本發明一實施例之風力發電機的效能評 估方法的步驟 S502〜S512 :本發明一實施例之利用支持向量資料描 述模型建立轉子基準狀態分佈的方法步驟 S602〜S614 :本發明一實施例之利用核主成份分析模 型建立轉子基準狀態分佈的方法步驟 17The ratio of Rate ' FAR ) (step S610) is compared with a preset ratio value to determine whether the ratio is greater than a preset scale value (step S612). If the calculated ratio is greater than the preset ratio value, return to step S602, reset the kernel parameter and the dimension of the multi-dimensional subspace, and recalculate the feature value of the kernel principal component analysis model and the feature vector corresponding to each feature value. On the other hand, if the calculated ratio is not greater than the preset scale value, the similarity index of the new running data and the state distribution of the 15 201245997 after the mapping is calculated by using the obtained feature value and the feature vector corresponding to each feature value (steps) S6i4). In summary, the method and device for evaluating the effectiveness of the wind power generator of the present invention establishes a reference state distribution of the rotor state through the environment and state data, and achieves performance judgment and fault detection by the similarity measure with the reference distribution. Among them, the quantifiable performance indicators transformed by the similarity measure can indicate the health of the transfer subsystem, reduce the monitoring cost, and simplify the process of separately installing the sensor required for the traditional abnormal detection. Although the present invention has been disclosed in the above embodiments, it is not intended to limit the invention to those skilled in the art, and may be modified and modified without departing from the spirit and scope of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a schematic diagram of conventional generator power monitoring. 2 is a schematic diagram of rotor performance evaluation according to an embodiment of the invention. 3 is a block diagram of a six-function evaluation device for a wind power generator according to an embodiment of the invention. Fig. 4 is a flow chart showing a six-function evaluation method of a wind power generator according to an embodiment of the present invention. > Figure 5 is a flow chart showing a method for establishing a rotor reference state distribution using a support vector data description model, in accordance with an embodiment of the invention. FIG. 6 is a flow chart of a method for establishing a rotor reference state distribution using a kernel principal component analysis model according to an embodiment of the invention. Knife 16 201245997 [Description of main component symbols] ίο: Power monitoring system 11, 21: Rotor system 12, 22: Gearbox 13, 23: Generator 14, 24: Evaluation device 20: Effectiveness evaluation system 30: Effectiveness evaluation device 31: The data extracting unit 32: the data mapping unit 33: the state distribution establishing unit 34: the similarity measuring unit S402 to S408: steps S502 to S512 of the performance evaluation method of the wind power generator according to an embodiment of the present invention: an embodiment of the present invention Method for establishing a rotor reference state distribution using a support vector data description model Steps S602 to S614: Method for establishing a rotor reference state distribution using a kernel principal component analysis model according to an embodiment of the present invention