TWI353076B - Method for estimating residual capacity of lead-ac - Google Patents
Method for estimating residual capacity of lead-ac Download PDFInfo
- Publication number
- TWI353076B TWI353076B TW096144353A TW96144353A TWI353076B TW I353076 B TWI353076 B TW I353076B TW 096144353 A TW096144353 A TW 096144353A TW 96144353 A TW96144353 A TW 96144353A TW I353076 B TWI353076 B TW I353076B
- Authority
- TW
- Taiwan
- Prior art keywords
- acid battery
- value
- cluster
- lead
- residual
- Prior art date
Links
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E60/00—Enabling technologies; Technologies with a potential or indirect contribution to GHG emissions mitigation
- Y02E60/10—Energy storage using batteries
Landscapes
- Charge And Discharge Circuits For Batteries Or The Like (AREA)
- Secondary Cells (AREA)
Description
1353076 九、發明說明: 【發明所屬之技術領域】 本發明是有關於一種鉛酸電池殘餘電量之估測方 法,且特別是有關於一種利用可拓類神經網路(Extension neural network, ENN)的估測技術來估測鉛酸電池殘餘電 量的方法。1353076 IX. Description of the Invention: [Technical Field] The present invention relates to a method for estimating residual power of a lead-acid battery, and more particularly to an extension neural network (ENN) Estimate techniques to estimate the residual charge of lead-acid batteries.
【先前技術】 電化學電池在電力系統中一直扮演著非常重要的角 色,其中鉛酸電池和充電器更由於可攜式通訊裝置、不斷 電電源系統以及電動車等產品之大量開發使用,使得其使 用量日益增加。船酸電池最大的缺點則在於很難去偵測電 池殘餘電量狀態。目前所發展出來之估測錯酸電池殘餘電 量的技術,主要的方法有下列幾種:[Prior Art] Electrochemical cells have always played a very important role in power systems. Lead-acid batteries and chargers have been developed and used in large quantities such as portable communication devices, uninterruptible power systems, and electric vehicles. Its use is increasing. The biggest disadvantage of the ship acid battery is that it is difficult to detect the state of the battery residual energy. At present, the main methods for estimating the residual power of the wrong acid battery are as follows:
(1 )開路電壓法(Open circuit voltage )、( 2 )電解液比重 法(Electrolyte specific gravity )、( 3 )負載電壓法([⑽糾 V〇ltage)、(4)内阻測定法(Internalresistance)以及(5) 庫倫汁法(Coulometnc measurement)等。這些電池殘餘 電量的估測方法皆有著因電池本身特性、或者是負載的特 性所引起的估測誤差。 马使胃審查委員便於了解本發明的技術特徵及功 效,兹對鉛酸電池的相關工作原理、特性以及可拓 原理先敘述如下: ’的 (一)錯酸電池的工作原理與特性 5 1353076(1) Open circuit voltage, (2) Electrolyte specific gravity, (3) Load voltage method ([(10) correction V〇ltage), (4) Internal resistance measurement (Internalresistance) And (5) Coulometnc measurement, etc. Estimates of the residual power of these batteries have estimation errors due to the characteristics of the battery itself or the characteristics of the load. The Ma Stomach Review Committee is easy to understand the technical features and effects of the present invention. The working principle, characteristics and extension principle of the lead-acid battery are described as follows: ─ (1) Working principle and characteristics of the wrong acid battery 5 1353076
錯酸電池基本上可分為兩大類:一為浸在水裏的鉛酸 電池(Flooded lead-acid battery );另一為密封的錯酸電池 (Sealed lead-acid battery)。其分別在於前者充電時會不 斷產生氣體並直接散出電池外,所以每隔一段時間便必須 補充水分。而後者在充電時,其正極的充電效率並不那麼 大,因此在充電較早階段所產生的氧氣,被利用於使充電 初期不會產生氫氣。亦即’在充電過程由正極所產生的氧 氣被負極所吸收,而直接與周圍的電解液自然反應,使負 極回復成硫酸錯,即為原來的放電生成物。 (二)錯酸電池的充放電反應 錯酸電池的充放電反應主要發生在極板處,正極板為 硫酸鉛的氧化反應,負極板為硫酸鉛的還原反應,其反應 方私式如式(1)所示。The wrong acid batteries can be basically divided into two categories: one is a flooded lead-acid battery immersed in water; the other is a sealed lead-acid battery. The difference is that when the former is charged, gas is continuously generated and directly discharged outside the battery, so it is necessary to replenish moisture at intervals. On the other hand, when the latter is charged, the charging efficiency of the positive electrode is not so large, so the oxygen generated at the early stage of charging is utilized so that hydrogen gas is not generated at the initial stage of charging. That is, the oxygen generated by the positive electrode during the charging process is absorbed by the negative electrode, and directly reacts with the surrounding electrolyte to return the negative electrode to sulfuric acid, which is the original discharge product. (II) Charge and discharge reaction of the wrong acid battery The charge and discharge reaction of the wrong acid battery mainly occurs at the plate, the positive plate is the oxidation reaction of lead sulfate, and the negative plate is the reduction reaction of lead sulfate. 1) shown.
Pb02 + 2H2S04 +Pb^ IPbSO, + 2H20 ( i )Pb02 + 2H2S04 +Pb^ IPbSO, + 2H20 ( i )
式(1)的左邊為正極反應,而右邊則為負極反應, 往右反應為放電反應,而往左邊則為充電反應。鉛酸電池 在放電過程中’化學反應會於正極(凡02)與負極(外) 極板上產生硫酸船(,使得極板附近電解液濃度下 降’電池電壓也隨之下降。 若在此時停止負載作用,會使得其他區域的電解液因 為擴散作用,導致極板附近之電解液上升,電池電壓也隨 之上升’約在2小時左右’電池電解液濃度擴散趨於均勻, 電池電壓上升速度漸平緩’當電池電壓達到穩態電壓值 時’這時所量測之電壓值即電池内部之電動勢。 1353076 (二)錯酸電池的等效電路 凊參照第1圖,其繪示為常用鉛酸電池之等效電路 圖°圖中電池電極與電池電解液之等效電阻12(兄,),電 池電極與電池電解液之介面電阻14 ( D,電池内部電容 ^( c ),其為電極之活性物質與電解液介面間的電氣容 里,而放電電流20 ( I )為電池内部流出之電流。 由第1圖之等效電路中,得到電池端電壓為式 (2 )所示如下: ~ Kc ~ (-^/„1 + Rinl)I + RinlIe CR,n' - Vce (2) CRin2 在穩態下,式(2)的電池端電壓方程式可簡化成式 (3 )所示如下: vb = Kc~I{Rin^Rin2) ( 3) 其中,内電阻(4 )可表示如式(4)所示如下: - + Rilt2) = (Voc _ Vb)/1=AV/1 ( 4 ) (四)可拓概論 物το可拓理論(Extensi〇n the〇ry)為大陸學者蔡文教 授於1983年所提出,其學說概念為經由研究事物的可拓 性及可拓之定則與方法,來解決矛盾問題,並且將其量化 為計算機可以處理之訊息。物元理論和可拓數學為可拓學 的兩大基礎,其中物元理論描述物元具有可拓性和變換= 之特性,而可拓數學概念則是以可拓集合及關聯函數作 演算核心》 可拓理論透過物元模型表示事物訊息,並經由物元變 換來表達事物「質」與「量」之轉換關係。因此,可由「量」、 1353076 質」與問題之影響關聯程度,明確顯示事物特徵之影響 \ 性。此外,可拓理論將古典數學由二位值邏輯延伸至連續 多值(continuous muhi-vaiue ),並利用關聯函數 . (Correlation functi〇n)表示事物之特質,亦即應用 ' 巾之實數描述—個元素隸屬於某—事物特性之程度,而此 值可稱為元素對於事物特性所屬集合之關聯度 (Membership grade) ° • 若正規化後之關聯度為1時,則表示此元素完全符合 該事物特性。若正規化後之關聯度為]時,表示元素^ . 於此事物特性。關聯度介於<-u>之間則是元素完全屬 於或是完全不屬於事物特性之強弱程度。 【發明内容】 本發明的目的是在提供—種斜酸電池殘餘電量之估 測方法’用以改善現有㈣電池殘餘電量的估測方法中, 其準確度不佳或估測速度慢等問題。 根據本發明之-種紹酸電池殘餘電量之估測方法,利 用類神經網路所具有之平行處理運算架構與學習能力,以 :可拓理論藉由關聯函數值之計算以進行分類處理之特 :’用以估測錯酸電池的殘餘電量。因此,可以 评估鉛酸電池的殘餘電量,且呈 、迷 介Η亜七丨埜丄I /、有冰'東時間短及對記憶體 =求小等功效’進而提昇可攜性裝置的使用穩定性。 依照本發明一較佳實施例’鉛酸電池殘餘電量 方法係透過可括理执杰4 ㈣先订建立鉛酸電池各種殘餘電量的 δ < 5 切 3076 可相物元模型,建立可拓物元模型之方式係利用所量 之斜酸電池的相關參數,進行錯酸電池的各種殘餘電 =類,並建立其對應每—殘餘電量等級的可拓物元模 根據此可拓物元模型建立一可抬類神經網路,其 ;類神經網路包含-輸入層以及-輸出層,輸入層包含福 —個輪入層節點’輸出層包含複數個物元,每— 二:出層節點’每一輸入層節點及每一輸出層節點間有: 分^结權重值,每-輸出層節點對應其中一殘餘電量等級 r接著進行可拓類神經網路之訓練程序。訓練可相 ::::輸二:訓練物元特徵樣本對其二個連結權重值 錄酸電池殘餘電量的估測,其輸 :進仃 點=_τ辨識後,僅會有其中之-輸出層節 等級j 輸出層節點所代表對應之殘餘電量 級,藉以判斷此待測資料屬於哪—個殘餘電量等級。 【實施方式】 種照第2圖,其繪示依照本發明一較佳實施例的一 =電池殘餘電量之估測方法的流_1 電里之估測方法,包含下列步驟: 戈餘 拓物透Γ拓料先行建立㉝㈣池各種殘餘電量的可 拓物—如步請所示。建立一配合此可拓物= 丄功〇76 =之可抬類神經網路架構,如步驟2 =經網路之訓練程序,以已知殘餘電量百分比:;了類 模型作為可相麵之可拓物元 230所-㈣的輪人樣本進行訓練,如^ 不。採行已訓練完成之可拓類 二 酸電池殘餘電量的估測,如步驟24。所示架構進仃錯 *法例…酸電池殘餘電量之估測 ⑴建立殘餘電量之可拓物元模型(步驟21〇) 若要準確的估測電池的殘餘 酸電池殘餘電量之·^ 、 先必須先建立鉛 戈餘電里之可拓物元模型(即 狀f樣本)。而在建立錯酸電池殘餘電量之可之 之前,需先取得錯酸電池之相關參數。 70、型 在本實施例令,以電子式負載對鉛酸 SHYKUANG BP12-12)作1A定電漭放 、尘唬 電阻測試儀器(TEST3 ) 驗’並使用内 變化每—分鐘錯酸電池内電阻之 邊化罝记錄下來,並於每一拄 〜 懕夕戀^ ㈣小時將負載移開以記錄開路電 ^魏值,所得之測試曲線分別如第3a圖及第3b圖所 由第h圖之鉛酸電池電虔在放電實驗下的變化曲 線’可看出其開路電壓值3〇1與其電池放電程度(殘餘電 量)是成一近似線性的關係’所以電池的開路電壓可作為 殘餘電量估測的參考。而第3b圖為叙酸電池放電期間内 電阻302之變化曲線。 為了使估測率更加準確’並將同—時間下之開路電塵 1353076 與内電阻值作相除運算得— 電量之-特η “亦將其當成殘餘 電池的開路電麼:雷在本實施例中,根據所量測之錯酸 阻)之相關來數及;路電流(開路電壓除以内電 型(亦即建立各種殘餘電= =殘餘電量等級分為十類,則可將此十種殘餘電量 類別及付號以第1表加以表示。 互殘餘電量為90% 殘餘電量為70〇/» 殘餘電量為50\ 玉21殘餘電量為30〇乂 量ίο種分類符號〇 殘餘電量芩8〇〇/〇 殘餘電量為4〇0/< -^λ_ΑΆΆ*_^6〇% ^ ----~~^1:__π 町、电里府 υ% 根據所量測之錯酸電池參數,將十種殘餘電量的可拓物元 模型以第2表予以表示。1中, 拓物兀 ” τ事物(R)代表鉛酸電池 十種殘餘電量分類之物元,尺=丨 "{尺1,尺2,...,尤。}為殘餘電量分 類集,每-物元内使用三個特徵元,分別為内電阻、 ,路電壓(〔)及短路電流((//u。各物元模型之^域 範圍為該特徵物元所對應之經典域4本實施例中,以各 殘餘電量分類之内電阻、開路電壓及短路電流分別所產生 的最大值及最小值為值域範圍之依據。 由於各分類物元之經典域是以電池之内電阻值、開路 電壓與短路電流的上下限值為輸入變數,因此鉛酸電池在 不同殘餘電量範圍内所產生之内電阻值、開路電磨與短路 11 1353076 電流的最大值及最小值,即為各分類物元之經典域的值 域。 第2表錯酸電池殘餘電量狀態之可祐物元模型。The left side of formula (1) is the positive electrode reaction, while the right side is the negative electrode reaction, the right side is the discharge reaction, and the left side is the charging reaction. During the discharge process of the lead-acid battery, the chemical reaction will produce a sulfuric acid boat on the positive electrode (of the 02) and the negative (outer) plate (so that the electrolyte concentration near the plate decreases) and the battery voltage also decreases. Stopping the load will cause the electrolyte in other areas to diffuse, causing the electrolyte near the plate to rise, and the battery voltage will rise accordingly. [About 2 hours or so.] The battery electrolyte concentration tends to be uniform, and the battery voltage rises. Gradually 'when the battery voltage reaches the steady-state voltage value', the measured voltage value is the electromotive force inside the battery. 1353076 (2) The equivalent circuit of the wrong acid battery 凊 Refer to Figure 1, which is shown as common lead acid. The equivalent circuit diagram of the battery in the figure is the equivalent resistance of the battery electrode and the battery electrolyte 12 (brother,), the interface resistance of the battery electrode and the battery electrolyte 14 (D, the internal capacitance of the battery ^ ( c ), which is the activity of the electrode The electrical capacity between the substance and the electrolyte interface, and the discharge current 20 (I) is the current flowing out of the battery. From the equivalent circuit of Figure 1, the battery terminal voltage is obtained as equation (2) The following is shown: ~ Kc ~ (-^/„1 + Rinl)I + RinlIe CR,n' - Vce (2) CRin2 In steady state, the battery terminal voltage equation of equation (2) can be simplified into equation (3) The following is shown: vb = Kc~I{Rin^Rin2) (3) where the internal resistance (4) can be expressed as shown in equation (4) as follows: - + Rilt2) = (Voc _ Vb) / 1 = AV / 1 (4) (IV) The extension theory το extension theory (Extensi〇n the〇ry) was proposed by the mainland scholar Professor Cai Wen in 1983. The concept of the theory is the extension of the study of things and the rules of extension. And methods to solve contradictions and quantify them into information that computers can process. Matter-element theory and extension mathematics are the two foundations of extenics, in which matter-element theory describes matter elements with extension and transformation = Characteristics, and the extensional mathematical concept is based on the extension set and the correlation function as the core of the calculation. The extension theory expresses the information of the thing through the matter-element model, and expresses the conversion relationship between the "quality" and the "quantity" through the matter-element transformation. Therefore, it is possible to clearly show the influence of the characteristics of the property by the degree of association between the "quantity" and the quality of the problem. In addition, the extension theory extends classical mathematics from two-valued logic to continuous muhi-vaiue, and uses the correlation function (Correlation functi〇n) to express the trait of the thing, that is, the application of the real description of the towel. The degree to which an element belongs to a certain feature, and this value can be called the membership grade of the element to which the feature belongs. ° • If the degree of association after normalization is 1, it means that the element is fully compliant with the The characteristics of things. If the degree of association after normalization is ], it indicates the element ^. The degree of association between <-u> is the degree to which the element is completely or completely unaffected by the characteristics of the thing. SUMMARY OF THE INVENTION The object of the present invention is to provide an estimation method for the residual electric quantity of a perylene acid battery, which is used to improve the residual electric quantity of the existing (four) battery, and the accuracy is poor or the estimation speed is slow. According to the method for estimating the residual electric quantity of a sour acid battery according to the present invention, the parallel processing architecture and learning ability of the neural network are utilized, and the extension theory is calculated by the calculation of the correlation function value. : 'To estimate the residual power of the wrong acid battery. Therefore, it is possible to evaluate the residual power of the lead-acid battery, and to improve the use of the portable device, such as the effect of the Η亜 Η亜 丨 丨 / / I /, the ice 'East time is short and the memory = small size' Sex. According to a preferred embodiment of the present invention, the method for the residual power of a lead-acid battery is to establish a δ < 5 cut 3076 phase-element matter model by establishing a residual charge of a lead-acid battery. The meta-model method uses the relevant parameters of the amount of the lyric acid battery to carry out the various residual electricity of the wrong acid battery, and establishes the extension element model corresponding to each of the residual power levels according to the extension matter model. A liftable neural network; the neural network includes an input layer and an output layer, and the input layer includes a Fu-rounded node. The output layer includes a plurality of matter elements, each - two: an outbound node. Each input layer node and each output layer node have: a weight value, and each of the output layer nodes corresponds to one of the residual power levels r and then a training program of the extension type neural network. The training can be phased::::transmission 2: the training element feature sample estimates the residual power of the two connected weights of the acid battery, and the input: the input point = _τ identification, there will only be one of them - the output layer The level of the j is the corresponding residual power level represented by the output layer node, so as to determine which of the remaining power levels the data to be tested belongs to. [Embodiment] FIG. 2 is a flow chart showing an estimation method of a method for estimating a residual battery power according to a preferred embodiment of the present invention, comprising the following steps: Through the extension of the material to establish the 33 (four) pool of various residual power of the extension - as shown in the steps. Establish a scalable neural network architecture with this extension = 丄 〇 76 =, such as step 2 = network training program, with a known percentage of residual power:; the class model as a comparable The extension of the 230-(4) round of the people sample training, such as ^ no. Estimate the residual power of the extended diacid battery that has been trained, as in step 24. The architecture shown is wrong*The law... Estimation of the residual capacity of the acid battery (1) Establishing the extensional matter model of the residual electricity (Step 21〇) To accurately estimate the residual acidity of the residual acid battery of the battery, First establish the extension matter element model (ie, the f-sample) in the lead Ge power. Before establishing the residual power of the wrong acid battery, it is necessary to obtain the relevant parameters of the wrong acid battery. 70. In the embodiment, the electronic load is applied to the lead acid SHYKUANG BP12-12) as a 1A constant electric discharge and dust mite resistance test instrument (TEST3) test and the internal resistance is changed every minute. The edge of the phlegm is recorded, and the load is removed at each 拄~懕夕恋^ (4) hours to record the open circuit value, and the obtained test curves are as shown in Figure 3a and Figure 3b, respectively. The change curve of the lead-acid battery electric raft under the discharge experiment can be seen that the open circuit voltage value of 3〇1 and its battery discharge degree (residual power) are in a nearly linear relationship, so the open circuit voltage of the battery can be estimated as the residual power. Reference. Figure 3b shows the variation of the resistance 302 during the discharge of the acid battery. In order to make the estimation rate more accurate' and divide the open circuit electric dust 1353076 with the internal resistance value separately - the electric quantity - special η "is also regarded as the open circuit of the residual battery? In the example, according to the correlation of the measured acid resistance, the circuit current (the open circuit voltage divided by the internal power type (that is, the establishment of various residual electricity = = residual power level is divided into ten categories, then ten kinds can be used) The residual electricity type and the payment number are shown in Table 1. The mutual residual electricity is 90%. The residual electricity is 70〇/» The residual electricity is 50\ Jade 21 The residual electricity is 30. ίο The classification symbol 〇 Residual power 芩 8〇 〇/〇 residual power is 4〇0/< -^λ_ΑΆΆ*_^6〇% ^ ----~~^1:__π 町, 电里府υ% According to the measured acid battery parameters, The extension matter element model of ten kinds of residual electricity is shown in Table 2. In 1 , the extension 兀 τ thing (R) represents the matter of ten kinds of residual electricity classification of lead-acid battery, ruler = 丨 "{尺1 , ruler 2, ..., especially.} is the residual electricity classification set, each feature element uses three characteristic elements, namely internal resistance, and road voltage ( And the short-circuit current ((//u. The domain range of each matter-element model is the classical domain corresponding to the feature element.) In this embodiment, the internal resistance, open-circuit voltage, and short-circuit current are classified by each residual power. The maximum and minimum values generated are the basis of the range of values. Since the classical domain of each class of matter is based on the resistance value of the battery, the open circuit voltage and the upper and lower limits of the short circuit current are input variables, the lead acid battery is in different residuals. The internal resistance value, open circuit electric grinder and short circuit 11 1353076 current maximum value and minimum value are the range of the classical domain of each class of matter. The second table of acid-acid battery residual state is blessing Matter element model.
殘餘電量類別 物元模型 K1 (90%) 尺=‘ K、,圪, (11.19,11.82) ' Vx, (12.71,12.82) VJK^ (1.081494,1.142091) ► K2 (80%) X,圪, (11.83,12.16) ' Vx, (12.59,12.69) VJK^ (1.036184,1.072696) > K3 (70%) K,,亀, (12.18,12.64) Vx, (12.49,12.58) v〇cIRin^ (0.988133,1.03284) > K4 (60%) R4=< K4, (12.65,13.69) VK, (12.33,12.46) (0-900657,0.98498) K5 (50%) X, (13.7,15.95) Vx, (12.18,12.32) Vcc/Rin^ (0.775445,0.940541) > K6 (40%) K6, (15.96,20.18) Vx, (11.87,12.18) v〇c/Rin< (0.588206,0.763158) * K7 (30%) ΚΊ,圪, (20.19,25.24) Vx, (11.7,11.86) Voc!Rin, (0.465533,0.58742) • K8 (20%) Ks,圪, (25.27,36.65) Vx, (11.59,11.74) v〇c/Rin> (0-316235,0.464583) » K9 (10%) 仏=< X,圪, (33.66,46.5) VK, (11.48,11.6) VJ^, (0.227112,0.316421) » K0 (0%) X, (46.7,96.7) Vx, (11.05,11.47) VJK^ (0.118925,0.24561) ► 12 1353076 第2表之物元模型可以 「Ar 』M式(5 )加以表示如下·_厂Ά吖’左=1,2”.·,〇 (5) C2, κ kl r 幻」Residual electricity class matter element model K1 (90%) Ruler = 'K,,圪, (11.19,11.82) ' Vx, (12.71,12.82) VJK^ (1.081494,1.142091) ► K2 (80%) X,圪, ( 11.83,12.16) ' Vx, (12.59,12.69) VJK^ (1.036184,1.072696) > K3 (70%) K,,亀, (12.18,12.64) Vx, (12.49,12.58) v〇cIRin^ (0.988133, 1.03284) > K4 (60%) R4=< K4, (12.65,13.69) VK, (12.33,12.46) (0-900657,0.98498) K5 (50%) X, (13.7,15.95) Vx, (12.18 ,12.32) Vcc/Rin^ (0.775445,0.940541) > K6 (40%) K6, (15.96,20.18) Vx, (11.87,12.18) v〇c/Rin< (0.588206,0.763158) * K7 (30%) ΚΊ,圪, (20.19,25.24) Vx, (11.7,11.86) Voc!Rin, (0.465533,0.58742) • K8 (20%) Ks,圪, (25.27,36.65) Vx, (11.59,11.74) v〇c /Rin> (0-316235,0.464583) » K9 (10%) 仏=< X,圪, (33.66,46.5) VK, (11.48,11.6) VJ^, (0.227112,0.316421) » K0 (0%) X, (46.7, 96.7) Vx, (11.05, 11.47) VJK^ (0.118925, 0.24561) ► 12 1353076 The matter-element model of the second table can be expressed as "Ar" M formula (5) as follows: _厂Ά吖' left =1,2".·,〇(5) C2, κ kl r illusion
:中,A為第灸個分類群集之物元模型,代表該鉛酸 ^也屬於第/個分類群集之物元模型Μ為該第㈣分類群 ”之物70¼型的名稱,代表第&個分類群集之殘餘電量類 為乂之第,特徵元,,网,2,3,為 '、·徵7LCy之經典域值,π表該錯酸電池屬於第灸個分 =群集之物元模型中相對應於特徵元c,量值,其中吣為 ,個特徵元對應第⑽類群集之權重最小值,而眸為 I個特徵元對應第_分類群集之權重最大值該可 拓物元模型的總數(即為分類的數目)。 (2 )建立可拓類神經網路架構(步驟22〇 ) 請參照第4圖’其繪示依照本發明一較佳實施例的一 種可拓類神經網路之架構圖。可拓類神經網路_包含— 輸入層㈣以及一輸出層侧。其中輸入層川包含複數 個輪入層節點41卜分別輸入物元特徵的量值(%〜尤p), 輸出層420包含複數個物元’每—物元對應一輸出層節點 (421。在輸出層42",每—次僅有一輸出層 節點421發生變動,藉以代表輸入物元特徵樣本之辨識結 果。 13 1353076 母一輸入層節點411及輪出層節點421間有 權重值其中—權重值:二個連結 元特徵之經典域最大值,另—權重〆、代表輸入物 特徵之矬典域最小值。 &代表輸入物元 輸入層節點數為物元模型的特徵 例中’輸入層節點數為3 (亦即為 n ’在本實施 開路電壓與短路電&丨^ ,的内電阻值、 型總數〇〜〇,本實施例中,輸出層^所數建立之物元模 池殘餘電量分成十類等級數 二數為1〇(即電 A太音丄 乐i表所示)〇 在本實苑例中’權重值430的初始 特徵元之經典域的最大值及最小值H值為輸入樣本 代表連接於^個輸人層節點與,· %與·別 及最大權重值。 固輸出層節點之最小 在本實施例中,所採用之 鲁 等級的訓練樣本輪入,藉由可拓距離 1為將母一殘餘電量 剛的計算㈣算其相對權録 ㈣fa·, 值依一定的演算方式重新調整確’僅錯誤的權重 路架構,於診斷(估測)階段時1 =可拓類神經網 殘餘電量分類’僅會有一輸出層 ^層節點對應- 節點所代表之殘餘電量等級。藉由輪以表示該 結權重值430進行訓練,以達』^特徵樣本對連 量估測。 續對銘酸電池的殘餘電 (3)可拓類神經網路之監督 設某-電池殘餘電量分類;丨:程序(步驟23〇) 類#級的訓練樣本為 14 z—卜丨,',...,、},其中%為 為尤p\ 練樣本總數,第/個樣本表示 1 ίΆ,...,<},《代表樣本 中㈣),p為第Η固樣本所屬之殘的在本實施例 了評估可抬類神經網路之識別二餘電量種類(等級)。為 誤差數目,則總誤 確性為整體估測之 " 、可表示如下: Ν Ρ (6) m % 可括類神經網路之學習 習權重值調整使得' ”·、監督式學習,其藉由學 或達到與目標值相同之輸出特性。較佳之辨識率,亦 ⑼參照第5圖’其繪示依照第2 二、.轉訓練的流程圖。可括類 ::行可拓類神 含: 之甽練學習步驟包 步驟221 ··設定連結於輪入層節點 權重值可—型之形,々=u,···, (\_1)0 (7) :式(7)中’Cy=1,2,3為'第·/·個特徵元,〆 為闕於特徵元ς之經典域值,㈣,町〉, 由訓練資料所決定。其中 士典蜮之範圍可 R,. u2C3, K) (8) 15 (9) 1353076 值,『"為第二:姓第7個特徵對應第“固分類之權重最小 步: 22/.徵對應第“固分類之權重最大值。In the middle, A is the matter-element model of the moxibustion classification cluster, which represents the name of the element 701⁄4 type of the elementary model of the first/category cluster, and the name of the 701⁄4 type, representing the & The residual electric energy class of the classification cluster is the first, characteristic element, and net, 2, 3, which is the classic domain value of ', · levy 7LCy, π table, the wrong acid battery belongs to the moxibustion point = cluster matter element model Corresponding to the feature element c, the magnitude, where 吣 is, the feature element corresponds to the minimum weight of the cluster of the (10) class, and 眸 is the weight of the feature element corresponding to the _ classification cluster. The total number of the classifications (ie, the number of classifications). (2) Establishing an extension-like neural network architecture (step 22). Referring to FIG. 4, an extended neural network is illustrated in accordance with a preferred embodiment of the present invention. The architecture diagram of the road. The extension type neural network _ contains - the input layer (4) and an output layer side, wherein the input layer includes a plurality of round-trip nodes 41 and the magnitude of the input matter element respectively (%~ especially p) The output layer 420 includes a plurality of matter elements 'each element corresponding to an output layer node (421. On the output 42", there is only one output layer node 421 changed every time, which represents the identification result of the input matter element feature sample. 13 1353076 The parent-input layer node 411 and the round-out layer node 421 have the weight value among them - the weight value: two The maximum value of the classic domain of the connected meta-features, and the weights 〆, the minimum values of the classical regions representing the characteristics of the input objects. & represents the number of input element inputs. The number of input nodes is the feature of the matter-element model. 3 (that is, n 'in the open circuit voltage and short circuit power & 丨 ^, the internal resistance value, the total number of types 〇 ~ 〇, in this embodiment, the output layer ^ number of the built-in material pool residual power Divided into ten categories, the number two is 1 〇 (that is, the electric A is too 丄 丄 i i). In the example of the real court, the maximum and minimum H values of the classical domain of the initial feature element of the weight value 430 The input sample represents the connection to the ^ input layer node and the % and the maximum weight value. The minimum of the solid output layer node In this embodiment, the training level of the Lu level used is rounded, by Extension distance 1 is the remaining amount of the mother Calculate (4) Calculate its relative power record (4) fa·, the value is re-adjusted according to a certain calculation method. It is only the wrong weight road structure. In the diagnosis (estimation) stage 1 = extension type neural network residual electricity classification 'only one output The layer layer node corresponds to the residual power level represented by the node. The training is performed by the wheel to represent the weighting value 430, so as to achieve the joint quantity estimation of the characteristic sample. Continued residual electricity of the eracid battery (3) Supervised neural network supervision set a certain - battery residual power classification; 丨: program (step 23 〇) class # training sample is 14 z-di, ',...,,}, where % is In particular, the total number of samples, the first sample represents 1 ίΆ,...,<}, "in the representative sample (4)), and p is the residual of the tamping sample. In this example, the evaluable nerve is evaluated. The network identifies two types of power (grade). For the number of errors, the total misrecognition is the overall estimate of ", which can be expressed as follows: Ν Ρ (6) m % The learning value of the neural network can be adjusted to make ' 》·, supervised learning, which borrows Learn or achieve the same output characteristics as the target value. The better recognition rate, also (9) refer to Figure 5, which shows the flow chart according to the second two, the training. Can be included: : Step 8 of 甽 学习 学习 学习 221 · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · =1, 2, 3 are '·································································· , K) (8) 15 (9) 1353076 value, "" is the second: the last name of the seventh feature corresponds to the "solid weight of the solid classification minimum step: 22 /. sign corresponding to the "solid classification weight maximum.
步驟222.計算每項權重群集之初始權中心㈣W cluster center) 〇 ={^l,Z.) 7 ( L a (10) (11) “1’ 2,..·,W = 1,2n (12)Step 222. Calculate the initial weight center of each weight cluster (4) W cluster center) 〇={^l,Z.) 7 ( L a (10) (11) “1′ 2,..·,W = 1,2n ( 12)
步驟2 2 3 ··讀取第f·個訓練樣本及其分類編號^。 XHd4,penc ⑼ /驟224_應用可拓距離(ED)計算訓練樣本^與^化 群集之距離’也就是計算訓練樣本與每—分類群集之 距離,其數學表示式如下: ^=Σ >ι ^ = 1, 2,..., η \x〇 ~ zkj\~(^!j-w^.)/2 厂 14) 清參照第6圖,其繪示為本實施例中所提之可拓距離 (ED)之示意圖。可拓距離(ED)可用以表示點X盥範 圍之距離,其中χ為群集之權中心,〆為權重群集 最小值,w"為權重群集最大值。由第6圖得知可拓距離; 因不同之數值範圍形成距離計算上之差異,因而產生如同 靈敏度數值變化之不同…般而言,若物元特徵之範圍較 大時,其意味著訓練資料較為廣泛模糊,因此表現在距離 計算上之靈敏度較低;相反地,若物元特徵之範圍較小 時,代表著所需資料樣本較為精確,所以可以表現出距離 16 1353076 計算上之高靈敏度。 步驟225 :經由計算比較後找到灸·,使得 五’若此時〆=〆即可拓距離最小之殘餘電量 類別與該訓練樣本所屬之殘餘電量類別相同)則演算程序 跳至步驟227 ;否則繼續執行步驟226。Step 2 2 3 ·· Read the f-th training sample and its classification number^. XHd4,penc (9) /Step 224_ Apply the extension distance (ED) to calculate the distance between the training sample and the cluster. That is, calculate the distance between the training sample and each cluster. The mathematical expression is as follows: ^=Σ > ι ^ = 1, 2,..., η \x〇~ zkj\~(^!jw^.)/2 Factory 14) Refer to Figure 6 for illustration, which shows the extension proposed in this embodiment. Schematic of distance (ED). The extension distance (ED) can be used to represent the distance of the point X盥 range, where χ is the weight of the cluster, 〆 is the weight cluster minimum, and w" is the weight cluster maximum. The extension distance is known from Fig. 6; the difference in the calculation of the distance is formed by different numerical ranges, and thus the difference in sensitivity value is generated. Generally, if the range of the matter element is large, it means training data. It is more ambiguous, so it is less sensitive in distance calculation; conversely, if the range of matter element is small, it means that the required data sample is more accurate, so it can show the high sensitivity of the distance 16 1353076. Step 225: After the comparison is found, the moxibustion is found, so that if the residual electric quantity category with the smallest extension distance is the same as the residual electric energy category to which the training sample belongs, the calculation program jumps to step 227; otherwise, continues Go to step 226.
步驟226:調整並更新p-认及々·_认群集之權重值如下: (a)更新;7-汍及认群集之權中心值,亦即更新訓練樣 本本身應屬分類群集(认群集)的權中心值以及更新訓 練樣本誤判對應之分類群集(Γ-认群集)的權中心值。 2Γ = β+η(«)Step 226: Adjust and update the weight values of the p-recognition and _· recognition clusters as follows: (a) update; 7-汍 and recognize the weight of the cluster center value, that is, the updated training sample itself should belong to the classification cluster (recognition cluster) The weight center value and the weight center value of the classification cluster (Γ-recognition cluster) corresponding to the update training sample misjudgment. 2Γ = β+η(«)
HA (16) 其中,為一更新後該訓練樣本所屬分類群集之權 中心值,為-更新前該訓練樣本所屬分類群集之權中 心值,為一更新後該錯誤輸出分類群集之 oidHA (16), wherein, after updating, the weighted center value of the classification cluster to which the training sample belongs is - the weight center value of the classification cluster to which the training sample belongs before updating, and the oid of the error output classification cluster after an update
〇為-更新前錯誤輸出分類群集之權中心值,"為一所設 疋之學習率’而¥為該訓練樣本之資料。 (b)更新p-认及灸·-㈣集之權錄,亦即更新訓練樣本 本身應屬分類群集(轉集)的權重值以及更新訓練樣 本誤判對應之分類群集(〆_认群集)的權重值。 <(ww)=<(oW)+7(^-4(oW)) (Π) 吣一 η(χ卜呓, 士) (ΐ8) 17 1353076 式(15 )〜式(1只、士 λ . ^ 中之7代表其學習率(Learning rate)。 在此步驟令’僅群隹 · 果户及々之權重在學習過程中進行調 整,而其匕權重值並 .s -, Λ Α 不改變。由於上述之特性,使得可拓 其他類神經網路架構擁有計算速度較快之 二:且於新的應用領域中具有較高之適應性。 凊參照第7a圖第7h固 _ 弟7b圖’其繪示為兩群集權重在學習 程序中進行調整之結果 & 禾的不忍圖。由圖中可以發現原始狀 m〇为—The weight of the pre-update error output cluster cluster center value, " is a set learning rate' and ¥ is the training sample data. (b) update the rights of the p-recognition and moxibustion--(4) episodes, that is, update the training sample itself to be the weight value of the classification cluster (transfer) and the classification cluster corresponding to the misjudgment of the training sample (〆_cognition cluster) Weights. <(ww)=<(oW)+7(^-4(oW)) (Π) 吣一η(χ卜呓,士) (ΐ8) 17 1353076 Equation (15)~式(1,士士7 of λ . ^ represents its learning rate. In this step, the weights of only the group, the household, and the 々 are adjusted during the learning process, and the weights of the 匕 are weighted and .s -, Λ Α Due to the above characteristics, the extension of other neural network architectures has the second fastest calculation speed: and it has higher adaptability in new application fields. 凊 Refer to Figure 7a, 7h solid _ brother 7b 'It is shown as the result of the adjustment of the two cluster weights in the learning program & Wo's intolerance. The original shape can be found in the figure m
態測試資料與群隼Δ /、之可拓距離()與原始狀態測試 資料與群集B之可拉路雜< λ女 柘距離(五Ζ)β )產生明顯之變化。且由 於吨 >吗(為調整後測試資料與群集&之可拓距 離’碼為調整後測試資料與群集B之可拓距離),因此訓 練樣本V斤屬群集由群集a(Za)變化至群集Μ。)。 V驟227 .重複步驟223至步驟226之演算程序,直 到所有的訓練樣本均已分類完畢’並且結束一學習批次 (Epoch ) ° 步驟228 :若分類程序已經達到收斂狀態,或是總誤 差率(尽)已達到預設之目標值則停止演算程序,否則返 回至步驟223。 在此以殘餘電量50%及40%舉例說明,分別列出其經 典域範圍,並根據步驟222之公式計算殘餘電量5〇%及 40%之初始權中心值。利用第2表及式(8)〜式(9)可得 殘餘電量50%各特徵之經典域範圍為: 〜〇% = 15.95,^j^〇%= 12.32 , W〇%= 0.940541 ^5〇%=13·7 吃 〇% =12.18 ^5〇./ο =0.775445 18 1353076The state test data and the group 隼 Δ /, the extension distance () and the original state test data and the cluster B's cola path < λ 柘 distance (five Ζ) β) produce significant changes. And because of the ton> (for the adjusted test data and the cluster & extension distance code is the extension distance of the adjusted test data and cluster B), the training sample V is clustered by the cluster a (Za) To the cluster. ). V. Step 227. Repeat the calculation procedure of steps 223 to 226 until all training samples have been classified 'and end a learning batch (Epoch). Step 228: If the classification procedure has reached a convergence state, or the total error rate If the preset target value has been reached, the calculation program is stopped, otherwise it returns to step 223. Here, the residual power is 50% and 40%, and the typical domain ranges are listed separately, and the initial weight center values of the residual powers of 5〇% and 40% are calculated according to the formula of step 222. Using the second table and equations (8) to (9), the classical domain range of 50% of the residual charge is: 〇% = 15.95, ^j^〇%= 12.32, W〇%= 0.940541 ^5〇 %=13·7 〇%=12.18^5〇./ο =0.775445 18 1353076
利用第2表及式(8)〜式(9)可得殘餘電量40%各 特徵之經典域範圍為: U.18 ,(=12.18,吣0% =0.763158 ^/40% =15.96 ^4〇«/„ = 11-87 5 ^/40%= 0.588206 利用式(11 )計算出殘餘電量50%各特徵之初始權中 心為: ^5〇〇/〇 = (K5〇% + ^5〇%)/2 = (15.95+13.7)/2=14.825 ^5〇〇/„ =(<5〇〇/„ + ^5〇〇/〇)/2 = (12.32+12.18)/2=12.25 ^/5〇〇/〇 = (Ko% + ^〇〇/„)/2 =(0-940541+0.775445)/2=0.85799 利用式(11)計算出殘餘電量40%各特徵之初始權中心Using the second table and equations (8) to (9), the classical domain range of each characteristic with a residual charge of 40% is: U.18, (=12.18, 吣0% =0.763158 ^/40% =15.96 ^4〇 «/„ = 11-87 5 ^/40%= 0.588206 Use the formula (11) to calculate the initial weight center of each characteristic of 50% of residual electricity: ^5〇〇/〇= (K5〇% + ^5〇%) /2 = (15.95+13.7)/2=14.825 ^5〇〇/„ =(<5〇〇/„ + ^5〇〇/〇)/2 = (12.32+12.18)/2=12.25 ^/5 〇〇/〇= (Ko% + ^〇〇/„)/2 =(0-940541+0.775445)/2=0.85799 Calculate the initial weight center of each characteristic of 40% of residual electricity using equation (11)
Zm=(K〇y^K〇%)/2 = (20.18+15.96)/2=18.07 ^40〇/〇 = «4〇〇/〇 + ^4〇〇J/2 = (12.18+11.87)/2=12.025 z/4〇〇/„ = (^/4〇〇/„ + ^4〇./o)/2 = (0.763158+0.588206)/2=0.675682 接著讀取訓練樣本,以一已知殘餘電量為50%之資料 為例,並根據步驟224之可拓距離公式,分別計算訓練資 料樣本與各分類之可拓距離值,可拓距離越小代表屬於該 分類區域可能性越大,反之則可能性越小。 取一已知實際殘餘電量狀態為50%之訓練樣本如下: '50%, Rin, 14.45 及= Kc, 12.26 _ Isc, 0.848443 19 1353076 利用式(14)計算訓練樣本與殘餘電量50%各特徵之 可拓距離:Zm=(K〇y^K〇%)/2 = (20.18+15.96)/2=18.07 ^40〇/〇= «4〇〇/〇+ ^4〇〇J/2 = (12.18+11.87)/ 2=12.025 z/4〇〇/„ = (^/4〇〇/„ + ^4〇./o)/2 = (0.763158+0.588206)/2=0.675682 Then read the training sample to a known residual Taking the data of 50% of electricity as an example, according to the extension distance formula of step 224, the extension distance values of the training data samples and each classification are respectively calculated. The smaller the extension distance, the greater the probability that the classification area belongs to the classification, and vice versa. The less likely it is. Take a training sample with a known actual residual charge status of 50% as follows: '50%, Rin, 14.45 and = Kc, 12.26 _ Isc, 0.848443 19 1353076 Use equation (14) to calculate the training sample and residual power 50% Extension distance:
ED R50% V50% |14.45 -14.825| - (15.95 -13.7) / 2 |(15.95-13.7)/2| |12.26 -12.25| - (12.32 -12.18)/2 |(12.32-12.18)/2| + 1 = 0.3333 + 1 = 0.14285 _|〇.848443-0.85799|-(0.940541-0.775445)/2 ED τ ^f\〇A 一 I +1 ,/50% β» |(0.940541-0.775445)/2| = 0.11565392 將計算出之各特徵值可拓距離相加,即可得訓練樣本 與50%殘餘電量之總可拓距離值為 V50% + ^^/5〇% = 0-59180392 同樣的,利用式(14 )計算訓練樣本與殘餘電量40% 各特徵之可拓距離:ED R50% V50% |14.45 -14.825| - (15.95 -13.7) / 2 |(15.95-13.7)/2| |12.26 -12.25| - (12.32 -12.18)/2 |(12.32-12.18)/2| + 1 = 0.3333 + 1 = 0.14285 _|〇.848443-0.85799|-(0.940541-0.775445)/2 ED τ ^f\〇A I I +1 , /50% β» |(0.940541-0.775445)/2| 0.11565392 Add the calculated extension values of each eigenvalue to obtain the total extension distance value of the training sample and 50% residual power V50% + ^^/5〇% = 0-59180392 Similarly, use the formula ( 14) Calculate the extension distance between the training sample and the residual electricity 40% of each feature:
ED R40% |14.45-18.07|-(20.18-15.96)/2 |(20.18-15.96)/2| + 1 = 1.715639 % Ύ40% 12.26-12.025 -(12.18-11.87)/2 J-;--;——— + 1 = 1.51613 (12.18-11.87)/2 zm _|〇.848443-0.675682|-(0.763158-0.588206)/2 EDiΛ(\〇/_ 一 i i 1 ΊΛ0% |(0.763158-0.588206)/2| 1.97495 將計算出之可拓距離相加,可得訓練樣本與40%殘餘 電量之總可柘距離值為 -^^40% = ^R40% + ^^K40% + -^/40% = 5.206719 訓練樣本與其他百分比殘餘電量之總可拓距離,亦可 20 2 1353076 依相同方式求得。根據總可拓距離計算,可得訓練樣本與 50%殘餘電量之總可拓距離為最小,故其權重中心及權重 值不須作更新,繼續讀取下一筆訓練樣本。但若該筆已知 為50%殘餘電量之訓練樣本,其與4〇%殘餘電量各特徵計 算所得之總可拓距離為最小的話,則須執行步驟226之權 中心及權重值更新。 其可利用式(15)〜式(16)作各特徵之權中心更新 如下(學習率設為7/=0.9):ED R40% |14.45-18.07|-(20.18-15.96)/2 |(20.18-15.96)/2| + 1 = 1.715639 % Ύ40% 12.26-12.025 -(12.18-11.87)/2 J-;--;- —— + 1 = 1.51613 (12.18-11.87)/2 zm _|〇.848443-0.675682|-(0.763158-0.588206)/2 EDiΛ(\〇/_ 一ii 1 ΊΛ0% |(0.763158-0.588206)/2| 1.97495 Adding the calculated extension distances, the total 柘 distance value of the training sample and 40% residual power is -^^40% = ^R40% + ^^K40% + -^/40% = 5.206719 Training The total extension distance between the sample and other percentage residual power can also be obtained in the same way as 20 2 1353076. According to the total extension distance calculation, the total extension distance of the training sample and 50% residual electricity is the smallest, so its weight The center and weight values do not need to be updated to continue reading the next training sample. However, if the pen is known as a 50% residual power training sample, the total extension distance calculated from the characteristics of the 4% residual power is the minimum. If necessary, the weight center and weight value update of step 226 must be performed. The weight center of each feature can be updated by using equations (15) to (16) as follows (learning rate is set to 7/=0.9):
zr7w, = 14·825 + 0.9(14.45 -14.825) = 14.4875 zv5〇% = 12.25 + 0.9(12.26 -12.25) = 12.259 z/w. = 0*85799 + 0.9(0.848443 - 0.85799) = 0.8493977 znRZ% = 18.07 - 0.9( 14.45 -18.07) = 21328 z;7〇% = 12.025 - 0.9(12.26 -12.025) = 11.8135 ^7〇% = 0-675682 - 0.9(0.848443 - 0.675682) = 0.520197 至於各權重值則可利用式(17)及式(18)作更新如 下:Zr7w, = 14·825 + 0.9(14.45 -14.825) = 14.4875 zv5〇% = 12.25 + 0.9(12.26 -12.25) = 12.259 z/w. = 0*85799 + 0.9(0.848443 - 0.85799) = 0.8493977 znRZ% = 18.07 - 0.9( 14.45 -18.07) = 21328 z;7〇% = 12.025 - 0.9(12.26 -12.025) = 11.8135 ^7〇% = 0-675682 - 0.9(0.848443 - 0.675682) = 0.520197 As for the weight values, the available formula (17) and (18) are updated as follows:
Ka^ = 15.96 - 0.9(14.45 -18.07) = 19.218 wr^%v) = 20.18 - 0.9(14.45 -18.07) = 23.438 如以上之更新程序,其餘權重值均以式(17)及式(is) 作更新,而更新後之權中心值及權重值再代回步驟225重 新估測,直到所有訓練資料均完成訓練程序。 (4 )可拓類神經網路之估測演算流程(步驟240 ) 請參照第8圖,其繪示係依照第2圖中之進行殘餘電 S ) 21 1353076 量估測的流程圖。當可拓類神經網路完成學習程序後,即 可進行估測分類,而其演算程序包含·· V驟241 .讀取可拓類神經網路之權重值矩陣。 /驟242 .利用式(u),計算每一分類群集之初始群 集權中心值。 步驟243 :讀取測試樣本 ' “1,乂2,···,(19) m /驟244.應用式(14)所提可拓距離(ed)之定 計算測試樣本與每一分類群集之距離。 /驟245 .找哥灸使得碑=她吨卜並且設定其相 •-輸出即點%· =1,冑以顯示測試樣本所屬群集類 別。若所有測試樣本均已分類完成則停止運算程序,否則 回到步驟243。 ^本實施例中,先以刚筆已知 資料對可拓類神經網路進行訓練,接= 3表所其特徵值之待測資料_^ ㈣機所選取之^^及其特徵值之待測資料 結果顯示,此方法了:、::表所示’由表中之測試 量。㈣去可正確呈現出待測鉛酸電池之殘餘電 22 1353076 第3表船酸電池殘餘電量狀態待估測之資料。 測試序號 内電阻 (mQ) 開路電壓 (V) 開路電壓/内電阻 (A) 實際殘餘電量 (%) 1 11.82 12.82 1.084602 90% 2 33.7 11.61 0.34451 20% 3 12.75 12.44 0.975686 60% 4 17.76 12.00 0.675676 40% 5 23.82 11.78 0.494542 30% 6 52.70 11.42 0.216698 0% 7 41.70 11.50 0.276978 10% 8 26.62 11.73 0.440646 20% 9 11.56 12.79 1.106401 90% 10 14.64 12.25 0.836749 50% 11 18.68 11.96 0.640257 40% 12 15.46 12.21 0.78978 50% 13 13.70 12.32 0.89927 60% 14 24.52 11.76 0.479608 30% 15 12.15 12.65 1.04115 80%Ka^ = 15.96 - 0.9(14.45 -18.07) = 19.218 wr^%v) = 20.18 - 0.9(14.45 -18.07) = 23.438 As the above update procedure, the remaining weight values are given by equations (17) and (is) The update, and the updated weight center value and weight value are then re-evaluated in step 225 until all training materials complete the training program. (4) Estimation calculation flow of extension type neural network (step 240) Please refer to Fig. 8, which is a flow chart showing the estimation of residual electric power S) 21 1353076 according to Fig. 2. When the extensional neural network completes the learning process, the estimation classification can be performed, and the calculation program includes the V. 241. The weight value matrix of the extension type neural network is read. /Step 242. Using equation (u), calculate the initial cluster center value for each cluster. Step 243: Read the test sample '1, 乂2, ···, (19) m / 244. Calculate the test sample and each classification cluster by the extension distance (ed) of the application formula (14) Distance. / 245. Find the moxibustion so that the monument = her tat and set its phase - output point % =1, 胄 to display the cluster category of the test sample belongs to. If all test samples have been classified, stop the operation program Otherwise, go back to step 243. In this embodiment, the extension type neural network is trained with the known data of the pen, and the data of the eigenvalue of the table is _^ (4) selected by the machine. ^ and its eigenvalues of the data to be tested show that this method: ::: The table shows 'the amount of test in the table. (4) to correctly display the residual electricity of the lead-acid battery to be tested 22 1353076 Table 3 Acid battery residual state status to be estimated. Test number internal resistance (mQ) Open circuit voltage (V) Open circuit voltage / internal resistance (A) Actual residual power (%) 1 11.82 12.82 1.084602 90% 2 33.7 11.61 0.34451 20% 3 12.75 12.44 0.975686 60% 4 17.76 12.00 0.675676 40% 5 23.82 11.78 0.494542 30% 6 52.70 11.42 0.216698 0% 7 41.70 11.50 0.276978 10% 8 26.62 11.73 0.440646 20% 9 11.56 12.79 1.106401 90% 10 14.64 12.25 0.836749 50% 11 18.68 11.96 0.640257 40% 12 15.46 12.21 0.78978 50% 13 13.70 12.32 0.89927 60% 14 24.52 11.76 0.479608 30% 15 12.15 12.65 1.04115 80%
第4表可拓類神經網路電量估測法之估測結果。 測 試 序 號 i 第/個樣本對應各殘餘電量分類之1 命出值 殘餘 電量 類別 判斷 結果 〇" (90%) 〇12 (80%) 0/3 (70%) 〇i4 (60%) 0/5 (50%) 〇i6 (40%) 〇π (30%) 〇i8 (20%) 〇i9 (10%) 〇iO (〇%) 1 (T) 0 0 0 0 0 0 0 0 0 90% 2 0 0 0 0 0 0 0 (ϊ) 0 0 20% 3 0 0 0 (ί) 0 0 0 0 0 0 60% 4 0 0 0 0 0 ① 0 0 0 0 40% 5 0 0 0 0 0 0 (Ϊ) 0 0 0 30% 6 0 0 0 0 0 0 0 0 0 (ϊ) 0% 7 0 0 0 0 0 0 0 0 (Τ) 0 10% 8 0 0 0 0 0 0 0 (i) 0 0 20% 9 (Ϊ) 0 0 0 0 0 0 0 0 0 90% 10 0 0 0 0 (ϊ) 0 0 0 0 0 50% 11 0 0 0 0 0 (ΐ) 0 0 0 0 40% 12 0 0 0 0 (i) 0 0 0 0 0 50% 13 0 0 0 (Ϊ) 0 0 0 0 0 0 60% 14 0 0 0 0 0 0 (Τ) 0 0 0 30% 15 0 0) 0 0 0 0 0 0 0 0 80%Table 4 is an estimate of the estimated neural network power estimation method. Test No. i The first sample corresponds to each residual power class. 1 Life Out Value Residual Power Category Judgment Result quot" (90%) 〇12 (80%) 0/3 (70%) 〇i4 (60%) 0/ 5 (50%) 〇i6 (40%) 〇π (30%) 〇i8 (20%) 〇i9 (10%) 〇iO (〇%) 1 (T) 0 0 0 0 0 0 0 0 0 90% 2 0 0 0 0 0 0 0 (ϊ) 0 0 20% 3 0 0 0 (ί) 0 0 0 0 0 0 60% 4 0 0 0 0 0 1 0 0 0 0 40% 5 0 0 0 0 0 0 (Ϊ) 0 0 0 30% 6 0 0 0 0 0 0 0 0 0 (ϊ) 0% 7 0 0 0 0 0 0 0 0 (Τ) 0 10% 8 0 0 0 0 0 0 0 (i) 0 0 20% 9 (Ϊ) 0 0 0 0 0 0 0 0 0 90% 10 0 0 0 0 (ϊ) 0 0 0 0 0 50% 11 0 0 0 0 0 (ΐ) 0 0 0 0 40% 12 0 0 0 0 (i) 0 0 0 0 0 50% 13 0 0 0 (Ϊ) 0 0 0 0 0 0 60% 14 0 0 0 0 0 0 (Τ) 0 0 0 30% 15 0 0) 0 0 0 0 0 0 0 0 80%
請參照第9圖,其繪示依照本發明一較佳實施例的一 23 2 1353076 ㈣酸電池殘餘電量之估測方法進行估測的結果曲線 圖。由於可拓類神經網路可藉由學f程序中,進行權重值 之修正,使件在1GG筆測試資料中估測錯誤筆數僅2筆, 並且只需2次疊代數即可達到高辨識準確率(帆)。 由於可拓類神經網路之初始權重設定已經考慮各估 測類別特徵數值之最大及最小值,所以其總誤差率:初奸Please refer to FIG. 9 , which is a graph showing the results of estimating the residual power of a 23 2 1353076 (four) acid battery according to a preferred embodiment of the present invention. Since the extension-like neural network can correct the weight value by learning the f program, the number of errors in the 1GG pen test data is estimated to be only 2, and only 2 times of the algebra can be used to achieve high recognition. Accuracy (sail). Since the initial weight setting of the extension-like neural network has taken into account the maximum and minimum values of the characteristic values of the estimated categories, the total error rate: first traits
值已較—般_經網料許多,故本實_巾所提之可括 類神經網路收斂逮度較其他方式快且準確。 雖然本發明已以一較佳實施例揭露如上,然其並 以限定本發明,任何熟f此技藝者,衫脫離本發明之於 當可作各種之更動與_,因此本發明之: 遵I巳圍當視後附之申請專利範圍所界定者為I。 、 I固八間早說明】The value is much better than that of the general network. Therefore, the convergence of the neural network can be faster and more accurate than other methods. Although the present invention has been described above in terms of a preferred embodiment, which is intended to be illustrative of the invention, the invention may be practiced otherwise. The definition of the scope of the patent application attached to it is I. , I solid eight early instructions]
=本發明之上述和其他目的、特徵、優點與實施例 牝更月顯易懂,所附圖式之詳細說明如下: 第1圖係繪示為一種常用之錯酸電池的等效電路圖 =2圖係繪示依照本發明—較佳實施例的鉛酸電池殘 '電里之估測方法的流程圖。 第3a圖係繪示依照一種鉛酸電池電壓 的變化曲線圖。 电貫驗下 :3b圖係繪示依照一種鉛酸電池電壓在放 電阻之變化曲線圖 』π 第4圖係繪示依照本發明一較佳實施例的一種可知類 24 1353076The above and other objects, features, advantages and embodiments of the present invention will be more readily understood. The detailed description of the drawings is as follows: FIG. 1 is an equivalent circuit diagram of a commonly used acid-acid battery. The figure is a flow chart showing an estimation method for the residual electric current of a lead-acid battery according to the preferred embodiment of the present invention. Figure 3a is a graph showing the change in voltage of a lead-acid battery. Under the electrical test: 3b is a graph showing the change of the voltage of the lead-acid battery in the discharge resistance. π Figure 4 shows a known class according to a preferred embodiment of the present invention.
神經網路之架構圖。 第5圖係繪示依照第2圖中 練的流程I 之進仃可拓類神經網路訓 第6圖係繪示依照本發明一較佳實施例的一種可拓距 離(ED )之示意圖。 第7a圖係繪示為兩群集權重在學習程序中調整前之 示意圖。 第7b圖係纷示為兩群集權重在學習程序中調整後 示意圖。 第8圖係繪示依照第2圖中之進行殘餘電量估測的 程圖。 之 流The architectural diagram of the neural network. Fig. 5 is a schematic diagram showing an extension distance (ED) according to a preferred embodiment of the present invention. Figure 7a is a schematic diagram showing the two cluster weights before adjustment in the learning program. Figure 7b is a schematic diagram showing the adjustment of the two cluster weights in the learning program. Fig. 8 is a diagram showing the calculation of the residual electric quantity in accordance with Fig. 2. Flow
第9圖係繪示依照本發明一較佳實施例的一種鉛酸電 池殘餘電量之估測方法進行估測的結果曲線圖。 【主要元件符號說明】 12 :電池電極與電池電解液14 :電池電極與電池電解液 之等效電阻 之介面電阻 16 :電池内部電容 20 :放電電流 210〜240 :步驟 221~228 :步驟 241〜245 :步驟 301 :開路電壓 302 :内電阻 410 :輸入層 420 :輸出層 430 :權重值 400 :可拓類神經網路 411 :輸入層節點 421 :輸出層節點 25Figure 9 is a graph showing the results of estimating the residual electricity amount of a lead-acid battery in accordance with a preferred embodiment of the present invention. [Main component symbol description] 12: Battery electrode and battery electrolyte 14: Interface resistance of battery electrode and battery electrolyte equivalent resistance: Battery internal capacitance 20: Discharge current 210 to 240: Steps 221 to 228: Step 241~ 245: Step 301: Open circuit voltage 302: Internal resistance 410: Input layer 420: Output layer 430: Weight value 400: Extension class neural network 411: Input layer node 421: Output layer node 25
Claims (1)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW096144353A TWI353076B (en) | 2007-11-22 | 2007-11-22 | Method for estimating residual capacity of lead-ac |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW096144353A TWI353076B (en) | 2007-11-22 | 2007-11-22 | Method for estimating residual capacity of lead-ac |
Publications (2)
Publication Number | Publication Date |
---|---|
TW200924257A TW200924257A (en) | 2009-06-01 |
TWI353076B true TWI353076B (en) | 2011-11-21 |
Family
ID=44728935
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW096144353A TWI353076B (en) | 2007-11-22 | 2007-11-22 | Method for estimating residual capacity of lead-ac |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI353076B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI461718B (en) * | 2012-08-01 | 2014-11-21 | Univ Ishou | Battery power test method |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI408387B (en) * | 2009-06-22 | 2013-09-11 | Nat Univ Chin Yi Technology | Inverter fault diagnosis method |
TWI395963B (en) * | 2010-01-05 | 2013-05-11 | State-of-health estimation method for lead-acid batteries | |
TWI395965B (en) * | 2010-06-04 | 2013-05-11 | Nat Univ Chin Yi Technology | Fuel cell faulty predicting system and its establishing method |
CN109410367B (en) * | 2018-10-16 | 2022-02-15 | 平安科技(深圳)有限公司 | Vehicle power supply management method, computer readable storage medium and server |
-
2007
- 2007-11-22 TW TW096144353A patent/TWI353076B/en not_active IP Right Cessation
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TWI461718B (en) * | 2012-08-01 | 2014-11-21 | Univ Ishou | Battery power test method |
Also Published As
Publication number | Publication date |
---|---|
TW200924257A (en) | 2009-06-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106842045B (en) | Battery multi-model fusion modeling method and battery management system based on self-adaptive weight method | |
She et al. | Offline and online blended machine learning for lithium-ion battery health state estimation | |
CN107066722B (en) | Electrochemical model-based combined estimation method for state of charge and state of health of power battery system | |
CN109143083B (en) | data-driven lithium ion battery lithium analysis diagnosis method for electric vehicle | |
CN104535932B (en) | Lithium ion battery charge state estimating method | |
Badeda et al. | Battery state estimation for lead-acid batteries under float charge conditions by impedance: Benchmark of common detection methods | |
TWI353076B (en) | Method for estimating residual capacity of lead-ac | |
Li et al. | On-line estimation method of lithium-ion battery health status based on PSO-SVM | |
Zhang et al. | Accurate prediction approach of SOH for lithium-ion batteries based on LSTM method | |
CN110210071B (en) | All-vanadium redox flow battery energy storage system operation optimization method based on genetic algorithm | |
CN113917334A (en) | Battery health state estimation method based on evolution LSTM self-encoder | |
Savard et al. | Accessible battery model with aging dependency | |
Zine et al. | Experimentally validated coulomb counting method for battery state-of-charge estimation under variable current profiles | |
Lei et al. | A novel temperature–hysteresis model for power battery of electric vehicles with an adaptive joint estimator on state of charge and power | |
Sadegh Kouhestani et al. | Prognosis and health management (PHM) of solid-state batteries: Perspectives, challenges, and opportunities | |
CN116165563A (en) | Ship power lithium battery health grade identification method based on neutron imaging | |
Shamim et al. | Valve regulated lead acid battery evaluation under peak shaving and frequency regulation duty cycles | |
He et al. | Short-term prediction of remaining life for lithium-ion battery based on adaptive hybrid model with long short-term memory neural network and optimized particle filter | |
Xiong et al. | A novel nonlinear decreasing step-bacterial foraging optimization algorithm and simulated annealing-back propagation model for long-term battery state of health estimation | |
Singh et al. | State of charge estimation techniques of Li-ion battery of electric vehicles | |
Tian et al. | Consistency evaluation of electric vehicle battery pack: multi-feature information fusion approach | |
Wang et al. | An efficient state-of-health estimation method for lithium-ion batteries based on feature-importance ranking strategy and PSO-GRNN algorithm | |
Xiong et al. | Improved cooperative competitive particle swarm optimization and nonlinear coefficient temperature decreasing simulated annealing-back propagation methods for state of health estimation of energy storage batteries | |
Wang et al. | State of charge estimation for lithium-bismuth liquid metal batteries | |
Zhang et al. | A comparative study of the LiFePO4 battery voltage models under grid energy storage operation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
MM4A | Annulment or lapse of patent due to non-payment of fees |