200939055 九、發明說明: 【發明所屬之技術領域】 本發明是有關於一種預測設施與設備故障的方法,且 特別疋有關於-種使用模糊類神經網路預測設施與 零件故障的方法。 /、叹请 【先前技術】 © %施與設備中若有損耗性零件,則每隔-段時間就必 須進行維修或更換,然而,零件的使用壽命因環境及材料 的特!·生而有所不同。一般來說,零件的使用壽命受到一些 使用因素的影響,而且在不同的情況下,各個影響因素的 f要性會有所改變’為避免設施與設備的零件於維護保養 月以貝壞或耗盡’影響其應有之功能,因此如能進—步主動 預測設施設備之零件可能損壞或耗盡之時間,即可防範於 未然。 、 【發明内容】 本發明的目的是在提供-種預測設施與設備故障的 法,以準確地估測設施設備零件的使用壽命。 、本發月較佳實施例,提出一種預測設施與設備 一障的方:¾:包含選擇—指定零#,係選擇—設*設備中之 零件作為指定零件。於選定了設施設備中之指定零件 則選擇會料設施設備+之指定料使用壽命的因 ,、,並將每一影響因素作為輸入變數。 200939055 心據所採用的影響因素4義模糊規則,用以預 〜叹備之指定零件較用壽命,模糊規則定義為 正'、娜ΧιΙ_ 勸 _XnISf^mENyISCj ^2^^模糊__路中設定的錄,即影響 指定零件使料命的因素,將參數設定其為指定Ϊ 件使用時間指定零件的❹頻率,^指定零件使 ❹200939055 IX. INSTRUCTIONS: TECHNICAL FIELD OF THE INVENTION The present invention relates to a method for predicting facility and equipment failures, and in particular to a method for predicting facility and component failure using a fuzzy neural network. /, sigh [previous technology] © % If there are lossy parts in the equipment, it must be repaired or replaced every other time. However, the service life of the parts is due to the environment and materials. Different. In general, the service life of parts is affected by some factors of use, and in different situations, the nature of each influencing factor will change 'to avoid the damage of the parts of the facility and equipment in the maintenance month. Doing 'affects the function that it should have, so if you can take the initiative to predict the time when the parts of the facility equipment may be damaged or exhausted, you can prevent it from happening. SUMMARY OF THE INVENTION It is an object of the present invention to provide a method for predicting facility and equipment failures to accurately estimate the service life of facility equipment components. In the preferred embodiment of the present month, a method for predicting facilities and equipment is proposed: 3⁄4: includes selection—specify zero#, and selects—sets the part in the device as the designated part. If the specified part in the facility equipment is selected, the factor of the service life of the facility equipment + is selected, and each influencing factor is taken as the input variable. 200939055 The influential factors used in the mental data are used to predict the useful life of the specified parts. The fuzzy rules are defined as positive ', NaΧιΙ_ persuasion_XnISf^mENyISCj ^2^^fuzzy__in the road setting Recording, that is, the factor that affects the life of the specified part, set the parameter to specify the part frequency of the specified part, and specify the part.
用的間隔時間’ X4為溫度’ &為澄度。〜W為隸屬函數, _y為根據各項影響因紊 的使用壽命。素所產生之結果,即為⑽預測零件 再根據模构規則,建立一個模糊類神經網路來修正此 模糊規則。其中,模糊類神經網路的建立方法包含決定— 輸入層、決定-模糊層、使用一介面層及使用一輸出層。 輸入層包含複數個神經元,神經元之個數對應輸入變數之 個數。模糊層包含複數個群組,每―個群組包含複數個神 經元,每-個群組中之每__個神經元代表—個模糊隸屬函 數。介面層包含複數個神經元,其係將模糊層之輸出值相 乘後輸出,輸出層接收介面層之輸出值並將其相乘後輸 出0 接著,依據一資料庫中之歷史維護資料,對模糊類神 經網路進行訓練,以建立指定零件之使用壽命預測模式。 然後,訓練完成之模糊類神經網路,即可用以進行預測設 施設備中之指定零件可能之使用壽命。 由上述本發明較佳實施例可知,應用本發明可防範設 施设備於維§蔓保養前即損壞或耗盡,而造成設施設備無法 200939055 運作,導致例行工作停止所產生的損失。 【實施方式】 、在本實施例中,所謂的設施與設備包含設施、設備與 像俱一大。卩为,設施係可指一楝建築或地方為了提供特定 的服務或特定的卫業用途;設備可為—項卫作或服務所需The interval used is 'X4 is the temperature' & ~W is a membership function, and _y is the service life of the dysfunction according to each influence. The result of the prime is (10) predicting the part and then based on the modular rules, a fuzzy neural network is established to correct the fuzzy rule. Among them, the establishment method of the fuzzy neural network includes the decision-input layer, the decision-fuzzy layer, the use of an interface layer, and the use of an output layer. The input layer contains a plurality of neurons, and the number of neurons corresponds to the number of input variables. The blur layer contains a plurality of groups, each group contains a plurality of neurons, and each __ neurons in each group represents a fuzzy membership function. The interface layer includes a plurality of neurons, which multiply and output the output values of the fuzzy layer, and the output layer receives the output values of the interface layer and multiplies them to output 0. Then, according to the historical maintenance data in a database, The fuzzy neural network is trained to establish a life prediction mode for the specified part. The trained fuzzy neural network can then be used to predict the likely lifetime of the specified part in the facility. It can be seen from the above preferred embodiments of the present invention that the application of the present invention can prevent the damage or depletion of the facility equipment before the maintenance of the maintenance vessel, and the facility equipment cannot operate 200939055, resulting in the loss of routine work. [Embodiment] In the present embodiment, the so-called facilities and equipment include facilities, equipment, and the like. In this case, a facility may refer to a building or place in order to provide a specific service or a specific industrial use; the equipment may be required for the maintenance or service
要的=具;傢倶可為家庭所用的器具,如桌椅、櫥櫃等。 1參照帛1圖’其、♦示&照本發明-較佳實施例的- 預測认施與&備故障的方法之流程圖^預測設施與設備 故障的方法200包含選擇一指定零件,如步驟21〇。選定 影響指定零件使用壽命的因素,如㈣220。定義-預測 指定零件使用壽命之模糊規則,如步驟230。建立-模糊 類神經網路’如步驟24G。訓練此模糊類神經網路,如步 驟250。推估設施設備中之指定零件可能之使用壽命,如 步驟260。 選擇-指定零件,如步驟21〇,係選擇一設施設備中 :零件作為指定零件。由於設施設備有許多的零件所構 、因此,-開始需選定對哪_個零件進行使用壽 施設射錢了—氮氣瓶料轉為例,在一 ,隨著使用的時間長短、使用的次數等因 時會耗ΐ定氮氣何時會耗盡。為了預測此氮氣瓶中氮氣何 :會:二因此選擇此氣氣瓶作為指定零件,來對其進行 便用哥命(即何時需更換)的預測。 選定影響指定零件使用壽命的因素,如步驟220。於 200939055 選定了設施設備中之指定零件後,必須選擇會影響設施設 備中之指定零件使用壽命的因素,並將每一影響因素作為 T輸入變數。在本實施例中,選定了指定零件使用時間、 指定零件的使用頻率、指定零件使用的間隔時間、溫度以 及溼度等,作為設施設備中之指定零件使用壽命的影響因 素。 根據所採用的影響因素,定義一預測指定零件使用壽 © 命之模糊規則,如步驟230 〇本實施例使用模糊類神經網 路作為預測方法,透過模糊類神經的學習模式,藉由歷史 資料自動調整各個影響因素的重要性’進而提升預測準確 度,預測指定零件(即設施設備零件)使用壽命之模糊規 則定義如式(1): IF X, IS AND x2 IS μ^) AND.·.·.·AND xn IS μη(χη) THENy IS Cj ( 1 ) 其中A,..·.A為模糊類神經網路中設定的參數,即影響指定 零件使用壽命的因素,將參數設定為A為指定零件使用時 φ 間,^為指定零件的使用頻率,七為指定零件使用的間隔 時間,&為溫度,々為溼度。…仏)為隸屬函數,少為根據 各項影響因素所產生之結果,即為c,所預測零件的使用壽 命°在本實施例中隸屬函數採用三角形隸屬函數。 參照第2圖’其續·示為第1圖中之模糊類神經網路之架 構圖。模糊類神經網路為結合模擬人類思考模式與人類神 經運作原理的人工網路,該網路同時具有人類思考與學習 的特性。根據式(1)建立一個模糊類神經網路來修正這 些模糊規則。在本實施例中,模糊類神經網路包含四個階 8 200939055 層’此四個階層分別為:輸入層(Input Layer ) 310、模糊 層(Fuzzify Layer) 320、介面層(Intermediate Layer) 330 及輸出層(Output Layer) 340,且其架構為一個含有j條模 糊規則的模糊類神經網路,將其進一步說明如下: 輸入層310包含η個神經元311,每一個神經元311配合 一個輸入變數。神經元311在接收輸入變數之訊號後,直接 傳遞給下一層(即模糊層320)中與此神經元311相連接之 其他的神經元。因此,輸入層310中第i個神經元的輸出值 為of >,其可表示如式(2 ): °,(1) = Pi for \ <ί<η ( 2 ) 輸入層310的神經元個數為輸入變數個數,亦即在輸 入層310中直接將輸入值凡傳送至模糊層32〇中,不做任何 改變。 模糊層320中共有J個群組321,每個群組321包含n個神 經元322,每個神經元322代表一個模糊隸屬函數,可分別 以式(3)、式(4)、式(5)加以表示。模糊層320中之神 經元322扮演著將輸入的數值轉換成隸屬度的角色。第』的 群組321中之第i個的神經元的輸出值為〇^,其為#所對應 的隸屬函數。模糊層神經元個數為輸入變數與模糊語意之 積。 請參照第3圖,其繪示為第2圖中之模糊類神經網路之 模糊層所使的三角形隸屬函數之示意圖。在本實施例中所 使用之隸屬函數定義如第3圖所示’其分別由三個模糊隸 屬函數所組成,分別以式(3 )、式(4 )、式(5 )加以表 200939055 示如下。 1 when x<2 °1=* (4-jc) when 2<x<4 2 (: 0 when x>4 1 when x<3 (λ: —3) when 3<x<5 2 2 °2j =< --— 2 when 5<x<7 (4) ^ 0 when x >7 0 when x <6 °ij =< (^-6) when 6 < x < 8 2 (. 5) 1 when λ: > 8If you want to use it, you can use it for household use, such as tables, chairs, cabinets, etc. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flow chart of a method for predicting a facility and equipment failure, and a method for predicting facility and equipment failures, including selecting a designated part, As in step 21〇. Select factors that affect the service life of the specified part, such as (4) 220. Definition - Forecast Specify the fuzzy rules for the life of the part, as in step 230. Establish a -fuzzy neural network as in step 24G. Train this fuzzy neural network, as in step 250. Estimate the possible service life of the specified part in the facility equipment, as in step 260. Select - specify the part, as in step 21, select a part in the facility: part as the specified part. Since the facilities and equipment have many parts, therefore, it is necessary to select which parts to use for the life-selling. The nitrogen bottle is turned into an example, one, with the length of use, the number of uses, etc. It will take time to determine when the nitrogen will be exhausted. In order to predict the nitrogen in the nitrogen bottle, it will: Secondly, the gas cylinder is selected as the designated part to predict the life of the gas (ie when it needs to be replaced). Select the factors that affect the service life of the specified part, as in step 220. After selecting the specified parts in the facility equipment in 200939055, you must select the factors that will affect the service life of the specified parts in the facility equipment, and use each influencing factor as the T input variable. In the present embodiment, the specified part use time, the frequency of use of the specified part, the interval time of the specified part, the temperature, and the humidity are selected as the influence factors of the service life of the specified part in the facility equipment. According to the influencing factors used, a fuzzy rule for predicting the use of the specified parts is defined, as in step 230. This embodiment uses a fuzzy neural network as a prediction method, through the fuzzy neural learning mode, by historical data automatically. Adjusting the importance of each influencing factor' to improve forecast accuracy and predicting the fuzzy rules for the service life of a specified part (ie, facility equipment parts) is defined as Equation (1): IF X, IS AND x2 IS μ^) AND.·.· .·AND xn IS μη(χη) THENy IS Cj ( 1 ) where A, ..·.A is the parameter set in the fuzzy neural network, that is, the factor that affects the service life of the specified part. Set the parameter to A for the specified When the part is used, φ, ^ is the frequency of use of the specified part, seven is the interval time of the specified part, & is temperature, 々 is humidity. ...仏) is a membership function, which is less than the result of each influencing factor, that is, c, the life of the predicted part. In this embodiment, the membership function uses a triangular membership function. Referring to Fig. 2, the continuation thereof is shown as a frame of the fuzzy neural network in Fig. 1. The fuzzy-like neural network is an artificial network that combines the simulation of human thinking patterns with the principles of human neural operation. The network also has the characteristics of human thinking and learning. A fuzzy neural network is established according to equation (1) to correct these fuzzy rules. In this embodiment, the fuzzy neural network includes four levels 8 200939055 layers. The four layers are: an input layer 310, a fuzzify layer 320, an intermediate layer 330, and The output layer (Output Layer) 340 is constructed as a fuzzy-like neural network containing j fuzzy rules, which is further described as follows: The input layer 310 includes n neurons 311, each of which cooperates with an input variable . After receiving the signal of the input variable, the neuron 311 passes directly to the other neurons in the next layer (i.e., the blur layer 320) that are connected to the neuron 311. Therefore, the output value of the i-th neuron in the input layer 310 is of >, which can be expressed as the equation (2): °, (1) = Pi for \ < ί < η ( 2 ) The nerve of the input layer 310 The number of elements is the number of input variables, that is, the input values are directly transferred to the blur layer 32〇 in the input layer 310 without any change. There are a total of J groups 321 in the fuzzy layer 320. Each group 321 includes n neurons 322, and each neuron 322 represents a fuzzy membership function, which can be expressed by equations (3), (4), and (5, respectively). ) to express it. The neuron 322 in the blur layer 320 acts as a character that converts the input value into membership. The output value of the i-th neuron in the group 321 of the 』 is 〇^, which is the membership function corresponding to #. The number of fuzzy layer neurons is the product of input variables and fuzzy semantics. Please refer to FIG. 3, which is a schematic diagram of a triangular membership function caused by the fuzzy layer of the fuzzy neural network in FIG. The membership function definitions used in this embodiment are as shown in Fig. 3, which are respectively composed of three fuzzy membership functions, which are represented by equations (3), (4), and (5), respectively. . 1 when x<2 °1=* (4-jc) when 2<x<4 2 (: 0 when x>4 1 when x<3 (λ: —3) when 3<x<5 2 2 °2j = < ---2 when 5<x<7 (4) ^ 0 when x >7 0 when x <6 °ij =< (^-6) when 6 < x < 8 2 (. 5 ) 1 when λ: > 8
❹ 凊參照第2圖’介面層330包含J個神經元331,每個神 經元331均代表每一個規則,神經元33 1個數與規則數相 同’且與模糊層320中的相對應神經元連結。這一層相當 於規則中的前提(if…)部份,神經元331可計算輸入變數 與規則的符合程度(Matching degree)。每個神經元331的 輸出值0ί3)為將所有的輸入值相乘’其可表示成式(6 ): "面層330神經元個數採用運用最廣泛之法則決定, 計算如式(7)所示。 其中為介面層310神經元個數’ TV為模糊層320神經元 ’乂為輸出層340神經元個數。 200939055 輸出層340僅包含一個神經元341,其對應一個輸出變 數卩可砲故障之時間。將輸入值相乘即為輸出層結果: 、 (8) 此層的神、玉元負責將輸出值進行解模糊,以獲得一明 確(Crisp)輸出值。 接著’對模糊類神經網路進行訓,練,係依據一資料庫 中之過去維護的歷史資料,執行錯誤倒傳遞演算法 (Error 〇 backpropagation)訓練模糊類神經網路,並建立零件使用 壽命推估模式。錯誤倒傳遞演算法係採用最陡坡降法來進 行均訓練’以使得能量函數(誤差函數)能夠到達最小化。 訓練完成之模糊類神經網路,即可用以進行推估設施 設備中之指定零件可能之使用壽命。在本實施例中,會將 預測設施與設備故障的方法2〇〇建立成一「預測故障模 組」,並架構於一資訊系統中。「預測故障模組」包含四個 子功能選項,分別為管理專案功能選項,係提供使用者新 ◎ 增或編輯專案;設定隸屬函數功能選項,係提供使用者設 定隸屬函數分布;訓練專案功能選項,係提供使用者指定 專案進行模糊類神經網路訓練;以及預測故障功能選項, 係提供使用者選擇專案後,推估設施設備零件的壽命。 其中’管理專案功能選項包含兩個子功能,分別為新 增專案功能選項以及編輯專案功能選項。新增專案功能選 項係提供使用者輸入新增專案檔名與各項參數,如:預測 零件名稱,預測因子及其最大最小值範圍、介面層神經元 數量等。編輯專案功能選項,係提供使用者對已存在的專 200939055 案進行編輯或修改,來修改現有專案内容。 於操作「預測故障模組」時,藉由使用新增專案功能 選項,提供使用者選擇特定零件並儲存為指定名稱,且可 選擇影響零件壽命因子與進行設定介面層神經元數量。於 新增專案時,系統會自動產生隸屬函數分布,此外,使用 者亦可使用設定隸屬函數功能修改隸屬函數分布。 訓練專案功能選項為依據過去維護歷史資料,提供使 Ο 用者運用模糊類神經網路建立零件壽命推估模式。預測故 障功能選項係提供使用者運用訓練完成之專案,推估指定 零件可能之壽命,提供作為維護管理之參考。 「預測故障模組」之操作方式為使用者可針對特定設 施設備之零件建立新專案,或依據需求修改過去專案,^ 案新建或修改後可依據歷史資料訓練模掏類神經網路,訓 練結果可用於推估設施設備零件可能的壽命。 雖然本發明已以一較佳實施例揭露如上,然其並非用 〇 以限定本發明,任何熟習此技藝者’在不脫離本發明之精 神^範圍内,當可作各種之更動與濁飾,因此本發明之保 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】 為讓本發明之上述和其他目的、特徵、優點與實施例 能更明顯易懂,所附圖式之詳細說明如下: 第1圖係綠示依照本發明一較佳實施例的一種預測設 施與設備故障的方法之流程圖。 12 200939055 第2圖係繪示為第1圖中之模糊類神經網路之架構 圖。 第3圖係繪示為第2圖中之模糊類神經網路之模糊層 所使的三角形隸屬函數之示意圖。 310 :輸入層 320 :模糊層 322 :神經元 331 :神經元 341 ··神經元❹ 凊 Referring to Figure 2, the interface layer 330 contains J neurons 331, each of which represents each rule, the number of neurons 33 is the same as the number of rules' and the corresponding neurons in the fuzzy layer 320 link. This layer is equivalent to the premise (if...) part of the rule, and the neuron 331 can calculate the matching degree of the input variable to the rule (Matching degree). The output value of each neuron 331 is 0 ί3) to multiply all input values, which can be expressed as equation (6): "The number of neurons in the surface layer 330 is determined by the most widely used rule, and the equation is as follows (7) ) shown. The number of neurons in the interface layer 310 'TV is the fuzzy layer 320 neurons 乂 is the number of neurons in the output layer 340. 200939055 The output layer 340 contains only one neuron 341, which corresponds to the time at which an output variable can fail. Multiplying the input values is the result of the output layer: (8) The gods and jade elements of this layer are responsible for defuzzing the output values to obtain a clear (Crisp) output value. Then, the training and practice of the fuzzy neural network is based on the past historical data maintained in a database, and the error 〇backpropagation is used to train the fuzzy neural network and establish the life of the part. Estimate the model. The error back-transfer algorithm uses the steepest slope method to perform the training so that the energy function (error function) can be minimized. A trained fuzzy-like neural network can be used to estimate the likely life of a given part of a facility. In this embodiment, the method 2 for predicting facility and equipment failure is established as a "predictive failure model" and is constructed in an information system. The "Predictive Fault Module" consists of four sub-function options, which are management project function options, which provide users with new or enhanced projects; set membership function options, provide user-defined membership function distribution; training project function options, The user-specified project is provided for fuzzy neural network training; and the predictive fault function option is provided to provide the user with the option to estimate the life of the facility equipment parts. The 'Manage Projects' option has two sub-features, the New Project feature option and the Edit Project feature option. The new project function option provides the user with the input of the new project file name and various parameters, such as the predicted part name, the predictor and its maximum and minimum range, and the number of interface layer neurons. Edit project options, which provide users with the ability to edit or modify existing 200939055 cases to modify existing project content. When operating the "Predictive Fault Module", the user is selected to select a specific part and stored as a specified name by using the Add Project Function option, and optionally affects the part life factor and sets the number of interface layer neurons. When a new project is added, the system automatically generates a membership function distribution. In addition, the user can also modify the membership function distribution by using the set membership function. The training project function option is based on past maintenance history data and provides a means for the user to use the fuzzy neural network to establish a part life estimation model. The predictive fault function option provides the user with a training-completed project to estimate the likely life of a given part and provide a reference for maintenance management. The operation mode of the "predictive fault module" is that the user can create a new project for the parts of the specific facility equipment, or modify the past project according to the requirements. After the new or modified case, the model neural network can be trained according to the historical data, and the training result Can be used to estimate the possible life of a facility's equipment parts. Although the present invention has been described above in terms of a preferred embodiment, it is not intended to limit the invention, and any skilled person can make various changes and neglects without departing from the spirit of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS The above and other objects, features, advantages and embodiments of the present invention will become more <RTIgt; A flow chart of a method for predicting facility and equipment failures. 12 200939055 Figure 2 is a diagram showing the architecture of the fuzzy neural network in Figure 1. Figure 3 is a schematic diagram showing the triangular membership function of the fuzzy layer of the fuzzy neural network in Figure 2. 310: Input layer 320: Blur layer 322: Neuron 331: Neuron 341 · Neurons
【主要元件符號說明】 200 :方法 210〜260 :步驟 311 :神經元 321 :群組 330 ··介面層 340 :輸出層[Main component symbol description] 200: Method 210 to 260: Step 311: Neuron 321 : Group 330 · Interface layer 340: Output layer
1313