201224823 p 六、發明說明： 【發明所屬之技術領域】 本發明係有關於一種基於影像之製程監控與診斷方 法，且特別有關於一種基於火焰影像之燃燒製程監控與異 常現象及原因診斷方法。 【先前技術】 锅爐系統係為目前化工廠、電廠或傳統製造工業中， 製程生產動力與熱能的來源，然而，受到近幾年國際油價 波動與環保意識抬頭的影響，以及對於工業安全的日益重 視，發展更有效率、排放氣體更符合環保標準、以及操作 更具安全性的燃燒監控系統，已成為鍋爐設備開發以及燃 燒製程監控的重要議題。 類似如鍋爐系統的工業燃燒系統，其運作的基本要求 在於建立穩定的燃燒火焰，不穩定的火焰通常導因於不好 的燃燒條件設定或動態控制。燃燒不穩定會降低熱效率， 也會引起爐膛熄火，甚至導致爆炸事故。 目前的預防實施方法是透過火檢器來判斷燃燒器是否 熄火以啟動燃燒系統保護裝置避免事故發生。然而，常見 的輻射光能式（UV/VIS/IR)火檢器只能檢測有無火焰，並需 設定靜態的門閥值且不具空間資訊，因此容易發生誤警 報，改良式的數位式火檢器仍無法有效解決因火焰飄移所 導致火焰有無的誤判斷，較新穎的作法是利用燃燒影像及 微處理器計算能力之全爐膛圖像式監測系統，但全爐膛圖 201224823 像式監測系統是對整個爐膛燃燒狀況進行判斷或監控，無 法識別火焰是否存在和其穩定性。 由於使用者尋求燃燒安全輔助感測系統之市場需求， 加上美國能源局於工業燃燒技術發展藍圖中也明白指出 『火焰穩定性感測器』之研發需求，因此為了產業需求與 延續火檢系統之技術發展趨勢，本專利開發了圖像式智能 化火焰診斷系統，可提供燃燒火焰之燃燒狀態資訊及即時 尾氣濃度資訊，並提供燃燒穩定性診斷和燃燒效率評估之 功能，在技術開發上，以減少不容易決定之人為參數設定 為目標，此專利技術讓燃燒火焰資訊透明化並具多功能 性，可以最小化成本與最大化使用效益。 【發明内容】 基於上述目的，本發明提供一種燃燒火焰診斷方法， 係先利用一個影像擷取裝置，得到執行一燃燒製程的一爐 膛内之一原始影像，其中該原始影像包含一火焰影像與一 背景影像。接著，利用一火焰識別技術將該原始影像中的 該火焰影像與該背景影像分離。之後，計算該火焰影像之 特徵。再依據該火焰影像的特徵，診斷該爐膛之穩定性與 燃燒效率，以監控該燃燒製程。 為使本發明之上述和其他目的、特徵、和優點能更明 顯易懂，下文特舉出實施例，並配合所附圖式，作詳細說 明如下： 201224823 【實施方式】 。 r方說明書提供不同的實施例來說明本發明不同實 ，方式的技術特徵。其中，實施例中的各元件之配置= 發明。且實施例中圖式標銳之 關聯性 了簡化說明，並非意指不同實施例之間的 診斷像之燃燒製程監趣與 ⑽—-個外拍控與診 溫環境下捕捉爐腔内的火焰影像，並 i 1置’於高 得到的特徵來進行火焰燃燒監控 析火焰影像所 態。 亥爐膛内的燃燒狀 第1圖顯示依據本發明實施例之 施本專利方法之硬體架構說明如圖:木構不意圖。實 土統η、-高溫攝影機13、一現場電；^二= 應用電腦（Application PC ) 17。 其中’燃燒系統11具有一爐膛。其、w 為外掛式或嵌入式高溫攝影機，其係用^攝影機13可以201224823 p VI. Description of the Invention: [Technical Field of the Invention] The present invention relates to an image-based process monitoring and diagnosis method, and particularly relates to a flame image-based combustion process monitoring and anomaly phenomenon and a cause diagnosis method. [Prior Art] Boiler system is the source of power and heat energy for process production in chemical plants, power plants or traditional manufacturing industries. However, it has been affected by fluctuations in international oil prices and environmental awareness in recent years, as well as the growing industrial safety. Paying attention to the development of more efficient, more environmentally friendly emission standards and safer combustion monitoring systems has become an important issue in boiler equipment development and combustion process monitoring. Similar to industrial combustion systems such as boiler systems, the basic requirement for operation is to establish a stable combustion flame, which is often caused by poor combustion conditions or dynamic control. Unstable combustion reduces thermal efficiency and can cause the furnace to stall and even cause an explosion. The current preventive implementation method is to use a fire detector to determine whether the burner is turned off to activate the combustion system protection device to avoid an accident. However, the common radiant light energy (UV/VIS/IR) fire detector can only detect the presence or absence of flame, and it is necessary to set a static threshold value without spatial information, so it is prone to false alarms. The improved digital fire detector Still can not effectively solve the misjudgment of the flame caused by the flame drift, the new method is to use the combustion image and the computing power of the full furnace image monitoring system, but the full furnace map 201224823 image monitoring system is the whole The burning condition of the hearth is judged or monitored, and the presence or absence of the flame and its stability cannot be recognized. As the user seeks the market demand for combustion safety assisted sensing systems, and the US Energy Administration also clearly pointed out the research and development requirements of the "flame stability sensor" in the industrial burning technology development blueprint, for the industrial demand and the continuous fire detection system Technology development trend, this patent developed an image-based intelligent flame diagnosis system, which can provide combustion state information of combustion flame and instantaneous exhaust gas concentration information, and provide combustion stability diagnosis and combustion efficiency evaluation function. In technology development, The goal is to reduce the number of people who are not easily determined. This patented technology makes the information of the combustion flame transparent and versatile, which can minimize the cost and maximize the use efficiency. SUMMARY OF THE INVENTION Based on the above objects, the present invention provides a method for diagnosing a combustion flame by first using an image capture device to obtain an original image of a furnace in which a combustion process is performed, wherein the original image includes a flame image and a Background image. The flame image in the original image is then separated from the background image using a flame recognition technique. After that, the characteristics of the flame image are calculated. Based on the characteristics of the flame image, the stability and combustion efficiency of the furnace are diagnosed to monitor the combustion process. The above and other objects, features, and advantages of the present invention will become more apparent from the embodiments of the invention. The R specification provides different embodiments to illustrate the technical features of the different embodiments of the present invention. Among them, the configuration of each element in the embodiment = invention. In the embodiment, the correlation between the diagrams and the sharpness of the drawings is simplified, and does not mean that the diagnosis between the different embodiments is like the combustion process monitoring and (10) - the external shooting control and the flame in the cavity are captured in the clinic environment. Image, and i 1 set the feature obtained at high to perform flame combustion monitoring to analyze the state of the flame image. The burning pattern in the furnace is shown in the first drawing. The hardware structure of the patented method according to the embodiment of the present invention is shown in the figure: wood structure is not intended. Real earth system η, - high temperature camera 13, a site power; ^ two = Application PC (Application PC) 17. Wherein the combustion system 11 has a furnace. The w is an external or embedded high temperature camera, which can be used with the camera 13
之爐膛内影像。高溫攝影機13所取得取传燃燒系統U 始旦/ /务甘4iS 爐腔内影像包含大 二像，其，取之爐膀内影像則由現場⑽ 加以分析，取付火焰影像並分析火培的特徵（例如燦 k的溫度、燃燒面積、燃燒重心、火培色声 ^ ·、'、 2㈣’進而可以診斷燃燒的狀態: ，、中，現％電腦15和應用電腦]7可以θ、 <分開的電腦系統， 201224823 也可以將現場電腦15和應用電腦17合一。 第2圖顯示依據本發明實施例之基於影像之燃燒製程 監控方法之流程圖。 步驟S21中，取得爐膛之原始影像。例如，利用一影 像擷取裝置例如是高溫攝影機，拍攝爐膛内部，以得到執 行燃燒製程的爐膛之一原始影像，其中原始影像包含一火 焰影像與一背景影像。 當影像擷取裝置擷取到原始影像之後，步驟S23中透 • 過一火焰識別技術將原始影像中的火焰影像與背景影像分 離（詳見第3圖），然後在步驟S25中計算火焰影像之特 徵，以作為後續監控及診斷的依據。在步驟S27中根據這 些火焰影像之特徵，診斷爐膛的穩定性和燃燒效率，最後 再依據步驟S27的診斷結果，於步驟S29中發出對應的警 報訊息。細部實施方法於後續詳細說明。 如上述，步驟S25之火焰特徵計算在考量減少人為參 數設定和影像擷取裝置限制情況下，火焰影像的特徵包括 ® 色彩資訊和幾何資訊兩大類。其中色彩資訊包含：以統計 分析得到的亮度值平均、亮度值變異、亮度峰態、亮度值 偏態、亮度熵值、均勻度、平均溫度等；或以輻射學方法 得到的溫度場計鼻貧訊，以及以動悲（頻譜）分.析得到的 火焰閃爍頻率資訊等其中之一。其中幾何資訊包含：火焰 分佈相關資訊及空間分佈等。火焰分佈相關資訊可以為火 焰長、寬與火焰喷射角度、火焰區域面積、火焰質量重心 位置等其中之一。而空間分佈則可以為2D-FFT、 201224823 2D-Wavelet等其中之一。此外，另可針對火焰影像之不同 區域之對比進行相對特徵計算，如：選擇區域之面積比例、 選擇區域之亮度值比例、選擇區域之亮度變異比值、火焰 燃燒區域能量、火焰内部與火焰全區面積比例等其中之 一，詳細計算方法請參考附件五。 第3圖顯示第2圖中的步驟S23之火焰識別方法流程 圆。 在本實施例中，列舉感官式火焰識別方法來做說明， 火焰識別方法也可利用HSV色彩模型或RGB高斯混合模 型等色彩模型來完成，也可利用門閥值法，並不以此為限。 感官式火焰識別方法係利用HSV色彩空間、色彩學和模糊 理論來建構燃燒影像火焰分割之技術。應用感官式火焰識 別技術的動機在於：（1)爐内環境單純；（2)特定燃料種類；（3) 具有不同之光學濾鏡；（4)要最小化人為設定。如果是應用 傳統閾值分割方法或RGB高斯混合模型來建構火焰的色 彩模型，則需要先以一張或複數張影像進行建模。而且， 影像擷取裝置可選用不同波長之濾鏡，使得收集之原始影 像會有色差。為了符合實際應用之限制與最小化人為設定 之目標，因此提出如第3圖所示之感官式火焰識別方法。 本發明之感官式火焰識別方法係基於人類視覺對色彩 直觀認知，將火焰色彩轉換至HSV空間，結合色彩學和模 糊演算法，智能化考量HSV，以迅速將火焰影像切割出來， 對於火焰影像而言，Η可代表不同燃燒物種所釋出之顏 色，S可代表不同溫度所呈現之飽和度，V代表煙所產生 201224823 不同的灰程度。 . 參見第3圖，在步驟伽中接收由影像 爐腔内得到的複數個原始影像（相同於步驟S21)。攝 為了節省火焰朗的計算㈣，可 影像中的特定區域作為分析對象。第3圖的二:; S303 ^slT2 ^ 二1 :人之後，再執行步驟，以選定原 定區域作為分析對象，進而節省火焰識別的 別，以產生複數個彩色物件。^識 =物件。舉例而言’爐壁或積灰通常呈二黑色i灰: 義為非感興 之色彩物件，產生單:!=，根據所找出的非感興趣 中矩陣的大小盥原始旦^+感興趣遮罩(mask)矩陣，其 個或複數個影像之二於步驟S504，儲存- 算得：=非感興趣遮草矩陣進行矩陣邏輯運 與步驟S3〇1 ;同也就是㈣8。步驟_， 複數個連續火焰之朴^像^料置減爐膛内得到的 個原始影像中進行隨機^，如^驟讀，從襲複數 則使用S5〇8〇，果如果不滿足隨機挑選條件， 、’、。果，即融合之感興趣遮罩作為特定區域選 201224823 定，反之，一旦隨機條件成 定為整張原始影像的尺寸區域選定範圍重新設 合的感興趣遮罩無法捕捉。避免攝影環境變化時，所融 本發明利用視覺化色彩物 定分析監測的範圍，如此力能，可以自動化決 學鏡頭表面遭到污染所造成去除影㈣取，光 障礙物錯，增加火㈣還相㈣壁等視覺 效減少火焰分割之計算時=徵#之正確性，更可以有The image inside the furnace. The high-temperature camera 13 obtains the transmission and combustion system U. The image in the furnace chamber contains a large image. The image in the furnace is analyzed by the site (10) to take the flame image and analyze the characteristics of the fire. (For example, the temperature of the can, the burning area, the burning center of gravity, the sound of the fire color ^ ·, ', 2 (four)' can then diagnose the state of combustion: ,, in, the current computer 15 and the application computer] 7 can be θ, < separate The computer system, 201224823 can also integrate the on-site computer 15 and the application computer 17. Figure 2 shows a flow chart of the image-based combustion process monitoring method according to an embodiment of the invention. In step S21, the original image of the furnace is obtained. Using an image capturing device such as a high temperature camera to photograph the inside of the furnace to obtain an original image of the furnace performing the combustion process, wherein the original image includes a flame image and a background image. When the image capturing device captures the original image Then, in step S23, the flame image is separated from the background image by a flame recognition technology (see Figure 3), and then in step The feature of the flame image is calculated in S25 as a basis for subsequent monitoring and diagnosis. In step S27, the stability and combustion efficiency of the furnace are diagnosed according to the characteristics of the flame images, and finally, according to the diagnosis result of step S27, in step S29. Corresponding alarm message is sent. The detailed implementation method is described in detail later. As described above, the flame feature calculation in step S25 is based on the consideration of reducing the artificial parameter setting and the image capturing device limitation, and the features of the flame image include: color information and geometric information. The color information includes: brightness value average obtained by statistical analysis, brightness value variation, brightness peak state, brightness value skewness, brightness entropy value, uniformity, average temperature, etc.; or temperature field meter obtained by radiological method Nasal poor news, and one of the flame flicker frequency information obtained by the sorrow (spectrum) analysis. The geometric information includes: information about the flame distribution and spatial distribution, etc. The information about the flame distribution can be the length and width of the flame. Flame spray angle, flame area, flame mass center position, etc. One, and the spatial distribution can be one of 2D-FFT, 201224823 2D-Wavelet, etc. In addition, relative feature calculation can be performed for comparison of different regions of the flame image, such as: area ratio of selected area, selection area The ratio of the brightness value, the ratio of the brightness variation of the selected area, the energy of the flame combustion area, the ratio of the inside of the flame to the area of the whole area of the flame, etc., please refer to Annex V for the detailed calculation method. Fig. 3 shows the flame of step S23 in Fig. 2. In the present embodiment, the sensory flame recognition method is used for illustration. The flame recognition method can also be completed by using a color model such as an HSV color model or an RGB Gaussian mixture model, and a gate threshold method can also be used. This is limited to the sensory flame recognition method using HSV color space, color theory and fuzzy theory to construct the technology of combustion image flame segmentation. The motivation for applying sensory flame identification techniques is: (1) the furnace environment is simple; (2) specific fuel types; (3) different optical filters; (4) minimizing artificial settings. If you are applying a traditional threshold segmentation method or an RGB Gaussian mixture model to construct a color model of a flame, you need to model one or multiple images first. Moreover, the image capture device can use different wavelength filters, so that the original image collected will have chromatic aberration. In order to meet the limitations of practical applications and to minimize the goal of artificial setting, a sensory flame recognition method as shown in Fig. 3 is proposed. The sensory flame recognition method of the invention is based on the human visual perception of color, transforms the flame color into the HSV space, and combines the color theory and the fuzzy algorithm to intelligently consider the HSV to quickly cut out the flame image for the flame image. In other words, Η can represent the color released by different burning species, S can represent the saturation of different temperatures, and V represents the different ash levels of 201224823 produced by the smoke. Referring to Fig. 3, a plurality of original images obtained from the image chamber are received in the step gamma (same as step S21). In order to save the calculation of the flame (four), a specific area in the image can be analyzed. Figure 3: 2: S303 ^slT2 ^ 2: After the person, the steps are executed to select the original area as the analysis object, thereby saving the flame recognition to generate a plurality of colored objects. ^ 识 = object. For example, 'the wall of the furnace or the ash is usually two black i gray: the color object is not a happy color, produce a single: !=, according to the size of the matrix in the non-interest found, 盥 original ^ ^ + interest cover A matrix of masks, one or more of the plurality of images in step S504, is stored - calculated: = non-interesting matte matrix for matrix logic operation with step S3 〇 1; the same is (four) 8. Step _, a plurality of continuous flames of the image of the continuous flame are reduced in the original image obtained in the furnace, such as ^ sudden reading, and the number of the attacking complex is S5 〇 8 〇, if the random selection condition is not satisfied, , ',. The fusion mask of interest is selected as the specific area 201224823. Conversely, once the random condition is determined to be the size range of the entire original image, the selected range re-set mask of interest cannot be captured. When avoiding changes in the photographic environment, the invention utilizes the scope of visual color analysis to monitor and monitor, so that the power can be automated to remove the shadow caused by contamination of the lens surface (four), light obstacles, increase fire (four) also When the phase (four) wall and other visual effects reduce the calculation of flame splitting, the correctness of the sign can be
作θ:二嗎煤系統中的積灰會影響锅爐的熱傳效率， 氣進行切動作，又會造絲損失， 所以吹灰時機的決定對於_效率簡重要。本發明之自 動化感興趣區域選定方法，若將非感興趣之色彩物件(積灰 視為監控之對象，針對爐灰增生的問題，此方法可以 爐灰增生偵測器，以決定適t的吹灰時機。For θ: the ash in the two-coal coal system will affect the heat transfer efficiency of the boiler, the gas will cut and the wire will be lost, so the decision of the soot timing is important for _ efficiency. The method for selecting an automatic region of interest according to the present invention, if a color object that is not of interest is regarded as an object of monitoring, for the problem of ash ash proliferation, the method can be used to determine the appropriate blowing Gray timing.
在步驟S303中，將原始影像（或特定區域之原始影像） 例如為RGB影像’轉換至Hsv的色彩空間，以產生一 mv 影像。HSV色彩模型可以將亮度和色彩資訊作分離，所以 可以k供如同人類之顏色感知。其中Hs V色彩空間的H表 示色度（hue )，S表示飽和度（saturati〇n)，v則表示明 亮度（value)。將RGB原始影像轉換至HSV的色彩空間 的方法可參見文獻（例如文獻：A. R. Smith, "Color GamutIn step S303, the original image (or the original image of the specific area), for example, the RGB image, is converted to the color space of Hsv to generate an mv image. The HSV color model separates the brightness and color information, so it can be used for human color perception. Where H of the Hs V color space represents hue, S represents saturation (saturati〇n), and v represents brightness (value). The method of converting RGB raw images to the color space of HSV can be found in the literature (eg literature: A. R. Smith, "Color Gamut
Transform Pairs," ACM SIGGRAPH Computer Graphics 12, pp.12-19, 1978.)。 步驟S304中’將原始影像中的每一像素（pixel )對應 10 201224823 於HSV影像執行模糊化程序，以建立模糊集合以及模糊規 貝|J ( Fuzzy Rule )，並區分切割Η、S與V隸屬函數 (Membership Function)之範圍。詳細來說，針對原始影 像中的每一個像素，將對應的HSV影像分別在Η、S與V 色彩空間中區分為多個模糊子集合，並利用這些模糊子集 合建立模糊規則，即分別在Η、S與V色彩空間中各擇一 模糊子集合以建立一條模糊規則，在本實施例中，將HSV 影像分別在Η、S與V色彩空間中區分為10、6與5個的 • 模糊子集合，因此可建立三百條模糊規則，但不以此為限。 此外，在區分切割Η、S與V隸屬函數之範圍時，係計算 隸屬度，即對應Η、S與V色彩空間中的每一模糊子集合 產生一量化的隸屬度。 在步驟S305中，執行模糊邏輯推論，利用建立之模糊 規則，進行顏色的推論，並依據色彩學中人眼對顏色區分 之程度，產生複數個分類結果，可將複數個模糊規則推論 為同一分類結果。詳細來說，依據建立的模糊規則，將每 ® 一模糊規則中的模糊子集合對應的隸屬度相乘，以產生一 推論值，並將這些產生的推論值分類至分類結果中。在本 實施例中，三百條的模糊規則即產生三百個推論值，依據 色彩學之分類，將火焰影像顏色區分為19個分類結果，並 將三百個推論值分類至19個分類結果中。要說的是，此處 所述之三百條規則與19個分類結果，僅爲方便說明之用， 並不限定於此。 在步驟S306中，執行解模糊化過程。為了增加顏色區 201224823 分精確性，本實施例係將三百個推論值分類至19個分類結 果中，並計算每一分類結果的推論值之總和，選出具有最 大值之分類結果作為該原始影像之像素之分類結果，即定 義該像素為具有其分類結果之顏色。 在步驟S307中，由原始影像中分離出火焰影像。經過 上述之步驟S306之後，原始影像中每一像素被定義為具有 一種分類結果之顏色，藉由這些分類結果可得知火焰影像 之區域。在本實施例中，將分類結果區分為19個，除了能 識別火焰内、火焰外圍以及爐壁等區域外，其針對有前處 籲 理之影像（例如加裝濾鏡後所拍攝到的影像，其火焰顏色 可能偏綠）亦可進行火焰影像之區分，而不需要重新建立 色彩模型。 本實施例中的19個分類結果包括：1白色(white)、2 淡灰色（light_grey)、3 深灰色（dark_grey)、4 黑色（black)、5 紅色(red)、6粉紅色(pink)、7深咖啡色（dark—brown)、8淺 咖啡色（light_brown)、9 深橘色（dark_orange)、10 淡橘色 (light_orange)、11 黃色(yellow)、12 橄欖綠(olive)、13 淡 鲁 綠色（light_green)、14 深綠色(dark_green)、15 藍綠色(teal)、 16 水綠色(aqua)、17 藍色(blue)、18 深粉紫色(dark_fucia)、 19淡粉紫色(light_fucia)。上述分類結果係為例示，本發明 實施並不以此為限。 例如’在重油燃燒的爐腫之原始影像中，屬於分類結 果之 1 白色（white)、5 紅色（red)、9 深橘色（dark_orange)、 10淡橘色（light_orange)、11黃色(yell〇w)的區域影像可以 12 201224823 判剛·马火焰影像 油燃燒的濟脖之…像；…機拍霄… 况J墟脛惑原始衫像中，屬於顏色分類】白 ㈣㈣、9深橘色（dark一orange)、1〇淡橘色（】ight⑽够）、 =黃色_〇W)、】3 I綠色(li氣抑叫的區域影像可以判 斷為火焰影像，藉由分類結果可以辨別出火焰影像。 本實施例技術可利用第3圖的感官式火焰識別技術， 为割出不同燃燒區域的火焰影像來定義相對的特徵，Transform Pairs, " ACM SIGGRAPH Computer Graphics 12, pp. 12-19, 1978.). In step S304, 'each pixel (pixel) in the original image corresponds to 10 201224823 to perform a blurring process on the HSV image to establish a fuzzy set and a fuzzy rule |J (Genuary Rule), and distinguish the cut Η, S and V affiliation The scope of the Membership Function. In detail, for each pixel in the original image, the corresponding HSV image is divided into multiple fuzzy subsets in the Η, S, and V color spaces, and the fuzzy rules are used to establish the fuzzy rules, that is, respectively And selecting a fuzzy subset in the S and V color spaces to establish a fuzzy rule. In this embodiment, the HSV images are respectively divided into 10, 6 and 5 in the Η, S and V color spaces. Collection, so you can create three hundred fuzzy rules, but not limited to this. In addition, when distinguishing the extents of the cut Η, S and V membership functions, the membership degree is calculated, that is, each fuzzy subset of the corresponding Η, S and V color spaces produces a quantized membership degree. In step S305, a fuzzy logic inference is performed, and the established fuzzy rule is used to perform color inference, and according to the degree of color distinction of the human eye in the colorology, a plurality of classification results are generated, and the plurality of fuzzy rules can be inferred into the same classification. result. In detail, according to the established fuzzy rule, the membership degrees corresponding to the fuzzy subsets in each fuzzy rule are multiplied to generate an inference value, and the generated inference values are classified into the classification result. In this embodiment, three hundred fuzzy rules generate three hundred inference values. According to the classification of color science, the flame image color is divided into 19 classification results, and three hundred inference values are classified into 19 classification results. in. It should be noted that the three hundred rules and 19 classification results described herein are for convenience of explanation and are not limited thereto. In step S306, a defuzzification process is performed. In order to increase the accuracy of the color zone 201224823, this embodiment classifies three hundred inference values into 19 classification results, and calculates the sum of the inference values of each classification result, and selects the classification result with the maximum value as the original image. The classification result of the pixel, that is, the pixel is defined as the color having the classification result. In step S307, the flame image is separated from the original image. After the above step S306, each pixel in the original image is defined as a color having a classification result, and the area of the flame image can be known by the classification result. In this embodiment, the classification result is divided into 19, in addition to recognizing the inside of the flame, the periphery of the flame, and the wall of the furnace, the image is directed to the frontal image (for example, the image captured after the filter is added). The flame color may be greenish. It can also distinguish between flame images without re-establishing the color model. The 19 classification results in this embodiment include: 1 white, 2 light gray, 3 dark gray, 4 black, 5 pink, 6 pink, 7 dark brown (light-brown), 8 light brown (light_brown), 9 dark orange (dark_orange), 10 light orange (light_orange), 11 yellow (yellow), 12 olive green (olive), 13 light green (light_green ), 14 dark green (green), 15 blue green (teal), 16 water green (aqua), 17 blue (blue), 18 dark pink (dark_fucia), 19 pale pink (light_fucia). The above classification results are exemplified, and the implementation of the present invention is not limited thereto. For example, 'in the original image of the burning of heavy oil, 1 is white (white), 5 red (red), 9 dark orange (dark_orange), 10 light orange (light_orange), 11 yellow (yell〇) w) The area image can be 12 201224823 Judgment of the horse flame image oil burning of the neck ... like; ... machine shot 霄 况 胫 J 胫 原始 original shirt image, belonging to the color classification] white (four) (four), 9 dark orange ( Dark one orange), 1 light orange (] ight (10) enough), = yellow _ 〇 W), 】 3 I green (li gas suppression area image can be judged as a flame image, by classification results can identify the flame image The technique of the embodiment can utilize the sensory flame recognition technology of FIG. 3 to define relative features for cutting out the flame images of different combustion regions.
強調智能化外，也可以有效抗爐膛環境之干擾。 若以重油燃燒為例，當氧氣不足的情況發生時，造成 燃燒成分無法即時反應，導致某些反應向外圍擴散或反摩 不完全，因此燃燒區域可以區分為⑴火焰内部：整個燃 燒最穩^之區域，通常與燃燒負荷有直接相關；（2)火焰 :圍：當火焰趨向不穩定情況，火焰開始會有閃爍 變化等狀況發生。 、 >見附件’其分別顯不在高空燃比情況和低空燃比 情況下之火焰内部和火焰外圍的火焰影像。 … 因為高溫區域是燃燒的骨幹，燃燒溫度越高通常燃燒 越穩疋’因此評量高溫區域和火焰影像的面積比或強度 比’都是姻火_、定性的重要指標，㈣是㈣系統， 火焰均勻度也是反應燃燒狀態好壞的另—個重要來數，通 =均句度越高’火焰間亮度差異越小，燃燒越穩定。在本 貫施例中，為了制本發明可以有效萃取出火焰影像所含 有的重要資訊’因此以固定燃料流量、不同過剩空氣比之 重油燃燒縣來實舰明线魏㈣財數之關聯性。 13 201224823Emphasis on intelligence, it can also effectively resist the interference of the furnace environment. If heavy oil combustion is taken as an example, when oxygen deficiency occurs, the combustion components cannot be reacted instantaneously, causing some reactions to spread to the periphery or incomplete anti-friction, so the combustion zone can be divided into (1) the inside of the flame: the whole combustion is the most stable ^ The area is usually directly related to the combustion load; (2) Flame: Encircling: When the flame tends to be unstable, the flame will start to flicker and other conditions. , > See Annex', which shows the flame image inside and outside the flame in the case of high air-fuel ratio and low air-fuel ratio. ... Because the high temperature area is the backbone of the combustion, the higher the combustion temperature, the more stable the combustion is. Therefore, the area ratio or intensity ratio of the high temperature area and the flame image is an important indicator of the marriage and the qualitative, and (4) is the system. The uniformity of the flame is also another important number of the reaction combustion state. The higher the average degree of the sentence is, the smaller the difference in brightness between the flames is, the more stable the combustion is. In the present embodiment, in order to manufacture the present invention, it is possible to effectively extract the important information contained in the flame image. Therefore, the correlation between the fixed fuel flow rate and the different excess air ratio of the heavy oil burning county to the real ship Mingxian Wei (four) fiscal number. 13 201224823
P 參見附件二之一，其顯示燃燒空氣過剩（左側圖片） 到燃燒空氣不足（右側圖片）的原始影像以及識別後的火 焰影像。 因為煙道氧濃度的取樣量測有時間延遲，所以在本實 施例中先利用相關性分析得知時間延遲約39秒，以利特徵 與氣體濃度的資料同步，參見附件二之二、二之三，根據 火焰影像特徵與尾氣氧濃度之觀察結果，重點整理如下： (1) 火焰影像特徵與燃燒狀態具高度相關性； (2) 内火焰面積幾乎固定，而外火焰面積隨著燃燒空氣 _ 減少而增大，同樣反映在面積比； (3) 操作條件改變時，内外火焰的平均亮度與平均溫度 幾乎不變，但總火焰的亮度和平均溫度，呈現與尾氣濃度 正相關，主要是内外平均時面積變化效應所貢獻； (4) 内外火焰的亮度變異比值反映了燃燒不穩定之趨 勢； (5) 空氣流量影響了整個燃燒火焰的質量中心位置，但 内火焰的Y軸質量中心位置沒有變動。 鲁 第4圖顯示第2圖中步驟S27之診斷爐膛的穩定性和 效率之方法流程圖。 步驟S41中，擷取於步驟S25中取得之火焰影像之特. 徵資訊。 步驟S42中，為了即時診斷燃燒效率，先利用一迴歸 模型，如最小平方法(PLS)或類神經網路，以火焰影像特徵 來建立尾氣濃度之即時預測值，以克服燃燒系統的輸送和 14 201224823 量測時延。即時尾氣預測模型係可以如文獻（例如文獻：H. Yu, J. F. MacGregor, "'Monitoring flames in an industrial boiler using multivariate image analysis,” AICHE Journal 50(7), pp. 1474-1483, 2004.)所述。 在步驟S43中，同時也將火焰影像之特徵透過一自適 應網路模糊推論系統（ANFIS)建立一單張火焰影像的穩 定性診斷結果。詳細來說’火焰影像之特徵輸入自適應網 路模糊推論系統後’會與一歷史事件資料庫進行比對，從 • 而將該火焰影像依燃燒狀態分類，產生單張火焰影像的穩 定性診斷結果。自適應網路模糊推論系統，係利用網路架 構，以達成自動調整隸屬函數，以及自動建立模糊if-then 規則的技術，例如可參考文獻中所述之自適應網路模糊推 論系統（例如文獻：J.-S. Roger Jang，C.T. Sun, E. Mizutani, Neuro-Fuzzy and Soft Computing: A Computational Approach to Learning and Machine Intelligence, Pearson Education Taiwan Ltd, 2004.)。 ® 在步驟S44中，利用當下的一張與前複數張火焰影像 的穩定性診斷結果連續性的透過自適應網路模糊推論系統 (ANFIS)以建立一動態火焰影像的穩定性診斷結果。在 ..本實施例中，係利用當下的一張與前3〜5張火焰影像的穩 定性診斷結果，以連續狀態輸入至自適應網路模糊推論系 統（ANFIS)，來做進一步的動態確認’此步驟可以避免 因為取樣時間和火焰隨機閃爍等因素造成誤診斷，並產生 動態火焰影像的穩定性與強勃性之診斷結果。 15 201224823 最唆於步驟& 和動態火焰旦 a,將複數個尾氣濃度之即時預測值 斷。即遷渦Γ、的穩定性診斷結果進行火焰燃燒效率診 穩定性診：結J應：路模糊推論系統推論動態火焰影像的 斷的指榫，，、，。、'Ή σ尾氣濃度之即時預測值作為效率診 /、以砰估火焰燃燒效率。 在本發明中，脾ό 火焰診斷，a 、將自適應網路模糊推論系統應用於燃燒 裝置的搭配繞火焰本身的閃爍動態和影像擷取 辨燒丰—、，難有決定性的特徵量化標準；（2)各式 :系統和樣本限制，難有完整的專㈣斷準則，因此本 从步驟S43的第一層自適應網路模糊推論系統來自動 建立該燃燒系統之影像特徵模糊量化標準和判斷準則，炎 可作為系統化專家經驗建置之方法，再以步驟S44中的第 一層自適應網路模糊推論系統來作進一步的動態確認，避 Μ為取樣時間和火㉟閃爍等因素造成誤診斷。 以下茲再列舉一實施例，用以進一步說明如何利用本 發明貫施例之燃燒製程監控與診斷方法來進行燃燒製程監 控與沴斷，但並非用以限定本發明。 本實施例以工業蒸氣鍋爐進行測試，其蒸氣容量最大 可達每小時15噸的蒸氣供給，操作條件設定係區分為小、 中、大三種負載。此外對於燃燒不穩定以及熄、火狀態，一 併進行實驗收集與燃燒狀態分析，影像來源則利用外拍式 的高溫攝影機來獲得，測試目標是利用火焰影像來準確地 識別出：i.穩定燃燒（三種負載）、ii.不穩定燃燒以及iii.熄火 等狀態。 201224823 在本實施例中選擇的3種特徵如附件 =別為火焰内部與火焰全區面積之比例、：全、域 度值平均'以及火焰内部之亮度網值 =之冗 ==及識別後的火焰影像外，也物不= 張上火r像的穩定性診斷結二動態= :負载、⑴大負載、⑷不穩定燃壤和(5)負:火二) 心’紅色原點即為燃燒影像的狀態值輸出。 火焰擾動所造成:離w的穩疋性#斷結果可以減少因為 碼的部份’可以藉由程式 碟片、：：々： 包含於實體媒體，如軟碟、光 、 碟、或是任何其他機器可讀取（如電腦可讀取） 成二中’當程式碼被機器，如電腦载入且執行時， 與本發明之裳置。程式碼也可以透過- ㈣送，盆中，繞、光纖、或是任何傳輸型態進 處==成用轉與本發明之裝置。當在-二. 應用特定邏輯電路之獨特裝置。 平作類似於 本么月已以較佳實施例揭露如上，然其並非用以 201224823 限定本發明，任何熟習此技藝者，在不脫離本發明之精神 和範圍内，當可作各種之更動與潤飾，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。 201224823 【圖式簡單說明】 第1圖顯示依據本發明實施例之硬體架構示意圖。 第2圖顯示依據本發明實施例之基於影像之燃燒製程 監控方法之流程圖。 第3圖顯示第2圖中的步驟S23之火焰識別方法流程 圖。 第4圖顯示第2圖中步驟S27之診斷爐膛的穩定性和 效率診斷方法流程圖。 第5圖係本發明之特定區域選定之一流程圖，用以實 現步驟S302。 【主要元件符號說明】 11〜燃燒系統； 13〜高溫攝影機； 15〜現場電腦（Field PC); 17〜應用電腦（Application PC); S21〜取得爐膛之原始影像； S23〜將原始影像中的火焰影像與背景影像分離； S25〜計算火焰影像之特徵； S27〜根據火焰影像之特徵，診斷爐膛的穩定性和燃燒 效率； S29〜發出對應的警報訊息； S301〜接收複數個原始影像； S302〜選定原始影像中的特定區域作為分析對象； 19 201224823 S303〜將原始影像轉換至HSV的色彩空間； S304〜執行模糊化程序； S305〜執行模糊邏輯推論； S306〜執行解模糊化過程； S307〜由原始影像中分離出火焰影像； S41〜擷取火焰影像之特徵資訊； S 4 2〜以火焰影像特徵來建立尾氣濃度之即時預測值； S43〜將火焰影像特徵透過自適應網路模糊推論系統建 立單張火焰影像的穩定性診斷結果； _ S 4 4〜利用當下的一張與前複數張火焰影像的穩定性診 斷結果連續性的透過自適應網路模糊推論系統以建立一動 態火焰影像的穩定性診斷結果； S45〜將複數個尾氣濃度之即時預測值和動態火焰影像 的穩定性診斷結果進行火焰燃燒效率診斷； S501〜色彩物件； S502〜選擇非感興趣的色彩物件； S503〜產生非感興趣遮罩； 籲 S504〜儲存複數個非感興趣遮罩； S505〜矩陣邏輯運算； S506〜複數個原始影像； .. S507〜是否滿足隨機選擇； S508〜融合之感興趣遮罩； S509〜完成區域選定。 20P See one of Annex II, which shows the original image of the excess combustion air (picture on the left) to the lack of combustion air (picture on the right) and the identified flame image. Since the sampling measurement of the flue oxygen concentration has a time delay, in the present embodiment, the correlation analysis is first used to know that the time delay is about 39 seconds, in order to synchronize the characteristics with the gas concentration data, see Annex II bis and II. Third, according to the observation of the characteristics of the flame image and the oxygen concentration of the exhaust gas, the key points are as follows: (1) The flame image features are highly correlated with the combustion state; (2) The inner flame area is almost fixed, and the outer flame area is related to the combustion air. Decrease and increase, also reflected in the area ratio; (3) When the operating conditions change, the average brightness and average temperature of the inner and outer flames are almost unchanged, but the brightness and average temperature of the total flame are positively correlated with the tail gas concentration, mainly inside and outside. (4) The ratio of the brightness variation of the inner and outer flames reflects the tendency of combustion instability; (5) The air flow affects the center of mass of the entire combustion flame, but the center of the Y-axis of the inner flame is not change. Lu Figure 4 shows a flow chart of the method for determining the stability and efficiency of the furnace in step S27 in Figure 2. In step S41, the feature information of the flame image obtained in step S25 is captured. In step S42, in order to diagnose the combustion efficiency in an instant, a regression model, such as a least squares method (PLS) or a neural network, is used to establish a real-time predicted value of the exhaust gas concentration by using the flame image feature to overcome the combustion system and 14 201224823 Measurement delay. The immediate exhaust gas prediction model can be as described in the literature (eg, literature: H. Yu, JF MacGregor, " 'Monitoring flames in an industrial boiler using multivariate image analysis,' AICHE Journal 50 (7), pp. 1474-1483, 2004.) In step S43, the feature of the flame image is also passed through an adaptive network fuzzy inference system (ANFIS) to establish a stability diagnosis result of a single flame image. In detail, the feature input adaptive of the flame image After the network fuzzy inference system is compared with a historical event database, the flame image is classified according to the combustion state, and the stability diagnosis result of the single flame image is generated. The adaptive network fuzzy inference system is Utilize the network architecture to achieve automatic adjustment of membership functions, and automatically establish fuzzy if-then rules. For example, refer to the adaptive network fuzzy inference system described in the literature (for example, J.-S. Roger Jang, CT Sun, E. Mizutani, Neuro-Fuzzy and Soft Computing: A Computational Approach to Learning and Machine Intelligence, Pearson Education Taiwan Ltd, 2004.). In step S44, an adaptive network fuzzy inference system (ANFIS) is used to establish a dynamic flame image using the stability of the stability of the current and the previous plurality of flame images. Stability diagnosis result. In this embodiment, the stability diagnosis result of the current one and the first 3 to 5 flame images is input to the adaptive network fuzzy inference system (ANFIS) in a continuous state. Do further dynamic confirmation' This step can avoid misdiagnosis due to factors such as sampling time and random flickering of the flame, and produce dynamic flame image stability and robustness. 15 201224823 Most steps & Once a, the instantaneous predicted value of the plurality of exhaust gas concentrations is broken. That is, the stability diagnosis result of the vortex, the flame combustion efficiency diagnosis and stability diagnosis: the knot J should: the path fuzzy inference system infers the broken index of the dynamic flame image , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , The adaptive network fuzzy inference system is applied to the combination of the combustion device and the flickering dynamics of the flame itself and the image capture and attenuation, and it is difficult to determine the characteristic quantization standard; (2) various types: system and sample limits, difficult There is a complete special (four) break criterion, so the first layer adaptive network fuzzy inference system of step S43 is used to automatically establish the fuzzy feature standard and judgment criterion of the image feature of the combustion system, and the fire can be established as a systematic expert experience. In the method, the first layer adaptive network fuzzy inference system in step S44 is used for further dynamic confirmation, which avoids misdiagnosis caused by factors such as sampling time and fire 35 flicker. An embodiment will now be further described to further illustrate how combustion process monitoring and diagnostics can be performed using the combustion process monitoring and diagnostic methods of the present invention, but is not intended to limit the invention. This embodiment was tested in an industrial steam boiler with a vapor capacity of up to 15 tons per hour of steam supply, and the operating conditions were divided into three types: small, medium and large. In addition, for the unstable combustion and the extinguished and fired state, the experimental collection and combustion state analysis are carried out together, and the image source is obtained by using an external shooting type high temperature camera. The test target is to accurately identify the flame image: i. Stable combustion (Three loads), ii. Unstable combustion, and iii. Flameout. 201224823 The three characteristics selected in this embodiment are the accessory = the ratio of the inside of the flame to the area of the flame, the total, the average of the domain value, and the brightness of the inside of the flame = the redundancy == and the identified Outside the flame image, the object does not = the stability of the image of the fire on the image of the r-image. Dynamics: : load, (1) large load, (4) unstable soil and (5) negative: fire 2) heart 'red origin is burning The status value of the image is output. Caused by flame disturbance: stability from w# The result of the break can be reduced because the part of the code can be used by the program disc, :::: included in physical media such as floppy, light, disc, or any other The machine can be read (such as a computer readable) into two 'when the code is loaded and executed by the machine, such as a computer, and the present invention. The code can also be transferred to the device of the present invention by - (4) sending, basin, winding, fiber, or any type of transmission. When in-two. Apply a unique device for a specific logic circuit. The present invention has been described in the above preferred embodiments, and it is not intended to limit the invention to 201224823. Anyone skilled in the art can make various changes and modifications 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. 201224823 [Simplified Schematic] FIG. 1 is a schematic diagram showing a hardware architecture according to an embodiment of the present invention. Figure 2 is a flow chart showing an image-based combustion process monitoring method in accordance with an embodiment of the present invention. Fig. 3 is a flow chart showing the flame identification method of the step S23 in Fig. 2. Fig. 4 is a flow chart showing the method of diagnosing the stability and efficiency of the diagnostic furnace of the step S27 in Fig. 2. Figure 5 is a flow chart showing one of the specific regions of the present invention for implementing step S302. [Main component symbol description] 11 ~ combustion system; 13 ~ high temperature camera; 15 ~ field computer (Field PC); 17 ~ application computer (Application PC); S21 ~ obtain the original image of the furnace; S23 ~ will be the flame in the original image The image is separated from the background image; S25~ calculates the characteristics of the flame image; S27~ diagnoses the stability and combustion efficiency of the furnace according to the characteristics of the flame image; S29~ sends a corresponding alarm message; S301~ receives a plurality of original images; S302~selects The specific area in the original image is used as the analysis object; 19 201224823 S303~ converts the original image to the color space of HSV; S304~ executes the blurring program; S305~ executes the fuzzy logic inference; S306~ performs the defuzzification process; S307~ is the original The flame image is separated from the image; S41~ extracts the characteristic information of the flame image; S 4 2~ establishes the instantaneous predicted value of the exhaust gas concentration by the flame image feature; S43~ establishes the flame image feature through the adaptive network fuzzy inference system The stability diagnosis result of the flame image; _ S 4 4~ using the current one and the previous plural The stability of the flame image is diagnosed continuously by the adaptive network fuzzy inference system to establish a stable diagnosis result of the dynamic flame image; S45~ the instantaneous predicted value of the plurality of exhaust gas concentrations and the stability diagnosis result of the dynamic flame image Perform flame combustion efficiency diagnosis; S501~color object; S502~ select non-interesting color object; S503~ generate non-interesting mask; call S504~ store multiple non-interest masks; S505~matrix logic operation; S506~ A plurality of original images; .. S507~ whether the random selection is satisfied; S508~ fusion of the interest mask; S509~ completion area selected. 20