Abstract
Based on a flame image processing technology, a fuzzy based temperature monitoring system in a rotary kiln was reported. In this paper, we propose a Fuzzy based flame analysis, which consider Red, Green and Blue intensity planes, to measure the temperature from the flame image. The proposed approach integrates RGB intensity as fuzzified input variables, temperature as defuzzified output variables and fuzzy inference rules based Mamdani models. Based on the color characteristics of burning flame, temperature of different flame zones are located using a fuzzy logic controller. The temperature level at hotspot area is the highest and through the fuzzy analysis we were able to identify hotspot area from the flame image. In order to evaluate the performance of the proposed method, quantitative metric such as f-measure has been used and it was found that the f-measure metric yields high accuracy for the hotspot area. The visual inspection of the results along with the f-measure values showed the superiority of our work. Experimental results indicate that the proposed approach can be applied to a high resolution video flame image.
Introduction
Kilns are often regarded as the heart of cement manufacturing plants. Plant profits are affected by the efficiency of these machines. One of the most important phases of cement production takes place in kilns. In order to produce cement with acceptable quality, it is important to make sure that the material has been properly burned. The process of sintering mainly relies on the conditions in the burning zone. The firing temperature, area of the flame and its propagation decides the clinkerisation process. Combustion zone temperature directly reflects the clinker quality. The burning zone temperature and the kiln load are continuously monitored by the experienced kiln operators for ensuring its quality. According to the variations in these two parameters, operators manually adjust the fuel flow rate, raw meal flow rate and the preheater air flows [1].
At present, the kiln operators are estimating burning zone temperature, which is about 1350°C to 1500°C by visually looking into the furnace flame and/or considering the back-end temperature measurement of the kiln. The kiln operators identify the individual zones such as the kiln wall zone, coal zone, material zones and the flame zone by analyzing the burning zone flame image. Burning zone temperature is directly proportional to the clinker quality [1]. Efficient recognition of burning state for sintering process of rotary kiln is important to the design of image-based intelligent control systems.
Current approaches such as consensus-based methods, temperature-based methods and temperature measurement using pyrometers etc, could not achieve satisfactory performance. The temperatures which are present in the two ends of a rotary kiln are measured, and the signals which are generated by measuring devices are used in guiding the operation of the rotary kiln. The temperature inside a rotating rotary kiln is measured by means of having at least one pyrometer mounted on the kiln shell. Simple temperature measurements on both ends of the rotary kiln are not sufficient for guiding or controlling the operation of such rotary kilns. Thermoelectric pyrometers (thermocouples) have been used which were fastened to the kiln shell, and which rotate with the kiln. The protective shield between the thermocouple and the raw material increases the time lag of temperature measurement, and produces an erratic temperature reading. Most of rotary kilns are still working under manual control with human operator observing the burning status. With the help of a standard video picture monitoring for the operator, the flame analysis system determines the flame shape, temperature distribution and the heat transfer that takes place within the sintering process.
Our idea is to characterize the flame’s shape and provide temperature measurement with the classical approach of digital image processing using fuzzy logic controller. Based upon a mathematical fuzzy inference engine it is possible to predict the output temperature variables using the current input intensity parameters. Flame images are obtained from the high temperature camera system used in the rotary kiln. A Fuzzy based flame image analysis considering all three intensity planes to measure different temperature zones from the flame image has been developed. The intensity values of each R, G and B planes are mapped to the fuzzy input membership functions. The output fuzzy sets, which is the set of temperature values, are mapped to the membership functions with the guidance of the kiln operator. Fuzzy rule base which is determined empirically, maps the fuzzy input variables to the fuzzy output. The Mamdani [2] model of fuzzy inference system (FIS) for the fuzzification and defuzzification process provides a fuzzy flame temperature mapping. In the flame images, taken using the high resolution camera, intensity of blue plane varies in the high temperature area. This is the reason why we consider all the three planes for temperature mapping using fuzzy logic.
In the Fuzzy based flame analysis, the input and the output are fuzzy variables. Input fuzzification variable depends up on the range of intensity values for each zone and the output defuzzification depends on the range of flame temperatures. Use of fuzzy logic in flame analysis provides better performance compared to the existing two color method [3]. The metric f-measure [4] is used for comparing the ground truth of the hotspot temperature area with defuzzified temperature output. In the result section the average f-measure values for the experimental data sets are tabled. The higher f-measure values shows the superioriority of our method over other methods.
This paper is organized as follows. The related works already reported in the literature is detailed in Section 2. Section 3 discuss about the proposed method of temperature mapping using fuzzy logic. Experimental results are shown in Section 4 and the conclusions are drawn in Section 5.
Literature review
A new system which performs the visual examination of the sintering process is desired for the continuous runnability of the rotary kiln. An image processing based thermographic approach is presented in this paper. Many flame image based methods have been proposed for the burning zone recognition in rotary kiln. One of the approaches uses a soft sensor model for online estimation of f-CaO (Calcium Oxide) content, where the raw data were pre-processed to distinguish the flame image regions of interest and remove process value outliers. A partial least square technique was applied for extracting the compressed score matrix features from the concatenated flame image features and the filtered process variables. Also feed-forward neural networks with random weights were employed as base learners [5].
In [6], use of charged coupled cameras (CCD) to capture digital image flames was described. And according to wine’s law of radiation, Radiation Energy Signal (RES) was obtained from the flame images. RES was sensitive to change in the combustion of the boiler. By analyzing the RES we can vary the fuel flow for combustion thus improving the combustion control. [7] is based on imaging based multi-functional instrumentation system for online monitoring. Here the software incorporates two color pyrometry, and power spectral analysis algorithm, to process the flame images.
In another method [8], flame image based burning state recognition system using a set of heterogenous features and fusion techniques is presented, where ensemble learner model with four types of base classifiers and five fusion operations are examined with comprehensive comparisons. A combustion working condition recognition method based on the Generalized Learning Vector Quantization [GLVQ] is presented in [9]. Kernel Principal Component Analysis (KPCA) method is adopted in this method to deduct input vector with high dimensionality so as to reduce the GLVQ target dimension and network scale. GLVQ neural network is trained by using normalized texture feature data.
In the eigen flame image based method [10], image segmentation is not being used, instead the features are extracted from the eigen flame images to classify the flame images using probabilistic neural network. But this method cannot work with images having poor quality. In [11], the use of state-of-the-art camera for documenting the continuous change in the sintering process is shown, where they use Nonlinear Model Predictive Control (NMPC) combined with flame analysis for gaining essential savings in the clinker burning process.
In the two color method [8], the flame image is captured using an industrial CCD camera with a cooling mechanism and dust cleaning arrangements. The method is based on the color feature based segmentation process. The image is captured in RGB space, which is filtered to sharpen and remove isolated pixels using the median filter, and then converted to Lab color space. Then the image is segmented to flame zone and material zone using the color difference in Lab color space (Delta E method). The temperature distribution of the flame is determined from the reconstructed gray-levels of the two primary color images, R and G images. Flame temperature was calculated by using the two color method. By this method, the flame temperature was calculated from the ratio of the monochromatic radiative intensity at two different and nearby wavelengths, with the assumption that the emissivity of the flame does not change significantly with the wavelength and the temperature.
Temperature mapping using fuzzy logic
The kiln is a rotating cylinder, rotating about its axis few times every minute. The axis of the cylinder is inclined at a slight angle, with a burner at its end. The rotation causes the slurry to pass from the cool end to hot end (burner end). The raw materials for the manufacture of cement (slurry) is fed to the kiln from one side. Coal is used as the fuel for the burning of the slurry. High temperature camera systems are used to capture quality picture of the burning process inside the kiln. These images help in determining multiple zone temperature inside the kiln.
In the proposed method we extract the color image from the video camera, and empirically determine the different temperature zones of the flame image. Fuzzy flame analysis takes the advantage of the fuzzy logic [12] for temperature mapping, where the fuzzy controllers work with intervals or linguistic variables. Here we divide the flame image into different temperature zones and map the intensity intervals of each zone to fuzzy input variables. With the help of the operator guidance the fuzzy rules were formulated to map the image zones with the temperature range. The kiln operator evaluates the high temperature area, i.e. hotspot area. Proposed method use f-measure metric to evaluate the performance of the proposed fuzzy temperature mapping, by comparing the hotspot area. Using fuzzy sets, accurate mapping of the input intensity values and output temperature values for each zone are performed. Also this method is not overly reliant on the operator knowledge as operator guidance is required only at the initializing phase.
Mamdani model of fuzzy inference system logic is employed in our paper. Here the variables are associated with the universes of discourse which are the non-fuzzy sets. These variables take on specific linguistic values which are indicated as fuzzy subsets of the universe. Given a subset A of X (A ⊂ X), A can be represented by a characteristic function. If the mapping is from X to a closed interval [0,1] then we have a fuzzy subset. Thus if A were a fuzzy subset of X, it could be represented by a membership function. If X is a non-fuzzy support set of a universe of discourse, A can then be equated to a linguistic value. Given two such linguistic values A1 and A2 on the same support set of X, logical combinations of A1 and A2 can be formed.
The high temperature camera installed in the rotary kiln provides with high quality images of the burning state of the kiln. For the flame image analysis we classify the flame image into different temperature zones. These zones are Wall Zone (W), Material Zones (M1, M2, M3, M4), Coal Zone (C), Plasma Zone (P), Flame Zone (F), and Hotspot Zone (H). We define W, M1, M2, M3, M4, C, P, F and H as the linguistic variables for the fuzzy system. A range of intensity values is associated with each Red, Green and Blue components of each zones of the image. We use fuzzy sets to represent the linguistic variables, and a fuzzy set x is characterized by its membership function μ
A
(x) → [0, 1]. Where A is the intensity values of each zones and most of the values were common for each of the RGB planes, with varying count. In our approach, trapezoidal membership function was selected after a series of experimental analysis for specifying the intensity range of different zones. Also the different temperature zones were having the same intensity values. We work with the membership functions represented by trapezoidal fuzzy numbers. The parameters a, b, c, d for the trapezoidal membership functions are the intensity values of the temperature zones. The fuzzy sets for the input temperature zones is defined as: Fuzzyset, N = {W, M1, M2, M3, M4, P, H, F}. The trapezoidal membership function is defined as:
The parameters for the trapezoidal membership functions for the various fuzzy sets for each of the Red, Green and the Blue planes were experimentally determined. The corresponding values are shown in Table 1.
The output temperature linguistic variables used in our method are Extremely Low, Very Low, Low, Intermediate, Medium, High, Very High and Extremely High. Fuzzy sets were used to represent these linguistic variables and is characterized by its membership function. Here we used the Gaussian membership function as the membership value never approaches zero for the Gaussian membership function. The Gaussian membership function in general form can be formulated as:
The process of writing the rule base for the FIS required the guidance from the kiln operator. After consulting with the operator we came up with the fuzzy rule base used for mapping the intensities, as shown below:
Rule 1: If Red is
Rule 2: If Red is
Rule 3: If Red is
Rule 4: If Red is
Rule 5: If Red is
Rule 6: If Red is
Rule 7: If Red is
Rule 8: If Red is
The input fuzzy membership function for each of the linguistic variables are shown in Fig. 1, for each Red, Green and Blue planes.
The Mamdani model maps the fuzzy input variables to the fuzzy output temperature variables. At this stage the process of defuzzification is performed. Defuzzification is the process of producing quantifiable results in fuzzy logic, given fuzzy sets and the corresponding membership degrees. During the defuzzification we are mapping each flame intensity values into appropriate crisp temperature value using the Mean of Maximum (MoM) [14] defuzzification method. Here the crisp output is the mean value of all points whose membership values are maximum.
The output of the proposed method is a thermography image, which consists of a colormap. The high temperature point in the flame image is the hotspot area. The kiln operator analyses all the zone temperature derived from the flame image. The performance analysis of the proposed method is evaluated using f-measure metric based on hotspot area segmentation. For measuring the f-measure, the hotspot section of the ground truth image was compared with the hotspot section of the output thermography image. K-means [11] clustering method was used to segment the hotspot section from the output thermography image. K-means provided with the accurate clustered image of the hotspot area. K-means is one of the simplest unsupervised learning algorithms which solves clustering problem effectively. The number of data clusters are fixed prior to the clustering. Main advantages of using the K-means is that it is fast, robust and easier to understand. It is relatively efficient and provides the best result when data sets are distinct. The f-measure can be viewed as a compromise between recall and precision. It is high only when both precision and recall is high. Assume an information retrieval (IR) system has recall R and precision P on a test document collection and an information need. The f-measure of the system is defined as the weighted harmonic mean of its precision and recall, that is, , where the weight αε [0, 1]. The balanced f-measure, commonly denoted as f1 or just f, equally weighs precision and recall, which means α = 1/2. The f1 measure can be written as . The hotspot area of both ground truth and the output is converted to binary image and then the f-measure is calculated. The f-measure is a measure of a test’s accuracy. Closer the value of f-measure to 1, better is the test accuracy.
The proposed algorithm is implemented in MATLAB R2014a. The computer system used for simulation is Intel Core i5 CPU at 2.5 GHz with 4 GB of RAM. High quality images were available from the high temperature cameras installed in the system. From the Cement manufacturing industry we were able to get two flame image data sets, data set 1 and 2. The resolution of the data sets 1 and 2 were the same at 720 × 576, with a frame rate of 25 frames per second. From the first data set, we performed our fuzzy based temperature mapping and the result of our work on the frame 175 is shown in the Fig. 2. In the figure the first column shows the input image. Next column is the ground truth of the hot spot area of the input image obtained from the kiln operator. The third column shows the thermography result of our method performed on the input image. The result is shown with a pseudo colormap for bringing clarity to the output. The temperature range of the flame image can be found out from the color map.
Figure 3 shows the result of our work on the data set 2. The figure shows the result of our method on frame 6 of our video data. The order in which the images are shown are input image, ground truth of the hotspot area and the thermography result of our work on the image. The thermography result shows the temperature of each of the zones which can be inferred from the temperature color map shown with the result. Visual comparison shows that the hotspot zone in the image is having the highest temperature range among all the zones.
Figure 4 shows the visual comparison between the ground truth of the hotspot area with the hotspot area found with the proposed method at different time interval. The frames 25, 50, 75, 200, 225 and the frame 250 were taken for the analysis. First column shows the ground truth value of the hotspot area, while second column shows the hotspot area obtained from the thermography result by using the k-means clustering algorithm.
Figure 5 shows the ground truth and the hotspot area segmented from the result of frames 1, 3, 6, 36, 37, 39 from the dataset 2. In Figure first column show the ground truth of the hotspot area and the next column shows hotspot area segmented from the result of the proposed method using the k-means algorithm. Figures 2 and 3 shows the results of temperature mapping algorithm performed in 2 datasets.
Table 3 shows the f-measure comparison of the hotspot areas in the two data sets. The f-measure values given in the table are an average of the computed f-measure values from the selected frames in the data sets. The f-measure values closer to 1 underlines the superiority of our method.
Conclusion
We have described in this paper a fuzzy method for the flame image segmentation and temperature mapping, for the cement manufacturing process. It can be seen that this method shows accurate temperature mapping of the flame images, which is very helpful in monitoring and controlling the temperature inside the rotary kiln for the cement manufacturing process. The f-measure values for the hot spot area shows the good performance of our method. As a future work we would try to improve the accuracy of our work and also try to create a generalized framework of our algorithm.