FLAME MONITORING USING IMAGE CLASSIFICATION

. This paper presents comparison of image classification methods for co-firing biomass and pulverized coal. Two classes of combustion – stable and unstable were defined for nine variants with different power value parameters and fixed amount biomass. Experimental results show that correct classification of images was achieved for the assumed variants.


Introduction
Combustion process is complex, nonlinear and non-stationary therefore analyzing is difficult. One method of diagnosis the combustion process is to use flame as the source of information. By analyzing the image of the flame can get information about the process almost without any delay. This is particularly important in the case of fuel combustion characterized by a high variability of physico-chemical properties. Current regulations require continuous increase in the proportion of biomass in order to obtain electricity. Co-firing of biomass and coal is the easiest way to use it but it is technologically difficult process due to the different characteristics of the components of the mixture combusted. Combustion tests were carried out on the test bench at the Institute of Energy. Measurements were made at the position of the camera perpendicular to the axis of the flame for different variants of power. After the initial analysis of images of the co-pulverized coal and biomass were determined sequence of images of stable and unstable combustion. Image classification method used to determine the state of the process. Article compared three methods for image classification with naive Bayesian classifier, K nearest neighbor and support vector machine with RBF kernel function [6].

Classification using Naive Bayes classifier
The Naive Bayes classifier is simple effective and has low computational requirements. Let i  be a input vector to classify and X be a defined class. In Bayes formula there are transform the probability ) (1) P(X) can be determined from training data. Given a conditional independence of components of the vector can be combined as follows: From equation (3) ) | ( i X P  can be calculated and i  can be classified with the highest A simple Bayes Classifier algorithm works as follows. Data sample can be represented by n dimensional feature vector. The can be m classes. Given an unknown data sample S the classifier will predict that S belongs to the class with the highest conditional probability on S. When classify sample S, for each class X i equation

Classification using K nearest neighbor
K nearest neighbor is a method for classifying objects using training examples in the feature space. From all machine learning algorithms K nearest neighbor is the simplest. Training process for this algorithm only consists of storing feature vectors and labels of the training images. In the classification process, the unlabelled query point is simply assigned to the label of its k nearest neighbors. Typically the object is classified based on the labels of its k nearest neighbors by majority vote. If K = 1, the object is simply classified as the class of the object nearest to it. When there are only two classes, K must be a odd integer. However, there can still be ties when k is an odd integer when performing multiclass classification. After we convert each image to a vector of fixed-length with real numbers, we used the most common distance function for K-NN which is Euclidean distance [4]: where x and y are histograms in X = R m . The process of K-NN classification is presented on figure 1. Classification depends of the K value of 1, 4, or 9. The query point can be a triangle at (1), a rectangle at (2) and a circle (3).

Classification using Support Vector Machine (SVM)
SVM is generalized linear (non-linear with RBF kernel) classifier and its characterized as classifier that can minimize the error. The main conception of SVM learning system is a separate hyper-plane by two parallel hyperplane. As shown in fig. 2, the black dots and the white dots are the training dataset which belong to two classes. A vector is mapped to the intertval to establish a hyper-plane of the largest higher-dimensional space. The optical plane H is found by maximizing the margin value 2/||w||. Hyperplanes H 1 and H , is applied to x, to make x become linearly dividable. A weight w and offset b satisfying the following criteria will be found: Assume that the equation of the optical hyperplane H (Fig. 1) is , then the distance of the data point in any of the two classes to the hyperplane is: A is to be found to maximize Then the search of the optimal plane H turns to a problem of a second order planning problem There are several kernels that can be used in support vector machines models. These include linear, polynomial, radial basis function (RBF) and sigmoid function: The most popular kernel types used in SVM is RBF [2,3,5,7,9].

Laboratory combustion facility
Combustion tests were done in a 0.5 MWth (megawatt of thermal) research facility, enabling scaled down (10:1) combustion conditions. The main part is a cylindrical combustion chamber of 0.7 m in diameter and 2.5 m long. Alow-NOx swirl burner about 0.1 m in diameter is mounted horizontally at the front wall. The stand is equipped with all the necessary supply systems: primary and secondary air, coal, and oil. Pulverized coal for combustion is prepared in advance and dumped into the coal feeder bunker. Biomass in a form of straw is mixed with coal after passing through the combustion chamber has two lateral inspection openings on both sides, which enable image acquisition. A high-speed camera with CMOS area scan sensor was placed near burner's nozzle, as shown in fig. 3. Flame images were transferred from the interior of the combustion chamber through a 0.7 m borescope. The camera was acquire 30 frames per second at its full resolution (12801024 pixels). The optical system was cooled with water jacket. Additionally, purging air was used to avoid dustiness of optical parts [8].

Combustion tests
Combustion testes were done for different variants (combinations) of the combustion facility, where thermal power (P th ) and excess air coefficient () were set independently for known biomass content (20%), where  is defined as quotient the mass of air to combust 1kg of fuel to mass of stoichiometric air. The exact values of thermal power and excess air coefficient are collected in table 1. The main goal of the tests are detected when co-firing is stable or unstable correlated with presented variants. Fig. 4-12 present example of nine variants co-firing coal and 20% of biomass (straw).

Results
In experiment were used first 10000 frame from all variants. Captured images were converted to 8-bit grayscale, thus pixel amplitude was ranging from 0 to 255.
In table 2, nine variants are tested in terms of stable combustion positive detection rates. It is defined as the number of correctly classified frames containing stable combustion divided by the total number of frames that contain stable combustion. There is compared Bayesian, SVM classifier and K-NN Classifier. Table 2 shows that classifier K nearest neighbor achieved bests results. Future work will be concentrate to implement combine classifier algorithm. The aim is to use the method, that is allow to monitor combustion process in real-time.