Methods Research Of Flame Image Multiple Features Extraction For BOF Steelmaking Blowing Data Prediction

Conberter blowing data prediction can control the converter end point effectively. With the rapid development of digital image processing technology, methods of calculating blowing data from flame image were concerned to many scholars. As the oxidation rate of chmical elements changed, the color, brightness, shape, length and other characteristics changed also. The key problem of converter end point is to construct appropriate features, which can accurately descript the changing process of the temperature, carbon and phosphorus.In the combustion process, the flame shows a rapid and dynamic changes and short-term steadystate moment. Flame boundary changes both in the macroscopic shape of its contor, but also reflects a significant turning point in the small scale changes. The flame image texture, which is a non-natural texture of the cycle, shows the characteristics of random microtexture.All these factors become the difficulty of the flame image feature value extraction. This article aims to expand more in-depth reaearch on methods to solve this difficult problem. The main research contents are as follows:(1) In the flame image collection process, the dust and impurities in the optical path may affect the clarity of the flame image. Color flame image pre-processing methods based on mathematical morphology were researched. On the basic of studying two value morphology and gray scale morphology image processing methods, Gary scale morphology processing method was extended to the color vector morphology. Method of vector morphology image processing based on difference formula in uniform space is proposed in this paper. The gap between the visual colors is proportional to the Euclidean distance in the Lab uniform color space coordinate, and this characteristic is used to calculate the color difference, which is used as the criterion of vector ordering. The color image quaternion model and morphology structure element quaternion model are established to delimit the new morphology operators, such as erode, dilation, open and close. The definition morphology image processing method was used to remove salt and pepper noise, boundary extraction, and was compared with some other method. The experiment result showed that it is more effective to use the proposed method to extract the borders and remove the salt and pepper noise.(2) The segmentation and chroma information extraction to the color flame images. Color similarity coefficient evaluation method and Euclidean distance method were done on the segmentation of flame and background respectively, and their performances were evaluated. The three order moment was calculated to the color flame based on the segmentation. GRNN was used to build the convert data prediction model based on color information. The recognition result was compared to the color mean method. The analysis proved that the flame third-order moment features has good speed and prediction accuracy, and can be used as the characteristic value of the chrominace information of the color flame.(3) In order to express the flame boundary bending complexity, which is used to describe the rate of oxidation of carbon in the molten pool. The polygon reconstruction was done to the flame boundary, which can remove the interference of inflection point and ensure the flame shape at the same time. The flame boundary complexity was extracted at different blowing period. GRNN was used to build the converting data forecast model based on flame boundary complexity. The recognition result was compared to the line moment invariant and circular boundary description, analysis and comparison result showed that this method is more effective. This method can be used to extract the flame boundary complex eigenvalue.(4) In order to express the flame texture roughness, which is used to describe the oxidation rate of the impurities in the molten pool and burning state. Gray differential statistics texture complexity description method was proposed. A kind of strategy suitable for random micro texture of the difference statistical was given. The gray scale difference histogram was established. The characteristic value was calculated. Since the entropy represent the concept of complexity, and fit to the application of this paper. GRNN was used to build the converting data foecast model based on entropy feature texture complexity. The recognition result was compared to the Laws texture and gray level co-occurrence matrix method, analysis and comparison of the effectiveness of the proposed method. This method can be used to extract the flame texture complexity.(5) Convert blowing data distinguish system was designed, which include hardware structure and software function. The hardware system structure and composition was designed. A callback function for flame image reading, segmentation, color feature extraction, boundary complexity feature extraction, texture complexity feature extraction were programmed on the MATLAB GUIDE visualization planform.GRNN was used to build the converter blowing data prediction model based on multi feature of flame image. A kind of correction method for the sharply changed data was proposed. The recognition result was compared to the BP neural network and showed the effectiveness and feasibility of this method.During the converter blowing process, flame image changed fast, and all kinds of natural characteristics showed a multiscale complex changes. This paper was aimed to study suitable flame image processing feature extraction method and theory, and to establish the relational model between flame image characteristics and converting data. This method can realize the data real time prediction at the actual blowing process, and is effective to the converter end-point control.