Boiler Combustion Flame Image Processing And State Recognition

The concurrent and accurate supervision of the state of the boiler combustion flame has greatmeaning in improving the combustion efficiency and prohibiting the occurrence of the potentialdanger. Therefore, studying the key techniques of flame image processing and state recognition hastheoretical and practical significance. On the basis of previous research results, researches on flameimage denoising, enhancement, thresholding, edge detection and state recognition have been done inthis thesis, and are described as follows:Firstly, a flame image denoising method based on dual-tree complex wavelet transform andhidden Markov tree (HMT) model is researched. Using multi-scale, multi-directional, andtranslational invariance features of dual-tree complex wavelet transform, this method is combinedwith HMT model and can accurately describe the correlation of dual-tree complex wavelet domaincoefficients between different scales. According to the quantitative analysis of experimental results,the method has achieved good denoising effect.Then, a boiler combustion flame image enhancement algorithm based on contourlet transformand parameterized logarithmic image processing model is proposed. The flame image is decomposedinto low-pass subband and band-pass directional subbands by the contourlet transform, and thelow-pass subband is enhanced by the PLIP Lee’s algorithm while band-pass directional subbands areprocessed by the nonlinear gain function. The parameters in the enhancement algorithm are selectedby the niche chaotic particle swarm optimization algorithm. The experimental results show that themethod can enhance flame image effectively.And then, a boiler combustion flame image multi-threshold selection method based on reciprocalcross entropy is realized. The definition of the reciprocal cross entropy is given and the singlethreshold selection method based on minimum reciprocal cross entropy is derived. Then it isgeneralized to multi-threshold selection and the multi-threshold selection method based on minimumreciprocal cross entropy and improved particle swarm optimization algorithm is given. Theexperimental results show that, compared with maximum Shannon entropy, gray entropy, Otsu andShannon cross entropy threshold selection method based on improved particle swarm optimization,the proposed method has obvious advantage.Next, an edge detection method of boiler combustion flame image based on anisotropicmathematical morphology is proposed. In the different pixels of the image, the morphology structural element is constructed according to the directional information of that pixel and the morphologygradient is computed. The edge is detected according to the result computed using morphologygradient. Compared with existing edge detection methods, the edge detected by the method proposedis more accurate and complete.Finally, a combustion flame image state recognition method based on multi-threshold selectionusing gray entropy and support vector machine is given. The flame image is segmented based on themulti-threshold selection method using gray entropy, and10characteristic parameters are extractedfrom the segmented image and support vector machine is trained using those characteristic parameters.The flame image is classified according to the extracted characteristic parameters and support vectormachine. The experimental results show that compared with the classifying method using the imagepixels as training samples, the method proposed has higher classifying rate.