Research On Burner Flame State Monitoring Technique Based On Image Processing

The real-time monitoring of the furnace flame in coal-fired power plants is crucial for the economy of power generation and the safe operation of boilers.Traditional fire detection techniques based on energy signals such as light,heat,and radiation can only detect the presence of flames and,therefore,gradually fail to meet the increasingly urgent demand for fine-tuning in thermal power generation.In this study,the features of the flame images from actual power plants are analyzed from multiple perspectives.A flame state classifier is designed using the support vector machine based on multi-dimensional feature fusion.Furthermore,an improved flame state classification method based on Inception deep convolutional network is proposed.The main work and innovation of the paper are as follows:(1)By deeply analyzing the characteristics of burner flame images,multi-dimensional features of the flame are extracted,and datasets are created.Firstly,targeted preprocessing such as image restoration and noise reduction filtering are carried out,and then the HSV color space color moments are used to describe the color characteristics of the flame.Geometric features such as the Unburned area,flame area,and flame irregularity are introduced.The flame texture features are extracted using the contrast,energy,and correlation of the gray-level co-occurrence matrix.The dynamic features of the flame are represented by the average brightness of the flame sequence and the variance of the center offset.All the obtained feature parameters are then composed into feature vectors that describe the flame state.Feature vectors of different state categories are combined to form a feature matrix,which is used as the flame feature dataset.At the same time,the preprocessed images of different categories of flames are used to create a flame image dataset.(2)The multi-class support vector machine and inception deep convolutional neural network models are constructed to achieve flame state classification based on manual and automatic feature extraction,respectively.For the former,the principal component analysis is used to reduce the dimensionality of the original feature matrix.The model is trained to classify flame images obtained under three different burner loads and achieves a classification accuracy of 99.76%after parameter optimization by improved particle swarm optimization algorithm.For the latter,system pruning was employed to optimize the network,resulting in a 2 times increase in convergence speed and a classification accuracy of 99.67%.To further improve the network’s convergence speed,an improved Siamese inception network model,which converted the flame state classification problem into an evaluation of state similarity,was proposed indirectly to achieve the classification objective.Experimental results showed that the proposed network architecture improved the convergence speed by 5.5 times without decreasing the classification accuracy.