Research On Boiler Combustion Status Monitoring Method With Multi-features Fusion

The safety and stability of combustion process is the priority in power plant.Effective monitoring of combustion status can fulfill the needs of plant for safety,economy,stability and environmental protection.With the integration of renewable energy generation into the power grid,it is challenging for the grid to operate safe and stable.Power generation of thermal power plant still dominates in China.Thermal power plant needs to undertake the responsibility of regulating and stabilizing the fluctuations brought by renewable energy generation.On the one hand,thermal power plant is affected frequently by the condition adjustment.It leads to unstable combustion and makes it difficult to maintain stable boiler output.On the other hand,renewable energy generation connected to the pow grid requires the plant to maintain stable operation under low and medium load condition for a long time.Therefore,real-time and efficient monitoring of the combustion status under low and medium load is significant to stabilize the combustion.It also achieves the task of deep load regulation.The paper focuses on the topic of combustion status monitoring of coal-fired boilers and for the lack of information in combustion process under low and medium load,the ineffective monitoring of combustion status and the difficulty of guaranteeing real-time monitoring,the paper examines the following aspects:1.Analyzing multiple factors affecting the combustion process in the boiler,a 2D Gabor-GLCM global texture feature extractor based on multi-scale directional is proposed for analyzing the combustion feature variations in a single frame of combustion images based on the boiler burner outlet combustion video images.The texture extractor extracts texture features that more effectively which describe the combustion process from the globe image without precise segmentation of flame locations.Meanwhile,the validity and sensitivity of the texture features was verified using correlation analysis under low and medium load conditions.2.Considering the fluidity and randomness of combustion process,the singleframe image texture features can only characterize part of the information of the combustion process and cannot reflect the dynamic features of combustion.Based on this,two different optical flow methods are proposed to construct combustion motion feature extractors.The pyramid modified bi-directional L-K sparse optical flow and dense inverse search fast optical flow are used to extract the dynamic features of the combustion process with changing working conditions.The quality and real-time performance of the dynamic features proposed by the two optical flow are evaluated by using three metrics:information entropy,correlation and algorithm running time.The experimental analysis shows that the changes of combustion motion features can accurately follow the changes of current working conditions and combustion state.The correlation between them and the related parameters also verifies the rationality of the extracted motion characteristics.3.Flue gas oxygen content is an important indicator to monitor the adequacy of combustion.It is one of the features that laterally reflects the combustion status,which is the focus of subsequent combustion adjustment.For the field oxygen content detection there are high cost,low accuracy,lag obvious and other problems.Based on this,a soft measurement model of oxygen content in Bi-GRU based on the fusion of combustion features and thermal engineering parameters is proposed.The model uses kernel principal component analysis to fuse and reduce the dimensionality of multiple combustion image features and secondary air volume.It uses the fused features to train a Bi-GRU neural network to establish an oxygen content measurement model.The simulation experiments show that this measurement model has higher measurement accuracy and better real-time performance compared with other soft measurement models.4.A semi-supervised combustion condition monitoring model based on semi-AP clustering method and RBFNN is proposed based on the above combustion texture features,dynamic features and oxygen content detection values.The model uses semiAP clustering to label unlabeled samples,expands the labeled sample set of the combustion monitoring model.And it mixes the real sample set with the pseudo-labeled sample set,and trains a RBFNN to build a semi-supervised combustion monitoring model.The semi-AP clustering can solve the problem of insufficient label samples of combustion state images,or too concentrated label samples.The experimental results show that the model can monitor the combustion state in real time and has a high recognition rate of unknown combustion state.Compared with other monitoring models,its identification accuracy of combustion state under low and medium loads is higher,and the monitoring results can provide guidance for subsequent low load stabilization combustion.