Feature Detection And Automatic Identification Of Powder Fuel (Biomass/Pulverized Coal) Based On Image Processing Techniques

Currently,coal is the main energy source widely used,while biomass fuels have a broad application prospect in thermal power generation.Different fuels have different physical and chemical properties,and exhibit complex but intrinsic patterns of combustion and pollution emission phenomena during application.Therefore,there is a need to identify fuel types and estimate the mass fraction of different components in fuel mixtures,and further provide a basis for the needs of calorific value prediction,pollution prediction,and combustion behavior prediction.However,traditional industrial identification methods have limitations such as time-consuming,expensive equipment and cumbersome technical routes.Image methods have been used by many research teams in fuel identification due to their simple principle,low cost,and feasibility to achieve online identification.However,in terms of fuel,most of the existing work has been done for pulverized coal;and in terms of image,most of the research has been done based on flame images.In summary,there is an application potential and research value to conduct research on images of biomass powder fuels that have not been burnt yet.In this study,we propose a method based on image processing techniques to identify the types of biomass powder fuels and to predict the mass fraction of different components of mixtures.The research content and main results are as follows.1)The current research status of the image method in the field of fuel identification and the related technologies and algorithms were investigated,and according to these,designed the experiment and set the overall expected objectives.2)An image dataset containing six different types of biomass powder fuels was established,and the data annotation of a large number of particles was also completed.Moreover,in the feature selection work,the corresponding feature selection rules were designed based on manual dimensionality reduction in order to make the adopted features could reflect the physical meaning of the information contained in the images.3)In the fuel type identification task,this study mainly applied the ensemble learning method,and found that it has good adaptability and high accuracy for different data characteristics.On the test set,the accuracy of the algorithm can reach 98.4%.Meanwhile,based on the combination formula in probability theory,an image partition strategy was designed to further improve the overall accuracy.For the algorithm proposed in this paper,its theoretical accuracy can reach 99.8%,while the actual accuracy is 99.3%.4)In the work of predicting the mass fraction of each component in the mixtures,three different methods were used in this study,completed modeling and testing work on four groups of representative mixed fuel from M1 to M4.In the feature extraction method,the Xgboost algorithm accomplishes feature selection under the condition that almost all features are not positively correlated with the mass fraction,by means of built-in feature selection rules,its RMSE was as low as 0.097 and MAE can reach 0.072,the R2 can exceed 0.84.In this research,based on fuel images,regression features Freg1 and Freg2 with strong positive correlation with the mass fraction were designed for the image processing method and the instance segmentation method.It was tested that for the image processing method,the average RMSE of the four mixed fuel was as low as 0.071,MAE was 0.059,and R2 up to 0.935.For the segmentation result of adhering and overlapping particles,the instance segmentation method was better than the image processing method,and its R2 value could exceed 0.99 when modeled under optimal conditions,but the method might exist the missing detection problem.Overall,based on image processing techniques,this study presents a method for powder fuel type identification and mass fraction prediction different from previous ones.It has the peculiarities of simple principle,low cost,non-invasive and non-destructive to experimental materials,which makes it feasible and effective in applications.Although the effectiveness in mass fraction prediction needs further improvement,the problems and differences in practical application scenarios have yet to be examined as well,this has provided a theoretical basis and a technical route to achieve the online identification and prediction of powder fuel.