Intelligent Characterization System for Char Combustion Reactivity using Deep Learning Alpana School of Computer&Systems Sciences Jawaharlal Nehru University New Delhi, India alpana.srk@gmail.com Satish Chand School of Computer&Systems Sciences Jawaharlal Nehru University New Delhi, India schand20@gmail.com Vivek Mishra Innovation Center of Coal Exploitation Hebei University of Engineering Handan, China drvmishra@hotmail.com Abstract—Coal is the primary source of fuel for generating electrical energy through the combustion of pulverized coal. Volatile compounds are released during this process, resulting in the creation of varieties of char elements. The morphology of char elements can be categorized into groups, which reflects the different level of coal reactiveness, that can be utilized to assess the effects of coal on burner operations. Currently, industries follow the manual and semi-automated microscopic methods to recognize the reactivity of char. These methods are time-consuming, subjective and restricted to small sample sizes only. The objective of this research is to provide intelligent char classification techniques with advantages in terms of speed, consistency, and accuracy. This paper attempts to propose an automated system for char reactivity classification using RES-NET deep learning method. The result shows an accuracy of 93.89% in less computational time. Hence, the proposed system may be suggested to industries for char classification. Keywords— coal, char, deep learning I. INTRODUCTION The most familiar technique to generate electricity in coal-burning plants is the combustion of pulverized coal. Chars are an outcome of the primary phase of the coal combustion process. The morphology of char can be analyzed through a microscope. Wall size, porosity, structures, and unfused material correspond to the morphological characteristics of the char. The morphology of char may be utilized to measure the coal’s reactivity through which the effectiveness of combustion, carbon dioxide and ash amount liberated into the atmosphere, is determined [1- 2]. During combustion, the reactivity of char depends upon changes in the structure and size of particles. The International Committee for Coal and Organic Petrology (ICCP), Combustion Working Group in Commission III, proposed nine types of char [3]. This categorization can be outlined as "reactive" and "non-reactive" into two main groups. The particles of char having high reactiveness morphologies are suitable for combustion. Experts typically study char morphology using a microscope. Char particles are soaked in resin until this can be done to create a stone. The wedges are refined to reveal the char particle upon the surfaces after the resin has been healed [4]. Instead, based on structural properties of the ICCP characterization, char particles from the block surface are detected, counted, and graded through an oiled microscope. Ultimately, frequencies per type of char are utilized to establish the reactivity of char. The type of char with the maximum occurrence is selected as reactive coal estimation. A professional examination of char reactiveness may be error-prone, subjective, and involves considerable time, as indicated by Wu et al. [5]. This is mandatory to examine and differentiate between 300 and 600 particles of each sample to ensure the reproducibility of findings. Nonetheless, several manual analyses can be done automatically in industrial processes using image processing. In coal industries, the coal characterization scheme allows the automation of coal quality estimation using color and textural properties [6], calculation of coarse coal ash content using color and texture characteristics [7], and particle size estimation and distribution on fine coals [8]. Likewise, to increase the precision of char investigations and minimize the processing times, automated software using image processing would be helpful. Few systems have been documented to classify coal samples into char form following the ICCP specifications wherever char elements are recognized inevitably and structural features are counted automatically [ 9–15]. Special effort is made in these systems to identify particles, as the classification depends heavily on morphological characteristics. Morphological characteristics are incorrectly calculated due to fused particles and fragile walls. Misclassified particles can, therefore, affect the study of char reactivity and familiarize faults in the characterization of coal qualities [16]. There have been many kinds of research done in the past on char analysis using viable software, for example, KS400 histogram-created approaches like Isodata [17-18] and Triangle technique [19]. Nevertheless, if noteworthy openings in detachment part the element wreckages, these approaches fail to perceive particles. A sliding window approach is implemented in other applications [20-21]. Some other researches have been also done based on candidate regions, color, texture area etc. [22-25]. Though the pieces of literature show the noteworthy results but these can be in the form of accuracy and computational time. To achieve better accuracy in less computational time, this paper attempts to propose the intelligent char characterization system. In this research, the automated system has been proposed for the classification of chars into two reactive groups using deep learning to avoid the manual features extraction process which is subjective and inaccurate. The research objective is to propose a scheme that is realized to be an effective method for char characterization with better accuracy in less computational time. Here, the simple convolutional neural network (CNN) and residual network (RES-NET [50]) has been used for the classification of two groups of char: class A and class B which is further named as a reactive and non- reactive group of chars. The comparative analysis of these deep learning models gives an acceptable result. II. MATERIALS AND METHODS A. Image Collection The char image samples have been collected from Crelling’s Petrographic Atlas of Coals and Carbons, Council of Scientific and Industrial Research (CSIR), India, supports this research under the Direct SRF scheme. 2020 7th International Conference on Signal Processing and Integrated Networks (SPIN) 978-1-7281-5475-6/20/$31.00 ©2020 IEEE 834