Using Deep Convolutional Neural Networks and Infrared Thermography to Identify Coal Quality and Gangue
Owing to the enormous demand for and growing large-scale use of coal in the China, India, and USA, speculation has arisen about possible hazards to environmental quality and human health. The contents of fly ash and volatile matter in low-quality coal are extremely harmful to the environment. As a r...
Guardado en:
| Autores principales: | , , , , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
IEEE
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/728f48959abc460386348c60da92ef24 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | Owing to the enormous demand for and growing large-scale use of coal in the China, India, and USA, speculation has arisen about possible hazards to environmental quality and human health. The contents of fly ash and volatile matter in low-quality coal are extremely harmful to the environment. As a result, there is still much to be explored regarding known hazards and harms to the natural environment of the Earth. For the detection of high-quality coal, we propose a new method of distinguishing coal quality in terms of different “coal types” (i.e., anthracite, bituminous coal, subbituminous coal and lignite) and efficiently separating gangue and rock from the production lines of coal preparation plants (CPPs) by exploiting infrared machine vision and convolutional neural networks (CNNs) for deep learning, which can make coal use less harmful to humans and nature and/or more useful for general welfare. In this article, we conducted two experiments. First experiment is to study the reactions of coal types, gangue and rock with infrared radiation at temperatures of 50°C, 70°C, 90°C, 110°C, and 150°C. Second experiment, several common CNN models (i.e., AlexNet, DarkNet-53, GoogLeNet, NasNet_Mobileb, ResNet-18, MobileNet-v2, Inception-v3 and DenseNet-201) are trained and tested to classify coal types and distinguish gangue and rock. We achieve a remarkable classification accuracy of 100% in these training and testing processes when employing the ResNet-18 and DenseNet-201 models. |
|---|