LKDPNet: Large-Kernel Depthwise-Pointwise convolution neural network in estimating coal ash content via data augmentation

Abstract

Product quality monitoring is one of the most critical demands in the coal industry. Conventional coal quality analysis is offline, laborious, and lagging behind coal production. Using machine vision for determining ash content in coal has been recently developed. However, there are some challenges in the model design due to its task complexity. A data augmentation method for a specific task was proposed to deal with the peculiarities of a small, unbalanced industrial dataset. For estimating ash content in coal from an image without background, we provided a trajectory going from a ResNet to an LKDPNet under the guidance of receptive field size. First, the model depth was determined by changing the stage ratio. Second, patch embedding was employed to substitute the stem cell of ResNet for downsampling. Third, a residual connection block of depthwise convolution followed by pointwise convolution was designed to replace the ResNet identify and conv blocks. The receptive field of the model was adjusted by increasing the kernel size of the depthwise convolution layer and removing the downsampling block. The results and visualization revealed that the proposed LKDPNet could significantly improve performance with an accuracy of 98.91% and an MAE of 0.082. It could pave the way to an online analysis of coal, thus accelerating intelligent coal production.