A CNN-based regression framework for estimating coal ash content on microscopic images

The overall pipeline for ash content estimation: Step 1: Microscopic images of coal ash content were collected from four scales. Step 2: Due to the limited dataset caused by label acquisition, a data synthetic method was proposed to stitch the four scale images. Step 3: To learn continuous target from imbalanced data and deal with missing data, LDS used Gaussian or Laplacian kernel for getting the smooth value to re-weight loss function. Step 4: The regression network was trained using Ranger optimizer and cosine annealing strategy was used to reduce the learning rate. The predictions were made on test data after training and then computed weighted MAE and CS to derive the final metrics. Step 5: The interpretation of individual predictions was used to explain the regressor in a faithful way, which provides a qualitative understanding of the relationship between the instance's components and the model’s prediction. Step 6: integrated gradients (IG) was used to explain the relationship between a model’s predictions in terms of its features. The results showed that the MAE of the regression model for predicting ash content was 0.31 on the 1,145 sets of test images, where 81.76% had a margin of error less than 0.5% of ash content and 96.25% less than 1.0%. This is promising for implementing online ash content prediction and realizing real-time adjustment for coal processing. [Display omitted] •A method of data synthesis was proposed to augment the dataset.•Label distribution smooth was used for imbalanced image regression.•A regression framework for estimating ash content was designed.•The explanation of the regression model for ash content estimation was made.•The regression model was visualized by integrated gradients. Coal ash content is an important criterion for evaluating coal quality. In recent years, the online ash measurement approach based on a convolutional neural network (CNN) has gotten a lot of attention. However, learning continuous targets from a small and unbalanced dataset is one of the biggest challenges for ash content estimation using CNN. In this paper, a CNN-based regression framework was proposed for rapidly estimating the ash content of coal. Firstly, data synthesis was performed to augment the limited dataset, and label distribution smoothing (LDS) was employed to alleviate the imbalance in datasets. Secondly, separable convolution (SC) and attention modules were introduced into multi-branch (MB) blocks of the backbone. SC was applied