特征(语言学)
人工智能
计算机科学
模式识别(心理学)
直方图
卷积神经网络
特征提取
特征选择
切片
煤
非线性系统
像素
深度学习
泡沫浮选
图像融合
绘图(图形)
融合
启发式
数据挖掘
工艺工程
图像(数学)
定向梯度直方图
图像处理
作者
Zhiping Wen,Sijie Yang,Xiuqi Chen,Maiqiang Zhou,Changchun Zhou
标识
DOI:10.1109/jsen.2026.3672611
摘要
Real-time monitoring of concentrate ash content is essential for effective control of coal flotation circuits. Current research predominantly focuses on optimizing convolutional neural network architectures to improve accuracy, yet neglects traditional ash-related features and interpretability, leading to excessive model complexity with marginal accuracy gains. This paper proposes a novel explanatory nonlinear feature fusion method for froth image analysis in coal flotation. The methodology integrates optimized convolutional feature slicing with traditional feature engineering, employing four machine learning techniques to predict concentrate ash content while elucidating feature contribution mechanisms through Shapley Additive exPlanations. In the feature fusion experiment, initial fusion features comprising 98 traditional features and 256 convolutional features demonstrated suboptimal performance in direct modeling. After feature selection, the XGBoost model achieved superior performance supported by 82 optimized features. These features exhibited descending importance from morphoscopic features to convolutional components, with the area feature showing the highest mean (|SHAP values|) (0.290631) and CNN59 the lowest mean (|SHAP values|) (0.004135). The optimized XGBoost model attained MAE=0.6139, RMSE=1.1547, and R²=0.7948. Force plot analysis revealed that morphoscopic and grayscale histogram features from traditional engineering were indispensable, while convolutional features exerted critical influences and statistical features contributed minimally. This investigation establishes a methodological framework for feature selection and modeling in coal flotation froth image analysis, providing theoretical and practical foundations for industrial applications.
科研通智能强力驱动
Strongly Powered by AbleSci AI