引用本文: 刘泽民, 程海勇, 毛明发, 李在利, 吴顺川, 姜关照, 孙伟, 刘伟铧. 基于3D卷积神经网络的膏体屈服应力预测[J]. 工程科学学报.
LIU Zemin, CHENG Haiyong, MAO Mingfa, LI Zaili, WU Shunchuan, JIANG Guanzhao, SUN Wei, LIU Weihua. Prediction of paste yield stress based on three-dimensional convolutional neural networks[J]. Chinese Journal of Engineering.
 Citation: LIU Zemin, CHENG Haiyong, MAO Mingfa, LI Zaili, WU Shunchuan, JIANG Guanzhao, SUN Wei, LIU Weihua. Prediction of paste yield stress based on three-dimensional convolutional neural networks[J]. Chinese Journal of Engineering.

## Prediction of paste yield stress based on three-dimensional convolutional neural networks

• 摘要: 膏体流变性能是膏体充填技术重要指标，是金属矿膏体充填工艺流程的重要工程参数. 本文提出一种基于3D卷积神经网络的膏体屈服应力预测方法，通过制定图像采集标准并研发图像采集装置采集图像数据集. 经Sobel算子实现膏体边缘检测、全图缩小等预处理，得到膏体图像数据集. 采用十折交叉验证方法划分数据集，避免因单次随机划分造成的偶然误差. 以膏体图像–屈服应力数据集为基础，利用3D卷积神经网络模型提取膏体纹理特征和时序信息等，又通过引入直方图均衡化算法的图像增强策略减少环境因素干扰，提高模型稳健性. 利用预处理后的数据集在3D卷积神经网络模型上做训练和测试，得到模型损失值曲线图和混淆矩阵. 将屈服应力模型预测结果进行分析，又引入卷积注意力机制嵌入到卷积神经网络实现模型优化，并对模型参数进行调整，模型预测平均准确率从93.26%提升至98.19%，论证了基于3D卷积神经网络的膏体屈服应力预测方法可行性. 经图像增强处理的数据集应用到各模型中，模型预测平均准确率均提升3%以上. 相比传统膏体流变测量方式，解决了传统膏体屈服应力测量操作复杂、外部因素扰动大、工程现场难以开展等问题.

Abstract: The rheological properties of paste are the foundation of the paste-filling process in metal mines, and paste yield stress is an important evaluation index for paste-filling technology. The change in ratio and concentration has a significant impact on the texture and appearance of paste slurry. Herein, a method for predicting the paste yield stress using three-dimensional convolutional neural networks (3D CNNs) is proposed through the development of image acquisition standards and an image acquisition device to collect image data sets based on a paste image data set. The Sobel operator is used to realize the pretreatment of paste edge detection and full size shrinking, and the paste image data set is obtained. The ten-fold cross-validation method is used to divide the data set to avoid accidental errors caused by a single random division. Based on the paste image–yield stress data set, the 3D CNNs model is used to extract the depth features and timing information on the paste. An image enhancement strategy for the histogram equalization algorithm is introduced to reduce the interference of environmental factors. The preprocessed data set is used for training and testing the 3D CNNs network model. In addition, the prediction accuracy of the yield stress model is analyzed: the convolutional attention block module is embedded into the CNN to optimize the model, and the introduction of channel attention and spatial attention enhances the ability of the model to perceive important areas in the image, which helps improve its ability to capture important information in the image and adjust the model parameters. The prediction accuracy of the model is increased from 93.26% to 98.19%, and the sample prediction error is within 20%, demonstrating the feasibility of paste yield stress prediction based on 3D CNNs. The image enhancement strategy using the histogram equalization algorithm can significantly improve the prediction accuracy of paste yield stress. The image enhancement strategy is applied to each model experiment, and the model prediction accuracy is improved by more than 3 percentage points. The developed image acquisition device and image acquisition standard can reduce the disturbance of environmental factors on image recognition and ensure the accuracy of paste yield stress prediction. Compared with the traditional paste rheological measurement method, the proposed method solves the problems of complex operation of traditional paste yield stress measurement, strong interference of external factors, and the difficulties associated with engineering sites.

/

• 分享
• 用微信扫码二维码

分享至好友和朋友圈