基于生成对抗网络的岩心薄片岩性智能识别方法

Intelligent lithology identification method of core slices based on generative adversarial network

传统人工识别沉积岩岩心薄片岩性的方法,需要大量的专业人员,耗时耗力,且鉴定结果也受个人感官认识、主观性等诸多因素影响。为此,提出了基于生成对抗网络的岩心薄片岩性智能识别方法:首先利用岩心薄片数据对生成的对抗模型进行对抗训练;然后用训练好的生成器生成模拟图像扩充数据集,扩充原始岩心薄片图像数据集,增加数据集的多样性,可以提高模型预测精度;再将判别器卷积层参数迁移至岩心薄片的岩性识别模型中,保留判别器提取的岩心薄片特征;最后训练模型中2个全连接层和softmax分类层,建立岩心薄片岩性识别模型。“WGAN+判别器参数迁移”的岩心薄片岩性识别方法进一步提高了岩性识别准确率。该模型可以智能识别出岩心薄片的岩性,实现对沉积岩岩性的智能分类。通过实验数据对比分析,该模型的准确率达到了94.93%,优于传统方法,具有较强的实践应用价值。

The traditional artificial identification method of of the lithology of sedimentary core slices requires a large number of professionals,which is time-consuming and labor-intensive,and the identification results are also affected by many factors such as personal sensory perception and subjectivity.Therefore,an intelligent lithology identification method of core slices based on generative adversarial network was proposed.First,the core slice data was used toconduct confrontation training on the generated adversarial model.The trained generator was then used to generate a simulated image expansion dataset to expand the original core slice image dataset to increase the diversity of the dataset,which can improve the model prediction accuracy.Then,the convolution layer parameters of the discriminator were transferred to the lithology identification model of the core slices,and the core slice features extracted by the discriminator were retained.Finally,two fully connected layers and softmax classification layer in the model were trained to establish the lithology identification model of core slices.The core slice lithology identification method of“WGAN+discriminator parameter migration”further improves the accuracy of lithology identification.This model can intelligently identify the lithology of core slices and realize the intelligent classification of lithology of sedimentary rocks.Through comparativeanalysis of experimental data,the accuracy of the model reaches 94.93%,which is better than traditional identification methods and has strong practical application value.