西安科技大学学报2025，Vol.45Issue(5)：904-915,12.DOI:10.13800/j.cnki.xakjdxxb.2025.0507
多尺度特征融合的岩性识别模型
A multi-scale feature fusion model for lithology identification
摘要
Abstract
To address the issues in the existing lithology recognition methods,such as insufficient fea-ture representation,limited model generalization ability and low recognition accuracy,a novel lithology recognition model based on multi-scale feature fusion was proposed,named MSFR-Net.A Feature Ex-traction Module(FEM)was designed from the raw logging curves.It combined Bidirectional Gated Re-current Units(BiGRU)and an even-odd sequence interaction mechanism to perform deep mining of multi-scale features from the logging curves.A Random Convolution Module was constructed,which uti-lized a dynamic convolution kernel parameter optimization strategy to effectively capture spatial correla-tions between strata features.Based on a decision module consisting of six base classifiers,a multi-clas-sifier collaborative decision mechanism was employed to improve the model's classification performance and complete the lithology recognition task.The results show that MSFR-Net performs better in such key metrics as accuracy,precision,recall,and F1-score in tests based on real logging data,compared to ten commonly used lithology recognition models,such as SVM and GRU.In the D-well experiment,MSFR-Net achieves a prediction accuracy of 95.1％for major lithology categories and 72.7％for mi-nor lithology categories,an improvement of 13.27％over SVM and 12.61％over LSTM.MSFR-Net,through the synergistic optimization of multi-scale feature fusion and ensemble learning strategies,effec-tively enhances the extraction of key geological features and the model's generalization performance,providing a new technical approach for lithology recognition.关键词
测井曲线/岩性识别/特征融合/Rocket网络/多分类器Key words
well log curves/lithology identification/feature fusion/Rocket network/multi-classifier分类
能源科技引用本文复制引用
贾澎涛,成宇超,蒋永杰,李娜..多尺度特征融合的岩性识别模型[J].西安科技大学学报,2025,45(5):904-915,12.基金项目
国家自然科学基金项目(62002285) （62002285）
