融合多模态特征与深度学习的沉积单元界限智能对比方法OA

To improve the automation,objectivity,and engineering efficiency of sedimentary unit boundary correlation in the Sanan development area of the Daqing oilfield,where dense well patterns and highly subdivided stratigraphic sequences make manual correlation inefficient,subjective,and inconsistent in complex intervals,an intelligent correlation method integrating multimodal features and deep reinforcement learning is developed.Using SP,GR,RLLD,and AC logging curves together with their roughness attributes,this study investigated collaborative representation of multimodal logging information,optimization of inter-well boundary matching paths,and automatic identification mechanisms for complex intervals.First,multi-source logging curves are synchronously sampled,standardized,smoothed,and organized into sliding-window inputs.Then,convolutional feature extraction,bidirectional long short-term memory(BiLSTM)temporal modeling,and attention-based adaptive fusion are employed to obtain integrated features for boundary identification.Furthermore,curve-roughness similarity measurement,a variable-window search strategy,and deep Q-network(DQN)-based sequential decision-making are combined to achieve automatic boundary matching and global path optimization between reference wells and target wells.The results show that,based on measured data from 1 500 wells in 23 blocks,the proposed method achieved an overall matching rate of 94.1%on 1 350 test wells,with an average processing time of approximately 52 s per well.Among 105 000 sedimentary unit boundary samples,78.1%of the absolute boundary-depth errors are controlled within 0.3 m,89.5%within 0.5 m,and 96.2%within 0.7 m.The comprehensive identification accuracy for six major marker beds is about 94.3%;among them,four marker beds achieved accuracies higher than 95%,while the remaining two reached about 91%～93%.Compared with the traditional correlation coefficient plus gradient method,as well as convolutional neural network(CNN)-BiLSTM and graph neural network(GNN)models,the proposed method outperforms comparative methods across all metrics including stratified identification accuracy,F1 score,intersection over union(IoU),overall matching rate,and processing efficiency.It is concluded that this method can effectively improve the accuracy,stability,and engineering applicability of sedimentary unit boundary correlation under large-scale well-network conditions,and can provide reliable technical support for fine stratigraphic correlation,reservoir characterization,and subsequent geological modeling.