基于对称双解码结构的多任务协同初至拾取框架OA

Data-driven intelligent first-arrival picking methods primarily rely on supervised learning using labeled data.However,a significant disparity exists between the representational capacities of mainstream local-focus la-beling methods and global-segmentation labeling methods.The former emphasizes the learning of local wave-form details of first arrivals,while the latter focuses on capturing global structural features of the wavefield,re-sulting in constrained learning perspectives for traditional single-task models.Consequently,this paper proposes a Multi-task collaborative first-arrival picking framework based on a symmetric dual-decoders structure(MT-SDD picking).First,by leveraging the complementary representational strengths of these two labeling strate-gies,the MT-SDD framework employs a dual-path decoder with a symmetrical configuration and incorporates a feature fusion module that integrates Transformer blocks with a multi-scale convolutional pyramid.Second,through a multi-stage optimization strategy,the framework progressively directs the model in transitioning from learning single-perspective features to multi-perspective feature fusion,ultimately achieving a substantial ad-vancement in both picking accuracy and stability.Finally,comprehensive ablation experiments substantiate the rationale and efficacy of the MT-SDD framework,demonstrating its superior performance in high-precision pick-ing and cross-site generalization capabilities.Validation results based on active seismic data from the Halfmile and Brunswick mining areas in Canada indicate that,in comparison to traditional single-task picking models,the proposed method reduces the average first-arrival picking error by 20%.Accuracy rates across various error tolerances have been consistently enhanced:at error thresholds of 2 ms,4 ms,8 ms,and 10 ms,the accuracy achieved 95.6%,97.8%,98.9%,and 99.2%,respectively.These figures represent improvements of 0.7%,0.3%,0.1%,and 0.1%over the single-task baseline method.