搜索

x
中国物理学会期刊

一种面向舰船磁场反演的全驻留矩阵化增强型烟花算法

A Full-Resident Matrix-based Enhanced Fireworks Algorithm for Ship Magnetic Field Inversion

PDF
导出引用
  • 针对舰船等效源建模中磁场反演问题的高维、多峰、非线性特征,以及由此带来的巨大计算负担,本文提出一种全驻留矩阵化增强型烟花算法(FRM-EFWA),旨在增强全局搜索能力并大幅提升计算效率。在搜索机制上,算法以增强型烟花算法(EFWA)为基础融合了引导型高斯变异、锦标赛选择以及无放回排序采样策略,在“大范围搜索”与“小范围寻优”之间实现了自适应平衡,有效提升了搜索效率与求解精度;消融实验证实了各算子改进对整体性能提升的实际贡献。在计算架构上,借鉴全矩阵化进化计算(MEC),将种群全部个体映射为扁平紧凑的统一高维张量,以矩阵化并行方式驱动全部迭代计算,并采用无分支逻辑控制,使算法全流程均在GPU显存内闭环完成,仅初始化与结果输出涉及CPU交互,从而大幅提升搜索与计算效率。等效磁源的磁偶极子阵列反演实验显示,本算法较传统CPU调度GPU架构计算效率提升约95倍,较主流CPU-GPU协同架构提升约1.71倍。针对36维磁偶极子阵列模型,重构磁场在观测点处的平均相对误差低至0.19%,平均绝对误差0.53 nT,重构磁场与目标磁场在观测平面上完全吻合。在抗异常数据测试中,算法能精准定位故障采样点;在抗随机噪声测试中,重构误差维持在底噪量级。电性源模型的磁场反演进一步证实,该方法在无降维策略的辅助下,具备直接于广域空间锁定全局最优解的寻优能力。本文为复杂舰船磁场的高保真快速反演提供了一条高效的算法解决方案。

     

    To address the high-dimensional, multi-peak, and nonlinear characteristics of magnetic field inversion in ship equivalent source modeling, as well as the substantial computational burden involved, this paper proposes a Full-Resident Matrix-based Enhanced Fireworks Algorithm (FRM-EFWA), aiming to enhance global search capability and significantly improve computational efficiency. In terms of the search mechanism, the algorithm integrates guided Gaussian mutation, tournament selection, and a sampling-without-replacement strategy into the Enhanced Fireworks Algorithm (EFWA) framework, achieving an adaptive balance between large-scale exploration and fine-grained exploitation, thereby effectively improving the ability to approach the global optimum. Ablation experiments confirm the actual contribution of each operator improvement to the overall performance enhancement. In terms of the computing architecture, drawing on Matrix Evolutionary Computation (MEC), all individuals in the population are mapped into a flat and compact unified high-dimensional tensor, with the entire iterative computation driven in a matrix-based parallel manner. By adopting branchless logic control, the entire algorithmic workflow is executed in a closed loop within GPU memory, with only initialization and final result output involving CPU interaction, thereby substantially increasing both search and computational efficiency. Magnetic dipole array inversion experiments for equivalent magnetic sources show that, compared with traditional CPU-scheduled GPU architectures, the computation efficiency of the proposed algorithm is improved by approximately 95 times, and by about 1.71 times compared with mainstream CPU-GPU cooperative architectures. For the 36-dimensional magnetic dipole array model, the average relative error of the reconstructed magnetic field at observation points is as low as 0.19%, with an average absolute error of 0.53 nT. In the anti-abnormal-data test, the algorithm can accurately locate faulty sampling points; in the anti-noise test, the reconstruction error remains at the noise-floor level. Magnetic field inversion of the electrical source model further confirms that, without the aid of dimensionality reduction strategies, the proposed method possesses the optimization capability to directly lock onto the global optimum within the wide-area search space. This paper provides an efficient algorithmic solution for the high-fidelity and rapid inversion of complex ship magnetic fields.

     

    目录

    /

    返回文章
    返回