【目的】黑盒优化算法广泛应用于各类复杂工程设计场景，但其固有的运算过程复杂、内部逻辑不透明，导致算力消耗大。在电机优化设计过程中，黑盒优化算法通过反复迭代运算和有限元仿真筛选最优结构参数，大幅拉长了设计周期，制约工程应用效率。传统代理模型虽能降低算力成本，但适配性较差，难以搭建适用于多工况的通用数据库，设计需求和应用场景出现小幅变动后，需重新采集数据、训练模型，存在泛化能力弱、数据复用率低的工程缺陷。【方法】针对上述问题，本文提出一种基于迁移代理模型(TSM)的黑盒优化算法。首先，依托迁移学习理论搭建全新TSM，通过模型自主筛选源域优质有效样本，深度挖掘数据内在特征，提升模型在目标域小样本条件下的预测精度与泛化能力。然后，将优化后的TSM与迭代优化算法深度融合，删减优化过程中的冗余计算步骤，进一步提升目标优化任务的执行效率。最后，通过仿真测试所提算法的实际应用效果。【结果】仿真结果表明，相较于传统优化算法，本文所提基于TSM的优化算法可有效降低算力损耗，显著提升电机参数优化的整体工作效率，同时能够加快模型迭代收敛速度，在多变的电机设计工况中具备更稳定的表现。【结论】本文所提基于TSM的优化算法，从数据利用与算法融合层面解决了传统黑盒优化算法和常规代理模型的固有短板，突破了传统模型场景适配性差、重复训练成本高的局限。所提算法能高效适配实际工程的多变设计需求，为电机智能优化设计以及同类复杂工程黑盒优化问题提供可靠、高效的技术支撑，具备极高的工程落地价值与推广前景。

[Objective] Black-box optimization algorithms are widely used in various complex engineering design scenarios. However, their inherent computational complexity and internal operational opacity result in intensive computing consumption. In the optimal design of motors, black-box optimization algorithms require numerous iterative computations and finite element simulations to screen optimal structural parameters, which significantly extends the design cycle and restricts engineering application efficiency. Although traditional surrogate models can reduce computational costs, they exhibit poor adaptability and struggle to construct universal databases for multiple operating conditions. Slight changes in design requirements or application scenarios necessitate data recollection and model retraining, leading to inherent engineering drawbacks including weak generalization capability and low data reusability. [Methods] To solve the above problems, a black-box optimization algorithm based on transfer surrogate model (TSM) was proposed. Firstly, a novel TSM was constructed based on transfer learning theory. By autonomously screening high-quality and effective samples from the source domain, the inherent data features were deeply mined to improve the prediction accuracy and generalization ability of the model under the small-sample conditions of the target domain. Subsequently, the optimized TSM was deeply integrated with the iterative optimization algorithm to eliminate redundant computational steps during the optimization process and further enhance the execution efficiency of the optimization task. Finally, simulation tests were conducted to verify the practical effectiveness of the proposed algorithm. [Results] The simulation results showed that, compared with the traditional optimization algorithm, the proposed optimization algorithm based on TSM effectively reduced computational consumption, significantly improved the overall efficiency of motor parameter optimization, and accelerated the iterative convergence speed of the model. The proposed algorithm exhibited superior stability under variable motor design conditions. [Conclusion] The proposed optimization algorithm based on TSM addresses the inherent shortcomings of traditional black-box optimization and conventional surrogate models from the perspectives of data utilization and algorithm integration. It breaks through the limitations of poor scenario adaptability and high retraining costs existing in traditional models. This scheme can efficiently adapt to the variable design requirements of practical engineering scenarios. It provides reliable and efficient technical support for the intelligent optimal design of motors and other complex engineering black-box optimization problems, and possesses excellent engineering applicability and popularization prospects.

国家自然科学基金(52407010)