特征迁移的细粒度产品形态智能决策方法

Intelligent Decision-Making of Fine-Grained Product Form with Feature Transfer

针对产品形态智能决策框架系统性不强、模型决策机制单一且历史样本数据量少等问题,提出一种基于混合迁移学习的细粒度产品形态智能决策方法.该方法将Swin Transformer和ResNets作为骨干网络设计了3个并行混合迁移学习子网络,包括产品形态识别网络(Form-CN)、产品形态深度回归评价网络(Form-REN)和产品形态分布拟合评估网络(Form-DFEN).首先应用Form-CN对产品进行细粒度形态分类判别,实现产品形态设计定位识别任务;其次应用Form-REN对产品整体形态语义进行预测评价;然后通过Form-DFEN对产品形态进行分布拟合评估;最后由Form-REN和Form-DFEN完成综合决策.以创建的手电钻数据集进行实验,并与其他经典模型进行比较,结果表明,所设计的3个网络分别取得了99.0%的准确率、0.2583的均方误差和84.3%的准确率;所提方法能够精细、高效地辅助设计师进行综合智能决策,为产品形态智能决策提供了一个更为系统的参考框架.

A fine-grained product form intelligent decision-making method based on hybrid transfer learning is proposed.The aim is to solve the problems of weak systematicity of the product form intelligent deci-sion-making framework,single decision mechanism of the model,and a small amount of historical sample data.The method uses Swin Transformer and ResNets as the backbone network to design three parallel hy-brid transfer learning sub-networks,including product form classification network(Form-CN),product form deep regression evaluation network(Form-REN)and product form distribution fitting evaluation network(Form-DFEN).Firstly,Form-CN is applied to classify the products with fine-grained form to achieve the product form design location identification.Secondly,Form-REN is applied to predict and evaluate the overall product form semantics.Then,Form-DFEN is used to evaluate the product form by distribution fit-ting.Finally,Form-REN and Form-DFEN are used to complete the integrated decision making.Experiments were performed on the created hand drill dataset and compared with other classical models.The results show that the three designed networks achieved 99.0%accuracy,0.2583 mean square error and 84.3%accuracy,respectively.The proposed method can finely and efficiently assist designers to make comprehensive intel-ligent decisions,which provides a more systematic reference framework for intelligent decision-making of product forms.