Data-Driven Modular Segmentation Of Complex Product And Prediction Of Design Change Link

Due to various factors such as customer demand and technological iteration,design change issues exist throughout the entire lifecycle of complex product development,which inevitably leads to an extension of the research and development and manufacturing cycle of complex products.Moreover,in the continuous evolution process of complex products,their complexity increases sharply,presenting characteristics such as increasingly complex structures,continuously increasing number of components,continuously increasing coefficient dimensions,and gradually strengthening coupling relationships,making the analysis of the propagation effects of design changes more complex.However,a reasonable modular segmentation method can effectively control the research and development cycle and complexity of complex products,and based on modular segmentation,further achieve accurate prediction of the design change link.Therefore,this thesis uses actual production data as production material and solves the above problems from a data-driven perspective through scientific methods.The main research work of this thesis is as follows:(1)Modularity acquisition of complex product networks based on implicit coupling features mining: The existing modular research methods have limitations in considering the non-linear coupling relationships between functional features of complex product components,resulting in imprecise module partitioning.Therefore,a complex product network modularization acquisition method based on implicit coupling feature mining is proposed to address this issue.Firstly,the complex product network graph is constructed based on the structural relationship and functional features of the components,and the Page Rank algorithm and cosine similarity are used to adaptively calculate the number of modules.Secondly,the Graph Auto-Encoder(GAE)is used to extract the implicit coupling features of the non-linear relationships between the components,enhancing the expression of component functional features.Finally,the Fuzzy C-Means(FCM)clustering algorithm is used to obtain the final modularization scheme.The effectiveness and superiority of the proposed method in module partitioning are verified using a self-cleaning filter as a case study.(2)Prediction for complex product design change link combining space-time characteristics: Based on the research content(1),first,a complex product feature network graph is constructed by expanding the connection relationships between modules and components based on the physical structural relationship and functional features of the complex product.Secondly,considering the lack of historical change data,the similarity between complex product design change prediction and information cascade prediction is taken into account,and an information cascade prediction model is constructed.The Topo-LSTM is used to extract the temporal features in the information propagation process,and the Graph Attention Network(GAT)is used to extract the overall spatial features of the network.Then,the Self-Attention Mechanism is used to fuse the temporal and spatial features to complete the construction of the information cascade prediction model.Finally,the constructed information cascade prediction model is migrated to the field of complex product design change and used to construct a complex product design change prediction model.The effectiveness of the proposed method is verified through an example.This thesis presents a complex product network modularization acquisition method based on implicit coupling feature mining and a complex product design change prediction algorithm that integrates temporal and spatial features to address the issues of modularization and design change prediction in complex products.Firstly,the proposed method solves the problem of subjective factors in the process of modularization partitioning of complex products.Secondly,it addresses the issue of insufficient historical data in complex product design change prediction.Experimental results demonstrate the effectiveness and superiority of the proposed algorithms.