A multimodal knowledge graph and two-stage reinforcement learning-based intelligent configuration approach of improvement design scheme for complex products

Within the modern production paradigm centred on mass customisation, efficient improvement and iterative design of complex products are vital for enhancing the competitive advantage of enterprises. However, their efficiency is severely hampered by challenges in converting heterogeneous and multimodal lifecycle data related to product design into integrated knowledge assets and intelligently configuring the optimal improvement design scheme. To address this, an improvement design scheme intelligent configuration (IDSIC) approach of complex products based on multimodal knowledge graph (MKG) and two-stage reinforcement learning (RL) is developed. First, an MKG construction method is proposed that integrated ontology, graph neural network (GNN), and TransR. Second, a two-stage RL mechanism that can leverage the rich semantic information of MKG to implement IDSIC is developed. The developed mechanism can identify a design baseline via constraint filtering and multi-objective optimisation, then determines component prioritisation and specifications. A case study on the improvement design of axle box bearings for high-speed electric multiple units (EMUs) validated the proposed approach. Compared with existing methods, the proposed approach enables more requirement-consistent, interaction-compatible, and robust improvement design scheme configuration. The results demonstrate its feasibility and effectiveness for complex product design with decomposable structures, explicit component interactions, and large specification spaces.Abbreviations: AC: Actor-critic; AI: Artificial intelligence; Bert: Bidirectional encoder representations from transformers; BiLSTM: Bidirectional long short-term memory; CAD: Computer-aided design; CBR: Case-based reasoning; CFRP: Carbon fibre reinforced polymer; CRF: Conditional random fields; D3QN: Double duelling deep Q network; EMUs: Electric multiple units; FAHP: Fuzzy analytic hierarchy process; FBS: Function-Behaviour-Structure; GAT: Graph attention network; GNN: Graph neural network; IDSIC: Improvement design scheme intelligent configuration; KG: Knowledge graph; LLMs: Large language models; MC: Monte Carlo; ML: Machine learning; MDP: Markov decision process modelling; ML: Machine learning; MKG: Multimodal knowledge graph; MRR: Mean reciprocal rank; NER: Named entity recognition; NLP: Natural language processing; OM: Operation and maintenance; OWL: Web ontology language; PPO: Proximal policy optimisation; RDF: Resource description framework; RL: Reinforcement learning