Research On Intelligent Decision-Making Models And Methods For Railway Vehicle Welding Manufacturing

The welded structure is an essential component of railway vehicles,and its manufacturing process involves various stages,such as structural and process design,welding production,and quality control.Decision-making in welding manufacturing involves multiple disciplines,stages,and data integration.Relying solely on manual decision-making is challenging to meet the needs of modern manufacturing for railway vehicles.Manual decision-making can lead to arbitrariness,uncertainty,limited reuse,and repetition of work,resulting in errors and information silos.To address these issues,introducing information technology and intelligent theory and methods into traditional welding manufacturing can provide decision support for the all-around,multi-level,and whole life cycle of railway vehicle welding intelligent manufacturing.This can enrich the theoretical system of welding decision-making,and transform and upgrade welding structure manufacturing mode,with crucial theoretical research value and practical engineering significance.Based on the background of decision-making in railway vehicle welding manufacturing,our paper focuses on several challenges,such as scattered decision-making systems and weak information association in engineering decision-making tasks,uncertain attribute relationships,and poor interpretability of structured data,as well as the limited research on knowledge acquisition,representation,and decision support of unstructured data and difficulties in modal information decision-making.In order to address these issues,many techniques are employed to develop an intelligent decision-making framework tailored to different data forms and characteristics,such as neighborhood rough sets,machine learning models,association rules,natural language processing,knowledge graphs,neural networks,and other related theories and methods.Our research covers three main aspects.(1)An interpretable fusion decision model based on neighborhood rough sets and XGBoost is proposed and developed.To address the issues of knowledge and rule induction,low decision-making accuracy,and poor interpretability in structured data decision-making for railway vehicle welding,an interpretable fusion decision model based on neighborhood rough sets and XGBoost is proposed.The model simplifies attributes using neighborhood rough set analysis,establishes relational associations between uncertain attributes,and combines these results with XGBoost to construct a high-quality decision model.Decision trees and SHAP analysis are applied to explain the model and address poor model interpretability.The engineering applicability and validity of the proposed model are demonstrated through the welding process documentation process.(2)Introducing natural language processing and knowledge graphs into the railway vehicle welding domain enables the acquisition,representation,and decision support of domain knowledge.The effective knowledge extraction models are obtained by comparing the training named entity recognition and relationship extraction models,respectively,to achieve automatic knowledge extraction from unstructured welded documents and construct domain knowledge graphs for knowledge representation and decision support.For structured data,a relationship discovery model is proposed based on neighborhood rough sets and association rules.Based on rough neighborhood sets,attribute associations are determined,and attribute value associations are established based on association rules to extract knowledge to support knowledge graph construction.An engineering application of the EN15805 series of standards is used as an example to construct the knowledge graph and to illustrate and validate its decision-making process.(3)A modal data decision-making model based on CNN and knowledge graph is proposed and implemented,and a three-tier decision-making framework with perceptual capabilities is constructed.A decision model is proposed based on CNN and the knowledge graph for the modal data and engineering decision problem.Based on this model,a three-layered decision framework is constructed,comprising a perception,data,and inference layer.The perceptual layer uses welded joint sketches to obtain image feature information through convolutional neural networks.The data layer utilizes natural language processing techniques to build a knowledge graph.The inference layer incorporates multiple models and methods to construct a hybrid inference system.In order to demonstrate the effectiveness of the proposed framework,a prototype system of welding decision support is developed for railway vehicle bogies.