Electric Drive for Locomotives

Quick Entry

Author Login
Expert Center
Editor Center
Legacy system entry

Quick Icon

Current Issue
Articles in Press
Back Issue Browse

2026年第3期

Digital Intelligence Integration and Intelligent Operation and Maintenance
Digital-intelligent transformation pathways for rail transit equipment enterprises AI导读

LIU Kean, XUE Haifeng, CHENG Zhen, MA Zhongming, CHEN Jian

DOI：10.13890/j.issn.1000-128X.2026.03.001

摘要：Rail transit equipment enterprises face systemic challenges in digital-intelligent transformation, including inadequate top-level planning, misaligned strategy execution, lagging process optimization, absent data standards, and fragmented system development. To address these challenges, this paper proposes EA-driven implementation pathways with enterprise architecture as the core top-level design tool.By constructing a five-in-one collaborative construction model integrating the business-data-application-technology-security (BDATS) framework, this paper detailedly designs construction pathways for business architecture, data architecture, application architecture, technology architecture, and security architecture, thereby forming a closed-loop mapping in which business defines data, data supports applications, and applications realize business objectives. Using CRRC as the practical subject has verified the feasibility and effectiveness of the proposed pathways. The results demonstrate that these pathways can effectively support corporate strategy implementation, enable data-driven decision making and intelligent business empowerment, and provide a reusable methodology and engineering practice reference for systematically advancing digital-intelligent transformation in rail transit equipment enterprises.
关键词：digital-intelligent transformation;rail transit;enterprise architecture;intelligently made in China

92

|

31

|

0

<HTML>
<L-PDF><WORD><Meta-XML>

<引用本文> <批量引用> 171875212 false

更新时间：2026-09-20
Research on evaluation methods for industrial large model AI导读

CHEN Jian, JIANG Dachen, FU Yunxiao

DOI：10.13890/j.issn.1000-128X.2026.03.002

摘要：In the context of the manufacturing industry's digital transformation driven by Industry 4.0, industrial large model (ILM) serve as core technical engines that can significantly improve enterprise production efficiency. However, their application must meet stringent requirements of high precision, low latency, and high robustness. ILM evaluation is key link for building a trusted intelligent industrial ecosystem and holds the core value of ensuring security, optimizing efficiency, unifying standards, and promoting industry-academia-research-application collaboration. This paper presents a full-life-cycle multi-dimensional evaluation framework, addressing five specific evaluation challenges: multi-model data and privacy protection, low latency, physical consistency, security classification, and decision traceability. Centered on an "Evaluation System Cube" and a "Pyramid Three-level Evaluation Model", this framework defines an evaluation metric system covering evaluation dimensions, life cycle stages and evaluation levels , etc. The methodology is implemented through a three-level method: simulation verification, hardware-in-the-loop simulation, and field pilot testing. Verification across multiple scenarios demonstrates that the framework can meet the evaluation requirements of ILMs throughout the full life cycle. For full-process verification, this study adopted a typical scenario from rail transit operation and maintenance: intelligent fault diagnosis of components that otherwise require visual inspection; the results show that the framework can effectively quantify ILM performance, security, and application value. These results demonstrate strong engineering feasibility and a foundation for cross-industry promotion, providing standardized methodological support for trustworthiness enhancement and large-scale ILM deployment.
关键词：industrial large model (ILM);evaluation system;full life cycle;three-level progressive evaluation;trusted intelligent industry

56

|

27

|

0

<HTML>
<L-PDF><WORD><Meta-XML>

<引用本文> <批量引用> 171875184 false

更新时间：2026-09-20
Methodology for constructing domain-specific models for unstructured knowledge data segmentation of rail transit AI导读

FENG Lijie, LIU Zhenbo, WANG Tian, ZHANG Qian

DOI：10.13890/j.issn.1000-128X.2026.03.003

摘要：Digital-intelligent transformation in the rail transit industry is hindered by poor retrieval and limited reuse caused by the unstructured storage of historical scanned archives and foreign language standards. To address this, this paper proposes a document structuring approach that combines layout semantic segmentation, multimodal extraction, hierarchical restoration, and graph modeling to segment rail transit documents into regions such as titles, body text, tables, and formulas. Integration with optical character recognition (OCR), formula parsing, and table restoration enables clause-level structured data storage. Furthermore, a traceable knowledge graph is constructed in the graph database to support knowledge graph-based retrieval augmented generation (KG-RAG) for question answering. Rather than redesigning the basic detection model, the proposed method integrates clause-level structuring with graph-enhanced retrieval for complex rail transit documents, supporting systematic conversion of unstructured documents into traceable knowledge units. Its effectiveness was validated by training a YOLO-based layout semantic segmentation model. Compared to conventional OCR (text recognition rate of approximately 68%) and general layout analysis models such as RapidLayout (overall recognition rate of 71.3%), the method achieves an overall recognition rate exceeding 80%, representing improvements by at least 12% and 8.7%, respectively. Specifically, the detection accuracy exceeds 90% for titles and body text; the OCR text recognition rate reaches approximately 95% (an approximately 27 percentage points improvement over conventional OCR); the formula parsing accuracy is 80%; and the table restoration accuracy is 85%. Further integration with KG-RAG yields retrieval performance: Recall@5 > 0.88, MRR@5 = 0.86, and recall completeness > 85%, with the ability to output clause numbers and node IDs for traceability. The retrieval accuracy is improved by more than 15% compared to conventional methods. These results demonstrate that high-quality document structuring enhances RAG retrieval performance and interpretability, providing a methodological pathway for the itemized management of existing rail transit documents and for the application of intelligent techniques.
关键词：document fragmentation;YOLO;OCR;graph database;RAG;artificial intelligence application;knowledge engineering

36

|

27

|

0

<HTML>
<L-PDF><WORD><Meta-XML>

<引用本文> <批量引用> 171875332 false

更新时间：2026-09-20
Design and implementation of a knowledge retrieval system based on two-stage fusion AI导读

LUO Shuting, LI Na, CHEN Rongzhen

DOI：10.13890/j.issn.1000-128X.2026.03.004

摘要：To address the challenge of balancing precise term matching and complex semantic understanding in conventional keyword-based and semantic retrieval for domain-specific knowledge bases operating under resource constraints in industrial intranet environments and with highly standardized terminology, this paper proposes a two-stage fusion retrieval algorithm framework tailored to industrial scenarios and presents its full-link engineering implementation in a knowledge retrieval system. Building on classic BM25 and vector retrieval, a two-stage progressive retrieval strategy was designed, which combines score-level λ-weighted linear fusion with ranking-level reciprocal ranking fusion (RRF). Through parameter sensitivity analysis on domain-specific datasets, optimal hyperparameters (λ=0.9, k=30) for the two-stage fusion were identified, achieving a synergistic balance between precise matching and semantic association. Additionally, a full-link retrieval system was developed, incorporating mechanisms for timed synchronization and incremental updating of remote data sources, to support the complete process from data processing to frontend presentation. Further effective integration with conventional retrieval algorithms enabled scenario-specific design and validated the engineering implementation of the proposed two-stage fusion retrieval framework, yielding an industrial knowledge retrieval solution that integrates the algorithmic strategy and system implementation.
关键词：two-stage fusion strategy;linear weighted fusion;reciprocal ranking fusion;industrial knowledge base;knowledge retrieval system

20

|

25

|

0

<HTML>
<L-PDF><WORD><Meta-XML>

<引用本文> <批量引用> 171874929 false

更新时间：2026-09-20