Chengke Wu

Post-approval environmental monitoring and audit (EM&A) are central to environmental impact assessment (EIA) effectiveness, but compliance evidence is often dispersed across manuals, monitoring reports, regulatory documents, and project records. This fragmentation delays exceedance checking and obscures how monitoring evidence is linked to thresholds, responsible parties, and response actions. This study develops an Environmental Impact Assessment Ontology (EIAO) implemented as a labeled property graph in Neo4j. The ontology integrates hazards, impacts, monitoring observations, regulatory thresholds, project entities, and mitigation measures, while rule-assisted queries support evidence retrieval, threshold checking, and traceable evidence-to-response pathways. The framework was evaluated at two scales. In a Hong Kong road project, seven compliance tasks required a mean of 1,827.3 s using manual workflows and 4.3 s using EIAO. The property-graph implementation also returned lower query times than a Resource Description Framework (RDF) benchmark for most retrieval- and path-oriented tasks, although RDF was faster for one rule-oriented task. Longitudinal validation transformed 124 public EM&A reports spanning November 2015 to June 2026 into 4,570 observations and a graph of 4,605 nodes and 9,175 relationships. The findings demonstrate how fragmented post-approval EIA records can be operationalized as auditable compliance evidence. These findings suggest that the monitoring–threshold–responsibility–response structure and document-to-graph workflow may also support post-approval EIA practice in other settings where compliance evidence remains document-based.

Purpose: This paper introduces a novel hybrid deep learning model aimed at enhancing human intention prediction in human–robot collaboration for utility tunnel maintenance. Recognizing the inherent dangers and the confined nature of utility tunnels, the study aims to advance safety and operational efficiency by improving robot assistants’ ability to accurately interpret and predict human-worker intentions. The purpose is to reduce human exposure to hazardous environments and optimize task execution through precise and timely robot actions. Design/methodology/approach: Our approach involves a hybrid deep learning architecture combining traditional time-series image classification with advanced semantic information extraction. The proposed hybrid intentions prediction deep learning model (HIPM) utilizes convolutional neural networks, long short-term memory networks and the contrastive language–image pre-training model for a comprehensive feature extraction and intention prediction. This integration enables the model to process visual and contextual data from dynamic and challenging tunnel environments, addressing the inadequacies of traditional vision-based and physical-based intention prediction methods. Findings: Empirical validation conducted on real-world utility tunnel data from Jiangsu Province, China, demonstrated that HIPM significantly outperforms traditional models. HIPM achieved a precision of 91.52% and a recall of 91.20%, indicating a high level of accuracy in predicting human intentions. The results underscore the model’s robustness and reliability, confirming its effectiveness in understanding and responding to complex human-worker behaviors in utility tunnel settings. Originality/value: The originality of this research lies in its novel integration of multimodal deep learning techniques to enhance the interpretability and adaptability of robots in human-robot collaborative environments. The HIPM model’s ability to interpret both human actions and environmental contexts presents a significant advancement over existing models, offering a more reliable and efficient approach to managing the safety and efficiency of utility tunnel maintenance operations. This study contributes a pioneering solution to the challenges of human intention prediction in one of the most hazardous and demanding industrial settings.

The rapid advancement of AI, especially large language models (LLMs), brings opportunities and challenges to industries. In construction, LLMs can enhance project coordination, support decision-making and reduce workload, but adoption is limited by hallucination, data security and domain complexity. This study investigates the current state of LLM adoption in China’s construction industry through a four-step approach, including a comprehensive literature review to identify potential drivers and barriers, questionnaire design and data collection for empirical analysis, and the application of the Entropy Weight Method (EWM) to quantify and rank the relative importance of each factor. The findings reveal that the top drivers originate at the company level, including strategic partnerships, internal research teams, and staff training—highlighting the central role of organizational readiness in enabling LLM integration. Conversely, the most critical barriers are embedded in the construction domain itself, including knowledge gaps, workflow integration, and data heterogeneity, which reflect structural limitations in the sector. Although LLM implementation remains in its early stages, survey responses show widespread optimism among stakeholders regarding its future potential. The study proposes several actionable strategies for both construction firms and policymakers to facilitate effective LLM adoption. Moreover, the identified drivers and barriers are not exclusive to construction but are also relevant to other digitally transforming sectors—such as manufacturing, healthcare, and energy—offering broader implications for AI adoption in complex, project-based environments.

Financial fraud detection in real-world scenarios presents significant challenges due to the subtlety and dispersion of evidence across complex, multi-year financial disclosures. In this work, we introduce a novel multi-agent reasoning framework AuditAgent, enhanced with auditing domain expertise, for fine-grained evidence chain localization in financial fraud cases. Leveraging an expert-annotated dataset constructed from enforcement documents and financial reports released by the China Securities Regulatory Commission, our approach integrates subject-level risk priors, a hybrid retrieval strategy, and specialized agent modules to efficiently identify and aggregate cross-report evidence. Extensive experiments demonstrate that our method substantially outperforms General-Purpose Agent paradigm in both recall and interpretability, establishing a new benchmark for automated, transparent financial forensics. Our results highlight the value of domain-specific reasoning and dataset construction for advancing robust financial fraud detection in practical, real-world regulatory applications.