Research on Construction Quality Risk Management of Urban Expressway Projects

Abstract

Urban expressway projects are critical components of modern transportation infrastructure, yet their construction quality is often threatened by multi-source, latent, and dynamic risks. Traditional expert-driven risk identification methods frequently suffer from subjective bias and low efficiency, failing to meet the rigorous management requirements of complex engineering environments. To address these challenges, this study proposes a robust risk assessment framework integrating Large Language Models (LLMs) and the Delphi method within a Bayesian Network (BN) structure. First, LLM technology is leveraged to perform semantic mining on extensive engineering texts, including construction specifications and project reports, to pre-identify potential risk factors. Second, the Delphi method is applied through multiple rounds of expert consultation to refine a comprehensive inventory comprising 32 risk factors across five dimensions: personnel, machinery, materials, methods, and environment. Finally, a BN-based evaluation model is developed, utilizing forward inference, backward diagnosis, and sensitivity analysis to quantify risk levels and pinpoint critical risk drivers. The framework was empirically validated using the T Expressway Project in Hangzhou as a case study. Results demonstrate that the model effectively transforms empirical management into precise, data-driven diagnosis, providing project managers with a quantitative tool for optimizing construction quality control and decision making in complex urban bridge projects.

1. Introduction

Urban expressways, as an important part of the modern road transportation system, serve as medium-to-high-speed transportation hubs that facilitate rapid short-to-medium-distance commuting within cities. The establishment of urban expressways helps vehicles maintain economical and environmentally friendly speeds, playing a crucial role in enhancing urban traffic safety and transportation efficiency. However, taking Hangzhou as an example, urban expressway projects are mostly located in central urban areas with extremely complex construction environments. These projects face issues such as confined sites, intricate underground pipelines, a high proportion of elevated bridges and tunnel engineering, and the frequent occurrence of highly difficult construction techniques. Coupled with the noise, dust, and prolonged interruption of existing traffic caused by construction, they often trigger significant public demands and pressure [1,2]. To expedite the schedule or shorten traffic closure times, some construction companies may ignore quality bottom lines, leading to superficial management processes and burying long-term structural safety hazards [3]. Therefore, conducting research on the construction quality risk management of Hangzhou’s expressway projects is of great significance for ensuring public safety and improving infrastructure investment benefits.

In the traditional construction quality management of urban expressways, risk identification relies heavily on expert-driven models such as the Delphi method or brainstorming. However, this resource-intensive management model has gradually exposed many incompatible limitations. On the one hand, it significantly consumes human and time resources, struggling to meet the practical needs of modern complex engineering for efficient management. On the other hand, the identification results are inevitably limited by the subjective cognitive biases of experts, resulting in a large amount of historical engineering data and tacit knowledge not being effectively utilized [4]. In addition, existing risk assessment methods mostly adopt linear weighting structures, such as the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation, which assume that indicators are independent of each other. These methods struggle to effectively characterize the complex correlations and dynamic evolutionary features among various factors within the construction quality risk system [5].

Currently, some scholars have innovatively incorporated Large Language Model (LLM) technology into the risk identification process. By leveraging its robust natural language processing and deep feature extraction capabilities, LLMs enable the efficient and automated extraction of latent risks from complex engineering texts. In the field of foundational research, traditional Natural Language Processing (NLP) techniques, such as word embedding and semantic matching, have been widely utilized to extract risk features from historical infrastructure projects and achieve automated risk classification through text similarity calculations [6,7]. However, conventional NLP often relies on keyword matching, making it difficult to process highly unstructured engineering corpora. Addressing these limitations, Ruhui Shi et al. designed a novel framework for the dynamic assessment and early warning of cross-border pipeline risks based on LLMs and knowledge graphs, significantly enhancing the level of cross-border pipeline risk management [8]. Chenyu Lu et al. proposed a novel framework leveraging Knowledge Graph reasoning and online Retrieval-Augmented Generation to assess multifaceted risks in the Sino-Russian pipeline network—ranging from geopolitical instability and international sanctions to complex cyber threats—ultimately improving the accuracy and timeliness of risk identification and response [9]. Similarly, Gongfan Chen et al. introduced a risk assessment framework where LLMs are used to refine and enhance the representation of risk categories; this model effectively proposes multimodal mitigation schemes by examining historical claims and supplemental agreements, thereby elevating overall project risk management [10]. Saika Wong et al. utilized LLMs combined with construction contract knowledge to build an enhanced language model for identifying contract risks, simulating the role of human experts during contract reviews and achieving high efficiency in contract auditing [11].

In terms of risk assessment modeling, Bayesian Networks (BNs) have been extensively applied in complex engineering domains due to their superior capability in handling uncertain information. Madihi M H et al. have integrated DEMATEL and ISM methods to systematically construct hierarchical structures of project risks based on expert knowledge, while introducing the Ranked Node Method (RNM) to alleviate the human burden of parameter assignment [12]. For larger and more intricate systems, Object-Oriented Bayesian Networks (OOBNs) have been proposed to achieve modular encapsulation of risk models, effectively addressing computational redundancy in large-scale networks by decomposing environment and operation elements into independent objects [13]. Furthermore, Sun X et al. have employed data-driven structural learning algorithms, such as Tree-Augmented Naive Bayes (TAN-BN), to automatically mine statistical dependencies among factors from historical accident cases [14].

Compared with recent advanced works, the distinct academic positioning of this study is highlighted by its unified, scalable technical chain. While existing studies in engineering text mining and LLM-assisted risk identification successfully automate knowledge extraction, they often stop at generating static, isolated risk inventories, failing to capture how these factors dynamically interact and propagate. Conversely, contemporary Bayesian construction risk modeling offers rigorous quantitative reasoning and predictive simulation, yet it relies heavily on tedious manual topology design or rigid data-driven algorithms that lack localized engineering physics and are highly sensitive to historical data gaps. Beyond these standalone and fragmented methodologies, this paper introduces a seamless technical pipeline that bridges the gap. By achieving a transition from LLM-driven semantic mining to Apriori-optimized structural topologies and fuzzy Noisy-OR parameterization, this research transforms unstructured qualitative data into a traceable, data–expert hybrid probabilistic network, offering a more balanced solution for infrastructure quality governance.

Driven by this integrated paradigm, this study presents a holistic approach for the secondary utilization of big data in the digital construction era. Specifically, an unredacted engineering text corpus is first processed via LLM semantic mining and refined through a Delphi-based expert consensus to establish a high-granularity risk inventory. To map these factors into a quantifiable system, the Apriori association rule algorithm is deployed to uncover statistical co-occurrences, which are subsequently translated into expert-justified directional dependencies to construct the BN topology. Backed by fuzzy Noisy-OR parameterization, the complete model enables dynamic probability updating, diagnostic reasoning, and critical sensitivity analysis. This technical chain not only bypasses the subjectivity of traditional risk estimation but also provides a scalable, precise decision support methodology for Hangzhou expressway projects.

2. Methodology

2.1. Construction of the Urban Expressway Project Construction Quality Risk Assessment Model

High-quality construction is the foundation for ensuring the long-term operation of urban infrastructure, and a scientific risk evaluation system is the key to achieving precise control of construction quality. Construction quality risks of urban expressway projects involve multiple dimensions such as personnel, machinery, materials, process methods, and the environment. The factors across these dimensions are mutually coupled and dynamically evolving. Traditional linear evaluation methods struggle to effectively depict their complex correlation features. Therefore, this study integrates LLM and the Delphi method to construct a construction quality risk assessment model based on BN structure, aiming to achieve a transition from “experience-based evaluation” to “precise diagnosis”.

The construction of this assessment model is divided into three phases. In Phase I, semantic mining of texts related to urban expressway projects is conducted based on LLM to preliminarily identify construction quality risk factors. In Phase II, the Delphi method is employed to conduct multiple rounds of expert modification on the preliminary risk inventory, forming a systematic and scientific list of risk factors. In Phase III, a risk assessment model is constructed based on BN theory. The network structure is optimized through association rule mining, and the quantitative assessment of risk levels and precise diagnosis of key risk factors are achieved using forward reasoning, backward reasoning, and sensitivity analysis. The establishment of this model provides a scientific method for the systematic identification and precise control of construction quality risks in urban expressway projects, serving as an important pathway to promote the high-quality development of urban infrastructure. The flow chart of evaluation model construction is shown in Figure 1.

2.2. Phase I: Preliminary Identification of Risk Factors Based on LLM

2.2.1. Acquisition of Textual Information

The acquisition of textual information is the foundation for utilizing LLM for preliminary risk identification. This study collects textual data related to the construction quality of Hangzhou expressways through multiple channels to ensure the authority, comprehensiveness, and relevance of the corpus. The acquired literature is mainly divided into the following three aspects:

• Industry norms and standards. Relevant norms for expressway construction issued by the state, Hangzhou city, and the industry, such as construction and quality acceptance norms, and engineering technical specifications, were collected. These norms define the acceptable standards and mandatory provisions for construction quality, possessing strong reference value. A total of 10 relevant documents were collected, with a total of about 221.5 thousand words.
• Materials and reports of Hangzhou expressway projects. Construction summaries and acceptance reports of expressway projects within Hangzhou in recent years were obtained through relevant companies. These texts record focal points related to construction quality, providing practical textual evidence. A total of five relevant documents were collected, with a total of about 28.5 thousand words.
• Relevant academic papers. Using databases such as CNKI and Wanfang Data, core journal and degree papers from the past decade were retrieved using keywords like “Hangzhou”, “urban expressway”, and “construction quality” to reflect the influencing factors of construction quality drawn from the research of other scholars. The above literature was organized to ultimately form the initial text corpus, providing an adequate data reserve for the LLM input. A total of 10 relevant documents were collected, with a total of about 144.7 thousand words.

After sorting out the above documents, the initial text corpus is finally formed, with a total of 25 documents, about 394.7 thousand words, providing sufficient data reserves for the input of LLM.

2.2.2. Mining of Risk Factors

Based on the acquisition of sufficient textual information, this study utilizes the semantic understanding and feature extraction capabilities of the LLM to conduct deep mining of the corpus [15]. This study employs Deepseek-R1 as the LLM tool [16]. The specific operational process is as follows:

• Construction of Prompt Engineering. To ensure the accuracy of the LLM in identifying risk factors and to avoid “hallucinations”, strict prompt engineering must be established. The LLM’s role is set as a “senior quality management expert for urban expressway projects,” and the goal is defined as “extracting influencing factors of construction quality risks for urban expressway projects,” identifying potential causes of quality defects from the input text paragraphs. Strict screening logic is set in the prompt, requiring the model to distinguish between “phenomena” (e.g., pavement cracks) and “risk factors” (e.g., loose control of the mix ratio), ensuring the mined content represents “risk factors” rather than “phenomena”. The detailed prompt words are shown in Appendix A.
• Text Preprocessing and Corpus Segmentation. Prior to LLM processing, raw texts underwent cleaning to remove HTML tags, headers, footers, and noise. Long documents were segmented into chunks of 500–800 words based on semantic integrity, with a 100-word overlap maintained between adjacent segments to prevent the loss of cross-contextual risk correlations.
• Semantic Scanning and Feature Identification. The preprocessed text is input into the LLM, leveraging its attention mechanism to capture negative quality descriptions related to “4M1E” (Man, Machine, Material, Method, Environment) within the text, and fix the temperature to 0 and Top_p to 0.9. Through semantic association, the model automatically identifies non-compliant behaviors, environmental mutations, and material failures during the construction process.
• Preliminary Clustering of Risk Items. Based on the extracted feature values, the LLM conducts preliminary merging and logical classification of scattered risk descriptions. The model automatically maps hundreds of identified risk data to the preset preliminary framework, forming an original mining list containing risk names, descriptions, and corresponding corpus sources. The risk factor categories are classified into five mutually independent and comprehensively covering dimensions, possessing broad universality: personnel, machinery, material, process method, and environment [17]. To mitigate the occurrence of hallucinations, a manual verification process was conducted by the researchers. Any risk factors not substantiated by the raw corpus were strictly excluded, ensuring that every identified factor is traceable and grounded in the original engineering documentation. The final compiled mining results consisted of 156 items. Some mining results are shown in

  Table 1. Preliminary excavation results of LLM.

  Risk Factor Category	Risk Factor	Risk Description/Consequence	Frequency
  Personnel Risk	Special operation personnel (welders, riggers, surveyors) without certificates or insufficient skills	The quality of key processes (welding, hoisting and measurement) is out of control, such as unqualified welds, hoisting accidents and linear deviation.	2
  Personnel Risk	The quality responsibility of the construction management personnel is not implemented, and the quality planning and disclosure are not effectively carried out	The quality management system is in vain, the quality control points are omitted, the construction personnel are not clear about the standards and risks, and the process is out of control.	3
  ……	……	……	……
  Environment Risk	Interference between underground pipelines and existing buildings	In the urban environment, the underground pipelines are disordered, close to existing buildings; the survey before construction is unclear or the protection is not effective, which may lead to the obstruction of excavation, the risk of foundation construction, and even affect the safety of surrounding facilities.	1

  .

2.2.3. Organization of Risk Factors

The mined risk factors are integrated, de-duplicated and semantically merged, and the risk descriptions of the same nature are merged. For example, “special operators (welders, riggers, surveyors) without certificates or insufficient skills” and “welders and nondestructive testing personnel without corresponding qualifications or working beyond the scope” can be combined into “insufficient qualifications and abilities of personnel in key positions”, and “survey and monitoring equipment (total station, level, etc.) are not verified or inaccurate” and “paving, rolling and other key construction equipment (such as pavers, rollers) are not configured or faulty” can be combined into “poor status of key construction machinery and equipment”, etc. All risk factors are structured and systematized, and finally a preliminary identification list of construction quality risk factors for the H city expressway project is formed. The list includes five first-class dimensions, and a total of 32 s-class risk factors, covering the key quality risk points from the project commencement preparation to the completion acceptance, as shown in

Table 2. Preliminary risk inventory of construction quality for Hangzhou expressway projects.

Risk Factor Category	ID	Risk Factor
Personnel Risk	R1	Insufficient qualifications and capabilities of key personnel
Personnel Risk	R2	Lax implementation of standards by construction and quality inspection personnel
Personnel Risk	R3	Inadequate quality and safety technical disclosures
Personnel Risk	R4	Dereliction of duty in on-site quality supervision and control
Personnel Risk	R5	Insufficient capability or stability of the core project management team
Machinery Risk	R6	Poor condition of key construction machinery and equipment
Machinery Risk	R7	Safety hazards in lifting equipment and rigging
Machinery Risk	R8	Inaccurate measurement and test monitoring equipment
Machinery Risk	R9	Lack of maintenance and calibration management for construction equipment
Material Risk	R10	Out of control inspection and acceptance of raw materials entering the site
Material Risk	R11	Loss of control over mixture production and construction temperatures
Material Risk	R12	Poor on-site storage of finished products, semi-finished products, and materials
Material Risk	R13	Unqualified mixture proportions and material gradations
Material Risk	R14	Improper management and use of connection and protective materials
Method Risk	R15	Improper processing techniques for foundation and special parts
Method Risk	R16	Defects in subgrade and pavement compaction process control
Method Risk	R17	Defects in pavement paving and joint construction processes
Method Risk	R18	Construction defects throughout the entire concrete engineering process
Method Risk	R19	Missing welding process control and quality inspection
Method Risk	R20	Defects in prestressing system construction technology
Method Risk	R21	Defects in pile foundation piling construction technology
Method Risk	R22	Defects in steel structure installation and coating processes
Method Risk	R23	Missing acceptance of concealed works and key procedures
Method Risk	R24	Insufficient preparation and argumentation of special construction plans
Method Risk	R25	Out of control third-party testing and experimental management
Method Risk	R26	Poor control of structural alignment and geometric dimensions
Environment Risk	R27	Forced construction under adverse weather conditions
Environment Risk	R28	Inadequate response to complex geological condition risks
Environment Risk	R29	Confined construction sites and cross-operation interference
Environment Risk	R30	Construction risks adjacent to existing facilities
Environment Risk	R31	Inadequate control of special operational environments
Environment Risk	R32	Lack of management for construction environmental protection and civilized construction

.

2.3. Phase II: Refinement of the Risk Factor Inventory Based on the Delphi Method

Based on the preliminary risk list identified through the LLM, to ensure the targeting and scientific validity of the research, the Delphi method must be utilized. By drawing on the knowledge and experience of relevant industry experts, the risk factor list is further refined and adjusted to perfect the construction quality risk factor inventory for Hangzhou expressways.

2.3.1. Expert Profile and Reliability Analysis

(1)

Basic Information of Experts

A total of 26 experts were recruited from various stakeholders within the Hangzhou construction industry for this study. These individuals demonstrated a willingness to participate and possessed specialized knowledge of urban expressway projects in Hangzhou. The contents of the survey questionnaires are shown in Appendix B. The statistical information of the experts is shown in

Table 3. Statistical profile of Delphi method experts.

Information	Category	Number	Percentage (%)
Type of employer	Owner	8	31
	Supervisor	6	23
	Contractor	6	23
	Designer	6	23
Professional title	Senior Engineer	15	58
	Engineer	9	35
	Other	2	8
Years of experience	0–10 years	3	12
	10–15 years	12	46
	15–20 years	8	31
	20 years and above	3	12
Number of Hangzhou expressway projects participated in	1–2	16	62
	3–4	7	27
	5 and above	3	12
Familiarity with survey content	Very familiar	10	38
	Relatively familiar	14	54
	Generally familiar	2	8

. The data in Table 3 indicate that although these experts are from Hangzhou, they cover the owner, design, construction and supervision parties, ensuring the independence of judgments to provide effective professional opinions and experiential support for this study.

(2)

Expert Reliability Analysis

Considering the existence of potential cognitive bias, such as anchoring or overconfidence, to ensure the scientific validity of the survey results, this study verifies expert reliability through indicators such as the expert enthusiasm coefficient, expert subjective reliability, and expert opinion coordination degree.

• Expert Enthusiasm Coefficient

The expert enthusiasm coefficient refers to the ratio of the actual number of participating experts to the total number of invited experts. This coefficient quantifies the degree of engagement among the surveyed experts in this study [18], calculated as follows [19]:

C = M_j/M,

(1)

where C is the expert enthusiasm coefficient; M_j is the number of experts participating in the survey; and M is the total number of selected experts. A total of two rounds of the Delphi method were conducted. In both rounds, 26 questionnaires were distributed and 26 were recovered, resulting in an enthusiasm coefficient of 100%. Moreover, the list of participating experts remained consistent across both rounds, indicating sufficient enthusiasm for this study.

2.

Expert Subjective Reliability

Expert subjective reliability reflects the depth of experts’ familiarity and understanding of the investigated issues, aiding in evaluating the reliability and validity of the survey results. In the two rounds of questionnaire surveys, the experts’ familiarity was rated on a 5-point scale, with specific grades and assignment standards shown in

Table 4. Assignment table for expert familiarity level.

Very Familiar	Relatively Familiar	Generally Familiar	Relatively Unfamiliar	Very Unfamiliar
5	4	3	2	1

.

Statistical analysis of expert familiarity across the two rounds, including mean and standard deviation, is presented in

Table 5. Calculation results of expert subjective reliability.

Expert Subjective Reliability	Mean	Standard Deviation
Round 1	4.308	0.618
Round 2	4.577	0.504

. The standard deviation measures data dispersion; a smaller standard deviation indicates more concentrated data. The calculations reveal that the mean familiarity coefficients for the two rounds were 4.308 and 4.577, with standard deviations of 0.618 and 0.504, respectively, indicating highly concentrated subjective reliability. The increased mean and decreased standard deviation in the second round suggest enhanced stability and reliability, reflecting improved consistency among experts.

3.

Expert Opinion Coordination Degree

The coordination degree reflects the consistency of expert evaluation results, typically measured by Kendall’s coefficient of concordance (W^a). A larger W^a value indicates less divergence among experts. If the significance level (P) of the coefficient is less than 0.05, the coordination of the evaluation is considered acceptable. The Kendall’s concordance test results for the two rounds are shown in

Table 6. Results of Kendall’s concordance test.

Expert Opinion Coordination Degree	Concordance Coefficient (Wa)	Chi-Square (χ2)	Significance (P)
Round 1	0.258	208.240	0.000
Round 2	0.291	234.802	0.000

.

As shown in Table 6, the concordance coefficient W^a for the first round was 0.258 with a significant p-value (0.000 < 0.05), indicating acceptable coordination. In the second round, W^a increased to 0.291, also significant (p = 0.000 < 0.05), further demonstrating enhanced consistency and data reliability. Considering that the panel comprised professionals from diverse disciplines, the inherent divergence in risk sensitivity among experts from different backgrounds is expected. Therefore, the observed W^a value was deemed sufficient for the requirements of this study. Because the convergence of both rounds met research requirements, this study concluded with two rounds of Delphi surveys.

2.3.2. First Round of Survey and Results

Using the preliminary LLM risk list as a reference, the first round combined open and structured questionnaires, asking experts to decide whether to “retain, modify, or eliminate” the identified factors and provide additional comments. To ensure the integrity of the findings, a strict anonymity protocol was maintained throughout the survey process. Experts were required to submit their responses independently, with their identities and judgments kept confidential from one another to mitigate potential interference or peer influence.

The screening criteria used “expert opinion concentration” (arithmetic mean > 4 points) and “expert opinion coordination” (coefficient of variation ≤ 25%) [20]. Factors meeting these criteria were retained. The statistical results are shown in

Table 7. Statistical results of the first-round expert survey.

Risk Factor	Mean	Coefficient of Variation (%)
Insufficient qualifications and capabilities of key personnel	4.615	18.46
Lax implementation of standards by construction and quality inspection personnel	4.308	25.23
Inadequate quality and safety technical disclosures	4.500	20.12
Dereliction of duty in on-site quality supervision and control	4.423	18.28
Insufficient capability or stability of the core project management team	4.385	20.48
Poor condition of key construction machinery and equipment	4.385	22.42
Safety hazards in lifting equipment and rigging	4.345	21.53
Inaccurate measurement and test monitoring equipment	4.269	22.52
Lack of maintenance and calibration management for construction equipment	4.346	21.53
Out of control inspection and acceptance of raw materials entering the site	4.538	18.93
Loss of control over mixture production and construction temperatures	4.500	18.05
Poor on-site storage of finished products, semi-finished products, and materials	4.308	20.52
Unqualified mixture proportions and material gradations	4.385	25.87
Improper management and use of connection and protective materials	4.385	21.47
Improper processing techniques for foundation and special parts	4.615	19.45
Defects in subgrade and pavement compaction process control	4.500	20.12
Defects in pavement paving and joint construction processes	4.385	19.43
Construction defects throughout the entire concrete engineering process	4.385	25.05
Missing welding process control and quality inspection	4.462	21.25
Defects in prestressing system construction technology	4.385	22.42
Defects in pile foundation piling construction technology	4.500	21.08
Defects in steel structure installation and coating processes	4.346	21.53
Missing acceptance of concealed works and key procedures	4.385	22.42
Insufficient preparation and argumentation of special construction plans	4.462	21.25
Out of control third-party testing and experimental management	4.308	24.36
Poor control of structural alignment and geometric dimensions	4.346	20.52
Forced construction under adverse weather conditions	4.615	17.42
Inadequate response to complex geological condition risks	4.423	24.07
Confined construction sites and cross-operation interference	4.231	24.39
Construction risks adjacent to existing facilities	4.385	23.33
Inadequate control of special operational environments	4.423	20.40
Lack of management for construction environmental protection and civilized construction	4.192	21.36

.

As shown in Table 7, all mean scores exceeded 4 points. However, the coefficients of variation for three factors (“Lax implementation of standards …”, “Unqualified mixture proportions …”, and “Construction defects throughout the entire concrete engineering process”) exceeded 25%, and thus were eliminated.

Experts also provided modification suggestions, summarized as follows:

(1)

R1 was modified to “Insufficient qualifications, experience, and practical capabilities of key personnel”;

(2)

R11 was modified to “Loss of control over mixture production, transportation, and construction temperatures”;

(3)

R12 was modified to “Poor on-site storage of finished products, semi-finished products, and hazardous materials”;

(4)

Added “Logical errors or improper connection of construction procedures” under “Method Risk”;

(5)

Added “Leakage hazards at key waterproof parts such as welds and construction joints” under “Method Risk”;

(6)

Added “Defects in traffic organization and diversion plans during construction” under “Environment Risk”. A new risk inventory was formed based on these adjustments.

2.3.3. Second Round of Survey and Results

The updated list was sent to experts for the second round. The statistical results are shown in

Table 8. Statistical results of the second-round expert survey.

Risk Factor	Mean	Coefficient of Variation (%)
Insufficient qualifications, experience, and practical capabilities of key personnel	4.731	9.56
Inadequate quality and safety technical disclosures	4.654	10.43
Dereliction of duty in on-site quality supervision and control	4.346	15.86
Insufficient capability or stability of the core project management team	4.308	18.30
Poor condition of key construction machinery and equipment	4.769	10.79
Safety hazards in lifting equipment and rigging	4.385	15.90
Inaccurate measurement and test monitoring equipment	4.615	13.81
Lack of maintenance and calibration management for construction equipment	4.385	19.43
Out of control inspection and acceptance of raw materials entering the site	4.500	15.71
Loss of control over mixture production, transportation, and construction temperatures	4.615	16.30
Poor on-site storage of finished products, semi-finished products, and hazardous materials	4.154	21.20
Improper management and use of connection and protective materials	4.231	20.40
Improper processing techniques for foundation and special parts	4.731	11.28
Defects in subgrade and pavement compaction process control	4.308	20.52
Defects in pavement paving and joint construction processes	4.577	15.35
Missing welding process control and quality inspection	4.615	12.37
Defects in prestressing system construction technology	4.808	10.22
Defects in pile foundation piling construction technology	4.654	13.51
Defects in steel structure installation and coating processes	4.731	9.56
Missing acceptance of concealed works and key procedures	4.538	16.76
Insufficient preparation and argumentation of special construction plans	4.462	18.19
Out of control third-party testing and experimental management	4.615	16.30
Poor control of structural alignment and geometric dimensions	4.654	16.01
Logical errors or improper connection of construction procedures	4.769	10.79
Leakage hazards at key waterproof parts such as welds and construction joints	4.423	18.28
Forced construction under adverse weather conditions	4.731	11.28
Inadequate response to complex geological condition risks	4.615	12.37
Confined construction sites and cross-operation interference	4.500	14.40
Construction risks adjacent to existing facilities	4.808	10.22
Inadequate control of special operational environments	4.654	13.51
Lack of management for construction environmental protection and civilized construction	4.654	12.07
Defects in traffic organization and diversion plans during construction	4.769	9.01

.

As indicated in Table 8, all factors achieved a mean > 4 and a coefficient of variation < 25%. Furthermore, no additional modifications or new risk factors were proposed by the experts, suggesting that the current inventory has reached information saturation and possesses sufficient rationality, precluding the need for further revision.

2.3.4. Construction Quality Risk Factor Inventory for Hangzhou Expressway Projects

The finalized inventory contains 32 risk factors, detailed in

Table 9. Construction quality risk factor inventory for Hangzhou expressway projects.

Category ID	Risk Factor Category	Risk Factor ID	Risk Factor
A	Personnel Risk	A1	Insufficient qualifications, experience, and practical capabilities of key personnel
		A2	Inadequate quality and safety technical disclosures
		A3	Dereliction of duty in on-site quality supervision and control
		A4	Insufficient capability or stability of the core project management team
B	Machinery Risk	B1	Poor condition of key construction machinery and equipment
		B2	Safety hazards in lifting equipment and rigging
		B3	Inaccurate measurement and test monitoring equipment
		B4	Lack of maintenance and calibration management for construction equipment
C	Material Risk	C1	Out of control inspection and acceptance of raw materials entering the site
		C2	Loss of control over mixture production, transportation, and construction temperatures
		C3	Poor on-site storage of finished products, semi-finished products, and hazardous materials
		C4	Improper management and use of connection and protective materials
D	Method Risk	D1	Improper processing techniques for foundation and special parts
		D2	Defects in subgrade and pavement compaction process control
		D3	Defects in pavement paving and joint construction processes
		D4	Missing welding process control and quality inspection
		D5	Defects in prestressing system construction technology
		D6	Defects in pile foundation piling construction technology
		D7	Defects in steel structure installation and coating processes
		D8	Missing acceptance of concealed works and key procedures
		D9	Insufficient preparation and argumentation of special construction plans
		D10	Out of control third-party testing and experimental management
		D11	Poor control of structural alignment and geometric dimensions
		D12	Logical errors or improper connection of construction procedures
		D13	Leakage hazards at key waterproof parts such as welds and construction joints
E	Environment Risk	E1	Forced construction under adverse weather conditions
		E2	Inadequate response to complex geological condition risks
		E3	Confined construction sites and cross-operation interference
		E4	Construction risks adjacent to existing facilities
		E5	Inadequate control of special operational environments
		E6	Lack of management for construction environmental protection and civilized construction
		E7	Defects in traffic organization and diversion plans during construction

.

2.4. Phase III: Construction of the Risk Assessment Model Based on BN Structure

2.4.1. BN Structure Construction

(1)

Initial Structure Construction

Based on the identified risk inventory, an initial BN structure for construction quality risks of the Hangzhou expressway project was established [21], as shown in Figure 2.

(2)

Structure Optimization

Based on Association Rule Theory, in reality, the occurrence of quality risks is often the result of the combined effect of multiple factors. Therefore, it is necessary to mine the influencing relationships, directions, and credibility among factors using Association Rule Theory, laying the foundation for network structure optimization. The Support–Confidence framework was applied [22]. The support of rule A → B is calculated as:

Support(A ⇒ B) = Count(A∪B)/N

(2)

The confidence is calculated as:

Confidence(A ⇒ B) = Support(A ⇒ B)/Support(A)

(3)

Given potential bias from extremely imbalanced data, Lift serves as a supplementary criterion [23]:

Lift(A ⇒ B) = Confidence(A ⇒ B)/Support(B) = P(B|A)/P(B)

(4)

A lift greater than 1 indicates a valid strong association rule. The Python 3.10 library mlxtend was used to execute the Apriori algorithm on the binarized matrix (values ≥ 4 converted to 1; <4 to 0). In order to clarify the confidence, scientific support and confidence thresholds, the sensitivity analysis was carried out to explore the distribution of association rules under different support and confidence thresholds, and analyze the change of the number of association rules with the change of the support and confidence thresholds. The analysis results are shown in Figure 3 and Figure 4, in which the slope of the line represents the change speed of the number of association rules [24].

The test results show that when the support is less than 0.5, a large number of redundant false associations will be generated, resulting in the complexity of the network topology and the difficulty of accurate reasoning. When the confidence is higher than 0.8, some potential directional risk propagation paths with important engineering significance will be incorrectly filtered. Therefore, the thresholds were set as Support = 0.5, Confidence = 0.8, and Lift = 1, which can balance the complexity and completeness of rules. The association rules based on the Apriori algorithm are extracted and analyzed by using Python programming language and its integrated mining library mlextend, and the direction is determined by logical reasoning combined with the actual construction engineering knowledge. Finally, a total of eight risk factor association rules are mined, as shown in

Table 10. Association rules table.

No.	Antecedent Risk Factor	Consequent Risk Factor	Support	Confidence	Lift
1	A1 Insufficient qualifications, experience, and practical capabilities of key personnel	D4 Missing welding process control and quality inspection	0.65	0.92	1.28
2	A1 Insufficient qualifications, experience, and practical capabilities of key personnel	D5 Defects in prestressing system construction technology	0.58	0.88	1.15
3	A2 Inadequate quality and safety technical disclosures	D12 Logical errors or improper connection of construction procedures	0.73	0.91	1.35
4	A3 Dereliction of duty in on-site quality supervision and control	D8 Missing acceptance of concealed works and key procedures	0.69	0.89	1.24
5	A4 Insufficient capability or stability of the core project management team	A3 Dereliction of duty in on-site quality supervision and control	0.55	0.81	1.10
6	B1 Poor condition of key construction machinery and equipment	D4 Missing welding process control and quality inspection	0.54	0.85	1.12
7	C1 Out of control inspection and acceptance of raw materials entering the site	D7 Defects in steel structure installation and coating processes	0.51	0.82	1.08
8	D9 Insufficient preparation and argumentation of special construction plans	D1 Improper processing techniques for foundation and special parts	0.62	0.86	1.19

.

It should be noted that the implementation of the Apriori algorithm is intended to uncover statistical co-occurrences and strong associations among risk factors within the project data. To maintain strict theoretical and epistemological rigor, the added arrows in the BN topology are interpreted not as statistically proven causal relationships, but as association-supported and expert-justified directional dependencies. While data mining provides the probabilistic coupling strength, the expert consensus sequentially validates the actual engineering mechanisms and determines the final causal directions. Therefore, to further demonstrate this paradigm, the engineering connotations and underlying mechanisms of the mined rules in Table 10 are elaborated as follows:

(1) “A1 Insufficient qualifications, experience, and practical capabilities of key personnel → D4 Missing welding process control and quality inspection” (i.e., A1 is the Antecedent Risk Factor, and D4 is the Consequent Risk Factor; the same applies below): Personnel capability constitutes a root-cause resource input that directly determines the execution quality of the construction process. According to the PDCA cycle, a lack of qualifications among welders or quality inspectors will fundamentally lead to an inability to identify technical essentials or result in missed inspections during construction, thereby triggering process dyscontrol. Since process deviations do not inversely cause changes in personnel qualifications, this pathway possesses logical irreversibility.

(2) “A1 Insufficient qualifications, experience, and practical capabilities of key personnel → D5 Defects in prestressing system construction technology”: Prestressing tensioning and grouting are special procedures with high professional technical requirements. A lack of practical experience among operators will directly lead to inaccurate tension control or uncompacted grouting, thereby inducing defects in the prestressing process. Conversely, localized process defects will not inversely force or cause a reduction in the basic qualifications of personnel, making the directional risk propagation path extend unidirectionally downward.

(3) “A2 Inadequate quality and safety technical disclosures → D12 Logical errors or improper connection of construction procedures”: Technical disclosure is the prerequisite information input phase for on-site execution. If the transmission of disclosure information attenuates or becomes a mere formality, frontline operational crews will develop comprehension deviations due to a lack of clear construction guidance, subsequently causing procedural disorder or disjointed cross-convergence in actual operations. The absence of information flow serves as a unidirectional trigger for physical execution errors in procedures.

(4) “A3 Dereliction of duty in on-site quality supervision and control → D8 Missing acceptance of concealed works and key procedures”: On-site supervision is a dynamic assurance mechanism for quality control. If a vacuum exists in the daily performance of management personnel, it will inevitably manifest in specific behaviors as the failure to conduct bystander supervision and acceptance of concealed works or key nodes according to specifications. “Management absence” is the cause, and “missing acceptance” is the effect; a significant managerial progression and irreversible relationship exist between the two in engineering practice.

(5) “A4 Insufficient capability or stability of the core project management team → A3 Dereliction of duty in on-site quality supervision and control”: The stability and comprehensive capability of the core team serve as the cornerstone for the operation of the project management system. Frequent team turnover or weak capabilities will lead to the collapse of the accountability system and management discontinuity, directly manifesting as the inability to effectively implement on-site supervision and control. Macro-level organizational state defects inevitably and unidirectionally trigger the failure of micro-level management behaviors.

(6) “B1 Poor condition of key construction machinery and equipment → D4 Missing welding process control and quality inspection”: Machinery and equipment are the physical carriers of engineering implementation. If key equipment such as welding machines and flaw detectors experience aging, operate with faults, or remain uncalibrated, this will lead to unstable welding parameter outputs or distorted inspection data, directly undermining the effectiveness of process control. As an objective hardware condition, the equipment state constitutes a prerequisite physical trigger for process deviations.

(7) “C1 Out of control inspection and acceptance of raw materials entering the site → D7 Defects in steel structure installation and coating processes”: Qualified raw materials are the foundation for ensuring subsequent processes. If a loss of control in the incoming inspection phase allows substandard steel components or anti-corrosion coatings to flow into the construction phase, this will directly lead to insufficient subsequent assembly precision or the failure of coating adhesion. The engineering material flow possesses a definitive unidirectional time sequence; the failure of front-end material control acts as a rigid origin causing back-end process defects.

(8) “D9 Insufficient preparation and argumentation of special construction plans → D1 Improper processing techniques for foundation and special parts”: Special schemes are the overarching documents guiding complex construction. If the scheme lacks targeted demonstration for special geological conditions during the preparation stage, the construction crew will have no basis to follow or will adopt incorrect construction processes when treating strata such as soft foundations. Based on the logic that “planning guides execution,” the inherent deficiency of the scheme unidirectionally determines the improper operation of on-site processes.

The data demonstrates high statistical significance [25]. Recognizing these links empowers managers to front-load resources into training (prevention at the source) rather than expending vast resources on post-defect detection and rework [26]. These directed edges were added to optimize the network [27,28,29], resulting in Figure 5.

2.4.2. Determination of BN Parameters

(1)

Quantifying Expert Evaluations

Based on Fuzzy Theory, Triangular Fuzzy Numbers (TFNs) were utilized to handle the subjectivity and uncertainty of expert evaluations [30,31,32]. The mapping of linguistic variables to TFNs is shown in

Table 11. Mapping table of fuzzy linguistic grades to triangular fuzzy numbers.

Fuzzy Linguistic Grade	Triangular Fuzzy Number
Extremely low	(0, 0.005, 0.01)
Low	(0.01, 0.025, 0.05)
Relatively low	(0.05, 0.1, 0.15)
Medium	(0.15, 0.2, 0.25)
Relatively high	(0.25, 0.35, 0.45)
High	(0.45, 0.6, 0.75)
Extremely high	(0.75, 0.875, 0.99)

.

The mean area method was selected for defuzzification:

F_i = (a_i + 2m_i + b_i)/4

(5)

(2)

Acquiring Subjective Expert Weights

In order to reduce the potential cognitive bias of experts, experts were given subjective and objective weights in the calculation process to improve the accuracy of BN parameters as much as possible. Subjective weights were assigned based on experts’ professional titles, years of experience, and familiarity (see

Table 12. Scoring criteria for subjective expert weight indicators.

Information	Category	Score
Professional title	Professor-level Senior Engineer	10
	Senior Engineer	8
	Engineer	6
	Assistant Engineer	4
	Technician	2
Years of experience	20 years and above	10
	15–20 years	8
	10–15 years	6
	5–10 years	4
	0–5 years	2
Familiarity with survey content	Very familiar	10
	Relatively familiar	8
	Generally familiar	6
	Relatively unfamiliar	4
	Very unfamiliar	2

) [33,34].

The subjective weight formula is:

$${\mathrm{\omega }}_{\mathrm{i}}^{\mathrm{s}}={\displaystyle \frac{{\mathrm{M}}_{\mathrm{i}}}{{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}}{\mathrm{M}}_{\mathrm{i}}}}$$

(6)

This was applied to the 26 experts.

(3)

Acquiring Objective Expert Weights

The DUOWA method was introduced to determine objective weights by calculating the similarity between individual evaluations and the mean [35]. The formulas are applied as follows:

$$a_v={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n}{a}_i, m_v={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n}{m}_i, b_v={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n}{b}_i$$

(7)

d(F_i,F_v)=(|a_i−a_v|+|m_i−m_v|+|b_i−b_v|)/3

(8)

$$\mathrm{s}({\mathrm{F}}_{\mathrm{i}},{\mathrm{F}}_{\mathrm{v}})=1-{\displaystyle \frac{\mathrm{d}({\mathrm{F}}_{\mathrm{i}},{\mathrm{F}}_{\mathrm{v}})}{{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}}\mathrm{d}({\mathrm{F}}_{\mathrm{i}},{\mathrm{F}}_{\mathrm{v}})}}$$

(9)

$${\mathrm{\omega }}_i^o={\displaystyle \frac{\mathrm{s}({\mathrm{F}}_{\mathrm{i}},{\mathrm{F}}_{\mathrm{v}})}{{\displaystyle {\sum }_{i=1}^n}\mathrm{s}({\mathrm{F}}_{\mathrm{i}},{\mathrm{F}}_{\mathrm{v}})}}$$

(10)

Taking the correlation of “Personnel Risk A → Construction Quality of Hangzhou Expressway Projects R” as an example, the objective weights of the 26 experts calculated according to the above steps are

$$\begin{array}{c}{\mathrm{\omega }}^{\mathrm{o}}=[0.039783438, 0.039783438, 0.037464093, 0.037464093, 0.035154847, 0.037464093, \hfill \\ 0.037464093, 0.039783438, 0.039783438, 0.039783438, 0.039783438, 0.039783438, \hfill \\ 0.037970153, 0.037464093, 0.036657316, 0.039783438, 0.039783438, 0.037970153, \hfill \\ 0.039783438, 0.037970153, 0.037464093, 0.039783438, 0.036657316, 0.039783438, \hfill \\ 0.037464093, 0.037970153].\hfill \end{array}$$

(11)

(1)

Acquiring Comprehensive Expert Weights

Euclidean distance was used to combine subjective and objective weights [36]:

$$\mathrm{d}({\mathrm{\omega }}^{\mathrm{s}},{\mathrm{\omega }}^{\mathrm{o}})=\sqrt{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{({\mathrm{\omega }}_{\mathrm{i}}^{\mathrm{s}}-{\mathrm{\omega }}_i^o)}^2}$$

(12)

$$\mathrm{d}({\mathrm{\omega }}_{\mathrm{i}}^{\mathrm{s}},{\mathrm{\omega }}_i^o)=\mathrm{a}-\mathrm{b}, \mathrm{a}+\mathrm{b}=1$$

(13)

$${{\mathrm{\omega }}_{\mathrm{i}}=\mathrm{a}\mathrm{\omega }}_{\mathrm{i}}^{\mathrm{s}}+\mathrm{b}{\mathrm{\omega }}_i^o$$

(14)

Taking the correlation of “Personnel Risk A → Construction Quality of Hangzhou Expressway Projects R” as an example, the final expert weights obtained according to the above steps are

$$\begin{array}{c}\mathrm{\omega }=\mathrm{a}{\mathrm{\omega }}^{\mathrm{s}}+\mathrm{b}{\mathrm{\omega }}^{\mathrm{o}}=[0.040312067, 0.035089651, 0.037438950, 0.037438950, 0.036311569, \hfill \\ 0.039179756, 0.040920561, 0.038571262, 0.036830456, 0.038571262, 0.036830456, \hfill \\ 0.042052873, 0.041167621, 0.035698145, 0.037045079, 0.042052873, 0.038571262, \hfill \\ 0.035945204, 0.040312067, 0.037686010, 0.042661367, 0.035089651, 0.038785884, \hfill \\ 0.038571262, 0.037438950, 0.039426815].\hfill \end{array}$$

(15)

Similarly, the expert weights after the subjective and objective combined weighting for all risk factor correlations can be calculated.

(2)

Determination of Prior Probabilities

Prior probabilities of root nodes were obtained through expert surveys, processed via TFNs and combined weighting.

(3)

Determination of Conditional Probability Distribution

In order to reduce the workload of experts and ensure the scientificity of data, the Leaky Noisy-OR model is implemented to circumvent the limitations of traditional complete conditional probability induction methods [37]. The model assumes that the effects of parent nodes on child nodes are independent. At the same time, considering that there are missing risk factors that cannot be fully listed in the actual project, the default node X_m is introduced to represent all missing factor sets [38]. The calculation method of the complete conditional probability distribution of non-root node Y considering the omission probability is

$$\mathrm{P}(\mathrm{Y}\leftarrow \mathrm{X})=1-(1-\mathrm{P}(\mathrm{Y}\leftarrow {\mathrm{X}}_{\mathrm{m}}))\prod _{\mathrm{i}:{\mathrm{X}}_{\mathrm{i}}\in {\mathrm{X}}_{\mathrm{T}}}(1-\mathrm{P}(\mathrm{Y}\leftarrow {\mathrm{X}}_{\mathrm{i}}))$$

(16)

The calculated probabilities are presented in

Table 13. Conditional probability distribution table for non-root nodes of the BN.

A	B	C	D	E	P(R = N)	P(R = Y)
N	N	N	N	N	0.99	0.01
N	N	N	N	Y	0.547006220	0.452993780
N	N	N	Y	N	0.495120081	0.504879919
N	N	N	Y	Y	0.273569458	0.726430542
N	N	Y	N	N	0.476566651	0.523433349
N	N	Y	N	Y	0.263318103	0.736681897
N	N	Y	Y	N	0.238341130	0.761658870
N	N	Y	Y	Y	0.131690991	0.868309009
N	Y	N	N	N	0.466294191	0.533705809
N	Y	N	N	Y	0.257642245	0.742357755
N	Y	N	Y	N	0.233203654	0.766796346
N	Y	N	Y	Y	0.128852373	0.871147627
N	Y	Y	N	N	0.224464910	0.775535090
N	Y	Y	N	Y	0.124023941	0.875976059
N	Y	Y	Y	N	0.112259681	0.887740319
N	Y	Y	Y	Y	0.062027014	0.937972986
Y	N	N	N	N	0.411731621	0.588268379
Y	N	N	N	Y	0.227494704	0.772505296
Y	N	N	Y	N	0.205915751	0.794084249
Y	N	N	Y	Y	0.113774946	0.886225054
Y	N	Y	N	N	0.198199555	0.801800445
Y	N	Y	N	Y	0.109511505	0.890488495
Y	N	Y	Y	N	0.099123818	0.900876182
Y	N	Y	Y	Y	0.054769035	0.945230965
Y	Y	N	N	N	0.193927336	0.806072664
Y	Y	N	N	Y	0.107150969	0.892849031
Y	Y	N	Y	N	0.096987190	0.903012810
Y	Y	N	Y	Y	0.053588481	0.946411519
Y	Y	Y	N	N	0.093352830	0.906647170
Y	Y	Y	N	Y	0.051580382	0.948419618
Y	Y	Y	Y	N	0.046687738	0.953312262
Y	Y	Y	Y	Y	0.025796447	0.974203553

.

3. Empirical Analysis

The T Expressway project (hereinafter referred to as Project T) serves as the case study, possessing complex structural and environmental characteristics. Project T started in July 2023 and is scheduled to be completed in April 2026, with a total investment of about 580 million yuan. The bridge works of Project T mainly include the main line viaduct, ramp bridge and river crossing bridge. The bridge structure types include prestressed concrete continuous box girder, steel–concrete composite girder and prestressed concrete hollow slab girder.

3.1. Determination of Prior Probabilities for Root Nodes

Another 19 experts who are familiar with Project T and have rich experience in construction and quality management were invited to evaluate the occurrence probability of 25 risk root nodes of Project T.

Table 14. Comprehensive expert weights for risk factor A1.

No.	Subjective Weight	Objective Weight	Comprehensive Weight
Expert 1	0.056074766	0.052945141	0.054549336
Expert 2	0.042056075	0.055209976	0.048467500
Expert 3	0.051401869	0.052576713	0.051974507
Expert 4	0.051401869	0.052185437	0.051783793
Expert 5	0.051401869	0.055209976	0.053258002
Expert 6	0.056074766	0.052945141	0.054549336
Expert 7	0.060747664	0.052945141	0.056944587
Expert 8	0.051401869	0.052185437	0.051783793
Expert 9	0.046728972	0.052576713	0.049579256
Expert 10	0.051401869	0.052945141	0.052154085
Expert 11	0.046728972	0.047692896	0.047198804
Expert 12	0.060747664	0.055209976	0.058048504
Expert 13	0.060747664	0.052576713	0.056765010
Expert 14	0.046728972	0.052185437	0.049388541
Expert 15	0.051401869	0.052576713	0.051974507
Expert 16	0.060747664	0.052185437	0.056574295
Expert 17	0.051401869	0.055209976	0.053258002
Expert 18	0.046728972	0.052945141	0.049758834
Expert 19	0.056074766	0.047692896	0.051989307

shows the comprehensive expert weights calculated for risk factor A1 based on the evaluation of 19 experts.

Prior probabilities of all root nodes were calculated and are listed in

Table 15. Prior probabilities of root nodes for Project T.

Root Node ID	Risk Factor	Y	N
A1	Insufficient qualifications, experience, and practical capabilities of key personnel	0.106494143	0.893505857
A2	Inadequate quality and safety technical disclosures	0.112128472	0.887871528
A4	Insufficient capability or stability of the core project management team	0.120255803	0.879744197
B1	Poor condition of key construction machinery and equipment	0.099581983	0.900418017
B2	Safety hazards in lifting equipment and rigging	0.110134512	0.889865488
B3	Inaccurate measurement and test monitoring equipment	0.097059264	0.902940736
B4	Lack of maintenance and calibration management for construction equipment	0.106122142	0.893877858
C1	Out of control inspection and acceptance of raw materials entering the site	0.112407822	0.887592178
C2	Loss of control over mixture production, transportation, and construction temperatures	0.106027254	0.893972746
C3	Poor on-site storage of finished products, semi-finished products, and hazardous materials	0.112498823	0.887501177
C4	Improper management and use of connection and protective materials	0.114793280	0.885206720
D2	Defects in subgrade and pavement compaction process control	0.112606426	0.887393574
D3	Defects in pavement paving and joint construction processes	0.120186871	0.879813129
D6	Defects in pile foundation piling construction technology	0.107505933	0.892494067
D9	Insufficient preparation and argumentation of special construction plans	0.112407822	0.887592178
D10	Out of control third-party testing and experimental management	0.100604735	0.899395265
D11	Poor control of structural alignment and geometric dimensions	0.111090464	0.888909536
D13	Leakage hazards at key waterproof parts such as welds and construction joints	0.103193727	0.896806273
E1	Forced construction under adverse weather conditions	0.104158053	0.895841947
E2	Inadequate response to complex geological condition risks	0.116697067	0.883302933
E3	Confined construction sites and cross-operation interference	0.123071250	0.876928750
E4	Construction risks adjacent to existing facilities	0.104050658	0.895949342
E5	Inadequate control of special operational environments	0.106387370	0.893612630
E6	Lack of management for construction environmental protection and civilized construction	0.109000627	0.890999373
E7	Defects in traffic organization and diversion plans during construction	0.098146047	0.901853953

.

3.2. Inference and Analysis of the BN Model

(1)

Forward Causal Reasoning

Using GeNIe 5.0 software, forward reasoning yielded a construction quality risk level of 66% for Project T, indicating a relatively high risk, requiring countermeasures according to Table 11 (Figure 6).

(2)

Backward Diagnostic Reasoning

By setting the target node “Construction Quality Risk” to state Y, backward reasoning obtained the posterior probabilities (Figure 7,

Table 16. Posterior probabilities of risk factors.

Root Node	Risk Factor	Posterior Probability
A4	Insufficient capability or stability of the core project management team	0.174904
E2	Inadequate response to complex geological condition risks	0.167554
E3	Confined construction sites and cross-operation interference	0.164696
E7	Defects in traffic organization and diversion plans during construction	0.161848
D2	Defects in subgrade and pavement compaction process control	0.158674
D11	Poor control of structural alignment and geometric dimensions	0.158344
A1	Insufficient qualifications, experience, and practical capabilities of key personnel	0.158189
C3	Poor on-site storage of finished products, semi-finished products, and hazardous materials	0.157949
D13	Leakage hazards at key waterproof parts such as welds and construction joints	0.157621
A2	Inadequate quality and safety technical disclosures	0.157272
E6	Lack of management for construction environmental protection and civilized construction	0.156922
D3	Defects in pavement paving and joint construction processes	0.154836
E1	Forced construction under adverse weather conditions	0.152606
C1	Out of control inspection and acceptance of raw materials entering the site	0.152125
D9	Insufficient preparation and argumentation of special construction plans	0.152125
E4	Construction risks adjacent to existing facilities	0.151886
D6	Defects in pile foundation piling construction technology	0.149602
B1	Poor condition of key construction machinery and equipment	0.147569
E5	Inadequate control of special operational environments	0.146329
C2	Loss of control over mixture production, transportation, and construction temperatures	0.146090
C4	Improper management and use of connection and protective materials	0.145769
B4	Lack of maintenance and calibration management for construction equipment	0.145518
B2	Safety hazards in lifting equipment and rigging	0.137065
D10	Out of control third-party testing and experimental management	0.136635
B3	Inaccurate measurement and test monitoring equipment	0.129800

).

Factors A4, E2, E3, and E7 exhibited the highest posterior probabilities, making them key control targets.

(3)

Sensitivity Analysis

Sensitivity analysis results are shown in Figure 8 and Figure 9, with detailed values in

Table 17. Sensitivity values of BN nodes.

Root Node	Risk Factor	Sensitivity
A1	Insufficient qualifications, experience, and practical capabilities of key personnel	0.1645420
A4	Insufficient capability or stability of the core project management team	0.1447970
A2	Inadequate quality and safety technical disclosures	0.1333410
B1	Poor condition of key construction machinery and equipment	0.1160250
C1	Out of control inspection and acceptance of raw materials entering the site	0.1095980
B2	Safety hazards in lifting equipment and rigging	0.0936718
B3	Inaccurate measurement and test monitoring equipment	0.0922844
C4	Improper management and use of connection and protective materials	0.0867231
B4	Lack of maintenance and calibration management for construction equipment	0.0850531
C2	Loss of control over mixture production, transportation, and construction temperatures	0.0810476
C3	Poor on-site storage of finished products, semi-finished products, and hazardous materials	0.0809737
D9	Insufficient preparation and argumentation of special construction plans	0.0776911
E2	Inadequate response to complex geological condition risks	0.0645270
D13	Leakage hazards at key waterproof parts such as welds and construction joints	0.0615199
E3	Confined construction sites and cross-operation interference	0.0607362
D2	Defects in subgrade and pavement compaction process control	0.0606976
E1	Forced construction under adverse weather conditions	0.0605541
E5	Inadequate control of special operational environments	0.0601408
D6	Defects in pile foundation piling construction technology	0.0600416
E7	Defects in traffic organization and diversion plans during construction	0.0599606
E6	Lack of management for construction environmental protection and civilized construction	0.0574572
D10	Out of control third-party testing and experimental management	0.0573131
D3	Defects in pavement paving and joint construction processes	0.0565326
E4	Construction risks adjacent to existing facilities	0.0563028
D11	Poor control of structural alignment and geometric dimensions	0.0540964

[39].

The top five most sensitive factors (A1, A4, A2, B1, C1) require proactive prevention to reduce overall risk occurrence probabilities [40].

4. Risk Prevention Suggestions and Implementation Effects

4.1. Risk Prevention Suggestions

Targeted countermeasures are proposed across five dimensions:

(1)

Personnel Management: Enforce strict access and dynamic assessment for key personnel, ensuring team stability and operative technical disclosures.

(2)

Machinery Equipment: Strengthen entrance inspections, standard lifting operations, and daily maintenance.

(3)

Material Control: Strictly control material acceptance, storage, and mixture temperature during transportation.

(4)

Environmental Response: Establish systematic mechanisms, optimize site layouts, and improve traffic diversion plans.

(5)

Method Optimization: Deepen special plans and strengthen concealed works acceptance.

4.2. Implementation Effects

After systematically implementing the above construction quality risk prevention and control scheme, Project T has achieved significant quantitative improvement in quality, cost, duration and other aspects. It achieved an initial weld pass rate of 97% (while the average rate of similar projects is generally 90%), reduced rework costs by approximately 600,000 CNY (0.1% of total cost), and adhered to the timeline, demonstrating feasibility and effectiveness. This reveals that the model contributed to cost reduction and quality improvement by providing quantitative decision support.

4.3. Discussion and Value Analysis

Compared to static models, this BN framework allows for logical reasoning and provides a basis for early risk identification. To ensure the objectivity of the validation process, the expert groups and scoring datasets used for model construction are decoupled from those used in the empirical analysis. Furthermore, Leave-One-Out Cross-Validation (LOOCV) was applied to the evaluation data provided by the 26 experts. It turns out that the BN model maintains high consistency across various sample combinations, thereby confirming its strong internal stability.

This study offers practical support for the management of urban transport infrastructure in Hangzhou. By identifying key risk nodes, the framework assists managers in allocating resources more effectively, which helps in reducing potential failures and associated rework. This approach contributes to improving project efficiency and supports local efforts toward more sustainable construction practices.

5. Conclusions and Discussion

5.1. Conclusions

This paper has conducted a study on the construction quality risk management of Hangzhou expressway projects. It identified and organized the construction quality risk factors of Hangzhou expressway projects, forming a comprehensive inventory of these risk factors. Subsequently, based on this risk factor inventory, a construction quality risk assessment model for Hangzhou expressway projects was constructed utilizing BN structure. This model was then applied to the risk evaluation of the T Expressway Project in Hangzhou. Based on the risk evaluation results, targeted construction quality improvement measures were proposed, achieving favorable risk management outcomes and ultimately verifying the scientificity and effectiveness of the research findings. This study has played a certain role in promoting the high-quality development of Hangzhou’s expressways. The research conclusions of this paper are as follows:

(1)

A localized construction quality risk inventory comprising five primary categories—personnel, machinery, material, method, and environment—and 32 discrete risk factors was successfully established. Deviating from traditional manual extraction, this study implemented a dual-stage filtering mechanism: a preliminary semantic mining utilizing a Large Language Model (LLM) over a localized corpus of industry specifications, engineering reports, and the academic literature, followed by Delphi-based consensus refinement. The resulting inventory provides a standardized ontological framework, minimizing human cognitive biases and serving as a foundational benchmark for quality control in high-density urban infrastructure projects.

(2)

A robust risk assessment model was established by mapping the text-derived risk factors into a parameterized BN topology. To resolve the “black box” challenge inherent in network structure definitions, the Apriori association rule algorithm was introduced to optimize directional edges based on statistical co-occurrences, thereby ensuring structural traceability. Root node prior probabilities and non-root conditional probabilities were subsequently quantified by integrating fuzzy set theory and combinatorial weighting with expert knowledge. This synthesis effectively captures the dynamic, non-linear correlations among intertwined risks, transforming qualitative textual logic into a rigorous probabilistic diagnostic tool.

(3)

The empirical diagnosis of Project T verified the practical scalability of the integrated framework. Through forward probability updating, backward diagnostic inference, and critical sensitivity analysis, the model isolated key risk drivers—such as mixture temperature loss—from background operational noise. Rather than yielding purely descriptive outputs, the model generated targeted risk response strategies across five dimensions: dynamic credential audits for personnel management, preventive maintenance for machinery, precise thermal tracking for material control, rapid response protocols for environmental shifts, and operational workflow optimization. These empirical findings demonstrate that the model provides precise, data-informed decision support, enabling project managers to shift from experience-driven reactive troubleshooting to proactive, resource-optimized risk mitigation in complex urban environments.

The methodological advancements and contributions of this research are primarily reflected in the following aspects: First, it breaks through the limitations of traditional risk management, which relies excessively on expert brainstorming, by incorporating LLM to achieve automated semantic mining from massive unstructured engineering texts. This significantly enhances the objectivity and granularity of risk identification. Second, to address the “black box” challenge in constructing BN topologies, this study innovatively introduces the Apriori association rule algorithm to optimize the network structure, ensuring the logical traceability and data-driven support for the directional edges between risk factors. Finally, through the integrated technical chain of “LLM-based identification, Apriori-based optimization, and fuzzy Noisy-OR parameterization,” this framework achieves a transition from experience-driven fuzzy estimation to data-informed precise diagnosis. The specific risk factors identified through empirical analysis, such as mixture temperature loss, demonstrate the practical applicability of this assessment framework in actual projects. Currently, the model has proven to offer valuable management support for urban expressway construction in Hangzhou, and its evaluation approach can also serve as a reference for similar projects in other regions.

5.2. Future Outlook

In summary, this study not only provides practical management suggestions for the quality control of Project T but also demonstrates a risk assessment approach that integrates data with expert experience. This work may serve as a useful reference for similar spatial infrastructure construction projects in high-density urban environments.

This paper has conducted research on the construction quality risk management of Hangzhou expressway projects and constructed a corresponding construction quality risk assessment model, achieving certain results. However, there is still some room for improvement in the research findings, primarily encompassing the following two aspects:

(1)

The depth and breadth of risk factor identification need to be expanded. This study primarily conducts risk factor mining based on textual materials such as industry specifications, project reports, and academic papers related to Hangzhou expressway projects. Although the combination of LLM and the Delphi method ensures the comprehensiveness of the risk inventory to a certain extent, omissions may still exist. Future research could further expand the scope of the corpus to include more diverse cases of expressway projects. Simultaneously, combining methods such as field investigations and in-depth interviews can continuously enrich and refine the risk factor inventory, thereby enhancing the universality of the research conclusions.

(2)

The diversification of research perspectives remains to be explored. This study primarily conducts research on the construction quality risk management of Hangzhou expressway projects from the perspective of the constructor. However, the construction of Hangzhou expressway projects involves multiple stakeholders, including owners, designers, constructors, supervisors, and material suppliers. There are differences among these parties regarding their cognition of quality risks, boundaries of responsibility, and management and control capabilities. Future research could attempt to introduce theories such as multi-party game theory and collaborative governance to construct a mechanism for shared risk bearing and collaborative management of construction quality risks in urban expressway projects with the participation of multiple subjects.