Safety risk identification is the prerequisite of the safety risk assessment. The existing literature, project cases, accident reports, and industry standards were often selected as texts for extracting the safety risks, but the previous scholar might select different frameworks to identify safety risks (27, 30). Some scholars used the framework of “machine-environment-management” (Framework 1) (23, 24, 29, 31), and they divided the identified second-level safety risks into machine-type, environment-type, and management-type safety risks. Machine-type safety risks refer to the safety risks arising from machine failure or misoperation; environment-type safety risks refer to the safety risks arising from poor geological and hydrologic conditions; and management-type safety risks focus on the organization or management aspects (e.g., technique arrangements, team or organization management). For instance, Hu et al. (23) used this framework to investigate the subway shield construction safety risks when under-crossing soft overburden layers, and identified geological complex conditions, underground water conditions, the minimum thickness of the overburden layer, the minimum radius of curvature, construction speed, distance from the surrounding environment, and construction management level as second-level safety risks; Zhai et al. (31) applied this view to examine the subway shield construction safety risks when being adjacent to an existing bridge and summed up the relevant second-level safety risks, involving geological and hydrological condition, shield construction parameters, tunnel conditions, bridge conditions, and organization and management risks. Some other scholars argued the subway shield construction is a complex system, and the “personnel-machine-environment” system was important when tackling this type of complex system problem, thus they used the framework of “personnel-machine-environment” (Framework 2). Compared to Framework 1, this framework incorporates the management-type safety risks into environment-type safety risks and highlights personnel's roles in safety risk controlling. Besides, the personnel-type safety risks are only related to the individual participating in construction and were further subdivided into worker-type safety risks and manager-type safety risks (5, 6). For example, Chen et al. (5) and Liu et al. (6) used this framework to examine subway shield construction safety risks when under-crossing complex overburden layers, subdivided environment-type safety risks into natural environment safety risks, manmade environment safety risks, management environment safety risks, and working environment safety risks, and identified 35 relevant second-level safety risks. Other scholars slightly modified Framework 2, they sorted out the management-type safety risks from the environment-type safety risks like Framework 1 and formed the framework of “personnel-machine-environment-management” (Framework 3). For instance, Pan et al. (13) and Wu et al. (32) utilized this framework to investigate the shield construction safety risks respectively. Pan et al.'s research identified 12 second-level safety risks, and Wu et al.'s study listed 15 second-level safety risks. Besides, there exists an all-inclusive framework, i.e., the “personnel-machine-material-method-environment” framework (Framework 4). Compared to the three frameworks mentioned earlier, this framework adds material-type and method-type safety risks. Material-type safety risks refer to the safety risks that arise from raw materials, components, semi-finished products, and finished products during construction (e.g., broken pipe segments, unqualified grouting materials); Method-type safety risks mainly focus on the construction technology aspects (e.g., shield tunneling technology and grouting technology). Li et al. (33) and Fan and Wang (34) examined the subway shield construction safety risk and collected the safety risks based on the framework. Above all, the four frameworks mentioned earlier were aligned with the total quality management theory of “man-machine-material-method-environment.” The differences between these four frameworks stem from the degree of the previous scholars' attention to these different aspects. The four frameworks provide the theoretical or literature basis for the identification framework applied in the study.

Literature on safety risk assessment focused on designing safety risk assessment models. The safety risk assessment model describes the calculation process of the risk assessment. As there exist multiple safety risks in the index framework, the two key questions in designing the assessment model are the weight-determining method and the safety risk measurement. As for the weight-determining method, earlier research selected the analytic hierarchy process (AHP) to determine the weight system. For instance, Li et al. (33) selected AHP to calculate the weights of safety risks when investigating the safety risks in slurry balancing shield construction. Gradually, some more objective techniques were introduced to mitigate the subjectivity in determining the weights. For instance, Fan and Wang (34) applied ISM-DEMATEL and Shapley value method to determine weights of safety risks to consider the relationships between different safety risks; Zhai et al. (31) selected a combinatorial weighting method by integrating G1 and CRITIC on shield construction safety risk when being adjacent to an existing bridge.

As for the safety risk measurement, previous literature can identify many quantitative methods. The fuzzy comprehensive evaluation method is widely applied in the evaluation of shield construction safety risks (33, 35). For instance, Ren et al. (35) chose this method to assess the total shield construction safety risks when being adjacent to an existing building. The matter-element method is another widely used technique (25, 34). This technique has advantages in determining the risk rating by connecting the risk and its risk criteria (4). At present, the Bayesian network (29, 36), cloud model (24, 32), and evidence reasoning theory (29) were also introduced into the area of safety risks to decrease the impact of uncertainty. Wu et al. (29) integrated fuzzy Bayesian and evidence theory to evaluate the subway shield construction safety risk when under-crossing an existing tunnel; Wu et al. (32) chose the cloud model to evaluate the shield construction safety risks; further, Chen et al. (24) applied the extension cloud theory and optimal cloud entropy to examine the shield construction safety risk when being adjacent to an existing building. Besides, Monte Carlo (31) and Systematic dynamic (SD) (13) were also applied to the shield construction safety risk assessment by simulating the probability sampling process and the dynamic relationships between different safety risks.

The above studies showed that an evaluation approach including a weight-determining method and safety risk measurement method was needed for a multifactor evaluation framework. For the weight-determining method, the key to the method selection is to mitigate the subjectivity caused by fewer experts. Confirmatory factor analysis (CFA) is an effective relationship extraction method based on more data collected by questionnaire (32); thus, the method can significantly reduce the impact of subjective judgment caused by a small number of experts. Therefore, this study applied the CFA to determine the weights of safety risks. Besides, evidence reasoning theory is a useful approach to integrate more information and reduce uncertainty (29), thus we selected fuzzy evidence reasoning to measure the value of a single safety risk and the onsite overall safety risk.

Although few studies investigated the safety risks during SSCUR, the frameworks identified in the previous literature can provide feasible frameworks to identify the safety risks. As explained earlier, Framework 3 and Framework 4 are more methodical frameworks. First, Framework 3 does not include material-type safety risks when compared to Framework 4. The reason is that there might be duplication in safety risk statistics in Framework 4, namely, non-standard materials cannot be used in the construction after multiple rounds of material inspections, while the material damage of the final-finished production is mostly caused by the non-standard construction operations, which are analyzed in personnel-type, machine-type, and method-type safety risks (32, 34). Second, Framework 3 distinguishes management-type safety risk from environment-type safety risk. The two are different. Management-type safety risks arise more from organization-level non-compliant behaviors. However, Framework 4 incorporates the method-type safety risk in its analysis views. Therefore, we adjusted Framework 3 based on Framework 4 and proposed the safety risk identification framework for SSCUR. The identification framework is presented in Figure 1. It is noted that personnel-type safety risks arise from inappropriate physical, psychological, and ability conditions, and non-compliant behaviors at the individual level; and the personnel include workers and managers.

Based on the aforementioned safety risk identification framework, the paper applied a two-step approach to collect the safety risks of SSCUR. First, we used a literature review to identify the related safety risks. Second, we arranged an expert group to evaluate and enrich the pre-identified safety risks.

Step 1. Safety risk identification based on literature review.

We selected CNKI and Scopus as retrieval databases and the searching strategy is ((“safety” OR “risk^*”) AND “shield construction”) OR [“shield construction” AND “river^*” AND (“safety” OR “river”)]. We also limited the type of literature to articles and reviews. The search shows 84 English articles and 75 Chinese articles. The researchers selected the final relevant studies by reading titles, abstracts, and texts. Two selection criteria were used to determine the relevance to the research, including: (a) removing the studies which are unrelated to shield construction, for example, the term “shield construction” and “safety” are all in the abstract of Yang et al.'s study (37), but this research is focused on the on-site shield spoil utilization; (b) removing the studies which are unrelated to shield construction safety or shield construction risks, for instance, Dai et al.'s (38) paper mentioned “shield construction” and “safety,” but the research is about the attitude and position prediction of shield machine in tunneling. After the exclusion process, 54 articles were retained, including 31 English articles and 23 Chinese articles. Safety risks were initially extracted from the retained articles based on the identification framework mentioned earlier. The preliminary safety risk list (including first-level safety risk categories and second-level safety risks) is presented in Appendix A.

Step 2. Safety risk evaluation and enrichment based on experts' group evaluation.

Twenty safety management experts were invited to evaluate the safety risk list gained in the previous step (see Appendix A). The invited 20 experts all have at least 5 years of construction experience in shield construction, and they include 2 professor-level experts, 5 senior engineers, and 13 site engineers. The experts were asked to mark whether the safety risks in the list were important or not and to propose any improvements to the safety risk list. If more than 10 experts marked the safety risk, then the safety risk was retained in the list. Based on the statistics on the experts' marks, safety control, failure of soil transport vehicle, safety management system, and high underground water were removed from the safety risk list. Besides, 5 experts were suggesting that incorporate safety skills and safety experience into safety competency. After the above-mentioned amendment, the final safety risk list of SSCUR was presented in Table 1.

Confirmatory factor analysis (CFA) is a widely used data analysis method belonging to factor analysis. Compared to the exploratory factor analysis, which is used to find out the common factor structure from the messy data, CFA aims at validating the feasibility of pre-identified common factor structure (i.e., dimension structure) (43, 44).

A questionnaire survey was chosen to collect data for further CFA (45). The questionnaire is presented in Appendix B. The questionnaire survey was conducted by using the Wenjuanxing platform (46) and the compiled online questionnaire was sent to onsite managers who have cooperation relationships with the researchers. A total of 232 questionnaires were distributed and 197 responses were retained after the validity test. The demographic information of 197 respondents was presented in Appendix C.

The retained data were first imported into SPSS 23 to test the reliability (47, 48). Cronbach's alpha was 0.942 (49), indicating the collected data have high internal consistency and reliability. Bartlett's test of sphericity produced a significant p-value (p < 0.001), showing that the data follow the normal distributions (50). The value of KMO was 0.921, suggesting the high correlations between the items and good sampling adequacy for factor analysis (39, 51).

Then, the collected data were imported into AMOS 23 to conduct a CFA (52, 53). The concept model was constructed based on the safety risk framework and it was presented in Figure 2. CFA shows a better fit (χ2/df = 2.33 < 3; CFI = 0.903) and the standard path coefficients can be gained, which are shown in Table 2. The standard path coefficients denote the relationship strength between different variables (54–57), namely the second-level safety risks and their first-level safety risk. Thus, based on the standard path coefficients, we calculated the weights of the safety risks, which are shown in Table 3.

The FER algorithm includes the follow-up five steps.

Step 1. Using triangular fuzzy numbers to represent safety risks.

Safety risk R can be expressed as R = P × S, that is, the production of the occurrence probability of a safety risk P and the consequence severity of a safety risk S. Due to the influence of uncertain factors, it is often difficult to measure the safety risks quantitatively. A practical and effective approach is to apply qualitative descriptions to represent safety risk grades (41). The occurrence probability of a safety risk can be qualitatively expressed on a verbal scale as “extremely low,” “low,” “relatively high,” “high,” and “extremely high.” The consequence severity can also be described as “no impact,” “minor,” “large,” “dangerous,” and “catastrophic.” In this paper, the verbal assessment grade descriptions of P and S are transformed into triangular fuzzy numbers, and the corresponding relationships were shown in Table 4.

Assuming that the occurrence probability P and the consequence severity S are represented by triangular fuzzy numbers, i.e., P~=(lP,mP,uP), S~=(lS,mS,uS), then the value of the corresponding safety risk is also a triangular fuzzy number, which can be represented by Equation 1 (58, 59).

Step 2. Constructing the fuzzy confidence structure model for the safety risks assessment grades.

Assuming that there exist N assessment grades for each safety risk and the corresponding membership functions are known, thus we can establish the fuzzy confidence structure of the safety risk assessment grades, which is presented by Equation 2.

In the Equation 2, FH_n denotes the fuzzy assessment grades; N denotes the numbers of the assessment grades; β_n denotes the confidence level of R at fuzzy assessment grades, besides, β_n≥0 and ∑n=1Nβn≤ 1.

In this article, we assumed that there exist five assessment grades for a safety risk and the corresponding membership functions adopted fuzzy triangular numbers. The definition and description of the safety risks grades are presented in Table 5. As such, the fuzzy confidence structure of the safety risks assessment grades can be expressed by Equation 3 and the membership functions of the 5 assessment grades are depicted in Figure 3.

Step 3. Calculating the confidence level of each safety risk.

Through this step, we can calculate the confidence structure of each safety risk (β_n) based on the triangular fuzzy values of each safety risk (R~). The calculating method includes the four rules below.

(1) Draw the membership curve based on R~;

(2) Find out the intersection points between the membership curve of R~ and the five safety risk grade membership curves in Figure 3, and calculate the corresponding ordinate of each intersection point (i.e., confidence value of the corresponding assessment grades). If there exist no intersection points, the confidence value on this level is zero.

(3) If there exist 2 intersection points between the membership curve of R~ and the five-grade membership curves, the larger ordinate will be used as confidence value;

(4) Sequentially determine the confidence values of a safety risk at the 5 assessment grades, and then standardize the five confidence values. As such the confidence structure of a safety risk Z(R) can be gained, which is expressed by Equation (4).

Step 4. Calculating onsite total safety risk based on evidence reasoning.

Assuming that there exists a risk assessment problem RE with L risk indexes r_i (i = 1, 2, …, L), the weights of these indexes r_i are ω_i, and every risk index follows the fuzzy confidential (Equation 8) (60).

In the above Equations, min denotes the basic fuzzy confidence value of risk index r_i at the fuzzy risk level of FH_n, miH denotes the uncertain risks due to lack of information, which includes m¯iHand m~iH.

Let mI(i)nrepresents the confidence values to the nth assessment grade of an upper-level risk index which the first I lower-level indexes support; and mI(i)H represents the retained probability after all the first i lower-level risk indexes are assigned to all the assessment grades, thus the recursive Equations of mI(i)n and mI(i)H can be presented in Equations 9–12 (60, 61).

Then, the fuzzy confidence values of an upper-level risk index βⁿ can be gained based on Equation 13.

Based on the aforementioned processes, we can calculate the fuzzy confidence values of all the second-level safety risks and the onsite total safety risks.

Step 5. Calculating the expected utility values of the safety risks.

Let u(H_n) represent the expected utility values of the different safety risk grades, then the total expected utility value of safety risks can be gained based on Equation 14 (59).

In Equation 14, u(Hn)=n5,n=1,2,3,⋯N.

Step 6. Determining the grades of the safety risks.

If the total expected utility value u(R)∈, n = 1, 2, 3, ⋯ , N, then the safety risk grade is n.

The subway interval between Xizhou Station and South Huanghe Road Station of Zhengzhou Subway line 12 crossed under the Qili River at K12 + 180~K12 + 270. The right line crosses under the Qili River approximately orthogonally, and the left line crosses under the Qili River at an angle of 80°. Currently, the cross-section of the Qili River is a compound cross-section, with an opening width of about 90m and a bottom width of about 50m. Two levels of slope protection are provided on both banks, with a main channel slope of 1:3 and an embankment slope of 1:3. The slope height is 6.4–7.0m. The river bottom is protected by dry masonry, while the first level slope is protected by a C20 concrete slab. The minimum vertical clear distance between the tunnel arch and the river bottom is approximately 18.2m. The section under-crossing is located in the quicksand layer and the underground water is located above the tunnel top.

The tunnel section under-crossing the Qili River is a highly risky section of the subway interval between Xizhou Station and South Huanghe Road Station. Before under-crossing the Qili River, the project management department invited an experts' group with 15 experts. These experts are from the expert team mentioned earlier (20 safety management experts selected for evaluating safety risks in Section 3.2) and they include 1 professor-level expert, 5 senior engineers, and 9 site engineers. The experts were asked to identify the safety risks while under-crossing the Qili River and then evaluate the occurrence probability and consequence severity of the identified safety risks.

After the onsite investigation, the experts analyzed the project documents and inquired about some project issues from the project managers. Then, a separate safety risk checklist (see Appendix D) was distributed to every expert. The safety risk checklist covers all the pre-identified safety risks in Table 1. The experts marked the safety risks they consider significant and the corresponding occurrence probability and consequence severity of these safety risks. Onsite management personnel collected all the checklists and calculated marked frequency, occurrence probability, and consequence severity of the second-level safety risks. If the marked frequency of a second-level safety risk is less than five, the second-level safety risk is removed. The level of the occurrence probability and consequence severity of a second-level safety risk were determined by the levels selected by most experts. According to the calculation, the selected safety risks by the experts' group were presented in Table 6, and the grades of the occurrence probability and consequence severity of the safety risks were presented in Table 7.

According to the transformation rule (see Table 4), the triangular fuzzy numbers of the safety risks can be calculated and the results are presented in Table 8. Following the transformation rules in step 3, we computed the confidence structure of the second-level safety risks, and the results are shown in Table 9.

Then the confidence structure of the first-level safety risks and the onsite total safety risk can be calculated by using Equations 5–14. The results are shown in Table 10.

We determined the risk grades of the safety risks by using Equation 14, and the results were presented in Table 11.

As can be seen in Table 11, safety inspection (M5), safety organization & duty, quicksand layer (MN3), and high-pressure phreatic water (HE1) are at the high grade, which shows that these safety risks are more likely to cause safety accidents during shield construction. Thus, the onsite manager should develop a special construction plan in advance to identify and prevent these safety risks. Most of the safety risks are medium-grade risks and the results indicate that some common safety countermeasures should be taken to address these safety risks.

As for the different first-level safety risks, although they are all at the medium grade, the management-type safety risks have the maximum expected utility value, the environment-type safety risks, the personnel-type safety risks, and the method-type safety risks follow closely behind, and the machine-type safety risks have the minimum expected utility value. It indicates that non-standard management is still the most important factor causing safety accidents in China. Onsite managers should focus on the control of management-type, environment-type, and personnel-type safety risks. Besides, when compared to worker-type safety risks, manager-type safety risks have higher expected utility values, which indicates that managers are the core of safety risks. The design of safety institutions and management regulations should consider how to improve managers' safety management competency and how to motivate the managers' behavior intentions.

The total onsite safety risk is at the medium grade, which indicates that overall, the safety risk is at an average level. It also can be inferred that the project management department has an average level of management competency on safety risks during SSCUR. More safety experiences should be collected and more safety training should be carried out so that its manager can quickly and gradually improve control and emergency ability on shield construction safety risks.