Systemic and dynamic risk analysis of drilling construction based on bayesian network and system dynamics model

Abstract

Drilling construction is a key means to obtain underground resources and information, and construction risk also directly affects the safety, economy and sustainability of engineering projects. Aiming at the limitations of traditional risk assessment methods in the dynamic and systematic aspects, this paper proposes a comprehensive analysis method combining Bayesian network (BN) and system dynamics (SD). First, through expert interviews and field investigations, an evaluation system consisting of 5 dimensions and 20 indicators was established, and the combination of entropy weight method and expert score was used to empower the system. Then the three-layer Bayesian network was constructed by GeNIe, and the system dynamics model with 5 feedback loops was established by Vensim-PLE to realize the static and dynamic coupling analysis of risk factors. The study found that equipment integrity (59%), geological conditions (55%) and operational skills (51%) were the leading risk factors. Finally, the model is applied to the actual scene, and the results are in agreement with the actual situation. This study provides a theoretical framework of probability prediction and process optimization for the systematic analysis of complex drilling construction, and provides a more comprehensive risk analysis for drilling safety under complex geological conditions.

Introduction

At present, the global mining depth adjustment has a significant impact on China’s geological exploration input, which is reflected in the optimization of investment structure, the diversification of mineral resource hotspots and the continuous increase of exploration proportion in the western region. As the core technical means of geological investigation, geological drilling is faced with complex challenges: its operating environment is both harsh in the field, variable geological conditions and complex equipment, resulting in multiple safety risks in the construction process^1,2,3. A typical case shows that in 2020, a drilling company broke the base of the pneumatic winch due to improper operation of equipment during operation, causing injuries to personnel. In May 2022, when an ultra-deep well in Tarim, Xinjiang was drilling into the Ordovician carbonate rock formation, the wellbore pressure imbalance was caused by high temperature and high pressure, resulting in drilling fluid loss and induced kick. The urgency of security risk prevention and control has become increasingly prominent.

The academic community has conducted multidimensional studies on drilling safety risks. For example, Ibrahim et al.⁴ built risk prevention and control strategies through the assessment of wellbore stability. Xu team⁵ established a drilling risk profile model to break through the bottleneck of small sample analysis; Zhao et al.⁶ reveal the instability mechanism of tilted seabed wellhead. In terms of methodological evolution, Huang⁷ applied the planning method and probability theory to optimize the construction scheme, and Dahlin⁸ applied the Monte-Carlo method to quantify the geological uncertainty. Since the twenty-first century, Fan Honghai⁹ developed a real-time monitoring system for complex working conditions in underground Wells, and Li Qi¹⁰ introduced analytic hierarchy process (AHP) to improve the accuracy of risk prediction. Zhou¹¹ achieved high-precision parameter prediction and fuzzy risk assessment under small sample data through the integration of deep forest algorithm and fuzzy set pair analysis, providing a reliable tool for the safety control of construction in complex strata. Zhou¹² innovatively combined the Rooting theory with the double-deep Q-network (DDQN), systematically identified 37 key influencing factors of metro construction safety risks, and quantified their mechanism of action through displacement importance.

Intelligent decision technology has also made breakthroughs in recent years. For example, Shehadeh’s team^{13,14,15,16,17} built a multi-objective optimization expert system to achieve significant optimization of engineering efficiency and cost. The ISSD algorithm developed by Alshboul et al.¹⁸ improved the accuracy of slope monitoring to 0.98. The innovative establishment of contract optimization framework¹⁹ reveals the strong correlation between contract stringency and maintenance efficiency; BIM-VR fusion technology²⁰ reduces design conflicts by 37% and improves collaboration efficiency by 62%. Therefore, the application of machine learning can better deal with some decision-making problems. In the field of risk modeling, Bayesian networks have demonstrated unique advantages. For instance, Meng et al.²¹ developed a dynamic Bayesian network (DBN) to construct a model for managing the accident chain of pressure drilling, and Sule et al.²² verified the dynamic reasoning advantage of DBN through real-time blowout risk monitoring. Tian et al.²³ verified the reliability of oil spill accident analysis. Ma²⁴ proposes a fusion method based on Principal Component Analysis (PCA) and Bayesian inference. By introducing parameter uncertainty and real-time data update mechanism, combined with Markov Chain Monte Carlo (MCMC) to optimize parameter estimation, the prediction accuracy and dynamic adaptability of the model have been significantly improved. Souvik Das²⁵ significantly improved the efficiency of reliability analysis for complex dynamic systems by integrating the modeling ability of BNN for accidental and cognitive uncertainties with the decoupling characteristics of PDEM.

System dynamics effectively reveals the risk evolution mechanism. For instance, Xie et al.²⁶ constructed an SD model for coal mine gas drainage to achieve stability prediction, and Wang et al.²⁷ combined fault tree and SD to quantify the dynamic probability of blowout fires. Syed Muhammad Taimur Shah²⁸ combined causal cycle diagrams with system dynamics methods to simulate the complex interaction of factors of safety hazards at construction sites in developing countries and formulate management strategies. Liu²⁹ proposed a dynamic risk assessment model for buried gas pipelines based on System Dynamics (SD), and verified the effectiveness of the model in characterizing the spatio-temporal risk evolution law and the “prevention and control” function through cases. However, the existing research still has limitations. Bayesian networks mostly focus on single risks such as pressure management and blowouts, while system dynamics are mainly applied in specific scenarios like gas drainage. It is difficult for them to comprehensively reflect the dynamic interaction among multiple elements including human, machine, environment, management and emergency response in drilling operations.

In this paper, a probability-dynamic dual-drive evaluation framework is proposed. By combining the advantages of DBN causal reasoning with the system evolution mechanism of SD model, a multi-level risk coupling model is constructed to realize the leap from single accident prediction to the risk situational awareness of the whole process. This method breaks through the limitation of traditional evaluation dimension and provides more comprehensive decision support for drilling safety under complex geological conditions.

Methods

This paper obtained data through questionnaire survey. After referring to the official website of the State Department of Emergency Management, the China Safety Production Network and the Code for Safe Construction of Geological Drilling (DB37/T11-2022) [Local standard of Shandong Province], the index system was determined by using fish-thorn map. After the drilling construction safety risk assessment index system was established, the expert scoring method and entropy weight method are used to assign and weight the risk factors in the index system, and evaluate the safety risk of each index. Then, the Bayesian network model is established according to the index system, the risk probability of each index is calculated, and the safety risk of drilling construction is further evaluated. Then the system dynamics model is established using the weights given by the entropy weight method. For a dynamic assessment of drilling construction safety risks, Fig. 1 illustrates the various stages of the research process, including data collection and analysis, development of models, and risk assessment in case studies.

Fig. 1

Technology roadmap.

Data source

Drilling construction safety assessment is to ensure the safety of drilling process, which usually covers drilling equipment and facilities, drilling equipment installation, drilling site electricity, safety protection, site management and other aspects. The Code for Safe Construction of Geological Drilling (DB37/T1811-2022) [Local Standard of Shandong Province]³⁰ provides relevant codes and standards. According to this standard, this paper determines five sub-risks in the index system: human, physical, environment, manage and emergency.

With reference to the cases published on the official website of the Ministry of Emergency Management and the China Production Safety Network³¹, the risk indicator system based on five sub-risks, human, physical, environmental, management and emergency response, was drawn through the analysis of actual cases.

Fishbone diagram analysis

In the risk analysis of drilling construction, through structured classification, the potential risk factors are systematically identified, the causal chain of accidents is visually displayed, and the team is helped to focus on the core causes and formulate targeted preventive measures, thus improving the systematic and comprehensive risk prevention and control. As shown in Fig. 2 below.

Fig. 2

Fishbone chart of drilling construction risk factors.

Index weight

According to the questionnaire survey method, five questionnaires were sent to 25 experts and scholars respectively, and the obtained data were returned to the experts after processing. In this way, continuous and complete data were obtained after three rounds of repetition, so as to carry out statistics on the data of the indicator system. A total of 20 evaluation indexes from C1 to C20 of the five sub-risks of human, physical, environmental, management and emergency response were obtained. Among them, a score above 80 was considered as high risk, a score between 60 and 80 was considered as medium risk, and a score below 60 was considered as low risk. However, for the problem of missing data, when dealing with missing data in Bayesian network parameter learning, we can use list deletion (for a small amount of MCAR data), data filling (such as mean or model prediction), expectation maximization (EM algorithm, Iterative optimization of parameters), Bayesian methods (such as MCMC joint estimation of parameters and missing values), or multiple filling (taking uncertainty into account), where non-random missing (MNAR) requires explicit modeling of missing mechanisms (such as the introduction of indicator variables); Method selection requires a balance of missing mechanisms (MCAR/MAR/MNAR), computational resources, and uncertainty requirements, often combining EM preliminary estimates with Bayesian refinement, or balancing efficiency and accuracy with a tool library such as pgmpy. The expert weight can be obtained by the following formula:

$$W_{{\exp {\text{ert}}}} { = }\frac{{\overline{s}_{i} }}{{\sum\nolimits_{k = 1}^{n} {\overline{s}_{k} } }}$$

The entropy weight method is used to determine the weight of drilling construction safety risk evaluation index³², and the weight of five aspects of drilling construction risk including human, physical, environment, manage and emergency is determined by information entropy. Calculate the information entropy of five indexes of drilling construction risk: human, material, environment, management and emergency, the information entropy E_j for the j attribute is:

$$E_{j} = - \frac{1}{\ln n}\sum\limits_{i = 1}^{n} {p_{ij} } \ln p_{ij}$$

In the formula: n indicates the number of indicators; $p_{ij} = \frac{{z_{ij} }}{{\sum\nolimits_{i = 1}^{n} {z_{ij} } }}$, z_ij is the normalized attribute value, i indicates the sequence number of the sample object, i = 1, 2, …, n; j = 1, 2, …, m.

The attribute importance of the j attribute D_j can be defined as:

$$D_{j} = 1 - E_{j}$$

The weight calculated by the entropy weight method and the weight scored by the expert are fused using the addition average:

$$W_{final} = 0.5W_{expert} + 0.5W_{entropy}$$

Uncertainty analysis

The construction of a reasonable and complete Bayesian network mainly includes the following three aspects: the determination of Bayesian network structure, the determination of basic information of network nodes, and the determination of Bayesian network parameters³³.

In this paper, the Bayesian network node represents the risk factors affecting the safety of drilling construction, and the state of each node is described as the risk size. Each node has three states, namely HIGH (H), MEDIUM (M) and LOW (L), indicating the HIGH (H), MEDIUM (M) and LOW (L) of the impact of risk factors on the safety of drilling construction. Under normal circumstances, the prior probability of each node can be obtained through questionnaire survey.

Qualitative analysis and quantitative evaluation are the two main parts of Bayesian network model. Qualitative analysis and quantitative evaluation can be carried out in three ways: modeling by using expert knowledge, modeling by using existing data, and modeling by combining expert knowledge and machine learning³⁴. Because there are many risk factors affecting the safety of drilling construction, the modeling method combining expert knowledge and machine learning is used to construct the model.

Bayesian network structure learning

Bayesian network structure learning methods are mainly divided into two categories: Bayesian network structure learning based on conditional independence test and search-based scoring method. K2 algorithm and mountain climbing method are two classical algorithms commonly used in search-based scoring methods. Taking human factors as an example, the optimal algorithm is selected for subsequent application through comparison experiments. The results of the two algorithms after parameter learning are compared as shown in Fig. 3 below.

Fig. 3

Results from mountain climbing (left) Results from K2 algorithm (right).

It can be seen from the results that there are differences between the two algorithms, because K2 algorithm can directly constrain the network structure through domain knowledge, so as to avoid generating dependency relationships that do not conform to professional logic. However, mountain climbing method only relies on greedy search of scoring function, which may fall into local optimality and cannot be directly integrated into a prior causal order. In drilling construction risk assessment, K2 algorithm realizes polynomial time complexity in computing efficiency through node sequence constraint and greedy search strategy, which is suitable for processing tens of variables in engineering scenarios. In terms of accuracy, its structure guidance based on domain knowledge and anti-overfitting characteristics of BDe score can generate a risk causal network with high confidence. This paper uses the Bayesian network analysis software GeNIe2.1 to construct the highway safety risk assessment model by K2 algorithm³⁵. The calculation process of K2 algorithm mainly includes the following two aspects:

• (1) Structure scoring function.

It is assumed that the nodes of Bayesian network structure have the same prior probability, the recorded data are independent and equally distributed, no data is lost in the data set, and each variable in the structure is a discrete variable^36,37. Based on these four assumptions, the scoring function is:

$$p\left( {G_{B} ,D} \right) = p\left( {G_{B} } \right)p\left( {\left. D \right|G_{B} } \right) = p\left( {G_{B} } \right) \cdot \prod\limits_{i = 1}^{n} {\prod\limits_{j = 1}^{{q_{i} }} {\frac{{(r_{i} - 1)!}}{{\left( {N_{ij} + r_{i} - 1} \right)!}}} } \cdot \prod\limits_{k = 1}^{{r_{i} }} {N_{ijk} } !$$

Among them, The range ofX_i $\left\{ {x_{i}^{1} , \ldots ,x_{i}^{{r_{i} }} } \right\},$ $N_{ij} = \sum\limits_{k = 1}^{{r_{i} }} {N_{ijk} } ,N_{ijk} {\text{ satisfies }}X_{i} = x_{i}^{k} {\text{ moreover }}\pi_{i} = j,$$p(G_{B} )$is the structure prior probability.

• (2) Set search algorithm.

K2 algorithm can greatly reduce the search space through two restrictions, such as a variable ordering ρ and a positive integer u. The optimal model φ meets two conditions: the number of parent nodes of any variable in φ does not exceed u, ρ is a topological order of φ.

In the search process, K2 algorithm starts from a boundless graph that contains all nodes but no directed edge connecting each node. K2 examines each variable node in ρ one by one in a certain order, determines the parent node of the node, and then adds the directed edge connecting the parent and child nodes.

Bayesian network parameter learning

In this paper, a complete data sample was obtained in the form of a questionnaire, and the maximum likelihood estimation method was used to complete the parameter learning³⁸.

Maximum likelihood estimation: A higher fit of a value to the data indicates that the value can be used as an estimate of the parameter θ, in general, conditional probability P(ϑ|θ₀θ₀) is used to indicate the degree of fitting between a possible value θ₀ of parameter θ and data ϑ. The greater the probability, the higher the degree of fitting between θ₀ and ϑ. Given θ, the conditional probability P(ϑ|θ) of data ϑ is called the likelihood of θ:

$$L(q|\vartheta ) = P(\vartheta |q)$$

If a data ϑ is given and the parameter θ varies on the domain of the data ϑ, then L(θ|ϑ) is a function of the parameter θ, called the likelihood function of θ. The value θ that maximizes L(θ|ϑ) is the maximum likelihood estimation of the parameter θ, or MLE:

$$\theta^{*} { = }\arg \mathop {\sup }\limits_{\theta } L\left( {\left. \theta \right|\vartheta } \right)$$

Basic principles of system dynamics model

System dynamics integrates feedback theory and systems analysis with computer simulation to model complex systems, exploring how internal interactions between factors (positive/negative feedback) shape system behavior³⁹. Changes in interconnected systems trigger material/energy exchanges, altering state variables that recursively influence system dynamics. This input–output interplay forms intricate feedback loops, jointly determining the system’s evolving state and behavioral patterns^40,41,42. The schematic diagram is shown in Fig. 4 below.

Fig. 4

System dynamics model concept diagram.

The behavior of the system is determined by its structure, and the feedback loop is the basis of the structure. The study of causality is the basis of system dynamics. The causal relationship between variables within the system is represented by a straight line or arc with an arrow, called a causal chain. The positive causal chain is represented by + , and A change in the dependent variable (A) causes a change in the effect variable (B) in the same direction, Negative causal bonds are represented by −, where A change in the dependent variable (A) causes an inverse change in the effect variable (B)^43,44. The schematic diagram is shown in Fig. 5 below.

Fig. 5

Causal chain.

Multiple causal chains are connected end to end to form a causal loop. The positive or negative of a causal loop depends on the number of negative causal chains in the loop⁴⁵.

Systematic and dynamic risk analysis

According to the fishbone chart analysis, a total of 20 evaluation indicators from C1 to C20 in 5 risks of human, physical, environmental, management and emergency response were obtained.

Establishment of evaluation index system

By using pre-risk analysis and fishbone diagram to analyze the factors that affect the occurrence of risk accidents in the drilling construction process, the factors that are easy to cause risk accidents are analyzed from five aspects: human, physical, environment, manage and emergency response, so as to lay a good foundation for the establishment of evaluation indicators.

Through the analysis of fishbone diagram in Fig. 2, drilling construction safety risks are divided into five sub-risks, namely human, physical, environment, manage and emergency, and a safety evaluation index system is established, as shown in Fig. 6 below. It is presented in tabular form as shown in Table 1 below.

Fig. 6

Drilling construction safety risk evaluation index system.

Table 1 Drilling construction safety risk evaluation index system.

According to a real case in the process of drilling construction, the risk factors affecting the safety of drilling construction are determined by using the fishbone diagram with reference to the data on the official website, including five aspects of human, physical, environment, manage and emergency response, and the evaluation index system of drilling construction safety risks is drawn.

Determine index weight

According to the determined index system, the expert scoring method and entropy weight method are used to objectively assign weights to the evaluation indicators. In order to minimize the subjectivity in the weight allocation process. In order to minimize the subjectivity in the weight allocation process. Through multiple rounds of expert anonymous score, the score result is modified, and the weight of expert score and the weight of entropy weight method are combined proportionally by direct addition average of linear combination. The human risk is shown in Table 2 below.

Table 2 Human risk assessment index score.

The score obtained by the expert scoring method is normalized to obtain the expert scoring weight W_expert, entropy is used to obtain the weight W_entropy⁴⁶, and the expert scoring weight table and the entropy weight table shown in Table 3, 4 are obtained, taking the human factor as an example.

Table 3 Human factor evaluation index expert score weight sum table.

Table 4 Human factor evaluation index entropy weight method weight sum table.

The weight calculated by the entropy weight method and the weight scored by the expert are fused by the addition average, and then the final weight is obtained by linear weighting, taking the human factor as the example shown in Table 5 below.

Table 5 Human factor evaluation index weight sum table.

From the above weight table, it can be seen that after the combination of expert scoring weight and entropy weight method for double weighting, it can be clearly seen that the evaluation index of operational skills has the largest weight for human factors, and also has a greater impact on the safety risk of drilling construction.4.2 Establish a risk assessment model.

A risk assessment model of drilling construction safety is established based on Bayesian network

Based on the analysis of drilling construction data, this paper constructs an index system of drilling construction safety risk evaluation from five aspects: human factor, physical factor, environment factor, manage factor and emergency factor, including 20 index factors, then, by using the questionnaire data, Bayesian network analysis software GeNIe and perfect integration of machine learning and expert knowledge, a drilling construction safety risk assessment model was built to ensure the accuracy of the Bayesian network model⁴⁷.

Determine risk status valuation

In this paper, the risk value of drilling construction was determined by the method of expert scoring, and the risk value was divided into three levels: high (H), medium (M) and low (L), so as to express the risk value of the risk factor. The data were normalized and the risk evaluation value of the risk factor was calculated as shown in Table 6 below.

Table 6 Statistical table of risk assessment of risk factors.

Structural learning

By means of Bayesian network analysis software GeNIe, this paper combines expert knowledge with relevant content of machine learning to build a drilling construction safety risk assessment model. In this paper, K2 algorithm is used for calculation, and structure learning is carried out after algorithm setting and background knowledge editing⁴⁸. Structure learning is a process of machine learning, and the Bayesian network structure of initial drilling construction safety risk assessment is obtained, as shown in Fig. 7 below.

Fig. 7

Drilling construction safety risk Bayesian network structure.

As can be seen from the Bayesian network structure of drilling construction safety risk in Figs. 4 and 5, the five main factors affecting the level of drilling construction safety risk are human factors, physical factors, environment factors, manage factors and emergency factors.

Parameter learning

On the basis of the constructed drilling construction safety risk Bayesian network structure, the parameter learning of the Bayesian network is carried out to obtain the conditional probability distribution table of each risk factor node in the network structure⁴⁹. After the value range and probability of node variables are defined, the imported data for normalization processing is matched with the constructed Bayesian network security risk assessment model, so that the constructed Bayesian network model is used as the object of parameter learning, and the results of parameter learning are output in the states of HIGH, MEDIUM and LOW. After the local sensitivity analysis, the parameters with high sensitivity are compared with the expert opinions to verify the accuracy of the node parameter algorithm. Take the human factor as an example, as shown in Fig. 8a below.

Fig. 8

(a) Local sensitivity analysis results. (b) Bayesian network parameter learning results.

Take C1 as an example, the HIGH risk probability is adjusted from 0.38 to 0.42, and the HIGH risk probability of the target node is changed from 0.33 to 0.34, and its sensitivity S = (0.34–0.33)/(0.42–0.38) = 0.25, indicating a low sensitivity. The influence of small changes of single parameters on the probability of target nodes is analyzed successively, and then compared according to expert opinions. The final learning results of parameters are shown in Fig. 8b below.

By comparing the above results of Bayesian network parameter learning for drilling construction safety risk evaluation with the obtained data, it is found that the risk factors that have a greater impact on the level of drilling construction safety risk obtained by parameter learning are basically consistent with the accident factors that lead to drilling construction accidents, indicating that the use of Bayesian network for drilling construction safety risk evaluation is scientific and reasonable. It can objectively and truly reflect the level of drilling construction safety risk.

Based on the inference analysis of the results of parameter learning, the influencing factors of drilling construction safety are further analyzed and evaluated, and more targeted suggestions are put forward for risk prevention and control based on the analysis results.

Backward reasoning

When the risk level reaches the highest, that is, the probability that the target node’s “risk level” is in a HIGH state is 100%. By using the software GeNIe, reverse reasoning analysis is carried out to analyze the influence degree of the intermediate node on the safety risk level of drilling construction. The inference results are shown in Fig. 9 below.

Fig. 9

Backward inference result.

Sensitivity analysis

Sensitivity analysis is a process of finding out the significant risk factors affecting the drilling construction safety system on the basis of reverse reasoning, and putting forward more specific suggestions and improvement measures according to the results of sensitivity analysis. The results of sensitivity analysis are shown in Fig. 10. GeNIe indicates the influence degree of sensitive nodes by changing the color of each node. Sensitivity analysis using GeNIe can also obtain specific sensitivity values of each network node, as shown in Fig. 11. In the figure, nodes marked in deep red are more sensitive factors affecting the safety of drilling construction. In addition, the sensitivity marked in light red is also higher. If the above nodes change, it will have a greater impact on the safety of drilling construction.

Fig. 10

Sensitivity analysis results (left).

Fig. 11

Time sensitive value (right).

System dynamics model

Establish system flow diagram

After evaluating the safety risks of drilling construction by using Bayesian networks, the system dynamics model is continued to carry out dynamic analysis. According to the causal relationship among various risk factors, the causal loop diagram shown in Fig. 12 is obtained by using Vensim-PLE software. “ + ” indicates that the cause variable has a positive promotion effect on the result variable, while “−” indicates that the cause variable has a reverse inhibition effect on the result variable.

Fig. 12

Drilling construction causal loop diagram.

On the basis of this causal loop diagram, the system dynamic flow diagram of drilling construction risk can be obtained, as shown in Fig. 13 below, further clarifying the relationship between various variables.

Fig. 13

Drilling construction risk system flow diagram.

Simulation result

According to the weights of various construction risk factors determined by entropy weight method, the main equations of system dynamics are established, and the step size is set as 1 month, and the system simulation is carried out for 6 months. In order to study the influence degree of each subsystem on the risk level of drilling construction, the initial risk value of each subsystem will be changed, when the initial risk values of other subsystems are unchanged, change the initial values of all risk factors under one subsystem at a time. In order to make the impact of each subsystem more distinguishable, the initial risk factor value of each subsystem is increased by 0.1 for simulation. Vensim-PLE software is used to obtain the results as shown in Fig. 14 below. After simulating the influence degree of each subsystem on the risk level of drilling construction, continue to simulate the influence degree of each secondary index on the risk level of its corresponding subsystem.

Fig. 14

Development trend of system risk under different scenarios.

Practical case application

The model is applied to a certain blowout risk scenario, and five factors affecting blowout accident risk (A mud density control, B casing integrity, C geological complexity, D emergency response ability, E management mode) were determined by pre-risk analysis and fishbone diagram analysis. The evaluation index system was established, and then the weights were determined by expert scoring method and entropy weight method. Bayesian network was used to conduct risk assessment on its five indicators, and the sensitivity results obtained were shown in the Fig. 15 below.

Fig. 15

Sensitivity analysis result.

The results showed that mud density control had the greatest impact on the risk of blowout, with the probability of blowout rising from 5 to 28% when mud density was below the threshold. The second is the geological complexity, the change of geological conditions (such as high pressure layer, fault) significantly increase the probability of blowout. The ability of emergency response has significant influence on the severity of consequences.

Then, the five indexes were evaluated dynamically by system dynamics, and the results were as shown in Fig. 16 below.

Fig. 16

System dynamics evaluation results.

It can be seen from Fig. 16 that the changes of risk trends are as follows: mud density control, geological complexity, emergency response ability, casing integrity, management mode. The results are in agreement with those obtained by Bayesian networks.

The results obtained from the model are compared with the results of the actual scenario, as shown in Table 7 below.

Table 7 The model result is compared with the actual scenario.

Results and discussion

This study reveals the dynamic characteristics of different risk factors by integrating fishbone diagram, entropy weight method, Bayesian network and system dynamics model to construct drilling construction safety risk assessment model.

First of all, in terms of methodology, this study uses the combination of expert scoring method (subjective weighting) and entropy weight method (objective weighting) to determine weights, which not only avoids the subjective bias of traditional analytic Hierarchy Process (AHP) over-relying on expert experience, but also makes up for the limitation of single entropy weight method ignoring domain knowledge. For example, as shown in Table 6, in human risk, the significant advantage of the weight of operational skills (0.884) reflects experts’ emphasis on technical ability, while the entropy weight rule strengthens the objective influence of real-time operating conditions of equipment through the degree of data variability (such as low equipment integrity entropy). In contrast, AHP-entropy weight method⁵⁰ also improves the reliability of the results through comprehensive weights in code security analysis, but does not involve dynamic risk transmission analysis. Traditional FTA fault tree analysis is mostly used for static causal chain analysis. In this study, Bayesian networks are introduced to reveal the weakening phenomenon of the role of emergency factors (B5 nodes) at the peak risk (LOW probability rises to 34%) through reverse reasoning. This is in contrast to the dominant emergency model based on oil drilling, highlighting the heterogeneity of risk transmission paths under different engineering scenarios.

Through the horizontal comparison of the weight distribution mechanism, the weight results of this study show that operating skills (0.884), equipment integrity (0.825) and geological conditions (0.738) are the dominant factors in each dimension, which is consistent with the core theory of “man–machine-environment”. However, the equipment reliability weight (0.589) is significantly lower than that of similar studies (such as the mean equipment reliability weight of 0.72 in power safety analysis), possibly because drilling operations are more sensitive to the real-time status of equipment, and the contribution of static reliability indicators (such as mean time to failure) is weakened. The weight of emergency agencies (0.676) exceeds that of traditional management factors (such as safety inspection 0.561), which is in contrast to the conclusion that traditional safety management studies focus on system construction. This phenomenon may reflect the dependence of the high-risk nature of drilling operations on immediate emergency response, and is methodologically complementary to the static management perspective of the six-dimension⁵¹ analysis of power safety hazards, which emphasizes “governance cycle” and “capital input”.

In terms of risk transmission mechanism, in Fig. 17 and Fig.18, the reverse reasoning of Bayesian networks exposes the weakening characteristics of the role of emergency factors (B5 nodes) in extreme risk states (LOW probability increased to 34%), which forms a significant difference from the leading model of emergency response in oil drilling scenarios, confirming the heterogeneity of risk paths in different engineering scenarios. This finding breaks through the static causal chain limitation of traditional FTA fault tree analysis, and provides a new perspective for dynamic emergency plan design. Compared with the “governance cycle” dimension emphasized in the field of power safety, drilling operations need to establish an emergency system linked with real-time conditions.

Fig. 17

Parameter learning results analysis.

Fig. 18

Reverse reasoning results analysis.

Fig. 19

The risk development trend of each sub-risk to the corresponding index under different schemes.

Through the dynamic deduction of the system dynamics model, as shown in Fig. 19, the temporal evolution law of risk factors is further quantified: the monthly average change rate of personnel skill risk is higher than that of equipment risk, which accurately matches the growth trend of personnel factor accidents revealed by the International Drilling Safety Association. In the sensitivity analysis in Fig. 20, the nodes marked as triangles (B1, B2, C7, B3, B4, B5) are sensitive factors affecting the safety of drilling construction. Their high sensitivity means that the fluctuation of the risk factor parameters has a great impact on the target nodes, which can directly affect the construction safety, and subtle changes may cause chain reactions. The mechanism that these sensitive factors affect the safety risk of drilling construction is usually as follows: operation skill factor, and its safety risk mechanism is operation error → equipment abnormal → accident trigger; Or lack of skills → delay in emergency response → escalation of incidents. Due to geological conditions, the safety risk mechanism is formation sudden change → equipment overload → structure failure. Or geological uncertainty → parameter misjudgment → chain risk. In order to mitigate the risks associated with highly sensitive factors in drilling construction, multi-level and targeted technical and management measures can be adopted. In terms of operational skills, the professional level of personnel is improved through precise training and intelligent auxiliary systems. For the uncertainty of geological conditions, it is necessary to rely on high-precision geological modeling and real-time monitoring technology dynamic optimization scheme. At the level of comprehensive management, the digital twin system is constructed to simulate the state of physical equipment synchronously to realize risk advance warning. The results of sensitivity analysis are consistent with the “not correctly realized” high-risk behavior identified by the HFACS model⁵⁰. However, this study further quantifies its probability threshold (such as 51% probability of HIGH operation skill) through Bayesian networks, thus enhancing the operability of decision making.

Fig. 20

Sensitivity analysis result.

As shown in Figs. 15 and 16 and Table 7, when the model is applied to the actual scene, the results obtained through the double evaluation of the Bayesian network and the system dynamics model are compared with the actual scene. Mud density control, geological complexity and emergency response ability are the three most important factors. It can be seen that the dual mechanism of Bayesian network and system dynamics is suitable for the collaborative management and control of complex geological conditions and multi-factor coupling risks in drilling construction.

The triple methodological innovation forms a breakthrough to the traditional risk assessment paradigm: the combined weighting mechanism solves the problem of subjective and objective information asymmetry, the Bayesian network realizes the visual analysis of the risk transmission path, and the system dynamics model reveals the time-varying law of the risk elements. This “static weights-dynamic conduction-temporal evolution” triple analytical framework demonstrates greater engineering adaptability and decision support than a single methodology such as AHP-entropy weights or traditional FTAs, especially in dealing with equipment sensitivity and geological mutability specific to drilling operations. The research confirms that the contribution of the six triangular sensitive nodes identified in the dynamic model to the system risk is much higher than that of the traditional static model, which provides a new theoretical tool for accurate risk prevention and control.

This paper combines several onshore cases to study the safety risk factors of drilling construction, and the research conclusions can provide a basic framework for different drilling scenarios. However, onshore and offshore risks also focus on different places, onshore risks are more focused on human operation and equipment maintenance, while offshore risks are compounded by complex geological and extreme environmental challenges, which need to be further optimized with specific environmental parameters. Regarding universality, operational skills: Both onshore and offshore, human error is always the main cause, but emergency response skills are more required at sea, such as rapid blowout closure. Equipment integrity: Equipment failure is a common risk across environments, but offshore equipment requires higher safety redundancy, such as dual BOP systems. Geological conditions: Geological complexity is common, but offshore exploration costs are high, data is limited, and risk prediction is more difficult.

Conclusion

Based on actual drilling engineering scenarios, this study builds an evaluation model that integrates subjective and objective empowerment and dynamic risk transmission, and forms innovative conclusions with engineering practice value:

• (1) The multi-dimensional risk assessment system based on portfolio weighting reveals the core risk elements of drilling operations. The evaluation system of 20 indicators constructed by entropy weight method shows that operation skills (0.8846), equipment integrity (0.8252) and geological conditions (0.7380) occupy the dominant position of human, object and environment dimensions respectively, which forms an empirical echo of the “man–machine-environment” theory. This mechanism of subjective and objective empowerment can effectively avoid the subjective deviation of traditional AHP method.

• (2) Bayesian network reverse inference exposes the scene dependence characteristic of emergency system. When the system risk is in the peak state (the HIGH probability of A1 node is 100%), the LOW probability of the emergency factor increases from 33 to 34% of the base value, which is significantly different from the dominant model of emergency response in the oil drilling scenario. This phenomenon reveals a special law of risk transmission in drilling operations: under extreme conditions, the chain reaction speed of geological mutation and equipment failure exceeds the conventional emergency response time, confirming the priority of real-time monitoring systems over post-disposal mechanisms.

• (3) System dynamics simulation quantifies the temporal evolution of risk factors. It is found that the monthly average change rate of personnel risk is higher than that of equipment risk, and its evolution curve is consistent with the fitting degree of the accident statistics of the International Drilling Safety Association. Compared with the HFACS model, this study transformed the traditional qualitative analysis into quantifiable risk early warning indicators by establishing a “weight-probability-dynamic” three-dimensional correlation matrix, and realized the accurate decision of risk prevention and control.

• (4) Methodology integration innovation lays a theoretical foundation for the development of intelligent security system. The “entropy weight—Bayes—system dynamics” hybrid model constructed in this study improves the risk identification accuracy of a single method and enhances the analysis ability of sensitive node contribution.

In the future research, the system risks and patterns can be identified by big data analysis, risk prediction and early warning can be conducted by machine learning, and intelligent decision support system can be developed to make fast and accurate safety decisions in complex situations, improve the safety risk assessment of drilling construction, and help ensure the life safety of construction personnel and the smooth progress of engineering projects.