Causal Analysis of Accidents During Wind Power Engineering Construction Using Bayesian Networks

In recent years, countries around the world have largely reached a consensus on reducing carbon dioxide emissions. With the continuous improvement of integrated utilization technologies for renewable energy, wind energy, alongside solar energy, has become a major source of renewable energy. To further enhance the utilization rate of wind resources and energy capture efficiency, the wind power industry is trending towards the construction of larger units and the development of wind power projects in areas with high altitude, low wind speed, and distant offshore locations [1]. The high demand for reliability in the wind power industry is reflected in the safety analysis of wind turbine units, and research on engineering safety is critical not only during the operation and maintenance phase but also during the preliminary design and construction phases to ensure the safety of the engineering systems. To ensure the safety of wind power construction, reduce the accident rate during the construction period, and protect the lives and property of personnel involved in wind power engineering construction, this paper focuses on the causal factors of accidents during the construction period. It incorporates human unsafe behaviors and unsafe states of objects as basic events into the fault tree structure, transforming these into a fault tree-BN model through GeNie for visualization, reverse diagnosis, and accident cause analysis. This lays the foundation for subsequent risk analysis and assessment, playing a significant role in preventing accidents.

Currently, many accident models are based on systems theory, describing the accident process as a complex, interrelated network of events rather than just simple causal chains. In 2005, scholar Lyu et al. [2] proposed the "2-4" Model of accident causation, combining personal behavior (human factor analysis) and organizational behavior (organizational analysis) based on previous studies and the current state of safety management in China. This research aims to establish a safety accident causation model for wind power engineering based on the "2-4" Model, analyzing its feasibility and effectiveness in the analysis of accidents during the construction phase of wind power projects, and summarizing unsafe actions and states from accident samples. The "2-4" Model has been applied in various fields including coal mine safety, construction electrical fire accidents, gas explosion accidents, high-altitude fall accidents, general aviation accidents, food and chemical accident statistical analysis, highway traffic accidents, and urban underground space pipeline collapses. Lyu et al. [2] compared ten different accident causation models, with the description of accident pathways evolving from linear to networked systems, although the "2-4" Model lacks in probabilistic analysis.

Scholars like Zhou et al. [3] have used dynamic BN models for risk assessment of offshore wind farms, establishing dynamic BN models in GeNie, and evaluating their effectiveness in risk analysis at different lifecycle stages of offshore wind farms. Other researchers, such as Zhang, have designed a complete Engineering, Procurement, and Construction (EPC) project management system covering project risk identification, analysis, response, and risk control management, proposing a Bayesian fault tree-based engineering risk assessment strategy for intelligent risk management and control of wind power projects [4]. Adedipe et al. [5] have systematically reviewed and evaluated the existing research on the use of BN models in the wind energy field, demonstrating their wide application from wind power and weather forecasting to risk management, fault diagnosis and prediction, structural analysis, reliability evaluation, and maintenance planning and updating. Yu et al. [6] developed a semi-qualitative risk model by combining BN with Evidence Reasoning (ER) methods to assess the risk of ship-turbine collisions, verifying the BN model through sensitivity analysis and other methods, indicating that the minimum passing distance is a key factor in the risk of collisions between ships and offshore wind turbines.

BN can demonstrate its potential applicability in modeling risks and reliability in different complex systems, considering different system levels and their interactions through causal methods. Thus, BN has significant potential research and practical value in the field of wind power engineering safety. While researchers have extensively applied BN for diverse risk assessments in offshore wind farms and have utilized them in fields spanning from wind energy forecasting, project risk management, to fault diagnosis of wind turbine equipment, structural analysis, reliability evaluation, and maintenance, there remains a notable gap in focusing on accident causation factors associated with onshore wind power engineering construction projects. Specifically, analyses based on the "2-4" Model, encompassing fault tree-BN modeling and the examination of the most critical causation chains, have not been sufficiently explored. Therefore, this study on the accident causation factors involved in the construction period of onshore wind power engineering projects based on BN has significant research value and innovation.

This paper primarily explores the following three issues: (1) How to incorporate human unsafe behaviors and unsafe states of objects as basic events into the fault tree structure diagram during the construction period of wind power engineering accidents. (2) To construct a fault tree-BN model, perform visualization with the help of GeNie, and analyze and diagnose the causes of accidents. (3) What are the key means and effective methods to reduce the occurrence of accidents during the construction period of wind power projects?

Given the short construction cycle, high intensity, and high-quality requirements of wind power engineering projects, along with the harsh natural conditions and adverse environments of wind farms, the safety management capabilities and awareness of construction personnel are relatively low, presenting unique challenges and difficulties in improving the safety management system of wind power projects. Based on the "2-4" Model, this paper establishes a fault tree-BN model of safety accident causation for onshore wind power engineering during the construction period, clarifying the network model of accidents occurring during this phase, which has significant theoretical and practical significance for improving the safety of wind power construction projects, and enhancing the safety level of wind power engineering construction.

3.1 Fault tree analysis method

Fault tree analysis is a graphical analysis method that identifies the causal and logical relationships between various factors related to an accident, tracing from the outcome to the causes. Starting with a specific accident or failure for analysis, it delves into the causes layer by layer until the basic causes, or basic events, are found. After constructing the model, the fault tree uses Boolean algebra to list its mathematical expressions. Boolean algebra is particularly suited to describing the accident process that takes two opposite states. The structure function describes the system's state, completely dependent on the condition of elements or components. It is usually assumed that at any time, elements, components, and the system can only be in a normal or failure state, and at any moment, the system's state is uniquely determined by the state of its elements or components. Assuming a fault tree system is composed of n basic events, the event state function can be defined as:

$\mathrm{x}=\left(\mathrm{x}_1, \mathrm{x}_2, \ldots, \mathrm{x}_{\mathrm{n}}\right)$

where, x_i is the state variable of the i-th basic event.

$x_i=\left\{\begin{array}{c}1 \text { indicates that event } \mathrm{i} \text { occurs }(\mathrm{i}=1,2, \ldots, \mathrm{n}) \\ 0 \text { indicates that event } \mathrm{i} \text { does not occur }(\mathrm{i}=1,2, \ldots, \mathrm{n})\end{array}\right.$

The state of the top event depends on the state of all basic events, i.e., y is a function of x:

$y=\varphi(x)$ is referred to as the fault tree's structure function. y=1 indicates the occurrence of the top event; y=0 indicates the non-occurrence of the top event. The fault tree is composed of events and logic gates, with events divided into bottom events, intermediate events, and top events. Bottom events are located at the lowest end of the fault tree structure diagram, consisting of basic events and undetermined events; intermediate events are the result events between the top event and bottom events; the top event is the target event of study, located at the top of the fault tree, with logic gates divided into AND and OR gates.

By analyzing human unsafe behaviors and unsafe states of objects during the construction period of wind power engineering through the "2-4" Model as basic events, they are categorized into human factors, equipment factors, environmental factors, and management factors as intermediate events, with accidents during the construction period of wind power engineering as the top event, to compile a fault tree.

3.2 BN modeling method

Compared to fault tree analysis, BN can perform bidirectional reasoning, but its structure learning is relatively complex. Converting a fault tree into a BN offers a clearer and more convenient approach. The correspondence between the bottom events, intermediate events, top events, and logic gates in the fault tree and the nodes and logical connections between nodes in the BN are as follows: (1) The bottom events of the fault tree correspond to the root nodes in the BN; (2) Intermediate events correspond to non-root nodes, and top events correspond to leaf nodes; (3) The input-output relationship of AND and OR gates in the fault tree corresponds to the direction of directed edges in the BN.

By using the above correspondences, the fault tree is converted into a BN. This paper utilizes Bayesian theory and the software tool GeNie for visualization operations and employs the EM algorithm for parameter learning to obtain the BN model. GeNie, a BN learning and inference software developed by Bayes Fusion, LLC, based on the Decision Systems Laboratory at the University of Pittsburgh, is a professional tool for BN modeling.

5.1 Model validation

To verify the reliability and predictability of the BN model, effective cross-validation of the network model is conducted. The Leave-One-Out Cross-Validation method is used to calculate the effectiveness of the prediction accuracy for each node in the model. The validation results are shown in Table 3. Among the 34 nodes, the precision of node E reached the highest at 0.947368, followed by F at 0.938235. Overall, the prediction accuracy for most nodes is above 0.8. This indicates that the BN model constructed in this paper has a high prediction accuracy and is suitable for reasoning and analyzing the causation relationships of accidents during the construction period of wind power engineering.

Table 3. Cross-validation results of the model

5.2 Parameter learning

Parameter learning in BN aims to determine the conditional probability distributions of node variables, quantifying the degree of dependency relationships between model node variables. Currently, the main parameter learning algorithms include the Bayesian method, the EM algorithm, and the Maximum Likelihood Estimation (MLE) method, etc. In this case, the EM algorithm is chosen mainly because it can perform MLE of parameters in the presence of missing data, making it well-suited for handling various types of incomplete data. The basic steps of the EM algorithm are as follows:

(1) Calculate $\mathrm{z}^{(\mathrm{i})}=\mathrm{E}\left[\mathrm{Z} \mid \mathrm{y}, \hat{\theta}^{(\mathrm{i})}\right]$;

(2) Expand the observed data y to (y, $\left.\mathrm{z}^{(\mathrm{i})}\right)$, maximize $\pi\left(\theta \mid y, z^{(i)}\right)$, denoting its maximum value as $\hat{\theta}^{(i+1)}$;

(3) Use $\hat{\theta}^{(\mathrm{i}+1)}$ and the results from step (1) to obtain $\mathrm{z}^{(\mathrm{i}+1)}$, which is then substituted into step (2), repeating this process until convergence criteria are met.

where, y represents the observed data, z represents the missing data or latent variable data, $\mathrm{p}(\mathrm{z} \mid \mathrm{y}, \hat{\theta})$ is the predictive distribution of $\mathrm{Z}$ given $y, \hat{\theta} ; \hat{\theta}^{(\mathrm{i})}$ is the estimate at the i-th iteration.

Through parameter learning of the BN, the conditional probabilities between nodes can be obtained. Due to the large number of nodes and limitations in space, this analysis will focus on the conditional probability distributions of key nodes H2 (Personal Protection), M4 (Accident Hazard Investigation), and M5 (Safety Risk Level Control).

From Table 4, it is known that when safety equipment fails, if personnel safety training is adequate and personnel are sufficiently staffed, then the probability of personal protection issues occurring is 0.5. However, even if personnel safety training is adequate, if staffing is insufficient when safety equipment fails, the probability of inadequate personal protection issues occurring is 0.78. If safety equipment fails and personnel safety training is inadequate, regardless of whether staffing is sufficient, personal protection issues are likely to arise. Inadequate personnel education and staffing can affect the adequacy of personal protection in wind power engineering.

When safety equipment is functioning properly, and personnel safety training is adequate with sufficient staffing, the probability of personal protection issues being adequately addressed is 0.88. If safety equipment is functioning properly but personnel safety training is inadequate, even with sufficient staffing, the probability of inadequate personal protection issues occurring is 0.84. If safety equipment is functioning properly, personnel safety training is inadequate, and staffing is insufficient, the probability of inadequate personal protection issues occurring is 0.71.

From Table 5, it is known that during abnormal weather conditions, if accident hazards are adequately addressed, the probability of failure in the wind turbine and its auxiliary equipment not occurring is 0.66; if accident hazards are not adequately addressed, the probability of failure occurring is 0.5. During normal weather conditions, if accident hazards are adequately addressed, the probability of failure in the wind turbine and its auxiliary equipment not occurring is 1.

Table 4. CPT for node H2

Table 5. CPT for child node T7

Table 6. CPT for child node H7

From Table 6, it is known that when equipment carries hazardous voltage and is close to a charged body, the probability of accidentally touching the charged body is 0.73. When equipment does not carry hazardous voltage, the distance to the charged body does not affect the occurrence of danger from accidentally touching the charged body.

5.3 Reverse inference of BN model nodes

Reverse inference in BN model nodes involves calculating the posterior probabilities of other node variables when the target node of the network model is known. By constructing a BN model and performing reverse inference on variables, the posterior probabilities of other node variables can be accurately determined, allowing for a more precise assessment and prediction of accident situations. Based on the constructed BN model, setting the root node "Wind Power Engineering Construction Accident (F)" as the evidence node, the posterior probability distributions of other nodes related to preventing accidents are derived. By analyzing and comparing the posterior probability values of various nodes, the most likely causes of wind power engineering construction accidents can be inferred, as shown in Table 7 and Figure 7.

Table 7. Posterior probability values of key node variables

From Table 7 and Figure 7, it is known that in the event of a wind power engineering construction accident, the probability of inadequate safety management supervision (M1 value as N) exceeds 70%, environmental factors (E value as Y) exceed 63%; the execution of safety management and organizational systems (M2 value as N) exceeds 61%, safety training, education, and guidance (M3 value as N) exceed 61%, management factors (M value as Y) exceed 56%; the inadequacy of personal protection (H2 value as N) exceeds 56%, human factors (H value as Y) exceed 53%; occurrence of abnormal changes in the workplace environment (E5 value as Y) exceeds 50%; inadequate safety risk level control (M5 value as N) exceeds 44%, and failure of lifting equipment (T2 value as N) is at 42%.

5.4 Node sensitivity analysis and maximum causation chain analysis

In the BN model, sensitivity analysis can reflect how changes in local parameters of the network model cause quantitative changes in the target node, thereby identifying sensitive factors within the model. The depth of the red color of the nodes is proportional to their sensitivity; the deeper the red, the higher the sensitivity. At the same time, the accident's maximum causation chain is used to find key risk factors, highlighting the risk causation chain, or the most likely key risk pathway leading to an accident, with a bold arrow. Setting node F as the target node, the variable sensitivity and the maximum causation chain analysis were performed, as shown in Figure 8.

From Figure 8, it is evident that the sensitivity levels of network model nodes vary, with nodes such as E1, E2, E4, E5, M3, M4, M5, T4, M2, and T5 showing high sensitivity. Focusing on these factors for prevention and being vigilant against accidents caused by geological reasons, observing road conditions before transporting wind turbine blades, avoiding operations in limited spaces, staying away from charged bodies, maintaining safety protection equipment, strengthening safety training, education, and guidance, addressing accident hazards in advance, paying attention to safety risk level control, and strictly implementing safety and organizational systems can effectively enhance the ability to resist accidents. In Figure 8, the longest causation chain leading to wind power engineering construction accidents includes E3 (Abnormal Weather) → T7 (Failure of the Wind Turbine and Its Auxiliary Equipment) → H10 (Sensitivity to Equipment Changes) → H (Human Factors) → F (Wind Power Engineering Accident) and M3 (Safety Training, Education, and Guidance) → H5 (Personnel Qualification) → T2 (Lifting Equipment) → T (Equipment Factors) → F (Wind Power Engineering Accident). Therefore, control measures should prioritize these key risk factors to reduce the probability of factors occurring in the maximum causation chain. During the construction of wind power projects, changes in weather or the environment should be closely monitored, along with changes in wind turbines and their auxiliary equipment. In the management process, personnel qualifications and safety training, education, and guidance before operations should be strengthened to timely detect safety hazards, enhance the safety construction awareness of operational and management personnel, ensure the safety of personnel and property, and effectively prevent accidents during the construction period of wind power engineering.

7.png

Figure 7. Distribution map of reverse inference for nodes in the BN model

8.png

Figure 8. Sensitivity analysis and maximum causation chain analysis of variables in the BN model