Mining Causal Chains for Tower Crane Accidents Using an Improved Transformer and Complex Network Model

Abstract

Tower crane structural failures remain a major safety concern on construction sites. To improve accident prevention, this study proposes an intelligent framework that combines an improved Transformer model with a Directional Interest Score (DIS) Apriori algorithm and complex-network analysis. A corpus of 535 tower crane accident reports (2002–2024) was compiled and annotated with causal and accident entities according to system–safety theory. Segment embeddings were introduced to the Transformer to reinforce boundary detection, enabling accurate extraction of causative factors and relation triples. The DIS-Apriori algorithm was then used to mine both positive and negative association rules while aggressively pruning irrelevant item sets. Eventually, causative factors were mapped into a weighted, directed complex network where edge weights reflect the absolute frequency difference between positive and negative rules, and edge directions correspond to their signs. Experiments show that the Transformer achieves higher precision and recall than baseline models, and DIS-Apriori substantially reduces unnecessary item-set complexity while preserving critical rules. Network analysis revealed five critical causal links and a closed-loop causal link that warrant priority intervention. The proposed method delivers a data-driven, explainable tool for pinpointing key risk sources and designing targeted mitigation strategies, offering practical value for intelligent safety management of tower cranes.

1. Introduction

The increasing automation of the construction industry underscores the key role of tower cranes as the primary means of transportation on construction sites [1,2]. However, accidents involving tower cranes occur frequently, often resulting in serious injuries, fatalities, and significant economic losses [3,4]. For tower crane accident cause analysis, previous research often focused on human factors. It is found that human factors are the key cause of tower crane safety risks [5,6]. Between 2018 and 2024, at least 156 deaths and millions in damages were reported in China, based on official statistics from the Ministry of Housing and Urban-Rural Development and the Ministry of Emergency Management. Human factors accounted for 43% of these incidents, followed by material failures at 31% [7]. It highlights the necessity of conducting research into the structural causes of accidents involving tower cranes.

Investigations such as those by Korean authorities identified procedural violations and unsafe behaviors as key risk factors [8,9]. Yang et al. used hierarchical models to explore human-technical interrelations in accident causes [10]. Besides, Lingard et al. proposed a method to measure the magnitude of risk factor safety coefficients based on non-singular weights, as well as a model to calculate an overall index to estimate the probability of safety risks occurring in tower cranes [11]. Zhang et al. applied decision-making experiments and centrality analysis to quantify causal relationships [12].

Recent efforts have begun integrating machine learning into safety analysis. Zhang et al. enhanced a random forest model to predict tower crane accidents with missing data using multiple imputation [13]. Fan et al. discussed the risk environments from the perspective of complex networks [14]. Xing et al. found that adaptive, mode-aware embedding and robustness strategies have improved performance in complex signal settings [15]. Ke et al. figured out that for long-form text, paragraph-level, content-adaptive recurrent architectures with probabilistic pooling have shown benefits for handling long dependencies and interpretability [16]. Milazzo et al. used systems thinking and complex network models to identify the key causal factors and pathways of tower crane accidents [17]. Seven key factors of tower crane accidents were identified, primarily related to stakeholder safety management, project safety management, and on-site personnel operations. Three critical causal links that precipitate tower-crane accidents were uncovered.

Earlier attempts to pair complex-network modeling with tower-crane accident analysis offered valuable insights, yet the absence of automated algorithms made model construction labor-intensive. Despite valuable contributions, existing research still faces three core limitations: (1) semantic parsing and feature extraction from unstructured accident reports remain manual, limiting efficiency and scalability; (2) existing algorithms such as Apriori only capture co-occurrence between causative factors but fail to assess directionality or strength; (3) many models lack adaptability to long-form, domain-specific texts like tower crane incident reports. These limitations restrict the depth and accuracy of safety risk analysis.

This study proposes a novel method that integrates a Transformer model and a Directional Interest Score (DIS) Apriori algorithm. The domain-adapted language model enables automatic extraction of causative entities and triples from long-form accident reports. The DIS-Apriori algorithm, improved with DIS metrics, quantifies both the direction and strength of associations. Together, these components enable the construction of a complex causation network, offering a robust framework for safety risk assessment in tower crane operations.

This study is structured as follows. Section 2 introduces the methods used in this study, including the improvement of the traditional Transformer model based on segment embedding, and the commonly used metrics of association rule mining (Apriori) algorithm: confidence and support. Section 3 is the corpus of 535 reports, the extraction of structured causative factors, network construction, and analysis of the results. Section 4 mainly introduces the conclusion of this study.

2. Methods

2.1. Transformer Models

The self-attention mechanism of the traditional Transformer model firstly enables the model to effectively capture the semantic associations among the words in the document when dealing with tower crane accident documents [1,2]. For example, it can associate the words “tower crane”, “overload”, “collapse”, etc., to understand the relationships in the accident descriptions and accurately identify the related entities. Secondly, compared with the traditional sequence model, which needs to process text sequentially, the Transformer can process in parallel, which greatly improves the efficiency of processing long documents related to tower crane accidents, and can quickly identify named entities in a large number of accident report texts. However, there are two shortcomings in the traditional Transformer model: (1) the traditional Transformer model cannot explicitly distinguish between different sentences in the input sequence in the multi-sentence processing task of natural language processing; (2) the model does not take into account the context of the text and the position information of the text in the process of converting the text into vectors.

To address the above two problems, this study adapts a Transformer-based architecture, incorporating both positional and segment embeddings, to effectively process and extract information from unstructured, long-form accident reports. On the one hand, to compensate for the problem of missing information that different positions of entities may produce different semantics, the model in this study introduces position embedding, which is used to represent the sequential relationship between entities. On the other hand, to distinguish two or more sentences in the input, segment embeddings are introduced, which are used to separate multiple utterances of a long text, see Figure 1.

The position embedding uses absolute position embedding based on trigonometric functions for injecting the position information of elements in the sequence into the model, see expressions from Equations (1)–(7) [3]:

$$PE(i,2k)=\mathrm{s}\mathrm{i}\mathrm{n}\left({\displaystyle \frac{i}{{\mathrm{10,000}}^{2k/d}}}\right)$$

(1)

$$PE(i,2k+1)=\mathrm{c}\mathrm{o}\mathrm{s}\left({\displaystyle \frac{i}{{\mathrm{10,000}}^{2k/d}}}\right)$$

(2)

where i denotes the position index in the sequence (e.g., the ith word), and k is the dimension index in the position embedding vector (from 0 → d − 1, with d being the model dimension, i.e., the word vector dimension).

In order to accelerate the gradient operation, encoding different dimensions by sine and cosine functions can capture the position information of different cycles, enabling the model to exploit position differences at different scales. The transformed formula is as follows:

$$PE(pos,2i)=\mathrm{s}\mathrm{i}\mathrm{n}\left({\displaystyle \frac{pos}{{\mathrm{10,000}}^{2i/d_{\mathrm{m}\mathrm{o}\mathrm{d}el}}}}\right)$$

(3)

$$PE(pos,2i+1)=\mathrm{c}\mathrm{o}\mathrm{s}\left({\displaystyle \frac{pos}{{\mathrm{10,000}}^{2i/d_{\mathrm{m}\mathrm{o}\mathrm{d}el}}}}\right)$$

(4)

where pos is the position index and i is the dimension index, which is the model hidden layer dimension.

$$\mathrm{s}\mathrm{i}\mathrm{n}(\alpha +\beta )=\mathrm{s}\mathrm{i}\mathrm{n}\alpha \mathrm{c}\mathrm{o}\mathrm{s} \beta$$

(5)

$$\mathrm{c}\mathrm{o}\mathrm{s}(\alpha +\beta )=\mathrm{c}\mathrm{o}\mathrm{s}\alpha \mathrm{c}\mathrm{o}\mathrm{s}\beta -\mathrm{s}\mathrm{i}\mathrm{n}\alpha \mathrm{s}\mathrm{i}\mathrm{n}\beta$$

(6)

$$PE\left(pos+k,2i\right)=PE\left(pos,2i\right)\times PE\left(k,2i+1\right)+PE\left(pos,2i+1\right)\times PE\left(k,2i\right)$$

(7)

$$PE\left(pos+k,2i+1\right)PE(pos,2i+1)\times PE(k,2i+1)-PE(pos,2i)\times PE(k,2i)$$

(8)

Segment embeddings are a learnable parameter matrix [4]:

$$E_{seg}\in R^{K\times d_{model}}$$

(9)

where K is the number of segment types (K = 2 in BERT, corresponding to segment 0 and segment 1), and $d_{model}$ is the model hidden layer dimension.

Since long texts usually have several paragraphs greater than 2, after extending the number of paragraph types, Equation (10) can distinguish between multiple paragraph or sentence positions. Specifically, K is extended to M, i.e., each sentence is assigned a unique paragraph ID (e.g., 0, 1, 2, …, M − 1). For an input containing M sentences, the token segment embedding for each sentence is as follows:

$${\mathbf{e}}_{\mathrm{s}\mathrm{e}\mathrm{g}}^{\left(i\right)}=E_{seg}\left[se{g}_{-}i{d}^{\left(i\right)}\right]$$

(10)

where ${seg\_id}^{\left(i\right)}$ is the sentence number to which the i-th token belongs.

The final input representation of the model is the sum of token embeddings, position embeddings, and segment embeddings:

$${\mathbf{h}}_{input}^{\left(i\right)}=e_{token}^{\left(i\right)}+e_{pos}^{\left(i\right)}+e_{seg}^{\left(i\right)}$$

(11)

2.2. Association Rule Mining

In association rule mining, the evaluation metrics of the Apriori algorithm are mainly used to measure the importance of frequent itemsets and the effectiveness of association rules. The core evaluation metrics and the calculation formula are as follows, see Equations (12) and (13):

(1)

Support

$$Support(X\to Y)={\displaystyle \frac{n\left(X\right)\cap n\left(Y\right)}{N}}$$

(12)

where $\mathrm{n}\left(X\right)\cap n\left(Y\right)$ is the number of transactions containing both X and Y, and N is the total number of transactions.

(2)

Confidence

$$Confidence(X\to Y)={\displaystyle \frac{\mathrm{n}\left(X\right)\cap n\left(Y\right)}{n\left(X\right)}}$$

(13)

where $\mathrm{n}\left(X\right)\cap n\left(Y\right)$ is the number of transactions containing both X and Y, and n(X) is the number of transactions containing X.

2.3. Steps of the Proposed Method

Figure 2 illustrates the whole process of mining and obtaining the content of tower crane accident documents based on the Transformer model illustrated in Section 2.1.

(1)

Step 1. Determine whether the tower crane incident reports are duplicated or not. If not duplicated, then proceed to remove the unstructured noise; if duplicated, then remove the duplicated reports before proceeding to the next step. Standardize the format of the text and maintain a uniform text format of accident description, accident cause analysis, accident handling, and lessons learned. The processed tower crane accident texts were aggregated into a domain-specific corpus for tower crane accidents.

(2)

Step 2. Based on the system safety theory, the tower crane accident causative entity and accident entity are divided into four levels, which are precisely labelled according to anthropogenic cause, material factors, management factors, and environmental factors. Then, pre-train the model to focus on the field of tower crane accidents.

(3)

Step 3. The Transformer model, illustrated in Section 2.1, is for parallel processing of tower crane secondary risk factor extraction and extracting causal relationship triples. The DIS metrics are introduced to improve the Apriori algorithm, which prunes the extracted causal triples, removes the invalid item sets, and retains the key strong association causal triples. Subsequently, the DIS Apriori algorithm retains the positive and negative association rules.

(4)

Step 4. Constructing a complex network model for the causes of structured accidents in tower cranes, identifying key causal links and closed-loop links. Then identify the key causal nodes by eigenvector centrality.

3. Results and Discussion

3.1. Construction of Tower Crane Accidents Corpus

3.1.1. Types of Physical Cause Tower Crane Accidents

This study collected 535 public reports of structured safety accidents of tower cranes in China during the last 22 years, from 2002 to 2024, as a corpus of tower crane accidents domain by crawler technology. The accident reports are published by local government departments and the Emergency Management Bureau on official websites, and include accident typology, severity classification, and causal failure points in tower crane systems.

The types of accidents are classified into six categories, i.e., lifting injuries, collapse accidents, elevated fall, object strike, mechanical damage, and other injuries (weather, electrocution factors), numbered A1–A6, respectively. Incident extraction and categorical counting methods were used to obtain the frequency of each type of accident, the frequency (%) being the ratio of the frequency of each type of accident to the total frequency. Statistical data on the types of accidental injuries are shown in

Table 1. Types of tower crane safety accidents and statistical data.

Type of Accident	Number	Frequency	Frequency (%)
Lifting injury	A1	167	31.21%
Collapse accident	A2	53	9.91%
Elevated fall	A3	142	26.54%
Object strike	A4	120	22.43%
Mechanical damage	A5	22	4.11%
Other injuries (weather, electrocution factors)	A6	31	5.80%

. The frequency of tower crane safety accidents for each year from 2002 to 2024, and the frequency of the most and least frequent types of accidents for each year, using the year as a category. As can be seen from Table 1, among the types of tower crane accidents and injuries, lifting injury caused the highest percentage of accidents, which is 31.21%, and belongs to the high incidence of accidents. Followed by the elevated fall, 26.54%. The lowest percentage of accidents was mechanical damage at 4.11%. Other injuries were 5.80%.

The severity of the 535 tower crane accidents was classified as Level-I accidents and Level-II accidents. Of these, 82% were Level-I accidents and 18% were Level-II accidents. In an in-depth study of the percentage of each accident type in the larger accidents, the data shows that lifting injuries account for 40.8% of the highest value, followed by object strikes (18%) and falls from heights (Figure 3).

For the lifting injuries that account for the highest percentage of accidents in Figure 3, this study considers the differences in the parts where accidents occur and divides tower cranes into foundation structure, tower body structure, jib structure, slewing structure, counter jib, jacking casing structure, safety device structure, hoisting structure, and external construction structure are nine parts, see Figure 4. The causes of tower crane safety accidents are analysed for the above nine parts.

3.1.2. Domain-Specific Corpus for Tower Crane Accidents

The construction of an exclusive corpus of tower crane accidents is divided into the following steps: i.e., collection of original tower crane accident reports, removing unstructured noise, standardizing text formatting, and aggregation of all texts to build a domain-specific corpus for tower crane accidents, see Figure 5. Because tower crane accident reports belong to long texts, it is particularly important to distinguish between the boundaries between sentences and the boundaries between paragraphs, which is why paragraph embedding is introduced in this study.

Based on the system safety theory [18,19,20,21], the causal entities and accident entities are precisely labelled according to the four levels of anthropogenic cause, material factors, management factors, and environmental factors in

Table 2. Precise labelling of causal and accidental entities.

Classification	Primary Causal Entities	Secondary Causal Entities	Accident Entities
Anthropogenic cause	Operator errors	Unlicensed operation/inadequate qualifications	Collision accidents from operational errors
		Illegal operations
		Fatigued work/distracted attention
		Improper emergency response
	Managerial deficiencies	Inadequate safety training	Overall tipping due to overloading and lifting
		Lack of safety technical disclosure
		Illegal command
Material factors	Equipment inherent defects	Design flaws	Tower collapse from structural fracture
		Manufacturing issues
	Component failure/aging	Wire rope abrasion, strand breakage/fracture	Load falling from wire rope fracture
		Limit switch (torque/height/range) failure
		Hydraulic leakage, brake failure
	Lack of safety devices	Uninstalled/unused anti-collision system	Collision from limit switch failure
		Anemometer failure
Management factors	System deficiencies and Implementation loopholes	Inadequate hazard inspection	Collapse from foundation settlement
		Chaotic subcontract management	Illegal lifting failure from uncontrolled plan approval
	Inadequate emergency and Monitoring	Unpracticed emergency plan	Unregistered operations from chaotic subcontract management
		Monitoring system failure
Environmental factors	Natural environmental risks	Extreme weather operations (winds ≥ Level 6, heavy rain, fog)	Tower tipping in strong winds
		Poor geology (unreinforced soft foundation, waterlogged base)
	Workplace hazards	Narrow site (obstacles within jib range)	Tower collapse from foundation settlement-induced tilt
		Insufficient lighting (obstructed night vision)
		Inadequate distance from high-voltage lines/buildings

. For example, for anthropogenic causes, the primary causal entities include operator errors and managerial deficiencies. Among them, the secondary causal entities of operator errors include unlicensed operation/inadequate qualifications, illegal operations, fatigued work/distracted attention, and improper emergency response. Accident entities for operator errors include collision accidents from operational errors and overall tipping due to overloading or lifting.

3.2. Identification of Causal Elements, Extraction of Causal Triples

3.2.1. Causal Relationship

Aiming at the terminology and causal expression characteristics of tower crane accident texts, a domain-adapted pre-training model is constructed. Specifically, the model is fine-tuned by incorporating the Tower Crane Accident Domain Corpus, the Tower Crane Domain Subsumption Lexicon3, and the Core Causality to improve the accuracy of the model’s Named Entity Recognition (NER) based on the Transformer Model [19]. The model judgement condition is set to mark a factor as “1” if it is mentioned in the accident report, and “0” if it is not involved, without considering its weight.

$$Aij=\{\begin{array}{cc}0,& \mathrm{Factor} \mathrm{A} \mathrm{was} \mathrm{not} \mathrm{mentioned} \mathrm{in} \mathrm{the} \mathrm{accident} \mathrm{report};\\ & \\ 1,& \mathrm{Factor} \mathrm{A} \mathrm{is} \mathrm{mentioned} \mathrm{in} \mathrm{the} \mathrm{accident} \mathrm{report};\end{array}$$

(14)

Based on the specific fine-tuning model and specific judgment rules, this study carries out the identification of the causal elements of tower crane accidents in the exclusive corpus of tower crane accidents. The Boolean data of partial judgment completion is shown in

Table 3. List of subsumed terms in the field of tower cranes.

Synonym	Universal Description
Not following rules, Violating operating regulations, Conduct violating operation rules, Violating regulations, Operating against regulations	Unauthorized operation	Controllable factor
Careless operation, Improper operation, Operational error, Execution error, Improper execution, Inaccurate execution, Careless work, Improper work	Execution Errors
Inaccurate execution, Careless work, Improper work	Poor physical and mental health
Lacking required qualification, Not possessing professional operating qualification, Not holding required professional certificate, Working without professional qualification, Lacking corresponding professional qualification	Insufficient professional qualifications
Not wearing a hard hat, Not wearing protection as required, Insufficient protective equipment, Safety rope and harness not in place, Emergency equipment not arranged	Improper wearing of protective gear
Construction technical briefing, Safety technical briefing, Missing briefing document	Failure to conduct technical briefings
Special construction plan, Construction organization design, Special scheme	Dedicated program error
Safety awareness, Safety cognition, Awareness	Insufficient security awareness
Safety education, Training work, Level of safety training	Perfunctory safety education and training
Inadequate on-site hazard identification, Ineffective implementation of hazard identification, Safety rectification not implemented, Failure to implement hazard identification system, Poor hazard management	Inadequate identification of hidden dangers
Repair, Maintenance, Fixing, Servicing, Care, Protective maintenance	Insufficient Maintenance
Device aging, System malfunction, Protective device ineffective, Device worn/aged, Structural damage, Device damaged	Device failure	Uncontrollable factor
Lifting overweight object, Overweight, Overload, Overcapacity, Excessive load	Overload
Work at height, Working at heights, Construction at heights	Work at Height
Toppling, Capsizing, Collapse, Toppling over, Falling over	Tower Crane Collapse
Poor visibility during operation, Low visibility, Poor range of vision	Poor Visibility
Rainfall, Rain, Rain/Snow, Snowfall, Snowfall, Rainy day, Snowy day, Heavy rain, Torrential rain, Heavy snow	Rain/Snow Weather
Device malfunction, Absence, Defect, No response, Failure, Ineffective	Device Malfunction
Strong wind, High wind speed, Windy, Strong wind force, High wind force	Inadequate Wind Resistance
……	……

.

There are five common causal relationships [22] (see

Table 4. Core causality.

Relationship Category	Specific Description	Definition	Model Recognition Example
Direct causation	Leads to, triggers, causes	Head entity directly triggers the occurrence of the tail entity	Violation of operating procedures → causes → lifting injury
Indirect causation	Facilitates, exacerbates, contributes to	Head entity indirectly leads to the tail entity via intermediate factors	Inadequate maintenance → facilitates → device aging → triggers → safety device failure
Conditional dependence	Due to, under…conditions	Head entity serves as a prerequisite for the occurrence of the tail entity	Due to rain/snow weather → under…conditions → low visibility
Synergism	Jointly causes with…	Multiple head entities act together to induce the tail entity	Moment imbalance and inadequate wind resistance → jointly cause → crane overthrow
Root cause attribution	Facilitates, exacerbates, contributes to	Head entity indirectly leads to the tail entity via intermediate factors	Inadequate maintenance → facilitates → device aging → triggers → safety device failure

): direct causation, indirect causation, conditional dependence, synergism, and root cause attribution. Among them, direct causality, indirect causality, conditional dependency, and synergy are all head entities affecting the tail entity, i.e., positive influence. The root cause–attribution relationship is a negative influence where the tail entity influences the head entity. In this case, the direction of influence between the head entity and the tail entity is reversed.

3.2.2. Extraction of Secondary Risk Factors

To further refine the perception of the causal risk of tower cranes, this study proposes the secondary risk of tower crane accident causation (see

Table A2. Frequency of texts with secondary risk factors for tower cranes.

Secondary Risk Factor	Frequency in Texts	Frequency (%)
Moment limiter malfunction	98	18.31%
Luffing mechanism brake failure	74	13.83%
Scaffold pipe falling	72	13.45%
Safety device failure due to illegal dismantling	57	10.65%
Fatigue and fracture of main limbs at tower base	49	9.16%
Uncontrolled vertical overlapping operations	48	8.97%
Imbalanced moment between front and rear Boom	47	8.79%
Non-standard connection at slewing standard sections	47	8.79%
Severe foundation water accumulation	46	8.60%
Insufficient foundation anti-overturning capacity	45	8.41%
Excessive dynamic load on jacking step	43	8.04%
Incomplete installation of tower crane connecting pins	40	7.48%
Jib slewing device failure	39	7.29%
Wire rope breakage for boom lashing	36	6.73%
Counter-jib fracture with severe downward tilt at distal end	36	6.73%
Cargo slippage due to uneven force distribution (e.g., steel bars)	35	6.54%
Unlocked pin holes at both ends	24	4.49%
Insufficient wind resistance	24	4.49%
Poor clutch engagement	22	4.11%
Fatigue in standard section steel	21	3.93%
Excessive free-standing height	18	3.36%
Unqualified welding quality	17	3.18%
Excessive spacing of high-formwork support uprights	9	1.68%
Total reports	535	/

, Appendix A). This study adopts the Transformer Model method (illustrated in Section 2.1) so that the sum of token embeddings, position embeddings, and segment embeddings determines the model inputs. Twenty-three secondary risks of tower crane accident causation were explored, such as the pin holes at the upper and lower ends not being locked and secured, the torque limiter being out of adjustment, and the brake failure of the luffing mechanism. Secondary risks are direct manifestations of primary risks (see Table 1), and the likelihood of a primary risk occurring increases when a secondary risk is present. Because this study focuses on analyzing the association between structural causative factors of tower cranes, the extracted secondary risk factors are screened to retain only those that are related to the structure of the tower crane equipment.

As seen from

Table A1. Boolean data set of secondary risk factors.

Accident Reports	Unlocked and Unsecured Pinholes at Both Ends	Moment Limiter Failure	Luffing Mechanism Brake Failure	Boom Slewing Device Failure	Insufficient Foundation Anti-Overturning Capacity	Imbalanced Moment Between Front/Rear Booms
Accident reports 1	0	1	1	0	0	0
Accident reports 2	0	0	0	0	0	0
Accident reports 3	0	0	0	0	0	0
Accident reports 4	0	0	0	0	0	0
Accident reports 5	0	0	0	1	0	0
Accident reports 6	1	0	0	0	0	0
Accident reports 7	0	0	0	0	0	0
Accident reports 8	0	0	0	0	0	0
Accident reports 9	0	0	0	0	0	0
Accident reports 10	1	1	0	1	0	1
Accident reports 12	0	0	0	0	0	1
Accident reports 13	0	0	0	0	0	0
Accident reports 14	0	0	0	0	0	0
Accident reports 15	0	1	0	0	0	1

Note: Due to the selection of more than 500 public accident reports in this study, space constraints have limited the selection of only some of the public accident reports to be placed in Table A1. The meaning of 0 and 1 is shown in Equation (13).

, any public accident report contained either a unique accident cause or multiple accident causes co-occurring. There were 256 crane accidents caused by a sole causal factor and 279 crane accidents that occurred under the combined influence of multiple causal factors. The results indicate that either a sole-accident causative factor or multiple-accident causative factors may lead to tower crane accidents. In addition, the frequency of occurrence of each accident causative factor in the total text was obtained as shown in Table A2, which can reflect the importance of accident causative factors to some extent.

3.3. Positive and Negative Rule Mining

3.3.1. Directional Interest Degree Improved Association Rule Algorithm

Association rules involve several concepts such as transactional database, frequent item sets, support, and confidence [23]. The core advantage of the traditional Apriori algorithm is to mine frequent item sets from a large amount of data and generate association rules. The core advantage of the traditional Apriori algorithm is to mine frequent item sets from a large amount of data and generate association rules. The previous study achieved the extraction of causal triples in tower crane accident documents based on the Transformer model (illustrated in Section 2.1). However, according to Table 4, the extracted causal triples are divided into positive and negative causal. Then, the extracted causal triples need to be mined by association rules without losing the causality orientation. The traditional Apriori algorithm only discovers the association between itemsets based on the frequent patterns in the data, without considering the semantic information of the itemsets. Thus, there is a limitation in dealing with the causal directionality of itemsets, which leads to the identification of causal triples from “causal A→causal B” to “{causal A, causal B}”. Based on the above problems, this study introduces the Directional Interest Score metrics to explore the positive and negative associations of rules in association rule mining and to measure the “actual interest value” of rules [24]. The formula is as follows:

$$DIS(X\to Y)={\displaystyle \frac{P⟨Y|X⟩-P⟨Y|\neg X⟩}{P⟨Y|X⟩+P⟨Y|\neg X⟩+\epsilon }}$$

(15)

When $P<Y|X>>P<Y|\neg X>$, $DIS\in (0, 1]$, the larger the value, the stronger the positive correlation; and vice versa, the stronger the negative correlation. ε (${10}^9$ this study) is the smoothing coefficient, which mainly serves to avoid the denominator being 0.

3.3.2. Comparison of the Directional Interest Degree Improved Association Rule Algorithm with Other Algorithms

The advantages of the DIS indicator are as follows: (1) in combination with the traditional Apriori algorithm, it can distinguish between positive and negative associations of binary itemsets by the positive and negative values of the indicator; (2) the item sets with DIS < 0.5 can be regarded as rules with unclear causal direction or weak influence, thus realizing the pruning strategy for partially invalid item sets. The association rule mining of the tower crane accident report causal triad dataset using the DIS Apriori algorithm is shown in Figure 6.

Overall, the number of item sets in all three curves shows a decreasing trend as the support threshold increases. The traditional Apriori algorithm only has positive rule folds, which indicates that the traditional Apriori algorithm cannot distinguish the positive and negative rules of the item sets. The DIS Apriori algorithm has both positive and negative rule folds. It indicates that the DIS Apriori algorithm can distinguish the positive and negative rules of the item sets, which provides support for preserving the positive and negative rules of causal ternary in the association rule mining process.

Focusing on the curve of the traditional Apriori algorithm, the number of item sets has exceeded 1000 in the initial state, and still retains more than 400 item sets when the support threshold reaches 0.3. Since the number of item sets is too high, the accuracy rate of the mined item sets is low, and the strong correlation between the causal elements of accidents cannot be obtained. Focusing on the DIS Apriori algorithm curve, due to the division of the item sets into positive rule item sets and negative rule item sets, the remaining strong association rules are less than 100 when the support threshold is 0.3. The initial item sets of the positive rule DIS Apriori algorithm curve are less than 700, and with the support threshold increasing, the number of item sets is always lower than that of the traditional Apriori algorithm curve. The initial number of item sets of the negative rule DIS Apriori algorithm curve is already below 400, and when the support threshold reaches 0.3, there are only 18 strong association rules, and the pruning effect is obvious.

In addition, it is found that the number of negative rules is always lower than that of positive rules, because only the root cause attribution relationship leads to the generation of negative rules. The DIS Apriori algorithm can not only differentiate the direction of the association rules, but also be an evaluation index of the strength of the influence between the rules. DIS value is greater than 0, and the bigger the value of DIS is, the greater the strength of the positive influence of the head entity on the tail entity in the accident causation triad. If the value of DIS is less than 0, the bigger the absolute value of DIS is, and the greater the strength of the influence of the head entity on the tail entity is. If the DIS value is less than 0, the larger the absolute value of the DIS value is, the larger the intensity of the negative influence of the head entity on the tail entity in the accident causation triad is, and the DIS value is close to 0, the interaction between the head entity and the tail entity is negligible.

The traditional Apriori algorithm does not introduce the DIS index, and its average DIS value always tends to be close to 0, and almost does not change with the confidence threshold. It indicates that the traditional Apriori algorithm is difficult to use to effectively identify the positive and negative association rules of the itemset, and the ability to distinguish between positive and negative associations is extremely weak. DIS Apriori algorithm increases the average DIS value significantly (from about 0.35 to nearly 0.9) as the confidence threshold increases from 0.65 to 0.9.

Figure 7 indicates that the algorithm can effectively identify positive and negative association rules, and the higher the confidence threshold, the stronger the ability to distinguish between positive and negative association rules, and can more accurately mine out valuable association rules. Comparing the two, the average DIS value of the DIS Apriori algorithm is always significantly higher than that of the traditional Apriori algorithm, which highlights its advantage in solving the problem of identifying positive and negative association rules and proves that the introduction of DIS has a key role in improving the performance of the Apriori algorithm in this field. According to Equation (13), if ε is too small or even infinitely close to 0, and P{Y∣X} + P{Y∣¬X} is also very small, the denominator tends to be close to 0. The calculated DIS value may have a great outlier due to the division operation. For example, the small difference of the numerator will be enlarged when the denominator is very small. Reflected on the curve, there will be abrupt fluctuating points, which will destroy the curve’s original smooth upward trend with the change of the confidence threshold. If ε is too large, the denominator is significantly “propped up”, which will compress the numerical range of DIS calculation results. No matter how big the difference between P{Y∣X} and P{Y∣¬X} is, the denominator is dominated by a large ε, the DIS value becomes smaller as a whole, and the curve will be shifted downward as a whole. Moreover, the difference of DIS values under different confidence thresholds is weakened, and the curve becomes “flatter”. This makes it difficult to reflect the normal law of confidence level’s influence on the average value of DIS.

3.3.3. Classification of Secondary Risk Areas of Tower Cranes

Considering the variability in the parts where the risk factors occur, the 23 secondary risk factors are further numbered according to the nine parts of the tower crane delineated in Figure 4. It can be found that the secondary risk factors occurring in different risk parts are shown in

Table A3. Risk factors for structured tower crane accidents.

Risk Area	Secondary Risk Factors	Serial Number	Explanation	Number of Text Occurrences of Secondary Risk Factors (Sentences)	Number of Occurrences of Text in the Risk Site (Sentence)
Foundation structure (B)	Insufficient foundation anti-overturning capacity	B1	Inadequate capacity to resist overturning forces, potentially causing full crane collapse	45	132
	Unqualified welding quality at foundation anchorage	B2	Substandard welding at foundation connections compromises structural stability	17
	Insufficient wind resistance	B3	Inadequate capacity to withstand wind loads, increasing hazard during high winds	24
	Severe foundation water accumulation	B4	Failure to drain construction wastewater leading to foundation failure/crane overturning	46
Tower body structure (C)	Fatigue in standard section steel	C1	Material fatigue in tower standard sections due to prolonged use, weakening structural integrity	21	110
	Fatigue and fracture of main load-bearing limbs at tower base	C2	Fatigue or fracture of primary load-bearing components at tower base, endangering structural safety	49
	Incomplete installation of tower crane connecting pins	C3	Failure to fully install connection pins per specifications, reducing joint reliability	40
Jib structure (D)	Luffing mechanism brake failure	D1	Failure of jib luffing brake causing uncontrolled trolley movement, affecting operations	74	120
	Unlocked pin holes at jib ends	D2	Unsecured pin holes at upper/lower jib ends may cause component loosening/detachment	24
	Poor clutch engagement	D3	Ineffective clutch engagement impairing power transmission and operational precision	22
Slewing structure (E)	Jib slewing device failure	E1	Malfunction preventing normal rotation of the jib	39	86
	Non-standard connection at slewing standard sections	E2	Improper connections between slewing standard sections compromising rotation stability	47
Counter-jib (F)	Imbalanced moment between front and rear boom	F1	Moment imbalance between jib and counter-jib, potentially causing crane overturn	47	83
	Counter-jib fracture with severe downward tilt at distal end	F2	Structural failure and tilting of counter-jib disrupting crane balance	36
Jacking casing structure (G)	Excessive dynamic load on jacking step	G1	Overloaded dynamic forces during jacking causing step damage or jacking accidents	43	61
	Excessive free-standing height	G2	Free-standing height exceeding safety limits increases instability risk	18
Safety device structure (H)	Safety device failure due to illegal dismantling	H1	Disabling of safety devices from non-compliant removal procedures	57	155
	Moment limiter malfunction	H2	Improperly calibrated moment limiter failing to restrict lifting moments	98
Hoisting structure (I)	Wire rope breakage for boom lashing	I1	Failure of boom lashing ropes may cause boom collapse with severe consequences	36	71
	Cargo slippage due to uneven force distribution (e.g., steel bars)	I2	Improper cargo slinging causing unbalanced load distribution and material drops during hoisting	35
External construction structure (J)	Uncontrolled vertical overlapping operations	J1	Unmanaged overlapping vertical work zones increasing collision risks	48	129
	Scaffold pipe falling	J2	Falling pipes from external scaffolding threatening workers/equipment below	72
	Excessive spacing of high-formwork support uprights	J3	Non-compliant spacing of uprights in tall formwork systems due to construction negligence	9

.

3.4. Structured Causal Modelling of Complex Networked Tower Cranes

3.4.1. Structured Causal Modelling of Complex Networked Tower Cranes

The 23 secondary risk factors in Table A3 and the six types of accidents in Table 1 are set as the network nodes of the complex network model. The number of occurrences of the ternary group of crane accident causative factors is the weight between any two causative factors, relying on the Pajek software to construct a complex network model of structured accident causative factors of the tower crane.

A positive impact network (see Figure A1) organizes associations around the nodes and uses edge weights to quantify the strength of facilitating effects between nodes, thereby showing how a source node promotes a target node in the complex network. A negative impact network (see Figure A2) is defined over the same nodes and uses weights to characterize the strength of inhibitory effects, effectively reflecting suppressive associations in the network. Based on the mutual impacts among structured causal factors of tower cranes, net influence is employed to measure the positive or negative association between any pair of nodes. The positive–negative influence (net-influence) network (see Figure 8) integrates the positive and negative impact views. The net impact is determined by the absolute difference between the positive and negative influence strengths for a node pair, providing a more comprehensive picture of overall relationships and interaction patterns among nodes.

As shown in Figure 8, the model consists of nodes, directed edges, and the weights of edges, and the 23 nodes belong to the significant causative factors of the tower crane accident. All 23 nodes are significant causal factors of tower crane accidents, and a single factor can lead to the occurrence of tower crane safety accidents. The severity of tower crane accidents will be significantly increased when two or more factors occur at the same time. Each directed edge indicates that the node at the end of the edge causes the node at the top of the arrow to occur, i.e., promotes the effect. The direction of the directed edge is determined based on the positive or negative difference in the frequency of occurrence of the triad of positive and negative association rules between the two nodes. Specifically, if the frequency difference is positive, it means that the head node positively influences the tail node, and the arrow points from the head node to the tail node. A negative difference in frequency means that the head node negatively affects the tail node, and the arrow points from the tail node to the head node. The weight of the edge is the absolute value of the difference in the frequency of occurrence of the positive and negative association rule triad between the two nodes, reflecting the strength of the influence of the head node on the tail node.

By using Dijkstra’s algorithm to obtain the maximum weight link as D2 → D1 → C2 → A5, i.e., unlocked pin holes at jib ends → luffing mechanism brake failure → fatigue and fracture of main load-bearing limbs at tower base → mechanical damage, with a weight value of 35. It indicates that the risk of tower crane accidents under this path has the lowest difficulty in occurrence, the highest severity after occurrence, and belongs to the key control causal chain. Higher weight link has C2 → A5 → A3 → G2, i.e., fatigue and fracture of main load-bearing limbs at tower base → mechanical damage → elevated fall → excessive free-standing height, the weight is 25. B4 → B1 → D1 → C2, i.e., severe foundation water accumulation → insufficient foundation anti-overturning capacity → luffing mechanism brake failure → fatigue and fracture of main load-bearing limbs at tower base, with a weight of 25; B2 → B1 → D1 → C2, i.e., unqualified welding quality at foundation anchorage → insufficient foundation anti-overturning capacity → luffing mechanism brake failure → fatigue and fracture of main load-bearing limbs at tower base. The weight value is 26; G1 → C3 → A3 → G2, i.e., excessive dynamic load on jacking step → incomplete installation of tower crane connecting pins → elevated fall → excessive free-standing height.

Not only do the highly weighted links need to be considered, but also the closed-loop links need to be controlled. Link G1 → H1 → I2 → G1, i.e., excessive dynamic load on jacking step → safety device failure due to illegal dismantling → cargo slippage due to uneven force distribution → excessive dynamic load on jacking step, with a weighting of up to 20. The closed loop between the links will result in a cyclic facilitation of the risk factors, with a very high likelihood of the risk of crane accidents occurring. It is not difficult to find that the starting point of this link is G1, and they should focus on the control of the G1 node, from the source, to cut off the circular link propagation.

Three experienced construction-safety experts were invited to validate the five critical causal links and the closed-loop link. Each expert gave a validity judgment for every link. A link was accepted if at least two of the three experts marked Agree. For the control sets attached to each path, experts also rated: Feasibility is the ease of implementation on a typical site. Expected risk reduction means the percentage drop in the path’s propagation probability after controls. Residual risk reflects remaining risk after controls, reported as categories. Using the 2-of-3 rule, all the links were accepted by the experts. Most control sets were rated high feasibility with low-to-moderate cost. As seen from

Table A4. Expert validation.

ID	Links	Description of Links	Acceptance	Feasibility (1–5)	Expected Risk Reduction	Residual Risk (Post-Control)	Final Priority (1–5)
1	D2 → D1 → C2 → A5	Unlocked pin holes at jib ends → luffing mechanism brake failure → fatigue and fracture of main load-bearing limbs at tower base → mechanical damage	Accepted (3/3)	4 [4,5]	35%	Medium–Low	5
2	C2 → A5 → A3 → G2	Fatigue and fracture of main load-bearing limbs at tower base → mechanical damage → elevated fall → excessive free-standing height	Accepted (2/3)	4 [3,4]	28%	Medium	4
3	B4 → B1 → D1 → C2	Severe foundation water accumulation → insufficient foundation anti-overturning capacity → luffing mechanism brake failure → fatigue and fracture of main load-bearing limbs at tower base	Accepted (3/3)	4 [4,5]	32%	Medium–Low	5
4	B2 → B1 → D1 → C2	Unqualified welding quality at foundation anchorage → insufficient foundation anti-overturning capacity → luffing mechanism brake failure → fatigue and fracture of main load-bearing limbs at tower base	Accepted (2/3)	4 [4,4]	25%	Medium	4
5	G1 → H3 → A3 → G2	Excessive dynamic load on jacking step → incomplete installation of tower crane connecting pins → elevated fall → excessive free-standing height	Accepted (2/3)	4 [3,4]	20%	Medium	3
6	G1 → H1 → I2 → G1	Excessive dynamic load on jacking step → safety device failure due to illegal dismantling → cargo slippage due to uneven force distribution → excessive dynamic load on jacking step	Accepted (2/3)	4 [3,4]	27%	Medium–Low	5

, the expected risk reduction was typically meaningful, and the residual risk after controls was usually medium-low to medium.

This study also proposed several targeted strategies. For instance, operators must be certified and regularly reassessed to reduce human error. For equipment failure, predictive maintenance and redundant interlocks are recommended. Environmental risks should be managed with wind sensors and predefined safety thresholds. Management oversight can be improved using digital inspection systems and routine audits. For the closed-loop link, a feedback mechanism is needed. Any failure should trigger re-inspection and retraining. These strategies support effective risk control based on causal link analysis.

According to the complex network model of six causal links, i.e., five critical causal links and one highly weighted cyclic closed link, the tower structure is a critical node in all five highly weighted links. Specifically, C2 (fatigue and fracture of main load-bearing limbs at tower base) is involved in the highest weighted link (35), C2 is involved in the higher weighted links (25, 26), and C3 (incomplete installation) is involved in the link with weight 20 of tower crane connecting pins. Therefore, the tower structure is the core transmission node of multiple causal chains, which is directly related to the serious consequences of mechanical injuries and main limb breakage, and is a high-risk area.

The lifting arm structure appears frequently in the three high-weight links. D2 (unlocked pin holes at jib ends) and D1 (luffing mechanism brake failure) are involved in the highest-weighted link (weight value 35), and D1 is involved in the higher-weighted links (weight values 25 and 26). Therefore, the risk of the lifting arm’s pin fixation and brake directly triggered the risk of the tower body structure, which is one of the key starting points for risk transfer.

The foundation structure is the starting node of the two higher-weighted links (weights 25 and 26). Involving B4 (severe foundation water accumulation) and B2 (unqualified welding quality at foundation anchorage), both lead to B1 (insufficient foundation anti-overturning capacity), which in turn leads to B1 (insufficient foundation anti-overturning capacity). Overturning Capacity, and then transferred to the crane arm, tower structure risk. Therefore, the foundation is the root of the stability of the tower crane, and its problems will lead to chain risk, which belongs to the higher-risk parts.

Jacking set frame structure is a key node in the three links, and is the core of the closed-loop link. The link of weight value 25 involves G2 (excessive free-standing height); the link of weight value 20 involves G1 (excessive dynamic load on jacking step)→G2; the closed loop link (weight value 20) takes G1 as the starting point, forming “G1→G2”. … →G1” cycle promotion, which is the source node of the cut-off cycle. Thus, the dynamic load and free height of the jacking sleeve directly trigger the connecting pin and the cyclic risk, which belongs to the higher-risk part.

The safety device structure and the lifting structure are the key intermediate nodes of the closed-loop link (weight value 20). They involve H1 (safety device failure due to illegal dismantling), I2 (cargo slippage due to uneven force distribution) respectively, and play a transferring role in the “G1 → H1 → I2 → G1” link plays a transmission role and promotes the amplification of the risk loop, which is an important node of the blocking loop and belongs to the higher risk part. Slewing Structure (E), Counter-Jib (F), and External Construction Structure (J) are not involved in all high-weight and closed-loop links, which are lower risk sites.

The physical path of risk propagation on the tower crane is analysed based on the closed-loop risk link (G1 → H1 → I2 → G1). The risk of jacking sleeve structure (G1: excessive dynamic load on jacking step) first affects the safety device structure (H1: safety device failure due to illegal dismantling) (e.g., failure of jacking safety locks and limiting devices); after the failure of the safety device, the After the failure of the safety device, the risk of the lifting structure (I2: cargo slippage due to uneven force distribution) is aggravated (failure of the linkage between the lifting mechanism and the safety device); the impact load generated by the dislodging of the cargo is reacted on the lifting sleeve, which again amplifies the dynamic load problem of G1, forming a closed loop.

3.4.2. Complex Network Model Degree Value Analysis

Eigenvector centrality considers that if there are many important nodes connected to a node in the network [25], then the importance of that node in the network corresponds to a higher degree of importance, and the distribution of eigenvector centrality of nodes is shown in Figure 9. The risk factor with the highest eigenvector centrality is E2 (non-standard connection at slewing standard sections), which is 0.45. It indicates that the non-standard connection at slewing standard sections is connected to many critical nodes, and its stability is crucial for the safety of the whole project. D1 (luffing mechanism brake failure), I2 (cargo slippage due to uneven force distribution), and C2 (fatigue and fracture of main load-bearing limbs at tower base) have eigenvectors. The degree of centrality is also high, which belongs to the key control of tower crane accident causation.

Fatigue and fracture of the main load-bearing limbs at the tower base (C2) belong to the tower body structure (C), as the vertical load-bearing core of the tower crane; fatigue fracture will directly lead to tower instability and collapse. High center degree comes from the tower body structure associated with the foundation, lifting arm, jacking, and other parts of the risk (such as the foundation instability aggravating the main limb force, lifting arm out of control impacting the main limbs), so it is a high-risk part of the tower crane. Cargo slippage due to uneven force distribution (I2) belongs to the hoisting structure (I); cargo dislodgement will have an impact on the main limbs. Cargo slippage due to uneven force distribution (I2) is classified as hoisting structure (I), and the impact load generated by the cargo dislodgement will damage the hoisting mechanism, and it is also prone to cause chain accidents such as fall from height and collision with tower body/boom. The high centrality comes from its risk linkage with safety devices (e.g., H1) and tower body (e.g., C2), which is the key to spreading the risk of hoisting operation to the whole structure. Therefore, it is also a high-risk part of the tower crane.

3.5. Ablation Study

To verify the contribution of the segment-embedding module to causal-factor extraction in tower-crane accident documents, we designed an ablation study based on a Transformer architecture, considering token embeddings, position embeddings, and segment embeddings.

Table 5. Ablation study on segment embeddings.

Model	Precision	Recall	F1 Score
Token embedding, position embedding, and segment embedding	92%	81%	86%
Token embedding and Position embedding	90%	71%	79%

reports precision, recall, and F1-score. On this task, the model with segment embeddings (token + position + segment) achieves 92% precision, 81% recall, and 86% F1, all of which outperform the model without segment embeddings (token + position: 90% precision, 71% recall, 79% F1). The gain in recall is particularly notable (+10 percentage points), indicating that segment embeddings help the model capture latent causal information in long reports and reduce missed detections. The small increase in precision (+2 points) shows that adding segment encoding improves feature separability without materially increasing false positives.

3.6. Contributions of the Algorithmic Improvements

The DIS-Apriori Algorithm proposed in this study goes beyond symmetric co-occurrence and makes direction-sensitive inferences among risk factors.

(1)

Accurate identification of directional relations in tower-crane risks.

In construction risk management, the standard Apriori uncovers only co-occurrence among factors. For example, it may report that severe weather and crane overturning frequently appear together, yet it cannot distinguish whether weather drives overturning or merely co-occurs with it. By computing a directional interestingness score for candidate rules X→Y, our method explicitly characterizes the direction of influence. Using historical tower-crane accident reports, DIS-Apriori yields rules such as severe weather → crane overturning, rather than a non-directional co-occurrence. This enables managers to trace risk back to plausible drivers and to design targeted preventive actions.

(2)

Improved precision of risk assessment.

Risk assessments based solely on co-occurrence frequencies ignore both the direction and the magnitude (positive or negative) of inter-factor effects, leading to biased prioritization. The proposed score quantifies the signed strength of influence between risk factors. In construction settings, this supports finer-grained evaluation of how specific factors affect safety and quality outcomes, thereby enabling more defensible risk grading and resource allocation.

4. Conclusions

This study integrates natural language processing technology with complex network theory to realize the transformation from unstructured accident text to structured causal knowledge. The approach enables earlier identification and proactive prevention of tower-crane safety risks. A corpus of 535 accident reports was first analyzed using a Transformer-based model (Section 2.1) to automatically extract causal entities and relation triples. Compared with manual processing, the pipeline reduces processing time and cost and mitigates human-induced bias. In addition, the traditional Apriori algorithm is improved by introducing the directional interest degree (DIS) index in the association rule mining process. Finally, a complex network model of structured accident causation of tower crane with directionality of association rules is constructed. The following conclusions are drawn from this study:

(1)

Intelligent and accurate identification of causes of tower crane safety accidents based on a Transformer model. Focusing on the field of tower crane-structured accidents, 23 structured causes of tower crane accidents are identified.

(2)

The structural parts of tower cranes are divided into nine categories, and based on the lack of refinement of the classification of primary risks, secondary risk factors are proposed. The structural causes of tower crane accidents are linked to the structural parts of tower cranes, and the structural parts with the highest probability of occurrence of each type of secondary risk factor are intelligently located to provide a scientific basis for the intelligent diagnosis of tower crane risks.

(3)

In the analysis of the tower crane accident causal triad, the improvement of the traditional Apriori algorithm based on the DIS Apriori algorithm is indeed effective in both pruning of frequent itemsets and identification of positive and negative association rules. The number of itemsets is significantly reduced, and the positive and negative association rules are completely retained. In the field of causal chain construction, on the one hand, the causal direction of any two nodes in the process of causal chain construction can be identified by the difference between the frequency of positive and negative association rules of any two nodes. On the other hand, the absolute value of the difference between the frequency of positive and negative association rules reflects the weight of the causal link.

(4)

Through the tower crane structured accident causation complex network model construction to obtain five key causative links and one high-weight value cyclic closed link, and gives the cyclic link risk blocking method, for the future tower crane accident risk prevention to provide ideas.

This study builds a static causal network based on the historical accident text. Future extensions will integrate time-series analysis and dynamic complex-network modeling using real-time monitoring data. Multi-modal data fusion and dynamic risk-evolution modeling will be pursued to enable continuous early warning and adaptive prevention across operating scenarios.