2.1 Optimization of civil aviation cabin safety

Currently, researchers are focusing on improving the ability to identify cabin safety risks through advanced technologies, as indicated below. Berkhahn [12] approached civil aviation cabin safety from the perspective of “data communication architecture for safety-critical cabin systems” and proposed, for the first time, a communication architecture featuring “triple redundancy + real-time fault self-healing.” This approach expanded the boundary of cabin safety from “single-point equipment reliability” to “system-level communication resilience”. Hsu and Liu [13] integrated “man–machine–environment–management–task” system factors through the five-M model, filling the gap in the systematic theoretical framework for cabin safety research. Zhang et al. [14] investigated evacuation efficiency in fire scenarios and proposed a design scheme to optimize cabin layout. Du et al. [15] established an emergency evacuation model for cabin passengers on the basis of cellular automaton theory, providing auxiliary data for the airworthiness certification of ARJ regional aircraft. Through research on the design of model-based complex system architectures in aircraft cabins, Ghanjaoud et al. [16] revealed via simulation results that the three key nodes in cabin illegal interference incidents are “cabin structure damage”, “insufficient airline training,” and “failure of airport police to promptly take over unruly passengers”. Čavka et al. [17] proposed the first “energy efficiency–safety” dual-objective design framework and introduced the “safety–energy efficiency index (SEI)” for multiobjective optimization of cabin design. Liu and Tong [18] constructed a cabin safety evaluation index system from six aspects—personnel, machinery, environment, management, information, and loading—via system engineering theory and conducted an empirical analysis. Cui et al. [19] established a response lag model for smoke detectors, providing theoretical support for optimizing the layout of cabin smoke detection equipment. Through theoretical models combined with engineering measurements, Belland et al. [20] systematically explored the safety and effectiveness boundaries of ultraviolet-C (UV-C) in cabins, promoting aviation epidemic prevention from “emergency response” to “active defense”. Wu et al. [21] developed a VR-based cabin lithium battery fire drill platform, and their research revealed that VR technology can effectively compensate for the shortcomings of traditional drills and increase practical training effects. Ling et al. [22] proposed an automatic cable segmentation method on the basis of knowledge distillation and context fusion, increasing the accuracy of cabin cable identification and providing technical support for preventing electrical faults and fire risks.

2.2 Civil aviation cabin safety management strategies

Studies have pointed out that comprehensive design and a sound safety management system are critical to increasing cabin safety, and different scholars have explored this topic from multiple dimensions: Molesworth and Burgess [23] approached the macro system of civil aviation cabin safety through “the effectiveness of auditory information,” prompting the industry to shift from “controlling equipment” to “empowering users” and providing an innovative “technology–policy–experience” trinity idea for cabin safety management. On the basis of the theory of planned behavior, Cui et al. [24] reported that risk perception, attitude, and behavioral control ability significantly affect the behavioral intentions of passengers and suggested the implementation of targeted safety education to reduce unsafe behavior. With a complex network model, Wang et al. [25] reported that controlling critical nodes such as crew members and the environment can reduce the probability of risk transmission and improve management levels. Çevik et al. [26] revealed the disregarded safety blind spot from “the uniqueness of cabin crew fatigue”: As pilot fatigue management becomes increasingly standardized, cabin crew fatigue is emerging as a new risk source. By quantifying the transmission mechanism of fatigue in service-safety operations, a “data-driven intervention plan” for cabin safety management can be provided. Through evolutionary game analysis, Wu et al. [27] clarified the decisive role of the safety disposal capabilities of airlines, government supervision, and the punishment intensity of airport public security in the emergency disposal effect of cabin disturbance incidents.

2.3 Aspects of civil aviation cabin personnel management

This research focuses on the effects of crew behavior and psychological factors on cabin safety, highlighting the importance of personnel management and training. For example, Zhang et al. [28] used a data-driven Bayesian network approach to conduct a risk analysis of civil aviation pilot factors from 2008 to 2020. Studies have shown that weather conditions and flight phases are relatively strongly correlated with the type of casualty in accidents, while human factors such as pilot decision-making are more likely to result in fatal accidents. Abdelhakim et al. [29] confirmed the necessity of food safety training from the “training–behavior safety” chain for the first time, expanding the boundary of cabin safety from traditional “equipment/emergency” to “daily operation compliance”. Kim et al. [30] reported that physical fatigue, psychological pressure, and complacency are the main causes of human errors among flight attendants and called for strengthening psychological and behavioral management. Through association rules and social network analysis, Liu et al. [31] reported that human factors account for the greatest proportion of the causes of cabin safety accidents, emphasizing that personnel management and training should be the focus.

2.4 Aspects of civil aviation system optimization

Currently, research on the causal analysis of aviation accidents is trending from single-factor identification to multidimensional and systematic analysis. In terms of research on causal systems, Guo et al. [32] applied the HFACS model combined with deep learning methods to analyze the causes of accidents during the approach and landing phases and clearly noted the key roles of human error, equipment failure, and organizational management defects in accidents. Yue and Li [33] constructed a causal analysis model of aviation accidents on the basis of reverse fuzzy Petri nets and identified the key factors in a man–machine–environment–management system through complex network analysis of 257 aviation accident investigation reports, providing new ideas for accident prevention and control. In terms of the application of text mining, Hu et al. [34] proposed a text mining method on the basis of an SVM and a DNN, which is used to classify NTSB accident reports and extract the text features of accident causes. Tanguy et al. [35] explored the aviation safety report analysis method based on NLP and proposed a complete set of tools for text classification, topic modeling, and similarity search, which can help aviation safety experts analyze accident reports efficiently.

Existing studies have shown that text mining technology has made considerable progress in the field of aviation safety. The integration of natural language processing, machine learning, and deep learning methods can effectively extract potential information from aviation accident reports, significantly increasing the accuracy of accident prediction and the efficiency of risk management and control. However, current research focuses mostly on traditional areas such as flight safety, air traffic control systems, and aircraft maintenance. Relatively few text mining studies target civil aviation cabin safety, and the application of related technologies is relatively backward. Existing cabin safety research focuses mainly on the macro level, such as the optimization of safety management systems and the improvement of operating procedures.

In this study, deep learning technology is applied to the risk management of civil aviation cabin safety to realize the automatic identification and accurate classification of cabin safety accidents and predict potential risk factors. Two innovative breakthroughs are achieved: From a research perspective, it is the first time that deep learning has been applied to the microrisk assessment of cabin safety; from a technical perspective, a full-process solution of “semantic understanding-intelligence assessment is constructed.” To provide a structured overview, Table 1 summarizes representative studies.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary of representative studies on aviation safety text mining and risk assessment.

https://doi.org/10.1371/journal.pone.0343678.t001

Overall, existing studies demonstrate the value of text mining and machine learning for aviation safety analysis; however, most work focuses on general flight safety domains (e.g., ASRS or NTSB reports) and emphasizes clustering or prediction rather than interpretable risk quantification. Moreover, relatively limited attention has been paid to cabin-specific safety narratives and the challenges of imbalanced outcome distributions. These gaps motivate the integrated framework and the CNN–LSTM–Attention-based risk assessment approach proposed in this study.

3.1 Methods for identifying risk sources of civil aviation cabin safety

In this paper, text mining technologies (a bag-of-words model and chi-square dimensionality reduction) are used to automatically extract key causal features from accident reports, addressing the problems of time-consuming manual processing and insufficient coverage. Furthermore, combined with the hazard source identification and risk analysis comprehensive evaluation list of the cabin service department, risk factors are converted to risk sources with distinct causal mechanisms.

3.2 Ethics statement

This study does not involve human participants, animal experiments, or other content requiring ethical review. The research is conducted in strict accordance with academic integrity and ethical norms, and thus no ethical approval was required.

3.3 Bag-of-words model

In the field of text mining and safety analysis, the bag-of-words (BOW) model is widely used for extracting text feature items because of its simplicity and efficiency. By counting the number high-frequency words in cabin safety accident reports, this model can effectively identify critical safety risk factors. Compared with complex grammatical structure analysis, keyword frequency statistics are more explainable and focus on the most representative causal features.

Assume that there is a vocabulary n (containing all unique words in the documents) in the cabin safety accident report database. Each word in text D corresponds to a feature value according to its position in the vocabulary. Each document D_i can be represented as a feature vector x_i, where each dimension corresponds to a word in the vocabulary, and its value represents the number of times the word appears in the document, as shown in Equation (1).

(1)

where f_j represents the occurrence frequency of the j-th word in the vocabulary in document D_i.

3.4 Chi-square statistic

Some high-frequency words extracted by the bag-of-words model are used as the initial feature items of the text. Owing to the high dimension of these feature items and the inclusion of a large amount of redundant information, which reduces computational efficiency, dimension reduction on the text feature items of aviation safety accident reports needs to be performed. In the text feature dimension reduction step, this study uses the chi-square statistic to conduct a thorough screening of risk keywords. This method quantifies the strength of the correlation between words and risk categories by calculating the degree of deviation between the actual observed frequency and the theoretical expected frequency of specific words in accident reports. The statistic is calculated by Equation (2).

(2)

where n represents the total number of texts; a is the frequency of texts that belong to category c_i and contain the feature term t; b denotes the frequency of texts that do not belong to category c_i but contain the feature term t; C indicates the frequency of texts that belong to category c_i but do not contain the feature term t; and d represents the frequency of texts that neither belong to category c_i nor contain the feature term t.

The value of the entire corpus can be calculated using Formula (3).

(3)

By setting the threshold value m, many meaningless feature items can be eliminated, achieving effective dimension reduction of feature items. This method can significantly improve the representation of text features while reducing computational complexity.

3.5 Civil aviation cabin safety risk level assessment model

To accurately assess the risk of civil aviation cabin safety incidents, after the influencing factors of civil aviation cabin safety risks are identified, a large amount of cabin safety incident data is used to construct a civil aviation cabin safety risk level assessment model based on CNN–LSTM attention. The model uses mainly a hierarchical feature coding architecture, and the overall architecture of the model is shown in Fig 1. The main idea is as follows: first, the original text is mapped into high-dimensional dense vectors through the word embedding layer to realize the distributed representation of semantic information; second, the multiscale CNN module is used to extract local semantic patterns, and feature dimension reduction and key feature selection are achieved through pooling operations; third, the LSTM network is used to perform temporal modeling on feature sequences, and the long-distance contextual dependencies are adaptively captured through the gating mechanism; last, the attention mechanism is used to dynamically generate attention weights according to different semantic subspaces, which are fused to create a discriminative feature representation. As shown in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Network structure diagram of CNN–LSTM–attention model.

The model processes input text through sequential layers: (1) an Embedding Layer using Word2Vec to create dense vector representations; (2) a CNN Layer with convolution and pooling to extract local features; (3) an LSTM Layer to capture long-range sequential dependencies; (4) an Attention Layer to weight important features; and (5) an Output Layer for final risk classification.

https://doi.org/10.1371/journal.pone.0343678.g001

3.6 CNN module

1. (1) Convolutional layer The core operation of the convolutional layer is convolution, whose purpose is to extract features from local regions via convolution kernels (filters). Given an input text sequence, each token is first mapped to a d-dimensional embedding. The resulting input to the CNN can be represented as a matrix of size L × d, where L denotes the sequence length and d denotes the embedding dimension. We apply one-dimensional convolution along the sequence dimension, with convolutional filters spanning k consecutive tokens and the full embedding dimension. This design allows the model to capture local n-gram features in the text sequence. The output c_i of the convolution operation can be expressed as Equation (4).
   (4)
   The convolution result c_i is subsequently subjected to a nonlinear transformation through the activation function to obtain the output a_i of the convolutional layer, which can be expressed as Formula (5).
   (5)
2. (2) Max pooling Max pooling is applied along the sequence dimension to aggregate the most salient features across different token positions, and its main functions are described as follows: First, max pooling reduces the spatial dimension of the feature maps through a sliding window mechanism, significantly decreasing the computational complexity of the model; second, max pooling selectively retains the strongest activation features in local regions, improving the ability of the model to identify key features; and last, the generalization performance of the model is improved by introducing translation invariance.
   Compared with average pooling, max pooling is more suitable for text processing tasks. Text features are often characterized by sparsity and local significance, and this method can more effectively capture discriminative key semantic units. The specific calculation process can be formalized as Equation (6), where max pooling is applied over the local feature regions to generate a condensed representation that is passed to subsequent layers.
   (6)
3. (3) Fully connected layer
   After several layers of convolution and pooling operations, the resulting multidimensional feature maps are flattened into a one-dimensional vector and input into the fully connected layer. The fully connected layer performs a linear transformation to map the features to the final classification space. Let the input vector of the fully connected layer be x, the weight matrix be W, and the bias be b. Afterward, the output after calculation can be formalized as Equation (7).
   (7)

3.7 LSTM layer

The LSTM network is composed mainly of three main gating units—the input gate, forget gate and output gate—and includes a candidate memory unit and current memory unit. The specific calculation formulas are presented as follows:

1. (1) Input gate: This gating unit is used to determine how much information from the current input x_t needs to be written into the memory unit, and its calculation process can be formalized as Formula (8).
   (8)
2. (2) Forget gate: This gating unit is used to determine how much information in the memory cell C_t−1 at the previous moment needs to be retained, and its calculation process can be formalized as Equation (9).
   (9)
3. (3) Candidate memory cell: This gating unit is used to generate new candidate information, which is combined with the input gate to determine whether it should be added to the memory cell. The calculation process can be formalized as Equation (10):
   (10)
4. (4) Memory cell update: The memory cell is updated by combining the forget gate and the input gate, and its calculation process can be formalized as Equation (11):
   (11)
   where ⊙ denotes the element-wise (Hadamard) product. (5) Output gate: This gating unit determines how much information in the current memory cell C_t is output to the next moment, and its calculation process can be formalized as Equation (12):
   (12)
5. (6) Hidden state: The final output hidden state h_t is jointly determined by the output gate and the current memory cell, and its calculation process can be formalized as Equation (13):
   (13)
   In the above formulas, x_t represents the current input; h_t−1 represents the hidden state at the previous moment; C_t−1 denotes the memory cell at the previous moment; W and U indicate the input weight matrix and the hidden state weight matrix, respectively; b represents the bias term; denotes the sigmoid activation function, whose output range is [0, 1] and is used to control the flow of information; and indicates the hyperbolic tangent activation function, whose output range is [−1, 1].

3.8 Attention layer

In the text reports of cabin safety accidents, different paragraphs and sentences vary in their importance. The attention mechanism can dynamically assign weights to each time segment of the text sequence to highlight key information. In this study, an additive attention mechanism is used to model the text reports of cabin safety accidents. With trainable parameters, this mechanism dynamically distributes attention weights to different text segments. Compared with the dot-product attention mechanism, the additive attention mechanism is more suitable for scenarios involving long texts with unevenly distributed importance. This mechanism can effectively capture important causal information and nonlinear relationships in accident reports while increasing classification accuracy and improving the explainability of the model. Its calculation process can be formalized as Equations (14)–(16):

(14)

(15)

(16)

Where h_t represents the output hidden state vector of the LSTM layer at time step W_s and b_s represent the trainable parameters. The attention weight is calculated, and then the context feature vector c is obtained through weighted summation.

3.9 Output layer

The output layer uses the softmax function to convert the output of the fully connected layer to a probability distribution of various categories, and its calculation process can be formalized as Equation (17):

(17)

4.1 Model training and evaluation details

To ensure the reproducibility of the proposed CNN–LSTM–Attention model, the main training and testing settings are described as follows.

Text preprocessing included tokenization and lowercasing. Stop words were not removed, as neural models with word embeddings can learn to down-weight non-informative tokens automatically.

For the deep learning models (CNN, CNN–LSTM, and CNN–LSTM–Attention), each token was mapped to a dense vector using Word2Vec embeddings. The embedding dimension was set to 300. All sequences were truncated or padded to a maximum length of 1000 tokens before being fed into the models.

To address the severe class imbalance among the 34 event outcomes, the original event categories were remapped into five risk levels (high, medium-high, moderate, medium-low, and low risk), and this risk-level stratification was used during model training and evaluation.

The dataset was split into training and test sets with a ratio of 70% to 30% using a stratified split to preserve the distribution of risk levels.

4.2 Data sources and processing

4.2.1 Data sources.

The experimental data were obtained from the online database of the Aviation Safety Reporting System (ASRS) at the National Aeronautics and Space Administration (NASA) Center. After filtering through the event assessment and date filters, 12,967 abnormal aviation events in civil aviation cabins from 2004 to 2024 were selected. After cleaning, 8,280 valid text data points were retained, encompassing 34 independent outcomes, as listed in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Description of event outcomes.

https://doi.org/10.1371/journal.pone.0343678.t002

4.3 Imbalance handling

The number of records for one outcome in the data report may be significantly greater than that of other categories. The distribution of outcomes for all hazardous events reported between 2004 and 2024 is shown in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of civil aviation cabin event outcomes.

The histogram displays the frequency of each outcome (x1 to x34) in the dataset (N = 8,280). The highly imbalanced distribution, with some outcomes occurring much more frequently than others, motivates the five-level risk stratification used in this study. For brevity, only outcomes with non-zero incident counts are shown.

https://doi.org/10.1371/journal.pone.0343678.g002

Owing to the significant class imbalance in cabin safety incident records, while traditional machine learning algorithms are usually based on the assumption of balanced data, such data skew severely affects model performance. Therefore, in this study, in accordance with international aviation safety classification standards, the incident consequences are divided into five-level risk systems (high risk, medium–high risk, moderate risk, medium–low risk, and low risk). The data distribution is reconstructed through a structured risk grading strategy to improve the ability of the model to identify key high-risk incidents (see Table 3).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Correspondence table of event outcome risk levels.

https://doi.org/10.1371/journal.pone.0343678.t003

To address the issue of highly imbalanced distribution among the 34 types of events in the original data, this study remaps 8,280 abnormal aviation events through a five-level risk stratification strategy. As shown in Fig 3, the risk level classification effectively improves the distribution of event categories to achieve a more balanced distribution. Ensuring the integrity of the main body is conducive to increasing the accuracy of the model.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of risk levels.

The bar chart illustrates the proportion of each risk category—High, Medium-High, Moderate, Medium-Low, and Low—after regrouping the 34 original event outcomes (N = 8,280). This stratification yields a more balanced distribution compared to the original data, mitigating class imbalance for model training.

https://doi.org/10.1371/journal.pone.0343678.g003

4.4 Model validation

4.4.1 Evaluation indicators.

In the multiclass classification task involving accident-causing factors and accident risk levels, to comprehensively assess the classification performance of the constructed model, the accuracy, precision, recall, and F1 score are selected as evaluation indicators. The structure of the confusion matrix introduced first is shown in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Structure of the confusion matrix.

https://doi.org/10.1371/journal.pone.0343678.t004

TP (true positive) samples predicted as positive that are actually positive. TN(True Negative) refers to samples that are predicted as negative and are actually negative. FP (false positive) samples predicted as positive that are actually negative. FN (false negative) samples predicted as negative that are actually positive. The accuracy, precision, recall, and F1 score are calculated through the confusion matrix, with the calculation formulas shown in Table 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Evaluation indicators.

https://doi.org/10.1371/journal.pone.0343678.t005

4.5 Comparative analysis results

In this study, 8,280 aviation anomaly events recorded between 2004 and 2024 were randomly divided into a training set and a test set at a ratio of 7:3. Specifically, 5,796 events (70%) were allocated to model training, and 2,484 events (30%) were used to assess the generalizability of the model. The training set data were then input into various models, including XGBoost, CNN, CNN–LSTM, and CNN–LSTM–attention, for training and prediction purposes. Table 6 presents the training accuracy of these four models, and Table 7 outlines their prediction and evaluation metrics. To emphasize the performance comparison, the performance indicators of each model are shown in a line chart in Fig 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Training accuracy of different models.

https://doi.org/10.1371/journal.pone.0343678.t006

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Comparison of model evaluation metrics based on the test set.

https://doi.org/10.1371/journal.pone.0343678.t007

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Comparison chart of model evaluation indicators based on the test set.

The line chart displays the Accuracy, Recall, and F1-score for four models: XGBoost, CNN, CNN–LSTM, and the proposed CNN–LSTM–Attention. The proposed model achieves the highest scores across all three metrics (Accuracy, Recall, F1-score), demonstrating its superior capability in classifying cabin safety incident risk levels compared to baseline approaches.

https://doi.org/10.1371/journal.pone.0343678.g004

1. (1) Analysis of the training accuracy of the model
   As shown in Table 6, the CNN–LSTM–attention model has significant advantages in the cabin safety text analysis task. Its training accuracy of 95.63% not only far exceeds that of the traditional XGBOOST model (90.44%) but is also 3.56% and 2.13% higher than that of the CNN (92.07%) and CNN–LSTM (93.50%), respectively. This advantage stems from its multilevel feature extraction capability: the CNN layer effectively captures local text patterns, the LSTM layer models long-distance contextual dependencies, and the attention mechanism dynamically focuses on key information, thereby achieving accurate modeling of the deep semantics of the text.
2. (2) Analysis of the generalizability of the model
   As shown in Table 7 and Fig 4, the accuracy, recall, and F1 score of XGBoost on the test set are inferior to those of the deep learning models (CNN, CNN–LSTM, and CNN–LSTM–attention). Although XGBoost remains an efficient benchmark method because of its advantages of fast computing speed and strong interpretability, it is limited in its ability to extract deep semantic features. The CNN efficiently extracts local text features through convolution operations, and its classification performance on the test set surpasses that of XGBoost. CNN–LSTM uses an LSTM structure to model the temporal dependencies of text, and its accuracy and F1 score are significantly better than those of the CNN. The CNN–LSTM–attention model uses an attention mechanism on this basis, which can dynamically assign weights to focus on key information, achieving better classification results on the test set and demonstrating better generalizability.

The CNN–LSTM–attention model in this study achieved the best performance, demonstrating the significant advantages of deep learning methods that combine local feature extraction, sequence modeling, and attention mechanisms in the risk assessment of civil aviation cabin safety.

4.6 Level-wise performance and error analysis

To provide a more detailed evaluation, Table 8 reports precision, recall, and F1-score for each of the five risk levels. The proposed model achieves strong and stable performance across all levels, with F1-scores ranging from 93.93% to 96.73%. In particular, the high-risk level achieves a precision of 97.58%, a recall of 95.65%, and an F1-score of 96.60%, indicating that the model can effectively identify severe cabin safety incidents. The macro-averaged and weighted-averaged F1-scores are 95.75% and 95.99%, respectively, suggesting robust overall performance despite class imbalance.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Level-wise performance for the five risk levels.

https://doi.org/10.1371/journal.pone.0343678.t008

Table 9 presents the confusion matrix (in sample counts) for the five-level risk classification task on the test set. Overall, correct predictions dominate the diagonal entries, showing strong discrimination among the five risk levels. For the high-risk category, 1494 out of 1562 cases are correctly classified, while the remaining errors are relatively small and mainly distributed across the other four levels (10 as moderately high risk, 21 as medium risk, 18 as moderately medium risk, and 19 as low risk). Similarly, the moderately high-risk and medium-risk categories show strong diagonal dominance (837/877 and 1475/1551 correct predictions, respectively).

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Confusion matrix for five-level risk classification.

https://doi.org/10.1371/journal.pone.0343678.t009

4.7 Risk assessment of civil aviation cabin safety

The texts related to risk sources in the test set are input into the constructed CNN–LSTM–attention model, and the predicted corresponding risk levels are shown in Table 10.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Risk level classification table.

https://doi.org/10.1371/journal.pone.0343678.t010

As shown in Table 10, the main aspects of high-risk accidents focus on the following key indicators: turbulence, smoke, electrical faults, burning smells, use of oxygen masks, head impact, QRH (calling of the quick reference handbook), medical emergency, abnormal engine, and related to flight attendant operations. All these factors pose direct and serious risks to aviation cabins and flight safety. For example, turbulence may cause personal injuries; smoke or smells may indicate fire risks; and engine or electrical faults may even have catastrophic consequences.

In contrast, medium- to high-risk factors such as gas leakage are slightly less harmful but still pose potential risks to health; medium-risk factors such as unusual smells, abnormal passenger behavior, and seat belt issues reflect more potential hidden dangers in operational management, whereas low-risk factors such as galley cart-related or policy implementation do not pose immediate risks but may increase risks in specific situations. Hierarchical risk distribution characteristics range from main system failures to operational management omissions.