An Analysis of Hybrid Management Strategies for Addressing Passenger Injuries and Equipment Failures in the Taipei Metro System: Enhancing Operational Quality and Resilience

Abstract

This study is the first to systematically integrate supervised machine learning (decision tree) and association rule mining techniques to analyze accident data from the Taipei Metro system, conducting a large-scale data-driven investigation into both passenger injury and train malfunction events. The research demonstrates strong novelty and practical contributions. In the passenger injury analysis, a dataset of 3331 cases was examined, from which two highly explanatory rules were extracted: (i) elderly passengers (aged > 61) involved in station incidents are more likely to suffer moderate to severe injuries; and (ii) younger passengers (aged ≤ 61) involved in escalator incidents during off-peak hours are also at higher risk of severe injury. This is the first study to quantitatively reveal the interactive effect of age and time of use on injury severity. In the train malfunction analysis, 1157 incidents with delays exceeding five minutes were analyzed. The study identified high-risk condition combinations—such as those involving rolling stock, power supply, communication, and signaling systems—associated with specific seasons and time periods (e.g., a lift value of 4.0 for power system failures during clear mornings from 06:00–12:00, and 3.27 for communication failures during summer evenings from 18:00–24:00). These findings were further cross-validated with maintenance records to uncover underlying causes, including brake system failures, cable aging, and automatic train operation (ATO) module malfunctions. Targeted preventive maintenance recommendations were proposed. Additionally, the study highlighted existing gaps in the completeness and consistency of maintenance records, recommending improvements in documentation standards and data auditing mechanisms. Overall, this research presents a new paradigm for intelligent metro system maintenance and safety prediction, offering substantial potential for broader adoption and practical application.

1. Introduction

In the densely populated greater Taipei area of Taiwan, both students and office workers are constantly engaged in their respective routines. In such a fast-paced urban environment, the development of a rapid, convenient, and safe transportation system is particularly vital. The Taipei rapid transit corporation (TRTC) serves as a key provider of transportation services for residents in this region [1,2]. As Taiwan’s first metropolitan rapid transit system, Taipei Metro is operated by TRTC, while the planning and construction of routes are primarily overseen by the department of rapid transit systems (DORTS) under the Taipei city government.

Since its inauguration in 1996, Taipei Metro has recorded a cumulative ridership exceeding 10 billion passenger trips, with daily ridership frequently surpassing two million. With increasing demand, the metro system has become a critical mode of transportation for residents in greater Taipei. In order to safeguard passenger rights and enhance overall service quality, TRTC has actively adopted big data analytics to improve operational safety and efficiency. For instance, track intrusion incidents—which were relatively frequent in the early stages—have been significantly reduced following the full installation of platform screen doors, and such incidents have become virtually nonexistent in recent years. However, passenger injuries still occasionally occur in non-train areas such as escalators and pedestrian walkways.

On the other hand, based on an analysis of 183 complete incident reports from the medium-capacity system (Brown line), 104 of the equipment failures were attributed to the train units themselves, while the remaining cases were caused by external factors or signal malfunctions. Through data-driven analysis and real-time response mechanisms, Taipei Metro has demonstrated its ability not only to swiftly resolve disruptions but also to effectively prevent future incidents.

As transportation modes continue to diversify, an increasing number of commuters are opting for public transit systems, with metro services becoming the primary choice for many [3,4,5,6]. The 2021 Taroko Express train accident [7], one of the most severe railway disasters in Taiwan’s history, resulted in significant casualties and sparked widespread concern over transportation safety. Whether in conventional railways, high-speed rail, or metro systems, the challenge of high passenger volume places non-negotiable demands on safety. Any accident could lead to severe consequences, undermining public trust and operational performance. Therefore, this study focuses on the Taipei Metro system, with particular emphasis on passenger injury incidents and train failure-related delays. The objective is to analyze potential areas for improvement in equipment and infrastructure, and to propose preventive and responsive strategies that can reduce accident rates and minimize service disruptions, thereby enhancing the overall safety and reliability of metro operations.

In the fields of management and engineering, decision trees and association rule mining are widely adopted analytical methods. The following provides a concise overview of recent research in this field. Decision tree algorithms have been widely applied across diverse engineering domains, ranging from data classification to the development of complex decision support systems, as demonstrated in numerous studies [8,9,10,11]. Decision trees are particularly valued for their intuitive structure and effectiveness in extracting actionable decision-making rules from large datasets. In 2024, Zein et al. addressed the accelerating pace of biomedical data generation and collection, which has led to increasingly distributed datasets across hospitals, research institutions, and other organizations. In response, they proposed PrivaTree, a highly efficient and privacy-preserving collaborative training protocol tailored for decision tree modeling on horizontally partitioned biomedical datasets [8]. Also in 2024, Liu et al. examined parking technologies within smart, interconnected transportation systems. Their work aimed to reduce computational complexity while enhancing adaptability and safety in vertical parking systems. They developed a data-driven trajectory planning algorithm based on decision trees, integrating numerical optimization with machine learning to accurately estimate and achieve optimal parking paths [9]. In 2023, Zhang et al. introduced a novel interpretable model for high-speed rail delay prediction. Their approach combined graph community neural networks with time-series fuzzy decision trees, effectively capturing spatiotemporal dynamics, community structures of train routes, and multiple influencing factors at high-speed rail stations [10]. Similarly, in 2023, Guglielmo et al. explored the growing use of behavior trees in robotics applications. They advanced this method by applying decision tree learning in conjunction with logic factorization techniques borrowed from circuit design. This approach improved the precision of behavioral modeling and supported further development in robotic systems [11].

Association rules: Studies [12,13,14] highlight the application of association rule mining techniques to uncover and interpret hidden patterns and correlations among frequently co-occurring items in large datasets. These rules have proven particularly valuable in domains such as market basket analysis and risk management. In 2024, Siswanto et al. introduced a novel algorithm known as set difference frequent pattern (SDFP-growth), which eliminates the need for manually predefining minimum support thresholds. The SDFP-growth algorithm performs dimensionality reduction on the original dataset and generates smaller first- and second-layer data subsets. Through this process, it can automatically produce frequent itemsets that satisfy optimal minimum support thresholds, thereby enhancing the accuracy and efficiency of data analysis [12]. In 2022, Darwish et al. addressed the issue of data cleansing in medical datasets by proposing a privacy-preserving mechanism based on metaheuristic optimization. Specifically, they applied a hybrid approach that combines tabu search with genetic algorithms to select appropriate itemsets for removal from transactions that meet sensitive negative conditions. The primary objective was to minimize modifications to the original database while supporting the integration of privacy preservation with data mining applications [13]. Also in 2022, Yu proposed a quantum algorithm for association rule mining, featuring high-speed search capabilities. The algorithm is designed to efficiently locate and extract frequent itemsets within transactional databases on medium-scale quantum devices or general-purpose fault-tolerant quantum computers. This work demonstrates the potential advantages of quantum computing in advancing the field of data mining [14].

In addition, this study compares the adopted decision tree algorithm with other commonly used classification algorithms, including random forest and gradient boosting [15,16], as shown in

Table 1. Comparison of decision tree, random forest, and gradient boosting algorithms.

Feature/Model	Decision Tree	Random Forest	Gradient Boosting
Basic Principle	Uses a single decision tree for classification or regression	Constructs multiple trees and aggregates their outputs via averaging or voting	Builds trees sequentially, correcting errors of previous models
Model Structure	Single model with high interpretability	Ensemble of multiple models, reducing overfitting	Sequential ensemble model, continuously optimizing residuals
Interpretability	High: easy to visualize and understand	Low: complex structure, difficult to interpret	Very low: difficult to trace individual decision paths
Training Time	Fast	Longer	Longest, especially with large datasets

. Given the emphasis on interpretability and practical applicability in risk prediction, the decision tree was selected as the primary classification model. It offers clear logic, ease of communication, and effective visualization, making it well-suited for real-world implementation.

Building upon the aforementioned literature, decision tree and association rule techniques have been widely adopted in the fields of management and engineering, demonstrating high efficiency and practicality in complex decision analysis and pattern mining tasks. This study focuses on the Taipei Metro system, where daily ridership continues to increase, underscoring its critical role in urban transportation. However, the occurrence of system failures and operational anomalies remains a challenge to service stability. To explore the underlying factors and emerging trends of such incidents, this study employs an integrated data mining approach using decision trees and association rules. The objective is to uncover hidden relationships between the causes of incidents and their associated conditions, thereby supporting accident prevention and enhancing response effectiveness. Ultimately, the aim is to reduce operational risks and minimize potential losses for the metro system.

The primary objectives and expected contributions of this research are as follows:

(i)

To analyze the probability of incident occurrence under specific conditions and identify influential factors.

(ii)

To propose actionable recommendations that assist the Taipei Metro corporation in strengthening its preventive mechanisms for passenger injury events.

(iii)

To provide predictive maintenance and early-warning strategies for train malfunctions, thereby improving overall system reliability and safety.

2. Introduction to Decision Trees and Association Rules

This study observes that the daily ridership of the Taipei Metro corporation has continued to increase, reflecting the growing demand for public transportation services. However, operational failures and anomalies still occur frequently within the metro system. To investigate the potential causes behind these incidents more thoroughly, in addition to traditional statistical charts for trend analysis, this study incorporates big data analytics to further examine the factors that may disrupt normal operations.

By applying data mining techniques such as decision trees and association rules, the research enables a rapid response and swift resolution of service disruptions when accidents occur. These methods are intended to reduce train delays, minimize operational losses, and ultimately enhance overall efficiency and service quality in metro system operations.

2.1. Introduction to Decision Trees

Decision trees are a supervised learning-based data mining technique that serve both feature extraction and classification purposes. The fundamental principle involves selecting the most discriminative input variables based on a target variable, using these to recursively split the dataset, and representing the resulting classification rules and hierarchical logic in a tree-like structure. Once trained and pruned, the decision tree model can be used not only for data exploration but also for predictive analysis, thereby revealing the layered relationships between the target variable and its associated predictors.

A decision tree operates by repeatedly partitioning the dataset into two subsets based on selected features, with the goal of maximizing homogeneity within subsets and enhancing differentiation between them. At each node, the algorithm chooses the most representative feature as the splitting criterion to improve the purity of the resulting subsets. Commonly used feature selection metrics include Information Gain and the Gini index [17,18,19], which assess both the importance and classification effectiveness of a feature.

In the structure of a decision tree, leaf nodes represent terminal nodes at the bottom level of the tree—nodes that are no longer split. In classification tasks, a leaf node corresponds to a specific class label, indicating that all instances falling into that node belong to the same category. In regression tasks, it corresponds to a predicted numerical value, which serves as the estimated output for all instances assigned to that node. Therefore, leaf nodes represent the final output predictions of the decision tree.

In this study, the decision tree method is adopted for data analysis and pattern construction. The process begins by feeding the training data into the model. If the stopping conditions are not yet met, the algorithm continues to search for the optimal splitting feature and generate new nodes. This process repeats until no further splits are possible, resulting in a fully grown decision tree. Subsequently, the testing data is input into the trained decision tree model to evaluate its classification accuracy and predictive performance. The overall construction and analysis procedure is illustrated in Figure 1.

A.

Introduction to pruning strategies

The primary objective of pruning is to prevent overfitting in decision tree models. During the training process, a decision tree may grow excessively complex by continuously splitting nodes in an effort to perfectly classify the training data. This often leads to an overfitted model that performs well on training data but lacks generalization capability for unseen data.

Pruning strategies for decision trees are generally categorized into two main types: pre-pruning and post-pruning [20].

(A)

Pre-pruning:

Pre-pruning involves introducing stopping criteria during the construction of the decision tree. When certain thresholds are met, further splitting is halted to restrict excessive tree growth. This method is typically applied during the training phase and requires tuning of hyperparameters to optimize model performance. Common pre-pruning conditions include the following:

(i)

The number of samples at a node is below a predefined threshold.

(ii)

The tree has reached a maximum specified depth.

(iii)

The difference in sample distribution between parent and child nodes is statistically insignificant.

While pre-pruning can effectively reduce the risk of overfitting, improperly chosen constraints may result in underfitting. Therefore, techniques such as cross-validation are essential for tuning hyperparameters and selecting the optimal model configuration to achieve the best predictive performance.

(B)

Post-pruning:

The key distinction between post-pruning and pre-pruning lies in their timing. Post-pruning is applied after the entire decision tree has been fully grown. It simplifies the model structure to improve its generalization on unseen data, particularly in cases of overfitting. The typical procedure involves removing certain leaf nodes and converting their parent nodes into new leaf nodes. This simplification helps streamline decision logic and avoids excessive fitting to the training data.

Post-pruning typically involves the following steps:

(i)

Select a target leaf node for pruning to generate a simplified decision tree structure.

(ii)

Evaluate the impact of the pruning on model performance using methods such as cross-validation.

(iii)

If the model’s generalization ability improves after the node is removed, retain the pruning operation and designate the parent node as a new leaf node.

Compared to pre-pruning, the advantage of post-pruning lies in its ability to preserve as much information from the training data as possible. It avoids premature termination of splits, which may otherwise degrade model performance. However, this approach is computationally more expensive, as it requires evaluating each node individually. Therefore, post-pruning is more suitable for analytical scenarios where high model accuracy is essential.

(C)

Minimum cost complexity pruning

Minimum cost complexity pruning is a widely used technique aimed at preventing overfitting in decision trees and enhancing their generalization capability on unseen data. This method utilizes a cost complexity function to jointly assess the model’s prediction error and structural complexity, thereby identifying optimal pruning points to simplify the tree.

The core objective is to minimize the structural complexity of the model while maintaining a high level of predictive accuracy. A common form of the cost function is expressed as follows:

$$R_{\alpha }\left(T\right)=R\left(T\right)+\alpha \left|\tilde{T}\right|$$

(1)

Here, R(T) represents the total prediction error of the decision tree T, $\left|\tilde{T}\right|$ denotes the number of leaf nodes in the tree, and α is the regularization parameter, which balances the trade-off between prediction error and model complexity.

When α is small, the algorithm tends to retain a more complex model to improve accuracy; conversely, a larger α favors a simpler tree structure to prevent overfitting.

By performing iterative testing and cross-validation with various values of α, a series of candidate pruning points can be identified. The pruning result that minimizes the cost complexity function R_α(T) is then selected as the final model structure.

(D)

Data balancing techniques

Oversampling is a commonly used data balancing technique in machine learning, primarily employed to address the issue of class imbalance. When one class in a dataset is significantly underrepresented compared to others, machine learning models tend to favor the majority class during training, resulting in reduced prediction accuracy for the minority class. To mitigate this issue, oversampling methods increase the number of samples in the minority class, thereby balancing the class distribution and enhancing the model’s ability to recognize underrepresented patterns.

Upsampling refers to the process of increasing the number of samples in the minority class [21], making it more comparable to the majority class in terms of sample size. This approach strengthens the learning of features specific to the minority class and improves prediction performance. Upsampling methods may involve either duplicating existing samples or generating synthetic samples using techniques such as synthetic minority oversampling technique (SMOTE).

In contrast, downsampling involves reducing the number of samples in the majority class [21] to achieve a relative balance with the minority class. Common approaches include random deletion of majority class samples or the selection of representative subsets based on specific criteria, such as distance metrics or information entropy. The goal of downsampling is to eliminate redundant information from the majority class, reduce computational burden during training, and enhance overall model efficiency and generalization performance.

B.

Decision tree classifier

The Decision tree classifier is Scikit-learn’s implementation of the decision tree algorithm tailored for classification tasks (see Appendix A) [22]. Its core principle involves recursively partitioning the dataset based on feature values to construct a hierarchical tree structure for classification. Users can configure various model parameters—such as maximum tree depth (max_depth), splitting criteria (e.g., Gini index or information gain), and minimum number of samples per node—to optimize the classifier’s accuracy and generalization performance.

In addition, the decision tree classifier supports visualization capabilities, enabling users to graphically display the full tree structure and node-splitting logic. This enhances model interpretability and facilitates result explanation. The classifier is capable of handling datasets containing both numerical and categorical features and can be applied to both multiclass and binary classification tasks. Its strengths lie in its ability to identify the most discriminative features from the data and to provide fast, interpretable predictions for new instances.

C.

Gini impurity

Gini impurity is one of the commonly used splitting criteria in decision tree algorithms, designed to measure the degree of impurity within a node’s data samples [22]. It evaluates how mixed the classes are within a given node, helping to determine whether the node should be further split.

Assuming a dataset D contains K distinct classes, and let p_j represent the probability that a randomly selected sample belongs to class j, the Gini impurity is defined as follows:

$$Gini\left(D\right)=1-\sum _{j=1}^k{p}_j^2$$

(2)

When a dataset D is split into two subsets D₁ and D₂ based on an attribute A, the weighted Gini impurity after the split can be calculated as

$${Gini}_A\left(D\right)={\displaystyle \frac{\left|D_1\right|}{\left|D\right|}}·Gini\left(D_1\right)+{\displaystyle \frac{\left|D_2\right|}{\left|D\right|}}·Gini\left(D_2\right)$$

(3)

The reduction in Gini impurity, representing the improvement in classification purity due to the split by attribute A, is defined as

$$△Gini\left(A\right)=Gini\left(D\right)-{Gini}_A\left(D\right)$$

(4)

A larger value of ΔGini(A) indicates that attribute A yields a more effective split at the node and is therefore a strong candidate for node division. The Gini impurity value ranges between 0 and 1. A value close to 0 implies high purity, meaning that most samples in the node belong to the same class. Conversely, a value near 1 indicates a more heterogeneous distribution, suggesting lower purity.

Compared to entropy, which is another commonly used splitting criterion, Gini impurity is computationally more efficient since it does not involve logarithmic calculations. Additionally, Gini impurity is often more sensitive to changes in the distribution of minority classes, making it particularly advantageous when handling imbalanced classification problems.

D.

Model evaluation metrics

To assess the performance and predictive capability of the classification model, this study adopts several commonly used evaluation metrics, as described below.

(A)

Accuracy:

Accuracy is a fundamental metric used to evaluate the proportion of correctly predicted samples within the test dataset. It reflects the overall correctness of the model’s predictions. The accuracy is calculated as the number of correctly predicted samples divided by the total number of test samples, then multiplied by 100 to express the result as a percentage:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{N}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{c}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t} \mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}}\times 100$$

(5)

(B)

Feature importance:

Feature importance is a metric used to assess the contribution of each input variable to the model’s prediction results. By analyzing feature importance, one can identify which variables have the most significant impact on the model’s classification decisions. A higher value indicates greater influence, making this metric useful for feature selection and model simplification in subsequent analyses.

(C)

Confusion matrix:

A confusion matrix is used to display the correspondence between the predicted outcomes and the actual labels of a classification model. It is a key tool for evaluating model performance across different classes [23,24,25]. In the matrix, rows represent actual classes while columns represent predicted classes. The matrix clearly illustrates classification results such as True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). The confusion matrix used in this study is shown in

Table 2. Confusion matrix.

		Actual Positive	Actual Negative
Predicted positive		TP (True positive)	FP (False positive, Type I error)
Predicted negative		FN (False negative, Type II error)	TN (True negative)

.

The four components of the confusion matrix are defined as follows:

(i)

True positive (TP): The number of samples that are actually positive and correctly predicted as positive by the model.

(ii)

True negative (TN): The number of samples that are actually negative and correctly predicted as negative by the model.

(iii)

False positive (FP): The number of samples that are actually negative but incorrectly predicted as positive by the model. This is also referred to as a Type I error.

(iv)

False negative (FN): The number of samples that are actually positive but incorrectly predicted as negative by the model. This is also known as a Type II error.

Additional commonly used classification evaluation metrics include the following:

Precision: Measures the proportion of correctly predicted positive samples among all samples predicted as positive. It is calculated as

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{TP}{TP+FP}}$$

(6)

Recall: Measures the proportion of actual positive samples that are correctly identified by the model. It is calculated as

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{TP}{TP+FN}}$$

(7)

2.2. Association Rules

Data mining refers to a technique that extracts meaningful patterns and rules from large datasets through automated or semi-automated processes. Among the various data mining methods, association rule analysis is one of the most common and widely applied approaches [26,27,28]. This method is particularly effective in uncovering potential relationships between events and their causes, revealing hidden correlations among variables, and providing actionable insights for decision support.

In this study, association rule mining is applied to analyze metro system incident records, aiming to identify potential causal relationships among various accident factors. The derived rules can serve as a reference for maintenance personnel, helping to reduce diagnosis time, improve repair efficiency and system quality, and ultimately achieve fault prevention and risk control.

The core concept of association rule analysis lies in discovering dependencies among attributes (fields) within large datasets. The evaluation of these rules is primarily based on three key indicators: Support, confidence, and lift. These metrics are essential for identifying statistically and practically significant rules. As illustrated in Figure 2, their definitions are as follows:

A.

Support:

Support measures the probability that both item X and item Y occur together. It is defined as

Support(X ⇒ Y) = P(X ∩ Y)

(8)

This represents the proportion of records where X and Y appear simultaneously out of the total dataset. A minimum support threshold is typically set to filter out infrequent and non-representative rules.

B.

Confidence:

Confidence indicates the conditional probability that Y occurs given that X has occurred. It is defined as

$$\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{f}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}(\mathrm{X}\Rightarrow \mathrm{Y})={\displaystyle \frac{\mathrm{P}(\mathrm{X}\cap \mathrm{Y})}{\mathrm{P}\left(\mathrm{X}\right)}}$$

(9)

C.

Lift:

Lift evaluates whether the occurrence of X significantly increases the likelihood of Y, compared to the baseline probability of Y occurring alone. It is defined as

$$\mathrm{L}\mathrm{i}\mathrm{f}\mathrm{t}(\mathrm{X}\Rightarrow \mathrm{Y})={\displaystyle \frac{\mathrm{P}\left(\mathrm{X}\right|\mathrm{Y})}{\mathrm{P}\left(\mathrm{Y}\right)}}$$

(10)

When Lift > 1, it suggests that the occurrence of X contributes positively to the occurrence of Y, indicating a significant and potentially meaningful association. Lift is especially useful for distinguishing causal relationships from mere statistical correlations.

Decision-makers can apply threshold values—such as minimum support, minimum confidence, and minimum lift—to filter and rank the discovered rules. This enables the selection of the most interpretable and practically valuable association rules, which can be used for operational decision-making or the design of early-warning systems.

2.3. Applications of Association Rules

Research on association rules generally focuses on two main objectives: (i) developing efficient algorithms for association rule mining, and (ii) accurately and rapidly extracting meaningful relationships from large-scale databases. In this study, association rule techniques are applied to the operation and maintenance management of the Taipei Metro corporation. The applications are detailed as follows:

A.

Electromechanical systems

Electromechanical systems include power supply, air conditioning, fire protection, elevators, and utility infrastructure. Power outages directly affect the functionality of all downstream equipment; air conditioning failures compromise passenger comfort; fire protection systems are critical to public safety and passenger life protection; elevator and utility malfunctions affect vertical mobility and basic services within stations. These systems are highly interdependent—failures in any component can impact metro operations and passenger safety.

B.

Rolling stock systems

Rolling stock forms the core of metro transportation and relies on the coordinated functioning of electrical, signaling, communications, civil, and electromechanical systems. The vehicle system is closely linked to track switches, third-rail power supply, and roadside equipment. Malfunctions in any of these subsystems may cause train service disruptions.

C

Operation control center

The operation control center (OCC) functions as the “central nervous system” of the metro system, responsible for coordinating subsystems such as rolling stock, electromechanical, civil, and electrical systems. Upon issuing commands, all subsystems must respond immediately. Any execution error may lead to serious operational incidents. Therefore, OCC staff must undergo psychological assessments and field-based training across all subsystems to ensure real-time decision-making and crisis response capabilities.

D.

Civil engineering systems

This system encompasses structural works, track infrastructure, and surrounding facilities, all of which are closely integrated with power, signaling, communication, and rolling stock systems. Design factors such as station layout, stair and ramp configuration, and track stability may directly or indirectly affect vehicle stability and passenger safety.

E.

Station operations

The responsiveness of station staff is crucial for both passenger service and emergency handling. In the event of equipment failure or a medical emergency, personnel must follow standard operating procedures (SOPs). For instance, contacting emergency services (119), assisting visually impaired passengers, or addressing abnormal behavior within train cars all require immediate coordination with the OCC to ensure rapid and risk-minimized responses.

F.

Electrical systems and integrated applications

Association rule analysis can be used to identify optimal interactions between the electromechanical system and other subsystems, thereby enhancing overall transportation safety. For example, when the fire alarm system is triggered, coordinated responses involving station staff, train operators, and the OCC must be executed promptly. In critical situations, operational strategies such as non-stopping trains or shuttle bus dispatch may be implemented to facilitate evacuation and alternative transport, thus reducing risks and mitigating potential impacts.

3. Hybrid Management Strategy Analysis

3.1. Analysis of Passenger Injury Patterns Using Decision Trees

As part of the research methodology, this study employs supervised learning—specifically, the decision tree algorithm from big data analytics—as a tool for further exploration of the problem. The decision tree model is implemented using Python 3.12 (see Appendix A). The objective is to apply decision tree analysis to classify and interpret patterns in Taipei Metro passenger injury incidents, with variables such as time period, age, gender, location of occurrence defined as feature variables (X), and injury severity as the target variable (Y).

To enhance the analyzability of the dataset, a data preprocessing stage was conducted. Time periods were categorized into peak, off-peak, and weekend classes; age was treated as a continuous variable; locations were classified into nine contextual categories (e.g., platform, train car, ground level); injury severity was divided into minor, moderate, and severe levels; and gender was treated as a binary variable. This preprocessing strategy facilitates the construction and interpretation of the decision tree model, as summarized in

Table 3. Feature variables (X) for passenger injury data in the metro system.

Feature Variables (X)	Category	Description
Period of time	Peak	Weekdays 07:00–09:00 and 17:00–19:30
	Off-peak	Non-peak weekday hours
	Weekend	Saturdays and Sundays
Age	–	Continuous variable
Location	Platform	Includes platform, platform doors, and track area
	Train	Inside the train car
	Floor	Station concourse or passageway floor
	Toilet	Restroom areas
	Escalator	Escalator zones
	Stairs	Stairways
	Elevator	Elevators
	Out-of-station	Areas outside the station premises
	Other	Other unspecified areas
Type of Incident	Pinch	Including pinching or being caught in a gap
	Sprain	Joint sprain or strain
	Insect bite	Bites from insects or mosquitoes
	Crowding	Includes pushing, collision, or compression
	Crush	Being struck by a falling object
	Attacked	Includes physical assault, biting, or stabbing
	Fall	Slips or falls
	Electric shock	Contact with electrical sources
	Step on	Stepping-related injuries
	Scrape	Includes abrasions, cuts, or bruises
	Discomfort	Physical discomfort or illness
	Other	Other unspecified causes
	Unknown	Cause cannot be determined
Sex	Male	–
	Female	–

and

Table 4. Target variable (Y) for passenger injury data in the metro system.

Degree of Injury	Description
Mild	Includes abrasions, swelling, or no visible injury
Moderate	Includes fractures or dislocations; also applies to individuals aged 60+ who were transported by ambulance
Severe	Includes major bleeding, unconsciousness, or similar conditions

.

First, the structured feature dataset must be imported into Python. Since machine learning models in Python cannot directly process categorical variables, it is necessary to convert both the target variable and the feature variables into dummy variables. The target variable, degree of injury, possesses a clear ordinal relationship and is thus treated as an ordinal variable. Therefore, this study utilizes the LabelEncoder module to encode the values “mild,” “moderate,” and “severe” into 0, 1, and 2, respectively.

In contrast, the feature variables are nominal (unordered categorical variables). To avoid the model mistakenly assuming an ordinal relationship, one-hot encoding is applied to transform them into dummy variables.

Once data preprocessing is completed, model training can proceed. Before building the model, the injury dataset is divided into a training set and a test set using an 80:20 split ratio. The training data is then fed into the DecisionTreeClassifier model to train and construct the decision tree.

Next, the trained model is used to make predictions on the test set, and the prediction accuracy is calculated. Experimental results indicate that the model achieved an accuracy of approximately 68%, which falls short of the ideal performance level. As a result, the study further analyzes the feature importance scores provided by the decision tree model to evaluate the contribution of each feature to the prediction outcome. This analysis guides feature selection and model optimization to improve overall predictive performance.

As shown in

Table 5. Feature importance of variables in the decision tree model.

Feature Variable	Feature Importance
Age	0.6506
Location_Escalator	0.0335
Sex_Male	0.0330
Location_Platform	0.0262
Time_Off-Peak	0.0259
Location_Stairs	0.0257
Location_Floor	0.0254
Sex_Female	0.0252
Location_Train	0.0244
Incident_Fall	0.0241
Time_Peak	0.0194
Incident_Pinch	0.0156
Time_Weekend	0.0102
Incident_Crowding	0.0100
Location_Toilet	0.0091
Location_Out-of-Station	0.0076
Incident_Scrape	0.0067
Location_Elevator	0.0056
Incident_Sprain	0.0041
Location_Other	0.0040
Incident_Insect_Bite	0.0031
Incident_Crush	0.0029
Incident_Unknown	0.0023
Incident_Other	0.0021
Incident_Attacked	0.0020
Incident_Step_On	0.0011
Incident_Discomfort	0.0000
Incident_Electric_Shock	0.0000

, feature importance is determined by the model based on how frequently and effectively each feature is used for splitting during training. The results reveal that the lower-ranked features are mostly from the variable “type of incident”, with some categories contributing nothing to the model (i.e., feature importance equals zero). For the sake of performance enhancement and model simplification, this study decides to remove the “type of incident” feature, thereby eliminating low-contribution variables that could negatively affect model accuracy.

In addition to simplifying feature variables through feature engineering, this study also identified another potential factor affecting the accuracy of the decision tree model: data imbalance. The injury data categorize the target variable into three levels of severity: (i) mild injury, which includes cases such as abrasions, swelling, or no visible external injury; (ii) moderate injury, which includes fractures, dislocations, or individuals aged 60 and above who were transported to the hospital by ambulance; and (iii) severe injury, which includes serious conditions such as heavy bleeding or loss of consciousness. As shown in

Table 6. Sample distribution by injury severity (Target variable).

Injury Severity Category	Number of Samples
Mild injury	2188
Moderate injury	1058
Severe injury	85

, among the total of 3331 records, 2188 are classified as mild injuries, 1058 as moderate injuries, and only 85 as severe injuries, indicating a clear imbalance in class distribution.

Data imbalance may cause the model to become biased toward predicting the majority class during training, thereby weakening its ability to recognize and classify minority classes. As a result, even if the model achieves acceptable overall accuracy, its predictive performance for underrepresented categories—such as severe injuries—may deteriorate significantly. This compromises the model’s capacity to accurately learn the distinguishing patterns of minority classes, thereby reducing its practical applicability and reliability in real-world deployments.

To address the issue of data imbalance, this study adopts an oversampling technique. Specifically, the method used is duplicate sampling, which involves randomly replicating samples from minority classes to increase their quantity, thereby achieving a more balanced class distribution.

Oversampling not only balances the dataset but also enhances the model’s ability to learn. In situations where class distributions are highly skewed, models tend to favor predictions toward the majority class, resulting in poor predictive performance for minority classes. By increasing the number of samples in the minority class, oversampling not only balances the dataset but also enhances the model’s ability to learn and accurately predict these underrepresented categories.

As shown in

Table 7. Sample distribution by injury severity after oversampling (Target variable).

Injury Severity Category	Number of Samples
Mild Injury	2188
Moderate injury	2188
Severe injury	2188

, this study duplicates the original samples from the severe injury and moderate injury classes to bring their sample sizes closer to that of the majority (mild injury) class. This increases the model’s sensitivity to minority classes during training and improves their relative contribution to the learning process. Consequently, this helps boost the model’s overall accuracy and stability in classifying minority cases.

After applying the aforementioned feature engineering and oversampling techniques, the dataset was reprocessed through the decision tree model for training and testing. The results show that the model’s accuracy improved to 81%, indicating a significant enhancement compared to the original model.

To further evaluate the overall performance of the decision tree classifier, several commonly used evaluation metrics were calculated and analyzed. First, a confusion matrix was constructed, as shown in

Table 8. Confusion matrix of the decision tree model.

Predicted\Actual	Mild Injury	Moderate Injury	Severe Injury
Mild injury	326	75	36
Moderate injury	62	312	64
Severe injury	0	0	438

, followed by the computation of key classification metrics such as precision and recall. These metrics provide a more comprehensive understanding of the model’s predictive performance across different injury severity levels.

Precision and recall values are summarized in

Table 9. Precision and recall by injury severity category.

Injury Severity Category	Precision	Recall
Mild injury	71%	81%
Moderate injury	75%	84%
Severe injury	100%	81%

. Overall, the model demonstrates strong performance across all categories in terms of both precision and recall, indicating that it possesses stable and reliable predictive capabilities for classifying cases of mild, moderate, and severe injuries.

Following the construction of the decision tree model, visualization was conducted to better understand its structure, as shown in Figure 3. The analysis reveals that the model is highly complex, with a depth of 20 layers and a total of 616 nodes. Such complexity not only makes the model difficult to interpret and explain but also hinders the extraction of clear, actionable rules.

To enhance the interpretability and conciseness of the model, this study further applied post-pruning to simplify the decision tree structure. Based on the pruned model, a set of representative and practically applicable classification rules was derived.

This study adopts minimal cost complexity pruning as the pruning method. This approach regulates the extent of pruning and structural simplification by adjusting the cost complexity hyperparameter α (alpha). After multiple rounds of testing and comparison, the optimal value of α was determined to be 0.05. This setting effectively simplifies the model structure, resulting in a more compact decision tree that facilitates subsequent rule extraction and analysis. Moreover, it enhances the interpretability and practical applicability of the model.

As shown in Figure 4, the structure of the decision tree has been significantly simplified following the pruning process, enabling clearer extraction of classification rules. This study further organizes the key classification rules and, based on their content, proposes preliminary recommendations for injury prevention strategies in the Taipei Metro system. Using a dataset of approximately 3331 passenger injury records, the model was trained to uncover potential relationships between various feature variables and the degree of injury, as well as to identify distribution patterns and potential risk factors associated with passenger injuries.

A.

Key extracted classification rules:

(A)

IF age ≤ 61 AND time period = off-peak AND location = escalator, then the injury is classified as severe.

(B)

IF age > 61 AND sex = male, then the injury is classified as severe.

(C)

IF age > 61 AND sex = female, then the injury severity is classified as moderate.

B.

Interpretation and practical implications:

(A)

Elderly male passengers (over 61 years old) show a higher likelihood of severe injuries. Enhanced safety education and targeted protection measures should be implemented for this demographic.

(B)

While elderly females exhibit slightly lower injury severity compared to males, the injuries are still predominantly moderate, suggesting that this group also represents a potential high-risk population.

(C)

Younger or middle-aged passengers using escalators during off-peak hours may also experience severe injuries. This underscores the importance of evaluating escalator safety and enhancing monitoring and maintenance measures during such time periods.

C.

Recommendations for improvement

(A)

Based on the rules extracted in this study, it was found that passengers aged over 61, regardless of gender, have a relatively higher likelihood of experiencing moderate or severe injuries in the metro system. It is recommended that metro operators enhance care and support measures for elderly passengers. These may include deploying additional staff during peak hours to proactively assist older adults with boarding and alighting, and offering guidance and physical support services. Furthermore, designated waiting zones or priority boarding areas for elderly passengers can be established, along with the installation of more seating to improve both safety and comfort during travel.

(B)

For passengers in younger age groups—including children, adolescents, adults, and middle-aged individuals—the risk of severe injury is also elevated when incidents occur in escalator areas during off-peak hours. This study speculates that individuals in these demographics often use electronic devices (e.g., smartphones, tablets, earphones) while commuting, which may lead to reduced attention and increase injury risk. In contrast, elderly passengers are less likely to use such devices and are typically more focused on their surroundings while moving.

Based on this inference, it is recommended that metro authorities strengthen escalator safety management, including the following actions:

(i)

Conduct regular inspections and maintenance of escalator equipment to ensure proper functioning and compliance with safety standards.

(ii)

Install clearly visible warning signs and pictograms reminding passengers to stay attentive and avoid distractions while using escalators.

(iii)

Consider installing smart monitoring systems in designated escalator zones to detect abnormal behaviors and provide real-time alerts.

3.2. Application of Association Rule Mining to Analyze EMU Fault Patterns

The research methodology is divided into two parts: (i) the medium-capacity system (Brown line), and (ii) the high-capacity system (including the Blue, Red, and Green lines).

For each system, the data was first cleaned and preprocessed, followed by the clear definition of target attributes and input attributes.

All fault events were initially categorized by their corresponding main subsystem, resulting in six major classes:

(i)

Power supply/electrical system faults.

(ii)

Signaling faults.

(iii)

Track faults.

(iv)

External factors.

(v)

Communication faults.

(vi)

EMU (Electric Multiple Unit) faults.

Among these, EMU fault was defined as the target attribute for prediction in this study, while season, time of day, and weather conditions were selected as input attributes.

The variable settings are as follows:

(i)

Season was categorized into four groups according to Taiwan’s climatic patterns:

Spring (March–May), Summer (June–August), Autumn (September–November), and Winter (December–February).

(ii)

Time of day was divided into four periods based on metro operating hours:

Early morning (00:00–06:00), morning (06:00–12:00), afternoon (12:00–18:00), and evening (18:00–24:00).

(iii)

Weather was classified into three categories based on data source:

Sunny, cloudy, and rainy.

This study aims to conduct data exploration and analysis based on the above classification framework, with the goal of uncovering potential associations between input attributes and the target attribute. The ultimate objective is to generate preventive recommendations that support fault prediction and risk management.

In contrast to the earlier section, which focused on classification analysis using decision trees, this part of the study emphasizes identifying the key associative factors underlying fault occurrences. Therefore, the association rule mining technique—widely used in big data analytics—is adopted to extract meaningful patterns from historical datasets. The derived rules are intended to provide practically applicable insights that can inform operational management and system improvement strategies for the Taipei Metro corporation.

The following sections present the results for the two subsystems, beginning with the analysis of the medium-capacity system (Brown line).

A.

Analysis of the medium-capacity system (Brown line)

In this study, a total of 183 valid records were compiled for the medium-capacity system (Brown line). Statistical analysis and association rule mining were conducted based on the defined input attributes. The relevant statistical results are presented in

Table 10. Statistical summary of input attributes for the medium-capacity system (Brown line).

Attribute	Category	Frequency
Season	Spring (Mar.–May)	35
	Summer (Jun.–Aug.)	71
	Autumn (Sep.–Nov.)	42
	Winter (Dec.–Feb.)	35
	Subtotal	183
Time of day	00:00–06:00	7
	06:00–12:00	77
	12:00–18:00	60
	18:00–24:00	39
Weather	Sunny	113
	Cloudy	19
	Rainy	51
Main system type	Power system fault	8
	Signal fault	12
	Track fault	11
	External factors	39
	Communication fault	9
	EMU fault	104

.

In the final stage of analysis, the minimum support threshold was set at 1.5%, the minimum confidence threshold at 10%, and only association rules with a lift greater than 1 were retained for further analysis. These three criteria were applied simultaneously to ensure the reliability and practical relevance of the extracted rules.

To enhance the readability and interpretability of the results, a color-coding scheme was applied based on the range of lift values:

(i)

Rules with a lift between 1.0 and 1.5 are highlighted in light blue.

(ii)

Those between 1.5 and 2.0 are in blue.

(iii)

Rules with a lift greater than 2.0 in dark blue.

After filtering, a total of 24 association rules met the specified conditions. The detailed results are presented in

Table 11. Association rule results for the medium-capacity system (Brown line). (Target attribute: six fault categories).

Attribute Combination	Power System (1)	Signal Fault (2)	Track Fault (3)	External Factor (4)	Communication Fault (5)	EMU Fault (6)
Sunny and 06:00–12:00	1.10	1.53				1.09
Sunny and 12:00–18:00				1.18	1.33
Sunny and 18:00–24:00					3.27
Rainy and 06:00–12:00	2.57
Rainy and 12:00–18:00						1.36
Rainy and 18:00–24:00				1.38
Sunny and Summer					1.54	1.11
Sunny and Autumn						1.18
Sunny and Winter	4.00			1.21
Rainy and Summer				2.73
Rainy and Autumn				2.48
Summer and 06:00–12:00						1.33
Summer and 12:00–18:00				2.05
Summer and 18:00–24:00					2.89
Autumn and 06:00–12:00						1.18
Autumn and 18:00–24:00				3.12
Winter and 06:00–12:00	3.60
Sunny and 06:00–12:00 and Summer						1.42
Sunny and 12:00–18:00 and Summer				1.93

.

This study applied association rule mining to investigate and summarize the causes of system faults in the medium-capacity system (Brown line). For each fault category, the corresponding association rules were computed and analyzed using three core indicators: Support, confidence, and lift.

After obtaining the preliminary analytical results, the study further cross-validated the findings by referencing actual maintenance records, relevant literature, and recent incident reports. Through the conditional combinations revealed by association rules, the potential correlations and contextual backgrounds were explored in depth.

Based on the results, a set of multi-faceted recommendations was proposed to assist the metro authority in optimizing maintenance routines, equipment inspections, and preventive mechanisms. Due to the large volume of data processed, only the most highly correlated and explanatorily significant rules are presented here.

Table 12. Association rule mining results for the medium-capacity system (Brown line): Support, confidence, and lift.

System Type	Special Attribute Combination	Support	Confidence	Lift
Power system	Sunny and 06:00–12:00	0.017	0.300	1.102
	Rainy and 06:00–12:00	0.017	0.300	2.571
	Sunny and Winter	0.022	0.400	4.000
	Winter and 06:00–12:00	0.017	0.300	3.600
Signal fault	Sunny and 06:00–12:00	0.028	0.417	1.531
External factors	Autumn and 18:00–24:00	0.022	0.121	3.116
Communication fault	Sunny and 12:00–18:00	0.017	0.272	1.327
	Sunny and 18:00–24:00	0.222	0.364	3.272
	Sunny and Summer	0.028	0.455	1.543
	Summer and 18:00–24:00	0.167	0.272	2.887
EMU fault	Sunny and 06:00–12:00	0.172	0.298	1.094
	Rainy and 12:00–18:00	0.061	0.106	1.359
	Sunny and Summer	0.189	0.327	1.111
	Sunny and Autumn	0.072	0.125	1.184
	Summer and 06:00–12:00	0.111	0.192	1.331
	Autumn and 06:00–12:00	0.083	0.144	1.183
	Sunny and 06:00–12:00 and Summer	0.100	0.173	1.416

provides a summary of the qualified association rules identified for the Brown Line system. Each conditional combination is listed alongside its respective values for support, confidence, and lift, serving as a practical reference for future operational planning and decision-making.

Recommendations for improvement of the medium-capacity system (Brown line).

(i)

Power system

This study identified that although the types of power system faults are diverse, failures tend to cluster between 06:00 and 12:00 on both sunny and rainy days. These time-weather combinations meet the minimum support (1.5%) and confidence (10%) thresholds, with particularly high lift values—3.6 and 4.0, respectively—under winter and sunny morning conditions. This suggests that failures in the power system are more likely to occur during the early morning start-up phase, potentially due to reduced system stability during cold starts.

Therefore, it is recommended that the Taipei Metro strengthens inspections and maintenance of power systems prior to morning train dispatch. Maintenance records should be thoroughly analyzed, with a focus on components that require frequent replacement, to appropriately shorten inspection intervals. Specifically, the following two measures are proposed for the propulsion power substation (PPSS):

Rapid fault response: When an abnormality is detected in the PPSS, a fault-handling procedure should be initiated immediately, aiming to restore service quickly, with backup systems in place to maintain short-term transport capacity.

Institutionalized regular inspections: Regularly assess the condition of PPSS components, replacing outdated or fragile parts in advance to improve overall system stability, reduce maintenance costs, and minimize risk of service interruptions.

(ii)

Signal faults

Signal faults in the medium-capacity system most frequently occur under sunny conditions between 06:00 and 12:00. As this system operates autonomously without drivers, many maintenance reports have linked these faults to limited system stability. It is hypothesized that such issues may result from initialization errors during early morning start-up or decreased operator attentiveness during the early hours. To mitigate these problems, it is recommended that the Taipei Metro (i) enhances the start-up stability of the autonomous system and (ii) provides targeted training and reinforces operational protocols for early-shift personnel.

(iii)

External factors

Data indicate a strong association between external factors and the 18:00–24:00 period in autumn. Further investigation revealed that many incidents were caused by passengers accidentally activating emergency handles inside train cars—categorized as uncontrollable events (only one was deemed controllable). Given that the Brown line operates on an unmanned system, this study proposes increasing the frequency of track inspections—especially during high-risk periods—and installing real-time video surveillance in vulnerable areas, and redesigning emergency handles with a multi-step triggering mechanism to reduce the risk of accidental activation and enhance the fault-tolerance of the system.

(iv)

Communication failures

Communication failures were primarily observed between 18:00 and 24:00 on sunny days during the summer, with significantly high lift values. Analysis reveals that switch malfunctions accounted for approximately 45% of these cases, likely due to high summer temperatures and peak evening passenger loads. It is recommended that the Taipei Metro evaluates the thermal tolerance of communication switches and upgrades components, or considers alternative suppliers if necessary. Additionally, maintenance frequency of the communication system should be increased before and during summer peak periods to facilitate preventive maintenance and reduce the likelihood of failures.

(v)

Electric multiple unit (EMU) failures

EMU-related failures represent the most prevalent fault type in the medium-capacity system. This study analyzed seven representative conditional combinations associated with such failures (as detailed in Table 12): Sunny and 06:00–12:00, Rainy and 12:00–18:00, Sunny and Summer, Sunny and Autumn, Summer and 06:00–12:00, Autumn and 06:00–12:00, and Sunny and 06:00–12:00 and Summer.

Four key observations and recommendations were derived, as follows:

• Sunny and 06:00–12:00 and Summer (with the highest lift value at 1.416)

The primary causes were identified as cable degradation and defective solder joints, jointly responsible for over 50% of incidents. These issues are likely caused by heat-induced insulation degradation and terminal desoldering. It is advised that preventive maintenance of electrical systems be strengthened during summer, particularly through prioritized inspections prior to morning dispatch.

• Rainy and 12:00–18:00

Data show that 54.54% of faults during this time were brake system-related. The wet environment in rainy weather likely increased braking system load, causing malfunctions to emerge in the early afternoon. Enhanced inspection and testing protocols for brake systems during rainy periods are recommended to ensure stable braking performance.

• Summer and 06:00–12:00

The predominant failure types were obstacle detection errors, signal anomalies, and cable damage, together comprising 75% of reported incidents. This highlights the substantial impact of summer morning weather conditions on EMU stability. Recommendations include upgrading cable insulation materials, improving the heat resistance of signal systems, and shortening pre-inspection intervals for related components.

• Conditions associated with the 06:00–12:00 period (e.g., Sunny, Rainy, Summer, Autumn)

The data indicate that the early morning period is a high-risk window for system faults. Although individual support and confidence metrics may not all meet significance thresholds, this time frame should be regarded as a critical risk period. The Taipei Metro is advised to prioritize system checks before morning departures to enhance the operational stability and safety of EMUs during the early stages of service.

B.

Analysis of the high-capacity system

Due to the incomplete weather records in the high-capacity system dataset—where more than half of the entries lacked specific weather information—this study excluded weather as a factor in the analysis. Instead, the research focused on season and time of day as the primary input attributes. After data filtering, a total of 978 valid records were obtained. The statistical summary based on the selected input attributes is presented in

Table 13. Statistical summary of input attributes for the high-capacity system.

Attribute Category	Classification	Frequency
Season	Spring (Mar.–May)	249
	Summer (Jun.–Aug.)	250
	Autumn (Sep.–Nov.)	250
	Winter (Dec.–Feb.)	229
	Subtotal	978
Time of day	00:00–06:00	48
	06:00–12:00	371
	12:00–18:00	315
	18:00–24:00	237
System category	Power system failure	46
	Signal system failure	120
	Track system failure	65
	External factors	329
	Communication failure	35
	Electric multiple unit (EMU) Failure	383

.

In this study, for the high-capacity system dataset, the minimum support was set at 1.5%, the minimum confidence at 7%, and the lift value was required to be greater than 1. Only association rules that simultaneously satisfied all three criteria were retained as results with substantial analytical value.

To enhance visual identification and facilitate data comparison, color coding was applied based on the range of lift values: rules with lift values between 1 and 1.5 were marked in light blue, those between 1.5 and 2.0 in blue, and those greater than 2.0 in dark blue. After filtering, a total of 23 association rules met the criteria. The summarized results are presented in

Table 14. Association rule statistics for the high-capacity system (Target attribute: Six types of system failures).

Special Attribute Combinations	Power System (1)	Signal Failure (2)	Track Failure (3)	External Factors (4)	Communication Failure (5)	EMU Failure (6)
Blue line and 12:00–18:00				1.54
Red line and 06:00–12:00			1.36			1.08
Red line and 12:00–18:00						1.23
Red line and 18:00–24:00						1.10
Blue line and Autumn				1.37
Red line and Spring		1.12				1.27
Red line and Summer						1.22
Spring and 06:00–12:00		1.57				1.11
Spring and 12:00–18:00						1.26
Spring and 18:00–24:00						1.01
Summer and 06:00–12:00			1.12			1.06
Summer and 12:00–18:00						1.28
Summer and 18:00–24:00			1.59
Autumn and 12:00–18:00			1.39
Winter & 06:00–12:00		1.44				1.02
Winter and 12:00–18:00			1.21
Red line and 12:00–18:00 and Spring						1.46
Red line and 12:00–18:00 and Summer						1.49

.

This study employed association rule mining to analyze and summarize the causes of system failures in high-capacity metro systems. For each failure type, the associated rule calculations were compiled, including three key indicators: support, confidence, and lift.

After obtaining the preliminary analysis results, further validation and categorization were conducted by cross-referencing actual maintenance records, relevant literature, and recent incidents. The rule-based attribute combinations revealed by the analysis were examined to explore their underlying correlations and contextual background.

Based on the findings, this study proposes a series of improvement strategies from multiple perspectives to assist the metro company in optimizing maintenance, equipment inspection, and preventive mechanisms. Due to the large volume of computed data, only a selection of highly relevant and interpretable results is presented.

Table 15. Association rule computation results for the high-capacity system (Support, confidence, and lift values).

System Type	Attribute Combination	Support	Confidence	Lift
Signal system failure	Red line and Spring	0.021	0.175	1.119
	Spring adn 06:00–12:00	0.016	0.133	1.571
	Winter and 06:00–12:00	0.015	0.125	1.438
Track failure	Red line and 06:00–12:00	0.019	0.292	1.361
External factors	Blue line and 12:00–18:00	0.029	0.088	1.539
	Blue line and Autumn	0.025	0.073	1.372
	Summer and 06:00–12:00	0.039	0.116	1.118
	Summer and 18:00–24:00	0.033	0.097	1.585
	Autumn and 12:00–18:00	0.038	0.112	1.392
	Winter and 12:00–18:00	0.033	0.100	1.211
EMU failure	Red line and 12:00–18:00 and Spring	0.033	0.084	1.459
	Red line and 12:00–18:00 and Summer	0.027	0.073	1.489

summarizes the qualifying association rule outcomes for the high-capacity system, detailing the support, confidence, and lift values for each condition combination as a reference for future operational planning and decision-making.

Recommendations for improvement of the high-capacity system.

(i)

Signal system failures

According to the data analysis, signal system failures were predominantly concentrated in the 06:00–12:00 time period during spring and winter. The association rules during these periods met the minimum support and confidence thresholds set by this study, as detailed in Table 15. The analysis revealed multiple causes, including signal anomalies, automatic train protection (ATP) system malfunctions, and poor circuit board connections—suggesting the faults are not attributable to a single factor.

Therefore, it is recommended that the Taipei Metro enhances the overall stability of the signal system, particularly by reinforcing inspection protocols for relevant equipment during spring and winter mornings. Additionally, a regular maintenance and inspection plan should be established, with preventive measures scheduled for high-risk periods to mitigate operational disruptions.

(ii)

Track failures

Track failures showed significant association only in the 06:00–12:00 period on the Red line, with the relevant rule listed in Table 15. Maintenance records indicate that 57% of the incidents were due to misaligned switch points. Other issues included aging components, mechanical malfunctions, or unidentified causes.

It is recommended that the Taipei Metro strengthens functionality tests and maintenance routines for switch points prior to the morning dispatch on the Red line. The inspection frequency should be shortened, and additional personnel should be assigned during high-risk periods for real-time support. It is also advisable to consider the preemptive replacement of problematic equipment or provide technical feedback to suppliers for performance improvement.

(iii)

External factors

Although the conditions associated with external factors met the basic association thresholds, the overall lift values were not significant. Data showed that 65% of controllable events were due to human error, primarily driver mistakes such as failure to properly close the cab door or incorrect station stopping. Unlike the medium-capacity system, where incidents were mostly passenger-induced, human errors in the high-capacity system were mainly caused by drivers. However, due to the absence of driver identification in the dataset, further verification of specific individuals or shifts was not possible.

It is recommended that the Taipei Metro enhances driver training, emphasizing operational protocols and accident simulation drills. A regular performance evaluation system should be implemented. From a hardware perspective, the human–machine interface design should be improved to reduce the likelihood of operational errors.

• (iv) EMU failures

Electric multiple unit (EMU) failures were the most frequent anomaly in the high-capacity system. The Red line during the 12:00–18:00 period in spring and summer showed the highest lift values, as detailed in Table 14. Subsystem analysis of failures revealed that on-board systems accounted for 21.14%, door systems 18.80%, brake systems 16.19%, propulsion systems 8.61%, auxiliary power systems 6.27%, and control systems only 1.83%. Further analysis identified the following main causes:

• Door system: Many failures were triggered by emergency stops due to improperly closed cab doors.
• On-board signaling system: Failures were often linked to automatic train operation (ATO) module instability, including system crashes, loose connections, and logic errors.
• Brake system: The causes varied; one case involved brake pressure failing to release, though maintenance logs did not specify the technical reasons.
• Drawing upon these findings, the study proposes the following improvement measures:
• Door System: It is recommended to install cab door closure indicators and strengthen driver training on door operation procedures. In addition, targeted retraining for personnel with frequent operational errors can help reduce the risk of incidents caused by human error.
• On-board signaling system: Conduct systematic testing and upgrades of the ATO modules, and consider adopting more stable alternatives to prevent cascading failures across subsystems.
• Brake system: Enhance preventive maintenance of electronic control units and require detailed failure documentation by technicians to support future analysis and system improvements.

Additionally, because failures are concentrated during high-temperature, high-load periods—particularly on spring and summer afternoons—it is recommended that the Taipei Metro assigns dedicated personnel for inspections and emergency response during these times to ensure operational safety and minimize the impact of unexpected incidents. Furthermore, the analytical framework and classification logic developed in this study are generalizable and may be applied to urban transit systems in other cities, especially those with similar incident categories and reporting mechanisms.

4. Conclusions

This study integrates supervised machine learning and association rule mining techniques to perform a comprehensive big data analysis of passenger injury incidents and train failure events within the Taipei Metro system. It proposes concrete and actionable optimization strategies, demonstrating both high novelty and practical contribution.

In the first part, the study innovatively applies a decision tree model to analyze 3,331 cases of passenger injuries. Through data preprocessing, oversampling, and feature engineering techniques, two highly discriminative rules were extracted, as follows. (i) Elderly passengers over the age of 61 are more likely to suffer moderate to severe injuries when involved in incidents within stations. (ii) Passengers aged 61 or below are more prone to serious injuries when using escalators during off-peak hours. These findings are among the first to quantitatively reveal the interaction between age and usage time in influencing injury severity. This fills a gap left by previous studies that primarily relied on surveys or descriptive statistics, and provides a new perspective for risk prediction and passenger service design.

The second part of the research focuses on exploring the causes of train failures. From 1157 records of operational delays exceeding five minutes, the study employs association rule analysis to uncover statistically significant conditional combinations related to failures, considering variables such as time period, season, metro line, and weather. The results not only identify the correlation between different subsystems (e.g., power supply, communication, and rolling stock) and their high-risk spatiotemporal patterns but are also cross-validated with maintenance records. This allowed for the classification of failure types (e.g., brake anomalies, door malfunctions, and switch-point failures) and the identification of predictive conditions. Notably, this research marks a pioneering application of association rule mining in metro maintenance decision-making, enabling a data-driven, proactive maintenance approach and providing a practical model for smart metro system management.

Moreover, this study reveals inconsistencies and missing data in the current maintenance record-keeping practices. It recommends enhancing documentation protocols and post-entry validation to improve data integrity and the accuracy of subsequent analyses. Regarding the threshold settings in association rule mining (such as minimum support and confidence), the study highlights their critical influence on rule discovery coverage. It thus recommends that future practical applications incorporate expert judgment and system-specific requirements to dynamically adjust parameter thresholds and uncover more insightful patterns.

In summary, this study demonstrates a cross-disciplinary integration by systematically applying machine learning and association rule techniques to metro incident analysis in Taiwan. It proposes a quantifiable and practically feasible risk identification and prevention framework. The approach not only enhances the safety and reliability of the metro system but also serves as a transferable reference model for risk assessment and maintenance optimization in other urban rail transit systems, with significant scalability and application potential.

Future research could extend this framework to other urban transit systems and incorporate real-time data streaming for dynamic risk prediction. Additionally, future studies may consider integrating passenger service quality and outsourced service outcomes into the analytical framework, with a focus on performance-based contract management. To enhance operational oversight, performance evaluation indicators should be thoroughly documented and systematically monitored. It is recommended to establish a periodic tracking and cross-verification mechanism—short-term (every 3 years), mid-term (every 5 years), and long-term (every 10 years)—to support regular performance reviews and enable data-driven adjustments based on cumulative records. In addition, it is important to consider regional differences. Metro systems may vary significantly in terms of passenger composition, infrastructure, outsourcing models, and managerial strategies. Therefore, prior to applying this framework in other contexts, localized data training and model adaptation would be essential. We also recommend extending this framework to metro systems in other major Asian cities—such as Tokyo, Seoul, and Singapore—to validate the applicability and robustness of the proposed risk prediction and classification mechanism across diverse urban environments.