Knowledge Graph Completion for High-Speed Railway Turnout Switch Machine Maintenance Based on the Multi-Level KBGC Model

Abstract

The incompleteness of the existing knowledge graphs in the railway domain creates information gaps, impacting their quality and effectiveness in the operation and maintenance of high-speed railway turnout switch machines. To address this, we propose a multi-layer model (KBGC) that combines KG-BERT, graph attention network (GAT), and Convolutional Embedding Network (ConvE) for knowledge graph completion in railway maintenance. KG-BERT fine-tunes a pre-trained BERT model to extract deep semantic features from entities and relationships, converting them into graph structures. GAT captures key structural relationships between nodes using an attention mechanism, producing enriched semantic and structural embeddings. Finally, ConvE reshapes and convolves these embeddings to learn complex entity interactions, enabling accurate link prediction. Extensive experiments on the HRTOM dataset, containing triplet data from high-speed railway turnout switch machines, demonstrate the model’s effectiveness, achieving an MRR of 50.8% and a Hits@10 of 60.7%. These findings show that the KBGC model significantly improves knowledge graph completion, aiding railway maintenance personnel in decision making and preventive maintenance, and providing new tools for railway maintenance applications.

1. Introduction

With the expansion of railway networks and increasing train speeds in China, the demand for reliable and safe signal infrastructure in the railway industry is rising [1]. There are many types of signal equipment failures, and their impact range is extensive [2]. As the actuator in the railway turnout switch system, the turnout switch machine converts electrical energy into mechanical energy through an internally mounted motor, driving the switch machine to control the switching of railway turnouts and enabling the train to switch and steer between different tracks. Its performance directly impacts the safety and efficiency of railway transportation [3], making it a key focus in high-speed railway signal equipment maintenance. It is crucial to detect all types of faults, as fault detection is instrumental in improving the level of intelligence of railway operations and optimizing their efficiency [4]. Numerous railway researchers have adopted data-driven approaches to analyze equipment failures. Reference [5] investigated the ZDJ9 switch conversion power curve, summarizing typical on-site failure types and causes, and developed a high-speed railway switch fault diagnosis model using the improved SSA-SVM algorithm. Reference [6] focused on the normal conversion and action curves of the S700 switch machine under failure conditions, using deep learning algorithms to extract image features, and achieving real-time fault diagnosis via the DCNN-SVM method. Reference [7] performed fault diagnosis using the sound of the switch machine during operation, addressing the limitation of traditional methods that rely solely on electrical signals without considering the physical characteristics of electromechanical systems. The abovementioned studies on switch equipment failures primarily focus on structured data. As high-speed railway operating mileage accumulates, substantial amounts of diverse data, including text, video, and images related to switch equipment operation and maintenance, have been collected, making unstructured data, particularly text data, the dominant component of big data. These text data contain critical information for ensuring railway safety. These switch operation and maintenance data are stored in an unstructured text format, serving as a knowledge base for future fault identification, diagnosis, and repair. However, the unstructured text data for high-speed turnout switch machine operations and maintenance are still recorded manually, resulting in issues such as ambiguity and irregularity in many fault records, challenges in computer processing, time-consuming manual analysis, and incomplete fault information mining [8]. When equipment faults occur, maintenance personnel must determine the cause based on specific conditions and relevant documents, resulting in relatively inefficient repairs. A knowledge graph is a knowledge base structured as “entity–relationship–entity” triplets, depicting concepts, entities, and their interrelationships in the physical world [9], thus effectively representing complex information. Knowledge graphs have been applied in fields such as power [10], aviation [11], and disaster management [12]. In the railway domain, Reference [13] organizes multi-source knowledge of high-speed trains to construct the schema layer of a high-speed train knowledge graph. Based on this conceptual schema, a data layer is developed to build domain-specific knowledge graphs for various stages of high-speed train operations. Reference [14] constructs a multi-domain integrated high-speed train maintainability knowledge graph, effectively integrating multi-source data to assist designers in train maintainability design and improve product quality. Reference [15] analyzes railway onboard fault maintenance logs and constructs a knowledge graph for railway onboard equipment faults, enhancing the knowledge discovery capability of fault logs and supporting intelligent maintenance. Reference [16] used semi-structured data from high-speed railway turnout equipment and applied a reinforcement learning-based approach to construct a knowledge graph for fault handling, thereby assisting maintenance personnel in decision making. However, the knowledge graphs generated through manual or semi-automated methods are often sparse and incomplete, leading to low-quality outputs with limited applicability. Research on knowledge graph completion in the railway domain is limited. Thus, filling the missing parts of the knowledge graph triplets is essential for enhancing the knowledge system, improving structural integrity, and providing a robust data foundation for decision support, preventive maintenance, and other railway maintenance applications.

Knowledge graph completion seeks to predict and fill missing elements in triplets [17], including entity and relation prediction. Entity prediction refers to identifying the missing head or tail entity type in a knowledge triplet, while relation prediction identifies the missing relationship type [18]. Reference [19] proposed the TransE model for knowledge graph completion based on translation distance. This model embeds entities and relationships into a vector space and uses vector addition to represent relationships, minimizing the distance between entities and their corresponding translations. Although simple, efficient, and easy to implement, the model performs poorly in handling complex relationships and does not fully consider semantic information [20]. Reference [21] proposed the RESCAL model, which is based on tensor decomposition. The model represents the knowledge graph as a high-order tensor and decomposes it to learn the low-dimensional embeddings of entities and relationships. However, it is susceptible to data sparsity [22]. Reference [23] introduced the ConvE model, a neural network-based approach that optimizes parameters through end-to-end training, capturing complex nonlinear relationships and higher-order features. However, the captured interaction relationships remain limited, focusing solely on the interaction information of stacked matrices, and the model fails to fully mine the semantic information between triplets [18]. Reference [24] proposed the KG-BERT knowledge completion model based on a pre-trained language model. This model leverages the pre-trained knowledge of the BERT module, which is highly effective in extracting textual semantic information and has demonstrated excellent performance in processing railway text data [25,26]. However, the model does not fully exploit the graph structure information of the knowledge graph, leading to suboptimal performance in knowledge graph completion. Although knowledge completion models based on graph neural networks (GNNs) utilize graph data representation that aligns with the inherent structure of knowledge graphs [27], GNNs achieve higher accuracy and greater robustness in graph-related tasks across various fields [28]. Nevertheless, the extraction of semantic information features remains insufficient [29]. The aforementioned models have limitations in handling complex relationships, semantic information, global graph structures, and local interaction features within knowledge graph triplets. Knowledge graph completion has been applied in various fields, including literature [30], agriculture [31], and disaster management [32]. In the railway sector, ensuring the completeness and accuracy of knowledge graphs is crucial for effective maintenance. Knowledge graph completion can fill gaps in the existing knowledge bases, enhance knowledge graph quality, and ensure that fault-related information is captured and utilized during the maintenance of core equipment like high-speed railway turnout switch machines. This significantly improves maintenance decision-making accuracy, reduces failure rates, and extends equipment lifespan.

In summary, a multi-level network model, referred to as the KBGC model, that integrates KG-BERT, GAT, and ConvE is proposed for the knowledge graph completion of high-speed railway turnout switch machine operation and maintenance, incorporating specialized railway terminology. The model first leverages the pre-trained KG-BERT to extract semantic information inherent in the knowledge triplet data, which includes complex railway terminology. Next, the GAT encoder employs an attention mechanism to process graph structure data, highlighting important nodes and effectively capturing dependencies between complex nodes. Finally, the ConvE decoder performs feature interactions, integrating semantic information with graph structure information to generate the optimal prediction value. A visual knowledge completion system for high-speed railway turnout switch machine operation and maintenance was developed by constructing a knowledge graph from the existing operation and maintenance data and applying the KBGC model for knowledge completion. This system supports on-site railway maintenance personnel in decision making and facilitates preventive maintenance.

2. Construction of High-Speed Rail Turnout Switch Machine Maintenance Knowledge Graph

During the construction of the high-speed railway turnout switch machine operation and maintenance knowledge graph, a combined top-down and bottom-up construction method is adopted to address the high level of specialization in the operational data. Natural language processing techniques, such as Chinese word segmentation and feature extraction, are employed to extract information from the text [33]. First, for entity extraction from unstructured data, the BERT-BiLSTM-CRF [34] model is introduced. By leveraging BERT’s robust text representation capabilities, text data are transformed into character-level embedding vectors. Contextual features in the text are captured using a bidirectional long short-term memory network (BiLSTM), while the conditional random field (CRF) layer learns internal dependencies between labels to achieve the optimal label sequence, ensuring accurate entity extraction. Next, relationship extraction is conducted using the BERT-CNN [35] model. BERT encodes the text data into word vectors, which are then fed into a convolutional neural network (CNN) for training. Finally, the model utilizes the fully connected and softmax layers to extract the semantic relationships between the target entities. For example, the text: “The automatic shutter of turnout J3 at #8 failed to disconnect the starting circuit at the first set of contacts after the turnout was in position, resulting in no indication for the turnout. It was restored after manually actuating the automatic shutter multiple times. The shutter was replaced during the maintenance window on the morning of the 30th. Following an on-site inspection by the manufacturer’s personnel, it was confirmed that the automatic shutter was stuck due to poor maintenance, which was classified as a signaling and communication responsibility”, extracts the entities “Automatic shutter of turnout J3 at #8”, “Failed to disconnect the starting circuit at the first set of contacts”, “No indication for the turnout”, “Manually actuating”, “Maintenance window”, “Replaced”, “Automatic shutter was stuck”, “Poor maintenance”, and the relationship “resulting in”. However, inconsistencies in recording methods for high-speed rail turnout switch machine maintenance data can result in varying descriptions of the same equipment, fault phenomena, and handling measures, causing duplicate entity extraction during knowledge graph construction. For instance, “turnout no indication” and “turnout lost indication” are named differently but represent the same equipment entity. To resolve this issue, a specialized dictionary is constructed based on the entities in the high-speed railway switch machine operation and maintenance data (as shown in Figure 1). The cosine similarity algorithm is then used to match the extracted entities with the dictionary, effectively linking them to their corresponding unique entities.

The processed knowledge triplet dataset, named the HRTOM dataset, is stored in the Neo4j graph database, completing the construction of the high-speed railway turnout switch machine maintenance knowledge graph. A partial visualization of this knowledge graph is shown in Figure 2.

The constructed knowledge graph consists of 891 entities, 15 relationships, and 10,027 triplets in total. To enhance the application of HRTOM triplet data in the operation and maintenance of high-speed railway turnout switch machines, the data were categorized into eight groups with the guidance of railway experts, as shown in

Table 1. Triplet type.

Head Entity Type	Relationship Type	Tail Entity Type	Example Triplets	Number of Triplets
Fault Location	Affiliation	Switch equipment	(Switch locking device, affiliation, and external locking and installation)	127
Fault Phenomenon	Cause	Fault Phenomenon	(Red light flashing during track switch, cause, and switch start circuit cutoff)	565
Maintenance Measure	Obtain	Maintenance Result	(Simulation experiment, obtain, and switch returns to normal)	282
Maintenance Condition	Permit	Maintenance Measure	(Maintenance window, permit, and inspect the contact block assembly of the switch machine)	1021
Maintenance Measure	Verify	Maintenance Experiment	(Inspection and disconnection, verify, and simulation experiment)	346
Equipment Phenomenon	Take	Maintenance Measure	(Irregular pointer, take, and joint repair by union departments)	1932
Fault Location	Occur	Fault Phenomenon	(Sliding plate, occur, and switch stock rail lacks lubrication)	663
Fault Phenomenon	Determine	Fault Nature	(Switch machine motor malfunction, determine, and poor maintenance)	278

.

3. High-Speed Railway Turnout Switch Machine Maintenance Knowledge Graph Completion Model

When completing the high-speed rail turnout switch machine maintenance knowledge graph, it is essential to consider both the semantic information and the graph structure of the maintenance data. Figure 3 illustrates the KBGC multi-level model for the knowledge graph completion of high-speed railway turnout switch machine operation and maintenance developed in this study. The model consists of three core modules—KG-BERT, GAT, and ConvE—designed to enhance knowledge graph completion accuracy through the deep processing of textual and structured information. First, the knowledge graph triplet (h, r, and t) is converted into a text sequence, and a fine-tuned KG-BERT model is applied to extract the deep semantic features of the entities and relationships within the operation and maintenance texts of high-speed railway turnout switch machines. Next, the graph attention network (GAT) layer is employed to process the relevance and importance of different nodes, effectively extracting structural information from the knowledge graph. Finally, the ConvE decoder module extracts higher-order interaction features through convolution operations, effectively completing the missing parts of the knowledge graph, which can subsequently be utilized for switch machine maintenance decision making.

3.1. KG-BERT Layer

The fine-tuned KG-BERT model utilizes BERT’s robust semantic extraction capabilities. In the high-speed railway turnout switch machine maintenance knowledge graph completion task, triplet data are first converted into text sequences and then transformed into vector representations via the BERT embedding layer, as shown in Figure 4. First, the triplet is converted into a text sequence with special tokens, where the [CLS] token is added at the beginning of the sequence and [SEP] is used to mark the end. Then, the pre-trained BERT model encodes the sequence, converting each token into token embedding, segment embedding, and position embedding, which are summed to form a combined embedding, represented by the following formula:

$$x_i=e_i+p_i+s_i$$

(1)

In the formula, $x_i$ represents the final embedding of the i-th word in the input sequence. Token embeddings decompose the text of the entities and relations into tokens using a tokenizer and convert them into vectors. Segment embeddings treat the head entity, relation, and tail entity as distinct segments to differentiate the logical parts of the input sequence. Positional embeddings provide positional information for each token in the sequence, enabling the model to deeply understand the semantic connections between the entities and relationships. Next, the combined embeddings are fed into a multi-layered Transformer encoder. The Transformer encoder utilizes a self-attention mechanism to represent each word in the input sequence as a weighted sum of all the other words. The formula is given as follows:

$$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}({\displaystyle \frac{Q{K}^T}{\sqrt{{\mathrm{d}}_k}}})V$$

(2)

In the formula, $Q$, $K$, and $V$ denote the query, key, and value vectors, respectively, generated by $\mathrm{d}$. Generate the contextual semantic representation for each token, extracting the semantic embedding vectors of the head entity, relationship, and tail entity ($Y_h$, $Y_r$, and $Y_t$). Finally, the generated embeddings are mapped onto the graph structure, where the head entity and tail entity serve as nodes, and the relationship acts as the edge connecting them, forming the graph structure ($h_h$, $h_r$, $h_t$).

3.2. GAT Layer

The GAT layer strengthens the associations between the entities in the graph structure derived from the KG-BERT layer using a graph attention mechanism. The structure of the model is illustrated in Figure 5. By calculating attention weights between different entities, the model dynamically adjusts the information flow, enabling it to consider not only local connection features but also to capture global information across the entire graph structure.

The core of GAT is the calculation of attention coefficients, which determine the importance of each node i to its neighboring node j in the graph. The initial representation of nodes and edges is obtained based on the adjacency matrix A and the node feature matrix h.

$$A=(a_{ij})+I_{N_0}$$

(3)

In the formula, A is the adjacency matrix, representing the connection relationship between node i and node j, and III is the identity matrix, used to add self-connections in the adjacency matrix. There are N nodes, each with a feature dimension of F, and the set of node features is represented as follows:

$$h=\{h_1,h_2,\dots ,h_N\}{h}_i\in R^F$$

(4)

To obtain an updated feature matrix, a weight-sharing parameter matrix $W\in R^{F^{\prime }\times F}$ is first applied to each node, Then, self-attention mechanisms are executed for each node to calculate the corresponding attention coefficients, denoted as $e_{ij}$, representing the importance of node j to node i:

$$e_{ij}=\alpha (W{h}_i,W{h}_j)$$

(5)

The specific calculation of the attention score is as follows:

$$\alpha (W{h}_i,W{h}_j)=\mathrm{LeakyReLU}(a^T[W{h}_i\left|\right|W{h}_j])$$

(6)

In the formula, $a$ is the vector used to calculate the attention score, || represents the vector concatenation operation, and LeakyReLU is the activation function. Next, the attention coefficients of all the neighboring nodes are normalized using the Softmax function, as shown in the formula:

$$a_{ij}={\displaystyle \frac{\exp (e_{ij})}{{\displaystyle {\sum }_{k\in N(i)}\exp (e_{ik})}}}$$

(7)

Finally, the features of the neighboring nodes are aggregated to obtain the new feature representation of the central node i, as given by the formula:

$$h_i^{l+1}=\sigma ({\displaystyle \sum _{j\in N_i}{a}_{ij}{h}_j^l{W}^{\prime }})$$

(8)

In the formula, $h_i^{l+1}$ and $h_i^l$ represent the feature vectors of node i after and before the update, respectively. $N_i$ is the set of neighboring nodes of node i, $W^{\prime }$ is the weight matrix, and $\sigma$ is the nonlinear activation function.

The GAT layer dynamically captures and aggregates graph structural information, generating node embeddings enriched with structural details. These embeddings are then passed as input to the ConvE layer for further processing and feature extraction.

3.3. ConvE Layer

The ConvE model is one of the most widely used decoders for evaluating the validity of knowledge graph triplets [36]. As illustrated in Figure 6, the model effectively processes interaction features between entities and relations using convolutional and fully connected layers to achieve precise link prediction. The ConvE layer fuses the output features of the GAT layer with relation-specific embeddings, integrating entity and relation vectors through element-wise concatenation and reshaping them into a comprehensive feature matrix, which is subsequently fed into the convolutional layer. Convolutional filters capture specific relationships in the data from various perspectives, extracting key local features of entity–relation interactions. The convolved feature map is flattened and transformed into a one-dimensional vector, which is then integrated and projected into the embedding space through the fully connected layer. Next, matrix multiplication is performed on the entity matrix and relation matrix to compute the triplet score. The scoring function for the triplet is defined as follows:

$$p(h,r,t)=\mathrm{ReLU}(\mathrm{vec}(\mathrm{ReLU}[{\mathrm{e}}_h;e_r]\ast \omega )W)){\mathrm{e}}_t$$

(9)

In the formula, $\omega$ represents a set of filters; “$\ast$” is the convolution operator; $\mathrm{vec}(·)$ transforms the tensor into a vector; and $\mathrm{ReLU}$ is the activation function. Finally, the Logistic Sigmoid function is used to generate the prediction values for entities and relations.

3.4. Loss Function

To optimize the model’s predictive ability, a negative sampling loss function is used. For each positive sample triplet (e.g., “the switch detector’s movable contact is broken, measures are taken, the contact group is replaced”), several negative samples are generated by randomly replacing entities or relations. The loss function adjusts the model parameters by comparing the positive and negative samples, ensuring that the positive samples score higher than the negative samples, thereby improving prediction accuracy. The loss function is formulated as follows:

$$L=-{\displaystyle \sum _{(h,r,t)\in S^{+}}\left[\lg (\sigma (f(h,r,t)))+{\displaystyle \sum _{(h^{\prime },\mathrm{r},t^{\prime })}\lg (\sigma (-f(h^{\prime },r,t^{\prime })))}\right]}$$

(10)

In the formula, $S^{+}$ represents the set of positive samples, $f(h,r,t)$ is the score of the model for the triplet $(h,r,t)$, and $\sigma$ represents the Sigmoid function.

4. Empirical Results and Discussion

4.1. Experimental Datasets and Evaluation Metrics

The experimental data involved in this paper are the on-site high-speed railway turnout switch machine maintenance data from various railway bureau jurisdictions, covering nearly five years of high-speed railway turnout switch machine maintenance data. They include the component-level causes of turnout switch machine equipment malfunctions and failures, such as phase failure protectors, indication rod insulation, switch machine contact groups, turnout grounding leakage devices, and close-contact detectors. The HRTOM dataset is randomly divided into training, validation, and test sets in a ratio of 8:1:1 to conduct comparative experiments on the model. Mean Reciprocal Rank (MRR) [37] and Hits@k [38], commonly used in knowledge graph completion research, are selected as the evaluation metrics.

$$\mathrm{MRR}={\displaystyle \frac{1}{\left|S\right|}}{\displaystyle \sum _{i=1}^{\left|S\right|}\frac{1}{ran{k}_i}}={\displaystyle \frac{1}{\left|S\right|}}\left({\displaystyle \frac{1}{ran{k}_1}}+{\displaystyle \frac{1}{ran{k}_2}}+\dots \dots +{\displaystyle \frac{1}{ran{k}_{\left|S\right|}}}\right)$$

(11)

In the formula, $S$ represents the number of triplets, and $ran{k}_i$ represents the predicted rank of the ith triplet. The value of MRR ranges from [0, 1], with values closer to 1 indicating more accurate model predictions.

$$\mathrm{Hits}@k={\displaystyle \frac{1}{\left|S\right|}}{\displaystyle \sum _{i=1}^{S}I(ran{k}_i\le k)}$$

(12)

In the formula, Hits@k refers to the average proportion of triplets ranked less than k in triplet prediction; I represents the indicator function.

4.2. Experimental Parameter Settings

The experimental environment consists of a CPU (Intel i9 12900K, Intel, Santa Clara, CA, USA), GPU (NVIDIA GeForce RTX3080Ti, NVIDIA Corporation, Santa Clara, CA, USA), and the operating platform is Pycharm2021.2. The code is written in Python 3.8 using the deep learning framework Pytorch. After multiple experiments, the main parameter settings for the model to achieve optimal performance are shown in

Table 2. Experimental parameter settings.

Parameter Name	Parameter Value
KG-BERT Learning Rate	3 × 10−5
KG-BERT Embedding Dimension	768
KG-BERT Batch Size	16
GAT Dropout Rate	0.3
Number of GAT Layers	1
Number of GAT Attention Heads	8
ConvE Convolution Kernel Size	3
Optimization Function	Adam

.

4.3. Comparative Experiments

To validate the performance of the proposed model in the knowledge graph completion task for high-speed railway turnout switch machine operation and maintenance, comparative experiments were conducted under identical conditions using various models, including translational distance-based models TransE and TransH, tensor decomposition-based models Rescal and DistMult, a neural network-based model ConvE, and a graph neural network-based model KBGAT. The results are presented in Figure 7.

As shown in the figure, the proposed model outperforms the compared models in terms of MRR and Hits@10 on the HRTOM dataset, demonstrating the superiority of the KBGC model. Compared to the translation-based model TransH, the proposed model shows a 6.2% and 10.2% improvement in MRR and Hits@10, respectively. Although TransH addresses TransE’s limitations in handling many-to-many relationships, it still struggles to capture the nonlinear interactions between relationships. The proposed model utilizes the GAT layer’s attention mechanism to effectively identify and reinforce the dependencies between key nodes, better capturing complex relationships and improving prediction accuracy. Compared to the tensor decomposition-based model RESCAL, the proposed model shows a 4.5% and 9.3% increase in MRR and Hits@10, respectively. This is because in the HRTOM dataset, certain entities and relationships appear infrequently, and RESCAL struggles to handle sparse data, leading to poor prediction performance. By leveraging the KG-BERT layer’s ability to extract deep semantic features and the GAT layer’s processing of structural information, the proposed model maintains high prediction accuracy even with sparse data. Compared to the graph neural network-based model KBGAT, which also incorporates graph attention networks, the proposed model achieves better results, with MRR and Hits@10 increasing by 5.6% and 3.1%, respectively. This indicates that the proposed model can more comprehensively extract contextual information from nodes, enhance semantic feature representation in triplets, and achieve superior performance in complex data scenarios. Specifically, as shown in Figure 8, for the triplet (Automatic contact breaker in turnout switch machine, Occurs, Breakage of the actuating roller), the proposed model, in addition to identifying “Breakage of the actuating roller”, also predicts other possible fault phenomena for the automatic contact breaker in turnout switch machine, such as “Sticking or jamming of the pivot shaft”, “Insufficient tensile force of the tension spring”, “Fracture of the moving contact holder”, and “Short circuit and arcing of the stationary contact”. The TransH model completion results include “Sticking or jamming of the pivot shaft”, “No indication of positioning”, “Fracture of the moving contact holder”, and “Short circuit and arcing of the stationary contact”. The KBGAT model completion results include “Sticking or jamming of the pivot shaft”, “Insufficient tensile force of the tension spring”, “Fracture of the moving contact holder”, and “Midway stalling of the turnout switch machine”. Clearly, “No indication of positioning” and “Midway stalling of the turnout switch machine” are not direct faults of the automatic switch itself, but secondary fault phenomena caused by it. Therefore, it can be concluded that the proposed model outperforms the other models in terms of fault completion accuracy and rationality.

4.4. Ablation Experiments

To investigate the role of each component in the KBGC model architecture for high-speed railway turnout switch machine operation and maintenance knowledge graph completion, and to analyze the impact of these components on the model’s overall performance, several ablation experiments were conducted on the HRTOM dataset under identical experimental conditions. The results of these experiments are presented in Figure 9.

(1)

GAT + ConvE: The KG-BERT pre-trained model is excluded, leaving the GAT and ConvE modules. GAT captures the graph structure information by updating entity representations through graph edges (relationships). ConvE subsequently extracts local features from the entity and relationship embeddings via convolution operations, thereby enhancing the model’s predictive capability.

(2)

KG-BERT + GAT: The ConvE module is removed, retaining the KG-BERT and GAT modules. KG-BERT embeds the entities and relationships from the triplets into a vector space, utilizing the pre-trained BERT model for prediction to capture complex semantic information. While KG-BERT encodes the input, GAT updates the entity representation by considering the influence of the neighboring entities on the target entity. The attention weights are used to determine the importance of different neighbors, providing more comprehensive modeling and prediction of entities and relationships.

(3)

KG-BERT + ConvE: The GAT module is excluded, leaving the KG-BERT and ConvE modules. KG-BERT utilizes the BERT pre-trained model to capture global semantic information in the knowledge graph, while ConvE aggregates the local feature information of the entities and relationships through convolution operations. Although this combination balances semantic understanding and local feature extraction, it lacks supplementary graph structure information.

(4)

KBGC: As the complete model in the experiment, the above three models are integrated to form a multi-level architecture. KG-BERT performs the global semantic encoding of entities and relationships using the BERT model, capturing complex semantic information. The GAT module extracts graph structure features from the knowledge graph, supplementing local structural information. The ConvE convolutional network then aggregates features through interactions between the entities and relationships, thus accomplishing the completion task.

As illustrated in Figure 8, the proposed model outperforms the other models in terms of performance metrics. Compared to the GAT + ConvE model, the proposed model shows a 1.2% and 5.1% increase in MRR and Hits@10, respectively. This is because the GAT + ConvE model’s ability to represent semantic information is limited when extracting entities and relationships from triplets, resulting in lower prediction accuracy. This indicates that incorporating the KG-BERT model as a pre-training strategy effectively enhances the extraction of intrinsic triplet features, optimizes the training efficiency of the graph attention network, and improves the representation of entity and relationship encoding, thereby boosting link prediction performance. Compared to the KG-BERT + GAT model, the proposed model shows a 1.7% and 4.4% increase in MRR and Hits@10, respectively. This is because without ConvE, the KG-BERT + GAT model struggles to extract effective local features from long-tail entities, resulting in poor performance when dealing with rare entities. Consequently, the model’s ability to generalize to long-tail entities and relationships declines during prediction. Compared to the KG-BERT + ConvE model, the proposed model shows a 2.9% and 3.9% increase in MRR and Hits@10, respectively. This is because without GAT, the KG-BERT + ConvE model can only capture dependencies for single-hop or short-path relationships. It fails to model long-path dependencies (e.g., multi-hop reasoning) or cross-entity relationship patterns, resulting in reduced prediction accuracy. The proposed model leverages graph structure, global semantic information, and convolutional features, thus achieving superior performance in triplet prediction tasks.

4.5. Model Generalization Analysis

To evaluate the generalization ability of the proposed model, experiments were performed on the public datasets FB15k-237 [39] and WN18RR [23]. The experimental results are presented in

Table 3. Experimental results of the model on public datasets.

Dataset	Model	MRR/%	Hits@1/%	Hits@3/%	Hits@10/%
FB15k-237	TransE	29.4	20.7	31.6	46.5
	TransH	27.1	19.8	30.3	44.2
	RESCAL	35.4	26.1	38.8	53.2
	DistMult	24.1	15.5	26.3	41.7
	ConvKB	25.3	16.0	27.9	42.2
	KBGAT	51.8	25.8	39.8	62.1
	KBGC	53.6	41.2	52.9	63.7
WN18RR	TransE	22.6	5.3	36.1	49.7
	TransH	45.9	9.8	39.3	51.7
	RESCAL	46.8	42.2	47.1	52
	DistMult	42.8	39.0	44.0	49
	ConvKB	24.7	5.4	36.4	52.4
	KBGAT	44	45	48.6	58.1
	KBGC	48.5	43.3	48.0	59.3

. The results indicate that the proposed model achieves strong performance, with the MRR values on FB15k-237 and WN18RR improving by 1.8% and 4.5%, respectively, compared to the state-of-the-art KBGAT model. These findings demonstrate that the proposed KBGC model exhibits excellent generalization ability.

5. Knowledge Completion Visualization Application

The model predicts the five most relevant missing entities based on their associations with the current entity and relationship. For example, a “Poor contact of turnout close contact detector” fault may cause the phenomenon of “Indication circuit cannot be connected”. After model completion, it may also result in other faults such as “Idling”, “ Improper Switching of Turnout”, “Turnout obstruction”, and “No indication for turnout.” Figure 10 illustrates three types of completion examples: (Fault Phenomenon, Cause, Fault Phenomenon), (Equipment Phenomenon, Take, Maintenance Measure), and (Fault Location, Occur, Fault Phenomenon).

To assist maintenance personnel in effectively utilizing the completed missing information, this study developed a knowledge completion system for high-speed railway turnout switch machine maintenance using the Flask framework, as illustrated in Figure 11. By inputting an entity and relationship from the triplet, the system will automatically identify the top five missing pieces of information most relevant to the current input and rank them in descending order of relevance (displayed from top to bottom in the system interface). By completing and inferring missing information from the existing data, the system predicts equipment operating conditions, failure causes, safety hazards, and maintenance measures. This assists railway maintenance personnel in better understanding turnout failure mechanisms, supports decision making, enables preventive maintenance, and accelerates the development of intelligent maintenance for high-speed railway fault handling.

6. Conclusions and Outlook

A knowledge graph completion model based on the KBGC framework is proposed to improve the maintenance of high-speed railway turnout switch machines. The model leverages deep learning techniques to comprehensively extract feature information from knowledge graph triplets. Its superiority and generalization capabilities are validated through comparisons with other models and evaluations on public datasets. By improving the completeness of the knowledge graph for railway switch machine maintenance, the model significantly enhances the accuracy and reliability of graph data, thereby strengthening intelligent decision-making and information services. Additionally, a user-friendly visualization system for the knowledge graph completion process was developed using the Flask framework, offering an intuitive interface for human–computer interaction. This system facilitates the practical application of the knowledge graph by enabling maintenance personnel to quickly access relevant information for predictive maintenance, thereby gaining a deeper understanding of fault mechanisms in high-speed railway turnouts and assisting them in making informed decisions to achieve effective preventive maintenance. In future work, we plan to optimize the current model by developing a lightweight version to reduce training time and computational costs, facilitating its deployment in large-scale knowledge graphs. Moreover, the integration of interpretability modules, such as case-based reasoning and multi-level attention visualization tools, will be prioritized to make the system outputs more comprehensible for non-technical users, such as railway maintenance personnel. This will enable them to better comprehend the model-generated results and make informed decisions based on these insights. Additionally, a feedback mechanism will be integrated into the knowledge graph completion system, allowing users to provide feedback on the prediction results, thereby improving the model’s performance and enhancing its ability to support precise and effective decision making for maintenance tasks.