Causality and Equipment Structure Enhanced Maintenance Plan Recommendation with Knowledge Graph Integration

Yanying Wang; Ying Cheng; Qinglin Qi; Zhiheng Zhao; George Q. Huang; Stefan Pickl; Fei Tao

doi:10.1016/j.eng.2025.11.036

Engineering ›› 2026, Vol. 63 ›› Issue (8) :137 -155. DOI: 10.1016/j.eng.2025.11.036

Research

research-article

Causality and Equipment Structure Enhanced Maintenance Plan Recommendation with Knowledge Graph Integration

Author information +

History +

PDF (4101KB)

Abstract

Recommending maintenance plans presents significant challenges due to the low standardization of maintenance records and unclear pathways for identifying appropriate plans. While knowledge graphs have been extensively researched for integrating and evolving maintenance data, these issues hinder the accurate recommendation of maintenance solutions within large-scale maintenance knowledge systems. This paper proposes a causality and equipment structure enhanced maintenance plan matching and recommendation (CEE-MPMR) method to address these challenges. The method leverages an unsupervised SimCSE model to normalize domain vocabulary in the absence of domain lexicon, and proposes a maintenance plan reasoning method based on RotatE. The proposed method achieves a maintenance plan matching accuracy of 90.80%, effectively improving the precision of maintenance plan recommendations. Finally, we applied and validated the approach on real-world data from a nuclear power enterprise and integrated the algorithm into a maintenance plan recommendation system, supporting intelligent analysis and decision-making for nuclear complex equipment maintenance.

Graphical abstract

Keywords

Maintenance plan recommendation / Knowledge graph / Fault causality and equipment structure / Knowledge reasoning / Knowledge graph embedding

Cite this article

Download citation ▾

Yanying Wang, Ying Cheng, Qinglin Qi, Zhiheng Zhao, George Q. Huang, Stefan Pickl, Fei Tao. Causality and Equipment Structure Enhanced Maintenance Plan Recommendation with Knowledge Graph Integration. Engineering, 2026, 63 (8) : 137-155 DOI:10.1016/j.eng.2025.11.036

登录浏览全文

4963

注册一个新账户忘记密码

1. Introduction

Equipment maintenance is an essential phase in the entire lifecycle of equipment and aimed at realizing its value, improving the economic efficiency of enterprises, and ensuring safety [1]. With the development of equipment complexity and information technology, equipment maintenance has gone through four stages: event-based reactive maintenance, time-based preventive maintenance, condition-based predictive maintenance, and data-driven predictive maintenance [2]. With advancements in sensors and the Internet of Things (IoT), large volumes of data have been accumulated [3]. Artificial intelligence now enables deeper insights into these data, fostering data-driven predictive maintenance. Compared with traditional maintenance strategies, predictive maintenance focuses more on the future health status of equipment and the collaboration between equipment, data, and operators, ultimately achieving intelligent maintenance on the basis of equipment history, current status, and future trends [4], [5].

In the current field of predictive maintenance, various analyses focusing on structured data such as operational state monitoring, the optimization of equipment operating parameters, and predictive fault warnings, have made significant progress [6], [7], [8]. However, the complexity of the knowledge involved in the equipment maintenance phase is significant [9]. Given that this knowledge directly influences personnel operations and decisions during maintenance, its importance cannot be overlooked. Currently, during the formation of complex equipment maintenance plans, owing to limitations in the technical capabilities of maintenance personnel, extensive data retrieval may be needed, and for complex faults, detailed documentation often necessitates technical support from experts, sometimes requiring several months of repeated surveys to develop a maintenance plan. The difficulty in obtaining complex equipment maintenance plans significantly reduces the maintenance efficiency [10], [11].

However, with the continuous improvement of enterprise informatization construction, enterprises have gradually established comprehensive information feedback and management systems [12], providing the possibility to explore the relationship between equipment faults and maintenance plans from accumulated equipment maintenance knowledge. In areas such as wind turbine maintenance [13] and aviation equipment maintenance [14], employing maintenance records to enhance maintenance question and answer (Q&A) quality has already yielded promising results. Equipment maintenance knowledge can be summarized into two major categories: procedural data and empirical data [10]. Taking nuclear power enterprise information management as an example, nuclear power enterprises have standardized the operational guidelines for regular maintenance and routine maintenance of large and complex equipment represented by nuclear circulation water pumps. They have established comprehensive experience feedback systems and maintenance work order systems for complex equipment maintenance, forming two types of data: procedural data and empirical data for equipment maintenance.

Formal concept analysis (FCA) formalizes conceptual knowledge using conceptual graphs [15]. This method converts the hierarchical logic of language into mathematical structures that can be processed by models, making it highly valuable in areas such as knowledge management and the semantic web. With the continued advancement of natural language processing (NLP) technologies, knowledge graph (KG), a type of semantic network designed to effectively organize and manage large-scale knowledge, have been widely applied in knowledge management, retrieval, recommendation, and decision support in recent years [16], [17]. Compared with traditional knowledge management methods, KGs offer more diverse information access categories and stronger knowledge reasoning capabilities. Therefore, in the industrial domain, many experts and scholars have conducted extensive research on various applications of KGs combined with manufacturing processes [18], [19].

In terms of knowledge management, many researchers have focused on distributed equipment maintenance knowledge. KG has been introduced to conduct knowledge mining on original maintenance texts, thereby achieving structured knowledge management and providing a solid knowledge foundation for the intelligent maintenance of complex equipment [20], [21], [22]. In the area of knowledge discovery, efforts have been made to use failure reports for failure diagnosis [23], troubleshoot long-chain failures [24], predict links in existing KGs to complete manufacturing KGs [25], [26], evolve KGs to enrich fault knowledge [27], discover potential solutions [28], and explore potential relationships among types of knowledge [29]. However, after the management and mining of large-scale knowledge have been achieved, determining how to quickly access knowledge, especially in the field of complex equipment maintenance, and determining how to explicitly recommend target content to users have become new challenges.

Thus, by analyzing the process of complex equipment maintenance, there are still two main issues in the current research on KG-based equipment maintenance plan recommendation:

(1) Contradiction between a low level of record standardization and a high level of maintenance accuracy requirements. Although the equipment maintenance KG effectively integrates various forms of maintenance knowledge, some terms in the maintenance KG are not standardized, and operators often use colloquial expressions when the maintenance needs are input. Therefore, there is an urgent need for a universal method for standardizing vocabulary in the equipment maintenance domain.

(2) Unclear maintenance plan discovery paths and multi-hop reasoning difficulties in equipment maintenance KGs. Although KG-based question answering (QA) and reasoning techniques continue to improve, they encounter limitations when applied to sparse, domain-specific graphs and to situations in which only limited amounts of domain data are available. Complicated knowledge-based QA (KEQA)-style methods [30] learn accurate predicate and head entity representations through supervised training and then infer the tail entity via an embedding-space relation, but these approaches are confined to single-hop questions and require large, manually annotated datasets. Neural logical-query models such as New Look [31] can cope with more complex logical patterns; however, they rely on millions of query-answer pairs generated from dense KGs and therefore do not transfer well to the sparser equipment-maintenance graphs studied in this article. Baselines using lightweight large language models (LLMs) exhibit a sharp decrease in answer accuracy because they lack access to essential information such as system codes and hierarchical device structure.

On the basis of the equipment maintenance KG, simple question matching related to equipment maintenance knowledge can be accomplished. However, the faults and maintenance plans within the KG do not always have a one-to-one correspondence. Often, obtaining a maintenance plan requires reasoning with both fault causality relationships and the equipment structure. Currently, this process relies on subjective judgments by operators, leading to a decrease in maintenance efficiency and accuracy.

Therefore, to address the above two issues, this paper contributes to the literature in the following main ways:

(1) This paper proposes an unsupervised SimCSE domain vocabulary normalization method. The incremental training of the unsupervised SimCSE model enhances the model's sensitivity to domain vocabulary, which is significant for the construction of domain knowledge bases. On this basis, a normalization process for KG and queries is proposed for knowledge inference. By standardizing and integrating nodes in the equipment maintenance KG, the redundancy of nodes and relationships in the knowledge base is effectively avoided, thus improving knowledge matching and reasoning efficiency.

(2) This paper proposes a maintenance plan recommendation algorithm based on RotatE embedding with fault causality and equipment structure enhancement (RotatE-CEE). By simulating the reasoning process of maintenance personnel, the method integrates fault causality with hierarchical equipment knowledge, and ranks candidate plans through RotatE-based similarity. Unlike the methods in Refs. [30], [31], RotatE-CEE performs the entire QA cycle without any additional supervision once the KG has been constructed; it supports multi-hop paths that mix causal and structural relations, and every recommended plan can be traced back to concrete nodes and edges in the KG.

2. Related work

2.1. Knowledge-based QA (KBQA)

KBQA is highly challenging. Current mainstream KBQA methods can be categorized into semantic parsing (SP)-based methods and information retrieval (IR)-based methods [32].

The core idea of SP methods is to convert natural language questions into a series of formal logical expressions. Owing to their good reasoning properties, these logical expressions are parsed bottom-up and eventually executed as queries in the KG [33]. One approach uses rule-based templates, where questions are parsed into logical expressions through predefined templates or syntactic rules [34]. Another approach uses query graphs, where syntactic dependency trees are employed to map questions to query graphs [35]. Both methods perform well on small-scale knowledge bases and offer good interpretability. However, they require extensive manual intervention for template definition and query graph construction, resulting in poor generalization. In addition to rule-based and query graph methods, recent developments in deep learning have led to the translation of questions into logical expressions as a machine translation task, employing Seq2Seq models to construct logical representations [36]. Compared with earlier methods, this approach reduces the reliance on manually defined rules but requires a substantial amount of labeled data, which is often insufficient in domain-specific QA tasks.

IR-based methods first link the entities mentioned in the question to corresponding entities in the KG and then extract subgraphs related to these entities as the candidate answer set. Finally, the ranking models are used to filter this set and produce the final answer. IR-based methods can be divided into two categories: ① Feature engineering-based methods utilize techniques that rely on syntactic analysis of the question to extract key components such as interrogative words, topic words, and verbs. Subgraphs related to the topic words are then extracted from the KG to generate candidate answers. Features are designed for both the question and the candidate answers, and a classifier is used to filter the final answer [37]. However, this approach requires precise feature design and struggles with complex questions. ② Representation learning-based methods introduce knowledge representation techniques to KBQA. The core idea is to map both questions and candidate answers into the same low-dimensional vector space and, through model training, maximize the similarity between the question vector and the answer vector [38]. This approach eliminates the need for syntactic parsing of questions, but owing to the simplistic embedding of entities and relations, it is inadequate for handling the interdependencies in complex questions and answers.

2.2. KG-based reasoning

Current KG-based question QA face significant challenges related to weak reasoning capabilities, which urgently need to be addressed. In the field of complex equipment maintenance, developing a maintenance plan is a complex process that involves reasoning based on multiple factors, such as equipment structure, operational logic, and fault causality. This reasoning process is often carried out by operators who rely on their experience and judgment. Therefore, enhancing the reasoning ability of models in KG-based QA and recommendation systems is a critical step toward achieving intelligent maintenance, particularly in industrial domains and equipment maintenance.

Reasoning based on KGs can be categorized into four approaches: rule-based, graph structure-based, representation learning-based, and neural network-based reasoning. Rule-based methods, derived from traditional deductive and inductive reasoning [39], offer strong interpretability but exhibit limited generalizability. Graph structure-based reasoning treats the KG as a connected graph, using random walk techniques to infer possible paths [40], although it performs poorly with infrequent relations. Neural network-based reasoning introduces graph neural networks into KG modeling, providing strong reasoning capabilities [41], but these methods are computationally complex and lack interpretability. Additionally, owing to the nature of neural networks, validating the results of these reasoning approaches for the existence of and potential for conflicts is difficult. Representation learning-based methods map the nodes and relations in a KG to low-dimensional vectors, enabling the prediction and inference of potential rules. These methods are widely applied because they balance interpretability and generalization [42].

A typical representation learning approach is the translational distance model, TransE [43]. TransE is the earliest and most well-known translational distance model and first introduced the method of representing entities and relations as vectors to perform knowledge reasoning. Owing to its limitations in various aspects, a series of improved TransE models have been proposed. In addition to the TransE series, another representation learning approach involves mapping triples to matrices and using matrix factorization for triple prediction, such as the relational scalable tensor embedding (RESCAL) [44] family of algorithms. These methods also offer some interpretability and fast computation but are limited in their reasoning ability, as they consider only constraints that satisfy factual conditions in the KG.

In KG embedding-based reasoning, in addition to representation learning methods, there are also approaches that integrate KG embeddings with neural networks. With respect to single-hop QA involving a single entity and a single relation, Huang et al. [30] employ neural networks to learn the representations of head entities, tail entities, and predicates within the KG embedding space, aligning them with the graph’s triples to determine the optimal response. While this method can capture multiple expressions of entities and predicates, its focus on single-hop queries makes it difficult to apply to equipment maintenance questions. For complex logical queries, a series of neural network-based reasoning methods represented by Refs. [31], [45], [46] map logical operations into the embedding space, thereby achieving multi-hop inference on large-scale KGs. However, learning such mappings requires extensive labeled data, which are not available in the context of nuclear power equipment maintenance addressed by this paper.

2.3. Research gaps

With the continuous advancement of KG-based QA and reasoning technologies, numerous studies have explored their application in efficient intelligent search and recommendation within industrial domains based on domain-specific KGs, such as distributed photovoltaic operation and maintenance [47] and wind farm equipment maintenance [48]. However, current applications of KG-based QA and reasoning in the field of complex equipment maintenance face several challenges, including insufficient annotated datasets, unstructured original data records, and difficulties in designing fault diagnosis and maintenance reasoning pathways. Consequently, many studies focus on matching general maintenance knowledge, such as answering questions about equipment components and operation methods, without advancing maintenance plans for specific faults [49], [50]. This underutilizes the knowledge reasoning capabilities of KGs.

On the basis of the above analysis, the primary reasons for the low accuracy in complex equipment maintenance plan matching can be summarized as follows: ① low standardization of knowledge in equipment maintenance records, leading to unclear information redundancy; and ② unclear paths for discovering maintenance plans, making multi-hop reasoning in equipment maintenance KGs challenging. To address these issues, this paper proposes a vocabulary normalization process for domain knowledge in the context of limited labeled data, aimed at resolving the standardization issues in domain knowledge representation. Furthermore, a query template-based plan matching method is introduced, along with a fault causality and equipment structure-enhanced RotatE [51] graph embedding-based maintenance plan recommendation method. These approaches ultimately enable intelligent recommendation for complex equipment maintenance.

3. Causality and equipment structure enhanced maintenance plan matching and recommendation

On the basis of the problem analysis above, the recommendation of complex equipment maintenance plans faces two main challenges: domain vocabulary standardization and multi-hop reasoning for maintenance plan discovery. Therefore, this paper proposes a causality and equipment structure enhanced maintenance plan matching and recommendation (CEE-MPMR) method, comprising three main components. The overall process is illustrated in Fig. 1.

The equipment maintenance knowledge discussed in this paper can be divided into two major categories: procedural data and empirical data. Procedural data, such as alarm manuals and maintenance procedures, provide operators with fundamental solutions for faults during reactive maintenance and guidelines for preventive maintenance. These are typically authored by domain experts. Empirical data, such as fault repair reports and experience feedback, cover faults discovered during reactive maintenance or periodic inspections. This type of data fills the gaps in fault types not addressed by procedural data and is typically recorded by operators.

Wang et al. [10] proposed a method based on dataspace techniques to extract knowledge from procedural data and empirical data and construct a KG. Using nuclear power equipment as an example, the method was used to build a nuclear power equipment maintenance KG based on industrial dataspace called IDS-KG. This paper further introduces a method that implements causality and equipment structure enhancement of maintenance plan matching and recommendation (CEE-MPMR). This approach leverages the integrated IDS-KG, which incorporates both procedural and empirical data, to provide precise maintenance plan recommendations for equipment maintenance personnel.

(1) Unsupervised SimCSE domain vocabulary normalization. To address the problem of inconsistent expression of domain vocabulary in IDS-KG and user input query statements, a standard vocabulary and entity library for the maintenance domain are first constructed, and a domain synonym library is generated through the maintenance domain knowledge extraction dataset and entity extraction model. Afterward, using publicly available datasets and the three aforementioned lexicons, an unsupervised SimCSE model is trained incrementally to standardize the domain vocabulary. Finally, a method based on the SimCSE model for IDS-KG and query statement normalization is proposed and the IDS-KG is normalized to form normalized nuclear power equipment maintenance KG (IDS-KG-QA), providing a foundation for subsequent maintenance plan matching and recommendation.

(2) Template-based basic question-answer matching. Owing to the relatively stable scope of issues and knowledge in the field of equipment maintenance, this section adopts a template-based approach to match several types of basic questions with one-hop answers. First, entity recognition is performed on the normalized queries, which are mapped to corresponding nodes in the normalized KG. Afterward, intent recognition based on templates is applied to the queries. If the query matches an intent recognition template, a query statement is generated based on the template and matched to an answer. However, if the templates constructed do not include the nodes or query intents of the query, answer inference based on RotatE is performed.

(3) Maintenance plan reasoning based on RotatE-CEE. To further match fault repair schemes that require multi-hop reasoning, this section first proposes an enhanced sampling method for RotatE KG embedding. It accurately embeds nodes and relationships from the normalized KG IDS-KG-QA. Then, by analyzing the maintenance plan generation process of operators, a maintenance plan reasoning method is proposed that integrates causality and equipment structure. Finally, candidate answers are selected based on the above method, and RotatE embedding is used to rank the candidate answers and match them with the query statement to obtain the maintenance plan recommendation result.

Before introducing CEE-MPMR, note that the proposed method is builds upon an existing IDS-KG. For details on entity and relation extraction, as well as the IDS-KG data sources and input formats, please refer to Ref. [10]. In the original IDS-KG, failure type, cause, and effect were all labeled as fault nodes, while detect point (i.e., various signal sensors) and manufacturing items were labeled as equipment nodes. Although Ref. [10] separates these categories for ease of construction, this paper focuses on fault causality and equipment structure, thus retaining Failure type, cause, and effect collectively as fault nodes, and detect point and manufacturing items collectively as equipment nodes. Tables 1 and 2 present the KG schema used in this article.

3.1. Unsupervised SimCSE domain vocabulary normalization

When query questions are embedded, considering that there are many synonymous expressions in maintenance requests, the conventional approach is to obtain the similarity between words by measuring the distance between word vectors, thus replacing similar words. However, the vocabulary in the field of complex equipment operation and maintenance differs from pre-trained word vectors in the public domain, and the current scale of available corpora in the complex equipment maintenance domain is small and which cannot support the training of word vector models. Therefore, relying solely on word vectors for synonym matching is not ideal in industrial scenarios. Moreover, due to the lack of publicly available datasets annotated with word similarity, it is difficult to apply supervised methods for vocabulary standardization and similarity calculation in the field of equipment maintenance. To address these issues, an unsupervised domain vocabulary standardization method is proposed in this section, which is based on SimCSE to resolve situations where vocabulary lists and domain lexicons are lacking.

3.1.1. Corpus collection for normalization

To standardize the terminology in the maintenance field, this section, using nuclear power complex equipment maintenance texts as an example, collects the training corpus for the model from the following sources:

(1) Domain standard vocabulary. In practical production, a set of codes or abbreviations is often used for more convenient reference to various equipment. If these standard codes appear in the knowledge base or KG, they are replaced with the corresponding standard terms from the table, which serve as a domain standard vocabulary. This vocabulary helps prepare for subsequent terminology normalization. An example of the standard domain vocabulary used in this section is shown in Table 3. In cases where such coding or abbreviation systems are not readily available, web scraping can be used to acquire the standard vocabulary from encyclopedic sources. While this may sacrifice some precision, it offers broader applicability.

(2) Domain entity library. Unlike the standard vocabulary, the domain entity library consists of colloquial terms often found in maintenance experience records rather than formalized procedural documents. These terms, referred to as “domain entities,” capture the informal expressions of equipment names and maintenance terms. To construct the domain entity library, this section employs a BERT-BiLSTM-CRF named entity recognition model to identify entities from “Fault Repair Record” and “Experience Feedback.”

(3) Domain synonym library. Due to the inconsistent terminology common in the maintenance field, a domain synonym library is essential for improving vocabulary standardization. To address the lack of synonym tables and lexical similarity datasets in the equipment maintenance domain, this section proposes using a Word2Vec model combined with manual identification to construct a basic synonym library for the maintenance field. The process, illustrated in Fig. 2, takes input from the domain standard vocabulary (as shown in Table 3) and the domain entity library obtained through named entity recognition, ultimately generating a domain synonym library.

3.1.2. Unsupervised SimCSE model for annotated data scarcity

The SimCSE model is a generative model for word and sentence vectors [52], which employs the idea of contrastive learning to generate word and sentence vectors without the need for supervised data. Contrastive learning was initially applied in the field of computer vision as a special form of unsupervised learning. Its goal is to minimize the distance between positive sample pairs while maximizing the distance between negative sample pairs through training. Data augmentation methods for images include cropping, rotating, and flipping the original images to create positive samples. However, data augmentation in NLP is more challenging. Common methods include word deletion and synonym replacement, but these approaches can lead to semantic changes when dealing with short texts. Therefore, SimCSE adopts a dropout-based method for data augmentation. The structure of the original unsupervised SimCSE model for data augmentation is shown in Fig. 3.

For sample $ x_{i}$, let its positive sample be denoted as $ x_{i}^{+}=x_{i}$. Then SimCSE inputs each sample twice into the encoder. Due to the transformer structure of the model, dropout layers exist in the model architecture. As dropout randomly deactivates neurons, the same sample generates different embeddings when input into the same model, resulting in the embeddings of the sample and its positive sample $ h_{i}$,$ h_{i}^{+}$, respectively. In the same batch, all other samples except for this sample are considered negative examples, thus, the positive and negative sample sets for this sample are obtained. For a mini-batch of N sentences, the loss function is defined as follows:

(1)$ l_{i}=-\log \frac{\mathrm{e}^{\operatorname{sim}\left(h_{i}^{z_{i}^{\prime}}, h_{i}^{z_{i}^{\prime}}\right) / \tau}}{\sum_{j=1}^{N} \mathrm{e}^{\operatorname{sim}\left(h_{j}^{z_{i}^{\prime}}, h_{j}^{z_{j}^{\prime}}\right) / \tau}}$

(2)$ \operatorname{sim}(\boldsymbol{a}, \boldsymbol{b})=\frac{\sum_{i=1}^{n}\left(a_{i} \cdot b_{i}\right)}{\sqrt{\sum_{i=1}^{n} a_{i}^{2}} \cdot \sqrt{\sum_{i=1}^{n} b_{i}^{2}}}$

Here, $ \operatorname{sim}\left(h_{i}^{z_{i}}, h_{i}^{z_{i}}\right) $ represents the cosine similarity between two vectors, and $ h_{i}^{z_{i}}$, $ h_{i}^{z_{i}^{\prime}}$ represents the embedding of positive and negative samples. n represents the dimensionality of the embeddings, and τ is used to adjust the model's attention to difficult samples. The smaller τ is, the more attention is given to difficult negative samples with very high similarity. τ is referred to as the temperature constant, l_i is the contrastive loss value for the a sample, a_i and b_i are the elements of vectors a and b in the cosine-similarity computation.

Although SimCSE can obtain word and sentence vector embeddings without the need for annotated data on lexical similarity, the vocabulary scale constructed in Section 3.1.1 is limited, and it is difficult to fine-tune and validate the model without annotated data. Therefore, in this section, we use an incremental training approach to train the model, as shown in Fig. 4. First, the model is trained on a publicly available dataset to obtain a well-performing Model 1 based on large-scale data training. Then, incremental training and fine-tuning of Model 1 are performed on the vocabulary constructed in Section 3.1.1. Finally, word vector embeddings specific to the equipment maintenance domain are obtained to measure lexical similarity.

3.1.3. Graph normalization and query optimization based on SimCSE

On the basis of the aforementioned corpus and model, a process is proposed for normalizing node terminology in the constructed graph and optimizing the input queries. This process primarily includes synonym replacement based on the domain synonym library and similarity calculation and replacement based on SimCSE. To ensure the accuracy of the graph normalization, manual verification of the nodes to be merged based on the similarity calculation is needed. The SimCSE-based graph normalization process is detailed in Algorithm 1. This process normalizes similar faults extracted from adjacent locations in the fault repair record, as shown in Fig. 5, ultimately resulting in a IDS-KG-QA.

In addition, the input queries need to be normalized. The process of query normalization based on SimCSE is outlined in Algorithm 2. This process involves optimizing the query expressions using synonyms and the normalized maintenance recommendation KG IDS-KG-QA. The goal is to match the query with nodes in the KG, providing a foundation for subsequent maintenance method matching and recommendation.

3.2. Template-based basic question-answer matching

Due to the lack of publicly available intent recognition datasets in the equipment maintenance domain and the relatively fixed scope of domain knowledge, this section adopts a template-based approach for intent recognition in queries. For common knowledge types in nuclear power equipment maintenance—such as querying faults from equipment, causes from faults, and faults from causes—in this work, seven types of query intent are designed, each with a corresponding template. Taking “Equipment-Fault” as an example, Table 4 provides the designed template and a sample query.

Based on the analyzed questions and corresponding input examples, query templates and answer generation templates are designed to facilitate basic answer matching. Using the query “What faults could occur in gearbox?” as an example, it is identified as an “Equipment-Fault” intent, and the system then generates the query template “MATCH (start)-[r:'有故障类型']->(end) WHERE start.name = '齿轮箱' RETURN start, r, end.” Based on the query templates, the surrounding nodes within one hop of the question entity node can be matched to return the corresponding answer in the IDS-KG-QA. Through this method, all basic Q&A tasks are addressed, in addition to maintenance plan matching.

3.3. Maintenance plan reasoning based on RotatE-CEE

To address the lack of direct links between faults and maintenance plans, this section proposes a maintenance plan reasoning method based on RotatE-CEE to recommend potential maintenance plan nodes based on the given queries.

3.3.1. Improved RotatE model

Due to the characteristics of the equipment maintenance KG, many-to-one, one-to-many, and many-to-many relationships are widespread in the graph, and the TransE model [43] often struggles to generate sufficiently discriminative embeddings—particularly for symmetrical or inverted fault relationships [29]. By performing rotational transformations in the complex plane, RotatE is better able to capture a variety of relational patterns [42]. Therefore, in this section, the RotatE model is adopted for KG embedding of the equipment maintenance KG.

RotatE is also a distance translation model [51], unlike TranE, RotatE proposes a method to map entities and relations to the complex plane. Based on the geometric interpretation of the complex plane, relations are defined as rotations from the head entity to the tail entity, thus avoiding the limitations of the TranE model in mapping complex relations. Specifically, for triple (h,r,t) in the KG, it is expected to have an embedding on the complex plane as follows:

(3)$ \boldsymbol{t}=\boldsymbol{h} \circ \boldsymbol{r}, \text { where }\left|r_{i}\right|=1 \text { and } \boldsymbol{h}, \boldsymbol{r}, \boldsymbol{t} \in \mathbb{C}^{d}.$

Here h, r, and t are all d-dimensional complex vectors. h and t are the mappings to the complex embedding vector of head and tail. ◦ represents the Hadamard product, that is, the element-wise multiplication of matrices. $ \mathbb{C}^{d}$ is the d-dimensional complex vector space. Therefore, for the embedding of entities and relations in each dimension, the following can be obtained:

(4)$t_{i}=h_{i} r_{i}, \text { where }\left|r_{i}\right|=1 \text { and } h_{i}, r_{i}, t_{i} \in \mathbb{C}.$

According to the definition of complex numbers, $r_{i}=\mathrm{e}^{i \theta_{r, i}}$, which means rotating the vector counterclockwise around the origin of the complex plane by an angle $\theta_{r, i}$. Therefore, the relationship can be seen as a rotation process from the head entity to the tail entity. Based on the above definition of triplet embedding, RotatE defines the distance function as follows:

(5)$f_{\mathrm{r}}(\boldsymbol{h}, \boldsymbol{t})=\|\boldsymbol{h} \circ \boldsymbol{r}-\boldsymbol{t}\|$

where f_r(⋅) is the distance function in RotatE.

Therefore, the distance defined by RotatE in the geometric plane can be represented as shown in Fig. 6.

(1) Self-adversarial negative sampling. To ensure that the model learns in the correct direction of the connections, the RotatE model introduces negative sampling during the training process. The model randomly replaces the head entity or tail entity from the correct triplets to generate a set of negative samples S′, where the distance of triplets in the negative samples serves as the negative term of the loss function.

Furthermore, to ensure that negative sampling remains effective as training progresses, RotatE introduces a method called self-adversarial negative sampling. The distribution of negative samples is defined as:

(6)$p\left(h_{j}^{\prime}, \boldsymbol{r}, t_{j}^{\prime} \mid\left\{\left(h_{i}, r_{i}, t_{i}\right)\right\}\right)=\frac{\exp \alpha f_{\mathrm{r}}\left(h_{j}^{\prime}, t_{j}^{\prime}\right)}{\sum_{i} \exp \alpha f_{\mathrm{r}}\left(h_{i}^{\prime}, t_{i}^{\prime}\right)}$

where α is a hyperparameter of the model, and represents the sampling temperature. From the equation above, it can be seen that the distribution of negative samples is influenced by the embeddings of entities and relations learned by the model, thereby ensuring that effective negative samples can still be collected as training progresses.

Ultimately, this distribution of negative samples is used as the weight of the negative samples in the loss function, making the loss function of RotatE as:

(7)$L=-\log \left(\sigma\left(\gamma-f_{\mathrm{r}}(\boldsymbol{h}, \boldsymbol{t})\right)\right)-\sum_{i=1}^{n} p\left(h_{i}^{\prime}, \boldsymbol{r}, t_{i}^{\prime}\right) \log \left(\sigma\left(f_{\mathrm{r}}\left(h_{i}^{\prime}, t_{i}^{\prime}\right)-\gamma\right)\right)$

where γ represents the fixed boundary distance, σ is the sigmoid function, and $f_{\mathrm{r}}\left(h_{i}^{\prime}, t_{i}^{\prime}\right)$ represents the distance of negative sample.

(2) Subsampling of low-frequency samples. In this section, in addition to the self-adversarial negative sampling method, due to the uneven distribution of triple relationships in IDS-KG-QA, a method for subsampling low-frequency samples is designed based on self-adversarial negative sampling. For each triple (h,r,t), its sampling weight is defined as $P\left(w_{(h, r, t)}\right)$, as shown in Eq. (9):

(8)$w_{(h, r, t)}=\sum(h, r, i)+\sum(j, r, t)$

(9)$P\left(w_{(h, r, t)}\right)=\sqrt{\frac{1}{w_{(h, r, t)}}}$

where $\sum(h, r, i)$ represents the number of occurrences of the head entity h and relation r appearing in the same triple, and $\sum(j, r, t)$ represents the number of occurrences of the tail entity t and relation r appearing in the same triple. By substituting the sampling weight into the loss calculation of a batch of triples, the loss function of RotatE can be rewritten as shown in Eq. (14).

(10)$p_{\text {score }}=\log \left(\sigma\left(\gamma-f_{\mathrm{r}}(\boldsymbol{h}, \boldsymbol{t})\right)\right)$

(11)$n_{\mathrm{score}}=\sum_{i=1}^{n} p\left(h_{i}^{\prime}, \boldsymbol{r}, t_{i}^{\prime}\right) \log \left(\sigma\left(f_{\mathrm{r}}\left(h_{i}^{\prime}, t_{i}^{\prime}\right)-\gamma\right)\right)$

(12)$\text{p}\_\text{sample} _{\text {loss }}=-\frac{\sum P\left(w_{(h, r, t)}\right) p_{\text {score }}}{\sum P\left(w_{(h, r, t)}\right)}$

(13)$\text{n}\_\text{sample}_{\text {loss }}=-\frac{\sum P\left(w_{\left(h^{\prime}, r, t^{\prime}\right)}\right) n_{\text {score }}}{\sum P\left(w_{\left(h^{\prime}, r, t^{\prime}\right)}\right)}$

(14)$\text{loss} =\frac{\left(\text { p_sample }{ }_{\text {loss }}+\text { n_sample }{ }_{\text {loss }}\right)}{2}$

where P_score is the positive-sample log-likelihood term for a true triple (h,r,t), n_score is the negative-sample term aggregated over n negative triples, p_sample_loss is the batch-level positive loss reweighted by the subsampling weights $ P\left(w_{(h, r, t)}\right)$, n_sample_loss is the batch-level negative loss reweighted by the same subsampling weights $ P\left(w_{(h, r, t)}\right)$, and loss is the final training objective.

According to the above equation, the higher the frequency of data related to the triple’s head and tail is, the smaller the sampling weight. Consequently, the proportion of the score of this triple decreases. Therefore, the method can effectively reduce the gap between high-frequency and low-frequency data.

3.3.2. Maintenance plan reasoning with fault causality and equipment structure enhancement

As depicted in Fig. 7, the IDS-KG-QA encompasses three types of nodes, which constitute the equipment set, fault set, and maintenance plan set. Given that IDS-KG-QA integrates a variety of equipment maintenance knowledge, this section simulates the thought process of operators from fault diagnosis to the derivation of a maintenance plan. Based on IDS-KG-QA, three reasoning paths for the design of maintenance plans are developed:

Path 1: Extract the fault directly from the query question and find the corresponding maintenance plans for that fault within a single-hop from the fault node, as shown in Path 1 of Fig. 7.

Path 2: Extract the fault directly from the query question and, through the fault causality chain, find the nearest fault cause with a maintenance plan; then, return the maintenance plans. This path begins with the original fault node, searches within the mar₁ hop range for fault nodes that have a “cause” relationship with the initial fault, and then compiles the associated PlanNodes into a PlanNodeSet. In doing so, it enables multi-hop reasoning for maintenance plan discovery based on fault causality, as shown in Path 2 of Fig. 7.

Path 3: Extract the equipment and fault from the query question, find similar faults from adjacent equipment, and then find the corresponding maintenance plans for that fault. This path starts from the original fault node, identifies the equipment containing that fault, and then searches within the mar₂ hop range from that equipment node for additional EquipNodes linked by an “isPartOf” relationship. Next, it determines the fault set of these adjacent equipment nodes and retrieves all PlanNodes linked by a “hasPlan” relationship to form a PlanNodeSet, enabling multi-hop inference across neighboring equipment, as shown in Path 3 of Fig. 7.

Based on these paths for discovering maintenance plans, the following problems exist with the template-based question-answer matching method:

Problem 1: The one-hop question matching method leads to the failure to discover maintenance plan nodes on the fault causality chain.

Problem 2: The reference role of adjacent equipment structures in the KG for maintenance plans is ignored.

Problem 3: When possible maintenance plans are recommended, measuring the distance between faults and possible maintenance plan nodes is difficult.

Based on the characteristics of the above problems, this section proposes a maintenance plan reasoning method based on RotatE-CEE. First, search for the corresponding nodes of faults and devices in the query are searched, and for nouns not in the node set of the KG, the SimCSE model is used to optimize the query statement to find the node that is semantically closest to the noun. Then, based on the graph search, the subgraph corresponding to the faults and devices in the query is searched, and possible maintenance plan nodes are inferred from both the fault causality chain and the equipment composition architecture. Finally, based on the RotatE model, the similarity between the inferred maintenance plan nodes and the query nodes is calculated, the distance between nodes through the KG structure is measured, and the most likely maintenance plan is returned. The specific process is shown in Algorithm 3.

4. Case study

This section first validates the ability of the proposed method to recognize synonyms and produce high-quality KG embeddings in the nuclear power equipment maintenance domain through the graph normalization experiment and maintenance plan reasoning experiment. Subsequently, Section 4.3 presents a series of model comparison experiments, which demonstrate the superiority of our maintenance plan recommendation approach in this domain.

All the experiments were performed on a machine equipped with an Intel i7-7700HQ CPU (Intel, USA) and a single NVIDIA Quadro T2000 GPU (NVIDIA, USA).

4.1. Graph normalization experiment

4.1.1. Datasets

The data used for training the SimCSE model in this section are divided into two categories: publicly available datasets and nuclear power equipment maintenance domain datasets [10], which are respectively used for the first-stage training and the second-stage incremental training of the SimCSE model, respectively.

For the publicly available datasets, the semantic textual similarity-B (STS-B) [53] and stanford natural language inference (SNLI) [54] datasets are selected. STS-B is a dataset used for text similarity training, whereas SNLI is used for natural language inference. Therefore, STS-B, which has similarity annotations, is used for model validation and fine-tuning, whereas SNLI enriches the data representation of the training dataset.

The datasets selected for the incremental training phase are from nuclear power equipment maintenance domain datasets [10], which focus on sentence-level similarity training and vocabulary-level similarity training.

The nuclear power equipment maintenance domain dataset comprises two data types—“Experience Feedback” and “Fault Repair Record”—collected from three nuclear power plants.

They are annotated at both the entity and relationship levels, following the annotation logic typical of open datasets. Specifically, “Experience Feedback” contains 89 raw entries, with an average of 327 characters per record, while “Fault Repair Record” consists of 154 raw entries, averaging 102 characters per record. The original records and collection templates are provided in Table 5, Table 6.

4.1.2. Graph normalization experiment

The word vectors trained by the SimCSE model are used for normalizing queries and the KG in this paper. To address the lack of annotated data, the SimCSE model is trained using a combination of publicly available data and domain-specific incremental training. In this article, the temperature hyperparameter is τ=0.05, the learning rate is lr=1×10^-5, and the dropout is set to d=0.3. For training, the number of training epochs and batch size is set to E=1 and B=64, respectively.

Given that the nuclear power complex equipment maintenance domain dataset in this section lacks similarity annotated data, performance validation of the model after incremental training is conducted using a synonym list. For each synonymous word pair wordS_i, SynWord_ij and word vector pair wordS_i_emb, SynWord_ij_emb are obtained using the SimCSE model. The similarity between the standard word and its synonym is subsequently calculated based on Eq. (15), and the average is computed. Since wordS_i,SynWord_ij represents a pair of synonymous words, the closer SynMean is to 1, the better the performance of the SimCSE model.

(15)$ \text { SynMean }=\sum_{i=1}^{M} \frac{\sum_{j=1}^{m} \operatorname{sim}\left(\text { wordS }_{i \text { enb }}, \operatorname{SynWord}_{i j_{\text {emb }}}\right) / g_{i}}{G}$

where g_i represents the number of synonyms for the standard word wordSi, and G represents the total number of synonym groups.

The experimental results indicate that after the model undergoes incremental training using nuclear power equipment maintenance domain data, the performance of domain synonym identification has improved to a certain extent. Specifically, when SimCSE is used for synonym recognition, the SynMean score increases from 0.8573 with the original SimCSE model to 0.8720 after incremental training.

By applying Algorithm 1 to normalize and merge nodes in the equipment maintenance KG IDS-KG, we finally obtain the KG IDS-KG-QA tailored for equipment maintenance plan recommendations. The graph size is optimized from 1017 nodes and 1671 relationships to 835 nodes and 1261 relationships. Compared with that before normalization, the node count decreases by 17.80%, and the relationship count decreases by 24.52%.

4.2. Maintenance plan reasoning experiment

4.2.1. Datasets and evaluation metrics

In this section, the nodes in IDS-KG-QA are used as the training dataset for the model. All the triples from the constructed IDS-KG-QA are exported, each node is encoded, and the triples are then rewritten in a format based on these node encodings.

The IDS-KG-QA contains a total of 835 nodes and 1261 relationships, comprising 471 labeled “cause,” 304 labeled “hasFault,” 270 labeled “hasOperation,” and 216 labeled “isPartOf.” These triples are proportionally split into training, validation, and test sets, resulting in 882 triples for training, 190 for validation, and 189 for testing. The underlying information is drawn from one year of accumulated experience feedback and maintenance procedures at three nuclear-power plants, ensuring that the dataset reflects the diversity of real-world maintenance practices. With respect to the graph complexity, a breadth first search (BFS) from every fault node shows that 83.15% of faults can reach at least one operation within 6 hops, and the average shortest distance is 3.13 hops. Among those shortest <fault, operation> paths, 76.76% involve a “cause” edge and 12.98% involve an “isPartOf” edge. These figures demonstrate that IDS-KG-QA and the query set derived from it genuinely require multi-hop reasoning, making the dataset suitable for evaluating maintenance plan reasoning methods.

Common evaluation metrics for KG embeddings include mean reciprocal ranking (MRR), mean rank (MR), and HITS@n.

To evaluate the results of KG embeddings, for the triple $ \left(h_{i}, r_{i}, t_{i}\right)$, by replacing the tail entity with any other entity, if there are q nodes in the graph, (q-1) new triples will be obtained. The distance between the original triple and the (q-1) new triples is calculated on the basis of Eq. (5), which are arranged in ascending order to obtain the ranked distance matrix RankedDistance, and the ranking of the true triple $ \left(h_{i}, r_{i}, t_{i}\right)$ within RankedDistance is determined as ranki. Based on the definition of ranki, detailed definitions of the above evaluation metrics are provided.

(16)$ \operatorname{MRR}=\frac{1}{|S|} \sum_{i=1}^{|S|}\left(\frac{1}{\operatorname{rank}_{1}}+\frac{1}{\operatorname{rank}_{2}}+\ldots+\frac{1}{\operatorname{rank}_{|S|}}\right)$

where S represents the set of triples, $ \left | S \right | $ represents the number of triples; this metric reflects the degree to which the embedding results align with the actual scenarios, with a larger MRR indicating better embedding effectiveness.

(17)$ \mathrm{MR}=\frac{1}{|S|} \sum_{i=1}^{|S|} \operatorname{rank}_{i}=\frac{1}{|S|}\left(\operatorname{rank}_{1}+\operatorname{rank}_{2}+\ldots+\operatorname{rank}_{|S|}\right)$

The symbols in this formula have the same meaning as MRR, with a smaller MR indicating better embedding effectiveness.

(18)$ \operatorname{HITS} @ n=\frac{1}{|S|} \sum_{i=1}^{|S|} \mathrm{II}\left(\operatorname{rank}_{i} \leqslant n\right)$

where II(·) represents the indicator function: if the condition is true, the function value is 1; otherwise, it is 0. Typically, values of 1, 3, and 10 correspond to the metrics HITS@1, HITS@3, and HITS@10, respectively. For example, HITS@3 means that the triple with ranki within the top 3 ranks is assigned a value of 1. A larger HITS@n value indicates better embedding effectiveness.

4.2.2. Comparative experiment on sampling strategies in the RotatE model

This section compares the training process and results of the RotatE model under four different scenarios: not using any special sampling strategy, solely employing low-frequency sample subsampling, solely utilizing self-adversarial negative sampling, and simultaneously employing both sampling strategies.

In Fig. 8, the MR and HITS@10 metrics on the validation set during the training process are compared. It can be observed that solely using low-frequency sample subsampling results in better metrics than not using any sampling strategy, as it considers the distribution of low-frequency samples, but the model convergence is slower. Using only self-adversarial negative sampling allows the model metrics to reach optimal values quickly, but since it ignores low-frequency samples, the performance does not significantly improve during training. In contrast, simultaneously employing both sampling strategies ensures the effectiveness of negative sample collection during training while considering the distribution of low-frequency samples, leading to better performance metrics. The training results of the models using different sampling strategies are illustrated in Fig. 9, which show improvements in MRR, HITS@1, HITS@3, HITS@10, and MR values.

4.2.3. Comparison with alternative KG embedding models

In this section, the improved RotatE implemented on our maintenance KG is evaluated by comparing it with two alternative embedding models: CompGCN [55], which is a primarily structure-based approach, and SimKGC [56], which integrates the pre-trained language model. Using the same dataset described in Section 4.2.1, the Hits@n scores are reported in Table 7 [55], [56].

(1) CompGCN [55]. This graph-convolutional model is designed to capture neighborhood patterns but works best when the KG exhibits clear, cluster-like relational communities. In IDS-KG-QA most relations are 1-to-1 or 1-to-N and form only weak communities, which leads to lower scores. Moreover, CompGCN requires more than four times as many trainable parameters as RotatE does, resulting in a significantly higher deployment cost on our limited GPU budget.

(2) SimKGC [56]. SimKGC enriches entity representations with a pre-trained language model. Because IDS-KG-QA contains only short textual labels, this linguistic advantage is greatly reduced. The additional semantics help the model retrieve more distant candidates, yielding a higher Hits@10, but the lack of strong geometric constraints makes the within-set ranking imprecise, so Hits@1 remains low.

(3) RotatE. As a translational KGE model, RotatE aligns naturally with the causal chains present in IDS-KG-QA and attains the best Hits@1 and competitive Hits@10, while using far fewer trainable parameters than the other two baselines do.

In summary, the KGE component in CEE-MPMR is pluggable, and practitioners can swap in any embedding model that better suits their graph. Nonetheless, in the specific case of IDS-KG-QA, RotatE remains the most appropriate choice, and we therefore retain the enhanced RotatE variant as the default in our system.

4.2.4. Parameter selection experiment for the RotatE model

To identify the optimal combination of hyperparameters for the model, this section compares model performance under different hyperparameter settings. The choice of batch size depends on the GPU performance and the size of the training data. In this study, a batch size of 128 and a negative sample size of 32 were selected based on the available GPU. MRR and HITS@10 were chosen as the evaluation metrics for model performance.

In this context, the margin distance gamma refers to the margin γ used in adversarial negative sampling as described in Eq. (7), and the adversarial temperature refers to the sampling temperature used in adversarial negative sampling α, as described in Eq. (6). The embedding dimension refers to the specified dimension d of the triplet embedding, as outlined in Eq. (3). As shown in Fig. 10, the model achieved the best performance when the margin distance gamma was set to 4, the adversarial temperature was set to 1, and the embedding dimension was set to 700. The final selected hyperparameter values are presented in Table 8. When the model trained with these parameters was used, the final HITS@10 reached 0.458, and the MRR value reached 0.362.

4.2.5. Maintenance plan recommendation results

To validate the effectiveness of the RotatE-CEE based maintenance plan reasoning method proposed in Section 3, fault records lacking “maintenance plan” from Section 3.1.1 of the “Fault Repair Record” are selected for maintenance plan reasoning in Fig. 11.

Path 1: Taking the “impeller mouth ring runout” fault as an example, the subgraph within one hop of this node is shown in Fig. 11(a-i). This fault has a directly corresponding maintenance plan, “replace the impeller,” which is returned as the corresponding maintenance plan node in the subgraph, referred to as Path 1 in Fig. 7.

Path 2: For the “bearing wear” fault, the subgraph within one hop of this node is shown in Fig. 11(a-ii). It can be seen that there is no directly corresponding maintenance plan for this fault, only one possible cause. Using the reasoning of Algorithm 3, the most likely maintenance plan is derived from the fault causality chain for the “bearing wear” fault. With a boundary parameter mar₁=3 and an answer range N_ans = 3, the recommended maintenance plan for “bearing wear” is “repair and straighten the pump shaft”. The reasoning path is shown in Fig. 11(b-ii), corresponding to Path 2 in Fig. 7.

Path 3: Again, using the “bearing wear” fault as an example, the subgraph within one hop of this node is shown in Fig. 11(b-i). In addition to similar fault maintenance plans in the fault causality chain, the reasoning of Algorithm 3 can also recommend potential maintenance plans from nearby equipment related to the “bearing” with a device boundary parameter mar₂=1 and an answer range N_ans = 3; the reasoning path is shown in Fig. 11(b-iii). Through related faults of the adjacent equipment “pump shaft” and its fault causality chain, combined with KG embedding results, the recommended maintenance plan is “replace the pump shaft and impeller” corresponding to Path 3 in Fig. 7.

4.3. Validation and comparison of maintenance plan recommendation methods

To validate the effectiveness of the proposed maintenance plan recommendation method, all fault nodes from the IDS-KG-QA dataset were extracted, and random queries were generated to determine the corresponding maintenance plans using the proposed reasoning method.

Specifically, 426 fault nodes are selected and 402 natural-language queries are generated by slot-filling and synonym substitution, thereby emulating the questions that technicians are likely to type in practice. The average query length is 13.89 Chinese words. The correctness of the answers returned by each method was evaluated manually by 2 nuclear power maintenance engineers, ensuring every query has a gold answer.

Based on these queries, comparisons were conducted among the template-based query method, the TransE-CEE-based method, the CEE-based method, the RotatE-CEE-based method, and the ChatGLM−6B-based [57] QA method. The comparison focused on the returned answers and the number of valid responses provided by each method.

The comparison models for Table 9 [41], [55], [56], [57] are described as follows:

Comparison Model 1: removes the reasoning phase and degenerates into a template-based query.

Comparison Model 2: omits the ranking stage from Fig. 7 and retains only subgraph searches for fault causality and equipment structure, returning the top N_ans answers.

Comparison Model 3: replaces the improved RotatE in Fig. 7 with different KGE models, including TransE [43], CompGCN [55] and SimKGC [56].

This section also includes a comparison with the ChatGLM−6B large language model [57], which can be deployed offline. While large language models have shown strong performance in QA tasks, nuclear power plants must often rely on smaller-scale offline models to protect sensitive operational data and accommodate hardware constraints. ChatGLM−6B is an open-source model developed using the GLM architecture, it can run on a standard GPU with approximately 6 GB of Video Memory (VRAM), making it feasible for deployment in real nuclear power environments. Compared with other large models that require more specialized hardware or continuous online connectivity, ChatGLM−6B offers a practical balance of performance and resource requirements, thus serving as a suitable comparison model for this system.

For each fault, different methods were used to recommend maintenance plans, and the number of valid responses was manually verified. The total number of returned answers, valid responses, ratio, and average response time are shown in Table 9. All the models are implemented in PyTorch. The Average response time reported in Table 9 are the average times taken from question input to answer generation. Except ChatGLM−6B-based method, these times were measured on a workstation equipped with an i7-7700HQ CPU and a single NVIDIA Quadro T2000 GPU.

The maintenance plan recommendation method based on ChatGLM−6B, on the other hand, achieved a much higher response rate of 97.26%. As a generative large language model, ChatGLM−6B generates coherent and logically structured text for all input queries, leading to a high response rate [58], [59]. However, due to the lack of fine-tuning with knowledge specific to nuclear power equipment maintenance, the model struggles to recognize queries with specialized expressions, such as equipment codes, and lacks systematic knowledge of nuclear power equipment architecture. As a result, its accuracy is relatively low. Furthermore, ChatGLM−6B has a slower response time compared to other methods. In this study, the INT4 version of the model, which has the lowest hardware requirements and was tested on a T4 GPU with 16 GB of VRAM in Colab, was used to simulate the hardware state in real-world scenarios. Despite this hardware configuration being superior to the workstation used by other methods, its average response time of 40.73 seconds was significantly higher than that of the other methods.

The template-based method achieved a response rate of 40.30% and an accuracy rate of 39.05%. Although both rates are relatively low, the responses generated by this method are mostly valid because the template-based approach returns maintenance plans that are directly connected to fault nodes in the KG, resulting in a close alignment between the response rate and accuracy.

The CEE-based maintenance plan recommendation method achieved a response rate of 84.08% and an accuracy rate of 60.95%. Its higher response rate than the template-based approach can be attributed to the inclusion of subgraph paths from fault causality and equipment structure during the reasoning phase. This allows the system to discover maintenance methods beyond those directly related to a given fault, including remedies for adjacent and contextually connected faults. However, without a ranking stage, the top N_ans answers often include redundant information, reducing the method’s overall usability compared with approaches that incorporate a ranking component.

Using CompGCN, SimKGC, and TransE as the embedding module increases both the response rate and the accuracy rate relative to the CEE-based method, but each still falls short of the RotatE-CEE, mirroring the result in Table 7. CompGCN is handicapped by its reliance on dense neighborhood structure, and the short textual labels in IDS-KG-QA reduce SimKGC’s semantic advantage. Furthermore, SimKGC requires BERT to be invoked at inference time to encode entity text, which increases latency given our limited GPU resources. In contrast, both TransE and RotatE can serve prebuilt embeddings directly from GPU memory, enabling them to respond more quickly. All the models are implemented in PyTorch. The average response time reported in Table 9 are the average times taken from question input to answer generation. These times were measured on a workstation equipped with an i7-7700HQ CPU and a single NVIDIA Quadro T2000 GPU.

In contrast, the maintenance plan recommendation method incorporating RotatE-CEE achieves both a high response rate and high accuracy. While its response rate is slightly lower than that of ChatGLM−6B, the generative nature of ChatGLM−6B allows it to return higher response rates, even for nontypical nuclear power equipment faults that are not recorded in the IDS-KG-QA. However, the accuracy rate of the RotatE-CEE-based approach is significantly higher than that of ChatGLM−6B and the other methods, largely because of its integration of equipment structures and fault causality. By leveraging RotatE-CEE, the system measures the similarity between the query’s intended content and the returned nodes, thereby offering users more reliable maintenance information. With respect to response time, the inclusion of the RotatE model makes the RotatE-CEE-based method slightly slower than the template-based and CEE-based approaches. Nevertheless, it still operates at approximately seven seconds, which is acceptable in an offline nuclear power equipment maintenance context.

In summary, the proposed method achieves the highest accuracy among the compared methods. Regarding the response rate, the generative LLM-based method (ChatGLM-6B) can always produce an answer, even when the KG lacks a relevant maintenance plan, so its response rate reaches 97.26%. However, because the model lacks specialized domain knowledge, many of its replies are hallucinations, yielding an accuracy of only 49.50%. The proposed method answers 94.53% of the queries, and the missing 2.73% corresponds mainly to atypical nuclear equipment faults whose device or fault code does not exist in IDS-KG-QA. As IDS-KG-QA is derived from historical experience and maintenance manuals, it omits entities never encountered before, in those cases our system returns no answer rather than an unreliable one. This phenomenon is shown in Table 10 [41], [55], [57]. Nevertheless, it still reaches an accuracy of 90.80%, confirming that the strong reliance on IDS-KG-QA makes the method well suited to the nuclear maintenance domain. As for response time, the KG embedding and ranking step makes our average latency 0.7 s slower than the template baseline; however, we consider this overhead acceptable because it increases both the response rate and accuracy.

A comparison of the output answers for different types of queries by various methods is shown in Table 10 [41], [57].

For standard equipment fault questions, such as “What should be done if the threads are damaged?”, the template-based method was unable to return an answer because the IDS-KG-QA does not contain directly associated knowledge for the fault. ChatGLM−6B generated several general maintenance plans, with some usable but not optimal content. The CEE-based method returns the top N_ans answers by searching for fault causality and equipment structure, but in the absence of a ranking phase, its results include irrelevant information. The TransE-based method returned unusable answers because of the inaccurate embedding of some nodes and relationships. In contrast, the RotatE-CEE-based method provided the most precise maintenance plans.

For typical nuclear power equipment fault questions, such as “What should be done if the packing gland is smoking?”, the template-based, CEE-based, TransE-based, and RotatE-CEE-based methods were all able to return specific maintenance plans. However, ChatGLM−6B, lacking a complete knowledge base for such specialized issues in the nuclear power domain, returned unusable answers.

For atypical nuclear power equipment fault questions, such as “What should be done if the chuck is overtightened?”, none of the template-based, CEE-based, TransE-based, or RotatE-CEE-based methods were able to infer or discover corresponding maintenance plans. The reason for this is that the equipment entity “chuck” mentioned in the query is not a typical component of nuclear circulating water pumps and was therefore not included in the IDS-KG-QA. This finding indicates that the equipment model architecture of the nuclear power equipment maintenance ontology does not cover the entity in the query, making it impossible to return an answer. ChatGLM−6B, however, generated a relatively general maintenance plan for this fault, with some usable content, making it the most optimal answer compared to the others.

4.4. Nuclear power key equipment maintenance plan recommendation system

Based on the above research, this paper developed a maintenance plan recommendation system for key nuclear power equipment. The system integrates the IDS-KG-QA and CEE-MPMR. It allows users to input questions in natural language, and the system provides the most likely maintenance plan, displayed below the search bar. Additionally, the system presents the reasoning path between the fault and the recommended maintenance plan, helping users understand the logic behind the recommendation. As shown in Fig. 12(a), when a user asks, “What to do if the pump shaft is worn? ” the system invokes the CEE-MPMR method to recommend maintenance plans based on the knowledge from IDS-KG-QA, providing two potential plans. The reasoning path for the maintenance plan recommendation is shown in Fig. 12(b). In addition to maintenance method queries, this module also supports other types of questions, such as equipment composition and fault causes.

Moreover, the answers returned by the system are validated by the user, as shown in Fig. 12(a). If a user is dissatisfied with an answer returned by CEE-MPMR, they can click “Wrong answer,” and the system then records the unsatisfactory Q&A pair, which is periodically reviewed by experts. Similar to the above procedure, any supplemental or corrected knowledge provided is extracted and incorporated into the IDS-KG.

If no answer is returned by the CEE-MPMR, the system will prompt the user to supply additional expert knowledge, as shown in Fig. 1. In the system interface displayed in Fig. 12(a), users with the appropriate expertise can click “Add new knowledge” to input new domain knowledge. After a knowledge extraction process, this information is added to the IDS-KG. The KG subsequently undergoes graph normalization before it is merged into the IDS-KG-QA. Only authorized experts can add new knowledge, ensuring that any modifications remain reliable.

5. Conclusion

To address the issue of low accuracy in nuclear power equipment maintenance plan matching, this paper proposes a CEE-MPMR method. The main contributions of this paper are as follows:

To address the low standardization of knowledge in equipment maintenance records, we introduce an unsupervised SimCSE-based domain vocabulary normalization method. Comparing with the original method, the proposed method effectively enhances the model’s sensitivity to domain-specific terminology even when domain lexicons are lacking. Based on this, the KG is normalized, reducing the node size by 17.80% and the relationship size by 24.52%, thus effectively avoiding redundancy in the knowledge base.

To address the difficulties in collecting negative samples and insufficient attention to low-frequency samples during the training process of RotatE embeddings, self-adversarial negative sampling and subsampling of low-frequency samples are applied in the training process. The experimental results show that these sampling strategies significantly improve model training efficiency and performance.

To overcome the challenges of multi-hop reasoning in maintenance plan recommendation, we propose a maintenance plan reasoning method based on RotatE-CEE Compared with other KG embedding-based QA methods, it emulates the reasoning process of maintenance personnel to subgraph-search candidate answers, effectively narrowing the subgraph scope and demonstrating stronger adaptability and interpretability in the equipment maintenance domain. Additionally, the improved RotatE model used here effectively represents the logical relationships (e.g., fault causality) in the maintenance KG. The model achieves 90.80% accuracy in response to nuclear power equipment maintenance queries. The case study shows that the proposed method results in greater domain-specific expertise and reliability than small-scale open-source large language models such as ChatGLM-6B, which are more suitable for offline deployment in nuclear power plants.

However, the proposed method still has several limitations: ① certain faults cannot be matched to the recommendations because the equipment ontology is incomplete; and ② numeric-type faults and parameter-setting plans are not yet reflected in real time within the KG.

In future work we will explore combining CEE-MPMR with retrieval-augmented generation (RAG) to increase the response rate while maintaining the reliability of the answers. Specifically, the explicit reasoning paths produced by CEE-MPMR act as the retriever, and the retrieved subgraph, together with real-time sensor values and calibration tables, is fed into a lightweight LLM. This enables the RAG-enhanced model to generate KG-based answers that are consistent with real-time numerical data. It is hoped that this strategy will offer a feasible pathway for deploying future industrial large models in real-world plants, paving the way for truly intelligent, data-driven maintenance.

References

Publishing order | Descend order by publishing year | Descend order by cited within

[1]	Foresti R, Rossi S, Magnani M, Guarino Lo Bianco C, Delmonte N. Smart society and artiﬁcial intelligence: big data scheduling and the global standard method applied to smart maintenance. Engineering 2020; 6(7):835-46.

[2]	Van Dinter R, Tekinerdogan B, Catal C. Predictive maintenance using digital twins: a systematic literature review. Inf Softw Technol 2022; 151:107008.

[3]	Yang J, Liu Y, Morgan PL. Human-machine interaction towards Industry 5.0: human-centric smart manufacturing. Digit Eng 2024; 2:100013.

[4]	Huang C, Bu S, Lee HH, Chan CH, Kong SW, Yung WKC. Prognostics and health management for predictive maintenance: a review. J Manuf Syst 2024; 75:78-101.

[5]	Tao F, Yi L, Wei Y. AI power for digital manufacturing. Digit Eng 2024; 2:100016.

[6]	Gölzer P, Fritzsche A. Data-driven operations management: organisational implications of the digital transformation in industrial practice. Prod Plann Control 2017; 28(16):1332-43.

[7]	Roh JJ, Hong P. Taxonomy of ERP integrations and performance outcomes: an exploratory study of manufacturing ﬁrms. Prod Plann Control 2015; 26 (8):617-36.

[8]	Liu G, Shen W, Gao L, Kusiak A. Knowledge transfer in fault diagnosis of rotary machines. IET Collab Intell Manuf 2022; 4(1):17-34.

[9]	Li P, Cheng K, Jiang P, Katchasuwanmanee K. Investigation on industrial dataspace for advanced machining workshops: enabling machining operations control with domain knowledge and application case studies. J Intell Manuf 2022; 33(1):103-19.

[10]	Wang Y, Cheng Y, Qi Q, Tao F. IDS-KG: an industrial dataspace-based knowledge graph construction approach for smart maintenance. J Ind Inf Integr 2024; 38:100566.

[11]	Zhang Y, Wang H, Shen W, Peng G. DuAK: reinforcement learning-based knowledge graph reasoning for steel surface defect detection. IEEE Trans Autom Sci Eng 2025; 22:557-69.

[12]	Kusiak A. Service manufacturing = process-as-a-service + manufacturing operations-as-a-service. J Intell Manuf 2020; 31(1):1-2.

[13]	Hu Z, Li X, Pan X, Wen S, Bao J. A question answering system for assembly process of wind turbines based on multi-modal knowledge graph and large language model. J Eng Des 2023; 36(7-9):1093-117.

[14]	Zhou B, Li X, Liu T, Xu K, Liu W, Bao J. CausalKGPT: industrial structure causal knowledge-enhanced large language model for cause analysis of quality problems in aerospace product manufacturing. Adv Eng Inform 2024; 59:102333.

[15]	Wille R. Concept lattices and conceptual knowledge systems. Comput Math Appl 1992; 23(6-9):493-515.

[16]	Wang S, Yang J, Yang B, Li D, Kang L. An intelligent quality control method for manufacturing processes based on a human-cyber-physical knowledge graph. Engineering 2024; 41:242-60.

[17]	Suvarna M, Yap KS, Yang W, Li J, Ng YT, Wang X. Cyber-physical production systems for data-driven, decentralized, and secure manufacturing-a perspective. Engineering 2021; 7(9):1212-23.

[18]	Li X, Zheng P, Bao J, Gao L, Xu X. Achieving cognitive mass personalization via the self-X cognitive manufacturing network: an industrial knowledge graph-and graph embedding-enabled pathway. Engineering 2023; 22:14-9.

[19]	Wan Y, Liu Y, Chen Z, Chen C, Li X, Hu F, et al. Making knowledge graphs work for smart manufacturing: research topics, applications and prospects. J Manuf Syst 2024; 76:103-32.

[20]	Zhao M, Wang H, Guo J, Liu D, Xie C, Liu Q, et al. Construction of an industrial knowledge graph for unstructured chinese text learning. Appl Sci 2019; 9 (13):2720.

[21]	Liu C, Yang S. Using text mining to establish knowledge graph from accident/incident reports in risk assessment. Expert Syst Appl 2022; 207:117991.

[22]	Tian J, Song H, Sheng G, Jiang X. An event knowledge graph system for the operation and maintenance of power equipment. IET Gener Transm Distrib 2022; 16(21):4291-303.

[23]	Li X, Zhang F, Li Q, Zhou B, Bao J. Exploiting a knowledge hypergraph for modeling multi-nary relations in fault diagnosis reports. Adv Eng Inform 2023; 57:102084.

[24]

Wen S, Chen Y, Pan X, Zhuang W, Li X. Enhancing fault troubleshooting through human-machine collaboration:a multi-stage reasoning approach. In: Proceedings of the 2024 IEEE 20th International Conference on Automation Science and Engineering (CASE); 2024 Aug 28-Sep 1; Bari, Italy. Piscataway: IEEE; 2024. p. 460-7.

[25]	Liu B, Chen CH, Wang Z. A multi-hierarchical aggregation-based graph convolutional network for industrial knowledge graph embedding towards cognitive intelligent manufacturing. J Manuf Syst 2024; 76:320-32.

[26]	Li Y, Liu X, Starly B. Manufacturing service capability prediction with graph neural networks. J Manuf Syst 2024; 74:291-301.

[27]	Han H, Wang J, Wang X, Chen S. Construction and evoluti on of fault diagnosis knowledge graph in industrial process. IEEE Trans Instrum Meas 2022; 71:1-12.

[28]	Xia L, Liang Y, Leng J, Zheng P. Maintenance planning recommendation of complex industrial equipment based on knowledge graph and graph neural network. Reliab Eng Syst Saf 2023; 232:109068.

[29]	Hogan A, Blomqvist E, Cochez M, de Melo G, Gutierrez C, Kirrane S, et al. Knowledge graphs. ACM Comput Surv 2022; 54(4):1-37.

[30]	Huang X, Zhang J, Li D, Li P. Knowledge graph embedding based question answering. In:Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining; 2019 Feb 11-15; Melbourne, VIC, Australia. New York City: Association for Computing Machinery; 2019. p. 105-13.

[31]	Liu L, Du B, Ji H, Zhai C, Tong H. Neural-answering logical queries on knowledge graphs. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining; 2021 Aug 14-18; Singapore. New York City: Association for Computing Machinery; 2021. p. 1087-97.

[32]	Lan Y, He G, Jiang J, Jiang J, Zhao WX, Wen JR. Complex knowledge base question answering: a survey. IEEE Trans Knowl Data Eng 2023; 35 (11):11196-215.

[33]	Berant J, Chou A, Frostig R, Liang P. Semantic parsing on freebase from question-answer pairs. In: Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing; 2013 Oct 18-21; Washington, DC, USA. Stroudsburg: Association for Computational Linguistics; 2013. p. 1533-44.

[34]	Liang P, Jordan MI, Klein D. Learning dependency-based compositional semantics. Comput Linguist 2013; 39(2):389-446.

[35]

Perez-Beltrachini L, Jain P, Monti E, Lapata M. Semantic parsing for conversational question answering over knowledge graphs. In: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics; 2023 May 2-6; Dubrovnik, Croatia. Stroudsburg: Association for Computational Linguistics; 2023. p. 2507-22.

[36]	Dong L, Lapata M. Language to logical form with neural attention. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); 2016 Aug 7-12; Berlin, Germany. Strou dsburg: Association for Computational Linguistics; 2016. p. 33-43.

[37]	Peng Y. Query-driven knowledge graph construction using question answering and multimodal fusion. In:Proceedings of the Companion Proceedings of the ACM Web Conference 2023; 2023 Apr 30-May 4; Austin, TX, USA. New York City: Association for Computing Machinery; 2023. p. 1119-26.

[38]	Zhu X, Gao W, Li T, Yao W, Deng H. Event-centric hierarchical hyperbolic graph for multi-hop question answering over knowledge graphs. Eng Appl Artif Intell 2024;133(Pt B):107971.

[39]

Galárraga LA, Teflioudi C, Hose K, Suchanek F. AMIE:association rule mining under incomplete evidence in ontological knowledge bases. In:Proceedings of the 22nd International Conference on World Wide Web; 2013 May 13-17; Rio de Janeiro, Brazil. New York City: Association for Computing Machinery; 2013. p. 413-22.

[40]	Lao N, Cohen WW. Relational retrieval using a combination of path-constrained random walks. Mach Learn 2010; 81(1):53-67.

[41]	Schlichtkrull M, Kipf TN, Bloem P, van den Berg R, Titov I, Welling M. Modeling relational data with graph convolutional networks. In: Gangemi A, Navigli R, Vidal ME, Hitzler P, Troncy R, Hollink L, The semantic web. Cham: Springer; 2018. p. 593-607.

[42]	Huang Q, Li G. Knowledge graph based reasoning in medical image analysis: a scoping review. Comput Biol Med 2024; 182:109100.

[43]

Bordes A, Usunier N, Garcia-Durán A, Weston J, Yakhnenko O. Translating embeddings for modeling multi-relational data. In: Proceedings of the 27th International Conference on Neural Information Processing Systems; Lake Tahoe, NV USA. New York City: Association for Computing Machinery; 2013. p. 2787-95.

[44]

Nickel M, Tresp V, Kriegel HP. A three-way model for collective learning on multi-relational data. In: Proceedings of the 28th International Conference on International Conference on Machine Learning; 2011 Jun 28-Jul 2; Washington, DC USA. New York City: Association for Computing Machinery; 2011. p. 809-16.

[45]	Ren H, Hu W, Leskovec J. Query2box:reasoning over knowledge graphs in vector space using box embeddings. In: Proceedings of the 2020 International Conference on Learning Representations, ICLR 2020; 2020 Apr 26-30; Addis Ababa, Ethiopia. ICLP; 2020.

[46]	Hamilton W, Bajaj P, Zitnik M, Jurafsky D, Leskovec J. Embedding logical queries on knowledge graphs. In:Proceedings of the 32nd International Conference on Neural Information Processing Systems; 2018 Dec 3-8; Montréal, Canada. New York City: Curran Associates Inc.; 2018. p. 2030-41.

[47]	Ou Y, Su Y, Jin J, Fei T. Matching method for distributed photovoltaic maintenance scheme based on knowledge graph. Comput Integr Manuf Sys 2021; 27:1860-70. Chinese.

[48]	Ding Y, Li H, Zhu F, Wang Z, Peng W, Xie M. A semi-supervised failure knowledge graph construction method for decision support in operations and maintenance. IEEE Trans Industr Inform 2024; 20(3):3104-14.

[49]	Liu X, Wang L, Guo Y, Zhang B,Xia X. A knowledge graph based remanufacturing equipment resource modeling method. Digit Twin 2024;4:11.

[50]	Xu Q, Zhou G, Zhang C, Chang F, Huang Q, Zhang M, et al. A digital twin framework for nuclear power equipment maintenance: design, prototyping, and preliminary validation. Digit Twin 2025;2:14.

[51]	Sun Z, Deng ZH, Nie JY, Tang J. RotatE: knowledge graph embedding by relational rotation in complex space. In: Proceedings of the International Conference on Learning Representations (ICLR 2019) ; 2019 May 6-9; New Orleans, LA, USA. ICLR; 2019.

[52]	Gao T, Yao X, Chen D. SimCSE:simple contrastive learning of sentence embeddings. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing; 2021 Nov 7-11; Punta Cana, Dominican Republic. Stroudsburg: Association for Computational Linguistics; 2021. p. 6894-910.

[53]

Cer D, Diab M, Agirre E, Lopez-Gazpio I, Specia L. SemEval- 2017 Task 1:semantic textual similarity multilingual and crosslingual focused evaluation. In: Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017); 2017 Aug 3-4; Vancouver, Canada. Stroudsburg: Association for Computational Linguistics; 2017. p. 1-14.

[54]

Bowman SR, Angeli G, Potts C, Manning CD. A large annotated corpus for learning natural language inference. In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing; 2015 Sep 17-21; Lisbon, Portugal. Stroudsburg: Association for Computational Linguistics; 2015. p. 632-42.

[55]	Vashishth S, Sanyal S, Nitin V, Talukdar P.Composition-based multi-relational graph convolutional networks. In: Proceedings of the 8th International Conference on Learning Representations, ICLR 2020; 2020 Apr 26-30; Addis Ababa, Ethiopia. Appleton: ICLR; 2020. p. 1-15.

[56]

Wang L, Zhao W, Wei Z, Liu J. SimKGC:simple contrastive knowledge graph completion with pre-trained language models. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); 2022 May 22-27; Dublin, Ireland. Stroudsburg: Association for Computational Linguistics; 2022. p. 4281-94.

[57]	Zeng A, Liu X, Du Z, Wang Z, Lai H, Ding M, et al. GLM-130B:an open bilingual pre-trained model. In: Proceedings of the Eleventh International Conference on Learning Representations, ICLR 2023; 2023 May 1-5; Kigali, Rwanda. Appleton: ICLR; 2023. p. 1-56.

[58]	Wang H, Li J, Wu H, Hovy E, Sun Y. Pre-trained language models and their applications. Engineering 2023; 25:51-65.

[59]	Kusiak A. Generative artiﬁcial intelligence in smart manufacturing. J Intell Manuf 2025; 36:1-3.