The Fukushima nuclear accident was caused by the excessive reactivity of nuclear fuel rods, prompting recent efforts to reduce fuel reactivity through the application of chromium (Cr) coatings. Maintaining a consistent Cr coating thickness is essential to ensure nuclear safety, which requires accurate measurement methods. Although various techniques have been proposed for measuring Cr coating thickness, explaining the relationship between measured signals and the actual physical thickness remains a significant challenge. Ideally, acquiring large-scale datasets would enable the discovery of such physical relationships, but this is difficult in practice. To address this issue, this paper proposes an approach that leverages data augmentation and artificial intelligence (AI) methods to enable accurate thickness prediction using limited data from ECT time-series signals. In particular, a predictive model based on a multi-head attention long short-term memory (LSTM) architecture was developed, and multiple time-series data augmentation methods were applied to enhance performance despite data scarcity. The results demonstrate that the proposed method not only improves measurement reliability in the context of nuclear safety, but also provides a practical framework for guiding sensor development in other fields through the integration of AI and data augmentation.

This study presents a hybrid approach for the automated recognition and classification of line objects in piping and instrumentation diagrams (P&IDs), with the goal of supporting the digital transformation of chemical process design and operation. By integrating Deep Learning (DL) techniques with rule-based methods, the proposed approach extracts flow and signal paths from legacy P&ID images, enabling applications such as process simulation, safety verification, and control logic validation. The approach consists of two stages. In the first stage, all line objects in a P&ID are detected and categorized into lines with special signs and continuous lines. A DL model identifies directional arrows and determines the overall flow structure. In the second stage, the continuous lines are further classified into dimension, extension, and leader lines using the rule-based algorithms, according to their functional characteristics. The method was tested on 30 P&ID sheets from Project A and two from Project B. Initially, the model trained on Project A data achieved precision and recall rates of 95.02% and 93.09%, respectively. On Project B, the performance dropped to 88.92% and 84.76% due to domain shift. After applying transfer learning using the four additional Project B sheets, the performance improved to 95.32% precision and 91.55% recall. These results demonstrate the potential of the proposed approach for accurate and scalable conversion of P&ID data into structured formats, contributing to smart plant design and engineering data integration.

This study addresses the problem of transforming unstructured engineering design requirements into machine-interpretable knowledge for automated verification. Conventional engineering verification approaches typically assume that design knowledge is already formalized, creating a critical bottleneck when requirements are described in natural language documents. To overcome these limitations, this study presents an automated approach for converting unstructured documents into Computer-Aided Design (CAD)-verifiable design rules. Design variables and requirements are first extracted and consolidated to resolve duplication arising from differences in expression. To enable reliable knowledge alignment under linguistic variability, document-level variables are semantically mapped to standardized parameters using a Sentence-BERT (SBERT)-based similarity model and a codebook. The extracted requirements are formalized into symbolic mathematical representations. The resulting rules are organized into a three-layer architecture inspired by a reinterpretation of the Requirement–Applicability–Selection–Exception (RASE) structure, enabling dynamic selection of applicable rules according to the relevant design features. The experimental results demonstrate that the proposed framework can reliably transform document-level requirements into a consistent and reusable design knowledge structure, achieving substantial coverage within the executable requirement scope and complete resolution of redundant variables. Moreover, the automated verification results on CAD models were fully consistent with manual verification outcomes.

Automated CAD–CAE integration, which links computer-aided design (CAD) and computer-aided engineering (CAE), is one of the key strategies for reducing repetitive tasks in the design process. However, setting analysis parameters and assigning boundary conditions for variously designed shapes require expert knowledge and precise recognition of topological elements, remaining a major challenge to full automation. To address these issues, this study proposes an automated integration framework. The framework utilizes a small language model (SLM) and prompt engineering to extract analysis and validation parameters from unstructured documents. Additionally, boundary faces are selected through design feature recognition. The parameter extraction performance was evaluated using four language models developed by different providers and trained at different scales. The evaluation showed consistent extraction accuracy across all models. Boundary condition assignment was verified using CAD models including cup-anemometer and snap-fit hook features, achieving 100% accuracy in all experiments. The effectiveness of automation was assessed by comparing the number of manual inputs and the total analysis time. The experimental results showed reductions of 97.06% in manual inputs and 63.92% in analysis time per CAD model. These experimental results demonstrate the effectiveness of a framework that enables the entire analysis process to be executed without human intervention. The study provides a practical solution to the challenges of parameter setting and boundary condition assignment. Furthermore, the proposed method presents an innovative approach that can be generally applied to performance analysis of complex and diverse CAD models and experimentally verifies the benefits achievable through full automation.

The demand for appearance diversity in concept development often results in repetitive, labour-intensive tasks. Generative artificial intelligence (AI) offers a promising solution for automating and enhancing appearance design. Effective integration generative AI into appearance design workflows requires capabilities in generation, segmentation, and editing of 3D appearance elements. This paper systematically surveys generative AI technologies within the appearance design stage, explicitly focusing on three core domains: generation, segmentation, and editing. Recent generative techniques rapidly visualize 3D appearance concepts from textual inputs, minimizing manual effort. Segmentation techniques separate appearance components within generated models, enabling targeted refinements and selective modifications. Editing methods support detailed geometry, texture, colour, and attribute adjustments of existing concepts, facilitating iterative design refinement. While generative AI effectively automates initial ideation and simplifies complex modifications, seamless integration among generation, segmentation, and editing processes remains challenging. Future research should focus on unified frameworks and computational efficiency improvements. By surveying these core areas, this paper emphasizes generative AI’s potential to support or partially replace traditional manual appearance tasks, thereby enhancing creativity and efficiency.

This study proposes a 3D point cloud registration method that achieves high registration performance without the need for geometric feature extraction or correspondence matching. The proposed method first performs an initial alignment by estimating the principal axes between the source and the target using principal component analysis (PCA). Additionally, a plane constraint is incorporated to mitigate PCA instability by restricting rotation to a single axis when a dominant planar patch is detected. It then refines the alignment through a multi-stage Bayesian optimization process that separately optimizes rotation and translation in the 6 degrees of freedom (6-DOF) transformation space while progressively shrinking the search space. This method maintains stable registration performance without relying on feature matching and effectively achieves a balance between low computation time and high accuracy compared to conventional 6-DOF global optimization approaches. Experimental results on both custom single-object datasets and the KITTI Odometry dataset show that, although some limitations exist for geometrically symmetric objects, the proposed method generally demonstrates superior accuracy and real-time performance. This confirms the potential of Bayesian optimization as an effective and extensible method for real-time point cloud alignment.

With the advancement of automation in industrial equipment, the need for precise and real-time process monitoring has become increasingly critical. Digital twins and artificial intelligence (AI) have emerged as promising technologies for addressing this need, and their application to industrial process monitoring has been actively investigated. This study proposes an interoperable approach for integrating AI models with physical assets by leveraging the Asset Administration Shell (AAS) and Open Neural Network Exchange (ONNX) standards and demonstrates its implementation as an AI-powered digital twin in a real industrial process. In the proposed architecture, the AAS meta-model provides a standardized information structure for the digital representation of physical assets and the interface between the physical and cyber worlds. In parallel, AI models are integrated in the ONNX format as framework-independent inference modules, enabling the operational status of assets to be predicted across heterogeneous deployment environments. By combining AAS-based asset representation with ONNX-based AI inference, the proposed digital twin moves beyond passive asset representation toward reusable, interoperable, and AI-enabled decision support. The proposed methodology achieved a 95.47

The Asset Administration Shell (AAS) of Industry 4.0 is a key technology for enabling information interoperability among heterogeneous assets in smart manufacturing. This study proposes an AAS-based digital twin framework that ensures both model reusability and application extensibility. The framework defines a neutral model based on the AAS meta-model linked with external reference data, allowing the digital twin to flexibly respond to changes in asset configuration. This approach improves the reusability of asset data. In addition, multiple application functions can be integrated without modifying the neutral model, which guarantees the extensibility of the digital twin system. To validate the proposed framework, digital twins of industrial robotic systems performing remote underwater cutting, wire arc additive manufacturing, and pneumatic transfer were implemented. When the configuration of the robotic system was changed, 84 % of the digital twin data were reused. Furthermore, a structural implementation efficiency of 87 % was achieved, enabling the flexible development of two digital twin systems—real-time operational monitoring and 3D operational simulation—without structural modification.

Due to the limited usability of unstructured piping and instrumentation diagrams (P&IDs) in plant projects, most companies manually digitize them into digital P&IDs. To automate the digitization of P&IDs, this study proposes a graph neural network (GNN)-based method to classify detected continuous lines. We present the graph structure that represents the connections of continuous lines and the process of generating these graphs. Additionally, we introduce the continuous line classification network (ContLineNet), which classifies the nodes of the continuous line connection graph into eight classes. Experiments were conducted using P&IDs from three different companies. P&IDs containing 7,371 continuous lines from one specific company were used to train ContLineNet. The trained ContLineNet achieved average values of 96.968% for precision, 96.778% for recall, and 96.700% for the F1 score on 1,924 continuous lines across five P&IDs from three companies. These experimental results demonstrate the applicability of our method to various data.

In the product development process, it is essential to identify potential design errors in design data and correct them appropriately before production begins. This paper proposes a method for rule-based design verification of mechanical parts through dynamic selection of rule subsets. First, a knowledgebase structure centered on design features was defined to specify rules from design requirements. Then, a method was developed to dynamically compose verification functions using the verification equations contained in the selected design rules. These rules are applied to the design parameter data extracted from design data such as 3D CAD models. The proposed method allows for the easy expansion of applicable parts by simply adding the necessary rules to the knowledgebase, without changing the structure of the design rule knowledgebase or expanding the verification system. Design verification experiments were conducted on anemometers and snap-fit hooks to verify the proposed method's effectiveness. The experiments successfully performed design verification for both parts, confirming that the proposed method can be flexibly applied to various parts.

Due to the limited time available during the bidding process, construction companies may fail to identify risky terms in the contract document before submission. This paper proposes a method that uses natural language processing (NLP) models, such as dependency parser and bidirectional encoder representations from transformers (BERT), to disassemble and simplify sentences in a contract document and automatically generate rules for identifying risk sentences. The sentence disassembly process is conducted in the following order: adjunct separation, parallel structure separation, and subject-verb-object separation. Subsequently, risk sentence identification rules are automatically generated through the input of risk terms. The performance of the proposed method is verified using the generated rules. In the experiments, the accuracies of sentence disassembly and risk sentence identification were 95.5 % and 91.3 %, respectively. The proposed method can assist experts in reviewing contracts, significantly reducing the time required to generate new identification rules.

Numerous studies have focused on digitizing piping and instrumentation diagrams (P&IDs) to enhance their applicability across industries. Even with the application of digitization technology, correcting errors in object recognition remains time-consuming, and because correction is performed manually, errors may persist. To address these issues, we propose a novel method for inspecting object recognition results in P&IDs. For unrecognized object inspection, patches are generated, and a deep learning-based classifier identifies missing elements. For misrecognized object inspection, optimal inspection methods are applied depending on the type of error, enabling effective detection of misrecognized objects. Specifically, the proposed misrecognition inspection methods include deep learning-based feature vector similarity calculation, text error detection based on distance-based detection, and line error detection through intersection-case inspection method. The proposed method was validated through experiments conducted using P&IDs from actual industrial sites. The results showed that unrecognized object inspection achieved a recall of 100%. Misrecognized object inspection achieved an accuracy of 99.2% and an F1 score of 96.7% for symbols, an accuracy of 95.8% and an F1 score of 97.3% for text, and 100% accuracy and F1 score for lines. Overall, error correction time was reduced by approximately 40%.

Design feature recognition plays a crucial role in digital manufacturing and is a key technology in automatic design verification. Traditional methods and deep learning approaches provide various strategies for feature recognition. However, these methods primarily address part classification or machining feature recognition, with limited research focusing on design feature recognition. To address this gap, a novel deep learning network called the design feature graph attention network (DFGAT) was proposed specifically for design feature recognition. In this study, the original boundary representation (B-rep) model is first converted into graph representation. Design feature recognition is then achieved using the DFGAT, which is based on the GAT. Additionally, the dataset generation process was generalized to efficiently train the deep learning model. To validate the performance of the DFGAT, experiments were conducted to recognize the representative faces of design features, such as snap-fit hooks, cups, and plates, in the EIF_Panel, Real_Panel, and Anemometer models. The experiments demonstrated F1-scores of 0.9924, 0.9982, and 1.0000.

The interior of a nuclear reactor, filled with water and classified as a medium-level radiation area, is inaccessible to humans, requiring underwater remote cutting during decommissioning. However, the cutting process generates bubbles and light, hindering camera-based monitoring and necessitating status determination through sensor data. This study introduces an adaptive weighted parallel 1D-DenseNet that integrates pressure and hydrophone sensor data in both time and frequency domains to distinguish between cutting and idle states. Time-series data are transformed into the frequency domain via fast Fourier Transform (FFT), generating four inputs: raw and FFT signals for each sensor. These inputs are processed through a parallel network, where the outputs of 1D-DenseNet are multiplied by adaptive weights, concatenated, and passed through a fully connected layer for status determination. The proposed method achieves higher computational efficiency than conventional time-frequency approaches and seamlessly integrates additional sensors. Experimental results on a validation dataset show an accuracy of 98.85% and an F1-score of 0.9888. Comparisons with baseline models and an ablation study confirm its superior performance. The proposed model offers an effective solution for monitoring underwater cutting processes during nuclear reactor decommissioning

The topological and geometric information in B-rep models is essential for various applications, including feature recognition, design verification, and automatic tool path generation. In particular, representing surfaces using an analytic format offers advantages for utilizing geometric information such as surface type, axis of rotation, and radius. Models composed solely of non-uniform rational B-spline (NURBS) surfaces often have limitations in extracting this type of information. To address this issue, this study proposes a method for identifying and replacing NURBS surfaces that can be represented analytically. In the first step, the feasibility of substituting an NURBS surface with an analytic surface is determined by examining the characteristics of its isocurves. In the second step, an equivalent analytic surface is created and used to replace the existing NURBS surface. To validate the proposed method, surface replacement experiments were conducted on multiple test cases. The results showed that, among 10 176 NURBS surfaces identified for possible analytic replacement across 37 B-rep models, 9163 surfaces were successfully substituted in a stable manner. In addition, an experiment comparing feature recognition in the pre- and post-replacement models was conducted. As a result, none of the 12 design features were recognized in the pre-replacement model, whereas all were correctly detected afterward, confirming that the replacement process significantly enhances feature recognition accuracy.

Topological elements form the basis for tasks such as geometric calculations, feature analysis, and direct modeling in 3D CAD systems. Handling these elements is also essential in various automated systems. This study proposes a method to search for topological elements within a boundary representation (B-rep) model by employing topological queries. To address complex scenarios that are difficult to handle using a single query, a topological query procedure that sequentially executes a predefined set of topological queries is used. To verify the effectiveness of the proposed method, experiments were conducted on Test Cases 1, 2, and 3, confirming the successful search of all target topological elements. Furthermore, tests on modified Snap-fit hook A and Bridge B models demonstrated that the same queries remained effective, provided the topological relationships and geometric constraints expressed in the query were preserved. In addition, a search time comparison showed that the proposed method reduced search time by over 90 % compared to manual processes. Finally, in an experiment involving participants with varying levels of programming proficiency, the results indicated that, for a developer with high programming skills, writing topological queries reduced the time required to search for a single topological element by more than 95 % compared to writing the program code.

Recent studies propose deep learning-based methods to recognize symbols and text in Piping and Instrumentation Diagrams (P&ID). However, existing approaches use complex processes with separate models for symbol detection, text detection, and text recognition. We propose an integrated model combining symbol-text detection and text recognition modules using a text spotting method. Our model extracts text region features encoded with local character information, enabling a lightweight text recognition module that reduces processing time. The integrated approach allows end-to-end learning between modules, facilitating semantic information transmission and improving overall performance compared to multi-model architecture. When tested on industrial P&ID images, our model achieved high performance with an IoU threshold of 0.5: maximum precision of 0.9763/0.9527, recall of 0.9521/0.9075, and F1 score of 0.9640/0.9295 for symbol-text detection/text recognition.

In the livestock industry, reducing the cost of raising animals while maintaining quality is of utmost importance. Growth management is crucial for livestock quality control, with weight being a primary focus. This study proposes a method for estimating the weight of a breed sow from point clouds using a convolutional neural network (CNN). The proposed method consists of two stages: first, 2D images are generated from point clouds of a walking breed sow; next, these images are used as input in a CNN model to estimate the weight. The point clouds are divided into two halves along a longitudinal plane. Then, distances are calculated between each element of the 128 × 64 rectangular mesh created on the aforementioned plane and the corresponding points in the divided point cloud. A 2D image is generated by varying the grayscale intensity according to the calculated distances. Three types of CNNs, namely, VGG16, DenseNet121, and EfficientNet B0, were employed to estimate the weight using the generated 2D images as input. The CNNs were trained using approximately 150,000 images of 71 pigs. To verify the effectiveness of the method, the trained networks were used to conduct weight estimation experiments on test cases. The results of the experiments demonstrate a high weight estimation accuracy, with an error rate as low as 1.35%.

Recently, the digitization of piping and instrumentation diagrams (P&IDs) has become increasingly necessary in the plant industry. In previous studies on line recognition in P&IDs, lines were classified according to line signs after line detection. However, detailed studies on the classification of continuous lines are limited. Herein, a rule-based method for classifying continuous lines in a P&ID is proposed. This method uses the shape and positional relationships between objects in the P&ID to classify continuous lines into eight types: drain, annotation, spec break, dimension, extension, leader, thick continuous, and continuous line. A prototype system was developed to validate the proposed method, and line classification experiments were conducted on actual high-density industrial P&ID systems. The experimental results indicated that the average precision, recall, and F1 score for the five P&IDs were 96.300%, 97.945%, and 96.415%, respectively. These results indicate the excellent performance of the proposed method in classifying continuous lines for various P&IDs.

Reusing existing CAD part models in product development can significantly shorten design time and reduce costs. However, due to the extensive number of existing CAD part models within the Product Data Management (PDM) system, designers often spend a considerable amount of time retrieving models that meet their specific requirements. This study introduces a method to retrieve parts that comply with design specifications using a relational design rule-embedded bill of materials (BOM). Parameters are extracted from CAD part models in the database to generate a list of parts with properties. The compliance of each part with the design specifications is verified utilizing the relational design rule-embedded BOM, and CAD part models meeting the specifications are identified and retrieved. To validate this approach, an experiment was conducted to retrieve CAD part models needed for designing a hinge assembly center of a refrigerator. The experiment successfully retrieved 24 CAD part models that complied with design specifications out of a total of 139 CAD part models from six different part types. Notably, this retrieval process was completed within 1 min and 45 s.