With the rapid and transformative advances of large language models (LLMs) in natural language processing, the capabilities of these models in knowledge integration and reasoning have opened new technological pathways for intelligent industrial applications. This review systematically surveys key adaptation techniques, representative application scenarios, and future development trends of LLMs in industrial scenarios. It also provides an integrated overview of their application paradigms and technical characteristics across core industrial processes. Key adaptation techniques for industrial scenarios are first analyzed, including prompt engineering, retrieval-augmented generation (RAG), and parameter-efficient fine-tuning, together with a summary of commonly used evaluation metrics and LLM-based assessment approaches. Representative practices of LLMs are then systematically reviewed across the product lifecycle, covering product design, process planning, production and manufacturing, as well as operation and maintenance. The effectiveness of LLMs in addressing practical industrial problems, facilitating technological innovation, and improving application performance is examined. Finally, major challenges currently encountered in industrial applications of LLMs are identified, including the scarcity of high-quality datasets, limited multimodal fusion capability, insufficient domain specificity, reliability concerns, constrained interpretability, and the lack of standardized evaluation frameworks. Corresponding future research directions are outlined, such as the development of data augmentation and secure sharing mechanisms, the exploration of novel model architectures, and the establishment of intelligent evaluation systems. Overall, this review provides a comprehensive reference for systematic investigations of LLM applications across the entire industrial process and offers theoretical foundations and methodological guidance for both academic research and engineering practice.

Large language models (LLMs) have shown remarkable performances in generic question-answering (QA) but often suffer from domain gaps and outdated knowledge in smart manufacturing (SM). Retrieval-augmented generation (RAG) based on LLMs has emerged as a potential approach by incorporating an external knowledge base. However, conventional vector-based RAG delivers rapid responses but often returns contextually vague results, while knowledge graph (KG)-based methods offer structured relational reasoning at the expense of scalability and efficiency. To address these challenges, a hybrid KG-Vector RAG framework that systematically integrates structured KG metadata with unstructured vector retrieval is proposed. Firstly, a metadata-enriched KG was constructed from domain corpora by systematically extracting and indexing structured information to capture essential domain-specific relationships. Secondly, semantic alignment was achieved by injecting domain-specific constraints to refine and enhance the contextual relevance of the knowledge representations. Lastly, a layered hybrid retrieval strategy was employed that combined the explicit reasoning capabilities of the KG with the efficient search power of vector-based similarity methods, and the resulting outputs were integrated via prompt engineering to generate comprehensive, context-aware responses. Evaluated on design for additive manufacturing (DfAM) tasks, the proposed approach achieved 77.8% exact match accuracy and 76.5% context precision. This study establishes a new paradigm for industrial LLM systems, which demonstrates that hybrid symbolic-neural architectures can overcome the precision-scalability trade-off in mission-critical manufacturing applications. Experimental results indicated that integrating structured KG information with vector-based retrieval and prompt engineering can enhance retrieval accuracy, contextual relevance, and efficiency in LLM-based Q&A systems for SM.

Translating natural language inquiries into executable Cypher queries (text-to-Cypher) is a persistent bottleneck for non-technical teams relying on knowledge graphs (KGs) in fast-changing industrial settings. Rule and template converters need frequent updates as schemas evolve, while supervised and fine-tuned parsers require recurring training. This study proposes a schema-guided prompting approach, namely text-to-Cypher with semantic schema (T2CSS), to align large language models (LLMs) with domain knowledge for producing accurate Cypher. T2CSS distils a domain ontology into a lightweight semantic schema and uses adaptive filtering to inject the relevant subgraph and essential Cypher rules into the prompt for constraining generation and reducing schema-agnostic errors. This design keeps the prompt focused and within context length limits while providing the necessary domain grounding. Comparative experiments demonstrate that T2CSS with GPT-4 outperformed baseline models and achieved 86 % accuracy in producing correct Cypher queries. In practice, this study reduces retraining and maintenance effort, shortens turnaround times, and broadens KG access for non-experts. • A T2CSS prompting approach that guides LLMs with the domain schema is proposed. • A systematic semantic schema to cover multifaceted concepts is designed. • An information filtering mechanism to select the relevant information is proposed. • Results achieve 86 % accuracy in translating user inquiries to Cypher statements.

This paper proposes a novel approach for generating artificial thermal images for induction motor faults using Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (cWGAN-GP) frameworks. Traditional fault classification methods based on vibration signals often require extensive preprocessing and are more susceptible to noise. In contrast, thermal images offer easier classification and require less preprocessing. However, challenges arise due to the limited availability of thermal images representing different fault conditions and data confidentiality. To overcome these challenges, this paper introduces the utilisation of WGAN-GP and cWGAN-GP with health condition labels to create high-quality thermal images artificially. The results demonstrate that the cWGAN-GP approach is superior in generating thermal images that closely resemble real images of induction motors under various health conditions with a Maximum Mean Discrepancy (MMD) score of 1.023 compared to 1.078 using WGAN-GP. Furthermore, cWGAN-GP requires less training time (7.25 hours to train all health conditions classes) compared to WGAN-GP (12 hours to train the Inner fault class only) using NVIDIA V100. In addition to using EMD and MMD metrics for quantitative analysis of the GAN model, the evaluation process incorporated the expertise of a pre-trained CNN model, namely AlexNet, to assess cWGAN-GP's discriminative capabilities of the generated samples and their alignment with the real thermal images, which resulted in an overall accuracy of 98.41%. Therefore, these proposed approaches offer a promising solution to address the lack of public datasets containing induction motor thermal images representing different health states. By leveraging these models, it will be feasible to enhance induction motor condition monitoring systems and improve the process of fault diagnosis.

Product appearance is one of the crucial factors that influence consumers’ purchase decisions. The attractiveness of product appearance is mainly determined by the inherent aesthetics of the design composition related to the arrangement of visual design elements. Hence, it is critical to study and improve the arrangement of visual design elements for product appearance design. Strategies that apply aesthetic design principles to assist designers in effectively arranging visual design elements are widely acknowledged in both academia and industry. However, applying aesthetic design principles relies heavily on the designer’s perception and experience, while it is rather challenging for novice designers. Meanwhile, it is hard to measure and quantify design aesthetics in designing artefacts when designers refer to existing successful designs. In this regard, this study aims to introduce a method that assists designers in applying aesthetic design principles to improve the attractiveness of product appearance. Furthermore, formulas for aesthetic measurement based on aesthetic design principles are also developed, and it makes an early attempt to provide quantified aesthetic measurements of design artefacts. A case study on camera design was conducted to demonstrate the merits of the proposed method where the improved strategies for the camera appearance design offer insights for concept generation in product appearance design based on aesthetic design principles.

Food image segmentation is a critical and indispensible task for developing health-related applications such as estimating food calories and nutrients. Existing food image segmentation models are underperforming due to two reasons: (1) there is a lack of high quality food image datasets with fine-grained ingredient labels and pixel-wise location masks---the existing datasets either carry coarse ingredient labels or are small in size; and (2) the complex appearance of food makes it difficult to localize and recognize ingredients in food images, e.g., the ingredients may overlap one another in the same image, and the identical ingredient may appear distinctly in different food images.

Abstract Industrial Internet of Things (IIoT) is performed based on the multiple sourced data collection, communication, management and analysis from the industrial environment. The data can be generated at every point in the manufacturing production process by real-time monitoring, connection and interaction in the industrial field through various data sensing devices, which creates a big data environment for the industry. To collect, transfer, store and analyse such a big data efficiently and economically, several challenges have imposed to the conventional big data solution, such as high unreliable latency, massive energy consumption, and inadequate security. In order to address these issues, edge computing, as an emerging technique, has been researched and developed in different industries. This paper aims to propose a novel framework for the intelligent IIoT, named Industrial Internet of Learning (IIoL). It is built using an industrial wireless communication network called Low-power wide-area network (LPWAN). By applying edge computing technologies in the LPWAN, the high-intensity computing load is distributed to edge sides, which integrates the computing resource of edge devices to lighten the computational complexity in the central. It cannot only reduce the energy consumption of processing and storing big data but also low the risk of cyber-attacks. Additionally, in the proposed framework, the information and knowledge are discovered and generated from different parts of the system, including smart sensors, smart gateways and cloud. Under this framework, a pervasive knowledge network can be established to improve all the devices in the system. Finally, the proposed concept and framework were validated by two real industrial cases, which were the health prognosis and management of a water plant and asset monitoring and management of an automobile factory.

Several decades of development in the fields of robotics and automation have resulted in human-robot interaction is commonplace, and the subject of intense study. These interactions are particularly prevalent in manufacturing, where human operators (HOs) have been employed in numerous robotics and automation tasks. The presence of HOs continues to be a source of uncertainty in such systems, despite the study of human factors, in an attempt to better understand these variations in performance. Concurrent developments in intelligent manufacturing present opportunities for adaptability within robotic control. This article examines relevant human factors and develops a framework for integrating the necessary elements of intelligent control and data processing to provide appropriate adaptability to robotic elements, consequently improving collaborative interaction with human colleagues. A neural network-based learning approach is used to predict the influence on human task performance and use these predictions to make informed changes to programed behavior, and a methodology developed to explore the application of learning techniques to this area further. This article is supported by an example case study, in which a simulation model is used to explore the application of the developed system, and its performance in a real-world production scenario. The simulation results reveal that adaptability can be realized with some relatively simple techniques and models if applied in the right manner and that such adaptability is helpful to tackle the issue of performance disparity in manufacturing operations. Note to Practitioners-This article presents research into the application of intelligent methodologies to this problem and builds a framework to describe how this information can be captured, generated, and used within manufacturing production processes. This framework helps identify which areas require further research and serves as a basis for the development of a methodology, by which a control system may enable adaptable behavior to reduce the impact of human performance variation and improve human-machine interaction (HMI). This article also presents a simulation-based case study to support the development and evaluate the presented control system on a representative real-world problem. The methodology makes use of a machine-learning approach to identify the complex influence of several identified human factors on human performance. This knowledge can be used to adjust the robotic behavior to match the predicted performance of multiple different operators over different scenarios. This adaptability reduces performance disparity by reducing idle times and enabling leaner production through workpiece-in-progress reduction. Future work will focus on expanding the intelligent capabilities of the proposed system to deal with uncertainty and improve decision-making ability.

In this paper, we propose RNN-Capsule, a capsule model based on Recurrent Neural Network (RNN) for sentiment analysis. For a given problem, one capsule is built for each sentiment category e.g., 'positive' and 'negative'. Each capsule has an attribute, a state, and three modules: representation module, probability module, and reconstruction module. The attribute of a capsule is the assigned sentiment category. Given an instance encoded in hidden vectors by a typical RNN, the representation module builds capsule representation by the attention mechanism. Based on capsule representation, the probability module computes the capsule's state probability. A capsule's state is active if its state probability is the largest among all capsules for the given instance, and inactive otherwise. On two benchmark datasets (i.e., Movie Review and Stanford Sentiment Treebank) and one proprietary dataset (i.e., Hospital Feedback), we show that RNN-Capsule achieves state-of-the-art performance on sentiment classification. More importantly, without using any linguistic knowledge, RNN-Capsule is capable of outputting words with sentiment tendencies reflecting capsules' attributes. The words well reflect the domain specificity of the dataset.

In this paper, we try to predict which category will be less accurately classified compared with other categories in a classification task that involves multiple categories. The categories with poor predicted performance will be identified before any classifiers are trained and additional steps can be taken to address the predicted poor accuracies of these categories. Inspired by the work on query performance prediction in ad-hoc retrieval, we propose to predict classification performance using two measures, namely, category size and category coherence. Our experiments on 20-Newsgroup and Reuters-21578 datasets show that the Spearman rank correlation coefficient between the predicted rank of classification performance and the expected classification accuracy is as high as 0.9.