With the continued advancement of Artificial Intelligence (AI) technology in Customized Manufacturing (CM), the current intelligence model, which separates ‘perception’ from ‘execution,’ lacks adaptability and generalizability. Embodied Intelligence (EI), an emerging technology emphasizing real-time environmental interaction and feedback, is expected to enable integrated intelligent manufacturing systems characterized by ‘perception-cognition-execution-feedback,’ enhancing the performance and intelligence of multifactor systems like the Internet of Manufacturing Things (IoMT), equipment management, and resource scheduling. However, current AI systems in CM remain isolated and lack methods for environmental interaction with multi-source perception and feedback, hindering the development of autonomous, evolving intelligent systems. To address the environmental interaction bottlenecks among diverse production elements in CM, this paper proposes a Circular Embodied Intelligence Manufacturing (CEIM) architecture aimed at enabling the fusion of heterogeneous production-element information. The objective is to enhance environmental perception and establish autonomous system evolution mechanisms, thereby optimizing production decision-making. By integrating sensor data and AI model outputs, and leveraging the advanced reasoning capabilities of large language models (LLMs), CEIM facilitates the semantic fusion of multiple production elements—including IoMT, intelligent devices, and manufacturing resources—to enable the deployment of EI applications in CM. The implementation of CEIM is illustrated and validated through a case study on a customized gift box packaging platform. Finally, this paper discusses the opportunities and challenges of applying EI in manufacturing and aims to provide insights into the future development of EI-oriented manufacturing systems.

Presently, resource-constrained devices in manufacturing Internet of Things (MIoT), such as sensors and radio frequency identification (RFID) devices, collect a large amount of privacy-sensitive data. However, weak passwords and vulnerable encryption capabilities in MIoT have often become loopholes of security risks. To this end, this paper proposes a novel secure data-sharing scheme based on the integration of blockchain and fusion of both real and fake data to address the data security requirements of resource-constrained MIoT devices. First, a computing resource collaboration architecture is designed, thereby enabling these devices to interface with multiple devices with full resource to implement flexible resource scheduling. Then, a resource-assistance mechanism is devised through blockchain-based smart contracts by utilizing the idle resources of full nodes to complement the computational tasks launched by resource-constrained nodes. In addition, polygon semantic rules are proposed to improve the security of private data. Subsequently, real data artifacts are generated by data tampering to achieve privacy cover, avoiding the consumption of computing resources in traditional encryption algorithms. Finally, the feasibility of the proposed scheme is verified on a customized candy production line. The experimental results validate that data protection using polygon semantic rules can prevent actual data from being peeped. Moreover, the results also indicate that the proposed method can obtain resource assistance from other nodes through the resource compensation mechanism.

In the wind turbine remote fault diagnosis, sensor data is susceptible to low-quality phenomena such as missing and damaged data due to communication delays, environmental noise, and sensor faults. These issues decrease the accuracy of fault diagnostic models, necessitating a solution to enhance model robustness under non-ideal sensor data conditions. Hence, a robust fault diagnostic scheme based on adaptive noise filtering and useful feature-domain enhancement (UFDE) is proposed in this paper to improve the stability of fault diagnostic performance. An interference identification branch is designed to analyze sensor data from a high-dimensional and multi-level perspective, automatically identifying and localizing feature noise during training. Subsequently, a UFDE mechanism containing three feature mapping modes is created, using adaptive mapping and filling of fault features in the neighborhood to eliminate feature noise and enhance the useful feature domain. This process improves the representation of fault features under non-ideal sensor data conditions, such as noise interference and data defects, thereby enhancing the fault diagnostic model’s robustness. Finally, under non-ideal sensor data conditions, comparative experiments with advanced fault diagnostic methods demonstrate that the proposed method exhibits minimal fluctuations in diagnostic accuracy and achieves the highest correctness rate, validating its robustness.

With the development of artificial intelligence, machine vision technology based on deep learning is an effective way to improve production efficiency. Because of the rapid update of the automobile manufacturing industry and the large variety of products, the learning time and the number of learning samples of the deep learning model are limited, which brings great difficulties to the recognition of components. Therefore, considering the economic benefits of enterprises, this paper proposes an intelligent component recognition method appropriate for small datasets, aiming to explore an automatic system for component recognition suitable for industrial manufacturing environments. The method completes the generation of the dataset through the system architecture with the potential for automation and the image cropping method based on feature detection and then designs a deep learning network based on coarse‐fine‐grained feature fusion to generate an intelligent recognition model of components. Finally, the designed network achieves an accuracy of 95.11%, and compared with the traditional classical network on multiple datasets, the designed network has better performance. Thus, the proposed method can improve the production flexibility of the automobile manufacturing industry and improve equipment intelligence.

To address the issue in zero-shot fault diagnosis (ZSFD) where fault attribute definitions (FADs) rely heavily on manual design and the accuracy of FAD depends on the expertise of developers, this article embedded expert knowledge into deep learning network, proposed a ZSFD method based on depth correlation feature extraction network (DCFEN), and automatically constructed FAD. Taking advantage of the periodic characteristics of bearing fault signals and the advantages of correlation analysis operation (CAO) in periodic signal analysis, DCFEN extracts the periodic characteristics of input signals in multiple dimensions by integrating CAO with deep learning. In addition, a soft-threshold-based feature percolation mechanism and FAD evaluation function are designed to generate the attributes related to bearing faults. The experimental results show that the FADs established by DCFEN are accurate, and the fault diagnosis performance of the proposed ZSFD is superior to the existing methods in unseen scenarios.