Journal of Intelligent Manufacturing最新文献

Human motion prediction is crucial for facilitating human–robot collaboration in customized assembly tasks. However, existing research primarily focuses on predicting limited human motions using static global information, which fails to address the highly stochastic nature of customized assembly operations in a given region. To address this, we propose a dynamic scenario-enhanced diverse human motion prediction network that extracts dynamic collaborative features to predict highly stochastic customized assembly operations. In this paper, we present a multi-level feature adaptation network that generates information for dynamically manipulating objects. This is accomplished by extracting multi-attribute features at different levels, including multi-channel gaze tracking, multi-scale object affordance detection, and multi-modal object’s 6 degree-of-freedom pose estimation. Notably, we employ gaze tracking to locate the collaborative space accurately. Furthermore, we introduce a multi-step feedback-refined diffusion sampling network specifically designed for predicting highly stochastic customized assembly operations. This network refines the outcomes of our proposed multi-weight diffusion sampling strategy to better align with the target distribution. Additionally, we develop a feedback regulatory mechanism that incorporates ground truth information in each prediction step to ensure the reliability of the results. Finally, the effectiveness of the proposed method was demonstrated through comparative experiments and validation of assembly tasks in a laboratory environment.

Designing efficient closed-loop control algorithms is a key issue in Additive Manufacturing (AM), as various aspects of the AM process require continuous monitoring and regulation, with temperature being a particularly significant factor. Here we study closed-loop control for the temperatures in the extruder of a Material Extrusion AM system, specifically a Big Area Additive Manufacturing (BAAM) system. Previous approaches for temperature control in AM either require the knowledge of exact model parameters, or involve discretisation of the state and action spaces to employ traditional data-driven control techniques. On the other hand, modern algorithms that can handle continuous state and action space problems require a large number of hyperparameter tuning to ensure good performance. In this work, we circumvent the above limitations by making use of a state space temperature model while focusing on both model-based and data-driven methods. We adopt the Linear Quadratic Tracking (LQT) framework and utilise the quadratic structure of the value function in the model-based analytical solution to produce a data-driven approximation formula for the optimal controller. We demonstrate these approaches using a simulator of the temperature evolution in the extruder of a BAAM system and perform an in-depth comparison of the performance of these methods. We find that we can learn an effective controller using solely simulated input–output process data. Our approach achieves parity in performance compared to model-based controllers and so lessens the need for estimating a large number of parameters of the often intricate and complicated process model. We believe this result is an important step towards achieving autonomous intelligent manufacturing.

Automatic tool condition monitoring becomes crucial in metal cutting because tool wear impacts the final product’s quality. The optical microscope approach for assessing tool wear is offline, time-consuming, and subject to measurement error by humans. To accomplish this, the machine must be stopped, and the tool must be removed, which causes downtime. As a result, numerous research attempts have been made to develop robust systems for direct tool wear measurement during machining. Therefore, the proposed work focused on developing a direct tool condition monitoring system using machine vision to calculate tool wear parameters, specifically flank wear. The cutting tool insert images are collected using a machine vision setup equipped with an industrial camera, bi-telecentric lens, and a proper illumination system during the machining of AISI 4140 steel. The comparative analysis of image processing algorithms for tool wear measurement is proposed under the selected machining environment. The wear boundary is extracted using digital image processing tools such as image enhancement, image segmentation, image morphology operation, and edge detection. The wear amount on the tool insert is extracted and recorded using the Hough line transformation function and pixel scanning. The comparison of results revealed the measurement accuracy and repeatability of the proposed image processing algorithm with a maximum of 6.25% and minimum of 1.10% error compared to manual measurement. Hence, the proposed approach eliminates manual measurements and improves the machining productivity.

Production scheduling has a significant role when optimizing production objectives such as production efficiency, resource utilization, cost control, energy-saving, and emission reduction. Currently, deep reinforcement learning-based production scheduling methods achieve roughly equivalent precision as the widely used meta-heuristic algorithms while exhibiting higher efficiency, along with powerful generalization abilities. Therefore, this new paradigm has drawn much attention and plenty of research results have been reported. By reviewing available deep reinforcement learning models for the job shop scheduling problems, the typical design patterns and pattern combinations of the common components, i.e., agent, environment, state, action, and reward, were identified. Around this essential contribution, the architecture and procedure of training deep reinforcement learning scheduling models and applying resultant scheduling solvers were generalized. Furthermore, the key evaluation indicators were summarized and the promising research areas were outlined. This work surveys several deep reinforcement learning models for a range of production scheduling problems.

Process monitoring and anomaly detection in manufacturing production systems are of vital importance in the consistency, reliability, and quality of manufacturing operations, which motivates the development of a large number of anomaly detection methods. In this work, a comparison study on representative unsupervised X-ray image based anomaly detection methods is conducted. Statistical, physical, and deep learning based dimension reduction methods and different anomaly detection criteria are considered and compared. Case studies on real-world X-ray image data with simulated anomalies and real anomalies are both carried out. Grey level co-occurrence matrix with Hotelling (T^2) statistic as detection criteria achieves the best performance on the simulated anomalies case with an overall detection accuracy of 96%. Principal component analysis with reconstruction error as criteria obtains the highest detection rate of 90.6% in the real anomalies case. Considering the ever-growing availability of image data in intelligent manufacturing processes, this study will provide very useful knowledge and find a broad audience.

Anomaly detection in multivariate time-series data is critical for monitoring asset conditions, enabling prompt fault detection and diagnosis to mitigate damage, reduce downtime and enhance safety. Existing literature predominately emphasises temporal dependencies in single-channel data, often overlooking interrelations between features in multivariate time-series data and across multiple channels. This paper introduces G-BOCPD, a novel graphical model-based annotation method designed to automatically detect anomalies in multi-channel multivariate time-series data. To address internal and external dependencies, G-BOCPD proposes a hybridisation of the graphical lasso and expectation maximisation algorithms. This approach detects anomalies in multi-channel multivariate time-series by identifying segments with diverse behaviours and patterns, which are then annotated to highlight variations. The method alternates between estimating the concentration matrix, which represents dependencies between variables, using the graphical lasso algorithm, and annotating segments through a minimal path clustering method for a comprehensive understanding of variations. To demonstrate its effectiveness, G-BOCPD is applied to multichannel time-series obtained from: (i) Diesel Multiple Unit train engines exhibiting faulty behaviours; and (ii) a group of train doors at various degradation stages. Empirical evidence highlights G-BOCPD's superior performance compared to previous approaches in terms of precision, recall and F1-score.

This study introduces a novel approach using Physics-Informed Neural Networks (PINN) to predict weld line visibility in injection-molded components based on process parameters. Leveraging PINNs, the research aims to minimize experimental tests and numerical simulations, thus reducing computational efforts, to make the classification models for surface defects more easily implementable in an industrial environment. By correlating weld line visibility with the Frozen Layer Ratio (FLR) threshold, identified through limited experimental data and simulations, the study generates synthetic datasets for pre-training neural networks. This study demonstrates that a quality classification model pre-trained with PINN-generated datasets achieves comparable performance to a randomly initialized network in terms of Recall and Area Under the Curve (AUC) metrics, with a substantial reduction of 78% in the need for experimental points. Furthermore, it achieves similar accuracy levels with 74% fewer experimental points. The results demonstrate the robustness and accuracy of neural networks pre-trained with PINNs in predicting weld line visibility, offering a promising approach to minimizing experimental efforts and computational resources.

In this manuscript, we present a comprehensive overview of true digital twin applications within the manufacturing industry, specifically delving into advancements in sheet metal forming. A true digital twin is a virtual representation of a physical process or production system, enabling bidirectional data exchange between the physical and digital domains and facilitating real-time optimization of performance and decision-making through synchronized data from sensors. Hence, we will highlight the difference between Industry 4.0 and the digital twin concept, which is considered synonymous with Industry 5.0. Additionally, we will be outlining the relationship between the true digital twin and Zero Defect Manufacturing. In manufacturing processes, including sheet metal stamping, the advantages of high production speed, cost-effective tooling, and consistent component production are counterbalanced by the challenge of dimensional variability in finished parts, which is influenced by process parameters. Data collection, storage, and analysis are essential for understanding manufactured parts variability, and leveraging true digital twins ensures high-quality parts production and processes optimization.

Teleoperation, which is a specific mode of human–robot collaboration enabling a human operator to provide instructions and monitor the actions of the robot remotely, has proved beneficial for application to hazardous and unstructured manufacturing environments. Despite the design of a command channel from human operators to robots, most existing studies on teleoperation fail to focus on the design of the feedback channel from the robot to the human operator, which plays a crucial role in reducing the cognitive load, particularly in precise and concentrated manufacturing tasks. This paper focuses on designing a feedback channel for the cognitive interface between a human operator and a robot considering human cognition. Current studies on human–robot cognitive interfaces in robot teleoperation are extensively surveyed. Further, the modalities of human cognition that foster understanding and transparency during teleoperation are identified. In addition, the human–robot cognitive interface, which utilizes the proposed multi-modal feedback channel, is developed on a teleoperated robotic grasping system as a case study. Finally, a series of experiments based on different modal feedback channels are conducted to demonstrate the effectiveness of enhancing the performance of the teleoperated grasping of fragile products and reducing the cognitive load via the objective aspects of experimental results and the subjective aspects of operator feedback.

Pushing is a fundamental ability in mobile robotics for transporting objects when grippers are not applicable. A successful “box-pushing” requires good coordination between model prediction, pushing strategy, and motion planning, therefore presents a well-known challenge in mobile robot transportation community. However, current research often focuses on local planning for altering push direction, while global planning remains inadequate. This can lead to inefficient pushing trajectories, especially in narrow passages where robots may unintentionally push the box into a dead end due to the lack of robust global path. To address this, we propose the use of stable pushing as an effective technique and develop a unique global planning approach based on the hybrid A* algorithm. We enhance the hybrid A* algorithm by modifying the node expansion approach and incorporating a mechanism for predicting push direction, enabling the system to adapt to changing push side behavior and discover optimal pathways. Extensive simulations validate our system’s effectiveness in handling complex scenarios with limited passageways. As a result, our method significantly improves the robot’s capability to generate superior global paths for box-pushing, mitigating wasteful trajectories and enhancing overall performance.