Journal of Intelligent Manufacturing最新文献

This study introduces a new industrial visual inspection method that emphasizes the application of computer-aided design (CAD) models. This method significantly reduces the dependence on acquiring and annotating extensive real-scene data, subsequently expediting the development of visual inspection models. The paper highlights two pivotal contributions. Firstly, we introduce a configurable 3D rendering technology that digitally simulates different states of the product, achieving automatic batch generation and labeling of training data. This feature distinguishes our work from existing methods. Secondly, we designed a domain generalization method based on second-order statistics. This approach effectively addresses the domain shift challenge between synthetic and actual production data, enhancing the model’s generalization capabilities. This represents a noteworthy advancement in the field as it boosts the model’s adaptability to real-world scenarios. Our method has demonstrated impressive performance, achieving accuracy rates of 94.30(%), 96.75(%), and 97.35(%) on component model classification, motor defect recognition, and rotating motor brush holder datasets, respectively. These results not only validate the efficacy of our domain generalization method but also underscore the potential of using CAD model data for industrial visual inspection. In summary, our research has created a new method for integrating industrial visual inspection into digital twin ecosystems, highlighting the potential for significant improvements in this field.

Machine learning models play a crucial role in smart manufacturing by revolutionizing industrial automation so as to boost productivity and product quality. However, the reliability of these models often faces challenges from factors such as data drift, concept drift, adversarial attacks, and increasing model complexity. In addressing these challenges, this paper proposes a novel approach called Reliability Improved Machine Learning (RIML), which leverages on prior knowledge by incorporating it into the machine learning pipeline through a secondary output that is easily verifiable and assessable within the application domain. Built upon the Knowledge-embedded Machine Learning (KML) framework, RIML differs from conventional strategies by modifying the model’s architecture. In its implementation, additional layers were introduced, specifically designed to identify and discard misclassified cases to improve the model’s reliability. RIML’s efficacy was successfully demonstrated through a simulated dataset and three real use-case studies, namely, a general walk/run scenario, an industry-related case using metro railway dataset, and a smart manufacturing application on gas detection. The promising results highlighted RIML’s ability to significantly reduce misclassifications, thereby enhancing model reliability in diverse real-world scenarios.

In the current scenario, smart scheduling has become an essential requirement to generate dynamic schedules, prescribe, and adjust scheduling plans in response to dynamic events such as machine failures, unpredictable demand, customer order cancellations, worker unavailability, and mass customization. Such scheduling techniques must also take advantage of intelligence continuously being built for next-generation manufacturing systems. This study presents a systematic literature review on smart scheduling, analysing 123 identified literature from 2010 to May 2024 using the PRISMA technique. The analysis includes scientometric and content analysis to identify paradigm shifts in development (concepts, methodologies, practices) along with their maturity levels, and provides recommendations for the next generation of smart scheduling. This study is significant for advancing knowledge and addressing current and future needs/requirements in smart scheduling. This would serve as a reference in understanding the maturity status of various developments, assist researchers and practitioners in identifying research gaps, and direct future advancements in the smart scheduling domain.

Additive manufacturing (AM) or 3-dimensional (3D) printing processes have been adopted in several industrial sectors including aerospace, automotive, medical, architecture, arts and design, food, and construction for the past few decades due to their numerous advantages over other conventional subtractive manufacturing processes. However, some flaws and defects associated with 3D-printed components hinder its extensive adoption in industries. Therefore, real-time detection and elimination of these defects by analyzing the defects-causing process parameters is very important to obtain a defect-free final component. While global efforts are in progress to develop defect detection techniques with the rise of Industry 4.0, there is still a limited scope of comprehensive research that encapsulates various defect detection techniques in the AM sector on a global scale. Thus, this systematic review explores defects in parts manufactured via metallic and non-metallic AM processes. It covers traditional defect detection methods and extends to recent advanced machine learning (ML) and deep learning (DL) based techniques. The paper also delves into challenges associated with the implementation of ML and DL approaches for defect detection, providing a comprehensive understanding of the current state and future directions in AM research.

As the basic manufacturing capabilities provide unit of the production system, the intelligent level of the CNC machine tool will affect the realization of intelligent manufacturing. Academia has carried out a lot of intelligent research on CNC machine tool from technical perspective, but there still needs a systematic cognitive model to promote the construction of cognitive abilities, to support the intelligent realization and continuous improvement of CNC machine tool. Therefore, this paper proposes a three-part, seven-layer cognitive model based on cognitive informatics to promote the construction of cognitive abilities and the intelligent transformation of CNC machine tool. Firstly, a systematic multi-layer cognitive model is proposed, and each cognitive layer is introduced to promote the different cognitive abilities construction of CNC machine tool. Then, this paper introduces the cognitive analysis loop and the cognitive learning loop contained in the multi-layer cognitive model, which can promote the construction of the adaptive and continuous learning abilities of CNC machine tool. The evaluation indicators of the intelligence machine tool are given, which is used to evaluate machine tool intelligence model. Furthermore, the cognitive enabling technologies of the multi-layer cognitive model for intelligent machine tool is presented, which supports the realization of cognitive abilities such as analysis, decision making, and learning. Finally, the feasibility of the proposed systematic multi-layer cognitive model is verified by the developed computable digital twin platform and comparison before and after implementation for intelligent machine tool.

Current equipment fault diagnosis faces challenges due to the difficulties in arranging sensors to collect effective data and obtaining diverse fault data for studying fault mechanisms. The lack of data results in disconnection between data from different spaces, posing a challenge to forming a closed loop of data and hindering the development of digital twin (DT) driven fault diagnosis (FD). To address these issues, a new DT paradigm Digital-Triplet is proposed. This paradigm comprises three entities: a physical entity, a semi-physical entity, and a virtual entity. A semi-physical entity is created by implementing the "six-D" process on the physical entity. A new six dimensional structure is formed through the addition of the semi-physical entity. The new structure streamlines the construction of fault datasets, enhances sensor data acquisition, and tightly links different data spaces, thereby promoting the application of DT in equipment FD. Subsequently, the elevator is selected as a case study to illustrate the Digital-Triplet framework in detail. The results demonstrate that the Digital-Triplet framework can effectively expand the fault dataset and improve data collection efficiency through optimized sensor placement, thereby promoting fault diagnosis.

In remanufacturing, humans perform visual inspection tasks manually. In doing so, human inspectors implicitly solve variants of visual acquisition planning problems. Nowadays, solutions to these problems are computed based on the object geometry of the object to be inspected. In remanufacturing, however, there are often many product variants, and the existence of geometric object models cannot be assumed. This makes it difficult to plan and solve visual acquisition planning problems for the automated execution of visual inspection tasks. Reinforcement learning offers the possibility of learning and reproducing human inspection behavior and solving the visual inspection problem, even for problems in which no object geometry is available. To investigate reinforcement learning as a solution, a simple simulation environment is developed, allowing the execution of reproducible and controllable experiments. Different reinforcement learning agent modeling alternatives are developed and compared for solving the derived visual planning problems. The results of this work show that reinforcement learning agents can solve the derived visual planning problems in use cases without available object geometry by using domain-specific prior knowledge. Our proposed framework is available open source under the following link: https://github.com/Jarrypho/View-Planning-Simulation.

In real industry, it is often difficult to obtain large-scale labeled data. Existing Convolutional Neural Network (CNN)-based fault diagnosis methods often struggle to achieve accurate diagnoses of machine conditions due to the scarcity of labeled data, hindering the ability of models to develop strong inductive biases. We propose a plug-and-play auxiliary method, random convolution layer (RCL), to improve the generalization performance of the fault diagnosis models. This method delves into the fundamental commonalities across diverse tasks and varying network structures, thereby enhancing the diversity of samples to establish a more robust source domain environment. The RCL preserves the dimensional nature of the data in the time domain while randomly altering the kernel sizes during convolution operations, thus generating new data without compromising global information. During the training process, the newly generated data is mixed with the original data and fed into the fault diagnosis model. RCL is incorporated as a module into the inputs of different fault diagnosis models, and its effectiveness is validated on three public datasets as well as a self-built testbed. The results show that the present auxiliary method improves the domain generalization performance of the baselines, and can improve the accuracy of the corresponding fault diagnosis models. Our code is available at https://github.com/zhiqan/Random-convolution-layer.

Data-driven, physics-based, and hybrid prognosis frameworks can be developed to estimate remaining useful life, depending on the availability of condition monitoring sensor data and physics-governing equations. No systematic study is available that shows the comparative performance of these frameworks. The present study, for the first time, attempts to show how these three frameworks can be developed under different scenarios and assumptions. The data-driven prognosis framework is developed using an accelerometer signal and an Artificial Intelligence-based random forest regression (RFR) model. A pit growth model inspired by the Paris crack growth law has been used for physics-based prognosis framework development. In this framework, sensor data is needed to know the gear’s current health status, as the prognosis framework can't be developed purely on physics. A hybrid prognosis framework is developed using two alternate approaches: one in which current health status is obtained directly from a visual inspection camera and the other in which this status is indirectly inferred from the accelerometer sensor data. In each case, the RUL prediction is made using a physics-based pit growth model coupled with the current health status obtained from either of the two approaches mentioned. To enhance the prediction accuracy, Bayesian inference is used to update the physics-based pit growth model parameters in both hybrid frameworks. Data obtained from five run-to-failure experiments performed on a specially designed gearbox test setup are used to show the comparative performance of these frameworks. The strengths and weaknesses of each of the frameworks are discussed based on the type of data requirement, model definition, parameter estimation, and prediction error.

Identifying tool wear state is essential for machine operators as it assists in informed decisions for timely tool replacement and subsequent machining operations. As each wear state corresponds to a unique mitigation strategy, timely identification is vital while implementing solutions to minimize tool wear. The paper presents a novel Human Guided-eXplainable Artificial Intelligence (HG-XAI) approach for identifying the tool wear state by integrating human intelligence and eXplainable AI with a pre-trained Convolutional Neural Network (CNN), Efficient-Net-b0 model. The tool wear states were identified based on different wear mechanisms during the machining of IN718. The study considers four distinct tool wear states, i.e., Flank, Flank+BUE, Flank+Face, and Chipping, representing abrasion, adhesion, diffusion, and fracture wear mechanisms. The image-based datasets were created to depict various tool wear states by machining IN718 at varying surface speeds. The effectiveness of the proposed HG-XAI approach was evaluated by comparing its prediction accuracy with a standalone Efficient-Net-b0 model lacking human intelligence and XAI. Further, the scalability of the HG-XAI approach was examined by predicting wear states from images acquired at different cutting parameters. The results from the present study showed that the HG-XAI approach can predict the tool wear state with an accuracy of 93.08% and is scalable to variations in cutting conditions. Also, the proposed approach can be extended while developing vision-based on-machine tool wear monitoring systems.