Automatic surface defect recognition is very important to ensure product quality in modern manufacturing. In recent studies, deep learning-based methods develop rapidly because of its end-to-end training, but the difficult in collecting massive samples limits the application of these methods. Due to the intra-class variations and inter-class similarities in surface defect, it is difficult for the methods to extract features precisely. The knowledge graph can represent the relations of defect samples to improve the ability of extracting features. Therefore, this paper proposes a knowledge graph-guided convolutional neural network (KCNN) for surface defect recognition. Firstly, KCNN defines a knowledge graph by computing cosine distance of defect samples, which can represent the relations of defect samples. Secondly, KCNN uses the knowledge graph to define an extra cost function to guide the training of CNN, which can improve the abilities of extracting features of defect samples. The proposed method is tested on a benchmark dataset and the results show that KCNN achieves as high as 99.79% of the prediction accuracy on small sample training dataset (300 samples) without increasing the recognition time. Compared with other deep learning methods, KCNN achieves the best performances.
© 2020 IEEE.

Scheduling of flexible job shop has been researched over several decades and continues to attract the interests of many scholars. But in the real manufacturing system, dynamic events such as machine failures are major issues. In this paper, an improved Q-learning algorithm with double-layer actions is proposed to solve the dynamic flexible job-shop scheduling problem (DFJSP) considering machine failures. The initial scheduling scheme is obtained by Genetic Algorithm (GA), and the rescheduling strategy is acquired by the Agent of the proposed Q-learning based on dispatching rules. The agent of Q-learning is able to select both operations and alternative machines optimally when machine failure occurs. To testify this approach, experiments are designed and performed based on Mk03 problem of FJSP. Results demonstrate that the optimal rescheduling strategy varies in different machine failure status. And compared with adopting a single dispatching rule all the time, the proposed Q-learning can reduce time of delay in a frequent dynamic environment, which shows that agent-based method is suitable for DFJSP. © 2019 IEEE.

Steelmaking-refining-Continuous Casting (SCC) scheduling is an important industrial scheduling problem, which is also a NP-hard combinatorial optimization problem. Efficient SCC scheduling methods can significantly improve the productivity efficiency. In this paper, an efficient Iterated Local Search (ILS) algorithm is devised for SCC scheduling. The ILS starts from an initial solution with certain quality, then executes a multiswap-based local search, subsequently carries out a cycle of perturbation with probability, neighborhood changing, Simulated Annealing (SA)-based local search, and acceptance criterion until termination condition is reached. The SA-based local search is devised to enhance the local search ability of the ILS, while the perturbation with probability is designed to guide the search to a promising area. Comparison experiments with several scheduling algorithms have shown the strength and effectiveness of the ILS.
© 2019 IEEE.

Intelligent manufacturing plays a significant role in Industry4.0. Dynamic shop scheduling is a key problem and hot researchtopic in the intelligent manufacturing systems, which is NP-hard.However, traditional shop scheduling mode, dynamic eventprediction approach, scheduling model and schedulingalgorithm, cannot cope with increasingly complicated problemsunder kinds of scales production disruptions in the real-worldproduction. To deal with these problems, this paper proposes anew joint data-model driven dynamic scheduling architecture forintelligent workshop. The architecture includes four new and keycharacteristics in the aspects of scheduling mode, dynamic eventprediction, scheduling model and algorithm. More specifically,the new scheduling mode introduces data analytics methods toquickly and accurately deal with the dynamic events encounteredin the production process. The new prediction model improvesthe deep learning method, and further applies it predict thedynamic events accurately to provide reliable input to thedynamic scheduling. The new scheduling model proposes a newhybrid rescheduling and inverse scheduling model, which cancope with almost scales of abnormal production problems. Thenew scheduling algorithm hybridizes dynamic programming andintelligent optimization algorithm, which can overcome thedisadvantages of the two methods based on the analysis ofsolution space. The dynamic programming is employed toprovide high-quality initial solutions for the intelligentoptimization algorithm by reducing the computation timegreatly. To sum up, the presented architecture is a new attempt to understand the problem domain knowledge and broaden thesolving idea, which can also provide new theories andtechnologies to manufacturing system optimization and promotethe applications of the theoretical results. © ASME 2019 14th International Manufacturing Science and Engineering Conference. All rights reserved.

The fault diagnosis is very important for the modern industry. Due to machine working conditions changing frequently, most of current fault diagnosis models built on the training (source) domain can't perform well in test (target) domain. In addition, in test domain, there are few labeled data to adjust model to be adaptive to test working conditions. Domain adaptation, as one type of transfer learning, can be used to solve this problem. This paper proposes a novel fault diagnosis method using unsupervised transfer learning based on adversarial network. In this method, deep neural network is used to extract feature of fault signal while the adversarial network is used to accomplish the transfer learning process. Firstly, the fault signal is converted into RGB images as inputs of networks. Then, the adversarial training methods are used, which includes three training processes: the regular training process using source data, the maximum discrepancy training process and the minimum discrepancy training process. These three steps are adversarial to each other to adjust the model to be more adaptive. The method is tested on motor bearing dataset provided by Case Western Reserve University (CWRU). The prediction accuracies are better than other four comparison methods. © 2019 IEEE.

The flexible job-shop scheduling problem (FJSP) is known as an important problem in manufacturing systems. Many methods have been proposed to solve this problem. The iterated greedy (IG) is one of those algorithms that are widely used in simpler shop scheduling problems. This research proposes a new Telescopic Population approach (TP) to assist the IG in solving the FJSP. The use of TP approach with IG provides an effective method that is also easier to reproduce. The performance of TP with IG proves that the new population approach effectively improves the performance of IG.
© 2018 IEEE.

Fault diagnosis has attracted great attention on preventing the serious consequences from happening and guaranteeing the stability and reliability of machinery equipment. With the rapid development of artificial intelligence, Deep Learning (DL) based approaches begin to play great importance in the field of fault diagnosis. In this research, we proposed an image conversion pre-processing method to transform the time-domain signals of fault diagnosis into 2D images. And a designed structure of Generative Adversarial Networks (GAN) modeled by Convolutional Neural Network (CNN) is proposed to make the classification of fault. Datasets with different capacities are also experimented to study the performance of GAN on limited data. The results illustrate the potential of GAN on the small sample classification of fault diagnosis. © 2018 IEEE.

Fault diagnosis plays a vital role in the modern industry. In this research, a joint vibration signal analysis and deep learning method for fault diagnosis is proposed. The vibration signal analysis is a well-established technique for condition monitoring, and deep learning has shown its potential in fault diagnosis. In the proposed method, the time-frequency technique, named as S transform, is applied to transfer the vibration signals to images, and then an improved convolutional neural network (CNN) is applied to classify these images. The results show the proposed method has achieved the significant improvement.
© 2018 The Authors. Published by Elsevier B.V.

Permutation flow shop scheduling problem (PFSP) exists widely in the manufacturing system. The traditional researches often considered setup time as one part of processing time. However, this assumption does not meet the requirement of the "more varieties, small batches" production model. Therefore, the PFSP with setup time becomes an urgent problem to be solved. This paper proposes an effective hybrid algorithm combining genetic algorithm and variable neighborhood search together. Several neighborhood structures are employed. A set of benchmarks has been used to evaluate its performance. By comparing with some famous algorithms, the merits of proposed method can be shown clearly.
© 2018 The Authors. Published by Elsevier B.V.