Adaptive Production Scheduling Approach With Closed-loop Layers And Cascaded Phases For Smart Manufacturing

Production scheduling is the core of production operation management as well as the key technology which determines the economic benefits and competitiveness of manufacturing enterprises.It is also a hot research topic in system engineering,industrial engineering and other fields.In the era of smart manufacturing,the more complicated and uncertain production environment causes great challenges for its theoretical research and industrial practice.As a result,how to improve the adaptive ability of production scheduling to the uncertain environment,so that manufacturing system can run smoothly and continuously,has become an important research topic.After first analyzing the production scheduling problem of smart manufacturing and systematically reviewing the theory of dynamic scheduling,this thesis takes limitations of traditional dynamic scheduling methods into account and develops a systematic adaptive scheduling solution.The methods of robust scheduling,adaptive rescheduling and real time scheduling are then studied using data driven and machine learning techniques.In order to comprehensively improve the proactive decision-making ability,realtime perception ability,dynamic optimization ability and online learning ability of production scheduling,an adaptive scheduling architecture with closed-loop layers and cascaded phases is established.The adaptive scheduling architecture contains three scheduling phases:robust scheduling phase,real-time monitoring and online scheduling phase,and scheduling knowledge online learning/updating phase.Based on the proposed closed-loop and cascaded mechanism,the adaptive scheduling optimization of "robust scheduling before execution-real-time monitoring during execution and online scheduling-scheduling knowledge online learning/updating after execution" is then achieved,which covers the whole process of scheduling execution.The robust scheduling phase aims to generate an initial robust scheduling scheme with anti-interference capability.For its multi-objective optimization problem,a scenario planning-and entropy weight-based multi-objective robust scheduling method(MORO)and a slack-based robust scheduling rule(SR)are proposed.The MORO method uses the scenario planning-based robustness measurement.Firstly,the feasible scheduling strategy set is obtained in the no-disturbance environment.The uncertain processing times of operations are then introduced.And the robust scheduling strategy with anti-interference ability is selected through the entropy weight-based multi-objective robust scheduling model.In the SR method,a scheduling rule considering the robustness of scheduling schemes is designed based on the slack time-based robustness measurement.It is demonstrated by comparing with the common single and multiple objective robust scheduling methods that the proposed method can effectively improve the global robustness of the scheduling strategy on multiple performance indexes without compromising production performance.The real-time monitoring and online scheduling phase aims to achieve timely response to disturbances and adaptive adjustment of scheduling strategies.An adaptive rescheduling approach based on Cyber Physical Production System(CPPS)is defined,which integrates disturbance identification and real-time scheduling.The deep neural network and transfer learning are used to train the performance prediction networks under each working condition.The disturbance online identification driven by perception and prediction is then achieved.In view of the problem of low data utilization rate in data-driven real-time scheduling method,the full-sample-driven scheduling knowledge training method is addressed,and a real-time scheduling method based on Long Short Term Memory(LSTM)is proposed.Compared with other real-time scheduling methods,the effectiveness of the proposed adaptive rescheduling method is verified under different types and intensities of disturbances.The scheduling knowledge online learning/updating phase aims to improve the continuous effectiveness of scheduling knowledge in new production scenarios.To solve the problem of scheduling knowledge online learning and updating,a real-time scheduling method based on online updating machanism of scheduling knowledge(OU-SK)and a real-time scheduling method based on Conventional Neural Networks and Asynchronous Advanced Actor Critic Algorithm(CNN-A3C)are proposed.The OU-SK method focuses on the updating of tabular scheduling knowledge.Three update strategies are designed:adding new knowledge,deleting knowledge and updating knowledge attribute parameters.And the online update of scheduling knowledge is realized by the online update judgment mechanism based on binary tree.The CNN-A3C method focuses on the online learning of network scheduling knowledge.A3C is introduced to train the scheduling knowledge.Furthermore,we design four mechanisms:a slidewindow-based two-dimensional state perception mechanism,an adaptive reward function that considers multiple objectives and automatically adjusts to dynamic events,a continuous action space based on composite dispatching rules and release strategies,and actor-critic networks based on convolutional neural networks(CNNs).Compared with traditional A3C method and scheduling rule method in different disturbance scenarios,the adaptability and superiority of the proposed method in uncertain environment are verified.Based on the adaptive scheduling architecture with closed-loop layers and cascaded phases,an adaptive scheduling prototype system integrating the above scheduling optimization methods is developed.Comprehensive case studies are then carried out on a semiconductor manufacturing system model named MiniFAB and an aircraft engine assembly line named AEAL.The simulation results further verify the effectiveness and the potential of industrial application of the proposed adaptive scheduling architecture with closed-loop layers and cascaded phases.Finally,the research of this thesis is summarized.Promising future research topics on collaborative optimization of multi-scheduling tasks with multi-source complex uncertainties and autonomous scheduling based on multi-agent deep reinforcement learning are discussed.