Model-Driven Learning-Based Memetic Algorithms For Green Distributed Flexible Job Shop Scheduling Problem

Manufacturing industry has a vital impact on the economy,employment and people’s livelihood of industrial countries.Along with the advancement of intelligent manufacturing strategy,enterprises are required to make digital transformation,enhance their production capacity and enhance their international competitiveness through intelligent technology.The core of intelligent manufacturing is intelligent shop scheduling,which assists enterprises to make production scheduling schemes through modeling and intelligent algorithms.This thesis abstracted the actual production scene as a green distributed heterogeneous flexible job shop scheduling problem(GDHFJSP)with minimizing makespan and total energy consumption(TEC).To better understand the features of DHFJSP,this work studies from green flexible job shop scheduling problem(GFJSP)to homogeneous distributed GFJSP(GDFJSP)and then to GDHFJSP.Several MILP models are built,and the model-driven and learning-based memetic algorithms are proposed to solve them.Finally,a practical engineering problem is used to verify the effectiveness of proposed algorithms.Specific research contents are as follows:(1)A GFJSP model was constructed with minimizing the makespan and TEC,and a Q learning and model-based memetic algorithm(LRVMA)was proposed.The main characteristics of LRVMA are as follows: based on problem features,various enhancement strategies are designed according to the knowledge,such as hybrid heuristic initialization strategy,model-driven local search strategy and heuristic greedy energy saving strategy.Combined with Q learning,a parameter self-learning model based on metric feedback is designed to guide the algorithm to evolve in the direction of better distribution.It was tested on the published T2 FJSP benchmark,and it was verified that Q learning parameter adaptive strategy can enhance the population distribution,and greedy energy saving strategy can effectively reduce the TEC in the production workshop.(2)An homegeneous GDFJSP was constructed with minimizing the makespan and TEC,and a surprisingly popular algorithm and model-based meme algorithm(SPAMA)was proposed.The main features of SPAMA are as follows: based on disjunctive graph theory,four kinds of local search operators based on critical path and critical block are designed.Inspired by the decoder theory of full active scheduling,the energy-saving strategy of full active scheduling is designed.Inspired by the emergent popular algorithm,an operator selection model with adaptive probability correction is designed.SPAMA was tested on classical flexible job-shop scheduling benchmarks Mk and DP.It is verified that the success rate of local search can be effectively improved based on critical path design operator,the full active scheduling decoding can effectively reduce the shop energy consumption,and the unexpected popular algorithm can enhance the selection probability of low-weight efficient operator and accelerate the algorithm convergence.(3)A GDHFJSP was constructed to study the minimization of makespan and TEC,and a deep Q-learning and model-based meme algorithm(DQPEA)was proposed.The main characteristics of DQPEA are as follows: By analyzing the characteristics of plant heterogeneity and machine flexibility,nine kinds of knowledge-based local search operators are designed.An improved framework of meme evolution is designed to balance the computational resources of global and local search through multi-population collaboration.An operator selection model based on deep Q learning is proposed.The mapping between solutions and operators is established by neural network to customize the optimal operator for all solutions.DQPEA was tested on 20 different scale test questions,and the effectiveness and robustness of the proposed algorithm were verified by comparing with the latest correlation algorithms.It has good performance in GDHFJSP.(4)The theory and method proposed in this thesis is applied to the blanking workshop scene of a state-owned large-scale engineering equipment production enterprise.Based on the actual engineering data,the problem is abstracted into DHFJSP problem,and the proposed algorithm and improved strategy are used to solve the problem.The results verify the effectiveness of the proposed algorithm on the actual problem.This thesis studies the characteristics of distributed flexible job shop scheduling problem model from GFJSP to homegeneous GDFJSP and then to GDHFJSP,which enricfies the research of flexible job shop scheduling problem.Secondly,the model driven learning MAs are designed to solve different types of problems.The learning strategies range from Q learning to surprisingly popular learning and then to deep Q learning.A large number of experiments have verified the effectiveness of the proposed algorithms and further enriched the research results of meme algorithms in terms of methods.Finally,the improved algorithms are applied to solve the optimization problem of distributed blanking shop of large engineering equipment,and the high efficiency of the improved algorithm is demonstrated.In application,it can provide a reliable optimization tool for distributed blanking shop of large engineering equipment.So the research of this thesis has good theoretical and practical significance.