WO2022099596A1 - 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统 - Google Patents

一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统 Download PDF

Info

Publication number
WO2022099596A1
WO2022099596A1 PCT/CN2020/128622 CN2020128622W WO2022099596A1 WO 2022099596 A1 WO2022099596 A1 WO 2022099596A1 CN 2020128622 W CN2020128622 W CN 2020128622W WO 2022099596 A1 WO2022099596 A1 WO 2022099596A1
Authority
WO
WIPO (PCT)
Prior art keywords
optimization
production
dynamic
scheduling
information
Prior art date
Application number
PCT/CN2020/128622
Other languages
English (en)
French (fr)
Inventor
何再兴
超越
杨东升
杨之乐
Original Assignee
浙江大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 浙江大学 filed Critical 浙江大学
Priority to US17/781,370 priority Critical patent/US20220413455A1/en
Priority to PCT/CN2020/128622 priority patent/WO2022099596A1/zh
Publication of WO2022099596A1 publication Critical patent/WO2022099596A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/41865—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by job scheduling, process planning, material flow
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32247—Real time scheduler
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32252—Scheduling production, machining, job shop
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32329—Real time learning scheduler, uses ANN, fuzzy
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32334—Use of reinforcement learning, agent acts, receives reward

Definitions

  • the invention belongs to the field of industrial personalized customized production system scheduling, and in particular relates to an adaptive learning intelligent scheduling unified computing framework and system for industrial personalized customized production.
  • Personalized customized production has the characteristics of rapid changes in production demand, complex customized processing process, high frequency of dynamic events, many categories, large amount of customized information, and difficulty in coordinated operation of many equipment.
  • the traditional personalized customized production unified computing framework and system are not intelligent enough. Scheduling optimization goals, etc., the existing personalized customization production line only liberates repetitive manual labor, but fails to solve the burden of mental labor. Therefore, on the basis of the existing computing framework and system, it is necessary to introduce the dynamic event classification optimization strategy in the production process and further improve the intelligence level of the personalized customized production line.
  • the purpose of the present invention is to overcome the deficiencies of the prior art, and to provide a self-adaptive learning intelligent scheduling unified computing framework and system for industrial personalized customized production.
  • a two-step method is adopted in static scheduling planning, dynamic scheduling planning and equipment deployment, that is, first classification and then targeted optimization to improve the optimization efficiency and effect, and further enhance the personalization.
  • the intelligent level of customized production forms an intelligent decision-making chain and minimizes labor costs.
  • macro strategy modeling is used to improve the integration of equipment and modules in personalized customized production. coordination between.
  • the present invention adopts the following technical solutions:
  • the system reads the real-time update information of the industrial big data module.
  • the optimal static scheduling planning module in the system generates the optimal static scheduling planning according to the customized task information, equipment information and material information.
  • the system calls the dynamic event interpretation module to dynamically monitor the event queue until a processable dynamic event is read.
  • the dynamic event interpretation module Based on the deep reinforcement learning technology, combined with the data information of the industrial big data module, the dynamic event interpretation module automatically classifies the dynamic events, generates classification labels, and decides the specific content of the corresponding optimization goals and constraints according to the classification labels.
  • the dynamic scheduling optimization module automatically receives the optimization objective decision scheme generated in step 2. Based on the deep reinforcement learning technology, the dynamic scheduling optimization module automatically selects different optimization algorithms according to the classification labels, optimization objectives and constraints, and selects appropriate methods for each algorithm. The parameters are optimized until the corresponding optimization objectives and constraints are satisfied, and the final dynamic scheduling optimization scheme is formed.
  • the scheduling deployment and evaluation module automatically receives the dynamic scheduling optimization scheme generated in step 3, adopts the method of deep neural network, evaluates the optimization scheme, and obtains the optimization scheme from the aspects of efficiency, energy consumption, deployment complexity, stability, etc. Score, if the score is higher than the score standard, the equipment scheduling and deployment optimization will be carried out on the spatial scale, combined with the optimization plan generated in the previous steps, according to the current industrial production situation and the distribution of production line equipment. The corresponding deployment optimization algorithm is automatically selected, and the equipment scheduling sequence is generated.
  • the system automatically deploys automatically according to the deployment sequence generated in step 4, and the industrial big data module receives equipment scheduling change information in real time, and updates the current industrial production environment.
  • the four modules included in the framework proposed by the present invention form a multi-module collaborative dynamic decision-making chain, and the entire process is completely completed by the computer without the participation of staff.
  • the system solves the impact of the current dynamic event, continues the optimized optimal static planning, and calls the dynamic event interpretation module to dynamically monitor the event queue, and performs the cycle from step 2 to step 5 until all the processing is completed. of customized production tasks.
  • the optimal static scheduling planning module includes acquiring the information provided by the industrial big data module; using the deep neural network semi-supervised learning method to extract the characteristics of the industrial production environment, and marking part of the labeled information in advance combined with other unlabeled information. Training, let the semi-supervised model return, learn how to classify information according to features, let the system learn to decide the corresponding static target weight parameters and constraints in different production environments, and select the most matching static optimization algorithm according to the decision content to ensure that the system It can adapt to a variety of complex working conditions in personalized customized production.
  • the whole process adopts the advanced offline generation method, and the optimal static planning of the global production task can be obtained before the production line runs.
  • the system includes a library of static optimization algorithms, including a variety of classic static optimization algorithms, including quality prediction methods, optimization algorithms, heuristic methods, maximum inheritance algorithms, genetic algorithms, etc. and their derived algorithms. Users can also customize according to their own needs and actual production. According to the environmental characteristics, the algorithm library is updated. After the algorithm library is updated, the system will also consider the user's personal algorithm in the decision-making process.
  • the static goals include the smallest maximum completion time, the lowest energy consumption, and the lowest production and processing costs.
  • the decision-making of static goals is mainly the comprehensive weight analysis and optimization of multiple static goals, and the decision-making generates static multi-objective comprehensive weight parameters.
  • the optimal static scheduling plan is a scheduling plan that is generated on the global personalized customized production task and considered from the macro level and conforms to the minimum maximum completion time, the minimum total energy consumption, and the minimum processing cost.
  • the optimization algorithm of the optimal static scheduling plan is independently selected by the system according to the current environment and customized production tasks. It is an important part of the macro strategy modeling of the intelligent production line proposed by the present invention.
  • the dynamic event interpretation module includes monitoring the production data information provided by the industrial big data module, and monitoring dynamic events that can be dynamically scheduled and optimized; based on the deep reinforcement learning technology, it is classified according to the content of the dynamic events, and the current industrial production information is used as the agent.
  • the reinforcement learning environment allows the agent to observe the characteristics of dynamic events in the current environment, set corresponding incentives, and strengthen the action of classifying dynamic events.
  • the categories include production line time deviation, machine failure, material problems, contract events, etc. According to the current industrial production environment data and event classification results provided by the industrial big data module, the multi-objective dynamic weight parameters and corresponding dynamic scheduling optimization under the current event are automatically determined. constraint.
  • Dynamic objectives include the smallest maximum completion time, the lowest energy consumption, the lowest production and processing costs, the highest flexibility in production plan adjustment, and the highest scheduling stability (ie, the lowest change compared to the optimal static scheduling plan), etc.
  • the decision-making of dynamic objectives mainly includes In order to analyze and optimize the comprehensive weight of multiple dynamic targets, the comprehensive weight of multiple dynamic targets is finally generated.
  • the dynamic scheduling optimization module includes a dynamic optimization algorithm selection strategy model and an optimization algorithm parameter adaptive learning method. Considering the characteristics of various types of dynamic events and the difficulty in determining the laws, the system uses deep reinforcement learning to generate dynamic optimization algorithm selection strategy models, and uses supervised learning to allow the system to adaptively match optimization algorithms and parameter optimization methods.
  • the dynamic optimization algorithm selection strategy model uses a variety of optimization algorithms, current event classification labels, and industrial production environment as the agent's reinforcement learning environment, allowing the agent to observe the optimization results gap between different optimization algorithms, set corresponding incentives, and strengthen its optimization algorithm Choose an action.
  • the dynamic scheduling optimization algorithm corresponding to different types of dynamic events is decided according to the situation.
  • the system includes a library of dynamic optimization algorithms, including a variety of classical dynamic optimization algorithms, including multi-agent algorithms, simulated annealing algorithms, hybrid particle swarm optimization, tabu search algorithms, genetic algorithms and their variants. After updating the algorithm library, the user's personal algorithm will also be taken into account in the system decision-making process.
  • the present invention extracts the characteristics of the optimization algorithms by means of deep learning, specifically, by means of deep neural network supervised learning, automatically categorizes the optimization algorithms, and selects a suitable parameter optimization method in the parameter optimization algorithm library. Users can also update the parameter optimization algorithm library, and the system can automatically extract individual algorithm features for classification.
  • the scheduling deployment and planning module includes an adaptive solution evaluation strategy based on deep reinforcement learning, and calculates the score for the received optimization solution, and performs the next scheduling deployment if the score is higher than the specified score threshold.
  • the deep learning method is used to extract the production environment and the characteristics of the scheduling scheme, adopt reinforcement learning to independently decide the weight parameters and related constraints of the current deployment optimization target, and select an appropriate optimization algorithm to finally generate a deployment sequence.
  • the system includes a deployment optimization algorithm library, which contains a variety of classic equipment deployment optimization algorithms, including adaptive genetic algorithm, ant colony optimization algorithm, multi-objective genetic algorithm, particle swarm optimization and other algorithm categories, as well as its derivative algorithms.
  • the algorithm library can be updated according to its own needs and the characteristics of the actual production environment. After the algorithm library update is completed, the user's personal algorithm will also be taken into account in the system decision-making process.
  • the optimization objectives considered in the deployment optimization process of the present invention mainly include transportation overhead, equipment delay overhead, production capacity, deployment complexity, etc.
  • the system proposed by the present invention mainly makes decisions based on the weight ratio of multiple optimization objectives.
  • the real-time update of the industrial big data module refers to all data information generated and monitored in the actual industrial production process, including dynamic event information, material information, production equipment information, personalized customized production task information, and current factory operation information.
  • the multi-module collaborative dynamic decision-making chain refers to the unified computing framework proposed by the present invention, which can be formed only by the industrial big data module receiving external data in real time, combined with the dynamic event interpretation module, dynamic scheduling optimization module, scheduling deployment and evaluation module.
  • the closed-loop decision-making path all decision-making tasks are automatically completed by the computer if there is no manual issuance of high-level decision-making instructions in the whole process.
  • the unified computing framework proposed by the present invention only needs to receive data to complete the automatic planning and deployment of equipment and processes on the intelligent production line.
  • the macro strategy modeling means that the operation and optimization of the intelligent production line can be considered from the overall production environment of the entire factory, instead of focusing only on the current process or current events. In the process of dealing with random dynamic events, it is still Pay attention to the entire production environment and individual custom production tasks. It includes macro-strategy modeling in terms of production tasks and macro-strategy modeling in terms of factors of production.
  • the macro strategy modeling of production tasks means that the entire framework first automatically selects a static optimization algorithm according to the global production task provided by the user after the task is initialized, and generates an optimal static plan based on the global production task, which is then optimized in the subsequent local dynamic scheduling.
  • the optimal static planning will be referred to, and changes to the global optimal static planning will be minimized to ensure that the overall production tasks can be taken into account during the operation of the overall framework.
  • the macro-strategy modeling in terms of production factors means that during the operation of the computing framework, the information of the industrial big data module will be updated in real time, and other modules can access the current information at any time.
  • the information contains all the information of the entire factory production environment.
  • the present invention proposes In the running process, the framework is not limited to the information of some equipment or processes involved in the current event, but will refer to the current information of the entire production environment, so as to consider the impact on other production processes and equipment.
  • the present invention in view of the characteristics of high frequency of dynamic events and various types of events in personalized customized production, classifies the three levels of static scheduling planning, dynamic scheduling planning and equipment deployment on the production line, and then optimizes them in a targeted manner.
  • deep reinforcement learning semi-supervised learning, supervised learning and other deep neural network adaptive learning technologies, and proposed a unified computing framework for intelligent personalized customized production lines with fully automatic adaptive learning.
  • the present invention has the following beneficial effects:
  • the unified computing framework and system proposed by the present invention adopts a multi-step method to perform static scheduling optimization, dynamic scheduling optimization and equipment deployment optimization, and can select the static scheduling optimization problem and planning deployment problem in a targeted manner in the face of different situations. Different optimization algorithms are used for optimization to improve the efficiency and effect of optimization.
  • the closed-loop intelligent decision-making chain of the multi-module collaborative mechanism proposed by the present invention can realize the full operation of the personalized customization production line without the high-level external manual intervention. automatic running.
  • the present invention proposes a macro strategy model, which considers the overall production tasks and production data in the scheduling process, improves the coordination between modules and equipment in the framework, and is more adaptable to complex and diverse personalized customized production tasks.
  • Fig. 1 is the system operation flow chart of the present invention
  • Fig. 2 is the framework model diagram of the present invention
  • Figure 3 is a schematic diagram of a static optimization algorithm based on SSL
  • Fig. 4 is the schematic diagram of dynamic event classification based on DRL
  • Fig. 5 is the principle diagram of optimization algorithm selection based on DRL
  • Fig. 6 is the Gantt chart of optimization result example
  • Figure 7 is a schematic diagram of a multi-module collaborative closed-loop decision chain.
  • the present invention establishes a unified calculation method for personalized customized production adaptive learning intelligent scheduling Framework and system, as shown in Figure 2, the framework is mainly composed of the following five modules:
  • Optimal static scheduling planning module Combined with industrial production environment and customized production tasks, it can make independent decisions to generate static optimization goals and constraints, and select appropriate algorithms from the static optimization algorithm library to generate the global optimal static state offline before production starts. Scheduling planning.
  • Dynamic event interpretation module dynamically monitor the event sequence in the production process, generate different classification labels for dynamic event information, and automatically determine the corresponding optimization goals and constraints based on the classification results;
  • Dynamic scheduling optimization module According to the optimization objectives and constraints generated by the dynamic event interpretation module, the corresponding optimization methods are automatically selected, so as to obtain better optimization results on the basis of satisfying the optimization objectives and constraints. Adaptive learning optimization parameters through deep learning technology and generating the final optimization plan;
  • Scheduling deployment and evaluation module Receive the optimization scheme generated by the dynamic scheduling optimization module and evaluate it. If the evaluation result meets the conditions, perform scheduling and deployment optimization and generate a deployment sequence;
  • Real-time update of industrial big data module including dynamic event information, operation information, material information, equipment information, customized task information and other factory data, it is the medium of collaborative operation between modules, and provides data foundation for other modules.
  • the present invention proposes a framework technology route as follows:
  • the system reads the real-time update information of the industrial big data module.
  • the optimal static scheduling planning module in the system generates the optimal static scheduling planning according to the customized task information, equipment information and material information.
  • the system calls the dynamic event interpretation module to dynamically monitor the event queue until a processable dynamic event is read.
  • this step specifically after the present invention starts to run, first interpret relevant information, extract characteristics of industrial production environment based on semi-supervised learning of deep neural network, and perform training by labeling a part of labeled information in advance and combining other unlabeled information. , let the semi-supervised model return, learn how to classify information according to features, and the classification result is the corresponding algorithm serial number.
  • the schematic diagram is shown in Figure 3. According to the actual production situation and personalized customized production tasks, the static optimization goals and constraints are independently determined. Then, an appropriate static optimization algorithm is selected, and the optimal static scheduling plan is generated in combination with the global production tasks. If there is no dynamic event disturbance during the production process, the production line will run according to the optimal static plan.
  • the dynamic event interpretation module monitors the dynamic events in the production process in real time. Go to the next step after detecting a dynamic event that can be processed.
  • the static objectives include the smallest maximum completion time, the lowest energy consumption, and the lowest production and processing cost.
  • the decision of the static objective is the comprehensive weight analysis and optimization of multiple static objectives, and the decision generates static multi-objective comprehensive weight parameters.
  • the dynamic event interpretation module Based on the deep reinforcement learning technology, combined with the data information of the industrial big data module, the dynamic event interpretation module automatically classifies the dynamic events, generates classification labels, and decides the specific content of the corresponding optimization goals and constraints according to the classification labels.
  • the monitored dynamic events are first interpreted.
  • the principle model diagram As shown in Figure 4, the current industrial production information is used as the agent's reinforcement learning environment, so that the agent can observe the characteristics of dynamic events in the current environment, set corresponding incentives, and strengthen its dynamic event classification actions. Generate classification labels for events by interpreting the characteristics of dynamic events, including production line time deviations, machine failures, material problems, contract events, etc., or generate multiple labels simultaneously under one dynamic event.
  • the multi-objective dynamic optimization weight parameters and corresponding constraints are automatically determined according to the generated tags and combined with the current industrial production situation information.
  • the dynamic optimization equation and the constraint equation together constitute the dynamic scheduling optimization scheme.
  • the dynamic scheduling optimization module automatically receives the optimization objective decision scheme generated in step 2. Based on the deep reinforcement learning technology, the dynamic scheduling optimization module automatically selects different optimization algorithms according to the classification labels, optimization objectives and constraints, and selects appropriate methods for each algorithm. The parameters are optimized until the corresponding optimization objectives and constraints are satisfied, and the final dynamic scheduling optimization scheme is formed.
  • this step specifically, after receiving the optimization scheme generated in the previous steps, first select the corresponding optimization algorithm through the dynamic optimization algorithm selection strategy model.
  • the principle is shown in Figure 5.
  • Information is used as an agent reinforcement learning environment, allowing the agent to observe the optimization results gap between different optimization algorithms, set corresponding incentives, and strengthen the actions selected by its optimization algorithm.
  • the goal of the dynamic scheduling optimization algorithm corresponding to different types of dynamic events is achieved.
  • the characteristics of the optimization algorithms are extracted by means of deep neural network supervised learning, the optimization algorithms are automatically classified, and an appropriate parameter optimization method is selected from the parameter optimization algorithm library. Users can also update the parameter optimization algorithm library, and the system can automatically extract individual algorithm features for classification. Finally, the parameter optimization of the optimization algorithm is realized, and the final dynamic scheduling optimization scheme is generated.
  • the related optimization example Gantt chart is shown in Figure 6
  • the scheduling deployment and evaluation module automatically receives the dynamic scheduling optimization scheme generated in step 3, adopts the method of deep neural network, evaluates the optimization scheme, and obtains the optimization scheme from the aspects of efficiency, energy consumption, deployment complexity, stability, etc. Score, if the score is higher than the score standard, the equipment scheduling and deployment optimization will be carried out on the spatial scale, combined with the optimization plan generated in the previous steps, according to the current industrial production situation and the distribution of production line equipment. The corresponding deployment optimization algorithm is automatically selected, and the equipment scheduling sequence is generated.
  • the scheduling deployment and evaluation module first generates an adaptive evaluation scheme through autonomous decision-making, aiming at the optimization scheme, in terms of efficiency, energy consumption, planning and deployment flexibility If the score exceeds the score threshold, the equipment deployment will be optimized for the dynamic scheduling optimization scheme.
  • the deployment process refer to the static scheduling scheme that has been generated or Dynamic scheduling scheme, combined with the current industrial production environment information, through deep reinforcement learning technology, the schematic diagram is shown in Figure 5, taking the current industrial production environment and scheduling scheme as the environment, using deep learning to extract the characteristics of the production environment and scheduling scheme, using reinforcement
  • the learning method autonomously decides the weight parameters and related constraints of the current deployment optimization target, and selects an appropriate optimization algorithm, and finally generates a deployment sequence.
  • the system automatically deploys automatically according to the deployment sequence generated in step 4, and the industrial big data module receives equipment scheduling change information in real time, and updates the current industrial production environment.
  • the four modules included in the framework proposed by the present invention form a multi-module collaborative dynamic decision-making chain, and the entire process is completely made by the computer to complete the decision without the participation of staff.
  • the system solves the impact of the current dynamic event, continues the optimized optimal static planning, and calls the dynamic event interpretation module to dynamically monitor the event queue, and performs the cycle from step 2 to step 5 until all the processing is completed. of customized production tasks.
  • the multi-module collaborative dynamic decision-making chain from the scheduling deployment and evaluation module to the industrial big data module builds a closed-loop decision-making path.
  • the intelligent level of the production line is improved, the manual decision-making labor intensity of the production line is reduced, and it is more suitable for the characteristics of complex customized production and processing processes and complex dynamic disturbances.
  • the personalized customized production macro strategy model is established from the two aspects of production tasks and production data, so that the system is more suitable for the coordinated operation of equipment in personalized customized production. It is characterized by high difficulty and large amount of information on production tasks.
  • the present invention is based on deep neural network and reinforcement learning, and is suitable for the characteristics of complex process in personalized customized production, large amount of customized information, many types of dynamic events and high frequency, and can be used in static scheduling planning, dynamic scheduling planning and equipment deployment.
  • a multi-step approach is adopted, that is, first classification and then targeted optimization, to improve the optimization efficiency and effect, and propose a unified computing framework for intelligent personalized customized production lines with fully automatic adaptive learning, which can effectively improve industrial personalization. Customize productivity and minimize labor costs.

Landscapes

  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Automation & Control Theory (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Manufacturing & Machinery (AREA)
  • Quality & Reliability (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • General Factory Administration (AREA)

Abstract

一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，基于深度神经网络和强化学习，以底层的工业大数据模块为信息基础，首先面向全局定制生产任务自动决策选择最合适的优化算法，生成全局最优静态调度规划；实时监听工厂当前动态事件；若未监测到需要进行动态调度优化的动态事件，则顺序执行全局最优静态规划；当监测到需要进行动态调度优化的动态事件冲击时，对当前动态事件信息进行解读并分类，针对不同的动态事件类型自动匹配相应的优化算法进行动态调度优化；后续模块对动态调度优化方案进行评估，根据评估结果选择重新生成优化方案或者基于方案自动决策选择最合适的优化算法，并生成设备部署序列进行自动部署。基于深度强化学习，针对个性化定制生产中工序复杂，定制信息量大，动态事件类别多、频率高的特点，在静态调度规划、动态调度规划以及设备部署等三方面均采用两步走的方式，即先分类决策再有针对性地优化，提高优化效率和效果，提出了全自动自适应学习的智能个性化定制生产线统一计算框架及系统，该系统更加适应个性化定制生产的特点，能够有效提升个性化定制生产效率，并最大限度降低人工决策成本。

Description

一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统 技术领域
本发明属于工业个性化定制生产系统调度领域，尤其涉及一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统。
背景技术
这里的陈述仅提供与本发明相关的背景技术，而不必然地构成现有技术。
当前智能制造技术的兴起让工业发达国家纷纷意识到制造业的广阔前景，中国政府发布《中国制造2025》来指引我国的经济转型发展，通过建设智能制造系统以促进制造产业发展。而个性化定制生产是智能制造的一种极为重要的模式，在个性化定制生产任务中，用户会介入产品的生产过程，以获得自己定制的个人属性强烈的商品或获得与其个人需求匹配的产品或服务。随着经济发展以及人们的多元消费需求增加，个性化定制服务逐渐成为客户的主流选择。
个性化定制生产具有生产需求变化快、定制加工流程复杂、动态事件频率高，类别多、定制信息量大，众多设备协调运作难度大等特点。发明人发现，在传统的个性化定制生产统一计算框架中，面对多种生产需求，往往只采用统一的调度优化方式进行优化，难以很好地适应个性化定制生产中生产过程高度离散、复杂以及动态事件量多且高频的特点，在实际应用场景中，传统个性化定制生产统一计算框架及系统，智能化程度不够，面对繁杂多样的定制信息，往往需要人工决策生产调度规划，制定调度优化目标等，现有的个性化定制生产线仅仅解放了重复性体力劳动，而没能很好地解决脑力劳动的负担。因此在现有计算框架及系统的基础上，引入生产过程中动态事件分类优化策略，并进一步提升个性化定制生产线智能化水平很有必要。
发明内容
本发明的目的是为了克服现有技术的不足，提供一种面向工业个性化定制 生产的自适应学习智能调度统一计算框架及系统，针对个性化定制生产中工序复杂，定制信息量大，动态事件类别多、频率高的特点，在静态调度规划、动态调度规划以及设备部署等三方面均采用两步的方式，即先分类再有针对性地优化，提高优化效率和效果，同时进一步提升个性化定制生产的智能化水平，形成智能决策链，最大程度降低人工劳动成本，针对定制信息量大、设备协调运作难度大等特点，通过宏观策略建模，提升个性化定制生产各设备及各模块之间的协调性。
为实现上述目的，本发明采用如下技术方案：
S1、系统读取实时更新工业大数据模块的信息，在生产开始前，系统中最优静态调度规划模块就根据其中的定制任务信息、设备信息、物料信息生成最优静态调度规划，加工开始后，系统调用动态事件解读模块动态监听事件队列，直到读取到可处理的动态事件。
S2、基于深度强化学习技术，结合工业大数据模块的数据信息，动态事件解读模块对动态事件进行自动分类，生成分类标签，并根据分类标签决策相应的优化目标及约束的具体内容。
S3、动态调度优化模块自动接收步骤2生成的优化目标决策方案，基于深度强化学习技术，动态调度优化模块根据分类标签以及优化目标和约束自动选择不同的优化算法，并选择合适的方法对算法各参数进行优化，直到满足对应的优化目标和约束，形成最终的动态调度优化方案。
S4、调度部署及评估模块自动接收步骤3生成的动态调度优化方案，采用深度神经网络的方法，评估优化方案，从效率、能耗、部署复杂度、稳定性等多方面计算获取到优化方案的得分，若得分高于分数标准，则在空间尺度上进行设备调度部署优化，结合前述步骤生成的优化方案，根据当前工业生产情况及产线设备分布情况。自动选择相应部署优化算法，并生成设备调度序列。
S5、系统自动根据步骤4生成的部署序列进行自动部署，工业大数据模块实时接收设备调度变化信息，对当前工业生产环境进行更新。至此，本发明所提出的框架包含的四个模块形成多模块协同动态决策链，整个流程完全由计算 机完成决策无需工作人员参与。完成一次循环后，系统解决当前动态事件带来的影响，继续延续优化后的最优静态规划，并调用动态事件解读模块动态监听事件队列，进行从步骤2到步骤5的循环，直到处理完所有的个性化定制生产任务。
所述最优静态调度规划模块，包括获取工业大数据模块提供的信息；采用深度神经网络半监督学习方式，提取工业生产环境特征，通过提前标注一部分有标签的信息结合其他没有进行标注的信息进行训练，让半监督模型回归，学习如何根据特征对信息进行分类，让系统学习在不同生产环境下决策出相应的静态目标权重参数和约束，根据决策内容，选择最匹配的静态优化算法，保证系统能够自适应个性化定制生产中的多种复杂工况。整个过程采用超前离线式生成的方式，在生产线运行前即可得到全局生产任务的最优静态规划。系统包含静态优化算法库，内含多种经典静态优化算法，包括质量预测法、最优化算法、启发式方法、最大继承算法、遗传算法等及其衍生算法，用户也可根据自身需求和实际生产环境特征，对算法库进行更新，完成算法库更新后系统在决策过程中也会将用户个人算法考虑在内。
静态目标包括最大完工时间最小、能源消耗最低以及生产加工成本最低等，静态目标的决策主要为对多静态目标的综合权重分析及优化，决策生成静态多目标综合权重参数。
所述最优静态调度规划，是在全局个性化定制生产任务上生成的从宏观层面考虑的符合最大完工时间最小、能源消耗总和最低、加工成本最低的调度规划方案。最优静态调度规划的优化算法由系统根据当前环境和个性化定制生产任务自主决策选择。是本发明提出的智能生产线宏观策略建模重要组成部分。
所述动态事件解读模块，包括监听工业大数据模块提供的生产数据信息，监测可进行动态调度优化的动态事件；基于深度强化学习技术根据动态事件的内容进行分类，以当前工业生产信息作为智能体强化学习环境，令智能体观察当前环境下动态事件的特征，设置相应的激励，强化其动态事件分类的动作。类别包括生产线时间偏差、机器故障、物料问题、合同事件等；根据工业大数 据模块提供的当前工业生产环境数据以及事件分类结果，自动决策出当前事件下动态调度优化的多目标动态权重参数和对应约束。
动态目标包括最大完工时间最小、能源消耗最低、生产加工成本最低、生产计划调整灵活性最高、调度稳定性最高(即相较于最优静态调度规划的变化最低)等，动态目标的决策主要包括为对多动态目标的综合权重分析及优化，最终生成多动态目标综合权重。
多动态目标综合权重，其形式如下式表示：
minZ＝α·f norm(T)+β·f norm(E)+γ·f norm(C)+(1-δ·f norm(P))
+(1-ε·f norm(Q))
其中α表示时间权重参数；β表示能耗权重参数；γ表示成本权重参数；δ表示灵活性权重参数；ε表示稳定性权重参数；T表示车间生产时间，E表示车间生产能耗，C表示车间生产成本，P表示调度方案灵活性，Q表示调度方案稳定性，f norm(x)表示归一化函数，将时间、能耗、成本、灵活性和稳定性等5个指标量纲统一，且函数值在0到1之间。各权重参数值也在0到1之间。
所述动态调度优化模块包含动态优化算法选择策略模型以及优化算法参数自适应学习方法。考虑到动态事件种类多样、规律难以确定的特点，本系统采用深度强化学习的方式生成动态优化算法选择策略模型，并以监督学习的方式让系统自适应匹配优化算法与参数优化方法。
动态优化算法选择策略模型以多种优化算法和当前事件分类标签、工业生产环境作为智能体强化学习环境，令智能体观察不同优化算法之间的优化结果差距，设置相应的激励，强化其优化算法选择动作。分情况决策不同类别动态事件对应的动态调度优化算法。系统包含动态优化算法库，内含多种经典动态优化算法，包括多智能体算法、模拟退火算法、混合粒子群算法、禁忌搜索算法以及遗传算法及其变种，用户可根据自身需求和实际生产环境特征，对算法库进行更新，完成算法库更新后，系统决策过程中也会将用户个人算法考虑在内。
针对不同的优化算法，本发明通过深度学习的方式来提取优化算法的特征， 具体以深度神经网络监督学习的方式，对优化算法自动归类，在参数优化算法库中选择合适的参数优化方法。用户也可以对参数优化算法库进行更新，系统可自动提取个人算法特征进行分类。
所述调度部署及规划模块，包括基于深度强化学习的自适应方案评估策略，针对接收到的优化方案计算其得分，分数高于规定分数阈值则进行下一步的调度部署。在部署过程中，参照前述步骤生成的静态调度方案或动态调度方案，结合当前工业生产环境信息，通过深度强化学习技术，以当前工业生产环境和调度方案为环境，采用深度学习的方式提取生产环境及调度方案的特征，采用强化学习的方式自主决策当前的部署优化目标权重参数和相关约束，并选择合适的优化算法，最终生成部署序列。系统包含部署优化算法库，内含多种经典的设备部署优化算法，包括自适应遗传算法、蚁群优化算法、多目标遗传算法、粒子群算法等算法大类，还包含其衍生算法，用户也可根据自身需求和实际生产环境特征，对算法库进行更新，完成算法库更新后，系统决策过程中也会将用户个人算法考虑在内。
本发明在部署优化过程中考虑的优化目标主要包括运输开销、设备时延开销、生产能力、部署复杂度等方面，本发明所提出的系统主要在多优化目标的权重比例做决策。
所述实时更新工业大数据模块是指实际工业生产过程中产生和监测到的所有数据信息，包括动态事件信息、物料信息、生产设备信息、个性化定制生产任务信息、工厂当前运营信息等。
所述多模块协同动态决策链，是指本发明提出的统一计算框架，只需工业大数据模块实时接收外部数据，结合动态事件解读模块、动态调度优化模块、调度部署及评估模块，即可形成闭环决策通路，整个过程中若无人工发出高级决策指令，全部决策任务均由计算机自动完成。依靠闭环决策链，本发明提出的统一计算框架只需接收数据，即可完成智能生产线上的设备及工序的自动规划部署。
所述宏观策略建模，是指从整个工厂生产环境的整体来可考虑智能生产线 的运行及优化，而非将注意力仅放在当前工序或者当前事件上，在处理随机动态事件过程中，仍要注意整个生产环境及个性化定制生产任务。包括生产任务方面的宏观策略建模和生产要素方面的宏观策略建模。
生产任务方面的宏观策略建模是指整个框架在任务初始化后首先根据用户提供的全局生产任务自动选择静态优化算法，并生成一个基于全局生产任务的最优静态规划，在后续的局部动态调度优化调整中，将会参考最优静态规划，尽量减小对全局最优静态规划的改变，以保证整体框架运行中能够兼顾总体生产任务。
生产要素方面的宏观策略建模是指在计算框架运行过程中，会实时更新工业大数据模块的信息，其他模块可以随时取用当前信息，该信息包含整个工厂生产环境的全部信息，本发明提出的框架在运行过程中不仅仅局限于当前事件所涉及到部分设备或工序的信息，而会参考当前整个生产环境的信息，从而考虑对其他生产工序及设备的影响。
针对个性化定制生产中工序复杂，定制信息量大，动态事件类别多、频率高的特点，在静态调度规划、动态调度规划以及设备部署等三方面均采用两步走的方式，即先分类再有针对性地优化，
本发明基于深度学习，针对个性化定制生产中动态事件发生频率高、事件类型多样的特点，对生产线上在静态调度规划、动态调度规划以及设备部署等三个层级进行分类再有针对性地优化，结合深度强化学习、半监督学习、监督学习等多种深度神经网络自适应学习技术，提出了全自动自适应学习的智能个性化定制生产线统一计算框架。
由于采用上述技术方案，本发明具有以下有益效果：
本发明提出的统一计算框架及系统，采用多步走的方法进行静态调度优化、动态调度优化以及设备部署优化，对静态调度优化问题和规划部署问题，面对不同情况，能够有针对性地选择不同的优化算法进行优化，提升优化的效率和效果。
针对传统个性化定制生产线智能化水平不足，人工决策劳动成本高的特点， 本发明提出的多模块协同机制闭环智能决策链，在无外部人工高级干预的情况下，能够实现个性化定制生产线的全自动运行。
此外本发明提出了宏观策略模型，在调度过程中考虑整体生产任务和生产数据，提升该框架内各模块以及各设备间的协调性，同时更适应复杂多样的个性化定制生产任务。
附图说明
构成本申请的一部分的说明书附图用来提供对本申请的进一步理解，本申请的示意性实施例及其说明用于解释本申请，并不构成对本申请的限定。
图1是本发明系统运行流程图；
图2是本发明框架模型图；
图3基于SSL的静态优化算法原理图；
图4是基于DRL的动态事件分类原理图；
图5是基于DRL的优化算法选择原理图；
图6是优化结果实例甘特图；
图7是多模块协同闭环决策链示意图。
具体实施方式
应该指出，以下详细说明都是例示性的，旨在对本申请提供进一步的说明。除非另有指明，本文使用的所有技术和科学术语具有与本申请所属技术领域的普通技术人员通常理解的相同含义。
以下结合附图对本发明作进一步的详细描述。应当理解，此处所描述的具体实施例仅用以解释本发明，并不用于限定本发明。
参照图1，按照本发明的发明内容完整方法实施的实施例机器实施过程如下：
针对个性化定制生产中工序流程复杂、定制生产信息量大，动态事件频率高、类型多以及当前生产线智能程度不足的情况，本发明建立了一种面向个性化定制生产自适应学习智能调度统一计算框架及系统，如图2所示，该框架主要由以下五个模块组成：
最优静态调度规划模块：结合工业生产环境及个性化定制生产任务，自主决策生成静态优化目标及约束，并在静态优化算法库中选择合适的算法，在生产开始前离线式生成全局最优静态调度规划。
动态事件解读模块：动态监听生产过程中的事件序列，针对动态事件信息生成不同的分类标签，基于分类结果自动决策出相应的优化目标以及约束；
动态调度优化模块：根据动态事件解读模块生成的优化目标和约束，自动选择相应的优化方法，以便在满足优化目标和约束的基础上获得更好的优化效果，在选择相应优化方法的基础上，通过深度学习技术自适应学习优化参数并生成最终的优化方案；
调度部署及评估模块：接收动态调度优化模块生成的优化方案，并进行评估，若评估结果满足条件，则进行调度部署优化，并生成部署序列；
实时更新工业大数据模块：包含动态事件信息、运营信息、物料信息、设备信息以及定制任务信息等工厂数据，是各模块间协同作业的媒介，为其他模块提供数据基础。
本发明提出框架技术路线如下：
S1、系统读取实时更新工业大数据模块的信息，在生产开始前，系统中最优静态调度规划模块就根据其中的定制任务信息、设备信息、物料信息生成最优静态调度规划，加工开始后，系统调用动态事件解读模块动态监听事件队列，直到读取到可处理的动态事件。
在本步骤中，具体的在本发明开始运行后，首先解读相关信息，基于深度神经网络半监督学习，提取工业生产环境特征，通过提前标注一部分有标签的信息结合其他没有进行标注的信息进行训练，让半监督模型回归，学习如何根据特征对信息进行分类，分类结果即为对应算法序号，原理图如图3所示，根据实际生产情况和个性化定制生产任务自主决策静态优化目标和约束，进而选择合适的静态优化算法，结合全局生产任务生成最优静态调度规划，若在生产过程中没有动态事件扰动，生产线将会按照最优静态规划运行。在运行过程中，由动态事件解读模块实时监听生产过程中的动态事件。在监测到可处理的动态 事件后进入下一步。
静态目标包括最大完工时间最小、能源消耗最低以及生产加工成本最低等，静态目标的决策为对多静态目标的综合权重分析及优化，决策生成静态多目标综合权重参数。
S2、基于深度强化学习技术，结合工业大数据模块的数据信息，动态事件解读模块对动态事件进行自动分类，生成分类标签，并根据分类标签决策相应的优化目标及约束的具体内容。
在本步骤中，具体的，当监听到可处理的动态事件后，由于个性化定制生产过程中动态事件种类多样，所以先对监听到的动态事件进行解读，基于深度强化学习技术，原理模型图如图4，以当前工业生产信息作为智能体强化学习环境，令智能体观察当前环境下动态事件的特征，设置相应的激励，强化其动态事件分类的动作。通过解读动态事件的特征来生成事件的分类标签，标签包括生产线时间偏差、机器故障、物料问题、合同事件等，或者在一个动态事件下同时生成多个标签。在本步骤中，还会根据生成的标签，结合当前工业生产情况信息，自动决策出多目标动态优化权重参数和对应约束。动态优化方程和约束方程共同组成动态调度优化方案。
多目标动态权重综合优化的形式，可参考以下公式：
minZ＝α·f norm(T)+β·f norm(E)+γ·f norm(C)+(1-δ·f norm(P))
+(1-ε·f norm(Q))
其中α表示时间权重参数；β表示能耗权重参数；γ表示成本权重参数；δ表示灵活性权重参数；ε表示稳定性权重参数；T表示车间生产时间，E表示车间生产能耗，C表示车间生产成本，P表示调度方案灵活性，Q表示调度方案稳定性，f norm(x)表示归一化函数，将时间、能耗、成本、灵活性和稳定性等5个指标量纲统一，且函数值在0到1之间。各权重参数值也在0到1之间。
S3、动态调度优化模块自动接收步骤2生成的优化目标决策方案，基于深度强化学习技术，动态调度优化模块根据分类标签以及优化目标和约束自动选择不同的优化算法，并选择合适的方法对算法各参数进行优化，直到满足对应 的优化目标和约束，形成最终的动态调度优化方案。
在本步骤中，具体的，在接收到前序步骤生成的优化方案之后，首先通过动态优化算法选择策略模型选择对应的优化算法，其原理如图5所示，以当前事件分类标签、工业生产信息作为智能体强化学习环境，令智能体观察不同优化算法之间的优化结果差距，设置相应的激励，强化其优化算法选择的动作。达到分情况决策不同类别动态事件对应动态调度优化算法的目标。
在本步骤中，针对不同的优化算法，以深度神经网络监督学习的方式提取优化算法的特征，对优化算法自动归类，在参数优化算法库中选择合适的参数优化方法。用户也可以对参数优化算法库进行更新，系统可自动提取个人算法特征进行分类。最终实现优化算法的参数优化，生成最终的动态调度优化方案。相关的优化实例甘特图如图6
S4、调度部署及评估模块自动接收步骤3生成的动态调度优化方案，采用深度神经网络的方法，评估优化方案，从效率、能耗、部署复杂度、稳定性等多方面计算获取到优化方案的得分，若得分高于分数标准，则在空间尺度上进行设备调度部署优化，结合前述步骤生成的优化方案，根据当前工业生产情况及产线设备分布情况。自动选择相应部署优化算法，并生成设备调度序列。
在本步骤中，具体的，在获取到前述步骤生成的动态调度优化方案后，调度部署及评估模块首先通过自主决策生成的自适应评估方案，针对优化方案，在效率、能耗、规划部署灵活性以及相较于最优静态调度方案变化稳定性等方面进行评分，若得分超过分数阈值，则针对动态调度优化方案进行设备部署方面的优化，在部署过程中，参照已经生成的静态调度方案或动态调度方案，结合当前工业生产环境信息，通过深度强化学习技术，原理图如图5，以当前工业生产环境和调度方案为环境，采用深度学习的方式提取生产环境及调度方案的特征，采用强化学习的方式自主决策当前的部署优化目标权重参数和相关约束，并选择合适的优化算法，最终生成部署序列。
S5、系统自动根据步骤4生成的部署序列进行自动部署，工业大数据模块实时接收设备调度变化信息，对当前工业生产环境进行更新。至此，本发明所 提出的框架包含的四个模块形成多模块协同动态决策链，整个流程完全由计算机完成决策无需工作人员参与。完成一次循环后，系统解决当前动态事件带来的影响，继续延续优化后的最优静态规划，并调用动态事件解读模块动态监听事件队列，进行从步骤2到步骤5的循环，直到处理完所有的个性化定制生产任务。
在本步骤中，具体的，在得到设备部署序列后，进行自动化部署，并将当前生产情况改变信息实时更新至工业大数据模块，形成从工业大数据模块，动态事件解读模块，动态优化调度模块，调度部署及评估模块再到工业大数据模块的多模块协同动态决策链，如图7，构建闭环决策通路，整个过程中若无人工发出高级决策指令，全部决策任务均由计算机自动完成。依靠闭环决策链，提升产线智能化水平，降低生产线的人工决策劳动强度，且更适应个性化定制生产加工流程复杂、动态扰动繁杂的特点。
上述实施中，分别通过建立全局最优静态规划以及工业大数据模块实时更新，从生产任务和生产数据两个方面建立个性化定制生产宏观策略模型，让系统更加适应个性化定制生产中设备协调运作难度大以及生产任务信息量大的特点。
由此实施可见，本发明基于深度神经网络及强化学习，针对个性化定制生产中工序复杂，定制信息量大，动态事件类别多、频率高的特点，在静态调度规划、动态调度规划以及设备部署等三方面均采用多步走的方式，即先分类再有针对性地优化，提高优化效率和效果，提出了全自动自适应学习的智能个性化定制生产线统一计算框架，能够有效提升工业个性化定制生产效率，并最大限度降低人工成本。
以上仅为本发明的具体实施例，但本发明的技术特征并不局限于此。任何以本发明为基础，为解决基本相同的技术问题，实现基本相同的技术效果，所作出地简单变化、等同替换或者修饰等，皆涵盖于本发明的保护范围之中。

Claims (8)

  1. 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于，包括以下步骤：
    S1、系统读取实时更新工业大数据模块的信息，在生产开始前，系统中最优静态调度规划模块就根据其中的定制任务信息、设备信息、物料信息生成最优静态调度规划，加工开始后，系统调用动态事件解读模块动态监听事件队列，直到读取到可处理的动态事件；
    S2、基于深度强化学习技术，结合工业大数据模块的数据信息，动态事件解读模块对动态事件进行自动分类，生成分类标签，并根据分类标签决策相应的优化目标及约束的具体内容；
    S3、动态调度优化模块自动接收步骤2生成的优化目标决策方案，基于深度强化学习技术，动态调度优化模块根据分类标签以及优化目标和约束自动选择不同的优化算法，并选择合适的方法对算法各参数进行优化，直到满足对应的优化目标和约束，形成最终的动态调度优化方案；
    S4、调度部署及评估模块自动接收步骤3生成的动态调度优化方案，采用深度神经网络的方法，评估优化方案，从效率、能耗、部署复杂度、稳定性等多方面计算获取到优化方案的得分，若得分高于分数标准，则在空间尺度上进行设备调度部署优化，结合前述步骤生成的优化方案，根据当前工业生产情况及产线设备分布情况。自动选择相应部署优化算法，并生成设备调度序列；
    S5、系统自动根据步骤4生成的部署序列进行自动部署，工业大数据模块实时接收设备调度变化信息，对当前工业生产环境进行更新。至此，本发明所提出的框架包含的四个模块形成多模块协同动态决策链，整个流程完全由计算机完成决策无需工作人员参与；完成一次循环后，系统解决当前动态事件带来的影响，继续延续优化后的最优静态规划，并调用动态事件解读模块动态监听事件队列，进行从步骤2到步骤5的循环，直到处理完所有的个性化定制生产任务；
  2. 根据权利要求1所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述最优静态调度规划模块，包括获取工业大数据模块提供的信息；采用深度神经网络半监督学习方式，提取工业生产环境特征，通过提前标注一部分有标签的信息结合其他没有进行标注的信息进行训练，让半监督模型回归，学习如何根据特征对信息进行分类，从而让系统在不同生产环境下决策出相应的静态目标权重参数和约束，根据决策内容，选择最匹配的静态优化算法，保证系统能够自适应个性化定制生产中的多种复杂工况。整个过程采用超前离线式生成的方式，在生产线运行前即可得到全局生产任务的最优静态规划。系统包含静态优化算法库，内含多种经典静态优化算法，包括质量预测法、最优化算法、启发式方法、最大继承算法、遗传算法等及其衍生算法，用户也可根据自身需求和实际生产环境特征，对算法库进行更新，完成算法库更新后系统在决策过程中也会将用户个人算法考虑在内；
    静态目标包括最大完工时间最小、能源消耗最低以及生产加工成本最低等，静态目标的决策主要为对多静态目标的综合权重分析及优化，决策生成静态多目标综合权重参数；
    所述最优静态调度规划，是在全局个性化定制生产任务上生成的从宏观层面考虑的符合最大完工时间最小、能源消耗总和最低、加工成本最低的调度规划方案。最优静态调度规划的优化算法由系统根据当前环境和个性化定制生产任务自主决策选择。是本发明提出的智能生产线宏观策略建模重要组成部分。
  3. 根据权利要求1所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述动态事件解读模块，包括监听工业大数据模块提供的生产数据信息，监测可进行动态调度优化的动态事件；基于深度强化学习技术根据动态事件的内容进行分类，以当前工业生产信息作为智能体强化学习环境，令智能体观察当前环境下动态事件的特征， 设置相应的激励，强化其动态事件分类的动作。类别包括生产线时间偏差、机器故障、物料问题、合同事件等；根据工业大数据模块提供的当前工业生产环境数据以及事件分类结果，自动决策出当前事件下动态调度优化的多目标动态权重参数和对应约束；
    动态目标包括最大完工时间最小、能源消耗最低、生产加工成本最低、生产计划调整灵活性最高、调度稳定性最高(即相较于最优静态调度规划的变化最低)等，动态目标的决策主要包括为对多动态目标的综合权重分析及优化，最终生成多动态目标综合权重；
    多动态目标综合权重，其形式如下式表示：
    minZ＝α·f norm(T)+β·f norm(E)+γ·f norm(C)+(1-δ·f norm(P))+(1-ε·f norm(Q))
    其中α表示时间权重参数；β表示能耗权重参数；γ表示成本权重参数；δ表示灵活性权重参数；ε表示稳定性权重参数；T表示车间生产时间，E表示车间生产能耗，C表示车间生产成本，P表示调度方案灵活性，Q表示调度方案稳定性，f norm(x)表示归一化函数，将时间、能耗、成本、灵活性和稳定性等5个指标量纲统一，且函数值在0到1之间。各权重参数值也在0到1之间。
  4. 根据权利要求1所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述动态调度优化模块包含动态优化算法选择策略模型以及优化算法参数自适应学习方法。考虑到动态事件种类多样、规律难以确定的特点，本系统采用深度强化学习的方式生成动态优化算法选择策略模型，并以监督学习的方式让系统自适应匹配优化算法与参数优化方法；
    动态优化算法选择策略模型以多种优化算法和当前事件分类标签、工业生产环境作为智能体强化学习环境，令智能体观察不同优化算法之间的优化结果差距，设置相应的激励，强化其优化算法选择动作。分情况决策不同类别动态事件对应的动态调度优化算法。系统包含动态优化算法库，内含 多种经典动态优化算法，包括多智能体算法、模拟退火算法、混合粒子群算法、禁忌搜索算法以及遗传算法及其变种，用户可根据自身需求和实际生产环境特征，对算法库进行更新，完成算法库更新后，系统决策过程中也会将用户个人算法考虑在内；
    针对不同的优化算法，本发明通过深度学习的方式来提取优化算法的特征，具体以深度神经网络监督学习的方式，对优化算法自动归类，在参数优化算法库中选择合适的参数优化方法。用户也可以对参数优化算法库进行更新，系统可自动提取个人算法特征进行分类。
  5. 根据权利要求1所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述调度部署及规划模块，包括基于深度强化学习的自适应方案评估策略，针对接收到的优化方案计算其得分，分数高于规定分数阈值则进行下一步的调度部署。在部署过程中，参照前述步骤生成的静态调度方案或动态调度方案，结合当前工业生产环境信息，通过深度强化学习技术，以当前工业生产环境和调度方案为环境，采用深度学习的方式提取生产环境及调度方案的特征，采用强化学习的方式自主决策当前的部署优化目标权重参数和相关约束，并选择合适的优化算法，最终生成部署序列。系统包含部署优化算法库，内含多种经典的设备部署优化算法，包括自适应遗传算法、蚁群优化算法、多目标遗传算法、粒子群算法等算法大类，还包含其衍生算法，用户也可根据自身需求和实际生产环境特征，对算法库进行更新，完成算法库更新后，系统决策过程中也会将用户个人算法考虑在内；部署优化过程中考虑的优化目标主要包括运输开销、设备时延开销、生产能力、部署复杂度等方面，系统主要在多优化目标的权重比例做决策。
  6. 根据权利要求1所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述实时更新工业大数据模块是指实 际工业生产过程中产生和监测到的所有数据信息，包括动态事件信息、物料信息、生产设备信息、个性化定制生产任务信息、工厂当前运营信息等。
  7. 根据权利要求2所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述多模块协同动态决策链，是指本发明提出的统一计算框架，只需工业大数据模块实时接收外部数据，结合动态事件解读模块、动态调度优化模块、调度部署及评估模块，即可形成闭环决策通路，整个过程中若无人工发出高级决策指令，全部决策任务均由计算机自动完成。依靠闭环决策链，本发明提出的统一计算框架只需接收数据，即可完成智能生产线上的设备及工序的自动规划部署。
  8. 根据权利要求2所述一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统，其特征在于：所述宏观策略建模，是指从整个工厂生产环境的整体来可考虑智能生产线的运行及优化，而非将注意力仅放在当前工序或者当前事件上，在处理随机动态事件过程中，仍要注意整个生产环境及个性化定制生产任务。包括生产任务方面的宏观策略建模和生产要素方面的宏观策略建模；
    生产任务方面的宏观策略建模是指整个框架在任务初始化后首先根据用户提供的全局生产任务自动选择静态优化算法，并生成一个基于全局生产任务的最优静态规划，在后续的局部动态调度优化调整中，将会参考最优静态规划，尽量减小对全局最优静态规划的改变，以保证整体框架运行中能够兼顾总体生产任务；
    生产要素方面的宏观策略建模是指在计算框架运行过程中，会实时更新工业大数据模块的信息，其他模块可以随时取用当前信息，该信息包含整个工厂生产环境的全部信息，本发明提出的框架在运行过程中不仅仅局限于当前事件所涉及到部分设备或工序的信息，而会参考当前整个生产环境的信息，从而考虑对其他生产工序及设备的影响。
PCT/CN2020/128622 2020-11-13 2020-11-13 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统 WO2022099596A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US17/781,370 US20220413455A1 (en) 2020-11-13 2020-11-13 Adaptive-learning intelligent scheduling unified computing frame and system for industrial personalized customized production
PCT/CN2020/128622 WO2022099596A1 (zh) 2020-11-13 2020-11-13 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/128622 WO2022099596A1 (zh) 2020-11-13 2020-11-13 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统

Publications (1)

Publication Number Publication Date
WO2022099596A1 true WO2022099596A1 (zh) 2022-05-19

Family

ID=81602004

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/128622 WO2022099596A1 (zh) 2020-11-13 2020-11-13 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统

Country Status (2)

Country Link
US (1) US20220413455A1 (zh)
WO (1) WO2022099596A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114881366A (zh) * 2022-06-13 2022-08-09 科大智能物联技术股份有限公司 一种基于pas与opcs交互的多生产线协同调度方法及系统
CN115204851A (zh) * 2022-09-19 2022-10-18 北京中创碳投科技有限公司 一种智慧核算模型的效果评估和优化方法及系统
CN115426319A (zh) * 2022-08-30 2022-12-02 上海飞机制造有限公司 一种网络资源调度系统
CN116151599A (zh) * 2023-04-21 2023-05-23 湖南维胜科技有限公司 一种基于深度强化学习的排产数据处理方法
CN117422585A (zh) * 2023-12-15 2024-01-19 中山市三乐电子有限公司 电感生产制造控制系统的状态监控方法及系统
CN117495210A (zh) * 2024-01-02 2024-02-02 衡水烨通建设工程有限公司 公路混凝土施工质量管理系统
CN117522176A (zh) * 2024-01-08 2024-02-06 中国电子科技集团公司第十五研究所 基于多智能体系统的多项目资源调度评估系统

Families Citing this family (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116088451A (zh) * 2023-02-15 2023-05-09 陕西西微测控工程有限公司 一种采用模块集群方式实现板纸品种切换的方法
CN116681320B (zh) * 2023-04-07 2023-12-05 中宬建设管理有限公司 一种工程监控一体化的智能管理方法及系统
CN116594358B (zh) * 2023-04-20 2024-01-02 暨南大学 基于强化学习的多层工厂车间调度方法
CN116300772B (zh) * 2023-05-18 2023-09-15 智昌科技集团股份有限公司 基于机器人集群的全产业链制造方法及系统
CN116268521B (zh) * 2023-05-22 2023-08-15 首域科技(杭州)有限公司 一种加料机自适应智能控制系统
CN116339955B (zh) * 2023-05-25 2023-08-11 中国人民解放军国防科技大学 计算换通信框架的局部优化方法、装置和计算机设备
CN116843916B (zh) * 2023-07-25 2024-02-23 湖北仕上电子科技有限公司 基于人工智能的板材干燥设备控制系统的优化方法及系统
CN117041040B (zh) * 2023-08-16 2024-05-10 宁夏隆合科技有限公司 一种基于智能调度算法的指挥中心分布式布局系统
CN116866084B (zh) * 2023-08-30 2023-11-21 国网山东省电力公司信息通信公司 基于强化学习的入侵响应决策方法及系统
CN116862206B (zh) * 2023-09-04 2023-12-15 宁德思客琦智能装备有限公司 一种自适应优化生产计划的分布式制造执行的方法和系统
CN116979622B (zh) * 2023-09-22 2024-01-02 华夏天信智能物联股份有限公司 用于煤矿能耗优化的变频一体机控制方法及系统
CN117006593B (zh) * 2023-09-28 2023-12-08 博纳环境设备(太仓)有限公司 基于需求的工业空调出风控制方法及系统
CN117114088B (zh) * 2023-10-17 2024-01-19 安徽大学 一种基于统一ai框架的深度强化学习智能决策平台
CN117114506B (zh) * 2023-10-17 2024-02-20 长春易加科技有限公司 一种智慧工厂质量检测与分析方法
CN117395685A (zh) * 2023-10-30 2024-01-12 中国铁建电气化局集团有限公司 一种基于人工智能的高铁无线网络优化系统
CN117289671B (zh) * 2023-11-27 2024-02-02 博纯材料股份有限公司 高阶锗烷提纯生产控制系统的状态监控方法及系统
CN117311295B (zh) * 2023-11-28 2024-01-30 深圳百通玄武技术有限公司 基于无线网络设备的生产质量提升方法及系统
CN117395166B (zh) * 2023-12-11 2024-02-13 四海良田(天津)智能科技有限公司 基于物联网的智能农业管理平台
CN117472587B (zh) * 2023-12-26 2024-03-01 广东奥飞数据科技股份有限公司 一种ai智算中心的资源调度系统
CN117522246B (zh) * 2024-01-05 2024-04-30 深圳市兆航物流有限公司 基于ai算法及大数据分析的物流中心优化调度系统
CN117829554B (zh) * 2024-03-05 2024-05-14 山东商业职业技术学院 智能感知成品修复决策支持系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110224816A1 (en) * 2010-03-12 2011-09-15 Pereira Ana Maria Dias Madureira Multi-agent system for distributed manufacturing scheduling with genetic algorithms and tabu search
CN104951425A (zh) * 2015-07-20 2015-09-30 东北大学 一种基于深度学习的云服务性能自适应动作类型选择方法
CN110737529A (zh) * 2019-09-05 2020-01-31 北京理工大学 一种面向短时多变大数据作业集群调度自适应性配置方法
CN111199272A (zh) * 2019-12-30 2020-05-26 同济大学 一种面向智能车间的自适应调度方法
CN111882215A (zh) * 2020-07-27 2020-11-03 武汉理工大学 一种含有agv的个性化定制柔性作业车间调度方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110224816A1 (en) * 2010-03-12 2011-09-15 Pereira Ana Maria Dias Madureira Multi-agent system for distributed manufacturing scheduling with genetic algorithms and tabu search
CN104951425A (zh) * 2015-07-20 2015-09-30 东北大学 一种基于深度学习的云服务性能自适应动作类型选择方法
CN110737529A (zh) * 2019-09-05 2020-01-31 北京理工大学 一种面向短时多变大数据作业集群调度自适应性配置方法
CN111199272A (zh) * 2019-12-30 2020-05-26 同济大学 一种面向智能车间的自适应调度方法
CN111882215A (zh) * 2020-07-27 2020-11-03 武汉理工大学 一种含有agv的个性化定制柔性作业车间调度方法

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114881366A (zh) * 2022-06-13 2022-08-09 科大智能物联技术股份有限公司 一种基于pas与opcs交互的多生产线协同调度方法及系统
CN115426319A (zh) * 2022-08-30 2022-12-02 上海飞机制造有限公司 一种网络资源调度系统
CN115204851A (zh) * 2022-09-19 2022-10-18 北京中创碳投科技有限公司 一种智慧核算模型的效果评估和优化方法及系统
CN116151599A (zh) * 2023-04-21 2023-05-23 湖南维胜科技有限公司 一种基于深度强化学习的排产数据处理方法
CN117422585A (zh) * 2023-12-15 2024-01-19 中山市三乐电子有限公司 电感生产制造控制系统的状态监控方法及系统
CN117422585B (zh) * 2023-12-15 2024-05-17 中山市三乐电子有限公司 电感生产制造控制系统的状态监控方法及系统
CN117495210A (zh) * 2024-01-02 2024-02-02 衡水烨通建设工程有限公司 公路混凝土施工质量管理系统
CN117495210B (zh) * 2024-01-02 2024-03-12 衡水烨通建设工程有限公司 公路混凝土施工质量管理系统
CN117522176A (zh) * 2024-01-08 2024-02-06 中国电子科技集团公司第十五研究所 基于多智能体系统的多项目资源调度评估系统
CN117522176B (zh) * 2024-01-08 2024-04-12 中国电子科技集团公司第十五研究所 基于多智能体系统的多项目资源调度评估系统

Also Published As

Publication number Publication date
US20220413455A1 (en) 2022-12-29

Similar Documents

Publication Publication Date Title
WO2022099596A1 (zh) 一种面向工业个性化定制生产的自适应学习智能调度统一计算框架及系统
CN102402716B (zh) 智能生产决策支持系统
CN111929748B (zh) 一种气象要素预报方法和系统
CN109343995A (zh) 基于多源异构数据融合、机器学习及客服机器人的智能运维分析系统
Han et al. Information-utilization-method-assisted multimodal multiobjective optimization and application to credit card fraud detection
Ouyang et al. A multi-stage intelligent system for unit commitment
CN109934422A (zh) 一种基于时间序列数据分析的神经网络风速预测方法
CN107121926A (zh) 一种基于深度学习的工业机器人可靠性建模方法
CN112990485A (zh) 基于强化学习的知识策略选择方法与装置
Malaga et al. Benchmarking smart manufacturing drivers using Grey TOPSIS and COPRAS-G approaches
CN110852387A (zh) 一种能源互联网超实时状态研判算法
CN117575108B (zh) 一种化工厂能源数据分析系统
CN111308975A (zh) 基于数字孪生技术的智能制造系统孪生智能体构建系统及方法
CN106845696A (zh) 一种智能优化水资源配置方法
Wei et al. A combination forecasting method of grey neural network based on genetic algorithm
CN114707613B (zh) 基于分层的深度策略梯度网络的电网调控方法
Wang et al. Human-machine cooperation based adaptive scheduling for a smart shop floor
CN109684471B (zh) 一种ai智能文本处理系统在新零售领域的应用方法
Ma et al. The research and application of a dynamic dispatching strategy selection approach based on BPSO-SVM for semiconductor production line
Chen et al. Performance evaluation of online recruitment enterprises based on intuitionistic fuzzy set and TOPSIS
Akagic et al. Deep reinforcement learning in smart grid: Progress and prospects
Fan et al. Framework and key technologies of human-machine hybrid-augmented intelligence system for large-scale power grid dispatching and control
CN111766839B (zh) 一种智能车间调度知识自适应更新的计算机实现系统
Li Research on the application of artificial intelligence in diagnosing power system faults
Tian et al. Macroeconomic early warning method based on support vector machine under multi-sensor data fusion technology

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20961153

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20961153

Country of ref document: EP

Kind code of ref document: A1