WO2024114121A1 - 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 - Google Patents
一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 Download PDFInfo
- Publication number
- WO2024114121A1 WO2024114121A1 PCT/CN2023/124030 CN2023124030W WO2024114121A1 WO 2024114121 A1 WO2024114121 A1 WO 2024114121A1 CN 2023124030 W CN2023124030 W CN 2023124030W WO 2024114121 A1 WO2024114121 A1 WO 2024114121A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- model
- task
- meta
- neural network
- target
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 20
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 14
- 238000004519 manufacturing process Methods 0.000 claims abstract description 70
- 238000012549 training Methods 0.000 claims abstract description 41
- 238000013528 artificial neural network Methods 0.000 claims abstract description 27
- 238000012360 testing method Methods 0.000 claims abstract description 19
- 238000013508 migration Methods 0.000 claims abstract description 10
- 230000005012 migration Effects 0.000 claims abstract description 10
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 238000003062 neural network model Methods 0.000 claims description 12
- 230000008859 change Effects 0.000 claims description 8
- 230000003068 static effect Effects 0.000 claims description 8
- 230000002787 reinforcement Effects 0.000 claims description 6
- 238000002474 experimental method Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 230000007613 environmental effect Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims 1
- 101001121408 Homo sapiens L-amino-acid oxidase Proteins 0.000 description 1
- 102100026388 L-amino-acid oxidase Human genes 0.000 description 1
- 101100012902 Saccharomyces cerevisiae (strain ATCC 204508 / S288c) FIG2 gene Proteins 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000003780 insertion Methods 0.000 description 1
- 230000037431 insertion Effects 0.000 description 1
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/092—Reinforcement learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/096—Transfer learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Definitions
- the present invention relates to the field of computer data science and technology, and in particular to a method for constructing an artificial intelligence cross-platform model intelligent computing engine based on knowledge self-evolution.
- the technical problem to be solved by the present invention is to provide a method for constructing an artificial intelligence cross-platform model intelligent computing engine based on knowledge self-evolution, which can reduce the convergence time of model parameters and has great application value for the training of dynamic discrete manufacturing models that are disturbed over time in actual production.
- the present invention provides a method for constructing an artificial intelligence cross-platform model intelligent computing engine based on knowledge self-evolution, comprising the following steps:
- Step 1 Confirm the source time and the target time; divide the discrete manufacturing system data sets of these two different time periods according to certain division rules based on manual experience;
- Step 2 Initialize the dynamic discrete manufacturing system model
- Step 3 Preprocess the data and build a task pool; part of the data is used to train the source model, and the other part is used to train the target neural network;
- Step 4 Build a meta-learning framework, which is divided into training the meta-learning model and quickly adjusting the target neural network to achieve rapid migration between multiple tasks;
- Step 5 Change the target time and use the meta-learning framework to quickly migrate the trained neural network to the new task
- Step 6 iterate step 5 until the dynamic discrete manufacturing system model converges, and save the model parameters after convergence;
- Step 7 Use the dynamic discrete manufacturing system model for new environmental tasks and test its performance.
- step 1 the source time and the target time are confirmed; based on manual experience, according to certain division rules
- the discrete manufacturing system data sets of these two different time periods are divided into:
- Step 11 Select the source time s and the target time t, and input the discrete manufacturing data of these two time periods;
- Step 12 According to the rules of equal division, production cost or product quantity, based on manual experience, the data sets of the two time periods are divided into the optimal N s and N t static discrete manufacturing data sets respectively.
- step 2 initializing the dynamic discrete manufacturing system model is specifically as follows:
- Step 21 Based on the deep reinforcement learning algorithm, select a suitable deep reinforcement learning neural network Q to initialize the parameters ⁇ of the dynamic discrete manufacturing system model;
- Step 22 Define two hyperparameters ⁇ and ⁇ of the meta-learning algorithm. The specific values require multiple experiments.
- step 3 preprocessing the data to construct a task pool; a portion of the data is used to train the source model, and the other portion is used to train the target neural network, which specifically includes the following steps:
- Step 31 Based on the data divided in step 12, the N s categories at the source time are called meta-train classes, which are used to train the meta-learning model Q s , which represents the static model applicable to the current time and working conditions; the N t categories at the target time are called meta-test classes, which are used to train the target model Q t , which represents the discrete manufacturing system model applicable to the new time and new working conditions after dynamic parameter adjustment;
- Step 32 set the task extraction method to M way-K shot, and construct a data set for training the meta-learning model; randomly select Ms categories from the meta-train classes, and randomly select Ls samples from each category ( Ls > Ks ) to form a Task Ts , where Ks samples are randomly selected from each category as the training set of the current Task, called the support set, and the remaining Ms *( Ls - Ks ) samples are used as the test set of the current Task, called the query set.
- Tasks are randomly extracted from the meta-train classes repeatedly to form a Task pool consisting of several T, whose distribution is defined as p( Ts );
- Step 33 Construct a data set for training the target model. Randomly select M t categories from the meta-test classes, randomly select L t samples from each category (L t >K t ) to form a Task T t , where K t samples are randomly selected from each category as the training set of the current Task, called the support set, and the remaining M t *(L t -K t ) samples are used as the test set of the current Task, called the query set. Tasks are repeatedly randomly selected from the meta-test classes to form a Task pool consisting of several T, whose distribution is defined as p(T t ).
- a meta-learning framework is constructed, which is divided into training a meta-learning model and rapidly adjusting a target neural network, and realizing rapid migration between multiple tasks specifically includes the following steps:
- Step 41 training the meta-learning model Q s , specifically including the following steps:
- ⁇ ′ si is the parameter of the model Q s after updating based on Ti , is the loss gradient function for calculating ⁇ s based on Ti ;
- step (d) Repeat step (c) for each task in the batch for n a times, thus completing the first gradient update and obtaining the updated parameter ⁇ s ;
- Second gradient update Use the query set in each Task Ti in the batch to calculate the loss gradient of ⁇ s , and then calculate the total loss of the batch. Use the total loss to update the gradient.
- the update formula is as follows:
- step (f) Return to step (b) and resample the next batch;
- Step 42 After the training is completed, the initialization parameters ⁇ s of the neural network Q s are obtained, and the model parameters are dynamically adjusted according to the data set at the target time to adapt it to the new internal and external production environment.
- the training target model Q t specifically includes the following steps:
- ⁇ ′ ti is the parameter of the model Q t after updating based on Ti ;
- step (j) Based on each Task randomly selected in step (h), apply the update algorithm of step (i) to the parameter ⁇ b initialized in step (g) to obtain n t updated parameters ⁇ ′ ti , and average them to obtain the final model parameters.
- the formula is as follows:
- step 5 changing the target time and using the meta-learning framework to quickly migrate the trained neural network to the new task specifically includes the following steps:
- Step 51 Use the trained target neural network model Qt as the source model and the next time t+1 as the new target time.
- the next task is to quickly migrate the neural network model from time t to time t+1.
- Step 52 perform data preprocessing according to step 3 and build a task pool
- Step 53 According to step 4, obtain the parameter ⁇ t+1 of the new target neural network Q t +1 .
- step 6 iterating step 5 until the dynamic discrete manufacturing system model converges, and saving the model parameters after convergence specifically includes the following steps:
- Step 61 iterate step 5, and continuously obtain the neural network model Q t , Q t+1 , Q t+2 , etc. at the next moment until the model parameters converge, indicating that the system model can be well adapted to the production environment of different working conditions at different time periods, and can stably output the optimal decision no matter how the internal and external conditions change;
- Step 62 After the model converges, save the model parameters to obtain the final dynamic discrete manufacturing system model.
- step 7 the dynamic discrete manufacturing system model is used for the new environment task, and its performance is tested specifically as follows: if the system model can output the scheduling strategy well in the new environment and has higher efficiency than the original system, the result meets expectations and the training is completed; if the result does not meet expectations, return to step 1 and retrain.
- the present invention has small computational complexity and good generalization performance.
- a meta-learning framework is used to realize learning of common features of multiple categories, thereby improving the generalization performance of a dynamic discrete manufacturing model.
- the model parameters converge quickly and are highly transferable.
- By loading the optimized parameters trained by meta-learning as initialization parameters in a new task only a few training steps are required to complete the model parameters of the discrete manufacturing model at a new moment, and task migration is completed quickly. The more similar the new task is to the original task, the less time is required.
- this algorithm can realize rapid fine-tuning of the neural network and reduce the convergence time of the model parameters. It has great application value for the training of dynamic discrete manufacturing models that are disturbed over time in actual production.
- FIG1 is a schematic flow chart of the method of the present invention.
- FIG2 is a schematic diagram of the meta-learning framework structure of the present invention.
- a method for constructing an artificial intelligence cross-platform model intelligent computing engine based on knowledge self-evolution includes the following steps:
- Step 1 confirm the source time and target time; based on manual experience, divide the discrete manufacturing system data sets of these two different time periods according to certain division rules; select two data sets of different time periods from the dynamic discrete manufacturing production data.
- the data sets of these two time periods represent different workshop operating conditions and internal and external conditions.
- the dynamic parameters of the workshop system change accordingly. Therefore, the workshop system needs to achieve self-evolution and adjust the dynamic parameters by itself to adapt to the complex and changeable workshop production conditions.
- These two data sets are called the data sets of the source time and the target time, respectively.
- the data set of the source time represents the data used to train the static discrete manufacturing production model
- the data set of the target time represents the data after the workshop operating conditions and internal and external conditions change, indicating that the system model is applicable to the new time after the dynamic parameters are adjusted, so as to achieve the adaptive adjustment of the dynamic model.
- Step 11 Select the source time s and the target time t, and input the discrete manufacturing data of these two time periods;
- Step 12 According to the rules of equal division, production cost or product quantity, based on manual experience, the data sets of the two time periods are divided into the optimal N s and N t static discrete manufacturing data sets respectively.
- Step 2 Initialize the dynamic discrete manufacturing system model; specifically:
- Step 21 Based on the deep reinforcement learning algorithm, select a suitable deep reinforcement learning neural network Q to initialize the parameters ⁇ of the dynamic discrete manufacturing system model;
- Step 22 Define two hyperparameters ⁇ and ⁇ of the meta-learning algorithm. The specific values require multiple experiments.
- Step 3 Preprocess the data and build a task pool; part of the data is used to train the source model, and the other part is used to train the target neural network; the specific steps include:
- Step 31 Based on the data divided in step 12, the N s categories at the source time are called meta-train classes, which are used to train the meta-learning model Q s , which represents the static model applicable to the current time and working conditions; the N t categories at the target time are called meta-test classes, which are used to train the target model Q t , which represents the discrete manufacturing system model applicable to the new time and new working conditions after dynamic parameter adjustment;
- Step 32 Set the Task extraction method to M way-K shot, and construct a dataset for training the meta-learning model. Randomly select Ms categories from meta-train classes, randomly select Ls samples from each category ( Ls > Ks ) to form a Task Ts , where Ks samples are randomly selected from each category as the training set of the current Task, called the support set, and the remaining Ms *( Ls -Ks ) samples are used as the test set of the current Task, called the query set. Tasks are randomly selected from meta-train classes repeatedly to form a Task pool consisting of several Ts, whose distribution is defined as p( Ts );
- Step 33 Construct a data set for training the target model. Randomly select M t categories from the meta-test classes, randomly select L t samples from each category (L t >K t ) to form a Task T t , where K t samples are randomly selected from each category as the training set of the current Task, called the support set, and the remaining M t *(L t -K t ) samples are used as the test set of the current Task, called the query set. Tasks are repeatedly randomly selected from the meta-test classes to form a Task pool consisting of several T, whose distribution is defined as p(T t ).
- Step 4 Build a meta-learning framework, which is divided into training the meta-learning model and quickly adjusting the target neural network to achieve rapid migration between multiple tasks; specifically, the following steps are included:
- Step 41 training the meta-learning model Q s , specifically including the following steps:
- ⁇ ′ si is the parameter of the model Q s after updating based on Ti , is the loss gradient function for calculating ⁇ s based on Ti ;
- step (d) Repeat step (c) for each task in the batch for n a times, thus completing the first gradient update and obtaining the updated parameter ⁇ s ;
- Second gradient update Use the query set in each Task Ti in the batch to calculate the loss gradient of ⁇ s , and then calculate the total loss of the batch. Use the total loss to update the gradient.
- the update formula is as follows:
- step (f) Return to step (b) and resample the next batch;
- Step 42 After the training is completed, the initialization parameters ⁇ s of the neural network Q s are obtained, and the model parameters are dynamically adjusted according to the data set at the target time to adapt it to the new internal and external production environment.
- the training target model Q t specifically includes the following steps:
- ⁇ ′ ti is the parameter of the model Q t after updating based on Ti ;
- step (j) Based on each Task randomly selected in step (h), apply the update algorithm of step (i) to the parameter ⁇ b initialized in step (g) to obtain n t updated parameters ⁇ ′ ti , and average them to obtain the final model parameters, which are as follows:
- Step 5 Change the target time and use the meta-learning framework to quickly migrate the trained neural network to the new task; specifically, the following steps are included:
- Step 51 Take the trained target neural network model Qt as the source model and the next time t+1 as the new target time, and divide the data sets of these two time periods according to step 12.
- the next task is to quickly migrate the neural network model from time t to time t+1;
- Step 52 perform data preprocessing according to step 3 and build a task pool
- Step 53 According to step 4, obtain the parameter ⁇ t+1 of the new target neural network Q t +1 .
- Step 6 iterate step 5 until the dynamic discrete manufacturing system model converges, and save the model parameters after convergence; specifically, it includes the following steps:
- Step 61 iterate step 5, and continuously obtain the neural network model Q t , Q t+1 , Q t+2 , etc. at the next moment until
- the convergence of model parameters indicates that the system model can adapt well to the production environment of different working conditions in different time periods, and can stably output the optimal decision regardless of changes in internal and external conditions;
- Step 62 After the model converges, save the model parameters to obtain the final dynamic discrete manufacturing system model.
- Step 7 Use the dynamic discrete manufacturing system model for new environment tasks to test its performance. Specifically, if the system model can output scheduling strategies well in the new environment and has higher efficiency than the original system, the result is as expected and the training is completed. If the result is not as expected, return to step 1 and retrain.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Feedback Control In General (AREA)
Abstract
本发明公开了一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,包括如下步骤:确认源时刻和目标时刻,基于人工经验,按照划分规则将某一时刻的离散制造系统数据集划分为多个数据集;初始化动态离散制造系统模型;对数据进行预处理,构建任务池;构建元学习框架,分为训练元学习模型和快速调整目标神经网络,实现多任务之间的快速迁移;更换目标时刻,利用元学习框架将训练好的神经网络快速迁移到新的任务;迭代上一步骤直到动态离散制造系统模型收敛,收敛后保存模型参数;将动态离散制造系统模型用于新环境任务,测试其性能。本发明能够减少模型参数的收敛时间,对于实际生产中随时间扰动的动态离散制造模型的训练具有重大的意义。
Description
本发明涉及计算机数据科学技术领域,尤其是一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法。
在经济全球化的背景下,随着科学技术的进步和人民需求的提高,离散制造业的发生了翻天覆地的变化,插单销单、个性化定制、设备故障、物料统筹等突发事件使得生产过程复杂多变。而目前的离散制造系统在复杂动态场景中难以运用,往往需要人工干预。这些在特定静态场景下训练的系统模型难以应对实际生产过程中制造信息的频繁变动。由于车间实际运行工况和内外部条件的改变,导致车间系统动态参数或系统模型结构经常发生变化。所以迫切需要一种知识自演化的技术,能使离散制造模型随时间自行演化,适应目前的动态场景,实现真正的智能化动态离散制造系统。
发明内容
本发明所要解决的技术问题在于,提供一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,能够减少模型参数的收敛时间,对于实际生产中随时间扰动的动态离散制造模型的训练具有重大的应用价值。
为解决上述技术问题,本发明提供一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,包括如下步骤:
步骤1、确认源时刻和目标时刻;基于人工经验,按照一定的划分规则对这两个不同时间段的离散制造系统数据集进行划分;
步骤2、初始化动态离散制造系统模型;
步骤3、对数据进行预处理,构建任务池;一部分数据用于训练源模型,另一部分用于训练目标神经网络;
步骤4、构建一个元学习框架,分为训练元学习模型和快速调整目标神经网络,实现多任务之间的快速迁移;
步骤5、更换目标时刻,利用元学习框架将训练好的神经网络快速迁移到新的任务;
步骤6、迭代步骤5直到动态离散制造系统模型收敛,收敛后保存模型参数;
步骤7、将动态离散制造系统模型用于新环境任务,测试其性能。
优选的,步骤1中,确认源时刻和目标时刻;基于人工经验,按照一定的划分规则
对这两个不同时间段的离散制造系统数据集进行划分具体为:
步骤11、选择源时刻s和目标时刻t,输入这两个时间段的离散制造数据;
步骤12、按照等分、生产成本或产品数量等规则,基于人工经验,分别将这两个时间段的数据集划分为最优的Ns和Nt个静态离散制造数据集。
优选的,步骤2中,初始化动态离散制造系统模型具体为:
步骤21、基于深度强化学习算法,选择合适的深度强化学习神经网络Q,对动态离散制造系统模型的参数θ进行初始化;
步骤22、定义元学习算法的两个超参数α和β,具体取值需多次实验。
优选的,步骤3中,对数据进行预处理,构建任务池;一部分数据用于训练源模型,另一部分用于训练目标神经网络具体包括如下步骤:
步骤31、依据步骤12划分的数据,源时刻具有的Ns个类别称为meta-train classes,用于训练元学习模型Qs,表示适用于当前时刻和工况的静态模型;目标时刻具有的Nt个类别称为meta-test classes,用于训练目标模型Qt,其表示动态参数调整后适用于新时刻新工况的离散制造系统模型;
步骤32、Task抽取方法设置为M way-K shot,构建用于训练元学习模型的数据集;从meta-train cla es中随机选取Ms个类别,每个类别随机选取Ls个样本(Ls>Ks),组成一个Task Ts,其中每个类别中随机选取Ks个样本作为当前Task的训练集,称为支持集support set,剩余Ms*(Ls-Ks)个样本作为当前Task的测试集,称为查询集query set,从meta-train classes中如此反复随机抽取Task,构成由若干个T构成的Task池,其分布定义为p(Ts);
步骤33、构建用于训练目标模型的数据集,从meta-test classes中随机选取Mt个类别,每个类别随机选取Lt个样本(Lt>Kt),组成一个Task Tt,其中每个类别中随机选取Kt个样本作为当前Task的训练集,称为support set,剩余Mt*(Lt-Kt)个样本作为当前Task的测试集,称为query set,从meta-test classes中反复随机抽取Task,构成由若干个T构成的Task池,其分布定义为p(Tt)。
优选的,步骤4中,构建一个元学习框架,分为训练元学习模型和快速调整目标神经网络,实现多任务之间的快速迁移具体包括如下步骤:
步骤41、训练元学习模型Qs,具体包括如下步骤:
(a)随机初始化Qs的模型参数θs;
(b)随机从Task池中采样ns个Task T,形成一个batch,其中每一个Task Ti(i=1,2,3…,ns)满足分布Ti~p(Ts);
(c)利用batch中的某一个Task Ti中的support set,计算模型参数θs的梯度模型参数θs的更新公式如下:
上式中,θ′si为基于Ti更新后的模型Qs的参数,为基于Ti计算出θs的损失梯度函数;
(d)基于batch中的每一个Task重复步骤(c),执行次数为na,如此完成第一次梯度的更新,获得更新后的参数θs;
(e)第二次梯度的更新:利用batch中的每一个Task Ti中的query set,计算θs的损失梯度,进而计算batch的损失总和,利用该损失总和对梯度进行更新,其更新公式如下:
如此便完成第二次梯度的更新,结束模型在该batch上的训练;
(f)返回步骤(b),重新采样下一个batch;
步骤42、训练结束后获得神经网络Qs的初始化参数θs,根据目标时刻的数据集动态调整模型参数使其适应新的内外部生产环境,训练目标模型Qt具体包括如下步骤:
(g)初始化Qt的模型参数,将Qs的模型参数赋值Qt,即θt=θs;
(h)随机从Task池中采样nt个Task T,其中每一个Task Ti(i=1,2,3…,n)满足分布Ti~p(Tt);
(i)利用随机抽取的某一个Task Ti中的support set,更新模型参数θt的梯度模型参数θt的更新公式如下:
上式中,θ′ti为基于Ti更新后的模型Qt的参数;
(j)基于步(h)中随机抽取的每一个Task,分别对步骤(g)中初始化的参数θb运用步骤(i)的更新算法,获得nt个更新后的参数θ′ti,对其取平均获得最终的模型参数,
其公式如下:
以上便是目标模型Qt的训练过程,最终获得适用于目标时刻t的神经网络模型Qt,实现参数的动态调整和系统模型的快速迁移。
优选的,步骤5中,更换目标时刻,利用元学习框架将训练好的神经网络快速迁移到新的任务具体包括如下步骤:
步骤51、将训练好的目标神经网络模型Qt作为源模型,下一时刻t+1作为新的目标时刻,接下来的任务是将神经网络模型从时刻t快速迁移到时刻t+1;
步骤52、根据步骤3进行数据预处理,构建任务池;
步骤53、根据步骤4,获得新的目标神经网络Qt+1的参数θt+1。
优选的,步骤6中,迭代步骤5直到动态离散制造系统模型收敛,收敛后保存模型参数具体包括如下步骤:
步骤61、迭代步骤5,不断获取下一时刻的神经网络模型Qt、Qt+1、Qt+2…,直到模型参数收敛,表示该系统模型已经能很好的适应于不同时间段不同工况的生产环境,无论内外部条件如何变化,都能稳定的输出最优决策;
步骤62、模型收敛后,保存模型参数,获得最终的动态离散制造系统模型。
优选的,步骤7中,将动态离散制造系统模型用于新环境任务,测试其性能具体为:如果系统模型在新环境下能很好的输出调度策略,比原系统具有更高的效率,则结果符合预期,训练完成;如果结果不符合预期,则重新回到步骤1,重新训练。
本发明的有益效果为:本发明计算量小,泛化性能好,通过少量不同类型的样本组成的训练集,利用元学习框架实现对多个类别共同特征的学习,提高动态离散制造模型的泛化性能;模型参数收敛速度快,可迁移性强,通过在新任务中加载元学习训练的优化参数作为初始化参数,仅仅需训练几步即可完成离散制造模型在新时刻的模型参数,快速完成任务迁移,与原任务越相似的新任务所需的时间越少,相比随机初始化参数或加载已有的网络模型参数,通过这种算法能实现对神经网络的快速微调,减少模型参数的收敛时间,对于实际生产中随时间扰动的动态离散制造模型的训练具有重大的应用价值。
图1为本发明的方法流程示意图。
图2为本发明的元学习框架结构示意图。
如图1和2所示,一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,包括如下步骤:
步骤1、确认源时刻和目标时刻;基于人工经验,按照一定的划分规则对这两个不同时间段的离散制造系统数据集进行划分;从动态离散制造生产数据中选择两个不同时间段的数据集,这两个时间段的数据集代表了不同的车间运行工况和内外部条件,车间系统的动态参数由此发生变化,因此需要车间系统实现自演化,自行调整动态参数,适应复杂多变的车间生产工况。这两个数据集分别称为源时刻和目标时刻的数据集,源时刻的数据集代表用于训练静态离散制造生产模型的数据,目标时刻的数据集代表车间运行工况和内外部条件发生改变后的数据,表示动态参数调整后适用于新时刻的系统模型,达成动态模型的自适应调整。接下来基于人工经验,按照一定的划分规则(完全等分数据集、按照生产成本分割数据集成若干个对应成本相同的数据集或者按照生产的产品数量划分数据集)将源时刻的离散制造系统数据集划分为多个数据集。具体为:
步骤11、选择源时刻s和目标时刻t,输入这两个时间段的离散制造数据;
步骤12、按照等分、生产成本或产品数量等规则,基于人工经验,分别将这两个时间段的数据集划分为最优的Ns和Nt个静态离散制造数据集。
步骤2、初始化动态离散制造系统模型;具体为:
步骤21、基于深度强化学习算法,选择合适的深度强化学习神经网络Q,对动态离散制造系统模型的参数θ进行初始化;
步骤22、定义元学习算法的两个超参数α和β,具体取值需多次实验。
步骤3、对数据进行预处理,构建任务池;一部分数据用于训练源模型,另一部分用于训练目标神经网络;具体包括如下步骤:
步骤31、依据步骤12划分的数据,源时刻具有的Ns个类别称为meta-train classes,用于训练元学习模型Qs,表示适用于当前时刻和工况的静态模型;目标时刻具有的Nt个类别称为meta-test classes,用于训练目标模型Qt,其表示动态参数调整后适用于新时刻新工况的离散制造系统模型;
步骤32、Task抽取方法设置为M way-K shot,构建用于训练元学习模型的数据集;
从meta-train classes中随机选取Ms个类别,每个类别随机选取Ls个样本(Ls>Ks),组成一个Task Ts,其中每个类别中随机选取Ks个样本作为当前Task的训练集,称为支持集support set,剩余Ms*(Ls-Ks)个样本作为当前Task的测试集,称为查询集query set,从meta-train classes中如此反复随机抽取Task,构成由若干个T构成的Task池,其分布定义为p(Ts);
步骤33、构建用于训练目标模型的数据集,从meta-test classes中随机选取Mt个类别,每个类别随机选取Lt个样本(Lt>Kt),组成一个Task Tt,其中每个类别中随机选取Kt个样本作为当前Task的训练集,称为support set,剩余Mt*(Lt-Kt)个样本作为当前Task的测试集,称为query set,从meta-test classes中反复随机抽取Task,构成由若干个T构成的Task池,其分布定义为p(Tt)。
步骤4、构建一个元学习框架,分为训练元学习模型和快速调整目标神经网络,实现多任务之间的快速迁移;具体包括如下步骤:
步骤41、训练元学习模型Qs,具体包括如下步骤:
(a)随机初始化Qs的模型参数θs;
(b)随机从Task池中采样ns个Task T,形成一个batch,其中每一个Task Ti(i=1,2,3…,ns)满足分布Ti~p(Ts);
(c)利用batch中的某一个Task Ti中的support set,计算模型参数θs的梯度模型参数θs的更新公式如下:
上式中,θ′si为基于Ti更新后的模型Qs的参数,为基于Ti计算出θs的损失梯度函数;
(d)基于batch中的每一个Task重复步骤(c),执行次数为na,如此完成第一次梯度的更新,获得更新后的参数θs;
(e)第二次梯度的更新:利用batch中的每一个Task Ti中的query set,计算θs的损失梯度,进而计算batch的损失总和,利用该损失总和对梯度进行更新,其更新公式如下:
如此便完成第二次梯度的更新,结束模型在该batch上的训练;
(f)返回步骤(b),重新采样下一个batch;
步骤42、训练结束后获得神经网络Qs的初始化参数θs,根据目标时刻的数据集动态调整模型参数使其适应新的内外部生产环境,训练目标模型Qt具体包括如下步骤:
(g)初始化Qt的模型参数,将Qs的模型参数赋值Qt,即θt=θs;
(h)随机从Task池中采样nt个Task T,其中每一个Task Ti(i=1,2,3…,n)满足分布Ti~p(Tt);
(i)利用随机抽取的某一个Task Ti中的support set,更新模型参数θt的梯度模型参数θt的更新公式如下:
上式中,θ′ti为基于Ti更新后的模型Qt的参数;
(j)基于步(h)中随机抽取的每一个Task,分别对步骤(g)中初始化的参数θb运用步骤(i)的更新算法,获得nt个更新后的参数θ′ti,对其取平均获得最终的模型参数,其公式如下:
以上便是目标模型Qt的训练过程,最终获得适用于目标时刻t的神经网络模型Qt,实现参数的动态调整和系统模型的快速迁移。
步骤5、更换目标时刻,利用元学习框架将训练好的神经网络快速迁移到新的任务;具体包括如下步骤:
步骤51、将训练好的目标神经网络模型Qt作为源模型,下一时刻t+1作为新的目标时刻,根据步骤12划分这两个时间段的数据集。接下来的任务是将神经网络模型从时刻t快速迁移到时刻t+1;
步骤52、根据步骤3进行数据预处理,构建任务池;
步骤53、根据步骤4,获得新的目标神经网络Qt+1的参数θt+1。
步骤6、迭代步骤5直到动态离散制造系统模型收敛,收敛后保存模型参数;具体包括如下步骤:
步骤61、迭代步骤5,不断获取下一时刻的神经网络模型Qt、Qt+1、Qt+2…,直到
模型参数收敛,表示该系统模型已经能很好的适应于不同时间段不同工况的生产环境,无论内外部条件如何变化,都能稳定的输出最优决策;
步骤62、模型收敛后,保存模型参数,获得最终的动态离散制造系统模型。
步骤7、将动态离散制造系统模型用于新环境任务,测试其性能;具体为:如果系统模型在新环境下能很好的输出调度策略,比原系统具有更高的效率,则结果符合预期,训练完成;如果结果不符合预期,则重新回到步骤1,重新训练。
Claims (8)
- 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,包括如下步骤:步骤1、确认源时刻和目标时刻;基于人工经验,按照一定的划分规则对这两个不同时间段的离散制造系统数据集进行划分;步骤2、初始化动态离散制造系统模型;步骤3、对数据进行预处理,构建任务池;一部分数据用于训练源模型,另一部分用于训练目标神经网络;步骤4、构建一个元学习框架,分为训练元学习模型和快速调整目标神经网络,实现多任务之间的快速迁移;步骤5、更换目标时刻,利用元学习框架将训练好的神经网络快速迁移到新的任务;步骤6、迭代步骤5直到动态离散制造系统模型收敛,收敛后保存模型参数;步骤7、将动态离散制造系统模型用于新环境任务,测试其性能。
- 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,步骤1中,基于人工经验,按照一定的划分规则将某一时刻的离散制造系统数据集划分为多个数据集具体为:步骤11、选择源时刻s和目标时刻t,输入这两个时间段的离散制造数据;步骤12、按照等分、生产成本或产品数量规则,基于人工经验,分别将这两个时间段的数据集划分为最优的Ns和Nt个静态离散制造数据集。
- 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,步骤2中,初始化动态离散制造系统模型具体为:步骤21、基于深度强化学习算法,选择合适的深度强化学习神经网络Q,对动态离散制造系统模型的参数θ进行初始化;步骤22、定义元学习算法的两个超参数α和β,具体取值需多次实验。
- 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,步骤3中,对数据进行预处理,构建任务池;一部分数据用于训练源模型,另一部分用于训练目标神经网络具体包括如下步骤:步骤31、依据步骤12划分的数据,源时刻具有的Ns个类别称为meta-train classes,用于训练元学习模型Qs,表示适用于当前时刻和工况的静态模型;目标时刻具有的Nt个类别称为meta-test classes,用于训练目标模型Qt,其表示动态参数调整后适用于新 时刻新工况的离散制造系统模型;步骤32、Task抽取方法设置为M way-K shot,构建用于训练元学习模型的数据集;从meta-train classes中随机选取Ms个类别,每个类别随机选取Ls个样本(Ls>Ks),组成一个Task Ts,其中每个类别中随机选取Ks个样本作为当前Task的训练集,称为支持集support set,剩余Ms*(Ls-Ks)个样本作为当前Task的测试集,称为查询集query set,从meta-train classes中如此反复随机抽取Task,构成由若干个T构成的Task池,其分布定义为p(Ts);步骤33、构建用于训练目标模型的数据集,从meta-test classes中随机选取Mt个类别,每个类别随机选取Lt个样本(Lt>Kt),组成一个Task Tt,其中每个类别中随机选取Kt个样本作为当前Task的训练集,称为support set,剩余Mt*(Lt-Kt)个样本作为当前Task的测试集,称为query set,从meta-test classes中反复随机抽取Task,构成由若干个T构成的Task池,其分布定义为p(Tt)。
- 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,步骤4中,构建一个元学习框架,分为训练元学习模型和快速调整目标神经网络,实现多任务之间的快速迁移具体包括如下步骤:步骤41、训练元学习模型Qs,具体包括如下步骤:(a)随机初始化Qs的模型参数θs;(b)随机从Task池中采样ns个Task T,形成一个batch,其中每一个Task Ti(i=1,2,3…,ns)满足分布Ti~p(Ts);(c)利用batch中的某一个Task Ti中的support set,计算模型参数θs的梯度模型参数θs的更新公式如下:
上式中,θ′si为基于Ti更新后的模型Qs的参数,为基于Ti计算出θs的损失梯度函数;(d)基于batch中的每一个Task重复步骤(c),执行次数为na,如此完成第一次梯度的更新,获得更新后的参数θs;(e)第二次梯度的更新:利用batch中的每一个Task Ti中的query set,计算θs的损失梯度,进而计算batch的损失总和,利用该损失总和对梯度进行更新,其更新公式 如下:
如此便完成第二次梯度的更新,结束模型在该batch上的训练;(f)返回步骤(b),重新采样下一个batch;步骤42、训练结束后获得神经网络Qs的初始化参数θs,根据目标时刻的数据集动态调整模型参数使其适应新的内外部生产环境,训练目标模型Qt具体包括如下步骤:(g)初始化Qt的模型参数,将Qs的模型参数赋值Qt,即θt=θs;(h)随机从Task池中采样nt个Task T,其中每一个Task Ti(i=1,2,3…,n)满足分布Ti~p(Tt);(i)利用随机抽取的某一个Task Ti中的support set,更新模型参数θt的梯度模型参数θt的更新公式如下:
上式中,θ′ti为基于Ti更新后的模型Qt的参数;(j)基于步(h)中随机抽取的每一个Task,分别对步骤(g)中初始化的参数θb运用步骤(i)的更新算法,获得nt个更新后的参数θ′ti,对其取平均获得最终的模型参数,其公式如下:
以上便是目标模型Qt的训练过程,最终获得适用于目标时刻t的神经网络模型Qt,实现参数的动态调整和系统模型的快速迁移。 - 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,步骤5中,更换目标时刻,利用元学习框架将训练好的神经网络快速迁移到新的任务具体包括如下步骤:步骤51、将训练好的目标神经网络模型Qt作为源模型,下一时刻t+1作为新的目标时刻,接下来的任务是将神经网络模型从时刻t快速迁移到时刻t+1;步骤52、根据步骤3进行数据预处理,构建任务池;步骤53、根据步骤4,获得新的目标神经网络Qt+1的参数θt+1。
- 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建 方法,其特征在于,步骤6中,迭代步骤5直到动态离散制造系统模型收敛,收敛后保存模型参数具体包括如下步骤:步骤61、迭代步骤5,不断获取下一时刻的神经网络模型Qt、Qt+1、Qt+2…,直到模型参数收敛,表示该系统模型已经能很好的适应于不同时间段不同工况的生产环境,无论内外部条件如何变化,都能稳定的输出最优决策;步骤62、模型收敛后,保存模型参数,获得最终的动态离散制造系统模型。
- 如权利要求1所述的基于知识自演化的人工智能跨平台模型智能计算引擎构建方法,其特征在于,步骤7中,将动态离散制造系统模型用于新环境任务,测试其性能具体为:如果系统模型在新环境下能很好的输出调度策略,比原系统具有更高的效率,则结果符合预期,训练完成;如果结果不符合预期,则重新回到步骤1,重新训练。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211517991.1A CN115964934A (zh) | 2022-11-30 | 2022-11-30 | 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 |
CN202211517991.1 | 2022-11-30 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2024114121A1 true WO2024114121A1 (zh) | 2024-06-06 |
Family
ID=87358964
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2023/124030 WO2024114121A1 (zh) | 2022-11-30 | 2023-10-11 | 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN115964934A (zh) |
WO (1) | WO2024114121A1 (zh) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115964934A (zh) * | 2022-11-30 | 2023-04-14 | 南京邮电大学 | 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3965020A1 (en) * | 2020-09-07 | 2022-03-09 | Siemens Aktiengesellschaft | A model for a rapid continual anomaly detector |
CN114429009A (zh) * | 2022-04-07 | 2022-05-03 | 中国石油大学(华东) | 一种基于元迁移学习的小样本有杆泵井工况诊断方法 |
CN114663729A (zh) * | 2022-03-29 | 2022-06-24 | 南京工程学院 | 一种基于元学习的气缸套小样本缺陷检测方法 |
CN114879185A (zh) * | 2022-06-14 | 2022-08-09 | 中国人民解放军海军航空大学 | 基于任务经验迁移的雷达目标智能识别方法 |
CN115204249A (zh) * | 2022-05-13 | 2022-10-18 | 北京理工大学 | 一种基于竞争机制的群体智能元学习方法 |
CN115964934A (zh) * | 2022-11-30 | 2023-04-14 | 南京邮电大学 | 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 |
-
2022
- 2022-11-30 CN CN202211517991.1A patent/CN115964934A/zh active Pending
-
2023
- 2023-10-11 WO PCT/CN2023/124030 patent/WO2024114121A1/zh unknown
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3965020A1 (en) * | 2020-09-07 | 2022-03-09 | Siemens Aktiengesellschaft | A model for a rapid continual anomaly detector |
CN114663729A (zh) * | 2022-03-29 | 2022-06-24 | 南京工程学院 | 一种基于元学习的气缸套小样本缺陷检测方法 |
CN114429009A (zh) * | 2022-04-07 | 2022-05-03 | 中国石油大学(华东) | 一种基于元迁移学习的小样本有杆泵井工况诊断方法 |
CN115204249A (zh) * | 2022-05-13 | 2022-10-18 | 北京理工大学 | 一种基于竞争机制的群体智能元学习方法 |
CN114879185A (zh) * | 2022-06-14 | 2022-08-09 | 中国人民解放军海军航空大学 | 基于任务经验迁移的雷达目标智能识别方法 |
CN115964934A (zh) * | 2022-11-30 | 2023-04-14 | 南京邮电大学 | 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 |
Also Published As
Publication number | Publication date |
---|---|
CN115964934A (zh) | 2023-04-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110880036B (zh) | 神经网络压缩方法、装置、计算机设备及存储介质 | |
CN107729999A (zh) | 考虑矩阵相关性的深度神经网络压缩方法 | |
WO2024114121A1 (zh) | 一种基于知识自演化的人工智能跨平台模型智能计算引擎构建方法 | |
CN111563706A (zh) | 一种基于lstm网络的多变量物流货运量预测方法 | |
CN109829541A (zh) | 基于学习自动机的深度神经网络增量式训练方法及系统 | |
CN110503251B (zh) | 一种基于Stacking算法的非节假日负荷预测方法 | |
CN106909990A (zh) | 一种基于历史数据的预测方法及装置 | |
WO2021109644A1 (zh) | 一种基于元学习的混合动力车辆工况预测方法 | |
CN107230349B (zh) | 一种在线式实时短时间交通流预测方法 | |
CN116503676B (zh) | 一种基于知识蒸馏小样本增量学习的图片分类方法及系统 | |
CN111553118B (zh) | 基于强化学习的多维连续型优化变量全局优化方法 | |
CN111199272A (zh) | 一种面向智能车间的自适应调度方法 | |
CN108805257A (zh) | 一种基于参数范数的神经网络量化方法 | |
CN108665322A (zh) | 口粮消费量预测模型的构建方法、消费量预测方法及装置 | |
CN106997488A (zh) | 一种结合马尔科夫决策过程的动作知识提取方法 | |
CN114943345A (zh) | 基于主动学习和模型压缩的联邦学习全局模型训练方法 | |
US20200050982A1 (en) | Method and System for Predictive Modeling for Dynamically Scheduling Resource Allocation | |
CN113095484A (zh) | 基于lstm神经网络的股票价格预测方法 | |
CN108921281A (zh) | 一种基于深度网络和对抗技术的领域适配方法 | |
CN113609762A (zh) | 一种基于gru-mtl的电冷热负荷联合预测方法及系统 | |
CN110188978B (zh) | 一种基于深度学习的大学生专业推荐方法 | |
WO2024113585A1 (zh) | 一种面向离散制造系统的智能交互式决策方法 | |
CN108090564A (zh) | 基于网络权重初始和最终状态差异的冗余权重去除方法 | |
CN109086868B (zh) | 一种抽象图像情感识别方法 | |
CN115713386A (zh) | 一种多源信息融合的商品推荐方法及系统 |
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: 23896277 Country of ref document: EP Kind code of ref document: A1 |