WO2024113420A1 - 一种基于分布式学习的刀具寿命预测方法 - Google Patents

一种基于分布式学习的刀具寿命预测方法 Download PDF

Info

Publication number
WO2024113420A1
WO2024113420A1 PCT/CN2022/138819 CN2022138819W WO2024113420A1 WO 2024113420 A1 WO2024113420 A1 WO 2024113420A1 CN 2022138819 W CN2022138819 W CN 2022138819W WO 2024113420 A1 WO2024113420 A1 WO 2024113420A1
Authority
WO
WIPO (PCT)
Prior art keywords
terminal
local
model
target
data
Prior art date
Application number
PCT/CN2022/138819
Other languages
English (en)
French (fr)
Inventor
杨之乐
吴承科
刘祥飞
郭媛君
冯伟
Original Assignee
深圳先进技术研究院
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 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Publication of WO2024113420A1 publication Critical patent/WO2024113420A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/27Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • BPERFORMING OPERATIONS; TRANSPORTING
    • B23MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
    • B23QDETAILS, COMPONENTS, OR ACCESSORIES FOR MACHINE TOOLS, e.g. ARRANGEMENTS FOR COPYING OR CONTROLLING; MACHINE TOOLS IN GENERAL CHARACTERISED BY THE CONSTRUCTION OF PARTICULAR DETAILS OR COMPONENTS; COMBINATIONS OR ASSOCIATIONS OF METAL-WORKING MACHINES, NOT DIRECTED TO A PARTICULAR RESULT
    • B23Q17/00Arrangements for observing, indicating or measuring on machine tools
    • B23Q17/09Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool
    • B23Q17/0995Tool life management
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00Details relating to CAD techniques
    • G06F2111/02CAD in a network environment, e.g. collaborative CAD or distributed simulation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04Ageing analysis or optimisation against ageing
    • YGENERAL 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
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02PCLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30Computing systems specially adapted for manufacturing

Definitions

  • the present invention relates to the field of artificial intelligence technology, and in particular to a tool life prediction method based on distributed learning.
  • the tools used on CNC machine tools have a certain service life, and the service life is highly correlated with the tool material, performance parameters, historical defect rate, and material of the workpiece being machined.
  • tool users may use a variety of CNC machine tools to process workpieces of various shapes and materials, which requires a large amount of diverse data for training to ensure the accuracy of the prediction results.
  • manufacturers are unwilling to share these data.
  • CNC machine tool users cannot obtain the manufacturer's data for training their own neural network models, and the lack of training data leads to poor accuracy in model prediction results.
  • the present invention provides a tool life prediction method based on distributed learning, aiming to solve the problem in the prior art that the user of CNC machine tools has little training data, resulting in poor accuracy of the tool life prediction results of the model.
  • a first aspect of the present invention provides a tool life prediction method based on distributed learning, the method comprising:
  • the first terminal inputs the tool feature data in the local data into the local model held locally, obtains the tool life prediction result output by the local model, obtains the local training loss according to the tool life prediction result and the tool life label in the local data, and sends the local training loss to the second terminal;
  • the second terminal obtains the local training losses respectively sent by each of the first terminals, and updates the parameters of the target model based on each of the local training losses, wherein the structure of the target model is the same as the structure of the local model;
  • the second terminal predicts tool life based on the target model.
  • the tool life prediction method based on distributed learning, wherein the local model includes a feature extraction module and a prediction module; the first terminal inputs the tool feature data in the local data into the local model held locally, and obtains the tool life prediction result output by the local model, including:
  • the first terminal inputs the tool feature data into the feature extraction module to obtain a feature vector output by the feature extraction module;
  • the first terminal inputs the feature vector into the prediction module to obtain the tool life prediction result output by the prediction module.
  • the tool life prediction method based on distributed learning, wherein the second terminal updates the parameters of the target model based on each of the local training losses comprises:
  • the first terminal determines a target vector based on the feature vector, and sends the target vector to the second terminal;
  • the second terminal obtains the target vectors sent by each of the first terminals, inputs each of the target vectors into an attention model, and obtains a weight corresponding to each of the target vectors outputted from the attention model;
  • the second terminal performs weighted summation on each of the local training losses according to the weights corresponding to each of the target vectors to obtain a target training loss;
  • the second terminal updates parameters of the target model based on the target training loss.
  • the tool life prediction method based on distributed learning wherein the first terminal determines a target vector based on the feature vector, comprising:
  • the first terminal inputs the feature vector into a weight feature extraction model to obtain the target vector output by the weight feature extraction model.
  • the tool life prediction method based on distributed learning wherein the second terminal updates the parameters of the target model based on the target training loss, comprises:
  • the second terminal calculates a gradient based on the target training loss and uses a gradient descent method to update parameters of the target model and the attention model.
  • the tool life prediction method based on distributed learning wherein, after the second terminal updates the parameters of the target model based on the target training loss, the method further includes:
  • the second terminal calculates a gradient based on the target training loss, decomposes the gradient according to the weights corresponding to the target vectors, and sends the decomposed results to the first terminals respectively;
  • the first terminal updates the parameters of the local model based on the received decomposition result of the gradient.
  • the tool life prediction method based on distributed learning wherein the first terminal updates the parameters of the local model based on the received disassembly result of the gradient, comprises:
  • the first terminal updates parameters of the local model and the weight feature extraction model based on the received decomposition result of the gradient.
  • the method further includes:
  • the first terminal receives the data template published by the second terminal, sorts local data according to the data template, and obtains the tool feature data.
  • a second aspect of the present invention provides a tool life prediction system based on distributed learning, comprising a plurality of first terminals and a second terminal;
  • the first terminal is used to input tool feature data in the local data into a local model held locally, obtain a tool life prediction result output by the local model, obtain a local training loss based on the tool life prediction result and the tool life label in the local data, and send the local training loss to the second terminal;
  • the second terminal is used to obtain the local training losses sent by each of the first terminals, and update the parameters of the target model based on each of the local training losses.
  • the second terminal predicts the tool life based on the target model, and the structure of the target model is the same as that of the local model.
  • the present invention provides a tool life prediction method based on distributed learning.
  • each first terminal locally inputs the locally held tool feature data into the local model for training to obtain a local training loss, and then sends the local training loss to the second terminal.
  • the second terminal updates the parameters of the target model with the same structure as the local model according to the local training losses sent by each first terminal, and predicts the tool life based on the target model.
  • the tool manufacturer can use the data held by the first terminal locally, and only send the training loss to the second terminal used by the CNC machine tool user.
  • the model of the CNC machine tool user is trained using the data of multiple tool manufacturers, thereby improving the accuracy of the model used by the CNC machine tool user in predicting tool life.
  • FIG1 is a flow chart of an embodiment of a tool life prediction method based on distributed learning provided by the present invention
  • FIG. 2 is a structural principle diagram of an embodiment of a tool life prediction system based on distributed learning provided by the present invention.
  • the steps include:
  • the first terminal inputs the tool feature data in the local data into the local model held locally, obtains the tool life prediction result output by the local model, obtains the local training loss according to the tool life prediction result and the tool life label in the local data, and sends the local training loss to the second terminal;
  • the second terminal obtains the local training losses respectively sent by each of the first terminals, and updates the parameters of the target model based on each of the local training losses.
  • the structure of the target model is the same as that of the local model.
  • the first terminal may be a terminal controlled by a tool manufacturer, and the local data includes tool life data held by the tool manufacturer, including tool feature data, such as the material, performance parameters, historical defect rate, machine tool using the tool, material of the experimentally processed workpiece, etc. of the tool, and also includes a tool life label corresponding to the tool feature data, that is, the actual life of the tool.
  • the tool manufacturer stores the local data in the first terminal.
  • the second terminal is a terminal controlled by the user of the CNC machine tool. The second terminal itself holds limited tool life data, and the data for model training is insufficient.
  • the tool manufacturer specializes in tool production and conducts a large number of experiments, which can supplement the model training data and improve the performance of the model.
  • first terminals There can be multiple first terminals, that is, multiple tool manufacturers can participate in assisting CNC machine tool users in model training.
  • the second terminal can send the target model to each of the first terminals.
  • the first terminal After receiving the target model, the first terminal saves it locally as the local model.
  • the first terminal performs model training on the local model based on local data, obtains training loss, and then feeds it back to the second terminal.
  • the model prediction data is actually data processing, and the data input to the model needs to be processed into a unified format.
  • the method further includes:
  • the first terminal receives the data template published by the second terminal, sorts local data according to the data template, and obtains the tool feature data.
  • the data template includes data types and corresponding positions.
  • the data input into the model can be processed into a matrix form, and each row of the matrix is a type of data.
  • the first terminal fills the data into the position corresponding to the type of data. For data types that are not available locally, fill in 0.
  • the local model includes a feature extraction module and a prediction module.
  • the first terminal inputs tool feature data in the local data into the local model held locally, and obtains the tool life prediction result output by the local model, including:
  • the first terminal inputs the tool feature data into the feature extraction module to obtain a feature vector output by the feature extraction module;
  • the first terminal inputs the feature vector into the prediction module to obtain the tool life prediction result output by the prediction module.
  • the feature vector can reflect the characteristics of the tool feature data. Specifically, different tool feature data have different contributions to model training. Therefore, in this embodiment, different first terminals are assigned different weights based on their different tool feature data. However, obviously, the tool feature data cannot be sent directly to the second terminal to assign weights, which will cause the leakage of privacy data. In one possible implementation, the first terminal sends the feature vector to the second terminal to assign weights to each of the first terminals. However, since the architecture of the local model and the target model is the same, it is possible for the second terminal to reverse the tool feature data held by the first terminal based on the feature vector. In order to protect privacy data, in this embodiment, the second terminal updates the parameters of the target model based on each of the local training losses, including:
  • the first terminal determines a target vector based on the feature vector, and sends the target vector to the first terminal;
  • the second terminal obtains the target vector sent by each of the second terminals, inputs each of the target vectors into an attention model, and obtains a weight corresponding to each of the target vectors outputted from the attention model;
  • the second terminal performs weighted summation on each of the local training losses according to the weights corresponding to each of the target vectors to obtain a target training loss;
  • the second terminal updates parameters of the target model based on the target training loss.
  • the first terminal processes the feature vector and then sends it to the second terminal, and the first terminal determines the target vector based on the feature vector, including:
  • the first terminal inputs the feature vector into a weight feature extraction model to obtain the target vector output by the weight feature extraction model.
  • the weight feature extraction model is local to the first terminal. Since the weight feature extraction model is local to the first terminal and will not be transmitted to the second terminal, the second terminal cannot know the structure and parameters of the weight feature extraction model. Therefore, after receiving the target vector, the second terminal cannot infer the tool feature data held by the first terminal.
  • the second terminal inputs the target vector received from each of the first terminals into the attention model.
  • the attention model adopts the attention mechanism and is a common weight assignment model that can assign weights to the input data.
  • the second terminal obtains the weights of each of the target vectors output by the attention model, it performs weighted summation on each of the local training losses according to the weights corresponding to each of the target vectors to obtain the target training loss.
  • the weight corresponding to the local training loss is the weight of the target vector corresponding to the local training loss.
  • the second terminal updates the parameters of the target model based on the target training loss. Specifically, the second terminal updates the parameters of the target model using a gradient descent method. The second terminal updates the parameters of the target model based on the target training loss, including:
  • the second terminal calculates a gradient based on the target training loss and uses a gradient descent method to update parameters of the target model and the attention model.
  • the attention model is also trained together with the target model, so that the attention model can achieve more accurate weight allocation.
  • the parameters of the local model held by each of the first terminals are also updated.
  • the method further includes:
  • the second terminal calculates a gradient based on the target training loss, decomposes the gradient according to the weights corresponding to the target vectors, and sends the decomposed results to the first terminals respectively;
  • the first terminal updates the parameters of the local model based on the received decomposition result of the gradient.
  • the second terminal After the second terminal calculates the gradient, it decomposes the gradient according to the weights corresponding to each target vector and sends it back to each first terminal accordingly.
  • the first terminal updates the local model based on the received gradient, so that the model update process of each first terminal is not only based on its own data, but also can learn the features of the data held by other first terminals, so as to achieve more accurate feature extraction and tool life prediction.
  • the parameters of the weight feature extraction model held locally by the first terminal are also updated together to ensure that the target vector extracted by the weight feature extraction model can accurately reflect the contribution of the feature vector of each first terminal, that is, the first terminal updates the parameters of the local model based on the decomposition result of the received gradient, including:
  • the first terminal updates parameters of the local model and the weight feature extraction model based on the received decomposition result of the gradient.
  • the method provided in this embodiment further includes the steps of:
  • S300 The second terminal predicts tool life based on the target model.
  • the second terminal inputs the tool feature data it holds into the target model, obtains the tool life prediction result output by the target model to realize tool life prediction. It can be understood that when the second terminal uses the target model to predict tool life, the data input into the target model should be consistent with the data template.
  • this embodiment provides a tool life prediction method based on distributed learning, wherein each first terminal locally inputs the locally held tool feature data into a local model for training to obtain a local training loss, and then sends the local training loss to the second terminal.
  • the second terminal updates the parameters of a target model having the same structure as the local model according to the local training losses sent by each first terminal, and predicts the tool life based on the target model.
  • the tool manufacturer can use the data held by the first terminal locally, and only send the training loss to the second terminal used by the CNC machine tool user.
  • the model of the CNC machine tool user is trained using the data of multiple tool manufacturers, thereby improving the accuracy of the model used by the CNC machine tool user in predicting tool life.
  • steps in the flowchart may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.
  • Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
  • the present invention also provides a tool life prediction system based on distributed learning.
  • the tool life prediction system based on distributed learning includes a plurality of first terminals and a second terminal;
  • the first terminal is used to input the tool feature data in the local data into the local model held locally, obtain the tool life prediction result output by the local model, and obtain the local training loss according to the tool life prediction result and the tool life label in the local data, and send the local training loss to the second terminal, as described in the first embodiment;
  • the second terminal is used to obtain the local training losses sent by each of the first terminals respectively, and update the parameters of the target model based on each of the local training losses.
  • the second terminal predicts the tool life based on the target model.
  • the structure of the target model is the same as that of the local model, as described in Example 1.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • General Engineering & Computer Science (AREA)
  • Geometry (AREA)
  • Computer Hardware Design (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Mechanical Engineering (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种基于分布式学习的刀具寿命预测方法,包括:第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取本地模型输出的刀具寿命预测结果,并根据刀具寿命预测结果和本地数据中的刀具寿命标签得到本地训练损失,将本地训练损失发送至第二终端;第二终端获取各个第一终端分别发送的本地训练损失,基于各个本地训练损失更新目标模型的参数,目标模型的结构和本地模型的结构相同;第二终端基于目标模型进行刀具寿命预测。本发明可以在刀具生产厂商不会泄露本地数据的前提下,利用到多个刀具生产厂商的数据进行了数控机床使用方的模型的训练,提升数控机床使用方使用的模型预测刀具寿命的准确性。

Description

一种基于分布式学习的刀具寿命预测方法 技术领域
本发明涉及人工智能技术领域,特别涉及一种基于分布式学习的刀具寿命预测方法。
背景技术
数控机床上使用的刀具有一定的使用寿命,而且使用寿命和刀具本身的材质、性能参数、历史缺陷率以及加工工件的材质等具有很高的关联性,目前,已有采用神经网络进行刀具寿命预测的方法,但是,刀具使用方可能会在多种数控机床,加工多种不同形态材质的工件,这就需要大量多样的数据来进行训练以保证预测结果的准确性。而鉴于刀具本身的性能参数、历史缺陷率属于刀具生产厂商的秘密,包括刀具生产中的试验数据,生产厂商并不愿意分享这些数据,数控机床使用方无法获取到生产厂商的数据用于训练自己的神经网络模型,训练数据少导致模型预测结果准确度差。
因此,现有技术还有待改进和提高。
技术问题
针对现有技术的上述缺陷,本发明提供一种基于分布式学习的刀具寿命预测方法,旨在解决现有技术中数控机床使用方训练数据少导致模型进行刀具寿命预测的结果准确度差的问题。
技术解决方案
为了解决上述技术问题,本发明所采用的技术方案如下:
本发明的第一方面,提供一种基于分布式学习的刀具寿命预测方法,所述方法包括:
第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,并根据所述刀具寿命预测结果和所述本地数据中的刀具寿命标签得到本地训练损失,将所述本地训练损失发送至第二终端;
所述第二终端获取各个所述第一终端分别发送的所述本地训练损失,基于各个所述本地训练损失更新目标模型的参数,所述目标模型的结构和所述本地模型的结构相同;
所述第二终端基于所述目标模型进行刀具寿命预测。
所述的基于分布式学习的刀具寿命预测方法,其中,所述本地模型包括特征提取模块和预测模块;所述第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,包括:
所述第一终端将所述刀具特征数据输入至所述特征提取模块,获取所述特征提取模块输出的特征向量;
所述第一终端将所述特征向量输入至所述预测模块,获取所述预测模块输出的所述刀具寿命预测结果。
所述的基于分布式学习的刀具寿命预测方法,其中,所述第二终端基于各个所述本地训练损失更新目标模型的参数,包括:
所述第一终端基于所述特征向量确定目标向量,将所述目标向量发送至所述第二终端;
所述第二终端获取各个所述第一终端发送的所述目标向量,将各个所述目标向量输入至注意力模型中,获取所述注意力模型中输出的各个所述目标向量对应的权重;
所述第二终端根据各个所述目标向量对应的权重对各个所述本地训练损失进行加权求和,得到目标训练损失;
所述第二终端基于所述目标训练损失更新所述目标模型的参数。
所述的基于分布式学习的刀具寿命预测方法,其中,所述第一终端基于所述特征向量确定目标向量,包括:
所述第一终端将所述特征向量输入至权重特征提取模型,获取所述权重特征提取模型输出的所述目标向量。
所述的基于分布式学习的刀具寿命预测方法,其中,所述第二终端基于所述目标训练损失更新所述目标模型的参数,包括:
所述第二终端基于所述目标训练损失计算梯度,采用梯度下降法更新所述目标模型和所述注意力模型的参数。
所述的基于分布式学习的刀具寿命预测方法,其中,所述第二终端基于所述目标训练损失更新所述目标模型的参数之后,所述方法还包括:
所述第二终端基于所述目标训练损失计算梯度,将所述梯度按照各个所述目标向量对应的权重进行拆解,将拆解后的结果分别发送至各个所述第一终端;
所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型的参数。
所述的基于分布式学习的刀具寿命预测方法,其中,所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型的参数,包括:
所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型和所述权重特征提取模型的参数。
所述的基于分布式学习的刀具寿命预测方法,其中,所述第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型之前,所述方法还包括:
所述第一终端接收所述第二终端发布的数据模板,根据所述数据模板整理本地数据,得到所述刀具特征数据。
本发明的第二方面,提供一种基于分布式学习的刀具寿命预测系统,包括包括多个第一终端和第二终端;
所述第一终端用于将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,并根据所述刀具寿命预测结果和所述本地数据中的刀具寿命标签得到本地训练损失,将所述本地训练损失发送至所述第二终端;
所述第二终端用于获取各个所述第一终端分别发送的所述本地训练损失,基于各个所述本地训练损失更新目标模型的参数,所述第二终端基于所述目标模型进行刀具寿命预测,所述目标模型的结构和所述本地模型的结构相同。
有益效果
与现有技术相比,本发明提供了一种基于分布式学习的刀具寿命预测方法,所述的基于分布式学习的刀具寿命预测方法,各个第一终端在本地将本地持有的刀具特征数据输入至本地模型中进行训练,得到本地训练损失,再将本地训练损失发送给第二终端,第二终端根据各个第一终端发送的本地训练损失更新与本地模型结构相同的目标模型的参数,基于目标模型进行刀具寿命预测,这样,刀具生产厂商可以利用第一终端在本地使用自己持有的数据,只是将训练损失发送给数控机床使用方使用的第二终端,在刀具生产厂商不会泄露本地数据的前提下,利用到多个刀具生产厂商的数据进行了数控机床使用方的模型的训练,提升数控机床使用方使用的模型预测刀具寿命的准确性。
附图说明
图1为本发明提供的基于分布式学习的刀具寿命预测方法的实施例的流程图;
图2为本发明提供的基于分布式学习的刀具寿命预测系统的实施例的结构原理图。
本发明的实施方式
为使本发明的目的、技术方案及效果更加清楚、明确,以下参照附图并举实施例对本发明进一步详细说明。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。
实施例一
如图1所示,所述基于分布式学习的刀具寿命预测方法的一个实施例中,包括步骤:
S100、第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,并根据所述刀具寿命预测结果和所述本地数据中的刀具寿命标签得到本地训练损失,将所述本地训练损失发送至第二终端;
S200、所述第二终端获取各个所述第一终端分别发送的所述本地训练损失,基于各个所述本地训练损失更新目标模型的参数,所述目标模型的结构和所述本地模型的结构相同。
所述第一终端可以为刀具生产厂商控制的终端,所述本地数据中包括刀具生产厂商持有的刀具寿命数据,包括刀具特征数据,例如刀具的材质、性能参数、历史缺陷率、使用刀具的机床、实验加工工件的材质等,还包括刀具特征数据对应的刀具寿命标签,即刀具的实际寿命。刀具生产厂商将所述本地数据存储在所述第一终端中。所述第二终端为数控机床使用方控制的终端,所述第二终端自身持有的刀具寿命数据有限,用于模型训练的数据不足,而刀具生产厂商专门从事刀具生产,会进行大量的实验,可以补充模型训练数据,提升模型的性能。
所述第一终端可以有多个,即多个刀具生产厂商都可以参与协助数控机床使用方的模型训练。所述第二终端在制定所述目标模型的结构后,可以将所述目标模型发送至各个所述第一终端,所述第一终端接收到所述目标模型后保存在本地作为所述本地模型,所述第一终端基于本地数据在所述本地模型上进行模型训练,得到训练损失,再反馈给所述第二终端。
进一步地,模型预测数据实际上是对数据进行处理,需要将输入至模型的数据处理为统一的格式,所述第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型之前,所述方法还包括:
所述第一终端接收所述第二终端发布的数据模板,根据所述数据模板整理本地数据,得到所述刀具特征数据。
所述数据模板中包括数据类型和对应的位置,一般来说,可以将输入至模型的数据处理为矩阵的形式,矩阵的每一行为一种数据,所述第一终端接收到所述数据模板后,将数据填入至该类型数据对应的位置中,对于本地没有的数据类型,填入0。
所述本地模型包括特征提取模块和预测模块,所述第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,包括:
所述第一终端将所述刀具特征数据输入至所述特征提取模块,获取所述特征提取模块输出的特征向量;
所述第一终端将所述特征向量输入至所述预测模块,获取所述预测模块输出的所述刀具寿命预测结果。
所述特征向量可以反映所述刀具特征数据的特征,具体地,不同的刀具特征数据,其对于模型训练的贡献度是不同的,因此,在本实施例中,对于不同的所述第一终端,基于其刀具特征数据的不同,赋予不同的权重。但是,显然地,不可以将所述刀具特征数据直接发送给所述第二终端去赋予权重,这会造成隐私数据的泄露。在一种可能的实现方式中,所述第一终端将所述特征向量发送给所述第二终端去对各个所述第一终端进行赋权,然而,由于所述本地模型和所述目标模型的架构是一样的,这样使得所述第二终端有可能根据所述特征向量倒推所述第一终端持有的所述刀具特征数据。为了保护隐私数据,在本实施例中,所述第二终端基于各个所述本地训练损失更新目标模型的参数,包括:
所述第一终端基于所述特征向量确定目标向量,将所述目标向量发送至所述第一终端;
所述第二终端获取各个所述第二终端发送的所述目标向量,将各个所述目标向量输入至注意力模型中,获取所述注意力模型中输出的各个所述目标向量对应的权重;
所述第二终端根据各个所述目标向量对应的权重对各个所述本地训练损失进行加权求和,得到目标训练损失;
所述第二终端基于所述目标训练损失更新所述目标模型的参数。
为了保证本地持有的隐私数据的安全性,在本实施例中,所述第一终端对于所述特征向量进行处理之后再发送给所述第二终端,所述第一终端基于所述特征向量确定目标向量,包括:
所述第一终端将所述特征向量输入至权重特征提取模型,获取所述权重特征提取模型输出的所述目标向量。
所述权重特征提取模型在所述第一终端的本地,由于所述权重特征提取模型在所述第一终端的本地,不会被传输至所述第二终端,所述第二终端无法知晓所述权重特征提取模型的结构和参数,因此,接收到所述目标向量后,无法推理得到所述第一终端持有的所述刀具特征数据。
所述第二终端将从各个所述第一终端处接收的所述目标向量输入至注意力模型中,注意力模型采用注意力机制,是常见的权重赋予模型,可以为输入数据赋予权重。所述第二终端获取到所述注意力模型输出的各个所述目标向量的权重后,根据各个所述目标向量对应的权重对各个所述本地训练损失进行加权求和,得到所述目标训练损失。具体地,对所述本地训练损失进行加权求和时,所述本地训练损失对应的权重就是该所述本地训练损失对应的所述目标向量的权重。
所述第二终端基于所述目标训练损失更新所述目标模型的参数,具体地,所述第二终端采用梯度下降法更新所述目标模型的参数,所述第二终端基于所述目标训练损失更新所述目标模型的参数,包括:
所述第二终端基于所述目标训练损失计算梯度,采用梯度下降法更新所述目标模型和所述注意力模型的参数。
在模型训练过程中,所述注意力模型也和所述目标模型一起被训练,使得所述注意力模型能够实现跟更准确的权重分配。
进一步地,为了提升所述目标模型的训练效率,各个所述第一终端持有的所述本地模型的参数也会更新,所述第二终端基于所述目标训练损失更新所述目标模型的参数之后,所述方法还包括:
所述第二终端基于所述目标训练损失计算梯度,将所述梯度按照各个所述目标向量对应的权重进行拆解,将拆解后的结果分别发送至各个所述第一终端;
所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型的参数。
所述第二终端计算出梯度后,将所述梯度按照各个所述目标向量对应的权重进行拆解,对应发回至各个所述第一终端。所述第一终端基于接收到的梯度,更新所述本地模型,这样,各个所述第一终端的模型更新过程中不仅仅是基于自身的数据,还可以学习到其他的所述第一终端持有的数据的特征,实现更准确的特征提取和刀具寿命预测。
进一步地,所述第一终端本地持有的所述权重特征提取模型的参数也一同被更新,保证所述权重特征提取模型提取的所述目标向量能够准确反应各个所述第一终端的所述特征向量的贡献度,即,所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型的参数,包括:
所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型和所述权重特征提取模型的参数。
再次参阅图1,本实施例提供的方法,还包括步骤:
S300、所述第二终端基于所述目标模型进行刀具寿命预测。
所述第二终端将自身持有的刀具特征数据输入至所述目标模型中,获取所述目标模型输出的刀具寿命预测结果以实现刀具寿命预测,可以理解,所述第二终端在利用所述目标模型进行刀具寿命预测时,输入至所述目标模型的数据应该与所述数据模板一致。
综上所述,本实施例提供一种基于分布式学习的刀具寿命预测方法,各个第一终端在本地将本地持有的刀具特征数据输入至本地模型中进行训练,得到本地训练损失,再将本地训练损失发送给第二终端,第二终端根据各个第一终端发送的本地训练损失更新与本地模型结构相同的目标模型的参数,基于目标模型进行刀具寿命预测,这样,刀具生产厂商可以利用第一终端在本地使用自己持有的数据,只是将训练损失发送给数控机床使用方使用的第二终端,在刀具生产厂商不会泄露本地数据的前提下,利用到多个刀具生产厂商的数据进行了数控机床使用方的模型的训练,提升数控机床使用方使用的模型预测刀具寿命的准确性。
应该理解的是,虽然本发明说明书附图中给出的的流程图中的各个步骤按照箭头的指示依次显示,但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,流程图中的至少一部分步骤可以包括多个子步骤或者多个阶段,这些子步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些子步骤或者阶段的执行顺序也不必然是依次进行,而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机程序来指令相关的硬件来完成,所述的计算机程序可存储于一非易失性计算机可读取计算机可读存储介质中,该计算机程序在执行时,可包括如上述各方法的实施例的流程。其中,本发明所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink) DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
实施例二
基于上述实施例,本发明还相应提供了一种基于分布式学习的刀具寿命预测系统,如图2所示,所述基于分布式学习的刀具寿命预测系统包括多个第一终端和第二终端;
所述第一终端用于将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,并根据所述刀具寿命预测结果和所述本地数据中的刀具寿命标签得到本地训练损失,将所述本地训练损失发送至所述第二终端,具体如实施例一中所述;
所述第二终端用于获取各个所述第一终端分别发送的所述本地训练损失,基于各个所述本地训练损失更新目标模型的参数,所述第二终端基于所述目标模型进行刀具寿命预测,所述目标模型的结构和所述本地模型的结构相同,具体如实施例一中所述。
最后应说明的是:以上实施例仅用以说明本发明的技术方案,而非对其限制;尽管参照前述实施例对本发明进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本发明各实施例技术方案的精神和范围。

Claims (9)

  1. 一种基于分布式学习的刀具寿命预测方法,其特征在于,所述方法包括:
    第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,并根据所述刀具寿命预测结果和所述本地数据中的刀具寿命标签得到本地训练损失,将所述本地训练损失发送至第二终端;
    所述第二终端获取各个所述第一终端分别发送的所述本地训练损失,基于各个所述本地训练损失更新目标模型的参数,所述目标模型的结构和所述本地模型的结构相同;
    所述第二终端基于所述目标模型进行刀具寿命预测。
  2. 根据权利要求1所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述本地模型包括特征提取模块和预测模块;所述第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,包括:
    所述第一终端将所述刀具特征数据输入至所述特征提取模块,获取所述特征提取模块输出的特征向量;
    所述第一终端将所述特征向量输入至所述预测模块,获取所述预测模块输出的所述刀具寿命预测结果。
  3. 根据权利要求2所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述第二终端基于各个所述本地训练损失更新目标模型的参数,包括:
    所述第一终端基于所述特征向量确定目标向量,将所述目标向量发送至所述第二终端;
    所述第二终端获取各个所述第一终端发送的所述目标向量,将各个所述目标向量输入至注意力模型中,获取所述注意力模型中输出的各个所述目标向量对应的权重;
    所述第二终端根据各个所述目标向量对应的权重对各个所述本地训练损失进行加权求和,得到目标训练损失;
    所述第二终端基于所述目标训练损失更新所述目标模型的参数。
  4. 根据权利要求3所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述第一终端基于所述特征向量确定目标向量,包括:
    所述第一终端将所述特征向量输入至权重特征提取模型,获取所述权重特征提取模型输出的所述目标向量。
  5. 根据权利要求3所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述第二终端基于所述目标训练损失更新所述目标模型的参数,包括:
    所述第二终端基于所述目标训练损失计算梯度,采用梯度下降法更新所述目标模型和所述注意力模型的参数。
  6. 根据权利要求4所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述第二终端基于所述目标训练损失更新所述目标模型的参数之后,所述方法还包括:
    所述第二终端基于所述目标训练损失计算梯度,将所述梯度按照各个所述目标向量对应的权重进行拆解,将拆解后的结果分别发送至各个所述第一终端;
    所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型的参数。
  7. 根据权利要求6所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型的参数,包括:
    所述第一终端基于接收到的所述梯度的拆解结果更新所述本地模型和所述权重特征提取模型的参数。
  8. 根据权利要求1所述的基于分布式学习的刀具寿命预测方法,其特征在于,所述第一终端将本地数据中的刀具特征数据输入至本地持有的本地模型之前,所述方法还包括:
    所述第一终端接收所述第二终端发布的数据模板,根据所述数据模板整理本地数据,得到所述刀具特征数据。
  9. 一种基于分布式学习的刀具寿命预测系统,其特征在于,所述系统包括多个第一终端和第二终端;
    所述第一终端用于将本地数据中的刀具特征数据输入至本地持有的本地模型中,获取所述本地模型输出的刀具寿命预测结果,并根据所述刀具寿命预测结果和所述本地数据中的刀具寿命标签得到本地训练损失,将所述本地训练损失发送至所述第二终端;
    所述第二终端用于获取各个所述第一终端分别发送的所述本地训练损失,基于各个所述本地训练损失更新目标模型的参数,所述第二终端基于所述目标模型进行刀具寿命预测,所述目标模型的结构和所述本地模型的结构相同。
PCT/CN2022/138819 2022-12-01 2022-12-13 一种基于分布式学习的刀具寿命预测方法 WO2024113420A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211525896.6 2022-12-01
CN202211525896.6A CN115600512B (zh) 2022-12-01 2022-12-01 一种基于分布式学习的刀具寿命预测方法

Publications (1)

Publication Number Publication Date
WO2024113420A1 true WO2024113420A1 (zh) 2024-06-06

Family

ID=84852931

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/138819 WO2024113420A1 (zh) 2022-12-01 2022-12-13 一种基于分布式学习的刀具寿命预测方法

Country Status (2)

Country Link
CN (1) CN115600512B (zh)
WO (1) WO2024113420A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116672094B (zh) * 2023-06-03 2023-10-13 江苏省人民医院(南京医科大学第一附属医院) 能量器械使用寿命分析系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109754105A (zh) * 2017-11-07 2019-05-14 华为技术有限公司 一种预测方法及终端、服务器
CN113536671A (zh) * 2021-07-09 2021-10-22 北京航空航天大学 一种基于lstm的锂电池寿命预测方法
CN113867263A (zh) * 2021-08-27 2021-12-31 大唐互联科技(武汉)有限公司 一种基于云边协同及机器学习的智能刀具管理系统
CN114930347A (zh) * 2020-02-03 2022-08-19 英特尔公司 用于无线边缘动态的分布式学习的系统和方法

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109465676B (zh) * 2018-09-13 2020-10-16 成都数之联科技有限公司 一种刀具寿命预测方法
CN110377984B (zh) * 2019-07-02 2023-08-18 中国科学院深圳先进技术研究院 一种工业设备剩余有效寿命预测方法、系统及电子设备
CN111178524B (zh) * 2019-12-24 2024-06-14 中国平安人寿保险股份有限公司 基于联邦学习的数据处理方法、装置、设备及介质
CN114650288B (zh) * 2020-12-02 2024-03-08 中国科学院深圳先进技术研究院 分布式训练方法及系统、终端设备、计算机可读存储介质
CN113536670B (zh) * 2021-07-09 2024-01-12 北京航空航天大学 一种群体产品剩余寿命预测的联邦迁移学习方法
CN114066100B (zh) * 2021-12-01 2024-06-25 深圳市新威尔电子有限公司 基于联邦学习的在线式电池循环寿命预测系统及方法
CN114818476B (zh) * 2022-04-01 2023-08-22 西南交通大学 一种应用于旋转机械设备寿命预测的联邦学习系统及方法
CN114676647B (zh) * 2022-05-30 2022-08-23 中科航迈数控软件(深圳)有限公司 一种基于深度学习方法的数控机床零部件寿命预测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109754105A (zh) * 2017-11-07 2019-05-14 华为技术有限公司 一种预测方法及终端、服务器
CN114930347A (zh) * 2020-02-03 2022-08-19 英特尔公司 用于无线边缘动态的分布式学习的系统和方法
CN113536671A (zh) * 2021-07-09 2021-10-22 北京航空航天大学 一种基于lstm的锂电池寿命预测方法
CN113867263A (zh) * 2021-08-27 2021-12-31 大唐互联科技(武汉)有限公司 一种基于云边协同及机器学习的智能刀具管理系统

Also Published As

Publication number Publication date
CN115600512A (zh) 2023-01-13
CN115600512B (zh) 2023-04-14

Similar Documents

Publication Publication Date Title
EP3811241B1 (en) Product performance prediction modeling method and apparatus, computer device, computer-readable storage medium, and product performance prediction method and prediction system
CN114296408A (zh) 一种数字孪生环境中的生产制造车间人工智能优化算法模型系统及其算法
WO2024113420A1 (zh) 一种基于分布式学习的刀具寿命预测方法
CN109978403B (zh) 一种产品装配过程的质量管控方法、装置及设备
US6678668B2 (en) System and method for complex process optimization and control
CN109472057B (zh) 基于跨工步隐式参数记忆的产品加工质量预测装置及方法
CN111913458B (zh) 一种基于深度学习的车间控制方法及系统
CN108491253A (zh) 一种计算任务处理方法以及边缘计算设备
CN114611379A (zh) 一种基于数据驱动的加工工艺节能规划方法
CN114881301A (zh) 生产线的仿真排产方法、系统、终端设备及存储介质
Lidberg et al. Applying aggregated line modeling techniques to optimize real world manufacturing systems
KR20230081501A (ko) 생산 계획 수립 장치 및 방법
CN111369130B (zh) 基于语义数据和知识推理的分布式自适应产线重构方法
CN110084462B (zh) 一种基于故障rgv智能二道工序加工调度方法
CN116414094A (zh) 一种焊接装配智能调度方法及系统
CN108573275B (zh) 一种在线分类微服务的构建方法
CN114240023A (zh) 一种火工品生产车间的流程再造方法及系统
Thamilselvan et al. Integrating genetic algorithm, tabu search and simulated annealing for job shop scheduling problem
Vosniakos et al. Neural network simulation metamodels and genetic algorithms in analysis and design of manufacturing cells
KR20230028466A (ko) 반도체 제조 환경의 자동화된 지원
Valeev et al. Intelligent Decision Support System of Operation Sequencing of Technological Processes Planning
CN113536464B (zh) 一种基于人工智能的尺寸控制方法及尺寸控制系统
CN115826506B (zh) 水处理控制系统及控制装置
AU2021105079A4 (en) The controlling machine tools system by using machine learning techniques.
CN116910984B (zh) 压射曲线参数优化方法、装置、电子设备和存储介质