CN115169617B - Mold maintenance prediction model training method, mold maintenance prediction method and system - Google Patents

Mold maintenance prediction model training method, mold maintenance prediction method and system Download PDF

Info

Publication number
CN115169617B
CN115169617B CN202211092921.6A CN202211092921A CN115169617B CN 115169617 B CN115169617 B CN 115169617B CN 202211092921 A CN202211092921 A CN 202211092921A CN 115169617 B CN115169617 B CN 115169617B
Authority
CN
China
Prior art keywords
product
model
mold
neural network
maintenance prediction
Prior art date
Legal status (The legal status 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 status listed.)
Active
Application number
CN202211092921.6A
Other languages
Chinese (zh)
Other versions
CN115169617A (en
Inventor
令狐彬
胡炳彰
周璠
许�鹏
鲍江宏
高磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei Zhongke Dihong Automation Co ltd
Original Assignee
Hefei Zhongke Dihong Automation Co ltd
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 Hefei Zhongke Dihong Automation Co ltd filed Critical Hefei Zhongke Dihong Automation Co ltd
Priority to CN202211092921.6A priority Critical patent/CN115169617B/en
Publication of CN115169617A publication Critical patent/CN115169617A/en
Application granted granted Critical
Publication of CN115169617B publication Critical patent/CN115169617B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/20—Administration of product repair or maintenance
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/04—Manufacturing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04—Ageing analysis or optimisation against ageing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20076—Probabilistic image processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Quality & Reliability (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • Marketing (AREA)
  • Tourism & Hospitality (AREA)
  • General Engineering & Computer Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Databases & Information Systems (AREA)
  • Geometry (AREA)
  • Manufacturing & Machinery (AREA)
  • Primary Health Care (AREA)
  • Mathematical Physics (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Feedback Control In General (AREA)
  • Moulds For Moulding Plastics Or The Like (AREA)

Abstract

A mold maintenance prediction model training method, a mold maintenance prediction method and a system. The die maintenance prediction model training method can train a die maintenance prediction model for accurately predicting die maintenance dynamics based on product quality data, and the die maintenance prediction model takes multi-modal characteristics formed by product image data and size data of a product produced by a die as the input of the die maintenance prediction model of the product, so that the characteristics are mutually complemented; predicting the trend of the characteristic vector of the product by adopting a multi-scale circulating neural network, predicting the wear state quantity of the mold at the future time by adopting a depth feedforward network, and drawing a change curve chart of the wear state quantity of the mold; and then, according to a preset die wear state quantity threshold value, calculating the remaining service life of the die at the current moment, and realizing the die maintenance prediction.

Description

Mold maintenance prediction model training method, mold maintenance prediction method and system
Technical Field
The invention relates to the field of health management of machining dies, in particular to a die maintenance prediction model training method, a die maintenance prediction method and a die maintenance prediction system.
Background
With the rapid development of internet technology in recent years, the deep integration of manufacturing industry and new generation information technology is not feasible, so that the production and processing flow becomes more automatic and intelligent, and the equipment fault diagnosis and maintenance prediction become key problems for realizing industrial intelligence. The mould is the most central processing element in the production and processing system, and it has direct influence to product quality, in case the mould takes place to damage the trouble and can't in time discover, light then direct influence product processingquality and production efficiency, seriously even lead to mill's machine to damage and harm staff's life safety, greatly increased enterprise's manufacturing cost, consequently accurate prediction mould remaining life and in time maintain, have important meaning to manufacturing enterprise.
Specifically, the mould maintenance prediction mainly realizes the online real-time monitoring of the quality of the processed product of the whole workshop equipment; identifying the abrasion state of the mold in time and mastering the abrasion change condition of the mold; accurately predicting the residual service life of the die and realizing die maintenance prediction; the intelligent early warning of the service life of the die is realized, the worker is informed to change the die in advance, and the old die is maintained, so that the fault time is reduced.
As processing time progresses, more and more process information is generated by the production shop, and how to extract information related to the life of the mold from the process data plays a crucial role in the maintenance prediction of the mold. The mainstream method for predicting the maintenance of the mold is basically based on a neural network algorithm, monitoring signal data such as vibration and the like collected in the processing process are input into a built neural network model for training after data preprocessing, feature extraction, dimension reduction and feature selection processing, and a prediction result of the residual life of the mold is output by using the trained model. Although the artificial intelligence technology is used for analyzing the processing signal data and obtaining good prediction effect, the method has certain defects that the internal attribute of the quality of the processed product is not considered, the internal relation between the quality characteristic data of the processed product and the health condition of the mould is not fully excavated, and the wear state information of the mould is not fully utilized and the influence of different characteristic parameters on the performance of the mould is not distinguished, so that the prediction precision of the model is influenced. In addition, the existing model lacks the characteristic of dynamically updating the residual life of the mold in real time on line, and dynamic maintenance prediction of the mold cannot be realized, so that further improvement of the performance of the model is limited.
Disclosure of Invention
In order to overcome the defect of predicting the residual life based on the vibration data of mold production in the prior art, the invention provides a mold maintenance prediction model training method which can train a mold maintenance prediction model for accurately and accurately predicting the maintenance dynamics of a mold based on product quality data.
The invention adopts the following technical scheme:
a method of training a mold maintenance prediction model for predicting a remaining life of a mold, the training method comprising the steps of:
s1, obtaining product quality data, wherein the product quality data comprise product images of the same product at different angles and multiple product size data, and the product size data are size data for displaying the shape characteristics of the product; constructing a marked sample, wherein the marked sample comprises product quality data of a product produced by the die at a certain time and the residual service life of the die at the time, and enablingiIndicates the serial number of the marked sample,y i indicates the production of the moldiThe remaining life of the product corresponding to each label sample,y i is recorded asiReal labeling of each labeled sample;
s2, constructing an initial model, wherein the initial model comprises a preprocessing module, a self-attention mechanism module, a multi-scale circulating neural network and a mould maintenance prediction module;
the preprocessing module is used for preprocessing the product images at different angles to obtain the feature vectors of the product images at all angles and construct a product image feature setF(ii) a The preprocessing module is used for preprocessing various product size data to obtain a product size feature setS’’；The preprocessing module is also used for collecting the product image characteristicsFAnd product size feature setS’’Weighted sum as sample dataXAnd outputting; order toX i Is shown asiSample data corresponding to each labeled sample;
F={f(1),f(2),…,f(r f ),…,f(n f )}；
S’’={s’’(1),s’’(2),…,s’’(r s ),…,s’’(n s )}；
wherein the content of the first and second substances,n f the number of image acquisition angles is represented,f(r f )represents from the firstr f Characteristic vectors corresponding to the product images acquired at the angles;n s the number of pieces of data representing the size of the product,s’’(r s )to show the product ofr s A dimensional feature;n s =n f ；1≤r f ≤n f ；1≤r s ≤n s ；
the input of the self-attention mechanism module is sample dataXOutput is asXFeature vector after self-attention mechanismX’；
The multi-scale recurrent neural network comprises a plurality of recurrent neural networks, each recurrent neural network is used for extracting the feature vector on different scales, and the input of the recurrent neural network is the feature vectorX’Or an output of itself;let the current time betAt a time of dayrThe input of the recurrent neural network isX’Its output is the current time eigenvectorX’’(r,t)(ii) a First, therThe input of the recurrent neural network isX’’(r,t’)Then, the output is the feature vector of the next timeX’’(r,t’+1)；t≤t’≤t+N- 1，NIs a set number of cycles; 1≤r≤R，RIs the number of recurrent neural networks;
the mould maintenance prediction module comprises a depth feedforward network and a mould service life prediction module; the deep feedforward network is based on the feature vector at each timeX’’(r,t+n)Predict the timet+nThe amount of the state of wear of the upper die,0≤n≤N(ii) a The mould life prediction module predicts the residual life of the mould according to a wear state quantity curve of the mould, and the wear state quantity curve of the mould is drawn by combining the wear state quantity of the mould at each moment;
the input of the initial model is the input of the preprocessing module, and the output of the initial model is the output of the mould life prediction module, namely the residual life of the mould;
and S3, enabling the initial model to automatically learn and label the sample to iterate the parameters until the set iteration condition is reached, fixing the parameters, and taking the initial model with the fixed parameters as a mold maintenance prediction model.
Preferably, both the preprocessing module and the self-attention mechanism module adopt fixed models, in the S3, the preprocessing module, the self-attention mechanism module and the multi-scale recurrent neural network form a pre-learning model, and the pre-learning model autonomously learns and marks samples to perform parameter iteration on the multi-scale recurrent neural network until the multi-scale recurrent neural network is fixed; then, learning the marked sample by using an initial model with fixed multi-scale cyclic neural network parameters so as to iterate the parameters of the mold maintenance prediction module;
the parameter iteration of the multi-scale recurrent neural network comprises the following steps:
SA31, setting transition parametersX’’’Order:
X’’’=w 1 ×X’’(1,t)+w 2 ×X’’(2,t)+…+w r ×X’’(r,t)+…+w R ×X’’(R,t)；
wherein the content of the first and second substances,w r represents a weight, 1≤r≤R，w 1 +w 2 +…+w r +…+w R =1；w r Self-adaptive adjustment is carried out through back propagation in the process that the pre-learning model autonomously learns the labeled samples;
SA32, enabling the pre-learning model to learn the marked samples, and performing parameter iteration on the multi-scale cyclic neural network in the pre-learning model by adopting an Adam optimizer in the learning process until the loss of the pre-learning model reaches the minimum value to fix the multi-scale cyclic neural network;
the loss of the pre-learning model is calculated using the following first loss function:
Figure 951798DEST_PATH_IMAGE001
；
wherein the content of the first and second substances,Lrepresenting the loss of the pre-learned model,Ma set of the most recent set of labeled samples representing pre-learning model learning,mis composed ofMThe number of the labeled samples in (1),iindicates the serial number of the marked sample,X i ’is shown asiSample data corresponding to each labeled sampleX i Feature vectors after the self-attention mechanism; sample dataX i The corresponding transition parameters are noted asX i ’’’When the input of the multi-scale circulation neural networkX’=X i ’When the temperature of the water is higher than the set temperature,X i ’’’=X’’’；λa weight representing a weight decay term in the first loss function,λis a set value;E(W,b)a weight decay term representing all weights and biases in the multi-scale recurrent neural network in the pre-learning model,Wandbweight vectors representing respectively multi-scale recurrent neural networks in a pre-learning modelAnd a bias vector.
Preferably, the parameter iteration method of the mold maintenance prediction module is as follows: substituting the multi-scale cyclic neural network with fixed parameters into the initial model, enabling the initial model to learn and label a sample, performing parameter iteration on a deep feedforward network in the initial model by adopting an Adam optimizer in the learning process until the loss of the initial model reaches the minimum, fixing the parameters of the deep feedforward network, and storing the initial model at the moment as a mould maintenance prediction model;
the loss function of the initial model is calculated by adopting the following second loss function:
Figure 498317DEST_PATH_IMAGE002
；
wherein the content of the first and second substances,L’which represents the loss of the initial model and,M’a set of the most recent set of labeled samples representing initial model learning,m’is composed ofM’The number of the labeled samples in (1),iindicates the serial number of the marked sample,y i is sample dataX i The corresponding real label is marked on the display screen,ŷ i representing sample dataX i Corresponding model label, when the input of the initial model isX i When it is, its output is recorded asŷ i ；λ’Representing the weight of the weighted decay term in the second loss function,λ’is a set value;E’(W’,b’)a weight decay term representing all weights and biases of the deep feed-forward network in the initial model,W’andb’respectively representing the weight vector and the bias vector of the deep feedforward network in the initial model.
Preferably, the set of product size characteristicsS’’The method comprises the following steps:
SD1, acquisitionn s Calculating the error between each item of size data and the item of size data of the standard product, and constructing a size error setS={s(1),s(2),…,s(r s ),…,s(n s )}；s(r s )To show the product ofr s Error corresponding to the item size data;
SD2, set of dimensional errorsSNormalizing each error in the error list to obtain a normalized size error setS’={s’(1),s’(2),…,s’(r s ),…,s’(n s )}，s’(r s )∈[0,1]；
SD3, pair setS’Carrying out embedding treatment to obtain a product size feature vector setS’’。
Preferably, the product image feature setFThe acquisition mode is as follows: within the set shooting timen f Collecting product images at different angles, preprocessing the images, and extracting the feature vectors of the preprocessed product images at all angles through a deep convolutional neural network so as to construct a product image feature setF(ii) a The shooting time is less than or equal to 1 minute.
Preferably, the sample data in S1XObtained according to the following formula:
X=pF+(1-p)S’’；
wherein the content of the first and second substances,p∈[0.5，0.6，0.7，0.8，0.9]。
preferably, the self-attention mechanism module obtains the sample data according to the following formulaXFeature vector ofX’；
Figure 78334DEST_PATH_IMAGE003
；
Wherein, the first and the second end of the pipe are connected with each other,Q、KandVboth of which represent a matrix of the image data,Q=W q X，K=W k X，V=W v X；W q 、W k andW v a parameter matrix representing linear mapping, wherein the three are intrinsic parameters of the self-attention mechanism;D k representing a matrix of vectorsKThe dimension (c) of (a) is,K T representing a matrix of vectorsKTransposing;softmaxrepresenting a normalized exponential function used to calculate the probability for all classes.
The invention also provides a mould maintenance prediction method, and the mould maintenance prediction model trained by the mould maintenance prediction model training method is adopted to predict the wear state quantity and the residual life of the mould, so that the real-time detection of the mould maintenance state is realized.
A method of predicting mold maintenance, comprising the steps of:
st1, obtaining a mould maintenance prediction model, wherein the mould maintenance prediction model is obtained by the mould maintenance prediction model training method;
st2, obtaining product quality data of a product currently produced by the mold to be predicted, inputting the product quality data into a mold maintenance prediction model, and obtaining the residual life output by the mold maintenance prediction model as an evaluation index of the mold maintenance prediction model.
The invention also provides a mould maintenance prediction system, and provides a carrier for popularization of the mould maintenance prediction model and the prediction method.
A mold maintenance prediction system includes a storage module having a mold maintenance prediction model stored therein and a computer program that, when executed, implements the mold maintenance prediction method.
Preferably, the system further comprises a processing module for executing the computer program to implement the mold maintenance prediction method.
The invention has the advantages that:
(1) Predicting the trend of the characteristic vector of the product by adopting a multi-scale circulating neural network, predicting the die wear state quantity at the future time by adopting a depth feedforward network, and drawing a die wear state quantity change curve chart; and calculating the residual service life of the mold at the current moment according to a preset threshold value of the abrasion state quantity of the mold, thereby realizing the maintenance prediction of the mold.
(2) The method uses the multi-modal characteristics composed of the product image data and the size data of the product produced by the mold as the input of the mold maintenance prediction model of the product, plays a role of mutual complementation on the characteristics, fully excavates the internal relation between the product quality data and the mold health condition, and effectively embeds the mold wear state information in the product characteristic vector extraction so as to ensure the accuracy of predicting the wear state quantity of the mold through the product quality data.
(3) The method applies the multi-scale recurrent neural network to an industrial scene, and predicts the characteristic data change at the future moment according to the product quality data; the method introduces a self-attention mechanism into the multi-scale cyclic neural network, fully considers the influence of the characteristics of different product quality data on the model prediction performance, and effectively improves the prediction precision of the model on the characteristic vector of the product at the future moment, thereby laying a foundation for predicting the wear state quantity trend of the mold at the next stage, realizing intelligent early warning of the service life of the mold, guiding the updating and optimization of the maintenance plan of the mold, effectively improving the production efficiency of enterprises, reducing the production cost and ensuring the product quality.
(4) The method monitors the quality data of the processed product in real time on line, and predicts the change of the abrasion state quantity of the die at the future time on line by using a depth feedforward network model based on the characteristic vector data of the product at the future time, so as to realize the online updating of the prediction result of the residual life of the die, thereby achieving the dynamic prediction effect of die maintenance.
Drawings
FIG. 1 is a flow chart of a method for training a mold maintenance prediction model;
FIG. 2 is a flow chart of a method for parameter iteration by a multi-scale recurrent neural network;
FIG. 3 is a set of product size featuresS’’Obtaining a flow chart;
FIG. 4 is a flow chart of a method of predicting mold maintenance;
FIG. 5 is a schematic view of the differencepA prediction curve of the corresponding mold maintenance prediction model;
FIG. 6 is a comparison of a wear state quantity curve labeled by the mold maintenance prediction model and an actual wear state quantity curve.
Detailed Description
Mould maintenance prediction model
The invention provides a mold maintenance prediction model which comprises a preprocessing module, a self-attention mechanism module, a multi-scale circulation neural network and a mold maintenance prediction module which are sequentially connected. The input of the mould maintenance prediction model is the input of the preprocessing module, and the output of the mould maintenance prediction model is the output of the mould service life prediction model, namely the residual service life of the mould.
The input of the preprocessing module is product images of a product produced by a mold at a certain moment and multiple product size data of the product, and the output of the preprocessing module is sample data obtained by weighting and summing a feature vector corresponding to the product image and a product size feature corresponding to the product size dataX。
The self-attention mechanism module is used for extracting sample dataXFeature vector after self-attention mechanismX’。
The multi-scale recurrent neural network comprises a plurality of recurrent neural networks, each recurrent neural network is used for extracting the feature vector on different scales, and the input of the recurrent neural network is the feature vectorX’Or an output of itself; let the current time betAt a time of dayrThe input of the recurrent neural network isX’Its output is the current time eigenvectorX’’(r,t)(ii) a First, therThe input of the recurrent neural network isX’’(r,t’)Then, the output is the feature vector of the next timeX’’(r,t’+1)；t≤t’≤t+N- 1，NIs a set number of cycles; 1≤r≤R，RIs the number of recurrent neural networks.
The mold maintenance prediction module includes a depth feed forward network and a mold life prediction module. The deep feedforward network is based on the feature vector at each timeX’’(r,t+n)Predict the timet+nThe amount of the state of wear of the upper die,0≤n≤N(ii) a And the mould life prediction module predicts the residual life of the mould according to a wear state quantity curve of the mould, and the wear state quantity curve of the mould is drawn by combining the wear state quantity of the mould at each moment.
Annotating a sample
The labeled sample in this embodiment can be written as a last pagex i ,y i },x i =(F 0 (i),S 0 (i))；x i As obtained from historical data of mold productioniProduct quality data corresponding to the product of each labeled sample,F 0 (i)is the firstiA set of product images corresponding to each annotated sample,S 0 (i))is the firstiA product size data set corresponding to each labeled sample;y i indicates the production of the moldiThe remaining life of the product corresponding to each label sample,y i is marked asiAnd (4) real labeling of each labeled sample.
Pre-processing module
The input of the preprocessing module is product images of different angles and a plurality of items of product size data, and the output of the preprocessing module is sample dataX。The product size data is size data showing the shape characteristics of the product, such as length, width, and the like.
X=pF+(1-p)S’’；
Wherein the content of the first and second substances,p∈[0.5，0.6，0.7，0.8，0.9]；Fthe method comprises the steps that a product image feature set is a set of feature vectors of product images of all angles obtained after input product images of different angles are preprocessed by a preprocessing module;S’’the method is a set of product size characteristics obtained after preprocessing input product size data by a preprocessing module.
Specifically, the method comprises the following steps:
F={f(1),f(2),…,f(r f ),…,f(n f )}；
S’’={s’’(1),s’’(2),…,s’’(r s ),…,s’’(n s )}；
wherein the content of the first and second substances,n f the number of image acquisition angles is represented,f(r f )represents from the firstr f Characteristic vectors corresponding to the product images acquired at an angle;n s the number of pieces of data representing the size of the product,s’’(r s )to express the productr s A dimensional feature;n s =n f ；1≤r f ≤n f ；1≤r s ≤n s 。
in this embodiment, the preprocessing module includes a depth convolution neural network that inputs product images at different angles and outputs feature vectors of the product images at the angles, and the preprocessing module extracts the feature vectors of the product images at the angles through the depth convolution neural network.
Referring to fig. 3, in this embodiment, the preprocessing module obtains the feature set of the product size by the following stepsS’’：
SD1, acquisitionn s Calculating the error between each item of size data and the item of size data of the standard product, and constructing a size error setS={s(1),s(2),…,s(r s ),…,s(n s )}；s(r s )To express the productr s Error associated with item size data, i.e. of productr s Item size data and standard product Nor s Error between the item size data; the product size data may be length, width, height, etc. of the product;
SD2, set of dimensional errorsSNormalizing the errors to obtain a normalized size error setS’={s’(1),s’(2),…,s’(r s ),…,s’(n s )}，s’(r s )∈[0,1]；
s’(r s )=s’(r s )/(s’(1)+s’(2)+…+s’(r s )+…+s’(n s ))；
SD3, set of pairsS’Carrying out embedding treatment to obtain a product size feature vector setS’’. The Embedding processing is a common data processing means in the field and is not described herein.
Self-attention mechanism module
In this embodiment, the self-attention mechanism module obtains sample data according to the following formulaXFeature vector ofX’；
Figure 620174DEST_PATH_IMAGE003
；
Wherein, the first and the second end of the pipe are connected with each other,Q、KandVeach of which represents a matrix of the image data,Q=W q X，K=W k X，V=W v X；W q 、W k andW v the parameter matrix representing the linear mapping, the three are intrinsic parameters of the attention-free mechanism, and the writing method is a fixed form, which is not described herein again;D k representing a vector matrixKThe dimension (c) of (a) is,K T representing a vector matrixKTransposing;softmaxrepresenting a normalized exponential function used to calculate the probability for all classes.
The self-attention mechanism module used in the present embodiment is a conventional technical means in the art, and is not specifically described again.
It should be noted that the deep neural convolution network used for extracting the image feature vector in the preprocessing module is an existing fixed model, and the normalization and data Embedding processing is an existing technical means, that is, the preprocessing module is a fixed model. The self-attention mechanism module is also a fixed model. Therefore, during the training of the mold maintenance prediction model, only the multi-scale cyclic neural network and the mold maintenance prediction module need to perform parameter iteration. In the embodiment, the parameters of the multi-scale recurrent neural network are first fixed, and after the parameters of the multi-scale recurrent neural network are fixed, the parameters of the mold maintenance prediction module are updated.
Multi-scale cyclic neural network training method
Referring to fig. 2, in the present embodiment, first, the preprocessing module, the self-attention mechanism module, and the multi-scale recurrent neural network form a pre-learning model, and the pre-learning model autonomously learns the labeled samples to perform parameter iteration on the multi-scale recurrent neural network until the loss of the pre-learning model reaches the minimum. Because the pre-learning model only performs parameter iteration on the multi-scale recurrent neural network in the process of learning the labeled sample, the loss of the pre-learning model is actually the loss of the multi-scale recurrent neural network.
In this embodiment, the loss of the multi-scale recurrent neural network, that is, the loss of the pre-learning model, is calculated according to the following first loss function:
Figure 729075DEST_PATH_IMAGE001
；
wherein, the first and the second end of the pipe are connected with each other,Lrepresenting the loss of the pre-learned model,Ma set of the most recent set of labeled examples representing pre-learning model learning,mis composed ofMThe number of the labeled samples in (1),iindicating the serial number of the marked sample;λrepresenting the weight of the weighted decay term in the first loss function,λthe value is a set value, and can be specifically between 0 and 1;E(W,b)a weight decay term representing all weights and biases in the multi-scale recurrent neural network,Wandbrespectively representing a weight vector and a bias vector in the multi-scale recurrent neural network;E(W,b)、Wandbupdating iterative parameter terms which are all multiscale recurrent neural networks, i.e.E(W,b)、WAndbare updated as the pre-learning model learns the labeled samples.
X i ’Denotes the firstiSample data corresponding to each labeled sampleX i Feature vectors after self-attention mechanism; sample dataX i Corresponding transition parameters are denoted byX i ’’’When the input of the multi-scale recurrent neural networkX’=X i ’When the temperature of the water is higher than the set temperature,X i ’’’=X’’’；
X’’’=w 1 ×X’’(1,t)+w 2 ×X’’(2,t)+…+w r ×X’’(r,t)+…+w R ×X’’(R,t)；
wherein the content of the first and second substances,w r represents a weight, 1≤r≤R，w 1 +w 2 +…+w r +…+w R =1；
Definition of the firstiMarking the sample data corresponding to each marked sample asX i ，X i The feature vectors after the self-attention mechanism are recorded asX i ’(ii) a The initial input of each recurrent neural network in the multi-scale recurrent neural network isX i ’Let us orderrA recurrent neural network input ofX i ’When it is, its output is recorded asX i ’’(r,t)(ii) a First, therThe input of the recurrent neural network isX i ’’(r,t’)Then, the output is the feature vector of the next timeX i ’’’(r,t’+1)。1≤r≤R，RIs the number of recurrent neural networks.
X i ’’’In the above formulaX’’’The same way of calculating (c) is:
X i ’’’=w 1 ×X i ’’(1,t)+w 2 ×X i ’’(2,t)+…+w r ×X i ’’(r,t)+…+w R ×X i ’’(R,t)。
in particular, the method comprises the following steps of,w 1 、w 2 、…、w r 、…、w R and self-adaptive adjustment is carried out through back propagation in the process of autonomously learning the labeled sample by the pre-learning model.
Training of mold maintenance prediction module
Referring to fig. 1, in the present embodiment, a preprocessing module, a self-attention mechanism module, a multi-scale recurrent neural network with fixed parameters, and a mold maintenance prediction module with initialized parameters are combined to construct an initial model, and the initial model is made to learn a labeled sample to update parameters of the mold maintenance prediction module until loss of the initial model, that is, loss of the mold maintenance prediction module, is minimized.
The loss function of the initial model is calculated by adopting the following second loss function:
Figure 356366DEST_PATH_IMAGE002
；
wherein the content of the first and second substances,L’which represents the loss of the initial model and,M’a set of the most recent set of labeled samples representing initial model learning,m’is composed ofM’The number of the labeled samples in (1),iindicates the serial number of the marked sample,y i is sample dataX i The corresponding real label is marked on the display screen,ŷ i representing sample dataX i Corresponding model label when the input of the initial model isX i When it is, its output is recorded asŷ i ；λ’Representing the weight of the weighted decay term in the second loss function,λ’the value is a set value, and can be specifically between 0 and 1;E’(W’,b’)a weight decay term representing all weights and biases in the model,W’andb’respectively representing a weight vector and a bias vector in the initial model;E’(W’,b’)、W’andb’are updated as the initial model learns the annotated sample.
It should be noted that the mold life prediction module in the mold maintenance prediction module is essentially a mapping function, that is, the remaining life corresponding to each wear state quantity is obtained according to the preset relationship between the preset wear state quantity and the remaining life; the training of the mold maintenance prediction module is essentially the training of the deep feed-forward network, only the parameters of the deep feed-forward network need to be iterated in the training process of the mold maintenance prediction module, and the loss is causedL’And is essentially a loss of the feed forward network.
It is noted that in the process of learning labeled samples by the neural network model to iterate parameters, a group of labeled samples are learned first, and then model loss is calculated; and if the model loss does not reach the set condition, such as minimum, continuing to learn the next group of labeled samples, and calculating the model loss again, so as to circulate. When the model loss is calculated, the loss is calculated by combining only one group of labeled samples which are learned newly at a time so as to evaluate the current accuracy degree of the neural network. Therefore, of the above first and second loss functions, the loss is calculated only from the latest set of labeled samples learned by the pre-learning/initial model.
Examples
In the present embodiment, the wear state amount is predicted for the maintenance state of a certain mold by using the mold maintenance prediction model described above, and the procedure is shown in fig. 4. In this embodiment, the predicted result is compared with the actual value. The training process of the mould maintenance prediction model is red,λthe value of the additive is 0.06,λ’the value was 0.05.
In this embodiment, when the multi-scale cyclic neural network and the mold maintenance prediction module are trained, the labeled sample is provided by historical production data of different molds, so as to improve the generalization of the mold maintenance prediction model.
In this embodiment, the pre-processing modules are differentpThe value of the wear state quantity has a direct influence on the prediction of the wear state quantity of the die, as shown in fig. 5.
In this embodiment, the mold maintenance prediction model is based on the product produced by the mold at each timeThe comparison of the product quality data of the product against the wear state quantity curve labeled for the mold and the actual wear state quantity curve is shown in fig. 6. As can be seen from FIG. 6, whenp= 0.7During the process, the trend of the wear state quantity curve marked by the model is consistent with the trend of the actual wear state quantity curve, the wear state quantity marked by the model is always slightly higher than the actual wear state quantity, and the difference value between the two is small, so that the fact that the mold maintenance prediction model is in the process of predicting the wear state quantity curve of the moldp=0.7The method has high prediction accuracy of the abrasion state quantity of the die, realizes high-accuracy prediction and safety evaluation of the abrasion state of the die, can be matched with a die life prediction module to realize high-accuracy prediction of the residual life of the die, ensures control over the production state of the die, and ensures the reliability of production.
The invention is not to be considered as limited to the specific embodiments shown and described, but is to be understood to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

Claims (10)

1. A method of training a mold maintenance prediction model for predicting a remaining life of a mold, the method comprising the steps of:
s1, obtaining product quality data, wherein the product quality data comprise product images of the same product at different angles and multiple product size data, and the product size data are size data for displaying the shape characteristics of the product; constructing a marked sample, wherein the marked sample comprises product quality data of a product produced by the die at a certain time and the residual service life of the die at the time, and enablingiIndicates that the sample number is marked,y i indicating the production of the moldiThe remaining life of the product corresponding to each label sample,y i is recorded asiTrue labeling of each labeled sample;
s2, constructing an initial model, wherein the initial model comprises a preprocessing module, a self-attention mechanism module, a multi-scale circulation neural network and a mold maintenance prediction module;
the preprocessing module is used for processingPreprocessing the product images at the same angle to obtain the feature vectors of the product images at all angles and construct a product image feature setF(ii) a The preprocessing module is used for preprocessing various product size data to obtain a product size feature setS’’(ii) a The preprocessing module is also used for collecting the product image characteristicsFAnd product size feature setS’’Weighted sum as sample dataXAnd outputting; order toX i Is shown asiSample data corresponding to each labeled sample;
F={f(1),f(2),…,f(r f ),…,f(n f )}；
S’’={s’’(1),s’’(2),…,s’’(r s ),…,s’’(n s )}；
wherein, the first and the second end of the pipe are connected with each other,n f the number of image acquisition angles is represented,f(r f )represents from the firstr f Characteristic vectors corresponding to the product images acquired at the angles;n s the number of pieces of data representing the size of the product,s’’(r s )to express the productr s A dimensional feature;n s =n f ；1≤r f ≤ n f ；1≤r s ≤n s ；
the input of the self-attention mechanism module is sample dataXOutput isXFeature vector after self-attention mechanismX’；
The multi-scale recurrent neural network comprises a plurality of recurrent neural networks, each recurrent neural network is used for extracting the feature vector on different scales, and the input of the recurrent neural network is the feature vectorX’Or an output of itself; let the current time betAt a time of dayrThe input of the recurrent neural network isX’Its output is the current time eigenvectorX’’(r,t)(ii) a First, therThe input of the recurrent neural network isX’’(r,t’)At the output of itIs the feature vector of the next timeX’’(r,t’+1)；t≤t’≤t+N-1，NIs a set number of cycles; 1≤r≤R，RIs the number of recurrent neural networks;
the mould maintenance prediction module comprises a depth feedforward network and a mould service life prediction module; the deep feedforward network is based on the feature vector at each timeX’’(r,t+n)Predict the timet+nThe amount of the wear state of the upper die,0≤n≤N(ii) a The mould life prediction module predicts the residual life of the mould according to a wear state quantity curve of the mould, and the wear state quantity curve of the mould is drawn by combining the wear state quantity of the mould at each moment;
the input of the initial model is the input of the preprocessing module, and the output of the initial model is the output of the mould life prediction module, namely the residual life of the mould;
and S3, enabling the initial model to automatically learn and label the sample to iterate the parameters until the set iteration condition is reached, fixing the parameters, and taking the initial model with the fixed parameters as a mold maintenance prediction model.
2. The training method of the mold maintenance prediction model according to claim 1, wherein the pre-processing module and the self-attention mechanism module both use fixed models, and in S3, the pre-processing module, the self-attention mechanism module and the multi-scale recurrent neural network are first made to constitute a pre-learning model, and the pre-learning model is made to autonomously learn labeled samples to perform parameter iteration on the multi-scale recurrent neural network until the multi-scale recurrent neural network is fixed; then, learning the marked sample by using an initial model with fixed multi-scale cyclic neural network parameters so as to iterate the parameters of the mold maintenance prediction module;
the parameter iteration of the multi-scale recurrent neural network comprises the following steps:
SA31, setting transition parametersX’’’Order:
X’’’=w 1 ×X’’(1,t)+w 2 ×X’’(2,t)+…+w r ×X’’(r,t)+…+w R ×X’’(R,t)；
wherein the content of the first and second substances,w r represents a weight, 1≤r≤R，w 1 +w 2 +…+w r +…+w R =1；w r Self-adaptive adjustment is carried out through back propagation in the process that the pre-learning model autonomously learns the labeled samples;
SA32, enabling the pre-learning model to learn the marked samples, and performing parameter iteration on the multi-scale cyclic neural network in the pre-learning model by adopting an Adam optimizer in the learning process until the loss of the pre-learning model reaches the minimum value to fix the multi-scale cyclic neural network;
the loss of the pre-learning model is calculated using the following first loss function:
Figure 110868DEST_PATH_IMAGE001
；
wherein, the first and the second end of the pipe are connected with each other,Lrepresenting the loss of the pre-learned model,Ma set of the most recent set of labeled samples representing pre-learning model learning,mis composed ofMThe number of the labeled samples in (1),iindicates that the sample number is marked,X i ’denotes the firstiSample data corresponding to each labeled sampleX i Feature vectors after the self-attention mechanism; sample dataX i The corresponding transition parameters are noted asX i ’’’When the input of the multi-scale circulation neural networkX’=X i ’When the temperature of the water is higher than the set temperature,X i ’’’=X’’’；λrepresenting the weight of the weighted decay term in the first loss function,λis a set value;E (W,b)a weight decay term representing all weights and biases in the multi-scale recurrent neural network in the pre-learning model,Wandband respectively representing a weight vector and a bias vector of the multi-scale recurrent neural network in the pre-learning model.
3. The method for training the mold maintenance prediction model according to claim 2, wherein the parameter iteration method of the mold maintenance prediction module is as follows: substituting the multi-scale cyclic neural network with fixed parameters into an initial model, enabling the initial model to learn a labeled sample, adopting an Adam optimizer to carry out parameter iteration on a depth feedforward network in the initial model in the learning process until the loss of the initial model reaches the minimum, fixing the parameters of the depth feedforward network, and storing the initial model at the moment as a mold maintenance prediction model;
the loss function of the initial model is calculated by adopting the following second loss function:
Figure 11828DEST_PATH_IMAGE002
；
wherein the content of the first and second substances,L’which represents the loss of the initial model and,M’a set of the most recent set of labeled samples representing initial model learning,m’is composed ofM’The number of the labeled samples in (1),iindicates the serial number of the marked sample,y i is sample dataX i The corresponding real label is marked on the display screen,ŷ i representing sample dataX i Corresponding model label when the input of the initial model isX i At that time, its output is recorded asŷ i ；λ’Representing the weight of the weighted decay term in the second loss function,λ’is a set value;E’(W’,b’)a weight decay term representing all weights and biases of the deep feed-forward network in the initial model,W’andb’respectively representing the weight vector and the bias vector of the deep feedforward network in the initial model.
4. The method of training a mold maintenance predictive model of claim 1, wherein the set of product size featuresS’’The method comprises the following steps:
SD1, acquisitionn s Calculating the error between each item of size data and the item of size data of the standard product to constructSet of dimensional errorsS={s(1),s(2),…,s(r s ),…,s(n s )}；s(r s )To express the productr s The error corresponding to the item size data;
SD2, set of dimensional errorsSNormalizing each error in the error list to obtain a normalized size error setS’={s’(1),s’(2),…,s’(r s ),…,s’(n s )}，s’(r s )∈[0,1]；
SD3, set of pairsS’Carrying out embedding treatment to obtain a product size feature vector setS’’。
5. The method of training a mold care prediction model of claim 1, wherein the set of product image featuresFThe acquisition mode is as follows: within the set shooting timen f Collecting product images at different angles, preprocessing the images, and extracting the feature vectors of the preprocessed product images at all angles through a deep convolutional neural network so as to construct a product image feature setF(ii) a The shooting time is less than or equal to 1 minute.
6. The method of claim 1, wherein the sample data in S1 is the same as the sample data in the predictive model for mold maintenance training methodXObtained according to the following formula:
X=pF+(1-p)S’’；
wherein the content of the first and second substances,p∈[0.5，0.6，0.7，0.8，0.9]。
7. the method of claim 1, wherein the self-attention mechanism module obtains sample data according to the following formulaXFeature vector ofX’；
Figure 306019DEST_PATH_IMAGE003
；
Wherein the content of the first and second substances,Q、KandVeach of which represents a matrix of the image data,Q=W q X，K=W k X，V=W v X；W q 、W k andW v a parameter matrix representing linear mapping, wherein the three are intrinsic parameters of a self-attention mechanism;D k representing a vector matrixKThe dimension (c) of (a) is,K T representing a vector matrixKTransposing;softmaxrepresenting a normalized exponential function used to calculate the probability for all classes.
8. A method of predicting mold maintenance, comprising the steps of:
st1, obtaining a mold maintenance prediction model obtained by the mold maintenance prediction model training method according to any one of claims 1 to 7;
and St2, acquiring product quality data of a product currently produced by the mold to be predicted, inputting the product quality data into a mold maintenance prediction model, and acquiring the residual life output by the mold maintenance prediction model as an evaluation index of the mold maintenance prediction model.
9. A mold maintenance prediction system comprising a memory module having a mold maintenance prediction model stored therein and a computer program that when executed is configured to implement the mold maintenance prediction method of claim 8.
10. The mold maintenance prediction system of claim 9, further comprising a processing module for executing the computer program to implement the mold maintenance prediction method of claim 8.
CN202211092921.6A 2022-09-08 2022-09-08 Mold maintenance prediction model training method, mold maintenance prediction method and system Active CN115169617B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211092921.6A CN115169617B (en) 2022-09-08 2022-09-08 Mold maintenance prediction model training method, mold maintenance prediction method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211092921.6A CN115169617B (en) 2022-09-08 2022-09-08 Mold maintenance prediction model training method, mold maintenance prediction method and system

Publications (2)

Publication Number Publication Date
CN115169617A CN115169617A (en) 2022-10-11
CN115169617B true CN115169617B (en) 2022-11-11

Family

ID=83481081

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211092921.6A Active CN115169617B (en) 2022-09-08 2022-09-08 Mold maintenance prediction model training method, mold maintenance prediction method and system

Country Status (1)

Country Link
CN (1) CN115169617B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002321032A (en) * 2001-04-27 2002-11-05 Toyota Motor Corp Metal die wear loss estimation apparatus, metal die wear loss estimation method, metal die wear loss estimation program, metal die life estimation apparatus, metal die life estimation method, metal die life estimation program, metal die wear loss detection apparatus and metal die life detection apparatus
JP2011221898A (en) * 2010-04-13 2011-11-04 Toyota Motor Corp Die wear predictor and production management system
CN103492834A (en) * 2010-09-15 2014-01-01 吴乃恩 Automatic apparatus for inspecting rotary parts and inspection method
CN110889633A (en) * 2019-11-28 2020-03-17 通鼎互联信息股份有限公司 Intelligent mold management system
CN111260212A (en) * 2020-01-14 2020-06-09 江苏古卓科技有限公司 Supply chain cooperative behavior analysis method for product design evolution
CN113469260A (en) * 2021-07-12 2021-10-01 天津理工大学 Visual description method based on convolutional neural network, attention mechanism and self-attention converter
CN113642414A (en) * 2021-07-19 2021-11-12 北京航空航天大学 Method for predicting residual service life of rolling bearing based on Transformer model

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11992952B2 (en) * 2020-10-29 2024-05-28 General Electric Company Systems and methods of servicing equipment

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002321032A (en) * 2001-04-27 2002-11-05 Toyota Motor Corp Metal die wear loss estimation apparatus, metal die wear loss estimation method, metal die wear loss estimation program, metal die life estimation apparatus, metal die life estimation method, metal die life estimation program, metal die wear loss detection apparatus and metal die life detection apparatus
JP2011221898A (en) * 2010-04-13 2011-11-04 Toyota Motor Corp Die wear predictor and production management system
CN103492834A (en) * 2010-09-15 2014-01-01 吴乃恩 Automatic apparatus for inspecting rotary parts and inspection method
CN110889633A (en) * 2019-11-28 2020-03-17 通鼎互联信息股份有限公司 Intelligent mold management system
CN111260212A (en) * 2020-01-14 2020-06-09 江苏古卓科技有限公司 Supply chain cooperative behavior analysis method for product design evolution
CN113469260A (en) * 2021-07-12 2021-10-01 天津理工大学 Visual description method based on convolutional neural network, attention mechanism and self-attention converter
CN113642414A (en) * 2021-07-19 2021-11-12 北京航空航天大学 Method for predicting residual service life of rolling bearing based on Transformer model

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于深度门控循环单元神经网络的刀具磨损状态实时监测方法;陈启鹏等;《计算机集成制造系统》;20200715(第07期);第1782-1793页 *

Also Published As

Publication number Publication date
CN115169617A (en) 2022-10-11

Similar Documents

Publication Publication Date Title
CN111127385B (en) Medical information cross-modal Hash coding learning method based on generative countermeasure network
CN109949286A (en) Method and apparatus for output information
CN113486578A (en) Method for predicting residual life of equipment in industrial process
CN113255573B (en) Pedestrian re-identification method based on mixed cluster center label learning and storage medium
Tian et al. CASDD: Automatic surface defect detection using a complementary adversarial network
CN115576267A (en) Hub machining dimension error correction method based on digital twinning
CN117237367B (en) Spiral blade thickness abrasion detection method and system based on machine vision
CN115169617B (en) Mold maintenance prediction model training method, mold maintenance prediction method and system
CN116720079A (en) Wind driven generator fault mode identification method and system based on multi-feature fusion
CN114372181B (en) Equipment production intelligent planning method based on multi-mode data
CN112763215A (en) Multi-working-condition online fault diagnosis method based on modular federal deep learning
CN117113086A (en) Energy storage unit load prediction method, system, electronic equipment and medium
CN116383747A (en) Anomaly detection method for generating countermeasure network based on multi-time scale depth convolution
CN110045691A (en) A kind of multitasking fault monitoring method of multi-source heterogeneous big data
CN116383757A (en) Bearing fault diagnosis method based on multi-scale feature fusion and migration learning
CN115456693A (en) Automatic evaluation method for automobile exterior design driven by big data
CN113158878B (en) Heterogeneous migration fault diagnosis method, system and model based on subspace
CN111401519B (en) Deep neural network unsupervised learning method based on similarity distance in object and between objects
CN114926702A (en) Small sample image classification method based on depth attention measurement
Niu et al. Learning trustworthy model from noisy labels based on rough set for surface defect detection
CN114548494B (en) Visual cost data prediction intelligent analysis system
CN117274750B (en) Knowledge distillation semi-automatic visual labeling method and system
CN117392440B (en) Textile fabric retrieval method and system based on tissue structure and color classification
CN113705729B (en) Garbage classification model modeling method, garbage classification device and medium
CN118035866A (en) Fault diagnosis method for aluminum extrusion equipment based on expert network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant