CN111415032A - Method for predicting production performance of polyester fiber precursor based on E L M-AE of transfer learning - Google Patents
Method for predicting production performance of polyester fiber precursor based on E L M-AE of transfer learning Download PDFInfo
- Publication number
- CN111415032A CN111415032A CN202010141610.9A CN202010141610A CN111415032A CN 111415032 A CN111415032 A CN 111415032A CN 202010141610 A CN202010141610 A CN 202010141610A CN 111415032 A CN111415032 A CN 111415032A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- representing
- output
- samples
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 119
- 238000013526 transfer learning Methods 0.000 title claims abstract description 74
- 239000000835 fiber Substances 0.000 title claims abstract description 59
- 229920000728 polyester Polymers 0.000 title claims abstract description 55
- 238000004519 manufacturing process Methods 0.000 title claims abstract description 42
- 239000002243 precursor Substances 0.000 title claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 11
- 238000000605 extraction Methods 0.000 claims abstract description 3
- 230000008569 process Effects 0.000 claims description 91
- 238000009987 spinning Methods 0.000 claims description 48
- 230000006870 function Effects 0.000 claims description 42
- 239000011159 matrix material Substances 0.000 claims description 41
- 238000013508 migration Methods 0.000 claims description 39
- 230000005012 migration Effects 0.000 claims description 39
- 230000004913 activation Effects 0.000 claims description 13
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 238000007664 blowing Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 239000004576 sand Substances 0.000 claims description 4
- 238000012546 transfer Methods 0.000 claims description 4
- 230000006399 behavior Effects 0.000 claims description 2
- 229920000642 polymer Polymers 0.000 description 8
- 238000010586 diagram Methods 0.000 description 2
- 239000007788 liquid Substances 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 239000000155 melt Substances 0.000 description 2
- 238000006116 polymerization reaction Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 239000004195 dipotassium inosinate Substances 0.000 description 1
- 238000010036 direct spinning Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 210000002364 input neuron Anatomy 0.000 description 1
- 238000002074 melt spinning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000004205 output neuron Anatomy 0.000 description 1
- 239000002994 raw material Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000012209 synthetic fiber Substances 0.000 description 1
- 229920002994 synthetic fiber Polymers 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
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
- G06Q10/04—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
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
-
- 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
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/067—Enterprise or organisation modelling
-
- 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
-
- 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
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Engineering & Computer Science (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Marketing (AREA)
- Development Economics (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Quality & Reliability (AREA)
- Educational Administration (AREA)
- Operations Research (AREA)
- Game Theory and Decision Science (AREA)
- Manufacturing & Machinery (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Nonwoven Fabrics (AREA)
- Feedback Control In General (AREA)
Abstract
The invention relates to a polyester fiber precursor production performance prediction method based on E L M-AE of Transfer learning, wherein an E L M network structure only has one hidden layer, the weight of input data and the bias of the hidden layer are randomly generated, and a shallow layer network has a certain deviation on data feature extraction, so that the characteristics of the E L M network structure for extracting data are deepened, and the multiple hidden layers are trained by adopting self-encoding (AE) to improve the training precision of a model.
Description
Technical Field
The invention belongs to the field of machine learning, and particularly relates to prediction of production performance of polyester fiber strands based on E L M-AE of transfer learning.
Background
The E L M neural network is an approximator of any function, the E L M neural network is a better approximator, and the model has the advantages of easiness in realization, high training speed and strong generalization capability.
An Auto Encoder (AE) is a neural network with only one hidden layer and input and output neurons with the same number of nodes. The self-encoder consists of two parts, wherein one part is an encoding part from an input layer to a hidden layer and is used for extracting the characteristics of input data; the other part is the decoding part, i.e. the hidden layer to the output layer, for reconstructing the data. The self-encoder learning is an unsupervised learning process, an output value is equal to an input value by using a back propagation algorithm, and the self-encoder learning has the functions of reducing noise of data and dimension for visualization and extracting effective data characteristics.
The conventional machine learning problem method assumes that data obeys the same distribution, however, in practical situations, due to the characteristics of difficult data acquisition, high acquisition cost, small number of acquired samples and the like, the migration learning uses data of an unused field with correlation for model learning and improves the generalization performance of the model.
Polyester fibers are synthetic fibers made by direct spinning or remelting spinning of polymers, referred to as PET fibers for short. Polyester fibers have good properties, such as high breaking strength, good elasticity, good heat resistance, wear resistance and light resistance, and stable fabric dimensions, and therefore, polyester fibers are widely used in the fields of industry, agriculture, clothing, home furnishing, and the like. The production yield of the polyester fiber is the second world in China, and is a country with high polyester fiber yield. Therefore, the quality of polyester fiber strand production is receiving attention from researchers. The production process of the polyester fiber protofilament is mainly divided into 3 processes, namely a polymer polymerization process, a melt conveying process and a spinning process. The polymerization process of the polymer is to compound raw materials at high temperature and high pressure to form the polymer in a molten state. The melt conveying part conveys the polymer in a molten state to the spinning process to prevent the polymer from physical change. The spinning process is a process of extruding a polymer in a molten state through a spinning assembly to form filamentous fibers. The melt spinning process of the polyester fiber is the most critical link in the whole production process and determines the quality of the nascent fiber. The spinning process is that the polymer in the molten state is extruded from the capillary holes of a spinneret plate by a metering pump to form liquid trickle, and then the liquid trickle is solidified into filaments by blowing cold air in the air. The polyester fibers with different specifications and varieties have different production lines, the number of spinning positions of each production line, the number of spinnerets in each spinning position and the specification of the spinnerets are different due to different product properties. The spun fiber obtained in the spinning process is fine in quality, the sensor is complex in design and layout and high in cost, and the phenomenon of broken filaments and flying filaments easily occurs in the data acquisition process. At present, because data in the polyester fiber spinning process is difficult to collect, the collection cost is high, the number of collected samples is small, and the like, a better polyester fiber spinning link production process model is difficult to establish. Therefore, the research of establishing a universal model aiming at the polyester fiber spinning process of production lines with different specifications has important significance.
Disclosure of Invention
The invention provides a polyester fiber precursor production performance prediction method based on E L M-AE of transfer learning, which is used for solving the problems that data is difficult to collect, few collected samples are obtained and the like, realizing the performance index prediction of a transfer learning model produced in the polyester fiber spinning process, improving the prediction precision of the model, reducing the production cost of the polyester fiber spinning process and improving the performance and quality of polyester fibers.
In order to achieve the aim, the invention adopts the technical scheme that the method for predicting the production performance of the polyester fiber precursor based on E L M-AE of transfer learning is characterized by comprising the following steps:
(1) establishing a data driving model;
(2) training the weights of the E L M-AE model of the transfer learning;
(3) inputting the production process characteristics of the polyester fiber protofilament to obtain the performance index of the polyester fiber;
the data-driven model is established as described above, and is an E L M data-driven model based on migration self-encoding, the structure of the E L M model is deepened through self-encoding to further extract data features, meanwhile, the source domain input data set and the target domain input data set are subjected to migration learning, and in the encoding process of the self-encoding model, the loss value calculated through the maximum mean difference method in the migration learning is added to the loss function of the self-encoding, the data-driven model is the E L M-AE model of the migration learning, which is referred to as T L-E L M-AE model for short, and the expression of the E L M-AE model target function of the migration learning is as follows:
the first two terms on the right side of the equation represent the minimum output weight of the E L M-AE model based on the transfer learning, the third term on the right side of the equation represents the depth extraction data characteristic information, and the fourth term on the right side of the equation represents the process of the transfer learning;
ETL-ELM-AEa loss function of the E L M-AE model representing the transfer learning, X represents input data of the E L M-AE model representing the transfer learning, and X ═ X { [ X ]S,xT},xSRepresenting a source domain input data set, xTRepresents the target domain input data set, ω represents the input weights of the E L M model, b represents the hidden layer bias of the E L M model,an activation function representing the hidden layer of E L M,representing the output of the E L M hidden layer, β representing the output weight of the E L M model, Y representing the output of the E L M-AE model of the transfer learning, hXInput data representing an AE model, f (-) represents an activation function of a hidden layer of the AE model, f (h)X) Represents the output of the hidden layer of the AE model, g (-) represents the activation function of the output layer of the AE model, g (f (h)X) Represents the output of the AE model output layer; MMD2(-) represents the maximum mean difference, used to calculate the difference between the source domain input data and the target domain input data;
and (3) establishing the E L M-AE model of the transfer learning by solving the loss function minimization of the E L M-AE model of the transfer learning and updating the weight of the E L M-AE model of the transfer learning.
And,
a) the source domain input data set and the target domain input data set are respectively collected from spinning process characteristics of two kinds of polyester fiber protofilament production; the spinning process characteristics of the target domain input data set comprise the spinning process characteristics of the source domain input data set, and the number of samples of each spinning process characteristic of the source domain input data set is far larger than that of each spinning process characteristic of the target domain input data set;
b) the source domain output data set and the target domain output data set are respectively collected from spinning performance indexes produced by two polyester fiber precursors; the spinning performance index of the target domain output data set is contained in the spinning performance index of the source domain output data set, and the number of samples of each spinning performance index of the source domain output data set is far larger than that of each spinning performance index of the target domain output data set;
adding the loss value calculated by the maximum mean difference method in the encoding process of the self-encoding model to the loss function of the self-encoding model, wherein the process is to establish the core content of the E L M-AE model of the transfer learning, and the aim is to calculate the minimum value of the loss function of the transfer self-encoding:
wherein E isTL-AELoss function, h, representing migration autocodeXInput data representing an AE model, f (-) represents an activation function of a hidden layer of the AE model, f (h)X) Represents the output of the hidden layer of the AE model, g (-) represents the activation function of the output layer of the AE model, g (f (h)X) Represents the output of the AE model output layer; x is the number ofSInput data, x, representing a source domainTInput data representing a target domain, MMD2(. Maximum Mean variance) is the Maximum Mean difference used to calculate the difference between the source domain input data and the target domain input data, and is calculated as follows:
wherein,representing a feature map of source domain input data samples, r representing an r-th sample point of the source domain input data, and p representing a number of samples of the source domain input data;a feature map representing target domain input samples, o represents the o-th sample point of the target domain input data, q represents the target domain inputThe number of samples of incoming data;
the updating formula of the trained weight and bias term is as follows:
wherein,the representation is a weight value that is migrated from the encoded update,indicating that the migration is from the old weight value of the code,a bias value representing a migrated self-encoded update,offset value representing migration since coding, αTL-AEFor migrating the learning rate of the self-coding, ETL-AEA loss function representing the migrated self-encoding,andthe partial derivatives of the loss function of the migrated self-encodings against weight and bias are shown separately.
As a preferred technical scheme:
the method for predicting the production performance of the polyester fiber precursor based on the E L M-AE of the transfer learning is characterized in that the E L M-AE model of the transfer learning is a neural network with a structure larger than 3 layers and comprises an input layer, an H layer hidden layer and an output layer, wherein the value of H is larger than or equal to 2.
The method for predicting the production performance of the polyester fiber strand based on E L M-AE of transfer learning is characterized in that the data set is specifically established as follows:
i, establishing a source domain input data set and a target domain input data set:
the input sample data collected from different production specification batches, namely the process characteristics of the polyester fiber spinning process, form a total input sample data set X ═ XS,xT},xSSource domain input dataset, x, for represented transfer learningTRepresenting a target domain input data set of transfer learning; a set of source domain input data samples isA set of target domain input data samples ism is the number of process features; thus, the total input sample data set X of N sets of inputs is ═ X1,x2,…,xN]N is p + q, p is the number of source domain samples, and q is the number of target domain samples; the number p of the source domain samples of the transfer learning is far larger than the number q of the target domain samples, and the matrix form of the data set is as follows:
each row of the matrix is a process characteristic, and m process characteristics are shared; each column of the matrix is a group of input data, and N groups of input data are provided in total;a value representing the type 1 process characteristic of the type 1 data of the source domain input sample,a value representing the mth process specific of the set 1 of data of the source domain input samples,a value representing a type 1 process characteristic of the p-th set of data of the source domain input samples,a value representing the mth process feature of the pth set of data of the source domain input sample;a value representing the type 1 process characteristic of the type 1 data of the target field input sample,a value representing the mth process feature of the set 1 data of the target field input sample,a value representing the type 1 process characteristic of the qth set of data of the target field input sample,a value representing the mth process feature of the qth set of data of the target field input sample;
standardizing a total input sample data set matrix X, and performing normalization on X in the input data set XSAnd xTRespectively standardizing, wherein a conversion formula of the data standardization treatment is as follows:
wherein i is 1, 2, …, m, r is 1, 2, …, p, o is 1, 2, …, q, m is the number of process features, r is the number of input samples of the source domain, and o is the number of input samples of the target domain;
inputting the numerical value of the ith process characteristic of the r group of data of the sample data set for the source domain;
inputting the standard deviation of the ith process characteristic of the sample data set for a source domain;
inputting numerical values of ith process characteristics of the r group of data standardized by the sample data set for a source domain;
inputting numerical values of ith process characteristics of the data of the No. data of the sample data set for the target domain;
inputting the standard deviation of the ith process characteristic of the sample data set for a target domain;
inputting numerical values of ith process characteristics of the data set of the No. o group after the standardization of the sample data set for the target domain;
normalized total input sample dataset matrix, i.e. eigenvector matrix X':
each row of the matrix has one process characteristic, and the total number of the process characteristics is m; each column of the matrix is a set of input dataTotal N ═ p + q sets of input data;normalized values representing the set 1 data type 1 process features of the source domain input samples,normalized values representing the mth process feature of the set 1 data of source domain input samples,normalized values representing the p-th set of data of type 1 process features of the source domain input samples,normalized values representing the mth set of data m process features of the source domain input sample;normalized values representing the set 1 data type 1 process features of the target field input sample,normalized values representing the mth process feature of the set 1 data of the target field input sample,normalized values representing the qth set of data of type 1 process features of the target field input sample,normalized values representing the mth set of data m process features of the target domain input sample;
II, establishing a source domain output data set and a target domain output data set:
the performance index of the polyester fiber, which is the output sample data collected in different production specification batches, is used for forming a total output sample data set Y={yS,yT},ySSource domain output dataset, y, representing transfer learningTRepresenting a target domain output data set of the transfer learning; a set of source domain output data samples isA set of target domain output data samples isl is the number of performance indexes; thus, N sets of outputs constitute a total set of output sample data Y ═ Y1,y2,…,yN]N is p + q, p is the number of source domain samples, and q is the number of target domain samples. The number p of the source domain samples of the transfer learning is far larger than the number q of the target domain samples, and the matrix form of the data set is as follows:
each row of the matrix has one performance index, the total performance index is l, each column of the matrix is provided with one group of output data, and the total number of the output data is N groups of output data;a value representing the 1 st performance indicator of the 1 st set of data of source domain output samples,a value representing the l performance indicator of the set 1 data of source domain output samples,a value representing the 1 st performance indicator of the p-th set of data of source domain output samples,a value representing a first performance indicator of a p-th set of data of source domain output samples;a value representing the 1 st performance indicator of the 1 st set of data of the target domain output samples,a value representing the l performance indicator of the set 1 data of target domain output samples,a value representing the 1 st performance indicator of the qth set of data of the target domain output samples,a value representing the l performance indicator for the q-th set of data of the target domain output sample;
standardizing the total output sample data set matrix Y, and performing Y in the output data set YSAnd yTThe conversion formula for respectively normalizing the data normalization processes is:
wherein i is 1, 2, …, l, r is 1, 2, …, p, o is 1, 2, …, q, l are numbers of performance indexes, r is number of source domain output samples, and o is number of target domain output samples;
outputting the numerical value of the ith performance index of the r group of data of the sample data set for the source domain;
outputting the standard deviation of the ith performance index of the sample data set for a source domain;
outputting the numerical value of the ith performance index of the r group of data after the sample data set is standardized for the source domain;
outputting the numerical value of the ith data performance index of the sample data set for the target domain;
outputting the standard deviation of the ith performance index of the sample data set for a target domain;
outputting the numerical value of the ith performance index of the group o data after the sample data set is standardized for the target domain;
the normalized total output sample data set matrix, i.e., the performance index matrix Y':
one performance index is used for each behavior of the matrix, and the total performance index is l; each column of the matrix is a group of input data, and N is p + q groups of output data;normalized values for the 1 st data performance index of the 1 st data set representing source domain output samples,representing source domain output samplesNormalized values for the first performance index for group 1 data,normalized values for the 1 st performance index of the p-th set of data representing source domain output samples,a normalized value representing a set p of data l performance indicators of the source domain output sample;normalized values for the 1 st data performance index of the 1 st data set representing the target domain output sample,normalized values for the l performance indicators of the set 1 data representing target domain output samples,normalized values for the 1 st performance index of the qth set of data representing the target domain output samples,normalized values representing the data l performance index of the q-th set of data of the target domain output samples.
The method for predicting the production performance of the polyester fiber precursors based on the E L M-AE based on the transfer learning is characterized in that the weights of the E L M-AE model based on the transfer learning are specifically that the process feature matrix X 'is used as an input sample matrix of the E L M-AE model based on the transfer learning, the output sample data set matrix Y' is used as an output sample matrix of the E L M-AE model based on the transfer learning, the process feature matrix is applied to the E L M-AE model based on the transfer learning, a training sample input data set is input to a first layer of the E L M-AE model based on the transfer learning, the link weights of each layer of the E L M-AE model based on the transfer learning are in a full-connection mode, the E L M-AE model based on the transfer learning has H hidden layers, the weights of a front H-1 layer of the model are obtained by transfer self-coding training and are used for deeply extracting the features of the training data and reducing the differences among different domains, and the weights of the H layer and the weights of the original E L M-AE model based on the transfer learning guarantee method.
The method for predicting the production performance of the polyester fiber precursor based on E L M-AE of transfer learning is characterized in that the number N of input sample groups is 200-1000, wherein N is p + q, p represents the number of samples of source domain data, q represents the number of target domain samples, the number p of the source domain data samples is far larger than the number of the target domain samples q, the number M of process characteristics in the spinning process is 1-8, and the number l of different performance indexes affecting the quality of the polyester fiber is 1-6.
The method for predicting the production performance of the polyester fiber precursor based on E L M-AE of transfer learning is characterized in that the number M of process characteristics of the spinning process is 4, namely the spinning speed, the spinning temperature, the blowing speed and the blowing temperature, and the number l of different performance indexes influencing the quality of the polyester fiber is 4, namely the half-time elongation (EYS 1.5.5), the elongation unevenness (EYSCV), the breaking strength (DT) and the elongation capability (DE).
The method for predicting the production performance of the polyester fiber precursor based on the E L M-AE obtained by the transfer learning is characterized in that a multilayer AE model is adopted for extracting the data characteristics of the E L M-AE model obtained by the transfer learning, the AE model calculates the connection weight for adjusting each layer of the network through an inverse error, and the aim is to make the output data of the AE model closer to the input data, namely minimize the loss function error of the AE model:
wherein E isAELoss function, X, representing input data and output data of AE modelAEInput data representing an AE model, f (-) is a hidden layer activation function of the AE model, g (-) is an output layer activation function of the AE model, WinAnd WoutConnecting weights for input and output layers of the AE model, binAnd boutInput and output layer bias terms for the AE model, HAEThe output representing the hidden layer of the AE model, i.e. the result of the encoding process of the AE model,represents the output of the AE model output layer, i.e., the result of the decoding process of the AE model;
by solving the loss function of the AE model, the updating formula of the weight and the bias term of the AE model is as follows:
wherein,the representation is a weight of the AE model update,the old weight of the AE model is represented,the offset representing the update of the AE model,representing the old offset of the AE model, αAEAs the learning rate of the AE model,andthe partial derivatives of the loss function of the AE model with respect to weight and bias are shown, respectively.
Drawings
FIG. 1 is a structural view of a polyester fiber strand productivity prediction method based on E L M-AE of migration learning;
FIG. 2 is a diagram of the AE neural network architecture;
FIG. 3 is a diagram of a neural network architecture for transfer learning E L M-AE;
FIG. 4 is a comparison of the actual value of the elongation at half-maximum (EYS 1.5.5) to the predicted result;
FIG. 5 is a comparison of the real values of the elongation unevenness (EYSCV) with the predicted results;
FIG. 6 is a comparison of the actual value of the breaking strength (DT) with the predicted result;
FIG. 7 is a comparison of the real value of the elongation ability (DE) with the predicted result.
Detailed Description
The invention will be further illustrated with reference to specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the present invention. Further, it should be understood that various changes or modifications of the present invention may be made by those skilled in the art after reading the teaching of the present invention, and such equivalents may fall within the scope of the present invention as defined in the appended claims.
Example 1
The method comprises the steps of collecting process characteristic data of four spinning processes including spinning temperature, spinning speed, blowing temperature and blowing speed of polyester fiber spinning as input of an E L M-AE model of migration learning, collecting performance indexes such as half-elongation (EYS.5), elongation-unevenness (EYSCV), breaking strength (DT) and elongation ability (DE) of the polyester fiber as output of an E L M-AE model of migration learning, collecting process characteristic data of two production lines A and B of different specifications of the polyester fiber in the production process of the polyester fiber, wherein the quantity of each input data and each output data collected by the production line A is 216, wherein the quantity of the data is shown as a source domain input data for establishing an E L M-AE model of migration learning, the data is shown as a source domain input data of the production line A632 and a source domain output data of the production line B is shown as an E L M-AE model of migration learning, the data is shown as a source domain input data of the production line A632 and the source domain output data of the E L M630M 6313M 6312, the source domain input data and the E632 as an E domain input data and the E3 as an E equivalent to the migration learning model, the migration learning result of migration learning model is shown as an E equivalent to a migration learning model, the migration learning model is shown as an E equivalent to a migration learning model of migration learning model, the migration learning model of migration learning model is shown as an E equivalent to the migration learning model, the migration learning model of migration learning model, the migration learning model is shown as an E equivalent to the migration learning model, the migration learning model of migration learning model, the migration learning model is shown as an E equivalent to the migration learning model, the migration learning model is shown as an E equivalent to the migration learning model, the migration learning model is shown as an equivalent to the migration learning model.
TABLE 1 number of samples collected for each process feature sample point of the input data set
TABLE 2 number of samples collected per performance index of the output data set
Table 3 number of acquisitions per sample point of the validation data set
Claims (7)
1. The polyester fiber protofilament production performance prediction method based on E L M-AE of transfer learning is characterized by comprising the steps of (1) establishing a data-driven model, (2) training the weight of the E L M-AE model of transfer learning, (3) inputting polyester fiber protofilament production process characteristics to obtain polyester fiber performance indexes;
the data driving model is established by establishing an E L M-AE data driving model based on transfer learning, the structure of the E L M model is deepened through self-coding to further extract data characteristics, meanwhile, transfer learning is carried out on a source domain input data set and a target domain input data set, in the coding process of the self-coding model, a loss value calculated through a maximum mean difference method in the transfer learning is added to a self-coding loss function, the data driving model is the E L M-AE model of the transfer learning, which is called T L-E L M-AE model for short, and the expression of the E L M-AE model target function of the transfer learning is as follows:
the first two terms on the right side of the equation represent the minimum output weight of the E L M-AE model based on the transfer learning, the third term on the right side of the equation represents the depth extraction data characteristic information, and the fourth term on the right side of the equation represents the process of the transfer learning;
ETL-ELM-AEa loss function of the E L M-AE model representing the transfer learning, X represents input data of the E L M-AE model representing the transfer learning, and X ═ X { [ X ]S,xT},xSRepresenting a source domain input data set, xTRepresents the target domain input data set, ω represents the input weights of the E L M model, b represents the hidden layer bias of the E L M model,an activation function representing the hidden layer of E L M,representing the output of the E L M hidden layer, β representing the output weight of the E L M model, Y representing the output of the E L M-AE model of the transfer learning, hXInput data representing an AE model, f (-) represents an activation function of a hidden layer of the AE model, f (h)X) Watch (A)Showing the output of the hidden layer of the AE model, g (-) showing the activation function of the output layer of the AE model, g (f (h)X) Represents the output of the AE model output layer; MMD2(-) represents the maximum mean difference, used to calculate the difference between the source domain input data and the target domain input data;
the method comprises the steps of establishing a transfer learning E L M-AE model by solving the loss function minimization of the transfer learning E L M-AE model and updating the weight of the transfer learning E L M-AE model;
and,
a) the source domain input data set and the target domain input data set are respectively collected from spinning process characteristics of two kinds of polyester fiber protofilament production; the spinning process characteristics of the target domain input data set comprise the spinning process characteristics of the source domain input data set, and the number of samples of each spinning process characteristic of the source domain input data set is far larger than that of each spinning process characteristic of the target domain input data set;
b) the source domain output data set and the target domain output data set are respectively collected from spinning performance indexes produced by two polyester fiber precursors; the spinning performance index of the target domain output data set is contained in the spinning performance index of the source domain output data set, and the number of samples of each spinning performance index of the source domain output data set is far larger than that of each spinning performance index of the target domain output data set;
adding the loss value calculated by the maximum mean difference method in the encoding process of the self-encoding model to the loss function of the self-encoding model, wherein the process is to establish the core content of the E L M-AE model of the transfer learning, and the aim is to calculate the minimum value of the loss function of the transfer self-encoding:
wherein E isTL-AELoss function, h, representing migration autocodeXInput data representing an AE model, f (-) represents an activation function of a hidden layer of the AE model, f (h)X) Represents the output of the hidden layer of the AE model, and g (-) represents the AE modelActivation function of the output layer, g (h)X) Represents the output of the AE model output layer; x is the number ofSInput data, x, representing a source domainTInput data representing a target domain, MMD2(. h) is the maximum mean difference, which is used to calculate the difference between the source domain input data and the target domain input data, and is calculated as follows:
wherein,representing a feature map of source domain input data samples, r representing an r-th sample point of the source domain input data, and p representing a number of samples of the source domain input data;representing a feature map of target domain input samples, o representing an o-th sample point of the target domain input data, q representing a number of samples of the target domain input data;
the updating formula of the trained weight and bias term is as follows:
wherein,the representation is a weight value that is migrated from the encoded update,indicating that the migration is from the old weight value of the code,a bias value representing a migrated self-encoded update,offset value representing migration since coding, αTL-AEFor migrating the learning rate of the self-coding, ETL-AEA loss function representing the migrated self-encoding,andthe partial derivatives of the loss function of the migrated self-encodings against weight and bias are shown separately.
2. The method for predicting the production performance of the polyester fiber precursors based on the transfer-learning E L M-AE of claim 1, wherein the transfer-learning E L M-AE model is a neural network with a structure larger than 3 layers and comprises an input layer, an H hidden layer and an output layer, and the value of H is larger than or equal to 2.
3. The method for predicting the production performance of polyester fiber strands based on E L M-AE obtained by transfer learning according to claim 1, wherein the data set is created by:
i, establishing a source domain input data set and a target domain input data set:
the input sample data collected from different production specification batches, namely the process characteristics of the polyester fiber spinning process, form a total input sample data set X ═ XS,xT},xSSource domain input dataset, x, for represented transfer learningTRepresenting a target domain input data set of transfer learning; a set of source domain input data samples isA set of target domain input data samples ism is the number of process features; thus, the total input sample data set X of N sets of inputs is ═ X1,x2,…,xN]N is p + q, p is the number of source domain samples, and q is the number of target domain samples; the number p of the source domain samples of the transfer learning is far larger than the number q of the target domain samples, and the matrix form of the data set is as follows:
each row of the matrix is a process characteristic, and m process characteristics are shared; each column of the matrix is a group of input data, and N groups of input data are provided in total;a value representing the type 1 process characteristic of the type 1 data of the source domain input sample,a value representing the mth process specific of the set 1 of data of the source domain input samples,a value representing a type 1 process characteristic of the p-th set of data of the source domain input samples,a value representing the mth process feature of the pth set of data of the source domain input sample;a value representing the type 1 process characteristic of the type 1 data of the target field input sample,data 1 of the group representing target field input samplesThe values of the m process characteristics are,a value representing the type 1 process characteristic of the qth set of data of the target field input sample,a value representing the mth process feature of the qth set of data of the target field input sample;
standardizing a total input sample data set matrix X, and performing normalization on X in the input data set XSAnd xTRespectively standardizing, wherein a conversion formula of the data standardization treatment is as follows:
wherein i is 1, 2, …, m, r is 1, 2, …, p, o is 1, 2, …, q, m is the number of process features, r is the number of input samples of the source domain, and o is the number of input samples of the target domain;
inputting the numerical value of the ith process characteristic of the r group of data of the sample data set for the source domain;
inputting the standard deviation of the ith process characteristic of the sample data set for a source domain;
input of ith process characteristics of data set of r group normalized by sample data set for source domainA numerical value;
inputting numerical values of ith process characteristics of the data of the No. data of the sample data set for the target domain;
inputting the standard deviation of the ith process characteristic of the sample data set for a target domain;
inputting numerical values of ith process characteristics of the data set of the No. o group after the standardization of the sample data set for the target domain;
normalized total input sample dataset matrix, i.e. eigenvector matrix X':
each row of the matrix has one process characteristic, and the total number of the process characteristics is m; each column of the matrix is a group of input data, and N is p + q groups of input data;normalized values representing the set 1 data type 1 process features of the source domain input samples,normalized values representing the mth process feature of the set 1 data of source domain input samples,normalized values representing the p-th set of data of type 1 process features of the source domain input samples,normalized values representing the mth set of data m process features of the source domain input sample;normalized values representing the set 1 data type 1 process features of the target field input sample,normalized values representing the mth process feature of the set 1 data of the target field input sample,normalized values representing the qth set of data of type 1 process features of the target field input sample,normalized values representing the mth set of data m process features of the target domain input sample;
II, establishing a source domain output data set and a target domain output data set:
the method comprises the steps of collecting output sample data of different production specification batches, namely the performance index of the polyester fiber, and forming a total output sample data set of Y ═ YS,yT},ySSource domain output dataset, y, representing transfer learningTRepresenting a target domain output data set of the transfer learning; a set of source domain output data samples isA set of target domain output data samples isl is the number of performance indexes; thus, N sets of outputs constitute a total set of output sample data Y ═ Y1,y2,…,yN]N is p + q, p is the number of source domain samples, and q is the number of target domain samples; the number p of the source domain samples of the transfer learning is far larger than the number q of the target domain samples, and the matrix form of the data set is as follows:
each row of the matrix has one performance index, the total performance index is l, each column of the matrix is provided with one group of output data, and the total number of the output data is N groups of output data;a value representing the 1 st performance indicator of the 1 st set of data of source domain output samples,a value representing the l performance indicator of the set 1 data of source domain output samples,a value representing the 1 st performance indicator of the p-th set of data of source domain output samples,a value representing a first performance indicator of a p-th set of data of source domain output samples;a value representing the 1 st performance indicator of the 1 st set of data of the target domain output samples,a value representing the l performance indicator of the set 1 data of target domain output samples,a value representing the 1 st performance indicator of the qth set of data of the target domain output samples,a value representing the l performance indicator for the q-th set of data of the target domain output sample;
standardizing the total output sample data set matrix Y, and performing Y in the output data set YSAnd yTThe conversion formula for respectively normalizing the data normalization processes is:
wherein i is 1, 2, …, l, r is 1, 2, …, p, o is 1, 2, …, q, l are numbers of performance indexes, r is number of source domain output samples, and o is number of target domain output samples;
outputting the numerical value of the ith performance index of the r group of data of the sample data set for the source domain;
outputting the standard deviation of the ith performance index of the sample data set for a source domain;
outputting the numerical value of the ith performance index of the r group of data after the sample data set is standardized for the source domain;
outputting the numerical value of the ith data performance index of the sample data set for the target domain;
outputting the standard deviation of the ith performance index of the sample data set for a target domain;
outputting the numerical value of the ith performance index of the group o data after the sample data set is standardized for the target domain;
the normalized total output sample data set matrix, i.e., the performance index matrix Y':
one performance index is used for each behavior of the matrix, and the total performance index is l; each column of the matrix is a group of input data, and N is p + q groups of output data;normalized values for the 1 st data performance index of the 1 st data set representing source domain output samples,a normalized value of the l performance indicator for the set 1 data representing source domain output samples,representing source domain output samplesThe normalized values of the 1 st performance index of the p-th group of data,a normalized value representing a set p of data l performance indicators of the source domain output sample;normalized values for the 1 st data performance index of the 1 st data set representing the target domain output sample,normalized values for the l performance indicators of the set 1 data representing target domain output samples,normalized values for the 1 st performance index of the qth set of data representing the target domain output samples,normalized values representing the data l performance index of the q-th set of data of the target domain output samples.
4. The method for predicting the production performance of polyester fiber strands based on E L M-AE in transfer learning of claim 3, wherein the weights for training the E L M-AE model in transfer learning are specifically that the process feature matrix X 'is used as an input sample matrix of the E L M-AE model in transfer learning, the output sample data set matrix Y' is used as an output sample matrix of the E L M-AE model in transfer learning, the E L M-AE model in transfer learning is applied, a training sample input data set is input to a first layer of the E L M-AE model in transfer learning, the link weights of each layer of the E L M-AE model in transfer learning are in a fully connected mode, the E L M-AE model in transfer learning shares H hidden layers, the weights of the front H-1 layer of the model are obtained by transfer self-coding training and are used for deeply extracting the features of the training data and reducing the performance difference between different domains, and the weight difference of the H layer and the output weight layer are calculated by the original E L M-AE model.
5. The method for predicting the production performance of polyester fiber precursors based on E L M-AE of transfer learning of claim 3, wherein the number N of input sample groups is 200-1000, wherein N is p + q, p represents the number of samples of source domain data, q represents the number of target domain samples, the number p of source domain data samples is far greater than the number of target domain samples q, the number M of process characteristics in the spinning process is 1-8, and the number l of different performance indexes affecting the quality of polyester fibers is 1-6.
6. The method for predicting the production performance of polyester fiber precursors based on E L M-AE in claim 5, wherein the number of process characteristics M of the spinning process is 4, which are the spinning speed, the spinning temperature, the blowing speed and the blowing temperature, respectively, and the number of different performance indexes l affecting the quality of polyester fibers is 4, which are the elongation at half maximum, the elongation at break, the breaking strength and the elongation ability, respectively.
7. The method for predicting production performance of polyester fiber precursors based on E L M-AE in claim 1, wherein the data characteristics of E L M-AE model are extracted by using a multi-layer AE model, the AE model calculates the connection weight for adjusting each layer by inverse error, and the aim is to make the output data of the AE model closer to the input data, namely the AE model loss function error is minimized:
wherein E isAELoss function, X, representing input data and output data of AE modelAEDenotes the AE modelType I data, f (-) is the hidden layer activation function of the AE model, g (-) is the output layer activation function of the AE model, WinAnd WoutConnecting weights for input and output layers of the AE model, binAnd boutInput and output layer bias terms for the AE model, HAEThe output representing the hidden layer of the AE model, i.e. the result of the encoding process of the AE model,represents the output of the AE model output layer, i.e., the result of the decoding process of the AE model;
by solving the loss function of the AE model, the updating formula of the weight and the bias term of the AE model is as follows:
wherein,the representation is a weight of the AE model update,the old weight of the AE model is represented,the offset representing the update of the AE model,representing the old offset of the AE model, αAEAs the learning rate of the AE model,andthe partial derivatives of the loss function of the AE model with respect to weight and bias are shown, respectively.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010141610.9A CN111415032B (en) | 2020-03-03 | 2020-03-03 | Method for predicting production performance of polyester fiber protofilament based on ELM-AE of transfer learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010141610.9A CN111415032B (en) | 2020-03-03 | 2020-03-03 | Method for predicting production performance of polyester fiber protofilament based on ELM-AE of transfer learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111415032A true CN111415032A (en) | 2020-07-14 |
CN111415032B CN111415032B (en) | 2022-04-29 |
Family
ID=71491157
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010141610.9A Active CN111415032B (en) | 2020-03-03 | 2020-03-03 | Method for predicting production performance of polyester fiber protofilament based on ELM-AE of transfer learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111415032B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113283481A (en) * | 2021-05-14 | 2021-08-20 | 群智未来人工智能科技研究院(无锡)有限公司 | Intelligent membrane pollution decision-making method based on knowledge type-two fuzzy |
CN115730734A (en) * | 2022-11-29 | 2023-03-03 | 广东工业大学 | Production line and equipment prediction method based on migration component regression |
CN117932232A (en) * | 2024-03-21 | 2024-04-26 | 南京信息工程大学 | Wind speed prediction system based on state identification RIME-DLEM multivariable time sequence prediction |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108152239A (en) * | 2017-12-13 | 2018-06-12 | 东北大学秦皇岛分校 | The sample composition content assaying method of feature based migration |
CN109060001A (en) * | 2018-05-29 | 2018-12-21 | 浙江工业大学 | A kind of multiple operating modes process soft-measuring modeling method based on feature transfer learning |
CN109787236A (en) * | 2019-01-28 | 2019-05-21 | 云南电网有限责任公司 | A kind of power system frequency Tendency Prediction method based on deep learning |
CN109858509A (en) * | 2018-11-05 | 2019-06-07 | 杭州电子科技大学 | Based on multilayer stochastic neural net single classifier method for detecting abnormality |
-
2020
- 2020-03-03 CN CN202010141610.9A patent/CN111415032B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108152239A (en) * | 2017-12-13 | 2018-06-12 | 东北大学秦皇岛分校 | The sample composition content assaying method of feature based migration |
CN109060001A (en) * | 2018-05-29 | 2018-12-21 | 浙江工业大学 | A kind of multiple operating modes process soft-measuring modeling method based on feature transfer learning |
CN109858509A (en) * | 2018-11-05 | 2019-06-07 | 杭州电子科技大学 | Based on multilayer stochastic neural net single classifier method for detecting abnormality |
CN109787236A (en) * | 2019-01-28 | 2019-05-21 | 云南电网有限责任公司 | A kind of power system frequency Tendency Prediction method based on deep learning |
Non-Patent Citations (2)
Title |
---|
JINXI ZHANG 等: "A Prediction Method Using Extreme Learning Machine with Immune Optimization", 《2017 11TH ASIAN CONTROL CONFERENCE (ASCC)》 * |
邓万宇 等: "基于ELM-AE的迁移学习算法", 《计算机与数字工程》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113283481A (en) * | 2021-05-14 | 2021-08-20 | 群智未来人工智能科技研究院(无锡)有限公司 | Intelligent membrane pollution decision-making method based on knowledge type-two fuzzy |
CN115730734A (en) * | 2022-11-29 | 2023-03-03 | 广东工业大学 | Production line and equipment prediction method based on migration component regression |
CN115730734B (en) * | 2022-11-29 | 2023-08-08 | 广东工业大学 | Production line and equipment prediction method based on migration component regression |
CN117932232A (en) * | 2024-03-21 | 2024-04-26 | 南京信息工程大学 | Wind speed prediction system based on state identification RIME-DLEM multivariable time sequence prediction |
CN117932232B (en) * | 2024-03-21 | 2024-05-28 | 南京信息工程大学 | Wind speed prediction system based on state identification RIME-DELM multivariable time sequence prediction |
Also Published As
Publication number | Publication date |
---|---|
CN111415032B (en) | 2022-04-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111415032B (en) | Method for predicting production performance of polyester fiber protofilament based on ELM-AE of transfer learning | |
CN107180155B (en) | A kind of disease forecasting system based on Manufacturing resource model | |
CN105975488B (en) | A kind of keyword query method based on theme class cluster unit in relational database | |
WO2024060381A1 (en) | Incremental device fault diagnosis method | |
Spiliopoulou et al. | Higher order mining: Modelling and mining the results of knowledge discovery | |
CN117407744B (en) | Multi-source data fusion method based on digital twin | |
CN118245523B (en) | Spinning optimization method and system of melt injection system | |
CN116167640B (en) | LCP film production quality detection data analysis method and system | |
Gillette et al. | Topological characterization of neuronal arbor morphology via sequence representation: II-global alignment | |
CN113239211A (en) | Reinforced learning knowledge graph reasoning method based on course learning | |
CN111048190A (en) | DRG grouping method based on artificial intelligence | |
CN113112796A (en) | Construction method and system of driving behavior characteristics and driving behavior analysis method | |
CN110331197A (en) | Application of the lncRNA in the product of preparation prediction Head and neck squamous cell carcinoma prognosis | |
CN110263380B (en) | Spinning process cascade modeling subsection interval parameter configuration method | |
CN115238962A (en) | Multi-production-line time sequence prediction method for performance indexes of polyester fibers in esterification stage | |
WO2014157750A1 (en) | Apparatus and method for providing causative factors for state of quality of effluent water from sewage treatment plant | |
CN106599431A (en) | FDY (fully drawn yarn) spinning technology for optimizing multi-model method based on mixture Gaussian weighting function | |
CN111137295B (en) | Driving tendency dynamic transition probability calculation method considering sad emotion | |
Ferriss | Does material well-being affect non-material well-being? | |
CN113377991A (en) | Image retrieval method based on most difficult positive and negative samples | |
CN113191397A (en) | Multi-dimensional signal feature fusion method based on maximum correlation entropy criterion | |
Hui Han et al. | Prediction method of carding process production quality based on digital twin technology | |
Bishop et al. | Comparison of quality-control rules used in clinical chemistry laboratories | |
CN115116616A (en) | Intra-group optimization based multiple interpolation breast cancer deletion data interpolation model | |
Shenoy et al. | Using poly (ethylene terephthalate) melt spinning simulation for process optimization |
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 |