CN112215351B - Enhanced multi-scale convolution neural network soft measurement method - Google Patents

Enhanced multi-scale convolution neural network soft measurement method Download PDF

Info

Publication number
CN112215351B
CN112215351B CN202010994589.7A CN202010994589A CN112215351B CN 112215351 B CN112215351 B CN 112215351B CN 202010994589 A CN202010994589 A CN 202010994589A CN 112215351 B CN112215351 B CN 112215351B
Authority
CN
China
Prior art keywords
data
sample
convolution
dimensional
neural network
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
CN202010994589.7A
Other languages
Chinese (zh)
Other versions
CN112215351A (en
Inventor
葛志强
江肖禹
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.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
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 Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202010994589.7A priority Critical patent/CN112215351B/en
Publication of CN112215351A publication Critical patent/CN112215351A/en
Application granted granted Critical
Publication of CN112215351B publication Critical patent/CN112215351B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/903—Querying
    • G06F16/90335—Query processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Computational Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Algebra (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an enhanced multi-scale convolution neural network soft measurement method, which carries out data enhancement by reconstructing coarse-grained characteristics and applies a convolution neural network to construct a regression model. Splicing process variable data of adjacent time periods to obtain a data matrix, replacing original fine-grained characteristics with new input of a sample, and solving the problems of unbalanced process variable sampling and inaccurate input and output correspondence of the sample; expanding the sample by using a generation countermeasure network, and performing data enhancement on the original data set by using a quantitative generated sample to solve the problem of local data loss in the process; the coarse-grained features are simultaneously extracted by using a one-dimensional convolution kernel and a two-dimensional convolution kernel, the dynamic property of variables and the time difference between the variables are considered, and the problem of time delay between process variables is solved; and reducing the data to one dimension through the alternative operation of convolution and pooling, and establishing a prediction model through a multilayer perceptron to realize real-time soft measurement of the target quality variable.

Description

Enhanced multi-scale convolution neural network soft measurement method
Technical Field
The invention belongs to the field of industrial process soft measurement, and particularly relates to an enhanced multi-scale convolution neural network soft measurement method.
Background
In the industry big data age, data-driven soft sensors have become an important means for guiding industrial production and control. However, in the actual industry, the means of acquiring different data are different, most process variables in production can be directly measured by a Distributed Control System (DCS), and some quality variables need to be analyzed by chemical means or specialized instruments to obtain values. Meanwhile, the process variable data and the quality variable data are input values and output values of one sample, respectively. The quality variable data of an industrial product is much less than the process variable data due to its high measurement costs. Therefore, the problem of unbalanced sampling of different variable data breaks the basic premise that the training sample needs to have an input value and an output value simultaneously in the traditional soft sensor modeling, and a large amount of process variable data is wasted.
In addition, in the data measurement process, because points distributed in production equipment by the DCS are different, data collected at the same time often do not correspond to the same batch of materials, and the problem of time delay among variable data is caused. Meanwhile, the time difference between the sampling of the quality variable data and the sampling of the process variable data also causes the problem that the input value and the output value of the sample are inaccurate correspondingly. Both of these problems also greatly limit the performance of soft sensors.
Based on the above three problems, the fine-grained characteristics (process variable data at a certain moment) cannot provide effective process information for soft sensor modeling, and thus a good soft measurement effect cannot be obtained. Meanwhile, when it is considered that the absolute amount of the quality variable data is small and a local loss is caused for some reason, the model is further badly influenced. Therefore, in industrial practice, a soft measurement method that can be used to specifically solve the above practical problems is very valuable.
Disclosure of Invention
The invention aims to provide an enhanced multi-scale convolution neural network soft measurement method aiming at the defects of the prior art. The invention realizes the prediction of quality variable data in industry based on the generation of a confrontation network and a Convolutional Neural Network (CNN) structure.
The purpose of the invention is realized by the following technical scheme: an enhanced multi-scale convolution neural network soft measurement method comprises the following steps:
step one, data preprocessing: let the mass variable be Y and the process variable be XiI is 1,2, …, m; where m is the number of process variables. Taking the quality variable data collected at the moment t as a sample StOutput value y oftAnd t is 1,2, …. Setting the duration of the material in the whole process as n, and dividing [ t-n, t]Process variable x collected over a period of timei·jJ-1, 2, …, n, spliced together as a sample StInput value of
Figure GDA0003537191810000021
Referred to as coarse-grained features. Then, the process variables are normalized by the maximum and minimum normalization to eliminate dimensional differences between the variables.
Step two, data enhancement: pairing samples S with a generating countermeasure networktCarrying out data enhancement to obtain a new generated sample S*。
Step three, feature extraction: the convolution operation employs both a 1 × U one-dimensional convolution kernel and a V × W two-dimensional convolution kernelConvolution kernel, for sample StAnd sample S*And performing feature extraction on the coarse granularity features after the normalization. Then, the feature maps extracted by the convolution kernels at different scales are fused through convolution of 1x 1. And then, performing dimension reduction processing on the extracted feature map by using a pooling layer. And completely converting the coarse-grained features into serialized one-dimensional features through repeated convolution and pooling operations.
Step four, regression modeling: and carrying out regression modeling on the serialized one-dimensional features through a multilayer perceptron comprising a single hidden layer. And training the model constructed in the third step to the fourth step through error back propagation, and correcting network parameters.
Step five, real-time prediction: and (4) predicting the quality variable corresponding to the real-time process variable by using the model trained in the step four aiming at the collected real-time process variable data in the actual production.
Further, in the second step, the input value D of the sample is usedtIs a data matrix, and the data value is a value, so the input data of the countermeasure network is firstly serialized and spliced to obtain a one-dimensional vector. And then, training of the generation countermeasure network is carried out to obtain one-dimensional generation data. Finally, the generated data is divided and reconstructed into a complete sample S*. Generating a sample S*Will compensate for the sample StData is locally missing for a certain period of time.
Further, in the third step, the convolutional layer is subjected to average pooling, and the result of pooling is as follows:
Figure GDA0003537191810000022
where b denotes the ordinal threshold of the activation values involved in pooling, RlRepresenting pooled fields in the ith feature map, k representing index values of activation values throughout the pool, rkAnd akRespectively, the ordinal and activation values of the activation value k.
Further, in the model constructed in the third to fourth steps, the activation function between the convolutional layer and the pooling uses a linear rectification function, and the formula is as follows:
f(x)＝max(0,x)
where x represents the in-and-out vector from the previous layer network.
The invention has the following beneficial effects:
1. by splicing the data of the process variables in the adjacent time periods, the process variable data which cannot participate in modeling originally is utilized. Meanwhile, fine-grained characteristics in the sample are converted into coarse-grained characteristics, information in process data is better reserved, and the problems of variable time delay and non-correspondence of input and output of the sample are solved.
2. The data enhancement is carried out on the samples by utilizing the generated countermeasure network, the absolute number of the samples is expanded, and the influence of local data loss on soft measurement modeling is prevented.
3. The multi-scale convolution kernel is used for pertinently extracting the coarse-grained characteristics, one-dimensional and two-dimensional convolution kernels are respectively considered in the aspects of the dynamic property and the time ductility of the data, and the characteristics of the process data are better met.
Drawings
FIG. 1 is a schematic diagram of data preprocessing;
FIG. 2 is a schematic diagram of data enhancement;
FIG. 3 is a schematic diagram of a soft measurement model structure;
FIG. 4 is a schematic diagram of a multi-scale convolution kernel.
Detailed Description
The enhanced multi-scale convolution kernel CNN soft measurement method of the present invention is further described in detail below with reference to specific embodiments.
The invention discloses an enhanced multi-scale convolutional neural network soft measurement method, which is used for realizing the prediction of quality variable data in industry based on the generation of a countermeasure network and a Convolutional Neural Network (CNN) structure, and comprises the following steps:
step one (data preprocessing): the data preprocessing process is shown in FIG. 1, and the quality variable is set to be Y and the process variable is set to be XiI is 1,2, …, m is the number of process variables. Taking the quality variable data collected at the moment t as a sample StOutput of (2)Value ytAnd t is 1,2, …. Setting the duration of the material in the whole process as n, and dividing [ t-n, t]Process variable x collected over a period of timei·jI 1,2, …, m, j 1,2, …, n, spliced together as sample StInput value of
Figure GDA0003537191810000031
The sample input obtained by reconstruction at this time is called coarse-grained feature Dt. Then, the process variables are normalized by the maximum and minimum normalization to eliminate dimensional differences between the variables.
Step two (data enhancement): data enhancement procedure As shown in FIG. 2, the samples S are paired with the generation countermeasure networktCarrying out data enhancement to obtain a new generated sample S*. Due to the input value D of the sampletIs a data matrix and the data value is a value, so that the input data D of the countermeasure network is generated firsttAnd carrying out serialization and splicing treatment to obtain a one-dimensional vector. And then, one-dimensional generated data is obtained by training the generation confrontation network. Finally, the generated data is divided and reconstructed into a complete sample S*. Generating a sample S*Will compensate for the sample StData is locally missing for a certain period of time.
Step three (feature extraction): as shown in FIG. 4, the multiscale convolution operation uses both a 1 × U one-dimensional convolution kernel and a V × W two-dimensional convolution kernel for the sample StAnd sample S*Extracting the characteristics of the medium and coarse granularity characteristics; u, V, W are selected in combination with the actual data situation. Unlike ordinary image data, the coarse-grained features of process variable data have different physical meanings on different coordinate axes. Therefore, we want to use multi-scale features to extract the coarse-grained features sufficiently to obtain complete process information. The one-dimensional convolution core extracts dynamic characteristics of the time sequence of each process variable, and the two-dimensional convolution core extracts time delay relation characteristics of the time sequence matrix of the adjacent process variable. And then fusing feature maps extracted by convolution kernels under different scales through convolution of 1x 1.
After the convolution operation, the feature map is subjected to dimensionality reduction by using the pooling layer. As shown in FIG. 3, the coarse-grained features D are transformed by repeated convolution and pooling operationstCompletely converting into serialized one-dimensional features. To make full use of the existing information, the convolutional layer was pooled evenly, and the pooling results are shown below:
Figure GDA0003537191810000041
where b denotes the ordinal threshold of the activation values involved in pooling, RlRepresenting pooled fields in the ith feature map, k representing index values of activation values throughout the pool, rkAnd akRespectively, the ordinal and activation values of the activation value k.
Step four (regression modeling): and carrying out regression modeling on the serialized one-dimensional features through a multilayer perceptron comprising a single hidden layer. And training the model through error back propagation, and correcting network parameters.
Step three and step four constitute the basic structure of the multi-scale convolution CNN soft measurement model shown in FIG. 3. The activation function between convolutional layer and pooling uses a linear rectification function (ReLU), whose formula is as follows:
f(x)＝max(0,x)
where x represents the in-and-out vector from the previous layer network. The ReLU enables the network to perform gradient descent and back propagation more efficiently, avoiding the problems of gradient explosion and gradient disappearance.
Step five (real-time prediction): and aiming at the fact that in actual production, collected real-time process variable data form coarse granularity characteristics, and the multi-scale convolution CNN soft measurement model trained in the third step, the fourth step is utilized to predict the coarse granularity characteristics in real time.
Combining a specific CO2Industrial example of an absorber tower to illustrate an enhanced multi-scale convolutional neural network soft measurement method. CO 22The absorption tower is an important part in the ammonia synthesis process, and CO in the mixed gas before being sent to the ammonia synthesis tower2The content of (a) may adversely affect the product. CO 22Absorption tower processThe equation contains 11 variables, as shown in Table 1. By modeling these variables, CO in the mixed gas is predicted2The content of (a).
Table 1: TE Process Fault Listing
Figure GDA0003537191810000042
Figure GDA0003537191810000051
In the data preprocessing stage, 30 sets of process variable data are spliced to obtain coarse-grained (30 × 11) features. Generation of a Confrontation network A Wasserstein generated Confrontation network (WGAN-GP) based on a gradient penalty was employed, generating 500 new samples. The convolutional layer uses one-dimensional convolution kernels of (1 × 3) and (1 × 5), and two-dimensional convolution kernels of (3 × 5) and (3 × 7), with a step size of 1. The size of the pooling layer was (2 × 2) with a step size of 2. The optimization algorithm adopts random gradient descent, and the learning rate is 0.01.
In the experiment, 1000 samples are collected, wherein the first 500 samples in time are taken as training samples, the second 500 samples are taken as testing samples, and 20% of the training samples are randomly deleted to simulate the problem of local data loss. Table 2 shows the predicted results of the method and other models, and from RMSE, the method achieves the best results.
Table 2: soft measurement model comparison
Soft measurement model Multilayer perceptron Convolutional neural network The invention
RMSE 0.01598 0.01298 0.01054

Claims (4)

1. An enhanced multi-scale convolution neural network soft measurement method is characterized by comprising the following steps:
step one, data preprocessing: let the mass variable be Y and the process variable be XiI is 1,2, …, m; where m is the number of process variables; taking the quality variable data collected at the moment t as a sample StOutput value y oftT is 1,2, …; setting the duration of the material in the whole process as n, and dividing [ t-n, t]Process variable x collected over a period of timei·jJ-1, 2, …, n, spliced together as a sample StInput value of
Figure FDA0003537191800000011
Referred to as coarse grain features; then, carrying out normalization operation on the process variables by using maximum and minimum normalization to eliminate dimensional difference between the variables;
step two, data enhancement: pairing samples S with a generating countermeasure networktCarrying out data enhancement to obtain a new generated sample S*；
Step three, feature extraction: the convolution operation simultaneously adopts a 1 XU one-dimensional convolution kernel and a V XW two-dimensional convolution kernel to perform convolution on the sample StAnd sample S*Performing feature extraction on the coarse-grained features after medium normalization; then fusing feature maps extracted by convolution kernels under different scales through convolution of 1x 1; then, performing dimensionality reduction on the extracted feature map by using a pooling layer; completely converting the coarse grain characteristics into serialized one-dimensional characteristics through repeated convolution and pooling operations;
step four, regression modeling: carrying out regression modeling on the serialized one-dimensional features through a multilayer perceptron comprising a single hidden layer; training the model constructed in the third step to the fourth step through error back propagation, and correcting network parameters;
step five, real-time prediction: and (4) predicting the quality variable corresponding to the real-time process variable by using the model trained in the step four aiming at the collected real-time process variable data in the actual production.
2. The enhanced multi-scale convolutional neural network soft measurement method of claim 1, wherein: in the second step, the input value D of the sample is usedtThe data matrix is used, and the data value is a value, so that input data of the countermeasure network is firstly serialized and spliced to obtain a one-dimensional vector; then, one-dimensional generated data is obtained through training of a generated countermeasure network; finally, the generated data is divided and reconstructed into a complete sample S*(ii) a Generating a sample S*Will compensate for the sample StData is locally missing for a certain period of time.
3. The enhanced multi-scale convolutional neural network soft measurement method of claim 1, wherein: in the third step, the convolution layer adopts average pooling, and the result of pooling is as follows:
Figure FDA0003537191800000012
where b denotes the ordinal threshold of the activation values involved in pooling, RlRepresenting pooled fields in the ith feature map, k representing index values of activation values throughout the pool, rkAnd akRespectively, the ordinal and activation values of the activation value k.
4. The enhanced multi-scale convolutional neural network soft measurement method of claim 1, wherein: in the model constructed in the third to fourth steps, the activation function between the convolutional layer and the pooling uses a linear rectification function, and the formula is as follows:
f(x)＝max(0,x)
where x represents the incoming and outgoing vectors from the previous layer of the network.
CN202010994589.7A 2020-09-21 2020-09-21 Enhanced multi-scale convolution neural network soft measurement method Active CN112215351B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010994589.7A CN112215351B (en) 2020-09-21 2020-09-21 Enhanced multi-scale convolution neural network soft measurement method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010994589.7A CN112215351B (en) 2020-09-21 2020-09-21 Enhanced multi-scale convolution neural network soft measurement method

Publications (2)

Publication Number Publication Date
CN112215351A CN112215351A (en) 2021-01-12
CN112215351B true CN112215351B (en) 2022-05-03

Family

ID=74050184

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010994589.7A Active CN112215351B (en) 2020-09-21 2020-09-21 Enhanced multi-scale convolution neural network soft measurement method

Country Status (1)

Country Link
CN (1) CN112215351B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113408578B (en) * 2021-05-13 2022-10-25 华南理工大学 Harmonic reducer fault diagnosis method and system based on generative countermeasure network

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960415B (en) * 2017-05-23 2021-04-20 上海寒武纪信息科技有限公司 Processing apparatus and processing system
CN109188903A (en) * 2018-08-20 2019-01-11 北京化工大学 The flexible measurement method of CNN furnace operation variable based on memory-enhancing effect optimization
CN110322077B (en) * 2019-07-10 2022-08-02 燕山大学 Cement raw material vertical mill raw material fineness index prediction method based on convolutional neural network
CN110378044B (en) * 2019-07-23 2021-06-11 燕山大学 Multi-time scale convolution neural network soft measurement method based on attention mechanism

Also Published As

Publication number Publication date
CN112215351A (en) 2021-01-12

Similar Documents

Publication Publication Date Title
CN110533631B (en) SAR image change detection method based on pyramid pooling twin network
CN106778682B (en) A kind of training method and its equipment of convolutional neural networks model
CN113159051B (en) Remote sensing image lightweight semantic segmentation method based on edge decoupling
CN106874688B (en) Intelligent lead compound based on convolutional neural networks finds method
CN109389171B (en) Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
CN109086799A (en) A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet
CN105825511A (en) Image background definition detection method based on deep learning
CN111539132B (en) Dynamic load time domain identification method based on convolutional neural network
CN112560966B (en) Polarized SAR image classification method, medium and equipment based on scattering map convolution network
CN109815855B (en) Electronic equipment automatic test method and system based on machine learning
CN115587985A (en) Method for dividing cell nucleus of histopathology image and normalizing dyeing style
CN111626296B (en) Medical image segmentation system and method based on deep neural network and terminal
Roewer-Despres et al. Towards finite element simulation using deep learning
Guo et al. Estimating cement compressive strength using three-dimensional microstructure images and deep belief network
CN112215351B (en) Enhanced multi-scale convolution neural network soft measurement method
CN114565594A (en) Image anomaly detection method based on soft mask contrast loss
CN112947080B (en) Scene parameter transformation-based intelligent decision model performance evaluation system
CN114170657A (en) Facial emotion recognition method integrating attention mechanism and high-order feature representation
Rethik et al. Attention Based Mapping for Plants Leaf to Classify Diseases using Vision Transformer
CN110288002B (en) Image classification method based on sparse orthogonal neural network
CN116883364A (en) Apple leaf disease identification method based on CNN and Transformer
CN111753995A (en) Local interpretable method based on gradient lifting tree
CN110554667A (en) convolutional Neural Network (CNN) based intermittent industrial process fault diagnosis
CN116129189A (en) Plant disease identification method, plant disease identification equipment, storage medium and plant disease identification device
CN116595313A (en) Fault diagnosis method and system based on multi-scale residual error attention 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