CN106862284B - A kind of cold rolled sheet signal mode knowledge method for distinguishing - Google Patents
A kind of cold rolled sheet signal mode knowledge method for distinguishing Download PDFInfo
- Publication number
- CN106862284B CN106862284B CN201710181131.8A CN201710181131A CN106862284B CN 106862284 B CN106862284 B CN 106862284B CN 201710181131 A CN201710181131 A CN 201710181131A CN 106862284 B CN106862284 B CN 106862284B
- Authority
- CN
- China
- Prior art keywords
- network
- plate shape
- quantum
- layer
- training
- 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
Links
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B21—MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21B—ROLLING OF METAL
- B21B38/00—Methods or devices for measuring, detecting or monitoring specially adapted for metal-rolling mills, e.g. position detection, inspection of the product
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B21—MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21B—ROLLING OF METAL
- B21B37/00—Control devices or methods specially adapted for metal-rolling mills or the work produced thereby
- B21B37/28—Control of flatness or profile during rolling of strip, sheets or plates
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B21—MECHANICAL METAL-WORKING WITHOUT ESSENTIALLY REMOVING MATERIAL; PUNCHING METAL
- B21B—ROLLING OF METAL
- B21B38/00—Methods or devices for measuring, detecting or monitoring specially adapted for metal-rolling mills, e.g. position detection, inspection of the product
- B21B38/04—Methods or devices for measuring, detecting or monitoring specially adapted for metal-rolling mills, e.g. position detection, inspection of the product for measuring thickness, width, diameter or other transverse dimensions of the product
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- Mechanical Engineering (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Physiology (AREA)
- Genetics & Genomics (AREA)
- Control Of Metal Rolling (AREA)
Abstract
A kind of cold rolled sheet signal mode knowledge method for distinguishing, this method content include the following steps:Each measuring section plate shape measurement value of cold-strip steel width direction for acquiring plate profile instrument on-line measurement, obtains each measuring section plate shape value;The neural network that the initial data that plate profile instrument exports is input to a n-layer mainly allows network to automatically extract feature to exclude artificial trace as feature extraction layer by training;It is identified into andante shape using the modified quantum nerve network based on genetic algorithm.The modified quantum nerve network of the multilayer excitation function of genetic algorithm optimization is applied in Flatness Pattern identification technology by the present invention, the training effectiveness for significantly improving network efficiently solves the precision encountered in conventional panels shape recognition methods and real-time is not ideal enough, the complicated network structure and training time is long and stability and the problems such as poor robustness.
Description
Technical field
The invention belongs to cold rolled sheet fields, are related to a kind of cold rolled sheet signal mode knowledge method for distinguishing.
Background technology
In cold rolled sheet production process, the pattern-recognition of cold rolled sheet shape is the important composition portion of cold rolled sheet shape control system
Point, there is vital effect to successive plates shape control effect to the accuracy of identification of flatness defect state in this link.It is logical
The stress signal data crossed in plate profile instrument acquisition production scene plank are recognized, and judge the class of current flatness defect state
Type, and the control mechanism of plate shape is fed back to, the degree of imperfection of plate shape is reduced by controlling the adjustment of executing agency, finally makes plate shape
Output meet the production standard of technique, so should seek and study the recognition methods of high precision plates shape defect mode.
Traditional Shape signal mode identification method is Factoring Polynomials method based on least square method and improved
Orthogonal polynomial regression decomposition method, these method interference rejection abilities are poor, in theory existing defects, do not meet flatness defect distribution
Essence, it is difficult to meet high-precision shape control demand.Plate shape recognition methods based on fuzzy theory is simple and effective, but precision
It is not ideal enough with real-time, practical application it is seldom.In recent years, experts and scholars in the art have studied based on neural network
Shape signal identify and application etc. technology, achieve extraordinary effect.But due to the complexity of neural network and
The deficiency of existing optimisation technique causes debating based on neural network filter system to know model structure complexity, and there are nets
The problems such as network training time is long, and stability is poor.
Plate shape refers to distribution situation of the Strip internal residual stress along plate width direction, and the task of plate shape identification is exactly that handle exists
One group of tension distribution centrifugal pump of line detection is mapped as less several characteristic parameters by certain processing, and can be preferably anti-
Reflect the classification situation of flatness defect.The quantum nerve network of multilayer excitation function uses the superposition of multiple Sigmoid functions, will be special
Sign space is divided into more multiple level, and to indicate more states, the general hidden layer excitation function by superposition has multiple and different amounts
Son interval, and each quantum interval corresponds to different quantum levels or ladder, you can to correspond to different space structures respectively.
In the training process of model, when sample data enters feature space, it need not be gone in entire feature space i.e. [0,1] range
Search for corresponding space structure, but the removal search in the space structure being layered, to make sample information quickly reflect
It is mapped on corresponding magnitude or ladder;Meanwhile if when the reality output of model is larger with desired output error, by appropriate
Learning algorithm adjustment quantum interval, change the quantum level divided or ladder width, to adapt to the knot of available data sample
Structure characteristic makes the uncertainty in sample data by quantum neural more preferably, to be quickly mapped on corresponding space structure
Network obtains and quantifies, and to greatly shorten the training time of model, improves the accuracy and convergence rate of model prediction result.
On the other hand, intelligent optimization algorithm has obtained quick development in nearest decades, passes through intelligent optimization method
It can obtain the optimal solution of many nonlinear optimal problems.Intelligent optimization algorithm technology is applied to quantum nerve network training
In, modeling accuracy and efficiency are improved, is the research direction rich in foreground and application value in Nonlinear Modeling field.This is also
High-precision and efficient cold rolled sheet signal on-line mode identification technology problem provide technical support and theoretical foundation.
In conclusion researching and developing with high-precision and efficient cold rolled sheet shape signal on-line mode recognition methods, for control
System processed provides reliable control foundation, to produce the cold-rolled plate and strip product of high quality, be further increase it is current cold
Roll a key technical problem urgently to be resolved hurrily of belt plate shape controlled level.
Invention content
The purpose of the present invention is to provide a kind of cold rolled sheet signal modes to know method for distinguishing, and this method can be solved effectively
The precision and real-time encountered using conventional panels shape mode identification method is inadequate, debates and knows model structure complexity and net training time
Long, the technical problems such as stability difference system can provide reliable control foundation, to improve the plate shape of cold-strip steel in order to control
It controls quality and strong guarantee is provided.
To solve the above-mentioned problems, technical solution provided by the invention is:
A kind of cold rolled sheet signal mode knowledge method for distinguishing, this method content include the following steps:
Step 1 acquires each measuring section plate shape measurement value of cold-strip steel width direction of plate profile instrument on-line measurement, obtains each survey
Measure section plate shape value;It is m to enable the number of measuring section, theiA measuring section plate shape measurement value is Fi;
The initial data that plate profile instrument exports is input to the neural network of a n-layer as feature extraction layer, mainly by step 2
Network is allowed to automatically extract feature to exclude artificial trace by training;Because plate shape mainly have left side wave, the right wave, middle wave,
Unrestrained eight kinds of defect modes in bilateral wave, right three points of waves, left three points of waves, four points of waves and side, so network output is made to be ai, i=1,
2,3...m, wherein m are output number, take m=8 here, then the calculation formula of each layer of network is as follows:
al i=f (Wlxl+bl), l=1,2 ... n
Wherein l is l layers of network;WlFor l layer network weights;xlIt is inputted for l layers;blIt is biased for network neural member;
Step 3 is identified using the modified quantum nerve network based on genetic algorithm into andante shape, by data processing module
Output aiAs the input of quantum nerve network, the output u of quantum nerve network1,u2,u3,u4For the degree of membership of plate shape;
If u1>0, display plate shape has left side wave, if u1<0, display plate shape has the right wave;
If u2>0, display plate shape has middle wave, if u2<0, display plate shape has bilateral wave;
If u3>0, display plate shape has left three points of waves, if u3<0, display plate shape has right three points of waves;
If u4>0, display plate shape has four points of waves, if u4<0, display plate shape has wave in side.
In step 3, the modified quantum nerve network based on genetic algorithm, process of establishing include following step
Suddenly:
1) control parameter of the quantum nerve network training study for Flatness Pattern identification is determined:
Specifically control parameter includes:For the initial population number NP of network training genetic algorithm, the maximum of genetic algorithm
Learn algebraically N, the desired value e and training sample number M of network training effect;
2) processing of training sample data:
The data that plate profile instrument exports are directly inputted to network, the multilayer neural network of network front end is allowed to carry out automated characterization
Extraction, is then input in the quantum nerve network of multilayer excitation function;
3) setting of multilayer excitation function modified quantum nerve network:
Multilayer excitation function modified quantum nerve network model uses fully connected topology, is by feature extraction layer, hidden
It is constituted containing layer and output layer, feature extraction layer is also referred to as input layer;Model hidden layer neuron excitation function uses Multiple-quantum energy level
Transforming function transformation function, each multiple level transforming function transformation function are a series of n with quantum spacing biassThe linear of a Sigmoid functions is folded
Add, i.e. multilayer excitation function;Therefore, the output y of hidden layer neuronhIt is represented by:
Wherein V is the hidden layer weights of multilayer excitation function quantum nerve network, and a is characterized the output of extract layer, θrFor
Quantum interval, nsFor quantum skip number;
4) by the weights of feature extraction layer network and biasing, the hidden layer weights of multilayer excitation function quantum nerve network, amount
Son interval and output layer weights are defined as the individual vector of genetic algorithm, and NP individual vector is determined according to equiprobability random distribution
Population initial value;
5) Optimization Learning of individual vector is carried out according to the mutation operation of genetic algorithm, crossover operation and selection operation;Make
The network parameter obtained with Optimization Learning configures the modified quantum nerve network parameter of multilayer excitation function, is instructed by M
Practice sample data and brings the modified quantum nerve network of multilayer excitation function into calculate output valve, then by calculating above-mentioned M
The quadratic sum of arithmetical difference between a corresponding 4M network output valve of training sample data and its actual value;The arithmetical difference is put down
Just and it is defined as error criterion functional value;Mean net if the desired value g that error criterion functional value is less than network training effect
Otherwise network training success should continue to train network, g that should be chosen according to the quantity of training sample and the scale of network;
6) network parameter of record training study obtains the modified amount of the multilayer excitation function of one 15 4 output of input
Sub- nerve net;
After establishing network according to above-mentioned steps, then into andante shape identify.
The present invention has the beneficial effect that:
1) data that plate profile instrument inputs are carried out automatically extracting feature by the neural network of a multilayer, are automatically processed,
Human disturbance factor is excluded, the precision of Network Recognition plate shape can be significantly improved.
2) the modified quantum nerve network of the multilayer excitation function of genetic algorithm optimization is applied to Flatness Pattern identification
In technology, the training effectiveness of network is significantly improved, efficiently solves the precision that is encountered in conventional panels shape recognition methods and in real time
Property the not ideal enough, the complicated network structure and training time is long and stability and the problems such as poor robustness, provided for Strip Shape Control
Reliable foundation.
Description of the drawings
Fig. 1 is implementing procedure figure of the present invention;
Fig. 2 is the modified quantum nerve network structural topology figure of type multilayer excitation function;
Fig. 3 is the modified quantum nerve network instruction of the multilayer excitation function based on genetic algorithm in one embodiment of the invention
Practice learning process figure;
Fig. 4 is obtained recognition effect figure after being identified into andante shape using the present invention.
Specific implementation mode
With reference to specific example and attached drawing, the present invention will be further described.
A kind of cold rolled sheet signal mode of the present invention knows method for distinguishing, and Fig. 1 is the flow chart of one embodiment of the invention, it
Include the following steps:
Step 1 acquires each measuring section plate shape measurement value of cold-strip steel width direction of plate profile instrument on-line measurement, obtains each survey
Measure section plate shape value;It is m, m=15, the to enable the number of measuring sectioniA measuring section plate shape measurement value is Fi;
The initial data that plate profile instrument exports is input to the neural network of a n-layer as feature extraction layer, mainly by step 2
Network is allowed to automatically extract feature to exclude artificial trace by training;Since plate shape mainly has left side wave, the right wave, centre
Unrestrained eight kinds of defect modes in unrestrained, bilateral wave, right three points of waves, left three points of waves, four points of waves and side, so network output is made to be ai, i=
1,2,3...m, wherein m are output number, take m=8 here;Then the calculation formula of each layer of network is as follows:
al i=f (Wlxl+bl), l=1,2 ... n
Wherein l is l layers of network;WlFor l layer network weights;xlIt is inputted for l layers;blIt is biased for network neural member;
Step 3 is identified using the modified quantum nerve network based on genetic algorithm into andante shape, by data processing module
Output aiInput as quantum nerve network;The output u of quantum nerve network1,u2,u3,u4For the degree of membership of plate shape;
If u1>0, display plate shape has left side wave, if u1<0, display plate shape has the right wave;
If u2>0, display plate shape has middle wave, if u2<0, display plate shape has bilateral wave;
If u3>0, display plate shape has left three points of waves, if u3<0, display plate shape has right three points of waves;
If u4>0, display plate shape has four points of waves, if u4<0, display plate shape has wave in side.
According to said program, the process of establishing of the modified quantum nerve network based on genetic algorithm is:
1) control parameter of the quantum nerve network training study for Flatness Pattern identification is determined:
Specifically control parameter includes:For the initial population number NP of network training genetic algorithm, the maximum of genetic algorithm
Learn algebraically N, the desired value e and training sample number M of network training effect;
2) processing of training sample data:
The data that plate profile instrument exports are directly inputted to network, the multilayer neural network of network front end is allowed to carry out automated characterization
Extraction, is then input in the quantum nerve network of multilayer excitation function.
3) setting of multilayer excitation function modified quantum nerve network:
Multilayer excitation function modified quantum nerve network model uses fully connected topology as shown in Figure 2, is by spy
It levies extract layer, hidden layer and output layer to constitute, feature extraction layer is also referred to as input layer;Model hidden layer neuron excitation function uses
Multiple-quantum energy level transforming function transformation function, each multiple level transforming function transformation function are a series of n with quantum spacing biassA Sigmoid letters
Several linear superpositions, i.e. multilayer excitation function;Therefore, the output y of hidden layer neuronhIt is represented by:
Wherein V is the hidden layer weights of multilayer excitation function modified quantum nerve network, and a is characterized the defeated of extract layer
Go out, θrFor quantum interval, nsFor quantum skip number;
4) by the weights of feature extraction layer network and biasing, the hidden layer weights of multilayer excitation function quantum nerve network, amount
Son interval and output layer weights are defined as the individual vector of genetic algorithm, and NP individual vector is determined according to equiprobability random distribution
Population initial value;
5) Optimization Learning of individual vector is carried out according to the mutation operation of genetic algorithm, crossover operation and selection operation;Make
The network parameter obtained with Optimization Learning configures the modified quantum nerve network parameter of multilayer excitation function, is instructed by M
Practice sample data and brings the modified quantum nerve network of multilayer excitation function into calculate output valve, then by calculating above-mentioned M
The quadratic sum of arithmetical difference between a corresponding 4M network output valve of training sample data and its actual value;The arithmetical difference is put down
Just and it is defined as error criterion functional value;Mean net if the desired value g that error criterion functional value is less than network training effect
Otherwise network training success should continue to train network, g that should be chosen according to the quantity of training sample and the scale of network;
6) network parameter of record training study obtains the modified amount of the multilayer excitation function of one 15 4 output of input
Sub- nerve net;Network is established according to above-mentioned steps and then is identified into andante shape.
The method of the present invention can be used for four-roller, six roller single chassis or multi-frame tandem mills.With six rolling of single chassis
For machine, the product of six-high cluster mill section rolling includes common plate, high-strength steel, part stainless steel and silicon steel etc..This example uses
Type be 900HC milling trains, the unit main technical performance index and device parameter are:
Mill speed:Max 300m/min, draught pressure:Max 8000KN, maximum rolling force square:60KN × m, curling
Power:4~80KN;Supplied materials thickness range:2.0~4.0mm, plate wide scope:460~780mm, finished product thickness:0.24~1.5mm;;
Working roll size:270/ 245 × 900mm of φ of φ, intermediate calender rolls size:340/ 320 × 920mm of φ of φ, support roller ruler
It is very little:850/ φ 790 of φ~× 850;
The maximum traversing amount of intermediate calender rolls:200mm, the positive/negative bending roller force of working roll (unilateral side):400/254.5KN.
Plate Profile Measuring System uses the self-control plate profile instrument in 15 channels.
The specific calculation process of Shape signal that this example carries out dead channels is:
(1) acquisition receives each measuring section plate shape measurement value in cold-strip steel direction of plate profile instrument on-line measurement.Plate profile instrument is cold
Roll has the measured zone of determining width, each measured zone to provide corresponding region configured with several on strip width direction
Plate shape value.Side is eaten from plate profile instrument fore side to transmission shares 15 effective measuring areas, measuring section hop count m-15, i.e. Fi, i=
1,2 ... 15 units are plate shape international unit I.The data that plate profile instrument measures are introduced directly into network and do not do artificial treatment.
(2) using plate profile instrument export initial data be input to a n-layer neural network be used as feature extraction layer, mainly
It is to allow network to automatically extract feature by training to exclude artificial trace.Since plate shape mainly has left side wave, the right wave, centre
Unrestrained eight kinds of defect modes in unrestrained, bilateral wave, right three points of waves, left three points of waves, four points of waves and side, so network output is made to be ai, i=
1,2,3...m, wherein m are output number, take m=8 here.Then the calculation formula of each layer of network is as follows:
al i=f (Wlxl+bl), l=1,2 ... n
Wherein l is l layers of network;WlFor l layer network weights;xlIt is inputted for l layers;blIt is biased for network neural member.
(3) it is identified into andante shape using the modified quantum nerve network based on genetic algorithm, by data processing module
Export aiInput as quantum nerve network;The output u of quantum nerve network1,u2,u3,u4For the degree of membership of plate shape;
If u1>0, display plate shape has left side wave, if u1<0, display plate shape has the right wave;
If u2>0, display plate shape has middle wave, if u2<0, display plate shape has bilateral wave;
If u3>0, display plate shape has left three points of waves, if u3<0, display plate shape has right three points of waves;
If u4>0, display plate shape has four points of waves, if u4<0, display plate shape has wave in side.
According to said program, as shown in figure 3, the process of establishing of the novel quantum nerve network based on genetic algorithm is:
1. determining the control parameter of the quantum nerve network training study for Flatness Pattern identification:
Design parameter includes:The initial population number NP of genetic algorithm for network training, the most university of genetic algorithm
Algebraically N is practised, the desired value of network training effect is e and training sample number is M;
2. the processing of training sample data:
The data that plate profile instrument exports are directly inputted to network, the multilayer neural network of network front end is allowed to carry out automated characterization
Extraction interest is then input in the quantum nerve network of multilayer excitation function.
3. the setting of the modified quantum nerve network of multilayer excitation function:
The modified quantum nerve network model of multilayer excitation function uses fully connected topology as shown in Figure 2, is
It is made of feature extraction layer (input layer), hidden layer and output layer.Model hidden layer neuron excitation function uses Multiple-quantum energy
Grade transforming function transformation function, each multiple level transforming function transformation function is a series of n with quantum spacing biassA Sigmoid functions it is linear
Superposition, i.e. multilayer excitation function;Therefore, the output y of hidden layer neuronhIt is represented by:
Wherein V is the hidden layer weights of multilayer excitation function modified quantum nerve network, and a is characterized the defeated of extract layer
Go out, θrFor quantum interval, nsFor quantum skip number.
4. by the weights of feature extraction layer network and biasing, the hidden layer of multilayer excitation function modified quantum nerve network is weighed
Value, quantum interval and output layer weights are defined as the individual vector of genetic algorithm, and NP are determined according to equiprobability random distribution
The population initial value of body vector;
5. according to the mutation operation of genetic algorithm, crossover operation and selection operation carry out the vectorial Optimization Learning of individual;Make
The network parameter obtained with Optimization Learning configures the improvement quantum nerve network parameter of multilayer excitation function, by M training
Sample data brings the modified quantum nerve network of multilayer excitation function into calculate output valve, then M above-mentioned by calculating
The quadratic sum of arithmetical difference between the corresponding 4M network output valve of training sample data and its actual value.Square of the arithmetical difference
Be defined as error criterion functional value;Mean network if the desired value g that error criterion functional value is less than network training effect
It trains successfully, otherwise, should continue to train network, g that should be chosen according to the quantity of training sample and the scale of network;
6. the network parameter of record training study obtains the modified amount of the multilayer excitation function of one 15 4 output of input
Sub- nerve net;
Network is established according to above-mentioned steps and then is identified into andante shape.
Fig. 4 gives this example and carries out obtained recognition effect figure after on-line mode identification.Fig. 4-a are Network Recognitions
Flatness defect curve distribution figure;Fig. 4-b are the left side wave components after identification;Fig. 4-c are the bilateral unrestrained components after identification;Fig. 4-d
It is three points of the left side wave after identification;Fig. 4-e be identification after side in wave.As can be seen from the figure technical solution energy proposed by the present invention
Enough identification missions for completing flatness defect well.Solve that the conventional method precision and real-time that encounter are not ideal enough, debate knowledge mould
The complicated technology for cutting too big net training time length, data characteristics processing human factor, stability and poor robustness of type is asked
Topic system can put forward reliable control foundation in order to control, strong guarantee is provided to improve cold rolled sheet steel quality.
Above example is merely to illustrate the Computation schema and feature of the present invention, and its object is to make technology in the art
Personnel can understand present disclosure and according to implement, and protection scope of the present invention is not limited to examples detailed above.So it is all according to
According to equivalent variations or modification made by disclosed principle, mentality of designing, within protection scope of the present invention.
Claims (2)
1. a kind of cold rolled sheet signal mode knows method for distinguishing, it is characterised in that:This method content includes the following steps:
Step 1 acquires each measuring section plate shape measurement value of cold-strip steel width direction of plate profile instrument on-line measurement, obtains each measuring section
Plate shape value;It is m to enable the number of measuring section, and ith measurement section plate shape measurement value is Fi;
The neural network that the initial data that plate profile instrument exports is input to a n-layer by step 2 is mainly led to as feature extraction layer
Crossing training allows network to automatically extract feature to exclude artificial trace;Because plate shape mainly has left side wave, the right wave, middle wave, double
Unrestrained eight kinds of defect modes in Bian Lang, right three points of waves, left three points of waves, four points of waves and side, so network output is made to be ai, i=1,2,
3...m, wherein m is output number, takes m=8 here, then the calculation formula of each layer of network is as follows:
al i=f (Wlxl+bl), l=1,2 ... n
Wherein l is l layers of network;WlFor l layer network weights;xlIt is inputted for l layers;blIt is biased for network neural member;
Step 3 is identified using the modified quantum nerve network based on genetic algorithm into andante shape, by the defeated of data processing module
Go out aiAs the input of quantum nerve network, the output u of quantum nerve network1,u2,u3,u4For the degree of membership of plate shape;
If u1>0, display plate shape has left side wave, if u1<0, display plate shape has the right wave;
If u2>0, display plate shape has middle wave, if u2<0, display plate shape has bilateral wave;
If u3>0, display plate shape has left three points of waves, if u3<0, display plate shape has right three points of waves;
If u4>0, display plate shape has four points of waves, if u4<0, display plate shape has wave in side.
2. a kind of cold rolled sheet signal mode according to claim 1 knows method for distinguishing, it is characterised in that:It is described to be based on losing
The modified quantum nerve network of propagation algorithm, the process of foundation include the following steps:
1) control parameter of the quantum nerve network training study for Flatness Pattern identification is determined:
Specifically control parameter includes:For the initial population number NP of network training genetic algorithm, the maximum study of genetic algorithm
Algebraically N, the desired value e and training sample number M of network training effect;
2) processing of training sample data:
The data that plate profile instrument exports are directly inputted to network, allows the multilayer neural network of network front end to carry out automated characterization and carries
It takes, is then input in the quantum nerve network of multilayer excitation function;
3) setting of multilayer excitation function modified quantum nerve network:
Multilayer excitation function modified quantum nerve network model uses fully connected topology, is by feature extraction layer, hidden layer
It is constituted with output layer, feature extraction layer is also referred to as input layer;Model hidden layer neuron excitation function is converted using Multiple-quantum energy level
Function, each multiple level transforming function transformation function are a series of n with quantum spacing biassThe linear superposition of a Sigmoid functions, i.e.,
Multilayer excitation function;Therefore, the output y of hidden layer neuronhIt is represented by:
Wherein V is the hidden layer weights of multilayer excitation function quantum nerve network, and a is characterized the output of extract layer, θrBetween quantum
Every nsFor quantum skip number;
It 4) will be between the weights of feature extraction layer network and biasing, the hidden layer weights of multilayer excitation function quantum nerve network, quantum
Every the individual vector for being defined as genetic algorithm with output layer weights, the kind of NP individual vector is determined according to equiprobability random distribution
Group's initial value;
5) Optimization Learning of individual vector is carried out according to the mutation operation of genetic algorithm, crossover operation and selection operation;Using excellent
Chemical acquistion to network parameter configure the modified quantum nerve network parameter of multilayer excitation function, by M trained sample
Notebook data brings the modified quantum nerve network of multilayer excitation function into calculate output valve, then by calculating above-mentioned M instruction
Practice the quadratic sum of the arithmetical difference between the corresponding 4M network output valve of sample data and its actual value;The quadratic sum of the arithmetical difference
It is defined as error criterion functional value;Mean that network is instructed if the desired value g that error criterion functional value is less than network training effect
Practice successfully, otherwise, should continue to train network, g that should be chosen according to the quantity of training sample and the scale of network;
6) network parameter of record training study obtains the modified quantum god of the multilayer excitation function of one 15 4 output of input
Through net;
After establishing network according to above-mentioned steps, then into andante shape identify.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710181131.8A CN106862284B (en) | 2017-03-24 | 2017-03-24 | A kind of cold rolled sheet signal mode knowledge method for distinguishing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710181131.8A CN106862284B (en) | 2017-03-24 | 2017-03-24 | A kind of cold rolled sheet signal mode knowledge method for distinguishing |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106862284A CN106862284A (en) | 2017-06-20 |
CN106862284B true CN106862284B (en) | 2018-09-04 |
Family
ID=59173517
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710181131.8A Active CN106862284B (en) | 2017-03-24 | 2017-03-24 | A kind of cold rolled sheet signal mode knowledge method for distinguishing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106862284B (en) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109063728A (en) * | 2018-06-20 | 2018-12-21 | 燕山大学 | A kind of fire image deep learning mode identification method |
CN109668842A (en) * | 2019-02-21 | 2019-04-23 | 岭南师范学院 | A kind of shellfish heavy metal pollution rapid detection method |
CN110348131A (en) * | 2019-07-15 | 2019-10-18 | 燕山大学 | A kind of FPGA implementation method of RBF plate shape identification model |
CN112270649A (en) * | 2020-09-30 | 2021-01-26 | 燕山大学 | Improved deep learning-based plate shape identification method |
CN113172097B (en) * | 2021-05-27 | 2022-07-29 | 燕山大学 | Cold-rolled strip shape pattern recognition method and system based on depth self-coding |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP3073613B2 (en) * | 1992-10-29 | 2000-08-07 | 株式会社神戸製鋼所 | Board removal method |
JP2984182B2 (en) * | 1994-04-22 | 1999-11-29 | 新日本製鐵株式会社 | Rolling mill logistics scheduling method |
JP2004148377A (en) * | 2002-10-31 | 2004-05-27 | Jfe Steel Kk | Finish rolling method |
CN100520786C (en) * | 2006-06-27 | 2009-07-29 | 宝山钢铁股份有限公司 | Processing parameter setting method of tension straightening machine set correct roller |
CN101846969A (en) * | 2009-03-23 | 2010-09-29 | 上海都峰智能科技有限公司 | Temper mill |
CN103341505B (en) * | 2013-06-25 | 2015-01-28 | 中冶南方工程技术有限公司 | Cold-rolled sheet shape signal online mode recognition method |
CN103745101A (en) * | 2013-12-30 | 2014-04-23 | 中冶南方(武汉)自动化有限公司 | Improved neural network algorithm based forecasting method of set value of rolling force of medium plate |
-
2017
- 2017-03-24 CN CN201710181131.8A patent/CN106862284B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN106862284A (en) | 2017-06-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106862284B (en) | A kind of cold rolled sheet signal mode knowledge method for distinguishing | |
CN104018831B (en) | A kind of fractured well reservoir evaluation methods | |
WO2023130666A1 (en) | Strip steel plate convexity prediction method based on data-driving and mechanism model fusion | |
CN107377634B (en) | A kind of hot-strip outlet Crown Prediction of Media method | |
CN108655186B (en) | Roll-force presetting method based on artificial neural network and mathematical model | |
CN108038300A (en) | Optical fiber state evaluating method based on improved membership function combination neutral net | |
CN108898215A (en) | A kind of sludge bulking INTELLIGENT IDENTIFICATION method based on two type fuzzy neural networks | |
CN110322014A (en) | A kind of finished cement specific surface area prediction technique based on BP neural network | |
CN103577888A (en) | Improved entropy weight AHP and application thereof | |
CN106825069B (en) | A kind of cold-strip steel high precision plates shape surface roughness on-line intelligence control method | |
CN106355540A (en) | Small- and medium-sized reservoir dam safety evaluating method based on GRA-BP (grey relational analysis and back propagation) neural network | |
CN110175763A (en) | Objective multiple attribute decision making (MADM) quality of voltage evaluation method based on fuzzy evaluation model | |
CN104785535A (en) | Cold rolling flatness quality judgment method based on fuzzy algorithm | |
CN104634265B (en) | A kind of mineral floating froth bed soft measurement method of thickness based on multiplex images Fusion Features | |
CN106056235A (en) | Power transmission grid efficiency and benefit detection method based on Klee method and matter element extension model | |
CN106841545A (en) | Based on the tealeaves organoleptic quality evaluations method for improving fuzzy comprehensive evaluation method | |
CN108053066A (en) | The Forecasting Methodology of adjustable electric load in industrial production user's energy management system | |
CN110479770A (en) | A kind of optimal control method of mill plate shape | |
CN105259754A (en) | Board thickness intelligent control method based on active learning | |
Li et al. | Modeling and validation of bending force for 6-high tandem cold rolling mill based on machine learning models | |
CN102886384A (en) | Flatness defect identification method of 20-roller Sendzimir rolling mill based on support vector machine | |
CN105290122B (en) | A kind of AGC system measurer for thickness of RBF network supports | |
CN110222825A (en) | A kind of finished cement specific surface area prediction technique and system | |
CN103316928B (en) | On-line cold-rolled flatness signal pattern recognition system | |
He et al. | Whole process prediction model of silicon steel strip on transverse thickness difference based on Takagi-Sugeno fuzzy 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 |