CN107247993A - Alloy designations recognition methods based on artificial neural network - Google Patents
Alloy designations recognition methods based on artificial neural network Download PDFInfo
- Publication number
- CN107247993A CN107247993A CN201710408317.2A CN201710408317A CN107247993A CN 107247993 A CN107247993 A CN 107247993A CN 201710408317 A CN201710408317 A CN 201710408317A CN 107247993 A CN107247993 A CN 107247993A
- Authority
- CN
- China
- Prior art keywords
- neural network
- trade mark
- artificial neural
- sample
- 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.)
- Pending
Links
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
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/20—Metals
Abstract
The invention discloses a kind of alloy designations recognition methods based on artificial neural network, this method includes：(1) foundation of artificial neural network.Suitable artificial neural network is set up according to the size in trade mark storehouse, the suitable artificial neural network includes neural network type, the number of plies, every layer of neuron number, transmission function, training function.(2) training of artificial neural network.The constituent content of element is paid close attention in the input of sample for emphasis, and sample is output as the trade mark identification code of the trade mark.The artificial neural network trained is preserved, it is standby.(3) emulation of artificial neural network.The neutral net that the measured data input of detected sample is trained, the output data of neutral net is rounded nearby, the numerical value after rounding is the trade mark identification code of detected sample.The trade mark of detected sample is quoted finally according to trade mark identification code.
Description
Technical field
The present invention relates to alloy detection field, more particularly to a kind of alloy designations identification side based on artificial neural network
Method.
Background technology
The trade mark is recognized, is referred to and is contained the actual measurement constituent content information and the element of several trades mark in trade mark storehouse of detected sample
Measure range information contrast, thus it is speculated that go out to detect the relevant informations such as the trade mark of object.Trade mark identification is with the hair of alloy detection industry
The New function gradually derived is opened up, the function is very easy to the identification and classification of alloy.
Existing trade mark recognition methods species is less, generally divides following several：(1) according to actual measurement constituent content value whether
Positioned at defining whether detection object is the trade mark as defined in the trade mark in content range, this method shortcoming is, if some element
Content can influence to detect the trade mark identification of object beyond the scope (even if exceeding and few).(2) improved method has, and will contain
Amount scope does appropriate amplification and defines whether detection object is the trade mark with multiple " scopes ".(1), (2) two class algorithms have above
Stronger contingency, recognition result easily the influence of examined error, be difficult to debug and safeguard.(3) board based on membership function
Number recognizer, some membership functions are constructed according to different situations, will be surveyed degree of the constituent content in the range of and be quantified,
This method has to be needed to consider the compatibility between different situations membership function in more outstanding accuracy and fault-tolerance, but the adjustment of parameter
Property, therefore not easy care.(4) trade mark recognizer based on coefficient correlation, its general principle calculates public for construction coefficient correlation
The similarity degree of constituent content and certain interval interior particular value of constituent content in trade mark storehouse, the calculation are surveyed in formula, the correlation coefficient charts levies in kind
The shortcoming of method is that in trade mark identification is calculated trace element accounts for identical weight with a great number of elements, causes the trade mark to recognize sometimes
There is mistake.
Trade mark identification problem also has another form：The trade mark identification problem of constituent content reference value is only provided.Also
It is that trade mark storehouse does not provide the element content range information of the trade mark and only provides constituent content reference value.Existing trade mark identification side
Method is not for the problem.
The content of the invention
For above-mentioned deficiency, the present invention provides a kind of alloy designations recognition methods based on artificial neural network, will be artificial
Neutral net is incorporated into trade mark identification problem, using the powerful learning ability of artificial neural network, fault-tolerant ability, casts aside complexity
Analysis process, the identification problem of alloy designations is converted into a black box problem, it is real by the powerful computing capability of computer
Show accurately identifying for alloy designations.
In order to achieve the above object, the technical solution adopted by the present invention is as follows：A kind of alloy based on artificial neural network
Trade mark recognition methods, specifically includes following steps：
(1) foundation of artificial neural network
Artificial neural network is set up according to the size in trade mark storehouse, the artificial neural network include neural network type, the number of plies,
Every layer of neuron number, transmission function, training function；
The type of neutral net is feedforward multitiered network or unity feedback network；
Neutral net has 1 input layer and 1 output layer respectively；In addition, hidden layers numbers are layer 2-3；
Every layer of neuron number is less than more than 10 100, and output layer neuron number is 1 or refreshing with input layer
It is identical through first number；
In terms of the selection of transmission function, hidden layer uses nonlinear transfer function, and output layer uses linear transfer function；
In terms of the selection of training function, training function is L-M optimized algorithms, quasi- newton BP algorithm, tension gradient decline
Method, gradient descent method or Scaled Conjugate Gradient Method.
(2) training of artificial neural network
Training sample is chosen, the constituent content of element is paid close attention in the input of sample for emphasis, and sample is output as the board of the trade mark
Number identification code, training obtains artificial neural network；Wherein pay close attention to element and refer to percentage composition owning more than 0.1% in alloy
Element；
(3) emulation of artificial neural network
The neutral net that the measured data input of detected sample is trained, the output data of neutral net is taken nearby
Whole, the numerical value after rounding is the trade mark identification code of detected sample, and detected sample is quoted finally according to trade mark identification code
The trade mark.
Further, the hidden layer uses nonlinear transfer function, and nonlinear transfer function is logarithmic or tangential type.
Further, there are following 4 kinds of approach in the source of training sample：
1) measured data, i.e., the test result of each trade mark actual sample；
2) some groups of analogue datas generated at random according to each trade mark constituent content interval；
3) according to the constituent content of each trade mark constituent content interval generation some groups of analogue datas incremented by successively；
4) various combination of 3 kinds of approach more than.
Beneficial effects of the present invention are as follows：
1) artificial neural network is incorporated into trade mark identification problem by this method, utilizes the powerful study of artificial neural network
Ability, fault-tolerant ability, cast aside the analysis process of complexity, the identification problem of alloy designations are converted into a black box problem, by
The excellent computing capability of computer, realizes accurately identifying for alloy designations.
2) this method has extremely strong adjustability, from neural network type, the number of plies, every layer of neuron number, transmission letter
Number, training function, can be adjusted to parameters such as the source of sample, sample sizes, so as to constantly extend the suitable of this method
Ying Xing.
Brief description of the drawings
The present invention is described further with reference to the accompanying drawings and examples：
Fig. 1 is flow chart of the invention；
Fig. 2 is the neural network structure figure of one embodiment of the invention；
Fig. 3 is the training error figure of one embodiment of the invention.
Embodiment
The present invention is described further with reference to the accompanying drawings and examples.
As shown in figure 1, a kind of alloy designations recognition methods based on artificial neural network, specifically includes following steps：
(1) foundation of artificial neural network
Artificial neural network is set up according to the size in trade mark storehouse, the artificial neural network include neural network type, the number of plies,
Every layer of neuron number, transmission function, training function；
The type of neutral net is feedforward multitiered network (BP networks) or unity feedback network (Hopfield networks) etc.；
Neutral net has 1 input layer and 1 output layer respectively；In addition, hidden layers numbers are layer 2-3；
Every layer of neuron number is less than more than 10 100, and appropriate tune is done according to the type and the number of plies of neutral net
It is whole；Input layer number depends in specific trade mark identification problem paying close attention to element number；Output layer neuron
Number be 1 (content is trade mark identification code (being corresponded with the trade mark)) or identical with input layer number (content is attached most importance to
The reference content of element is paid close attention to, the trade mark is speculated on this basis)；
In terms of the selection of transmission function, hidden layer uses nonlinear transfer function, and nonlinear transfer function is logarithmic
Or tangential type (tan-sigmoid) (log-sigmoid)；Output layer uses linear transfer function (purelin)；The hidden layer makes
Use nonlinear transfer function.
In terms of the selection of training function, training function is L-M optimized algorithms, quasi- newton BP algorithm, tension gradient decline
Method, gradient descent method or Scaled Conjugate Gradient Method.
(2) training of artificial neural network
The training of artificial neural network needs great amount of samples.There are following 4 kinds of approach in the source of sample：
1) measured data, i.e., the test result of each trade mark actual sample；
2) some groups of analogue datas generated at random according to each trade mark constituent content interval；
3) according to the constituent content of each trade mark constituent content interval generation some groups of analogue datas incremented by successively；
4) various combination of 3 kinds of approach more than.
The constituent content of element is paid close attention in the input of sample for emphasis, and sample is output as the trade mark identification code of the trade mark, trains
Obtain artificial neural network；Wherein pay close attention to element and refer to all elements that percentage composition in alloy is more than 0.1%；
Number of samples is easily restrained at least in the training of artificial neural network, and time consumption for training is few, but network adaptability is poor；Instead
May not then restrain and (not reach expected error level).Therefore repetition test is needed when training artificial neural network, control
Rational number of samples.
(3) emulation of artificial neural network
The neutral net that the measured data input of detected sample is trained, the output data of neutral net is taken nearby
Whole, the numerical value after rounding is the trade mark identification code of detected sample, and detected sample is quoted finally according to trade mark identification code
The trade mark.
Embodiment：
The present embodiment establishes a feedforward network, and the artificial neural network has 1 input layer, 1 output layer, 2
Hidden layer (every layer of neuron number is 20), hidden layer transmission function uses tanh S types transfer function (tansig), output layer
Transmission function uses linear transfer function (purelin), and training function uses Scaled Conjugate Gradient Method (trainscg), such as Fig. 2
It is shown.
Samples sources are in measured data (i.e. the test result of each trade mark actual sample in trade mark storehouse) in the present embodiment.Its institute
The trade mark storehouse being related to has 7 trades mark, and trade mark title is respectively：Trade mark A, trade mark B, trade mark C, trade mark D, trade mark E, trade mark F, board
Number G.Trade mark identification code is respectively 1,2,3,4,5,6,7.To 50 (45 surveys therein of each sample test in the trade mark storehouse
Examination data input as neutral net, in addition 5 test datas be used for result verification).Sample input is that a length is 10 (weights
Point concern element number is 10) vector, the vectorial each element characterizes the content value of element, and output data is trade mark mark
Know code.
The maximum iteration for designing the artificial neural network is 10000 times, and mean square error target is set to 10-8, will be each
Any 45 test results of sample are trained as input to artificial neural network, as shown in figure 3, after training terminates,
Square error is 5.24*10-8.
Emulated other 5 test datas of each trade mark as input respectively, after output data is rounded nearby
Judge whether the numerical value after rounding is consistent with trade mark identification code, then calculates relative error.Result of calculation is shown in Table 1.
The simulation result of table 1 and data analysis
Simulation result from 1,35 test datas of table is correct, and relative error can control 0.00125 with
Interior, the trade mark identification that the neutral net is used for the trade mark storehouse has extremely strong reliability.
Claims (3)
1. a kind of alloy designations recognition methods based on artificial neural network, it is characterised in that specifically include following steps：
(1) foundation of artificial neural network
Artificial neural network is set up according to the size in trade mark storehouse, the type, the number of plies, every layer of nerve of the artificial neural network is designed
First number, transmission function, training function；
The type of neutral net is feedforward multitiered network or unity feedback network；
Neutral net has 1 input layer and 1 output layer respectively；In addition, hidden layers numbers are layer 2-3；
Every layer of neuron number is less than more than 10 100, and output layer neuron number is 1 or and input layer
Number is identical；
In terms of the selection of transmission function, hidden layer uses nonlinear transfer function, and output layer uses linear transfer function；
In terms of the selection of training function, training function is L-M optimized algorithms, quasi- newton BP algorithm, tension gradient descent method, ladder
Spend descent method or Scaled Conjugate Gradient Method.
(2) training of artificial neural network
Training sample is chosen, the constituent content of element is paid close attention in the input of sample for emphasis, and sample is output as the trade mark mark of the trade mark
Know code, training obtains artificial neural network；Wherein pay close attention to element and refer to all members that percentage composition in alloy is more than 0.1%
Element；
(3) emulation of artificial neural network
The neutral net that the measured data input of detected sample is trained, the output data of neutral net is rounded nearby,
Numerical value after rounding is the trade mark identification code of detected sample, and the board of detected sample is quoted finally according to trade mark identification code
Number.
2. the alloy designations recognition methods according to claim 1 based on artificial neural network, it is characterised in that described hidden
Layer uses nonlinear transfer function, and nonlinear transfer function is logarithmic or tangential type.
3. the alloy designations recognition methods according to claim 1 based on artificial neural network, it is characterised in that training sample
There are following 4 kinds of approach in this source：
1) measured data, i.e., the test result of each trade mark actual sample；
2) some groups of analogue datas generated at random according to each trade mark constituent content interval；
3) according to the constituent content of each trade mark constituent content interval generation some groups of analogue datas incremented by successively；
4) various combination of 3 kinds of approach more than.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710408317.2A CN107247993A (en) | 2017-06-02 | 2017-06-02 | Alloy designations recognition methods based on artificial neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710408317.2A CN107247993A (en) | 2017-06-02 | 2017-06-02 | Alloy designations recognition methods based on artificial neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107247993A true CN107247993A (en) | 2017-10-13 |
Family
ID=60017733
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710408317.2A Pending CN107247993A (en) | 2017-06-02 | 2017-06-02 | Alloy designations recognition methods based on artificial neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107247993A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110389570A (en) * | 2018-04-19 | 2019-10-29 | 株洲中车时代电气股份有限公司 | A kind of locomotive traction system trouble-shooter and method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101793886A (en) * | 2010-03-24 | 2010-08-04 | 北京林业大学 | Sensory evaluation method for predicting fermented yogurt based on BP neural network |
| CN105678329A (en) * | 2016-01-04 | 2016-06-15 | 聚光科技(杭州)股份有限公司 | Method for identifying designations |
-
2017
- 2017-06-02 CN CN201710408317.2A patent/CN107247993A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101793886A (en) * | 2010-03-24 | 2010-08-04 | 北京林业大学 | Sensory evaluation method for predicting fermented yogurt based on BP neural network |
| CN105678329A (en) * | 2016-01-04 | 2016-06-15 | 聚光科技(杭州)股份有限公司 | Method for identifying designations |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110389570A (en) * | 2018-04-19 | 2019-10-29 | 株洲中车时代电气股份有限公司 | A kind of locomotive traction system trouble-shooter and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bennett et al. | Characterising performance of environmental models | |
| van Zyl | The psychometric properties of the emotional quotient inventory 2.0 in South Africa | |
| CN106874686A (en) | Index network of personal connections construction method | |
| CN110533007A (en) | A kind of vehicle-mounted strain of bridge influences the identification of line feature intelligent and extracting method | |
| Palaghianu | A tool for computing diversity and consideration on differences between diversity indices | |
| CN112836937A (en) | Flood disaster loss evaluation method based on entropy weight and BP neural network technology | |
| Prasanalakshmi et al. | Classification and prediction of student academic performance in king khalid university-a machine learning approach | |
| CN109685321A (en) | Event risk method for early warning, electronic equipment and medium based on data mining | |
| Turhan et al. | Estimation of student success with artificial neural networks | |
| Yin et al. | Concordance between ACT and ITED scores from different populations | |
| Richards et al. | Optimisation of competition indices using simulated annealing and artificial neural networks | |
| Leauprasert et al. | Intelligent machine learning techniques for condition assessment of power transformers | |
| Takekawa et al. | Role of layers and neurons in deep learning with the rectified linear unit | |
| CN107247993A (en) | Alloy designations recognition methods based on artificial neural network | |
| Kong et al. | A backpropagation neural network for sales forecasting | |
| Wongkhamdi et al. | A comparison of classical discriminant analysis and artificial neural networks in predicting student graduation outcomes | |
| Dorans | Contributions to the quantitative assessment of item, test, and score fairness | |
| CN110321951A (en) | A kind of VR simulated flight device evaluation of training method | |
| CN109272262A (en) | A kind of analysis method of natural language feature | |
| Derrick et al. | Methods for comparing the responses from a Likert question, with paired observations and independent observations in each of two samples | |
| Music et al. | Classification of leaf type using multilayer perceptron, naive bayes and support vector machine classifiers | |
| Benzer et al. | Examination of international PISA test results with artificial neural networks and regression methods | |
| Spagnolo et al. | Forensic metrology: uncertainty of measurements in forensic analysis | |
| Ginting et al. | 360 DEGREES AND K-NEAREST NEIGHBOUR METHODS FOR LECTURER PERFORMANCE APPRAISAL SYSTEMS: 360 DEGREES AND K-NEAREST NEIGHBOUR METHODS FOR LECTURER PERFORMANCE APPRAISAL SYSTEMS | |
| Golia | Differential Item Functioning classification for polytomously scored items |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171013 |