CN109983482A - Learning model generation method, learning model generating means, signal data method of discrimination, signal data discriminating gear and signal data discriminating program - Google Patents
Learning model generation method, learning model generating means, signal data method of discrimination, signal data discriminating gear and signal data discriminating program Download PDFInfo
- Publication number
- CN109983482A CN109983482A CN201780070171.3A CN201780070171A CN109983482A CN 109983482 A CN109983482 A CN 109983482A CN 201780070171 A CN201780070171 A CN 201780070171A CN 109983482 A CN109983482 A CN 109983482A
- Authority
- CN
- China
- Prior art keywords
- signal data
- distance
- mapping
- signal
- data
- 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
- 238000000034 method Methods 0.000 title claims description 54
- 238000013507 mapping Methods 0.000 claims abstract description 139
- 230000002159 abnormal effect Effects 0.000 claims abstract description 75
- 238000012549 training Methods 0.000 claims abstract description 72
- 241001269238 Data Species 0.000 claims abstract description 29
- 238000005070 sampling Methods 0.000 claims description 43
- 230000006870 function Effects 0.000 claims description 25
- 230000004069 differentiation Effects 0.000 claims description 24
- 238000004364 calculation method Methods 0.000 claims description 20
- 230000008569 process Effects 0.000 claims description 20
- 239000000284 extract Substances 0.000 claims description 7
- 239000002131 composite material Substances 0.000 claims description 5
- 230000007488 abnormal function Effects 0.000 claims description 2
- 238000012937 correction Methods 0.000 claims description 2
- 230000009897 systematic effect Effects 0.000 claims 2
- 238000012905 input function Methods 0.000 claims 1
- 238000013075 data extraction Methods 0.000 abstract description 2
- 238000001514 detection method Methods 0.000 description 18
- 238000013527 convolutional neural network Methods 0.000 description 17
- 238000012545 processing Methods 0.000 description 14
- 230000000007 visual effect Effects 0.000 description 10
- 238000010801 machine learning Methods 0.000 description 9
- 230000005856 abnormality Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000011176 pooling Methods 0.000 description 5
- 238000000605 extraction Methods 0.000 description 4
- 238000002360 preparation method Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000007429 general method Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 230000003647 oxidation Effects 0.000 description 1
- 238000007254 oxidation reaction Methods 0.000 description 1
- 230000000149 penetrating effect Effects 0.000 description 1
- 230000005236 sound signal Effects 0.000 description 1
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
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
Abstract
The signal data discriminating gear that variation strong robustness for non-intrinsically safes such as light source variations is provided and can accurately be differentiated with learning data amount less than in the past.Have: signal data input unit is inputted as the signal data for differentiating object；Feature Mapping generating unit, using according to be attached for it is normal or be abnormal training signal the learner that learn in advance of multiple sampled datas, generate Feature Mapping about signal data extraction characteristic quantity；Distance mapping generating unit, using the data of the multiple Feature Mappings generated according to the multiple sampled datas for being attached normal training signal and the Feature Mapping of signal data, the difference of Feature Mapping is taken between each sampled data and the combination of signal data to generate distance mapping；Distance value operational part finds out the distance between signal data and sampled data value according to distance mapping；And signal data judegment part, differentiate that the signal data is normal or abnormal according to distance value.
Description
Technical field
The present invention relates to for being applied to differentiate that signal data is normal goes back as visual examination based on view data
It is the generation of the learning model of abnormal discriminating gear and the signal data discriminating gear for applying learning model.
Background technique
In the past, it in order to cope with visual examination, the abnormality detection of product, has studied based on image obtained from shooting product
Inspection technique.Generally have by the way that the image of object product to be compared to differentiation object and produce with image obtained from shooting normal product
Product are methods as normal product or abnormal article, and as method of discrimination at this time, as representative example, there are difference inspections
Survey method and machine learning method.
Differential Detection method is not limited to visual examination and the method that is determined according to the difference of 2 similar signals.?
The various of visual examination, abnormality detection etc. can be coped with by the slight essential difference between 2 signals of detection by knowing
Problem.For example, identify object product whether include the visual examination of damage the problem of in, by measurement " to normal product into
Difference between image obtained from row camera shooting " and " from image obtained from the camera shooting of same viewpoint ", can determine the object of object
Whether body has damage.Identify that the whether normal abnormality detection problem of product of object can be by comparing " from sensor acquirement
Normal signal " and the difference of " current signal " solve.The methodology one of difference between such comparison two samplings
As be referred to as Differential Detection (Change detection).It can solve the problems, such as to be extensive by Differential Detection, as example
Son can enumerate the animation identification of supervision camera, long-range induction, automatic Pilot etc..
Machine learning method is learnt according to a large amount of study sampling for being attached normal or abnormal label information
The data of identification object are the model of normal value or exceptional value and the problem for being applied to visual examination, abnormality detection etc.
Method.Although in order to learn to need a large amount of sampled data and operation time, as long as once generating the study for completing study
Model is then able to carry out judgement with high accuracy.
As such technology for visual examination, such as surface quality described in Patent Document 1 has been proposed and comments
Valence device.The surface quality evaluating apparatus described in Patent Document 1 is for it can be considered that being formed in the oxidation overlay film on surface
The distribution of thickness and the device for automatically evaluating the quality of metal, using the evaluation based on machine learning.
Existing technical literature
Patent document
Patent document 1: Japanese Unexamined Patent Publication 2011-191252 bulletin
Summary of the invention
Above-mentioned Differential Detection method be in the field of computer vision before just known general method, with engineering
In the case that habit method compares, the study of model is not needed largely, so the effort for obtaining study sampling, use can be cut down by having
In assign the time of annotation the advantages of, but as one of the especially important project in Differential Detection, it can enumerate and suitably know
The problem of whether not obtained difference is substantially important difference.For example, there are between two images in visual examination
Difference not merely due to damage generates and the case where generated due to the non-intrinsically safes such as illumination changes, individual difference.Needle
It is the difference strong robustness for each sampling extraction for non-intrinsically safes such as light source variations to the typical method that the problem uses
Characteristic vector and calculate they difference method.But which type of difference is difference important in itself according to place
The content of the problem of reason and it is different, so using most of method of Differential Detection there are problems that according to need for it is each sample into
Exercise to export the pretreated problem of difference appropriate.For example, in the example of visual examination, in order to inhibit light source variation,
The difference between image is taken after having suitably adjusted brightness and contrast.However, this method, which exists, suitably to be pressed down
The shortcomings that making the variation that can not be coped in pretreatment.In addition, there is also be difficult to set the sets of values of the variation for the mankind
It is calculated as pretreated problem appropriate.
On the other hand, above-mentioned machine learning method is different from Differential Detection method, is expected to be obtained by study for light source
The model of the variation strong robustness of non-intrinsically safes such as change, but in order to make such model, need a large amount of learning data and
Indicate which position in signal is the detailed annotation data of exceptional value.In the presence of in order to prepare largely include abnormal position
The problem of practising data and assigning detailed annotation respectively and need great time and cost.
The present invention is to complete in view of the above problems, and its purpose is to provide one kind for non-intrinsically safes such as light source variations
Variation strong robustness and the signal data discriminating gear of differentiation can be accurately proceed with learning data amount less than in the past.
The present invention provides a kind of learning model generation method, differentiates that signal data is normal or abnormal for being applied to
Discriminating gear, which is characterized in that the generation method of learning model includes: Feature Mapping generation step, about being attached
For it is normal or be abnormal training signal multiple sampled datas, extract feature out about each sampled data using learner
Amount is to generate Feature Mapping；Distance mapping generation step, is normally to sample that about the training signal in multiple sampled datas
This combination or training signal is normal sampling and training signal is the combination of abnormal sampling, about the Feature Mapping
Difference is taken to generate distance mapping；Distance value calculation step finds out each combined distance value according to the distance mapping；Distance
It is worth discriminating step, the distance value is compared with scheduled threshold value, differentiates that the combination is less than the training signal of threshold value and is
Training signal more than normal sampling combination with one another or threshold value is normal sampling and training signal is abnormal sampling
Combination；And parameters revision step, so that the training letter of differentiation result and sampled data in the distance value discriminating step
Number the consistent mode of combination, correct the operational parameter of the learner to be learnt.
In addition, the generation method of learning model of the invention is characterized in that, in the distance value calculation step, pass through
Pond function (pooling function) operation distance value according to distance mapping.
In addition, the present invention provides a kind of learning model generating means, differentiate that signal data is normal goes back for being applied to
It is abnormal discriminating gear, which is characterized in that the generating means of learning model include: Feature Mapping generating unit, about quilt
Attached for it is normal or be abnormal training signal multiple sampled datas, extracted out using learner about each sampled data
Characteristic quantity generates Feature Mapping；Distance mapping generating unit is normal sampling about the training signal in multiple sampled datas
Combination with one another or training signal are normal sampling and training signal is the combination of abnormal sampling, are reflected about the feature
It penetrates and takes difference to generate distance mapping；Distance value operational part finds out each combined distance value according to the distance mapping；Distance
It is worth judegment part, the distance value is compared with scheduled threshold value, differentiates that the combination is less than the training signal of threshold value and is positive
Training signal more than normal sampling combination with one another or threshold value is normal sampling and training signal is abnormal sampling
Combination；And parameter correction unit, so that the group of the training signal of differentiation result and sampled data in the distance value judegment part
Consistent mode is closed, corrects the operational parameter of the learner to be learnt.
The present invention provides a kind of signal data method of discrimination characterized by comprising signal data input step, input
As the signal data for differentiating object；Feature Mapping generation step, using according to being attached to be normal or be abnormal instruction
Practice the learner that multiple sampled datas of signal are learnt in advance, extracts characteristic quantity out about the signal data to generate spy
Sign mapping；Distance mapping generation step, uses what is generated according to the multiple sampled datas being attached as normal training signal
The Feature Mapping of the data of multiple Feature Mappings and the signal data generated in the Feature Mapping generation step, is generating
The difference of Feature Mapping is taken between each sampled data of Feature Mapping and the combination of signal data to generate distance mapping；Distance value
Calculation step finds out the distance between signal data and sampled data value according to the distance mapping；And signal data differentiates
Step differentiates that the signal data is normal or abnormal according to the distance value.
The present invention provides a kind of signal data discriminating gear, which is characterized in that has: signal data input unit, and input is made
For the signal data for differentiating object；Feature Mapping generating unit, using according to being attached to be normal or be that abnormal training is believed
Number the learner that is learnt in advance of multiple sampled datas, reflected about signal data extraction characteristic quantity to generate feature
It penetrates；Distance mapping generating unit, uses the multiple spies generated according to the multiple sampled datas being attached as normal training signal
The data of mapping and the Feature Mapping of the signal data generated by the Feature Mapping generating unit are levied, Feature Mapping is being generated
The difference of Feature Mapping is taken between each sampled data and the combination of signal data to generate distance mapping；Distance value operational part, root
The distance between signal data and sampled data value are found out according to the distance mapping；And signal data judegment part, according to described
Distance value differentiates that the signal data is normal or abnormal.
In addition, signal data discriminating gear of the invention is characterized in that, the distance value operational part passes through composite unit
The distance value of distance value calculation process and signal data unit distance value calculation process and operation signal data, the composite unit
Distance value calculation process finds out each combined distance value according to the difference value of each feature of the distance mapping, the letter
Number unit distance value calculation process finds out signal number according to signal data and the distance value of sampled data all combined
According to the distance value of unit.
In addition, signal data discriminating gear of the invention is characterized in that, the combination list in the distance value operational part
In the distance value calculation process of position, by pond function according to apart from mapping operations distance value.
In addition, signal data discriminating gear of the invention is characterized in that, the signal number in the distance value operational part
According to the average of the distance value that in unit distance value calculation process, operation is all combined using average value as the distance of signal data
Value.
According to the present invention, it is directed to without as Differential Detection method being used to cope with the variation of the non-intrinsically safes such as light source variation
The pretreatment of input data, it is not necessary that a large amount of sampling of detailed annotation has been attached as previous machine learning method
Data, as long as being attached normal or abnormal such sampled data simply annotated with amount preparation less than in the past,
It is able to carry out high-precision study, so compared with the past can cut down great time and cost.In addition, by between sampling
Combination make data set, so the hits of data set increases compared to single sampling is used come the case where study.Especially
Be under the situation more considerably less than the quantity of abnormal article with normal condition, it is difficult using single sampling to learn
To prepare the sampling of many abnormal articles, but in the present invention the quantity of the pairing of normal product-abnormal article increase relatively more than normal product-
The quantity of the pairing of normal product.Therefore, even if also can be realized the high-precision of study under the considerably less situation of the quantity of abnormal article
Degreeization, high efficiency.In addition, study can be accurately proceed if the quantity of pairing is more, so compared to previous machine
Hits can be greatly decreased in the method for study.
Detailed description of the invention
Fig. 1 is the block diagram for showing the structure of signal data discriminating gear 10 of the invention.
Fig. 2 is the summary for showing the generation of learning model 17 used in the learner of Feature Mapping generating unit 12
Explanatory diagram.
Fig. 3 is saying for the summary until showing the distance value operation used in the differentiation in signal data discriminating gear 10
Bright figure.
Fig. 4 is the flow chart for showing the process of the generation of learning model.
Fig. 5 is the flow chart for showing the process of differentiation processing of signal data.
(symbol description)
10: signal data discriminating gear；11: signal data input unit；12: Feature Mapping generating unit；13: distance mapping life
At portion；14: distance value operational part；15: signal data judegment part；16: storage unit；17: learning model；18: training signal is attached
Add sampled data；19: signal data.
Specific embodiment
[the 1st embodiment]
Hereinafter, illustrating the example of the signal data discriminating gear of the 1st embodiment referring to attached drawing.
Fig. 1 is the block diagram for showing the structure of signal data discriminating gear 10 of the invention.In addition, signal data discriminating gear
10 are also possible to the device designed as special purpose machinery, and can be realized by general computer.In this case, it is set as
Signal data discriminating gear 10 have CPU that general computer would generally have (Central Processing Unit: in
Entreat arithmetic processing apparatus), GPU (Graphics Processing Unit: image processing apparatus), memory, hard disk drive
Deng storage equipment (diagram is omitted).In addition, can certainly be in order to make these general computers as the signal number of this example
It is functioned according to discriminating gear 10 and various processing is executed by program.
Signal data discriminating gear 10 at least has signal data input unit 11, Feature Mapping generating unit 12, distance mapping
Generating unit 13, distance value operational part 14, signal data judegment part 15 and storage unit 16.
It is normal that there is signal data input unit 11 input, which to want to differentiate in signal data discriminating gear 10 of the invention,
Or the function of abnormal signal data 19.The input of signal data 19 can be by wired mode or wireless mode in real time
It obtains and is inputted every time, can also be determined after by the storage to aftermentioned storage unit 16 of the signal data 19 of acquirement in differentiation
When read.
There is Feature Mapping generating unit 12 signal data 19 about differentiation object to extract signal data out according to learner
Characteristic quantity generate the function of Feature Mapping.As long as Feature Mapping can be generated, it can be arbitrary learner, such as examine
Consider and convolutional neural networks (CNN:Convolutional Neural Network) is used as learner.In the generation of Feature Mapping
In, such as in the case where convolutional neural networks, utilize the feature of input signal data and weight filter or process on the way
Mapping and weight filter and take inner product, process of convolution is repeated using raster scanning, to obtain Feature Mapping.In addition,
Learner in this feature mapping generating unit 12 is learnt in advance can generate such as be accurately proceed and final sentence
Another characteristic mapping, is stored in storage unit 16 as learning model 17.The generation of aftermentioned learning model 17 it is detailed in
Hold.
In addition, in the present invention, in order to differentiate signal data, needing to be attached for normal training signal
Multiple sampled datas also exist for each sampled data of these multiple sampled datas for being attached normal training signal
Feature Mapping is generated in Feature Mapping generating unit 12.Multiple training signal additional sample data 18 are stored in storage unit 16, can
With pre-generated Feature Mapping corresponding with training signal additional sample data 18 and store to storage unit 16, it can also be
The timing of the generation of the Feature Mapping of signal data also generates Feature Mapping for training signal additional sample data 18.
Distance mapping generating unit 13 has following function: believing using according to the training being attached as normal training signal
The data for multiple Feature Mappings that number additional sample data 18 generate and the Feature Mapping generated according to signal data, in training
The difference of Feature Mapping is taken between each sampled data of signal additional sample data 18 and the combination of signal data come generate away from
From mapping.
Distance value operational part 14, which has, to be mapped according to distance and finds out the distance between signal data and sampled data value
Function.
Signal data judegment part 15 has according to by between the signal data found out of distance value operational part 14 and sampled data
Distance value and to differentiate signal data be normal or abnormal function.Specifically, for example considering following method: will be by distance
The distance value that value operational part 14 is found out is compared with preset scheduled threshold value, is then determined as signal if it is less than threshold value
Data are normal, are abnormal if it is signal data more than threshold value is then determined as.In addition, the present invention can not only be applied to differentiate
Normally, abnormal situation, additionally it is possible to, can be using according to distance in the case where classification the case where applied to such as being classified
The method that the size of value classifies to signal data.
Storage unit 16 is stored with the signals number such as learning model 17, training signal additional sample data 18, signal data 19
According to information needed for the processing in discriminating gear 10.
Fig. 2 is the summary for showing the generation of learning model 17 used in the learner of Feature Mapping generating unit 12
Explanatory diagram.Although needing to use learning method corresponding with the learner used, learner herein is as an example, right
It is illustrated using the case where convolutional neural networks (hereinafter referred to as CNN).In addition, for example can be using loss letter in study
Number, can use triple loss function (Triplet loss function), twin loss function (siamese loss function), S
The various loss functions such as shape cross entropy loss function (sigmoid cross entropy loss function).In addition, as study
Last stage, CNN when to initial study properly (such as randomly) allocation of parameters.
In study, it will be attached the multiple sampled datas for the training signal for being normal (0) either abnormal (1) first
(X1~X4) it is separately input to CNN for generating Feature Mapping, generate respective Feature Mapping (h1~h4).In addition, in this example
In son, as long as since normal and exception can be distinguished, so used (0) and (1) this 2 as training signal, but
In the case where being such as classified as multiple projects, the type of training signal can also increase.
Next, for Feature Mapping (h corresponding with sampled data respectively1~h4), it is normally to adopt for training signal
Sample combination with one another or training signal are normal sampling and training signal is the combination of abnormal sampling, about Feature Mapping
Difference is taken to generate distance mapping, is mapped according to distance, each combined distance value (d of operation12、d13、d14、d23、d24).As
It is flat according to the method apart from mapping calculation distance value, such as using global maximum pond (global max pooling), the overall situation
Global pools (the global such as equal pond (global average pooling), the overall situation pond Lp (global Lp pooling)
Pooling) function.
Then, by each combined distance value (d12、d13、d14、d23、d24) compared with preset scheduled threshold value
Compared with the value that the distance value is normal combination with one another being then determined as if it is less than threshold value, if it is more than threshold value being then determined as this
Distance value is normal and abnormal combined value.Compare the training letter of the differentiation result and sampled data for such distance value
Number, evaluate discrimination precision.As an example, consider whether the accuracy rate differentiated becomes the probability more than making a reservation for.Differentiating essence
In the case that degree does not meet predetermined condition, the parameter of CNN is corrected, in addition from most from the beginning of carrying out study processing.Such repetition learning,
When discrimination precision meets predetermined condition, terminates study processing, obtain learning model.
Fig. 3 is saying for the summary until showing the distance value operation used in the differentiation in signal data discriminating gear 10
Bright figure.In the differentiation processing of signal data, first against be utilized learnt in advance obtained from learning model
CNN, input signal data (X) generate Feature Mapping (h).In addition, preparing to be attached to adopt for the multiple of normal training signal
Sample data (X1~XN), for each sampled data of these multiple sampled datas for being attached normal training signal, also exist
Feature Mapping is generated in Feature Mapping generating unit 12, obtains Feature Mapping (h1~hN)。
Next, with sampled data (X1~XN) corresponding Feature Mapping (h1~hN) and it is corresponding with signal data (X)
Between the combination of Feature Mapping (h), the difference of Feature Mapping is taken to generate distance mapping.Then, it is mapped according to the distance of generation,
Such as signal data (X) and each sampled data (X are obtained using global maximum pond function1~XN) the distance between value (d1~
dN).Then, it carries out finding out obtained distance value (d1~dN) the processing such as average value, obtain for used in the differentiation away from
From value d.Distance value d and scheduled threshold value are compared to carry out the differentiation of signal data.
Fig. 4 is the flow chart for showing the process of the generation of learning model.As shown in Fig. 4, in order to for learner
Study processing, it is necessary first to prepare the multiple sampled datas (step S11) for being attached training signal.About multiple hits
According to by training signal be normal sampled data and to be that abnormal sampled data prepares respectively multiple.Then, according to the more of preparation
A sampled data generates Feature Mapping (step S12).For example, carrying out Feature Mapping by extracting multiple characteristic quantities out using CNN
Generation.In addition, the parameter of initial CNN is for example randomly assigned.The parameter of the initial CNN also can be used outside other
See the parameter that the study of other CNN used in the problem of checking finishes.
Next, being normal combination with one another or normal and abnormal combination about training signal, Feature Mapping is taken
Difference maps (step S13) to generate distance.Then, it is mapped according to distance, each combined distance value (step S14) of operation.
For example, carrying out the operation of distance value at this time using global maximum pond function.By the distance value found out and scheduled threshold value into
Row compares, and normal, abnormal (step S15) is differentiated for each combination.
Finally, each combined differentiation result is compared with training signal, discrimination precision (step S16) is found out.So
Afterwards, differentiate whether discrimination precision meets scheduled condition (step S17).Here, not meeting scheduled condition in discrimination precision
In the case where, the parameter (step S18) for generating the learner such as CNN of Feature Mapping is corrected, later, carries out step S12
The study of~S16 is handled.In the case where being determined as that discrimination precision meets scheduled condition in step S17, the time is obtained
The learning model that is constituted of parameter of learner under point and terminate.
Fig. 5 is the flow chart for showing the process of differentiation processing of signal data.As shown in Fig. 5, for signal data
Differentiation processing, carries out the input (step S21) of signal data first.Then, using the learner example that learning model is utilized
Such as CNN, the Feature Mapping (step S22) about signal data is generated.For example, by using CNN extract out multiple characteristic quantities come into
The generation of row Feature Mapping.In addition, prepare to be attached multiple sampled datas for normal training signal, it is attached for these
Multiple sampled datas of normal training signal have been added also to generate Feature Mapping respectively.
Next, being that normally the respective Feature Mapping of multiple sampled datas and the feature of signal data are reflected in training signal
Difference is taken between penetrating, and generates each combined distance mapping (step S23).Then, each for each combination, according to distance
It maps and operation distance value (step S24).It is, for example, possible to use global maximum pond functions to carry out operation at this time.According to
To the distance value all combined, determine the distance value (step S25) of signal data.Here, for example carrying out all combining
Operation of the average value of distance value as the distance value of signal data.
Finally, the distance value of signal data is compared with scheduled threshold value, differentiate that signal data is normal or different
Often (step S26).For example, being determined as signal data is normally, if it is threshold if the distance value of signal data is less than threshold value
More than value, then it is abnormal for being determined as signal data.Then, output differentiates result and terminates (step S27).
The signal data discriminating gear 10 of this example being made of structure as described above has effect below.
(1) different from the method for previous Differential Detection, it is set as by being based on machine learning to the setting of feature extraction section
Learner and feature extraction is carried out by learning model, can so if carrying out feature extraction by learning model
Enough differentiations for realizing the variation strong robustness for non-intrinsically safes such as light source variations.
(2) different from the method for previous machine learning for directly providing sampling and identifying, by expressly catching and just
The difference of Chang Pin, even slight difference also can accurately be caught.
(3) only by whether including that simple annotation information as abnormal position is just able to carry out study in sampling.
That is, without annotating abnormal position in detail for each sampling, so being easy to carry out the preparation of the sampled data for study.
(4) data set is made by the combination between sampling, so compared with using the case where single sampling is to learn,
The hits of data set increases.Especially under the situation more considerably less than the quantity of abnormal article with normal condition, using single
Sampling learnt in the case where, it is difficult to prepare the sampling of many abnormal articles, but normal product-abnormal article is matched in the present invention
Pair quantity increase relatively more than the normal product of normal product-pairing quantity.Therefore, even if the shape considerably less in the quantity of abnormal article
Under condition, high precision int, the high efficiency of study also can be realized.In addition, can be accurately proceed if the quantity of pairing is more
Study, so hits can be greatly decreased compared with the method for previous machine learning.
That is, according to the signal data discriminating gear of this example, without being used to cope with light source variation as Differential Detection method
The pretreatment for input data of the variation of equal non-intrinsically safes, it is not necessary that be attached as previous machine learning method
The a large amount of sampled data of detailed annotation, as long as being attached normal or abnormal such letter with amount preparation less than in the past
The sampled data of single annotation, it will be able to carry out high-precision study, so it is compared with the past can cut down the great time and
Cost.
In the first embodiment described above, as shown in figure 3, be set as normal training signal additional sample data respectively with
Operation distance value between signal data finds out being averaged for the distance value all combined, is inferred according to the average value of distance value
It handles and is illustrated.It, can not only be using average but in the deduction, additionally it is possible to which various methods are applied according to situation.
For example, can also after finding out Feature Mapping using with based on local outlier factor (Local Outlier Factor) away from
It leaves school just associated abnormality detection system.In addition, if each sampling in data set and the distance between unknown sampling are considered as
The fractional function of weak identifier, then can also be using the study of the set such as pack (bagging).
In the first embodiment described above, it is set as carrying out differentiating that signal data is normal or abnormal inference process
Example and illustrate signal data discriminating gear 10, but can also carry out as signal data is classified as multiple projects
Inference process.In addition, as signal data, it also can be using various signal datas as object.It e.g. can be in base
In shoot the visual examination of image obtained from product, the detection of abnormal sound based on sound signal data, certification system, face
The technology applied in the various fields such as the differentiation and classification of the feature of the face in system.
[the 2nd embodiment]
In the first embodiment described above, in order to differentiate that the signal data of object is normal or abnormal, such as it is set as to ask
Distance value out is compared with preset scheduled threshold value, be then determined as if it is less than threshold value signal data be it is normal,
It is abnormal if it is signal data more than threshold value is then determined as, but the present invention can not only differentiate that the signal data of object is normal
It is or abnormal, additionally it is possible to be extended to and produce the exception of which type of type if it is abnormal then output.In this case, can
In a manner of exporting multiple distance values corresponding with the type of the exception according to signal data extension signal data differentiate
Portion 15 copes with.
Specifically, training signal in the case of only determining normal or abnormal, such as normal (0) and abnormal (1)
Distinguished in this way with 2 values, but can also according to abnormal type such as abnormal (1), abnormal (2) ..., abnormal (k) in this way differentiation instruction
Practice signal.In addition, prepare it is as shown in Figure 3 for obtaining multiple structures of distance value, as based on normal (0) and exception (1)
The distance value (1) of the comparison of training signal additional sample data and signal data, the training letter based on normal (0) and abnormal (2)
The distance value (2) of the comparison of number additional sample data and signal data ..., the training signal based on normal (0) and (k) extremely it is attached
Add the distance value (k) of the comparison of sampled data and signal data in this way, calculating distance value for abnormal each type.For
The each distance value of distance value (1)~distance value (k), such as by being compared to discriminate whether comprising the type with threshold value
It is abnormal.If being judged as normally, distinguishing that the signal data is normally, if about any about all abnormal types
Abnormal type and be judged as being exception, then distinguish the exception of the signal data and can determine abnormal type.
In addition, can also identify this point comprising multiple exceptions.
That is, the signal data illustrated in the 1st embodiment is arranged in and differentiates for the abnormal each type for differentiating object
Signal data input unit 11, Feature Mapping generating unit 12 in device 10, distance mapping generating unit 13, distance value operational part 14 with
And the structure of signal data judegment part 15, and each Feature Mapping generation is carried out using sampled data corresponding with abnormal type
The study of learner in portion 12, can be according to the differentiation of multiple signal data judegment parts of the quantity setting by abnormal type
As a result differentiate whether signal data is type that is normal and determining exception in an exceptional case.
Claims (10)
1. a kind of learning model generation method differentiates that signal data is normal or abnormal discriminating gear for being applied to,
Wherein, the generation method of learning model includes:
Feature Mapping generation step, about be attached for it is normal or be abnormal training signal multiple sampled datas, make
Characteristic quantity is extracted out to generate Feature Mapping about each sampled data with learner；
Distance mapping generation step is normal sampling combination with one another or instruction about the training signal in multiple sampled datas
Practice that signal is normal sampling and training signal is the combination of abnormal sampling, take difference about the Feature Mapping generate away from
From mapping；
Distance value calculation step finds out each combined distance value according to the distance mapping；
The distance value is compared with scheduled threshold value, differentiates that the combination is less than the instruction of threshold value by distance value discriminating step
Practicing the training signal that signal is normal sampling combination with one another or threshold value or more is that normally sampling and training signal are different
The combination of normal sampling；And
Parameters revision step, so that the combination of the training signal of differentiation result and sampled data in the distance value discriminating step
Consistent mode corrects the operational parameter of the learner to be learnt.
2. learning model generation method according to claim 1, wherein
In the distance value calculation step, by pond function according to apart from mapping operations distance value.
3. a kind of learning model generating means differentiate that signal data is normal or abnormal discriminating gear for being applied to,
Wherein, the generating means of learning model have:
Feature Mapping generating unit, about be attached for it is normal or be abnormal training signal multiple sampled datas, use
Learner extracts characteristic quantity out about each sampled data to generate Feature Mapping；
Distance mapping generating unit is normal sampling combination with one another or training about the training signal in multiple sampled datas
Signal is normal sampling and training signal is the combination of abnormal sampling, takes difference about the Feature Mapping to generate distance
Mapping；
Distance value operational part finds out each combined distance value according to the distance mapping；
The distance value is compared with scheduled threshold value, differentiates that the combination is less than the training of threshold value by distance value judegment part
Signal is that the training signal of normal sampling combination with one another or threshold value or more is normal sampling and training signal is abnormal
Sampling combination；And
Parameter correction unit, so that the differentiation result in the distance value judegment part is consistent with the combination of the training signal of sampled data
Mode, correct the operational parameter of the learner to be learnt.
4. a kind of signal data method of discrimination, comprising:
Signal data input step is inputted as the signal data for differentiating object；
Feature Mapping generation step, using according to be attached for it is normal or be abnormal training signal multiple sampled datas
The learner learnt in advance extracts characteristic quantity out about the signal data to generate Feature Mapping；
Distance mapping generation step, it is multiple using being generated according to the multiple sampled datas being attached as normal training signal
The Feature Mapping of the data of Feature Mapping and the signal data generated in the Feature Mapping generation step, is generating feature
The difference of Feature Mapping is taken between each sampled data of mapping and the combination of signal data to generate distance mapping；
Distance value calculation step finds out the distance between signal data and sampled data value according to the distance mapping；And
Signal data discriminating step differentiates that the signal data is normal or abnormal according to the distance value.
5. a kind of signal data discriminating gear, wherein have:
Signal data input unit is inputted as the signal data for differentiating object；
Feature Mapping generating unit, using according to be attached for it is normal or be abnormal training signal multiple sampled datas it is pre-
The learner first learnt extracts characteristic quantity out about the signal data to generate Feature Mapping；
Distance mapping generating unit, uses the multiple spies generated according to the multiple sampled datas being attached as normal training signal
The data of mapping and the Feature Mapping of the signal data generated by the Feature Mapping generating unit are levied, Feature Mapping is being generated
The difference of Feature Mapping is taken between each sampled data and the combination of signal data to generate distance mapping；
Distance value operational part finds out the distance between signal data and sampled data value according to the distance mapping；And
Signal data judegment part differentiates that the signal data is normal or abnormal according to the distance value.
6. signal data discriminating gear according to claim 5, wherein
The distance value operational part by composite unit distance value calculation process and signal data unit distance value calculation process come
The distance value of operation signal data, the composite unit distance value calculation process map according to the distance and find out each combination
Distance value, the signal data unit distance value calculation process according to signal data and the distance value of sampled data all combined and
Find out the distance value of signal data unit.
7. signal data discriminating gear according to claim 6, wherein
In the composite unit distance value calculation process in the distance value operational part, transported by pond function according to distance mapping
Calculate distance value.
8. according to signal data discriminating gear described in claim 6 or 7, wherein
In the signal data unit distance value calculation process in the distance value operational part, distance value that operation is all combined
Averagely and using average value as the distance value of signal data.
9. a kind of signal data discriminating gear, wherein
For the abnormal each type for differentiating object, signal number described in any one in the claim 5 to 8 is set
According to signal data input unit, Feature Mapping generating unit, distance mapping generating unit, distance value operational part and the letter in discriminating gear
The structure of number judegment part, and carried out in each Feature Mapping generating unit by sampled data corresponding with abnormal type
The study of learner can be sentenced according to the differentiation result of multiple signal data judegment parts of the quantity setting by abnormal type
Whether level signal data are type that is normal and determining exception in an exceptional case.
10. a kind of signal data discriminating program differentiates that the signal data of object is normal for differentiating for realizing computer
Or abnormal function, wherein
The signal data discriminating program implements function such as the computer:
Signal data input function is inputted as the signal data for differentiating object；
Feature Mapping systematic function, using according to be attached for it is normal or be abnormal training signal multiple sampled datas
The learner learnt in advance extracts characteristic quantity out about the signal data to generate Feature Mapping；
Distance mapping systematic function, it is multiple using being generated according to the multiple sampled datas being attached as normal training signal
The Feature Mapping of the data of Feature Mapping and the signal data generated by the Feature Mapping generating unit, is generating Feature Mapping
Each sampled data and signal data combination between take the difference of Feature Mapping generate distance mapping；
Distance value calculation function finds out the distance between signal data and sampled data value according to the distance mapping；And
Signal data discrimination function differentiates that the signal data is normal or abnormal according to the distance value.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2016-222489 | 2016-11-15 | ||
| JP2016222489A JP6216024B1 (en) | 2016-11-15 | 2016-11-15 | Trained model generation method and signal data discrimination device |
| PCT/JP2017/040841 WO2018092747A1 (en) | 2016-11-15 | 2017-11-14 | Learned model generation method, learned model generation device, signal data discrimination method, signal data discrimination device, and signal data discrimination program |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109983482A true CN109983482A (en) | 2019-07-05 |
Family
ID=60107420
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201780070171.3A Pending CN109983482A (en) | 2016-11-15 | 2017-11-14 | Learning model generation method, learning model generating means, signal data method of discrimination, signal data discriminating gear and signal data discriminating program |
Country Status (4)
| Country | Link |
|---|---|
| JP (1) | JP6216024B1 (en) |
| CN (1) | CN109983482A (en) |
| DE (1) | DE112017005011T5 (en) |
| WO (1) | WO2018092747A1 (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019229910A1 (en) * | 2018-05-30 | 2019-12-05 | 株式会社ウフル | Immigration inspection system in which face authentication is used, immigration inspection method, program, and authentication device |
| JP7070157B2 (en) * | 2018-06-29 | 2022-05-18 | 富士通株式会社 | Image processing program, image processing device and image processing method |
| JP7046760B2 (en) * | 2018-08-23 | 2022-04-04 | アンリツ株式会社 | Signal analysis device, signal analysis method, and signal analysis program |
| US11210335B2 (en) | 2018-09-27 | 2021-12-28 | Optim Corporation | System and method for judging situation of object |
| EP3637303B1 (en) * | 2018-10-09 | 2024-02-14 | Naver Corporation | Methods for generating a base of training images, for training a cnn and for detecting a poi change in a pair of inputted poi images using said cnn |
| JP2020112483A (en) * | 2019-01-15 | 2020-07-27 | 春江 姚 | Exterior appearance inspection system, calculation model construction method and calculation model construction program |
| JPWO2021002465A1 (en) * | 2019-07-03 | 2021-01-07 | ||
| WO2021033239A1 (en) * | 2019-08-19 | 2021-02-25 | 富士通株式会社 | Information processing device, information processing program, and information processing method |
| US20230306259A1 (en) | 2020-08-17 | 2023-09-28 | Nippon Telegraph And Telephone Corporation | Information processing apparatus, information processing method and program |
| JP7391907B2 (en) | 2021-03-16 | 2023-12-05 | 株式会社東芝 | Anomaly detection device, anomaly detection method, and anomaly detection program |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH1145337A (en) * | 1997-07-28 | 1999-02-16 | Matsushita Electric Works Ltd | Coating system |
| JP2001022926A (en) * | 1999-07-06 | 2001-01-26 | Mitsubishi Electric Corp | Method and device for image recognition |
| JP2002259948A (en) * | 2001-03-02 | 2002-09-13 | Tdk Corp | Next process determining method, inspection method and inspection device |
| JP2005156334A (en) * | 2003-11-25 | 2005-06-16 | Nec Tohoku Sangyo System Kk | Pseudo defective image automatic creation device and imaging inspection device |
| JP5546317B2 (en) * | 2010-03-31 | 2014-07-09 | 株式会社デンソーアイティーラボラトリ | Visual inspection device, visual inspection discriminator generation device, visual inspection discriminator generation method, and visual inspection discriminator generation computer program |
| JP6309887B2 (en) * | 2014-12-19 | 2018-04-11 | ヤフー株式会社 | Model generation apparatus, information distribution apparatus, model generation method, information distribution method, model generation program, and information distribution program |
| US9436895B1 (en) * | 2015-04-03 | 2016-09-06 | Mitsubishi Electric Research Laboratories, Inc. | Method for determining similarity of objects represented in images |
-
2016
- 2016-11-15 JP JP2016222489A patent/JP6216024B1/en active Active
-
2017
- 2017-11-14 WO PCT/JP2017/040841 patent/WO2018092747A1/en active Application Filing
- 2017-11-14 DE DE112017005011.1T patent/DE112017005011T5/en active Pending
- 2017-11-14 CN CN201780070171.3A patent/CN109983482A/en active Pending
Also Published As
| Publication number | Publication date |
|---|---|
| JP6216024B1 (en) | 2017-10-18 |
| WO2018092747A1 (en) | 2018-05-24 |
| DE112017005011T5 (en) | 2019-07-04 |
| JP2018081442A (en) | 2018-05-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109983482A (en) | Learning model generation method, learning model generating means, signal data method of discrimination, signal data discriminating gear and signal data discriminating program | |
| CN111179251B (en) | Defect detection system and method based on twin neural network and by utilizing template comparison | |
| CN101915769B (en) | Automatic optical inspection method for printed circuit board comprising resistance element | |
| FI3452959T3 (en) | Model construction in a neural network for object detection | |
| CN110069972A (en) | Automatic detection real world objects | |
| CN109284355B (en) | Method and device for correcting oral arithmetic questions in test paper | |
| CN110992349A (en) | Underground pipeline abnormity automatic positioning and identification method based on deep learning | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN110288032B (en) | Vehicle driving track type detection method and device | |
| CN105654066A (en) | Vehicle identification method and device | |
| CN107358264A (en) | A kind of method that graphical analysis is carried out based on machine learning algorithm | |
| JP2018026122A5 (en) | ||
| CN107194908A (en) | Image processing apparatus and image processing method | |
| CN108932712A (en) | A kind of rotor windings quality detecting system and method | |
| CN113761259A (en) | Image processing method and device and computer equipment | |
| CN108573289A (en) | Information processing unit, information processing method and recording medium | |
| CN115019294A (en) | Pointer instrument reading identification method and system | |
| CN105426926B (en) | A kind of couple of AMOLED carries out the method and device of detection classification | |
| Pillai | Student Engagement Detection in Classrooms through Computer Vision and Deep Learning: A Novel Approach Using YOLOv4 | |
| CN109102486B (en) | Surface defect detection method and device based on machine learning | |
| CN117115147A (en) | Textile detection method and system based on machine vision | |
| CN106096634B (en) | Fault detection method based on Adaptive windowing mental arithmetic method with interval halving algorithm | |
| KR20210122429A (en) | Method and System for Artificial Intelligence based Quality Inspection in Manufacturing Process using Machine Vision Deep Learning | |
| CN110017998A (en) | Vehicle checking method, device and equipment | |
| CN113450323B (en) | Quality detection method and device, electronic equipment and computer readable storage medium |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190705 |
|
| WD01 | Invention patent application deemed withdrawn after publication |