CN109345506A - A kind of hot spot based on convolutional neural networks and MARFE automatic testing method - Google Patents
A kind of hot spot based on convolutional neural networks and MARFE automatic testing method Download PDFInfo
- Publication number
- CN109345506A CN109345506A CN201810967404.6A CN201810967404A CN109345506A CN 109345506 A CN109345506 A CN 109345506A CN 201810967404 A CN201810967404 A CN 201810967404A CN 109345506 A CN109345506 A CN 109345506A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- layer
- convolutional neural
- feature
- marfe
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of hot spots based on convolutional neural networks and MARFE automatic testing method, include the following steps: input picture, image is pre-processed, feature extraction layer extracts the feature of image, feature is gone to classify using classifier, output category result.The invention proposes one kind can detect whether occur the method for Hot Spot and MARFE in plasma discharge processes automatically, reduces convolution kernel in design as far as possible, improves calculating speed.
Description
Technical field
The invention belongs to field of image processings more particularly to a kind of automatic based on the hot spot of convolutional neural networks and MARFE
Detection method.
Background technique
Tokamak device is a kind of fusion facility being widely studied at present.In order to realize the thermonuclear fusion that can be constrained,
Each main economic body has put into a large amount of fund on tokamak device.Currently, tokamak is undergoing by small-sized dress
Set the process that large-scale plant arrives Commercial Demonstration heap again.As tokamak device becomes greatly, plasma in device
Can be also more and more in the plasma of constraint, the thermic load on vacuum-chamber wall and its component is also increasing.The phase from device
The picture that machine takes, it can be seen that produce apparent hot spot in these positions.In general, experimenter also can be existing according to these
As judging whether to produce Hot Spot phenomenon.It but is a kind of extremely inefficient plan by human eye detection Hot spot phenomenon
Slightly.When experimenter needs to find some paradoxical discharge big guns for Hot Spot occur, one from vast as the open sea image data
Lookup is opened obviously extremely to waste time.Moreover, probably generating nearly million in present each round experiment by taking EAST as an example
Image.Paradoxical discharge big gun number is searched out from these pictures will become unrealistic.
With the continuous renewal upgrading of hardware resources such as the rise of artificial intelligence and GPU in recent years, it is based on neural network
The convolutional neural networks of model constantly push the development of computer vision field.In fields such as image classification, object detections, base
Very bright eye is showed in the algorithm of convolutional neural networks.With increasingly mature, the training convolutional nerve of convolutional neural networks algorithm
The detection that model carries out phenomena such as Hot Spot, MARFE on tokamak device is possibly realized.It is detected using convolutional neural networks
, mainly there is following consideration in phenomena such as Hot Spot, MARFE: 1, shape when Hot Spot, MARFE occur be it is irregular,
Judged by texture information very difficult；2, under Divertor configuration, the presence of X point to determine Hot by gray scale variation
There is the possibility of erroneous judgement in Spot and MARFE.
Summary of the invention
Phenomena such as in order to accurately detect Hot Spot, MARFE, the invention proposes a kind of based on convolutional neural networks
Hot spot and MARFE automatic testing method.This method, which can be effectively detected out, to be taken in a plasma discharge processes
Image whether there is Hot Spot and MARFE etc..
In order to solve technical problem above, present invention employs the following technical solutions:
A kind of hot spot based on convolutional neural networks and MARFE automatic testing method, it is characterised in that the following steps are included:
Step 1: input needs the image identified；
Step 2: image scaling zooms to 256*256 pixel size；
Step 3: image preprocessing, including median filtering is carried out to image；
Step 4: extracting full figure feature using convolutional neural networks；
Step 5: being classified using classifier to the feature extracted；
Step 6: output category result.
A kind of hot spot based on convolutional neural networks and MARFE automatic testing method, it is characterised in that: the step
In rapid 4, convolutional neural networks are divided into two convolution sums, two pond layers using three-layer coil product structure；Wherein first layer convolution kernel
Size is 5*5, and step-length 1, second layer convolution kernel size is 3*3, and step-length 1, convolution kernel number is 1, is had after two convolutional layers
One pond layer convolution kernel size is 3*3, and step-length 2, convolution kernel number is 2；Each convolution convolutional layer is made using Relu function
Nonlinear Processing is carried out for activation primitive；The each layer of feature extraction layer that Chi Huahou is obtained is divided into M*M small regions simultaneously
Spliced, is spliced into M*M*2 characteristic layer；Global pond is done on these small regions, obtains M*M*2 feature.
A kind of hot spot based on convolutional neural networks and MARFE automatic testing method, it is characterised in that: the step
In rapid 5, classifier uses support vector machines, neural network or Naive Bayes Classifier.It is selected according to the result of hands-on
It shows optimal classifier and carries out last classification work.
A kind of hot spot based on convolutional neural networks and MARFE automatic testing method, it is characterised in that: the step
In rapid 5, convolutional neural networks are trained as classifier using the neural network connected entirely；In training convergence, obtaining property
After feature extraction structure that can be more excellent, the feature of image is first extracted with feature extraction layer, and as input data
Collection, for training new classifier.
A kind of hot spot based on convolutional neural networks and MARFE automatic testing method have in tokamak of the present invention
There is advantage below:
The present invention provides the appearance of hot spot in automatic detection tokamak and MARFE etc., and this method can be with on-line checking, can also
To be used to screen the data for occurring hot spot and MARFE in historical data；Computationally, the present invention is excellent using classical CNN model
Change the quantity of each layer convolution kernel, and abandon the method connected entirely in bottleneck layer, using processing means similar with full convolution by mould
Shape parameter reduces；The reduction of parameter, which is avoided as much as possible, there is the case where over-fitting, while substantially reducing in calculation amount.
Detailed description of the invention
A part of attached drawing of the invention is constituted to be used to provide further understanding of the present invention.Schematic implementation of the invention
Example and its explanation are not constituted improper limitations of the present invention for explaining invention.In the accompanying drawings:
Fig. 1 be on a kind of tokamak device based on convolutional neural networks described in the embodiment of the present invention hot spot and MARFE from
The flow diagram of dynamic detection method.
Fig. 2 is the convolutional neural networks model structure for feature extraction.
Specific embodiment
In order to the technological means of the invention realized, creation characteristic, achieving the goal is easy to understand with effect, ties below
Conjunction is specifically illustrating and preferred embodiment, and the present invention is further explained.
Realization process of the invention is referring to fig. 1 and fig. 2.The present invention provides a kind of hot spot based on convolutional neural networks
With MARFE automatic testing method, the present invention extracts the feature of image using convolutional neural networks, realizes on tokamak device
The automatic detection of hot spot and MARFE, the technical solution realized include:
1, image preprocessing
Image preprocessing includes: the adjustment of image size, image filtering.Image is uniformly adjusted to 256* when image size adjusts
256 sizes.The image filtering stage uses Threshold segmentation and median filtering, handles the noise of image, improves the accuracy of result.
2, the design and training of convolutional neural networks
Convolutional neural networks are a kind of common and most commonly used neural network models of field of image recognition.Use convolution
Neural network can effectively extract high dimensional information from image, realize a series of functions such as image classification, detection.The present invention
The convolutional neural networks of middle design consider the requirement of later real-time, and each layer has only used a few convolution kernel.Pass through
Practice discovery, it is not high to last result promotion using down-sampling between convolutional layer.So the convolutional Neural in the present invention
Network introduces average pond layer after two convolutional layers, reduces parameter；Specifically:
Convolutional neural networks are divided into two convolution sums, two pond layers using three-layer coil product structure；Wherein first layer convolution kernel is big
Small is 5*5, and step-length 1, second layer convolution kernel size is 3*3, and step-length 1, convolution kernel number is 1, has one after two convolutional layers
A pond layer convolution kernel size is 3*3, and step-length 2, convolution kernel number is 2；Each convolution convolutional layer uses the conduct of Relu function
Activation primitive carries out Nonlinear Processing；It is divided into M*M small regions to go forward side by side on each layer of feature extraction layer that Chi Huahou is obtained
Row splicing, is spliced into M*M*2 characteristic layer；Global pond is done on these small regions, obtains M*M*2 feature.
Training sample is selected from EAST device experiment data.Contained in positive sample occur at various locations Hot Spot with
The picture of MARFE.Negative sample contains numerous, tests the Plasma picture of different phase.
3, classifier design
Classification is two classification problems in the present invention, available good as a result, in training using full Connection Neural Network
After restraining, obtaining the more excellent feature extraction structure of performance, the feature of image is first extracted with feature extraction layer, and made
For input data set, for training new classifier.But in two classification problems, can usually it be reached using support vector machines
To better result.The present invention also has trained SVM after training feature extraction layer.The support that the present invention designs to
Amount machine kernel function uses polynomial kernel.
4, final output classification results.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to restrict the invention, all in essence of the invention
Within mind and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (4)
1. a kind of hot spot based on convolutional neural networks and MARFE automatic testing method, it is characterised in that the following steps are included:
Step 1: input needs the image identified；
Step 2: image scaling；
Step 3: image preprocessing；
Step 4: extracting full figure feature using convolutional neural networks；
Step 5: being classified using classifier to the feature extracted；
Step 6: output category result.
2. a kind of hot spot based on convolutional neural networks according to claim 1 and MARFE automatic testing method, feature
Be: in the step 4, convolutional neural networks are divided into two convolution sums, two pond layers using three-layer coil product structure；Wherein
One layer of convolution kernel size is 5*5, step-length 1, and second layer convolution kernel size is 3*3, step-length 1, and convolution kernel number is 1, two
Having a pond layer convolution kernel size after convolutional layer is 3*3, and step-length 2, convolution kernel number is 2；Each convolution convolutional layer uses
Relu function carries out Nonlinear Processing as activation primitive；The each layer of feature extraction layer that Chi Huahou is obtained is divided into M*M
Small region is simultaneously spliced, and M*M*2 characteristic layer is spliced into；Global pond is done on these small regions, obtains M*M*2
A feature.
3. a kind of hot spot based on convolutional neural networks according to claim 1 and MARFE automatic testing method, feature
Be: in the step 5, classifier uses support vector machines, neural network or Naive Bayes Classifier.
4. a kind of hot spot based on convolutional neural networks according to claim 3 and MARFE automatic testing method, feature
It is: in the step 5, convolutional neural networks is trained as classifier using the neural network connected entirely；In training
After restraining, obtaining the more excellent feature extraction structure of performance, the feature of image is first extracted with feature extraction layer, and made
For input data set, for training new classifier.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810967404.6A CN109345506A (en) | 2018-08-23 | 2018-08-23 | A kind of hot spot based on convolutional neural networks and MARFE automatic testing method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810967404.6A CN109345506A (en) | 2018-08-23 | 2018-08-23 | A kind of hot spot based on convolutional neural networks and MARFE automatic testing method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109345506A true CN109345506A (en) | 2019-02-15 |
Family
ID=65296906
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810967404.6A Pending CN109345506A (en) | 2018-08-23 | 2018-08-23 | A kind of hot spot based on convolutional neural networks and MARFE automatic testing method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109345506A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111935892A (en) * | 2019-05-13 | 2020-11-13 | 中科智云科技有限公司 | Method and apparatus for measuring plasma state |
| CN115135990A (en) * | 2020-02-21 | 2022-09-30 | 国立研究开发法人产业技术总和研究所 | Sample analysis system, learned model generation method, and sample analysis method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096804A (en) * | 2010-12-08 | 2011-06-15 | 上海交通大学 | Method for recognizing image of carcinoma bone metastasis in bone scan |
| CN106127164A (en) * | 2016-06-29 | 2016-11-16 | 北京智芯原动科技有限公司 | The pedestrian detection method with convolutional neural networks and device is detected based on significance |
| CN107247949A (en) * | 2017-08-02 | 2017-10-13 | 北京智慧眼科技股份有限公司 | Face identification method, device and electronic equipment based on deep learning |
| CN107341518A (en) * | 2017-07-07 | 2017-11-10 | 东华理工大学 | A kind of image classification method based on convolutional neural networks |
| CN107657249A (en) * | 2017-10-26 | 2018-02-02 | 珠海习悦信息技术有限公司 | Method, apparatus, storage medium and the processor that Analysis On Multi-scale Features pedestrian identifies again |
| CN108021894A (en) * | 2017-12-07 | 2018-05-11 | 中国科学院合肥物质科学研究院 | A kind of wheat diseases detection method based on depth-size strategy study |
-
2018
- 2018-08-23 CN CN201810967404.6A patent/CN109345506A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096804A (en) * | 2010-12-08 | 2011-06-15 | 上海交通大学 | Method for recognizing image of carcinoma bone metastasis in bone scan |
| CN106127164A (en) * | 2016-06-29 | 2016-11-16 | 北京智芯原动科技有限公司 | The pedestrian detection method with convolutional neural networks and device is detected based on significance |
| CN107341518A (en) * | 2017-07-07 | 2017-11-10 | 东华理工大学 | A kind of image classification method based on convolutional neural networks |
| CN107247949A (en) * | 2017-08-02 | 2017-10-13 | 北京智慧眼科技股份有限公司 | Face identification method, device and electronic equipment based on deep learning |
| CN107657249A (en) * | 2017-10-26 | 2018-02-02 | 珠海习悦信息技术有限公司 | Method, apparatus, storage medium and the processor that Analysis On Multi-scale Features pedestrian identifies again |
| CN108021894A (en) * | 2017-12-07 | 2018-05-11 | 中国科学院合肥物质科学研究院 | A kind of wheat diseases detection method based on depth-size strategy study |
Non-Patent Citations (2)
| Title |
|---|
| 秦雅丽 等: "《EAST等离子体剖面数据可视化》", 《计算机***应用》 * |
| 许志贤: "《基于表面增强拉曼散射的福美双检测研究》", 《中国优秀硕士学位论文全文数据库 工程科技Ⅰ辑》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111935892A (en) * | 2019-05-13 | 2020-11-13 | 中科智云科技有限公司 | Method and apparatus for measuring plasma state |
| CN111935892B (en) * | 2019-05-13 | 2022-11-22 | 中科智云科技有限公司 | Method and apparatus for measuring plasma state |
| CN115135990A (en) * | 2020-02-21 | 2022-09-30 | 国立研究开发法人产业技术总和研究所 | Sample analysis system, learned model generation method, and sample analysis method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | A cascaded R-CNN with multiscale attention and imbalanced samples for traffic sign detection | |
| CN108876780B (en) | Bridge crack image crack detection method under complex background | |
| CN105512638B (en) | A kind of Face datection and alignment schemes based on fusion feature | |
| Li et al. | Sewer pipe defect detection via deep learning with local and global feature fusion | |
| CN106408562A (en) | Fundus image retinal vessel segmentation method and system based on deep learning | |
| CN114937179B (en) | Junk image classification method and device, electronic equipment and storage medium | |
| CN111767860A (en) | Method and terminal for realizing image recognition through convolutional neural network | |
| CN110852330A (en) | Behavior identification method based on single stage | |
| CN113487610B (en) | Herpes image recognition method and device, computer equipment and storage medium | |
| Mahmood et al. | A Comparative study of a new hand recognition model based on line of features and other techniques | |
| CN108875456A (en) | Object detection method, object detecting device and computer readable storage medium | |
| CN109345506A (en) | A kind of hot spot based on convolutional neural networks and MARFE automatic testing method | |
| Zhou | Feature extraction of human motion video based on virtual reality technology | |
| CN115457296A (en) | Structure extraction method oriented to non-stationary texture structure attributes | |
| CN113888505B (en) | Natural scene text detection method based on semantic segmentation | |
| Yang et al. | An improved algorithm for the detection of fastening targets based on machine vision | |
| Wang et al. | A deep learning-based experiment on forest wildfire detection in machine vision course | |
| CN113111804B (en) | Face detection method and device, electronic equipment and storage medium | |
| Xie et al. | Research on mtcnn face recognition system in low computing power scenarios | |
| Li et al. | Online rail fastener detection based on YOLO network | |
| CN112686128A (en) | Classroom desk detection method based on machine learning | |
| Aina et al. | Gesture recognition system for Nigerian tribal greeting postures using support vector machine | |
| Kołodziej et al. | Using deep learning to recognize the sign alphabet | |
| Xia et al. | Application Study of YOLOv5 Algorithm on Automotive Wheel Surface Defect Detection | |
| Yu et al. | PCB Defect Detection Based on GAN Data Generation with Self-attentive Mechanism |
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: 20190215 |