CN112597865A - Intelligent identification method for edge defects of hot-rolled strip steel - Google Patents
Intelligent identification method for edge defects of hot-rolled strip steel Download PDFInfo
- Publication number
- CN112597865A CN112597865A CN202011489459.4A CN202011489459A CN112597865A CN 112597865 A CN112597865 A CN 112597865A CN 202011489459 A CN202011489459 A CN 202011489459A CN 112597865 A CN112597865 A CN 112597865A
- Authority
- CN
- China
- Prior art keywords
- layer
- image data
- size
- data
- hot
- 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
- 230000007547 defect Effects 0.000 title claims abstract description 138
- 229910000831 Steel Inorganic materials 0.000 title claims abstract description 66
- 239000010959 steel Substances 0.000 title claims abstract description 66
- 238000000034 method Methods 0.000 title claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 56
- 238000001514 detection method Methods 0.000 claims abstract description 24
- 238000004519 manufacturing process Methods 0.000 claims abstract description 12
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 238000007689 inspection Methods 0.000 claims abstract description 5
- 238000011176 pooling Methods 0.000 claims description 37
- 230000006870 function Effects 0.000 claims description 34
- 230000004913 activation Effects 0.000 claims description 21
- 238000012360 testing method Methods 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 12
- 238000012795 verification Methods 0.000 claims description 12
- 230000003044 adaptive effect Effects 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 7
- 230000007246 mechanism Effects 0.000 claims description 5
- 210000002569 neuron Anatomy 0.000 claims description 5
- 239000002893 slag Substances 0.000 claims description 5
- 235000002566 Capsicum Nutrition 0.000 claims description 4
- 239000006002 Pepper Substances 0.000 claims description 4
- 235000016761 Piper aduncum Nutrition 0.000 claims description 4
- 235000017804 Piper guineense Nutrition 0.000 claims description 4
- 235000008184 Piper nigrum Nutrition 0.000 claims description 4
- 238000009826 distribution Methods 0.000 claims description 4
- 210000004205 output neuron Anatomy 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 150000003839 salts Chemical class 0.000 claims description 4
- 238000013519 translation Methods 0.000 claims description 4
- 238000003860 storage Methods 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 2
- 238000005520 cutting process Methods 0.000 claims description 2
- 239000011159 matrix material Substances 0.000 claims description 2
- 244000203593 Piper nigrum Species 0.000 claims 1
- 238000005457 optimization Methods 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 abstract description 5
- 238000013461 design Methods 0.000 abstract description 2
- 230000004044 response Effects 0.000 abstract description 2
- 238000000605 extraction Methods 0.000 description 6
- 241000722363 Piper Species 0.000 description 3
- 238000005098 hot rolling Methods 0.000 description 3
- 229910000976 Electrical steel Inorganic materials 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011143 downstream manufacturing Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an intelligent identification method for edge defects of hot-rolled strip steel, and aims to improve the detection capability of the edge defects of the hot-rolled strip steel. The invention firstly scientifically classifies common edge defects, establishes an edge defect image data set on the basis, designs a brand new network structure by taking a convolutional neural network as a core, establishes an intelligent identification model of the edge defects of the hot-rolled strip steel according to the network structure, and finally embeds the model into a surface inspection system to participate in guiding production. The method comprises the following steps: 1. collecting and preprocessing edge defect image data; 2. establishing an intelligent identification model for edge defects of hot-rolled strip steel; 3. training, verifying, tuning and predicting the recognition model; 4. and (4) the optimal recognition model is fused with the table inspection system in a cooperative manner. The method has the characteristics of simple model structure, high response speed, high identification precision and the like, and has important significance for improving the surface quality of the hot-rolled strip steel.
Description
Technical Field
The invention belongs to the field of surface quality detection of hot-rolled strip steel in the metallurgical rolling technology, and particularly relates to an intelligent identification method for edge defects of the hot-rolled strip steel.
Background
With the rapid development of industry and economy in China, the use requirement of high-quality hot-rolled strip steel is more and more increased, the surface quality is taken as one of the most important indexes of the quality of hot-rolled strip steel products, the appearance and the yield of the products are seriously influenced, the production of downstream processes is also influenced, and therefore, the research on the online identification and detection of the surface quality defects of the hot-rolled strip steel has important significance. The edge defects are one kind of surface defects, and the defects have strong similarity, so that the existing detection model and equipment are difficult to achieve a good detection effect. Therefore, the invention provides an intelligent identification method for the edge defects of the hot-rolled strip steel, which has important significance for improving the surface quality of the strip steel.
Related researches have been published in the field of surface defect detection of hot rolled strip steel at present, for example: 8 linear array CCD cameras with 1024 pixels are adopted as an image acquisition device, a defect detection and identification algorithm flow is provided according to the surface characteristics of the hot-rolled strip steel, and the method is applied to a 1700mm hot-rolling production line; the hot-rolled steel plate surface defect identification method based on Tetrolet transformation and dimension reduction of nuclear protection authority projection obtains 97.3846% identification rate of a defect sample library. The method for identifying the surface defects of the silicon steel by using a decision tree and expert experience classification is applied to the production line of the cold-rolled silicon steel for the Wuhan steel, and the on-line overall identification rate of serious defects reaches 76%. A hot rolled strip surface defect data set (NEU) is established, which comprises six common surface defects: the recognition accuracy of the anti-noise defect recognition algorithm provided by the method reaches 97.89%.
In summary, the existing surface defect detection equipment, theory and technology have basically satisfactory detection effects on typical defects with definite characteristics, such as cracks, slag inclusions, plaques, pocks, scales and scratch defects. However, for the edge defects of the hot-rolled strip steel with similar macroscopic morphology characteristics, an effective detection method is not available so far, and the defects need to be subdivided in a manual detection mode in the production process, so that the production efficiency is seriously reduced, and the labor intensity is increased.
The method comprises the steps of firstly dividing common edge defects into 5 types according to a defect generation mechanism and corresponding control measures, secondly constructing a brand new convolutional neural network structure, and further establishing an intelligent identification model of the edge defects of the hot-rolled strip steel on the basis. The model has the characteristics of high identification speed, high detection precision and good generalization capability, and has important significance for improving the surface quality of the hot-rolled strip steel.
Disclosure of Invention
The invention aims to provide an intelligent identification method for edge defects of hot-rolled strip steel. The edge part is a region with multiple surface defects of the hot rolled strip steel, and the phenomenon of false detection is easy to occur due to certain similarity of the appearance of the edge part defects. In order to improve the edge defect detection precision, the invention firstly carries out scientific and reasonable classification on common edge defects, establishes a defect image data set on the basis, secondly designs a new network structure by taking a convolutional neural network as a core, and finally establishes an intelligent identification model of the edge defects of the hot-rolled strip steel. The model has high speed and high precision for identifying the edge defect images, and has important significance for improving the edge quality of the hot-rolled strip steel on the production site.
In order to achieve the purpose, the invention provides an intelligent identification method for edge defects of hot-rolled strip steel, which comprises the following steps:
a. the method comprises the following steps of collecting and preprocessing a hot-rolled strip steel edge defect image data set, and comprises the following specific steps:
a1, dividing the edge defects of the hot-rolled strip steel into 5 types including black lines Bl, air holes Bb, cracks Ck, slag inclusion Ic and warping skin Wp according to the appearance, the generation mechanism and the generation positions of the edge defects;
a2, collecting an edge defect image of the hot-rolled strip steel: acquiring edge defect image data through a meter inspection instrument on a hot-rolled strip steel production field, and classifying the edge defect images according to classification standards;
a3, storing the classified edge defect image data into different files, and then adding labels to the image data;
a4, dividing the labeled defect image data set, and dividing the defect image data set according to the ratio of 4: 1 is divided into a data set dataset1 and a data set dataset2, wherein dataset2 is used as a test set to detect the recognition effect of the model;
a5, performing image enhancement processing on the data set dataset1 to expand the capacity of the defect image data set, so as to extract more characteristic information from the defect image data and improve the generalization capability and anti-interference capability of the model;
a6, and performing image enhancement processing on the data set dataset1 according to the ratio of 3: 1, dividing the training set into a train set and a validation set;
a7, preprocessing data of a training set train, a verification set validset and a test set dataset 2;
b. establishing an intelligent identification model of edge defects of hot-rolled strip steel:
b1, constructing and identifying the structure of a feature extractor featurein the model;
b2, establishing a structure of a recognition model Classifier;
c. training and predicting an intelligent recognition model of edge defects of hot-rolled strip steel, and specifically comprising the following steps:
c1, setting training parameters of the recognition model: setting batch training size batch-size, learning rate learning-rate and iteration times epoch of training, and selecting an optimizer and a Loss function;
c2, inputting the training set train set and the verification set validset preprocessed in the step a7 into the recognition model established in the step b for training, and finishing the training when the training iteration times reach the times set by epoch;
c3, outputting the loss and acc of the training set and the verification set of the recognition model, judging whether the model meets the condition of loss <0.5 and acc > 90%, if not, switching to the step c4, and if so, switching to the step c 5;
c4, properly adjusting the training parameters of the recognition model according to the training result, and then continuing training in the step c 2;
c5, storing the trained identification model parameters to obtain an intelligent identification model R for the edge defects of the hot-rolled strip steel;
c6, predicting the edge defect image by using the stored recognition model R.
Preferably, the image enhancement processing is performed on the data set in the step a5, and the specific steps include:
a51, symmetrically overturning the original edge defect image data and storing the image data into an original data set;
a52, rotating the original edge defect image data, randomly selecting an angle from the range of alpha more than 0 degrees and less than 90 degrees, and storing the angle into an original data set;
a53, adding Gaussian noise to the original edge defect image data and storing the image data into an original data set;
a54, performing plane pixel translation on the original edge defect image data, and storing the original edge defect image data into an original data set;
a55, adding salt and pepper noise to the original edge defect image data and storing the image data into an original data set;
a56, adjusting the brightness of the original edge defect image data and storing the image data into an original data set;
and a57, reducing the brightness of the original edge defect image data and storing the image data into an original data set.
Preferably, the specific steps of preprocessing the image before inputting the model in the step a7 include:
a71, performing center crop CenterCrop processing on the image data, and setting the size after crop to be BxB;
a72, in order to eliminate the influence caused by the difference of unit and scale between different image data, improve the comparability of the data so as to improve the convergence speed during model training, carry out Z-score standardization processing on the image data,x is a pixel point matrix of the image, and the mean value x of the image data is calculated according to xminAnd standard deviation xstdUsing the conversion formula x ═ x-xmin)/xstdSo that each image data is converted into data having a mean value of 0 and a standard deviation of 1, which shows a normal distribution.
Preferably, the process of constructing the feature extractor Features structure in the recognition model is as follows:
b11, determining the layer number F of the feature extractor Features to be 5;
b12, establishing a first layer Features _ layer1 of the feature extractor: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 1Set to 32, convolution kernel size kernel _ size1 1Set to 3 × 3, step size stride1 1Set to 2, padding1 1Is set to 1; calculating the convolved data through a ReLU activation function; then performing pooling operation, wherein the pooling layer is a Maxpool2d layer, and the kernel _ size is pooled2 1Set to 2 × 2, step size stride2 1Set to 2;
b13, establishing a second layer Features _ layer2 of the feature extractor: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 2Set to 96, convolution kernel size kernel _ size1 2Set to 5 × 5, step size stride1 2Set to 1, padding1 2Set to 2; calculating the convolved data through a ReLU activation function; then performing pooling operation, wherein the pooling layer is a Maxpool2d layer, and the kernel _ size is pooled2 2Set to 2 × 2, step size stride2 2Set to 2;
b14, establishing a third layer Features _ layer3 of the feature extractor: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 3Set to 192, convolution kernel size kernel _ size1 3Set to 3 × 3, step size stride1 3Set to 1, padding1 3Is set to 1; calculating the convolved data through a ReLU activation function, wherein a pooling layer is not added to the feature extractor of the layer;
b15, establishing a feature extractor fourth layer Features _ layer 4: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 4Set to 128, convolution kernel size kernel _ size1 4Set to 3 × 3, step size stride1 4Set to 1, padding1 4Is set to 1; calculating the convolved data through a ReLU activation function, wherein a pooling layer is not added to the feature extractor of the layer;
b16, constructing a fifth layer Features _ layer 5: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 5Set to 128, convolution kernel size kernel _ size1 5Set to 3 × 3, step size stride1 5Set to 1, padding1 5Is set to 1; calculating the convolved data through a ReLU activation function; then performing pooling operation, wherein the pooling layer is a Maxpool2d layer, and the kernel _ size is pooled2 5Set to 2 × 2, step size stride2 5Set to 2; after pooling, a layer of adaptive average pooling layer adaptive average pooling pool2d is added, the output _ size of the adaptive average pooling layer is set to 6 × 6, and the data size is pooled into 6 × 6 × 128 three-dimensional data.
Preferably, the process of establishing the Classifier structure of the recognition model Classifier is as follows:
b21, determining the layer number C of the Classifier to be 3;
b22, flattening the three-dimensional data with the data size of 6 multiplied by 128 processed in the step b16 into one-dimensional data by using a Flatten function before the three-dimensional data is input into a classifier, wherein the data length is 6 multiplied by 128, then passing through a dropout layer, and the rejection probability is set to be P1＝0.5；
b23, establishing a Classifier first layer Classifier _ layer1, adding a layer of Linear layer, inputting one-dimensional data of 6 multiplied by 128, setting the number of output neurons as 2304, calculating through a ReLU activation function, then passing through a layer of dropout, setting the rejection probability as P2＝0.5；
b24, establishing Classifier second layer Classifier _ layer 2: adding a layer of Linear layer, wherein the input is 2304, the number of output neurons is 2304, and the classifier layer is calculated by a ReLU activation function and has no dropout layer;
b25, establishing Classifier third layer Classifier _ layer 3: a layer of Linear layer is added, the input is 2304, and the number of neurons set by the output layer is the same as the number of classification task categories of the recognition model.
Preferably, the tuning step of step c4 is:
c41, adjusting the parameter learning rate learning-rate, batch training size batch-size and training iteration frequency epoch of the recognition model R according to the actual training effect;
c42, adjusting the size B multiplied by B of the center cutting in the step a 5;
c43, adjusting the layer structure of the feature extractor Features and the network parameters of each layer;
c44, adjusting the layer structure of the Classifier and the network parameters of each layer.
Preferably, the predicting step of the step c6 is:
c61, loading the recognition model R stored in the step c5, and inputting the test set dataset2 preprocessed in the step a7 into the recognition model R for prediction;
c62, outputting the prediction result of the recognition model R on the test set dataset2 and the class probability of each image data;
and c63, embedding the recognition model R into the hot-rolled strip steel surface detection system, carrying out real-time online defect detection according to the edge defect image of the strip steel shot by the camera, and storing the detection result in a database for storage and recording.
Compared with the prior art, the invention has the following beneficial effects:
1) according to the comprehensive influences of the appearance, mechanism occurrence positions and the like of the field defects, scientific and reasonable classification standards are established for the edge defects of the hot-rolled strip steel, and the edge defects are further classified;
2) the invention provides a brand-new convolutional neural network structure, and establishes an intelligent identification model of the edge defects of the hot-rolled strip steel by taking the convolutional neural network structure as a core, the model has high response speed and high identification precision, the detection speed of a single defect image meets the production requirement, and the intelligent identification model has important significance for improving the surface quality of the hot-rolled strip steel.
Drawings
FIG. 1 is an overall flow chart of an intelligent identification model for edge defects of hot-rolled strip steel;
FIG. 2 is an image of edge defects of a class 5 hot rolled strip;
FIG. 3 is image data of a black line type edge defect image after image enhancement;
FIG. 4 is a network structure diagram of an intelligent recognition model for edge defects of hot-rolled strip steel;
FIG. 5 is a plot of training error loss for recognition model R;
fig. 6 is a graph of the training accuracy acc of the recognition model R.
Detailed Description
The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It is to be noted that, unless otherwise specified, technical or scientific terms used herein shall have the ordinary meaning as understood by those skilled in the art to which the invention pertains.
As shown in figures 1-6, an intelligent identification method for edge defects of hot-rolled strip steel mainly comprises
a. The method comprises the following steps of collecting and preprocessing a hot-rolled strip steel edge defect image data set, and comprises the following specific steps:
a1, establishing scientific classification standards for the edge defects of the hot-rolled strip steel according to the comprehensive influence factors such as the appearance, the generation mechanism, the generation position and the like of the edge defects, and totally classifying the defects into 5 types: black lines Bl, air holes Bb, cracks Ck, slag inclusion Ic and warping skin Wp;
a2, collecting edge defects of the hot-rolled strip steel: the data of edge defect images are collected through a meter inspection instrument on the site of a hot rolling strip steel production line, the edge defect images are divided according to classification standards and retrograde classes, 5 types of edge defects collected on the site of the hot rolling production line are shown in figure 2, and figure 2 a-figure 2e respectively show a black line defect, a bubble defect, a crack defect, a slag inclusion defect and a skin tilting defect.
a3, storing the classified edge defect image data into different files, and then adding labels to the image data;
a4, dividing the labeled defect image data set, and dividing the defect image data set according to the ratio of 4: 1 into a data set dataset1 and a data set dataset2, wherein dataset2 serves as a test set to detect the recognition effect of the model;
a5, performing image enhancement processing on the data set dataset1, specifically comprising the following steps: the original data is symmetrically inverted, the rotation angle α is 45 °, gaussian noise is added, salt and pepper noise is added, planar pixel translation, brightness is increased, and brightness is decreased, and the original data is stored in an original defect image data set dataset1, the image data after image enhancement by taking a black line as an example is shown in fig. 3, and fig. 3a to 3h correspond to the original data, symmetric inversion, rotation angle, gaussian noise is added, salt and pepper noise is added, brightness is increased, brightness is decreased, and planar pixel translation, respectively. The volume of the data set dataset1 subjected to image enhancement is 7 times of that of the original data, so that the scale of training data is increased, and the generalization capability of the model can be improved;
a6, and performing image enhancement on the data set dataset1 according to the ratio of 3: 1, dividing the defect image into a training set train and a verification set validset, wherein the number of each divided defect image data set is shown in table 1;
TABLE 1 distribution of various defect images for training set, validation set, and test set
a7, preprocessing data of a training set train, a verification set validset and a test set dataset2, and specifically comprises the following steps:
a71, performing CenterCrop center cropping processing on the data set to obtain pictures with the size of 227 multiplied by 227;
72, converting the cut image data into data with a mean value of 0 and a standard deviation of 1 presenting normal distribution after Z-score standardization, eliminating the influence caused by unit and scale difference among different images by the data standardized by Z-score, and improving the comparability of the data and the convergence speed during model training;
b. the method comprises the following steps of establishing an intelligent recognition model for edge defects of hot-rolled strip steel, and specifically comprising the following steps:
b1, constructing and identifying the structure of a feature extractor featurein the model: determining the layer number F of the feature extractor Features to be 5, and determining the number out _ channel of convolution layers Conv2d convolution kernels in the first feature extraction layer Features _ layer11 1Set to 32, convolution kernel size kernel _ size1 1Set to 3 × 3, step size stride1 1Set to 2, padding1 1Setting the value to be 1, selecting a ReLU function by the convoluted activation function, selecting a Maxpool2d by the pooling mode, and pooling the kernel size kernel _ size2 1Set to 2 × 2, step size stride2 1Set to 2, the data becomes three-dimensional data of 57 × 57 × 32 size after passing through the Features _ layer 1; the number of convolution kernels out _ channel of convolution layer Conv2d in the second feature extraction layer Features _ layer21 2Set to 96, convolution kernel size kernel _ size1 2Set to 5 × 5, step size stride1 2Set to 1, padding1 2Setting the kernel size to be 2, selecting a ReLU function by the convolved activation function, selecting a Maxpool2d by the pooling mode, and pooling the kernel size kernel _ size2 2Set to 2 × 2, step size stride2 2Set to 2, the data becomes three-dimensional data of 28 × 28 × 96 size after passing through Features _ layer 2; number of convolution kernels out _ channel of convolution layer Conv2d in third feature extraction layer Features _ layer31 3Set to 192, convolution kernel size kernel _ size1 3Set to 3 × 3, step size stride1 3Set to 1, padding1 3Setting the value to be 1, selecting a ReLU function by the convoluted activation function, not performing pooling operation on the layer, and changing data into three-dimensional data with the size of 28 multiplied by 192 after the data passes through Features _ layer 3; fourth characteristicExtracting the number of convolution kernels out _ channel of convolution layer Conv2d in Features _ layer4 layer1 4Set to 128, convolution kernel size kernel _ size1 4Set to 3 × 3, step size stride1 4Set to 1, padding1 4Setting the value to be 1, selecting a ReLU function by the convoluted activation function, not performing pooling operation on the layer, and changing data into three-dimensional data with the size of 28 multiplied by 128 after the data passes through Features _ layer 4; number of convolution kernels out _ channel 2d in convolution layer conv2 _ d in feature extraction Features _ layer5 layer1 5Set to 128, convolution kernel size kernel _ size1 5Set to 3 × 3, step size stride1 5Set to 1, padding1 5Setting the value to be 1, selecting a ReLU function by the convoluted activation function, selecting a Maxpool2d by the pooling mode, and pooling the kernel size kernel _ size2 5Set to 2 × 2, step size stride2 5Setting the size of three-dimensional data obtained after pooling to be 2, wherein the size of the three-dimensional data obtained after pooling is 14 multiplied by 128, the data after pooling passes through a layer of adaptive average pooling adaptive _ avgpool2d, the output _ size of the adaptive average pooling layer is set to be 6 multiplied by 6, and the output data is pooled into three-dimensional data with the size of 6 multiplied by 128;
b2, establishing a structure of a Classifier of the recognition model Classifier: the image data is output to be 6 multiplied by 128 three-dimensional data after passing through a feature extraction layer, because the layer structure in the Classifier is a Linear layer, a Flatten function is used for carrying out flattening operation on the data after the image data is output from the feature extraction layer to enable the data to be changed into one-dimensional data, the data length is 6 multiplied by 128, a dropout layer is added before the image data enters a classification layer Classifier, and a rejection probability P is set10.5; the Classifier selects three layers of fully connected neural networks; the number of Linear layer neurons in the first classification layer is 2304, the ReLU function is selected as the activation function, then a dropout layer is added, and the rejection probability is set to be P20.5; the number of Linear layer neurons in the second classification layer is 2304, the ReLU function is selected by the activation function layer, and the dropout layer is not added; the third classification layer is an output Linear layer, and because the model identifies 5 defects, the neuron number of the output layer is 5, and the edge defects of the well-established hot-rolled strip steelThe structure of the trap recognition model is shown in FIG. 3;
c. training and predicting an intelligent recognition model of edge defects of hot-rolled strip steel, and specifically comprising the following steps:
c1, setting training parameters of the recognition model, setting the batch training size batch-size to be 64, setting the iteration time epoch of model training to be 500, selecting Adam by an optimizer, setting the learning rate learning-rate of the optimizer to be 0.0005, and selecting Cross EntropyLoss by a Loss function;
c2, inputting the training set train set and the verification set validset preprocessed in the step a7 into the recognition model established in the step b for training, and finishing the training when the training iteration times reach 500 times;
c3, outputting the loss and acc of the training set and the verification set of the recognition model, and making a loss map and acc image, wherein the loss error of the training set and the verification set of the recognition model after 500 times of iterative training is 0.1365 and 0.1981 respectively, and the accuracy of the training set and the verification set is 98.4 percent and 94.94 percent respectively, as shown in fig. 4 and 5. If yes, the step is carried out to step c 5;
c5, storing the trained identification model parameters to obtain an intelligent identification model R for the edge defects of the hot-rolled strip steel;
c6, predicting the edge defect image by using the stored recognition model R, which comprises the following steps:
c61, loading the recognition model R stored in the step c5, and inputting the test set dataset2 preprocessed in the step a7 into the recognition model R for prediction;
c62, outputting the prediction results of the recognition model R on the test set dataset2, and the class probability of each image data, wherein the prediction results of various types of defects are shown in Table 2
TABLE 2 prediction of various types of defects in the test set
And c63, embedding the recognition model R into the hot-rolled strip steel surface detection system, carrying out real-time online defect detection according to the edge defect image of the strip steel shot by the camera, and storing the detection result in a database for storage and recording.
The above-mentioned embodiments are only used for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features thereof can be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.
Claims (7)
1. The intelligent identification method for the edge defects of the hot-rolled strip steel is characterized by comprising the following steps of:
a. the method comprises the following steps of collecting and preprocessing a hot-rolled strip steel edge defect image data set, and comprises the following specific steps:
a1, dividing the edge defects of the hot-rolled strip steel into 5 types including black lines Bl, air holes Bb, cracks Ck, slag inclusion Ic and warping skin Wp according to the appearance, the generation mechanism and the generation positions of the edge defects;
a2, collecting an edge defect image of the hot-rolled strip steel: acquiring edge defect image data through a meter inspection instrument on a hot-rolled strip steel production field, and classifying the edge defect images according to classification standards;
a3, storing the classified edge defect image data into different files, and then adding labels to the image data;
a4, dividing the labeled defect image data set, and dividing the defect image data set according to the ratio of 4: 1 is divided into a data set dataset1 and a data set dataset2, wherein dataset2 is used as a test set to detect the recognition effect of the model;
a5, performing image enhancement processing on the data set dataset1 to expand the capacity of the defect image data set, so as to extract more characteristic information from the defect image data and improve the generalization capability and anti-interference capability of the model;
a6, and performing image enhancement processing on the data set dataset1 according to the ratio of 3: 1, dividing the training set into a train set and a validation set;
a7, preprocessing data of a training set train, a verification set validset and a test set dataset 2;
b. establishing an intelligent identification model for edge defects of hot-rolled strip steel;
b1, constructing and identifying the structure of a feature extractor featurein the model;
b2, establishing a structure of a recognition model Classifier;
c. training and predicting an intelligent recognition model of edge defects of hot-rolled strip steel, and specifically comprising the following steps:
c1, setting training parameters of the recognition model: setting batch training size batch-size, learning rate learning-rate and iteration times epoch of training, and selecting an optimizer and a Loss function;
c2, inputting the training set train set and the verification set validset preprocessed in the step a7 into the recognition model established in the step b for training, and finishing the training when the training iteration times reach the times set by epoch;
c3, outputting the loss and acc of the training set and the verification set of the recognition model, judging whether the model meets the condition of loss <0.5 and acc > 90%, if not, switching to the step c4, and if so, switching to the step c 5;
c4, properly adjusting the training parameters of the recognition model according to the training result, and then continuing training in the step c 2;
c5, storing the trained identification model parameters to obtain an intelligent identification model R for the edge defects of the hot-rolled strip steel;
c6, predicting the edge defect image by using the stored recognition model R.
2. The intelligent identification method for the edge defects of the hot-rolled strip steel according to claim 1, characterized in that the image enhancement processing is carried out on the data set in the step a5, and the specific steps comprise:
a51, symmetrically overturning the original edge defect image data and storing the image data into an original data set;
a52, rotating the original edge defect image data, randomly selecting an angle from the range of alpha more than 0 degrees and less than 90 degrees, and storing the angle into an original data set;
a53, adding Gaussian noise to the original edge defect image data and storing the image data into an original data set;
a54, performing plane pixel translation on the original edge defect image data, and storing the original edge defect image data into an original data set;
a55, adding salt and pepper noise to the original edge defect image data and storing the image data into an original data set;
a56, adjusting the brightness of the original edge defect image data and storing the image data into an original data set;
and a57, reducing the brightness of the original edge defect image data and storing the image data into an original data set.
3. The intelligent identification method for the edge defects of the hot-rolled strip steel according to claim 2, wherein the specific steps of preprocessing the image before the model is input in the step a7 comprise the following steps:
a71, performing center crop CenterCrop processing on the image data, and setting the size after crop to be BxB;
a72, in order to eliminate the influence caused by the difference of unit and scale between different image data and improve the comparability of the data so as to improve the convergence speed during model training, the image data is subjected to Z-score standardization processing, x is a pixel matrix of the image, and the mean value x of the image data is calculated according to xminAnd standard deviation xstdUsing the conversion formula x ═ x-xmin)/xstdSo that each image data is converted into data having a mean value of 0 and a standard deviation of 1, which shows a normal distribution.
4. The intelligent hot-rolled strip steel edge defect identification method as claimed in claim 3, wherein the process of constructing the feature extractor Features structure in the identification model is as follows:
b11, determining the layer number F of the feature extractor Features to be 5;
b12, establishing a first layer Features _ layer1 of the feature extractor: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 1Set to 32, convolution kernel size kernel _ size1 1Set to 3 × 3, step size stride1 1Set to 2, padding1 1Is set to 1; calculating the convolved data through a ReLU activation function; then performing pooling operation, wherein the pooling layer is a Maxpool2d layer, and the kernel _ size is pooled2 1Set to 2 × 2, step size stride2 1Set to 2;
b13, establishing a second layer Features _ layer2 of the feature extractor: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 2Set to 96, convolution kernel size kernel _ size1 2Set to 5 × 5, step size stride1 2Set to 1, padding1 2Set to 2; calculating the convolved data through a ReLU activation function; then performing pooling operation, wherein the pooling layer is a Maxpool2d layer, and the kernel _ size is pooled2 2Set to 2 × 2, step size stride2 2Set to 2;
b14, establishing a third layer Features _ layer3 of the feature extractor: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 3Set to 192, convolution kernel size kernel _ size1 3Set to 3 × 3, step size stride1 3Set to 1, padding1 3Is set to 1; calculating the convolved data through a ReLU activation function, wherein a pooling layer is not added to the feature extractor of the layer;
b15, establishing a feature extractor fourth layer Features _ layer 4: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 4Set to 128, convolution kernel size kernel _ size1 4Set to 3 × 3, step size stride1 4Set to 1, padding1 4Is set to 1; calculating the convolved data through a ReLU activation function, wherein a pooling layer is not added to the feature extractor of the layer;
b16, constructing a fifth layer Features _ layer 5: convolutional layer selection Conv2d, number of convolutional kernels out _ channel1 5Set to 128, convolution kernel size kernel _ size1 5Set to 3 × 3, step size stride1 5Set to 1, padding1 5Is set to 1; calculating the convolved data through a ReLU activation function; then performing pooling operation, wherein the pooling layer is a Maxpool2d layer, and the kernel _ size is pooled2 5Set to 2 × 2, step size stride2 5Set to 2; after pooling, a layer of adaptive average pooling layer adaptive average pooling pool2d is added, the output _ size of the adaptive average pooling layer is set to 6 × 6, and the data size is pooled into 6 × 6 × 128 three-dimensional data.
5. The intelligent hot-rolled strip steel edge defect identification method according to claim 4, wherein the process of establishing the Classiier structure of the identification model Classifier comprises the following steps:
b21, determining the layer number C of the Classifier to be 3;
b22, flattening the three-dimensional data with the data size of 6 multiplied by 128 processed in the step b16 into one-dimensional data by using a Flatten function before the three-dimensional data is input into a classifier, wherein the data length is 6 multiplied by 128, then passing through a dropout layer, and the rejection probability is set to be P1＝0.5；
b23, establishing a Classifier first layer Classifier _ layer1, adding a layer of Linear layer, inputting one-dimensional data of 6 multiplied by 128, setting the number of output neurons as 2304, calculating through a ReLU activation function, then passing through a layer of dropout, setting the rejection probability as P2＝0.5；
b24, establishing Classifier second layer Classifier _ layer 2: adding a layer of Linear layer, wherein the input is 2304, the number of output neurons is 2304, and the classifier layer is calculated by a ReLU activation function and has no dropout layer;
b25, establishing Classifier third layer Classifier _ layer 3: a layer of Linear layer is added, the input is 2304, and the number of neurons set by the output layer is the same as the number of classification task categories of the recognition model.
6. The intelligent identification method for the edge defects of the hot-rolled strip steel as claimed in claim 4, wherein the step c4 is an optimization step:
c41, adjusting the parameter learning rate learning-rate, batch training size batch-size and training iteration frequency epoch of the recognition model R according to the actual training effect;
c42, adjusting the size B multiplied by B of the center cutting in the step a 5;
c43, adjusting the layer structure of the feature extractor Features and the network parameters of each layer;
c44, adjusting the layer structure of the Classifier and the network parameters of each layer.
7. The intelligent identification method for the edge defects of the hot-rolled strip steel as claimed in claim 1, wherein the prediction step of the step c6 is as follows:
c61, loading the recognition model R stored in the step c5, and inputting the test set dataset2 preprocessed in the step a7 into the recognition model R for prediction;
c62, outputting the prediction result of the recognition model R on the test set dataset2 and the class probability of each image data;
and c63, embedding the recognition model R into the hot-rolled strip steel surface detection system, carrying out real-time online defect detection according to the edge defect image of the strip steel shot by the camera, and storing the detection result in a database for storage and recording.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011489459.4A CN112597865A (en) | 2020-12-16 | 2020-12-16 | Intelligent identification method for edge defects of hot-rolled strip steel |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011489459.4A CN112597865A (en) | 2020-12-16 | 2020-12-16 | Intelligent identification method for edge defects of hot-rolled strip steel |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112597865A true CN112597865A (en) | 2021-04-02 |
Family
ID=75196807
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011489459.4A Pending CN112597865A (en) | 2020-12-16 | 2020-12-16 | Intelligent identification method for edge defects of hot-rolled strip steel |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112597865A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113578976A (en) * | 2021-07-28 | 2021-11-02 | 北京首钢股份有限公司 | Production method and production system of hot-rolled electrical steel capable of controlling edge warping |
| CN113592024A (en) * | 2021-08-12 | 2021-11-02 | 燕山大学 | Cold-rolled copper strip surface defect recognition model training method, recognition method and system |
| CN114904915A (en) * | 2022-04-28 | 2022-08-16 | 首钢智新迁安电磁材料有限公司 | Hot-rolled coil shearing control method and device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109754017A (en) * | 2019-01-09 | 2019-05-14 | 西北工业大学 | Based on separable three-dimensional residual error network and transfer learning hyperspectral image classification method |
| CN110660052A (en) * | 2019-09-23 | 2020-01-07 | 武汉科技大学 | Hot-rolled strip steel surface defect detection method based on deep learning |
| CN110826609A (en) * | 2019-10-29 | 2020-02-21 | 华中科技大学 | Double-flow feature fusion image identification method based on reinforcement learning |
| CN111709265A (en) * | 2019-12-11 | 2020-09-25 | 深学科技(杭州)有限公司 | Camera monitoring state classification method based on attention mechanism residual error network |
| CN111898419A (en) * | 2020-06-17 | 2020-11-06 | 西安交通大学 | Partition landslide detection system and method based on cascade deep convolutional neural network |
-
2020
- 2020-12-16 CN CN202011489459.4A patent/CN112597865A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109754017A (en) * | 2019-01-09 | 2019-05-14 | 西北工业大学 | Based on separable three-dimensional residual error network and transfer learning hyperspectral image classification method |
| CN110660052A (en) * | 2019-09-23 | 2020-01-07 | 武汉科技大学 | Hot-rolled strip steel surface defect detection method based on deep learning |
| CN110826609A (en) * | 2019-10-29 | 2020-02-21 | 华中科技大学 | Double-flow feature fusion image identification method based on reinforcement learning |
| CN111709265A (en) * | 2019-12-11 | 2020-09-25 | 深学科技(杭州)有限公司 | Camera monitoring state classification method based on attention mechanism residual error network |
| CN111898419A (en) * | 2020-06-17 | 2020-11-06 | 西安交通大学 | Partition landslide detection system and method based on cascade deep convolutional neural network |
Non-Patent Citations (7)
| Title |
|---|
| 于显龙等: "《机械设备失效分析》", 31 October 2009, 上海科学技术出版社 * |
| 张亮军等: "基于Dense U-net方法的眼底彩色照片图像血管分割研究", 《生物医学工程研究》 * |
| 张飞: "《高性能Android开发技术》", 30 April 2019, 北京航空航天大学出版社 * |
| 杨卫华: "《智能眼科概论》", 31 October 2019, 湖北科学技术出版社2 * |
| 王东城等: ""数据驱动的热轧带钢边部线状缺陷智能预报模型"", 《钢铁》 * |
| 高敬鹏: "《深度学习：卷积神经网络技术与实践》", 30 June 2020, 机械工业出版社 * |
| 黄道鑫: "《提钒炼钢》", 31 January 2000, 冶金工业出版社 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113578976A (en) * | 2021-07-28 | 2021-11-02 | 北京首钢股份有限公司 | Production method and production system of hot-rolled electrical steel capable of controlling edge warping |
| CN113578976B (en) * | 2021-07-28 | 2023-09-01 | 北京首钢股份有限公司 | Production method and production system for hot-rolled electrical steel with edge skin tilting control function |
| CN113592024A (en) * | 2021-08-12 | 2021-11-02 | 燕山大学 | Cold-rolled copper strip surface defect recognition model training method, recognition method and system |
| CN113592024B (en) * | 2021-08-12 | 2024-05-28 | 燕山大学 | Cold-rolled copper strip surface defect identification model training method, cold-rolled copper strip surface defect identification method and cold-rolled copper strip surface defect identification system |
| CN114904915A (en) * | 2022-04-28 | 2022-08-16 | 首钢智新迁安电磁材料有限公司 | Hot-rolled coil shearing control method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112597865A (en) | Intelligent identification method for edge defects of hot-rolled strip steel | |
| CN114758288B (en) | Power distribution network engineering safety control detection method and device | |
| CN108765412A (en) | A kind of steel strip surface defect sorting technique | |
| CN110532946B (en) | Method for identifying axle type of green-traffic vehicle based on convolutional neural network | |
| CN110689524B (en) | No-reference online image definition evaluation method and system | |
| CN112184693A (en) | Intelligent detection method for weld defects of ray industrial negative | |
| CN109101976B (en) | Method for detecting surface defects of arc-extinguishing grid plate | |
| CN109975196B (en) | Reticulocyte detection method and system | |
| CN116883393B (en) | Metal surface defect detection method based on anchor frame-free target detection algorithm | |
| CN114419007B (en) | Defect type identification method and system for multi-strategy fusion deep learning network model | |
| CN112258470B (en) | Intelligent industrial image critical compression rate analysis system and method based on defect detection | |
| CN116012337A (en) | Hot rolled strip steel surface defect detection method based on improved YOLOv4 | |
| CN112819748A (en) | Training method and device for strip steel surface defect recognition model | |
| CN117315380B (en) | Deep learning-based pneumonia CT image classification method and system | |
| CN115829995A (en) | Cloth flaw detection method and system based on pixel-level multi-scale feature fusion | |
| CN115294089A (en) | Steel surface defect detection method based on improved YOLOv5 | |
| CN113221881A (en) | Multi-level smart phone screen defect detection method | |
| CN115546187A (en) | Agricultural pest and disease detection method and device based on YOLO v5 | |
| CN118071734B (en) | Mask defect positioning and classifying method and system based on machine vision | |
| CN117237599A (en) | Image target detection method and device | |
| CN113538342B (en) | Convolutional neural network-based aluminum aerosol can coating quality detection method | |
| CN112396580A (en) | Circular part defect detection method | |
| CN117350964A (en) | Cross-modal multi-level feature fusion-based power equipment detection method | |
| Nie et al. | Machine vision-based apple external quality grading | |
| CN115761606A (en) | Box electric energy meter identification method and device based on image processing |
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 |
Application publication date: 20210402 |
|
| RJ01 | Rejection of invention patent application after publication |