CN114881940A - Method for identifying head defects of high-temperature alloy bolt after hot heading - Google Patents
Method for identifying head defects of high-temperature alloy bolt after hot heading Download PDFInfo
- Publication number
- CN114881940A CN114881940A CN202210421323.2A CN202210421323A CN114881940A CN 114881940 A CN114881940 A CN 114881940A CN 202210421323 A CN202210421323 A CN 202210421323A CN 114881940 A CN114881940 A CN 114881940A
- Authority
- CN
- China
- Prior art keywords
- model
- neural network
- parameters
- data set
- image
- 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 71
- 238000000034 method Methods 0.000 title claims abstract description 40
- 229910045601 alloy Inorganic materials 0.000 title claims abstract description 22
- 239000000956 alloy Substances 0.000 title claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 37
- 238000003062 neural network model Methods 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 238000013528 artificial neural network Methods 0.000 claims description 31
- 230000006870 function Effects 0.000 claims description 24
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 5
- 238000011478 gradient descent method Methods 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 238000002372 labelling Methods 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 230000008034 disappearance Effects 0.000 claims description 2
- 238000005286 illumination Methods 0.000 claims description 2
- 230000000644 propagated effect Effects 0.000 claims description 2
- 230000000007 visual effect Effects 0.000 claims description 2
- 238000013461 design Methods 0.000 claims 1
- 238000013135 deep learning Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 238000001514 detection method Methods 0.000 description 4
- 238000011176 pooling Methods 0.000 description 4
- 230000008014 freezing Effects 0.000 description 3
- 238000007710 freezing Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000010438 heat treatment Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 229910000601 superalloy Inorganic materials 0.000 description 1
Images
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
- G06T7/0004—Industrial image inspection
- G06T7/0008—Industrial image inspection checking presence/absence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
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)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention provides a method for identifying defects of a hot-upset head of a high-temperature alloy bolt, namely a method for identifying defects of a hot-upset head of a high-temperature alloy bolt with difference in image brightness, which comprises the following implementation steps of: firstly, the method comprises the following steps: dividing a data set and preprocessing an image; II, secondly: establishing a residual error neural network model, and loading parameters of the model after pre-training; thirdly, the method comprises the following steps: replacing the output layer, and carrying out two times of training of the model; fourthly, the method comprises the following steps: verifying the model precision; through the steps, the problem of identifying the head defects of the bolts with different image brightness after hot heading is solved, the images with the brightness difference can be effectively classified under the condition of having the characteristic identification capability, the engineering practical situation is met, convenience is provided for engineering technicians, and important application value is achieved.
Description
Technical Field
The invention provides a method for identifying defects of a hot-upset head of a high-temperature alloy bolt, namely a method for identifying defects of a hot-upset head of a high-temperature alloy bolt with different image brightness, and relates to the technical field of computer vision and industrial automation, in particular to a method for identifying defects of a hot-upset head of a high-temperature alloy bolt with different image brightness.
Background
In manufacturing a superalloy bolt head, a hot heading process is typically used to upset the bolt head into shape. In the hot heading process, different bolt head defects can be caused by too low temperature, insufficient heading die depth and inaccurate heating position; in the actual production process, because the defects of the bolt head can be reflected on the appearance of the bolt head, field operators are often required to perform appearance detection to identify the defects of the bolt head; considering that the surface defect identification technology based on machine vision is a quality detection method without personnel participation, and can efficiently and quickly realize detection tasks; therefore, the defect identification technology based on machine vision is developed for identifying the defects on the surface of the head of the high-temperature alloy bolt subjected to hot heading, and has important significance;
the current defect identification technology based on machine vision mainly uses a deep learning network to carry out feature extraction and defect classification, the high precision of the deep learning network needs a large amount of training data to guarantee, and when the characteristics of the data change, the precision of the deep learning network is likely to be greatly reduced; in the actual hot heading process of the high-temperature alloy bolt, the high-temperature alloy bolt is usually a customized piece and has small batch, so that the generated defect data are few, and the training of a deep learning network cannot be supported; in addition, due to the influence of day and night, light rays in a workshop have large difference, and images acquired in the workshop have large brightness difference, so that the characteristics of data are changed, and the precision of a deep learning network is influenced;
in order to effectively identify the head defect of the high-temperature alloy bolt after hot upsetting based on machine vision under the actual condition that a small sample and image brightness are different, the invention provides a method for identifying the head defect of the high-temperature alloy bolt after hot upsetting based on the difference of the image brightness, and a method for identifying the surface defect by comprehensively applying a residual error neural network model, a cross entropy loss function and a transfer learning fine adjustment technology.
Disclosure of Invention
(1) The purpose of the invention is as follows: the invention provides a method for identifying the defects of the hot-upset head of a high-temperature alloy bolt aiming at the problem of identifying the defects of the hot-upset head of the bolt with different image brightness, namely a method for identifying the defects of the hot-upset head of the high-temperature alloy bolt with different image brightness, namely a method for identifying surface defects based on residual error neural network fine tuning; firstly, carrying out data set division and defect type marking on a picture of a bolt hot-upset head with image brightness difference, changing the size of the picture and carrying out normalization processing; step two, establishing a residual error neural network model, taking parameters obtained after pre-training as knowledge applicable to a target data set in a source data set, and loading the knowledge into the established model; freezing parameters of all layers except the output layer, replacing the output size of the output layer with the number of categories of the target data set, performing first round of training of the parameters of the output layer, then unfreezing the parameters of all layers in the model, and selecting a smaller learning rate to perform second round of integral training; fourthly, acquiring a head image of the bolt subjected to hot heading, and inputting the acquired image into the target residual error neural network model obtained in the third step for detection to obtain a predicted defect classification result;
(2) the technical scheme is as follows: based on the theory and thought, the invention provides a method for identifying the defects of the hot-upset head of the high-temperature alloy bolt, namely a method for identifying the defects of the hot-upset head of the high-temperature alloy bolt with difference in image brightness, which comprises the following specific implementation steps:
the method comprises the following steps: partitioning data sets and image pre-processing
Taking the surface defect image of the head after the hot heading of the bolt shot in a dark light scene as a source domain data set, and taking the surface defect image of the head after the hot heading of the bolt shot in a bright light scene as a target domain data set; the source domain data set and the target domain data set both contain various defect types, and the images of the source domain data set and the target domain data set are labeled with the defect types; then, each image is scaled to 224 x 224 pixel size, and then the data of each channel in the image is normalized,
wherein, said "X" is input "refers to the data value of the image in that channel; said "X out "means" X input "normalized data values; the mean refers to the mean value of the image in the channel, and the std represents the standard deviation of the image in the channel;
step two: building a residual neural network model, loading the pre-trained parameters of the model
Firstly, establishing a model container, referring to a structure of a ResNet50 residual neural network model, and adding six different modules from zero to five modules into the container, as shown in FIG. 1; the residual block is designed differently from other convolutional neural networks, namely the residual unit can be realized in a layer jump connection mode; in particular, assume that the net input to the l-th layer of the neural network is x l The function to be fitted by the neural network unit is H (·), then the residual structure x l+1 Can be expressed as
x l+1 ＝H(x l )＝x l +F(x l )#(4)
Wherein F (-) represents a residual function; then consider any two layers l 2 ＞l 1 The recursive expansion can be obtained, and,
in the formula:
l < th > representing a neural network 2 Net input of layers;
l < th > representing a neural network 1 Net input of layers;
thus, for deep level cell 2 It can be characterized as the sum of the outputs of the previous residual functions plus
In forward propagation, the input signal is propagated directly from any lower layer to a higher layer; the gradient of the final loss function L to a certain low-level output can be expanded into,
in the formula:
l < th > representing loss function L to neural network 1 Net input of layers
Partial derivatives of (a);
l < th > representing loss function to neural network 2 Net input of layers
Partial derivatives of (a);
l < th > representing a neural network 2 Net input of layers
To neural network 1 Net input of layers
Partial derivatives of (a);
l < th > representing a neural network 2 Lth of pre-layer residual function output sum to neural network 1 Net input of layers
Partial derivatives of (a); in the whole training process, the training device is provided with a training device,
the value can not be-1 all the time, so that the problem of gradient disappearance in a residual error network is ensured;
after the residual error neural network is established, because the number of the defect images of the head part of the bolt is less, the residual error neural network for defect identification cannot be directly trained, therefore, the parameters pre-trained on a large-scale data set are used as the initialization parameters of the model on the basis of the residual error neural network pre-trained on the large-scale public data set ImageNet; the ImageNet is a large public data set commonly used for researching visual object recognition methods;
step three: replacing the output layer and carrying out two times of training of the model
After a model with initialization parameters is established, because the final output layer contains specific classification parameters of ImageNet data sets for different images, the final output layer needs to be replaced by a full connection layer with the output size of the defect type of a target data set, and the parameters are initialized by uniform distribution; for a fully connected layer l fc The output data y, then the fully connected layer may be represented as,
in the formula:
parameters representing a fully connected layer;
a feature value representing an input; y represents output data;
in order to train only the parameters of the output layer, the gradient parameters of the other layers need to be frozen at first, and only the parameters of the full-connection layer are ensured to be updated in the training process; in the present invention, a cross entropy loss function (L) is selected cross_entropy ) As a loss function, the cross entropy loss function can be expressed as
In the formula: n represents the number of samples; m represents the number of categories; y is i，c For a sign function, y is when the true class of sample i is equal to c i，c Get 1 else y i，c Taking 0; p is a radical of formula i，c Representing the probability that sample i belongs to class c;
then, a random gradient descent method is utilized, a larger learning rate is set, and a source domain data set image (namely a head surface defect image after bolt hot heading shot in a dark scene) is adopted to train the model and update parameters; after the parameters of the full-link layer are trained and the primary effective classification precision is obtained, the knowledge of the data set can be assumed to be learned by the full-link layer of the model; however, other parameters in the model are still initialization parameters after pre-training, contain knowledge of the ImageNet data set, and are not completely suitable for the source domain data set and the target domain data set (namely, the defect image of the head surface of the bolt after hot heading, which is shot in a bright scene); therefore, the gradient of other layer parameters in the model is unfrozen, a random gradient descent method is utilized, meanwhile, the learning rate is set to be reduced, and the source domain data set image is adopted to carry out the integral training of the model and the updating of the parameters;
step four: the model precision verification collects the head defect image after the bolt is subjected to hot heading under any illumination condition without labeling, the image is preprocessed and then input into the residual error neural network model subjected to fine adjustment in the third step for testing to obtain a predicted defect classification result, and a defect identification result is obtained;
through the steps, the problem of recognizing the head defects of the bolts with different image brightness after hot heading is solved, and the images with brightness differences can be effectively classified under the condition of characteristic recognition capability through fine adjustment of the parameters of the pre-trained residual error neural network, so that the method is in accordance with the actual conditions of engineering, provides convenience for engineering technicians and has important application value.
(3) The advantages are that:
the invention provides a method for identifying the head defect of a high-temperature alloy bolt after hot heading, namely a method for identifying the head defect of the high-temperature alloy bolt after hot heading with difference in image brightness, wherein the parameters of a residual error neural network after pre-training are updated by using a parameter fine-tuning technology of transfer learning, so that the identification precision of the head defect of the bolt after hot heading is improved; in addition, the invention provides a new processing mode for the problem of identifying the surface defects of the head part after the bolt is subjected to hot heading, and the method can be applied to different types of surface defect classification problems and is easier to realize and convenient for engineering technicians to master and use.
Drawings
FIG. 1 is a flow chart of the method of the present invention
FIG. 2 is a diagram of a model architecture used in the present invention
The numbers, symbols and codes in FIG. 2 are illustrated as follows:
brackets (X, Y, Z) represent the size of the data, wherein X, Y and Z respectively represent the number of channels of the data, the width of the data and the height of the data; the numbers after the convolution layer respectively represent the size of the convolution kernel, the number of channels of output data and the convolution step length; the numbers behind the maximum pooling layer respectively represent the size of the pooling layer and the step length of pooling; the numbers behind the full connection layer respectively represent the number of channels of input data and the number of channels of output data; in the residual block 1, C denotes the number of channels of data, W denotes the width and height of data, C1 denotes the number of channels of output data, and S denotes the step size of convolution; in the residual block 2, C denotes the number of channels of data, and W denotes the width and height of the data.
Detailed Description
The invention provides a method for identifying defects of a hot-upset head of a high-temperature alloy bolt, namely a method for identifying defects of a hot-upset head of a high-temperature alloy bolt with different image brightness, namely a method for identifying defects of a hot-upset head of a bolt with different image brightness, wherein a flow chart is shown in figure 1; the implementation steps of the invention are further described in detail by taking the hot-headed head of the GH159 bolt as an example;
the method comprises the following steps: partitioning data sets and image pre-processing
The GH159 high-temperature alloy is often used for manufacturing a bolt special for an aeroengine, and due to the special material and the complex structure of the bolt head special for the aeroengine, the bolt head can be upset and formed only by using a hot upsetting process; acquiring two data sets with large brightness difference according to actual data acquisition conditions, wherein the two data sets comprise 185 defect pictures of the hot pier forming surface in a dark brightness scene and 179 defect pictures in a bright brightness scene; meanwhile, the defects of the head of the GH159 bolt subjected to hot heading are divided into 5 types: good appearance with no defects, slight folding, severe folding, unsaturations and unformed; then, each image is scaled to 224 multiplied by 224 pixel size, and then the data of each channel in the image stored by RGB format is normalized for the next time; finally, the image and the label thereof are stored in a data reader, and the batch size of the data reader is set to be 16 sheets at a time when the data reader reads the image, namely, the input size of the model is (16,3, 224) each time;
step two: building a residual neural network model, loading the pre-trained parameters of the model
Establishing a model container, and constructing modules from zero to five with convolution layers, pooling layers and the like by referring to the model structure shown in FIG. 2, wherein the construction of the residual block is shown in the right side of FIG. 2; then, sequentially connecting modules from zero to five according to the dotted line in the figure 2 to form a target residual error neural network model;
after a residual neural network is established, after the residual neural network is pre-trained on an ImageNet image data set, parameters of each module in a ResNet50 model capable of accurately identifying pictures are used as initialization parameters of the model, and the parameters are imported into a target model to be used as initialization parameters of the model;
step three: replacing the output layer and performing two times of training of the model
After a model with initialization parameters is established, replacing an original output layer of the model with a full connection layer with the output size of the defect type of a target data set, and initializing the model by using uniform distribution; then, freezing the gradient parameters of the other layers except the full connection layer; setting a loss function of the model as cross entropy loss, using an optimizer with descending random gradient, setting the learning rate to be 0.1, simultaneously, training every 10 steps, and reducing the learning rate to 0.2 times of the original learning rate, and carrying out 80 rounds of model training; the training result is expressed by the classification loss and the classification precision of the verification set;
on the basis of the classification result, the gradients of other layers in the model are unfrozen, the learning rate of the model is adjusted to 0.0001, and meanwhile, the learning rate is reduced to 0.8 time of the original learning rate every 10 steps of training; then, carrying out overall training of the model for 80 rounds and updating parameters;
step four: model accuracy verification
Respectively collecting a head defect image after hot heading of the bolt under dark brightness and bright brightness conditions without labeling, preprocessing the image, inputting the preprocessed image into the residual error neural network model subjected to fine adjustment in the third step, testing to obtain a predicted defect classification result, and obtaining a defect identification result; the test results are shown in table 1, and the classification accuracy of the luminance dark image data set obtained by direct training using the model with random initialization parameters is 51.16%, while the classification accuracy of the luminance bright image data set is 32.18%; initializing the model by using pre-training parameters, wherein the classification accuracy of the model obtained by training only the full connection layer on the brightness dark image data set is 95.34%, and the classification accuracy of the model obtained by training the full connection layer on the brightness bright image data set is 78.16%; the model obtained by initializing the model by using the pre-training parameters and performing fine adjustment twice has the classification accuracy of 98.84% on the brightness dark image data set and 93.10% on the brightness bright image data set;
TABLE 1
In conclusion, the GH159 bolt hot-upset head is taken as a case, and aiming at solving the problem of identifying the defects of the bolt hot-upset head with different image brightness, firstly, the images of the bolt hot-upset head with different image brightness are subjected to data set division and defect type marking, the size of the images is changed, and normalization processing is carried out; secondly, establishing a residual error neural network model, taking parameters obtained after pre-training as the knowledge of the target data set applicable to the source data set, and loading the model into the model; then, freezing parameters of all layers except the output layer, replacing the output size of the output layer to be the number of the categories of the target data set, and performing a first round of training of the parameters of the output layer; and finally, unfreezing the parameters of all layers in the model, and selecting a smaller learning rate to perform a second round of integral training to obtain a neural network model with the capability of extracting and classifying the surface defect characteristics of the hot-upset head of the bolt under a bright and dark scene, so that the identification precision of the hot-upset head of the bolt under the condition of difference of image brightness is effectively improved.
Claims (1)
1. A method for identifying the head defect of a high-temperature alloy bolt after hot heading is characterized by comprising the following steps: the method comprises the following steps:
the method comprises the following steps: partitioning data sets and image pre-processing
Taking the surface defect image of the head after the hot heading of the bolt shot in a dark light scene as a source domain data set, and taking the surface defect image of the head after the hot heading of the bolt shot in a bright light scene as a target domain data set; the source domain data set and the target domain data set both contain various defect types, and the images of the source domain data set and the target domain data set are labeled with the defect types; then, each image is scaled to 224 x 224 pixel size, and then the data of each channel in the image is normalized,
wherein, X input Refers to the data value of the image in the channel; x out Means X input A normalized data value; mean refers to the mean value of the image in the channel, std represents the standard deviation of the image in the channel;
step two: building a residual neural network model, loading the pre-trained parameters of the model
Firstly, establishing a model container, referring to a framework of a ResNet50 residual neural network model, and adding modules from zero to five, namely six different modules into the container; the residual block is different from the design of other convolutional neural networks, namely the residual unit is realized in a layer jump connection mode; specifically, let the net input of layer I of the neural network be x l The function to be fitted by the neural network unit is H (·), then the residual structure x l+1 Is expressed as
x l+1 ＝H(x l )＝x l +F(x l )
Wherein F (-) represents a residual function; then consider any two layers l 2 ＞l 1 And the result is obtained by recursive expansion,
in the formula:
l < th > representing a neural network 2 Net input of layers;
l < th > representing a neural network 1 Net input of layers;
thus, for deep level cell 2 Expressed as the sum of the outputs of the previous residual functions plus
In forward propagation, the input signal is propagated directly from any lower layer to a higher layer; the resulting loss function L spreads out the gradient of some low-level output as,
in the formula:
l < th > representing loss function L to neural network 1 Net input of layers
The partial derivatives of (1);
l < th > representing loss function L to neural network 2 Net input of layers
Partial derivatives of (a);
l < th > representing a neural network 2 Net input of layers
To neural network 1 Net input of layers
Partial derivatives of (a);
l < th > representing a neural network 2 Lth of pre-layer residual function output sum to neural network 1 Net input of layers
Partial derivatives of (a); in the whole training process, the training device is provided with a training device,
the value cannot be-1 all the time, so that gradient disappearance cannot occur in a residual error network;
after the residual error neural network is established, because the number of the defect images of the head part of the bolt is less, the residual error neural network for defect identification cannot be directly trained, therefore, the parameters pre-trained on a large-scale data set are used as the initialization parameters of the model on the basis of the residual error neural network pre-trained on the large-scale public data set ImageNet; ImageNet is a large public dataset for visual object recognition method studies;
step three: replacing the output layer and carrying out two times of training of the model
After a model with initialization parameters is established, because the final output layer contains specific classification parameters of ImageNet data sets for different images, the final output layer needs to be replaced by a full connection layer with the output size of the defect type of a target data set, and the parameters are initialized by uniform distribution; for a fully connected layer l fc And its output data, y, then the full connection layer is represented as,
in the formula:
parameters representing a fully connected layer;
a feature value representing an input; y represents the output data;
in order to train only the parameters of the output layer, the gradient parameters of the other layers need to be frozen at first, and only the parameters of the full-connection layer are ensured to be updated in the training process; selecting a cross entropy loss function L cross_entropy As a loss function, the cross entropy loss function is expressed as
In the formula: n represents the number of samples; m represents the number of categories; y is i，c For a sign function, y is when the true class of sample i is equal to c i，c Get 1 otherwise y i，c Taking 0; p is a radical of i，c Representing the probability that sample i belongs to class c;
then, setting a learning rate by using a random gradient descent method, and training and updating parameters by using a source domain data set image pair model;
after the parameters of the full-connection layer are trained and the primary effective classification precision is obtained, the full-connection layer of the model is set to learn the knowledge of the data set; however, other parameters in the model are still initialization parameters after pre-training, contain the knowledge of the ImageNet data set, and are not completely suitable for the source domain data set and the target domain data set; unfreezing gradients of other layer parameters in the model, reducing the learning rate by using a random gradient descent method, and performing overall training and parameter updating on the model by using a source domain data set image;
step four: model accuracy verification
And (4) acquiring a head defect image after the bolt is subjected to hot heading under any illumination condition, performing no labeling treatment, preprocessing the image, inputting the preprocessed image into the residual error neural network model subjected to fine adjustment in the third step, testing to obtain a predicted defect classification result, and obtaining a defect identification result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210421323.2A CN114881940A (en) | 2022-04-21 | 2022-04-21 | Method for identifying head defects of high-temperature alloy bolt after hot heading |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210421323.2A CN114881940A (en) | 2022-04-21 | 2022-04-21 | Method for identifying head defects of high-temperature alloy bolt after hot heading |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114881940A true CN114881940A (en) | 2022-08-09 |
Family
ID=82670848
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210421323.2A Pending CN114881940A (en) | 2022-04-21 | 2022-04-21 | Method for identifying head defects of high-temperature alloy bolt after hot heading |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114881940A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019169816A1 (en) * | 2018-03-09 | 2019-09-12 | 中山大学 | Deep neural network for fine recognition of vehicle attributes, and training method thereof |
| CN110728654A (en) * | 2019-09-06 | 2020-01-24 | 台州学院 | Automatic pipeline detection and classification method based on deep residual error neural network |
| CN112434646A (en) * | 2020-12-08 | 2021-03-02 | 浙江大学 | Finished tea quality identification method based on transfer learning and computer vision technology |
-
2022
- 2022-04-21 CN CN202210421323.2A patent/CN114881940A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019169816A1 (en) * | 2018-03-09 | 2019-09-12 | 中山大学 | Deep neural network for fine recognition of vehicle attributes, and training method thereof |
| CN110728654A (en) * | 2019-09-06 | 2020-01-24 | 台州学院 | Automatic pipeline detection and classification method based on deep residual error neural network |
| CN112434646A (en) * | 2020-12-08 | 2021-03-02 | 浙江大学 | Finished tea quality identification method based on transfer learning and computer vision technology |
Non-Patent Citations (1)
| Title |
|---|
| 黎磊等: "基于迁移学习的GH159螺栓热镦后头部缺陷识别", 《系统科学与数学》, 31 January 2022 (2022-01-31), pages 175 - 192 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
| CN113469953B (en) | Transmission line insulator defect detection method based on improved YOLOv4 algorithm | |
| CN110675462A (en) | Gray level image colorizing method based on convolutional neural network | |
| CN112749675A (en) | Potato disease identification method based on convolutional neural network | |
| CN116843650A (en) | SMT welding defect detection method and system integrating AOI detection and deep learning | |
| CN110363218A (en) | A kind of embryo's noninvasively estimating method and device | |
| CN111460947B (en) | BP neural network-based method and system for identifying metal minerals under microscope | |
| CN111461006B (en) | Optical remote sensing image tower position detection method based on deep migration learning | |
| CN115797808A (en) | Unmanned aerial vehicle inspection defect image identification method, system, device and medium | |
| CN116342536A (en) | Aluminum strip surface defect detection method, system and equipment based on lightweight model | |
| CN114549489A (en) | Carved lipstick quality inspection-oriented instance segmentation defect detection method | |
| CN112270404A (en) | Detection structure and method for bulge defect of fastener product based on ResNet64 network | |
| CN114881940A (en) | Method for identifying head defects of high-temperature alloy bolt after hot heading | |
| CN116958662A (en) | Steel belt defect classification method based on convolutional neural network | |
| CN115100451B (en) | Data expansion method for monitoring oil leakage of hydraulic pump | |
| CN115631186A (en) | Industrial element surface defect detection method based on double-branch neural network | |
| CN115049611A (en) | Continuous casting billet crack defect identification method based on improved yolov5 | |
| CN111046861B (en) | Method for identifying infrared image, method for constructing identification model and application | |
| CN114092441A (en) | Product surface defect detection method and system based on dual neural network | |
| CN114419529A (en) | Cross-modal pedestrian re-identification method and system based on distribution space alignment | |
| CN111652275B (en) | Sparse astragal identification model construction method, sparse astragal identification method and sparse astragal identification system | |
| CN117392440B (en) | Textile fabric retrieval method and system based on tissue structure and color classification | |
| CN112816408B (en) | Flaw detection method for optical lens | |
| CN112926619B (en) | High-precision underwater laser target recognition system | |
| CN117152687B (en) | Communication line state monitoring system |
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 |