CN115908843A - Superheat degree recognition model training method, recognition method, equipment and storage medium - Google Patents
Superheat degree recognition model training method, recognition method, equipment and storage medium Download PDFInfo
- Publication number
- CN115908843A CN115908843A CN202211377092.6A CN202211377092A CN115908843A CN 115908843 A CN115908843 A CN 115908843A CN 202211377092 A CN202211377092 A CN 202211377092A CN 115908843 A CN115908843 A CN 115908843A
- Authority
- CN
- China
- Prior art keywords
- module
- electrolyte
- yolo
- data set
- superheat degree
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000012549 training Methods 0.000 title claims abstract description 53
- 238000003860 storage Methods 0.000 title claims abstract description 9
- 239000003792 electrolyte Substances 0.000 claims abstract description 67
- 238000010586 diagram Methods 0.000 claims abstract description 26
- 238000007781 pre-processing Methods 0.000 claims abstract description 11
- 238000012545 processing Methods 0.000 claims abstract description 5
- 238000002372 labelling Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 31
- 238000001514 detection method Methods 0.000 claims description 21
- 230000004913 activation Effects 0.000 claims description 15
- 238000000605 extraction Methods 0.000 claims description 15
- 238000005070 sampling Methods 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 12
- 238000007499 fusion processing Methods 0.000 claims description 9
- 230000011218 segmentation Effects 0.000 claims description 9
- 230000004927 fusion Effects 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 8
- 230000009466 transformation Effects 0.000 claims description 3
- 230000008030 elimination Effects 0.000 claims description 2
- 238000003379 elimination reaction Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 230000003044 adaptive effect Effects 0.000 claims 1
- 238000013021 overheating Methods 0.000 abstract 1
- 239000000523 sample Substances 0.000 description 41
- XAGFODPZIPBFFR-UHFFFAOYSA-N aluminium Chemical compound [Al] XAGFODPZIPBFFR-UHFFFAOYSA-N 0.000 description 11
- 229910052782 aluminium Inorganic materials 0.000 description 11
- 238000005868 electrolysis reaction Methods 0.000 description 8
- 210000003128 head Anatomy 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 5
- 238000004519 manufacturing process Methods 0.000 description 5
- 238000005259 measurement Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 3
- 239000013078 crystal Substances 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000000877 morphologic effect Effects 0.000 description 2
- 239000013558 reference substance Substances 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical compound OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 description 1
- 241000282414 Homo sapiens Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 239000000155 melt Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a superheat degree recognition model training method, a superheat degree recognition method, equipment and a storage medium, wherein the training method comprises the steps of obtaining an electrolyte morphology image, and carrying out fire hole position marking to obtain a fire hole sample data set; performing background removal processing and category labeling on the electrolyte morphology image to obtain a category sample data set; preprocessing a fire hole sample data set to obtain a first characteristic diagram; the built YOLO-V5 network model comprises a backhaul part, a Neck part and a prediction layer part, wherein the backhaul part comprises a Focus layer, a MobileNet-V2 module and a first CSP module; training a YOLO-V5 network model by using the fire hole sample data set; preprocessing each sample in the class sample data set to obtain a second feature map; and training a MobileNet-V2 module of the YOLO-V5 network model by utilizing the class sample data set to obtain a final YOLO-V5 network model. The invention not only can automatically position the position of the fire hole, but also can accurately identify the overheating state.
Description
Technical Field
The invention belongs to the technical field of aluminum electrolysis, and particularly relates to an aluminum electrolysis cell electrolyte superheat degree state recognition model training method, recognition method, equipment and storage medium based on improved YOLO-V5, which are applied to recognition of electrolyte superheat degree states in an aluminum electrolysis production process.
Background
The superheat degree of the electrolyte of the aluminum electrolytic cell refers to the difference value between the temperature of the electrolyte and the temperature of primary crystals. At present, in the process of aluminum electrolysis production technology, the degree of superheat is one of key indexes reflecting the current production efficiency and quality of an aluminum electrolysis cell, and the proper degree of superheat is maintained, so that the current efficiency can be improved, the production process is stabilized, and the production energy consumption is reduced.
At present, the superheat degree measurement method needs to measure the electrolyte temperature and the primary crystal temperature respectively, the electrolyte temperature is generally measured on line by using a thermocouple or an infrared thermometer, and there are three methods for obtaining the primary crystal temperature:
one is to sample the electrolyte, and through chemical analysis, the electrolyte is obtained through calculation with empirical formula after knowing the components.
The second method is to sample the electrolyte and test the electrolyte by using a step curve method. No matter the first kind or the second kind, all carry out the off-line, its measurement process is comparatively complicated, and it is long to measure, all can't carry out real-time measurement, has great hysteresis quality.
The third is to design a reference substance probe to be inserted into the electrolyte melt and to analyze the temperature difference potential between the reference substance and the melt to measure the superheat degree. Although the third way can realize real-time measurement, the reference object probe is disposable, and a new probe is required for each measurement. Therefore, the superheat degree is judged on site basically by observing the information of the fire hole state of the aluminum electrolytic cell by experienced workers for a long time. However, since the subjective and random observation of the information of the fire hole state of the aluminum electrolytic cell by the naked eyes of workers is relatively large, the measurement levels are uneven, and the manpower is limited in time and space, a large amount of manpower resources are wasted, and the long-time monitoring and observation in 24h are difficult to realize.
In recent years, some experts and scholars hope to avoid the above disadvantages by computer vision technology instead of human beings, but the automation degree and the model reasoning speed of the current technology still have room for further improvement.
At present, two problems of aluminum electrolysis superheat degree state automatic identification through a computer vision technology are solved:
first, the degree of automation is insufficient. In the prior art, the appearance image of the electrolyte needs to be acquired by manually positioning the 'fire hole' opening, and the acquired image has a large amount of invalid background information, so that the key problem of how to realize unmanned automatic positioning of the 'fire hole' opening and only acquire the appearance image of the effective electrolyte is solved.
Secondly, the model parameter quantity is too large, the reasoning speed is too slow, and the single recognition work efficiency needs to be improved by improving the network reasoning speed.
Disclosure of Invention
The invention aims to provide a superheat degree recognition model training method, a superheat degree recognition method, superheat degree recognition equipment and a storage medium, and aims to solve the problems that the traditional technology cannot automatically position a fire hole, so that the automation degree is low, and the single recognition work efficiency is low.
The invention solves the technical problems through the following technical scheme: an electrolyte superheat recognition model training method comprises the following steps:
acquiring an electrolyte morphology image, and carrying out fire hole position marking on the electrolyte morphology image to obtain a fire hole sample data set; performing background removal processing and category labeling on the electrolyte morphology image to obtain a category sample data set;
preprocessing each sample in the fire hole sample data set to obtain a first characteristic diagram;
constructing a YOLO-V5 network model, wherein the YOLO-V5 network model comprises a backhaul part, a Neck part and a prediction layer part which are sequentially connected, and the backhaul part comprises a Focus layer, a MobileNet-V2 module and a first CSP module which are sequentially connected;
training the YOLO-V5 network model by using a fire hole sample data set formed by a first characteristic diagram to obtain a trained YOLO-V5 network model;
preprocessing each sample in the class sample data set to obtain a second feature map;
and training a MobileNet-V2 module of the trained YOLO-V5 network model by utilizing a class sample data set consisting of a second feature diagram, wherein parameters of the Focus layer, the first CSP module, the Neck part and the prediction layer part are unchanged, and obtaining the final YOLO-V5 network model.
Further, the specific implementation process of training the YOLO-V5 network model by using the fire hole sample data set composed of the first feature map is as follows:
performing feature extraction on the first feature map by using the Focus layer to obtain 304 × 304 feature maps;
performing feature extraction on the 304 × 304 feature map output by the Focus layer by using the MobileNet-V2 module to obtain 152 × 152, 76 × 76 and 38 × 38 feature maps;
segmenting the 152 x 152, 76 x 76 and 38 x 38 feature maps output by the MobileNet-V2 module by using the first CSP module, and performing merging operation through cross-stage layering operation to obtain merged 152 x 152, 76 x 76 and 38 x 38 feature maps;
using the heck part to perform upsampling and feature fusion on the 152 × 152, 76 × 76, 38 × 38 size feature maps output by the first CSP module to obtain 76 × 76, 38 × 38, 19 size feature maps;
and carrying out optimization and elimination of a large detection head and self-adaptive adjustment of an anchor frame on the size characteristic graphs of 76 × 76, 38 × 38 and 19 × 19 output by the Neck part by using the prediction layer part to obtain a target detection frame of the fire hole.
Further, the tack portion includes a second CSP module, a CBL module, and an SPP module;
the specific implementation process of performing upsampling and feature fusion on the 152 × 152, 76 × 76 and 38 × 38 feature maps output by the first CSP module by using the tack part is as follows:
performing segmentation and fusion processing on the 38 × 38 feature maps output by the first CSP module by using the second CSP module, and performing feature extraction by using the CBL module to obtain 38 × 38 feature maps;
utilizing the SPP module to carry out down-sampling on the 38 × 38 size characteristic map output by the CBL module to obtain a 19 × 19 size characteristic map;
the 19 × 19 feature maps output by the SPP module are subjected to up-sampling to obtain 38 × 38 feature maps, the second CSP module is used for carrying out segmentation and fusion processing on the 38 × 38 feature maps obtained by up-sampling, and the CBL module is used for carrying out feature extraction to obtain 38 × 38 feature maps;
and performing up-sampling on the 19 × 19 size characteristic graphs output by the SPP module to obtain 76 × 76 size characteristic graphs, performing segmentation and fusion processing on the 76 × 76 size characteristic graphs obtained by the up-sampling by the second CSP module, and performing characteristic extraction by the CBL module to obtain 76 × 76 size characteristic graphs.
Further, the prediction layer part comprises a convolution layer and 3 detection heads with the sizes of 76 × 76 × 255, 38 × 38 × 255 and 19 × 19 × 255 respectively;
in the prediction layer part, a 76 x 255 detection head for a large target is optimized and removed, an original anchor frame is adaptively adjusted to be [10,14,23,27,37,58], [81,82,135,169,344,319], and finally a target detection frame of a fire hole is output based on a loss function and back propagation.
Further, the MobileNet-V2 module includes an avgpool module, 7 bottleeck modules, a batch normalization layer, and an activation function layer, where the bottleeck module includes a channel-by-channel convolution layer and a point-by-point convolution layer.
Further, the expression of the normalized loss function of the batch normalization layer is:
in the formula,
to normalize the loss function, x (k) Is a loss function value after linear transformation of a batch normalization layer, E [ [ alpha ] ]]Representing the mean of the loss function values, var is the mean square operator.
Further, the activation function layer adopts an LeakyReLU activation function, and the specific expression of the activation function layer is as follows:
where f (i) is the LeakyReLU activation function, and i represents the profile input value.
Based on the same invention concept, the invention also provides an electrolyte superheat degree identification method, which comprises the following steps:
obtaining an electrolyte morphology image;
preprocessing the electrolyte morphology image to obtain a characteristic diagram;
training according to the training method of the electrolyte superheat degree recognition model to obtain an electrolyte superheat degree recognition model;
and identifying the position and the superheat degree of a fire hole of each characteristic diagram by adopting the electrolyte superheat degree identification model.
Based on the same inventive concept, the present invention also provides an electronic device, comprising:
at least one processor; and the number of the first and second groups,
a memory communicatively coupled to the at least one processor; wherein,
the memory stores instructions executable by the at least one processor, which when executed by the at least one processor, cause the at least one processor to perform steps in the electrolyte superheat recognition model training method as described above, or to perform steps in the electrolyte superheat recognition method as described above.
Based on the same inventive concept, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions that, when executed by at least one processor, cause the at least one processor to perform the steps in the electrolyte superheat recognition model training method as described above, or to perform the steps in the electrolyte superheat recognition method as described above.
Advantageous effects
Compared with the prior art, the invention has the advantages that:
according to the superheat degree recognition model training method, the recognition method, the equipment and the storage medium, provided by the invention, a main network is replaced by a MobileNet-V2 lightweight network structure from DarkNet on the basis of YOLO-V5; in the training process, firstly, the improved YOLO-V5 network model is trained for recognizing the position of the fire hole by using the fire hole sample data set, and then, mobileNet-V2 in the improved YOLO-V5 network model is trained for recognizing the superheat degree state independently by using the category sample data set, so that the accuracy of the superheat degree state recognition of the model is greatly improved, and the final YOLO-V5 network model not only can automatically position the position of the fire hole, but also can accurately recognize the superheat degree state.
According to the invention, a MobileNet-V2 lightweight network structure is adopted to replace DarkNet, the inference speed is improved while the identification precision is ensured, meanwhile, a large amount of invalid background information is automatically removed from the adopted image, only the effective electrolyte melt morphology image is retained, and the quick and accurate identification of the aluminum electrolysis superheat degree state is realized.
Drawings
In order to more clearly illustrate the technical solution of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only one embodiment of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
FIG. 1 is a flow chart of a training method of an electrolyte superheat degree recognition model in an embodiment of the invention;
FIG. 2 is a diagram of a backbone network structure of a YOLO-V5 network model in an embodiment of the present invention;
FIG. 3 is a block diagram of the Neck portion and the predicted layer portion of the YOLO-V5 network model in an embodiment of the present invention;
FIG. 4 is a schematic structural diagram of a MobileNet-V2 module in an embodiment of the present invention;
fig. 5 is a schematic diagram of the operation of a conventional 3 x 4 convolutional layer in an embodiment of the present invention;
FIG. 6 is a schematic diagram of a per-channel convolution operation in accordance with an embodiment of the present invention;
FIG. 7 is a schematic diagram of a point-by-point convolution operation in an embodiment of the present invention;
FIG. 8 is a curve of the superheat degree identification accuracy of the MobileNet-V2 model in the embodiment of the present invention;
FIG. 9 is a schematic diagram of the GIoU curve of the YOLO-V5 network model in the embodiment of the present invention;
FIG. 10 is a graph of the average accuracy parameter of the YOLO-V5 network model in an embodiment of the present invention.
Detailed Description
The technical solutions in the present invention are clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without inventive efforts based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.
The technical solution of the present application will be described in detail below with specific examples. The following several specific embodiments may be combined with each other, and details of the same or similar concepts or processes may not be repeated in some embodiments.
As shown in fig. 1, the method for training the electrolyte superheat degree recognition model according to the embodiment includes the following steps:
step 1: making a fire hole sample data set and a category sample data set
Step 1.1: making a fire hole sample data set: and shooting by using an industrial camera to obtain an electrolyte morphology image, and carrying out fire hole position marking on the electrolyte morphology image to obtain position labels of the upper left coordinate and the lower right coordinate of the fire hole area, thereby obtaining a fire hole sample data set.
In this embodiment, 24000 images of the electrolyte morphology are obtained, and the obtained fire hole sample data set includes 24000 fire hole samples. And (3) dividing a training set and a test set of the fire hole sample data set according to the proportion of 8.
Step 1.2: making a category sample data set: and (3) acquiring an electrolyte morphology image by utilizing the shooting of an industrial camera, and performing background removal processing and category labeling on the electrolyte morphology image to obtain a category sample data set.
The extraction of the morphological characteristics of the electrolyte by the YOLO-V5 network model is seriously influenced by the background, and when the class sample data set is manufactured, the image is cut and segmented, so that only the effective morphological image of the electrolyte melt is reserved, and the superheat degree state recognition can be accelerated. In this embodiment, the category labels include too high, normal, low, and too low. And (2) dividing the training set and the test set of the class sample data set according to the proportion of 8.
In this embodiment, the fire hole sample set is used for performing fire hole position recognition training on the YOLO-V5 network model, the category sample set performs individual superheat degree state recognition training only on the MobileNet-V2 module in the YOLO-V5 network model, and other part parameters of the YOLO-V5 network model are kept unchanged.
Step 2: and preprocessing each sample in the fire hole sample data set to obtain a first characteristic diagram.
Each fire hole sample has a size of 1920 × 1080, and each fire hole sample is preprocessed to obtain a 608 × 608 size first feature map.
And step 3: construction of YOLO-V5 network model
A YOLO-V5 network model is constructed, as shown in FIGS. 1 to 3, the YOLO-V5 network model comprises a backhaul part, a Neck part and a prediction layer part which are sequentially connected, and the backhaul part comprises a Focus layer, a MobileNet-V2 module and a first CSP module (namely a cross-stage local network) which are sequentially connected. The YOLO-V5 network model is improved on the basis of the traditional YOLO-V5 network, and the lightweight MobileNet-V2 is used for replacing DarkNet in a main network part, namely, the deep separable convolution is used for replacing the common convolution operation, so that the precision is ensured, and the inference speed is increased.
And 4, step 4: and (3) training the YOLO-V5 network model constructed in the step (2) by using the fire eye sample data set.
In this embodiment, the specific training process of the YOLO-V5 network model is as follows:
step 4.1: extracting features of the first feature map by using a Focus layer to obtain 304 × 304 feature maps;
step 4.2: performing feature extraction on the 304 × 304 feature map output by the Focus layer by using a MobileNet-V2 module to obtain 152 × 152, 76 × 76 and 38 × 38 feature maps;
step 4.3: the first CSP module is used for segmenting the 152 x 152, 76 x 76 and 38 x 38 size characteristic graphs output by the MobileNet-V2 module, and merging operation is carried out through cross-stage layering operation, so that merged 152 x 152, 76 x 76 and 38 x 38 size characteristic graphs are obtained;
step 4.4: using the Neck part to perform upsampling and feature fusion on the 152 × 152, 76 × 76 and 38 × 38 feature maps output by the first CSP module to obtain 76 × 76, 38 × 38 and 19 × 19 feature maps;
step 4.5: and (5) optimizing and eliminating the large detection head and adaptively adjusting the anchor frame by using the prediction layer part on the size characteristic graphs 76 × 76, 38 × 38 and 19 × 19 output by the Neck part to obtain the target detection frame of the fire hole.
The MobileNet-V2 module includes an avgpool module, seven bottomleck modules, a batch normalization layer, and an activation function layer, as shown in fig. 4, wherein the bottomleck module includes a 3 × 3 channel-by-channel convolution layer (as shown in fig. 6) and two 1 × 1 point-by-point convolution layers (as shown in fig. 7). By replacing the normal convolution operation with a deep separable convolution, the number of 8/9 convolution parameters can be reduced while ensuring the same performance. In the convolution operation process of the MobileNet-V2 module, firstly, a 1x 1 point-by-point convolution layer is used for carrying out dimension increasing operation, the feature diagram after dimension increasing is sent to a 3 x 3 channel-by-channel convolution layer, the feature diagram obtained from the 3 x 3 channel-by-channel convolution layer is input to the 1x 1 point-by-point convolution layer again for carrying out dimension reducing operation, and finally, not only can better model capacity be obtained, but also the operation amount of convolution operation can be further reduced, and the reasoning speed is improved.
The expression of the normalized loss function of the batch normalization layer is:
in the formula,
to normalize the loss function, x (k) Is the value of the loss function after linear transformation through the layer, E [ 2 ]]Representing the mean of the loss function values, var is the mean square operator.
The activation function layer adopts an LeakyReLU activation function, and the specific expression of the activation function layer is as follows:
where f (i) is the LeakyReLU activation function, and i represents the profile input value.
And the Neck part performs upsampling, feature fusion and other processing on tensor characteristic maps of different scales output by the Backbone part to obtain the tensor characteristic maps of different scales. As shown in fig. 2 and 3, the hack section includes a second CSP module (cross-stage local network), a CBL module (tandem convolutional layer module), and an SPP module (spatial pyramid pooling module). The specific implementation process of using the tack part to perform upsampling and feature fusion on the 152 × 152, 76 × 76, 38 × 38 feature maps output by the first CSP module is as follows:
step 4.41: the second CSP module is used for carrying out segmentation and fusion processing on the 38X 38 large and small feature maps output by the first CSP module, and then the CBL module is used for carrying out feature extraction to obtain 38X 38 large and small feature maps;
step 4.42: utilizing the SPP module to perform down-sampling on the 38 × 38 size characteristic map output by the CBL module in the step 4.21 to obtain a 19 × 19 size characteristic map;
step 4.43: performing up-sampling on the 19 × 19 size feature maps output by the SPP module in the step 4.22 to obtain 38 × 38 size feature maps, performing segmentation and fusion processing on the 38 × 38 size feature maps obtained by the up-sampling by the second CSP module, and performing feature extraction by the CBL module to obtain 38 × 38 size feature maps;
step 4.44: and (3) performing up-sampling on the 19 × 19 size feature map output by the SPP module in the step 4.22 to obtain a 76 × 76 size feature map, performing segmentation and fusion processing on the 76 × 76 size feature map obtained by the up-sampling by the second CSP module, and performing feature extraction by the CBL module to obtain a 76 × 76 size feature map.
In the Neck part, a cross-phase local network structure designed by a deep network (CSPnet) with enhanced learning ability is adopted to enhance the network feature fusion ability and reduce the size of a network model. Specifically, the characteristic diagram output by the main network is further extracted through a cross-stage local network and a convolutional layer module, and then the characteristic diagram is processed in 3 scales through an SPP module, so that an image characteristic matrix output by the main network can be better utilized, and an electrolyte morphology image characteristic diagram with more scales is obtained; and finally, transferring the processed tensor characteristic images with different scales to a prediction layer part.
As shown in fig. 3, the prediction layer part includes a convolution layer and 3 detection heads having sizes of 76 × 76 × 255, 38 × 38 × 255, and 19 × 19 × 255, respectively. Aiming at the problems of target stray distribution and small target pixel ratio caused by the view angle of an industrial camera, in a prediction layer part, a 76 x 255 detection head for a large target is optimized and removed, an original anchor frame (namely anchor box) is adaptively adjusted to be [10,14,23,27,37,58], [81,82,135,169,344 and 319], and finally, a target detection frame of a fire eye opening is output based on a loss function and back propagation.
And 5: and preprocessing each sample in the class sample data set to obtain a second characteristic diagram.
The size of each class sample is 1920 × 1080, and each class sample is preprocessed to obtain 608 × 608 second feature maps.
Step 6: and (4) independently training the MobileNet-V2 module in the YOLO-V5 network model trained in the step (4) by utilizing the class sample data set.
And (3) training a MobileNet-V2 module of the trained YOLO-V5 network model by utilizing a class sample data set formed by a second feature map, wherein parameters (such as weight values) of the Focus layer, the first CSP module, the Neck part and the prediction layer part are unchanged, namely the Focus layer, the first CSP module, the Neck part and the prediction layer part are not changed along with the training of the MobileNet-V2 module, and obtaining the final YOLO-V5 network model.
In this embodiment, the output result of the YOLO-V5 network model is evaluated by using generalized intersection-parallel ratio (GIOU), average precision (MAP), and inference speed. The generalized intersection ratio is a regression target frame loss function, and has the following characteristics when taken as an evaluation index: nonnegativity, symmetry and scale invariance, and the smaller the generalization intersection ratio is, the higher the output precision of the target frame is. The calculation formula of the generalized intersection ratio is as follows:
in the formula, GIoU represents generalized cross-over ratio, ioU represents cross-over ratio, a and B represent target detection frames of any two fire eyes, C represents a minimum square frame capable of enclosing a and B, | C \ a | B | represents the area of C minus the union of a and B, | C | represents the area of C, | a | B | represents the area of the union of a frame and B frame, | a | B | represents the area of the intersection of a frame and B frame.
The average precision is an index for measuring the detection precision of the multi-label image. In the multi-standard image detection task, more than one image is labeled, and the Average Precision (MAP) calculation Precision similar to that in information retrieval should be adopted. The larger the average accuracy value is, the higher the target detection accuracy is. The average precision is calculated by drawing a PR curve, namely a two-dimensional curve taking precision and recall as vertical and horizontal axis coordinates. Generally, precision is accuracy and recall is recall.
The inference speed is defined as the number of detectable images in one second. The faster the reasoning speed, the better the real-time performance of the target detection network.
Simulation experiment: the platform of the implementation mode is a Windows10 operating system, and the development environment is Pycharm Community Edition 2021.1.1x64. The experimental model is on a Pytroch 1.10.0 framework and adopts a MobileNet-V2 learning network. Model training is completed in an experimental environment of Nvdia 3060Ti (video memory 12G) GPU and CUDA 11.2.
The training parameters of the YOLO-V5 network model are set as: the number of training rounds is 200 rounds, bitch size is 16, namely, the images put in one time are 16 pictures, and the initial learning rate is 0.001.
FIG. 8 is a superheat state identification accuracy curve of MobileNet-V2, and when the iteration is about 160 rounds, the superheat state identification accuracy of the model can reach 99.6%, which is greatly improved compared with the previous 95% highest identification accuracy. In fig. 9, the ordinate is a generalized intersection ratio, the abscissa is the number of training rounds, and the YOLO-V5 network model has a smaller generalized intersection ratio than the conventional YOLO-V5 model in about 200 iterations, and the target frame output accuracy is higher. Fig. 10 is a graph of an average accuracy parameter of the YOLO V5 network model, a vertical axis of the graph is an average accuracy value, a horizontal axis of the graph is a number of training rounds, and after about 200 iterations, the average accuracy value of the YOLO V5 network model is 45.4, the average accuracy value of the traditional YOLO V5 is 37.2, and the target detection accuracy is higher. In the aspect of reasoning speed, the YOLO V5 network model can detect 158 pictures in one second, the traditional YOLO V5 network can detect 96 pictures in one second, the detection speed of the YOLO V5 network model is improved by 22%, and the real-time performance is better. Table 1 shows experimental comparison data of traditional YOLO-V5 and the YOLO V5 network model of the application in three aspects of parameter quantity, reasoning speed and average precision, and the application changes a main network in the YOLO-V5 into a lightweight MobileNet-V2 network, so that the parameter quantity is greatly reduced from 7.2M to 1.7M.
Table 1 experimental comparison data
| Model | Reference quantity (M) | Reasoning speed (ms) | Average Precision (MAP) |
| Conventional YOLO-V5 | 7.2 | 0.9 | 37.2 |
| This application YOLO-v5 | 1.7 | 0.63 | 45.4 |
Based on the same inventive concept, the embodiment of the invention also provides an electrolyte superheat degree identification method, which comprises the following steps:
step 1: obtaining an electrolyte morphology image;
and 2, step: preprocessing the electrolyte morphology image to obtain a characteristic diagram;
and 3, step 3: training according to the training method of the electrolyte superheat degree recognition model to obtain an electrolyte superheat degree recognition model;
and 4, step 4: and identifying the position and the superheat degree of a fire hole of each characteristic graph by adopting the electrolyte superheat degree identification model.
According to the superheat degree recognition model training method, the superheat degree recognition model training device and the superheat degree recognition model storage medium, a backbone network is replaced by a MobileNet-V2 lightweight network structure from DarkNet on the basis of YOLO-V5; in the training process, firstly, the improved YOLO-V5 network model is trained for recognizing the position of the fire hole by using the fire hole sample data set, and then, the MobileNet-V2 in the improved YOLO-V5 network model is trained for recognizing the superheat degree state independently by using the category sample data set, so that the accuracy of the recognition of the superheat degree state of the model is greatly improved, and the final YOLO-V5 network model not only can automatically position the position of the fire hole, but also can accurately recognize the superheat degree state. According to the invention, a MobileNet-V2 lightweight network structure is adopted to replace DarkNet, the inference speed is improved while the identification precision is ensured, meanwhile, a large amount of invalid background information is automatically removed from the adopted image, only the effective electrolyte melt morphology image is retained, and the quick and accurate identification of the aluminum electrolysis superheat degree state is realized.
The above disclosure is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of changes or modifications within the technical scope of the present invention, and shall be covered by the scope of the present invention.
Claims (10)
1. A training method for an electrolyte superheat degree recognition model is characterized by comprising the following steps:
acquiring an electrolyte morphology image, and carrying out fire hole position marking on the electrolyte morphology image to obtain a fire hole sample data set; performing background removal processing and category labeling on the electrolyte morphology image to obtain a category sample data set;
preprocessing each sample in the fire hole sample data set to obtain a first characteristic diagram;
constructing a YOLO-V5 network model, wherein the YOLO-V5 network model comprises a backhaul part, a Neck part and a prediction layer part which are sequentially connected, and the backhaul part comprises a Focus layer, a MobileNet-V2 module and a first CSP module which are sequentially connected;
training the YOLO-V5 network model by using a fire hole sample data set formed by a first characteristic diagram to obtain a trained YOLO-V5 network model;
preprocessing each sample in the class sample data set to obtain a second feature map;
and training a MobileNet-V2 module of the trained YOLO-V5 network model by utilizing a class sample data set consisting of a second feature diagram, wherein parameters of the Focus layer, the first CSP module, the Neck part and the prediction layer part are unchanged, and obtaining the final YOLO-V5 network model.
2. The training method of the electrolyte superheat degree recognition model according to claim 1, wherein the specific implementation process of training the YOLO-V5 network model by using a fire hole sample data set composed of a first feature map comprises the following steps:
performing feature extraction on the first feature map by using the Focus layer to obtain 304 × 304 feature maps;
performing feature extraction on the 304 × 304 size feature map output by the Focus layer by using the MobileNet-V2 module to obtain 152 × 152, 76 × 76 and 38 × 38 size feature maps;
the first CSP module is used for segmenting the 152 x 152, 76 x 76 and 38 x 38 size characteristic graphs output by the MobileNet-V2 module, and merging operation is carried out through cross-stage layering operation, so that merged 152 x 152, 76 x 76 and 38 x 38 size characteristic graphs are obtained;
using the heck part to perform upsampling and feature fusion on the 152 × 152, 76 × 76, 38 × 38 size feature maps output by the first CSP module to obtain 76 × 76, 38 × 38, 19 size feature maps;
and carrying out optimization elimination on the size characteristic graphs of 76 × 76, 38 × 38 and 19 × 19 output by the Neck part by using the prediction layer part, and carrying out self-adaptive adjustment on an anchor frame to obtain a target detection frame of the fire hole.
3. The training method of an electrolyte superheat recognition model according to claim 2, wherein the tack section includes a second CSP module, a CBL module, and an SPP module;
the specific implementation process of performing upsampling and feature fusion on the 152 × 152, 76 × 76 and 38 × 38 feature maps output by the first CSP module by using the tack part is as follows:
performing segmentation and fusion processing on the 38 × 38 feature maps output by the first CSP module by using the second CSP module, and performing feature extraction by using the CBL module to obtain 38 × 38 feature maps;
utilizing the SPP module to carry out down-sampling on the 38 × 38 size characteristic map output by the CBL module to obtain a 19 × 19 size characteristic map;
the 19 × 19 feature maps output by the SPP module are subjected to up-sampling to obtain 38 × 38 feature maps, the second CSP module is used for carrying out segmentation and fusion processing on the 38 × 38 feature maps obtained by up-sampling, and the CBL module is used for carrying out feature extraction to obtain 38 × 38 feature maps;
and performing up-sampling on the 19 × 19 size feature maps output by the SPP module to obtain 76 × 76 size feature maps, performing segmentation and fusion processing on the 76 × 76 size feature maps obtained by the up-sampling by the second CSP module, and performing feature extraction by the CBL module to obtain 76 × 76 size feature maps.
4. The training method of an electrolyte superheat degree recognition model according to claim 2, wherein the prediction layer portion includes a convolution layer and 3 detection heads having sizes of 76 × 76 × 255, 38 × 38 × 255, and 19 × 19 × 255, respectively;
in the prediction layer part, the 76 × 76 × 255 detection heads for large targets are removed in an optimized mode, the original anchor frame is adjusted to be [10,14,23,27,37,58], [81,82,135,169,344,319] in an adaptive mode, and finally the target detection frame of the fire hole is output based on a loss function and back propagation.
5. The training method of the electrolyte superheat degree identification model according to any one of claims 1 to 4, wherein the MobileNet-V2 module comprises an avgpool module, 7 bottleeck modules, a batch normalization layer and an activation function layer, and the bottleeck modules comprise channel-by-channel convolution layers and point-by-point convolution layers.
6. The training method of the electrolyte superheat recognition model according to claim 5, wherein the expression of the normalized loss function of the batch normalization layer is:
in the formula,
to normalize the loss function, x (k) Is a loss function value after linear transformation of a batch normalization layer, E [ [ alpha ] ]]Representing the mean of the loss function values, var is the mean square operator.
7. The training method of the electrolyte superheat degree recognition model according to claim 5, wherein the activation function layer adopts a LeakyReLU activation function, and the specific expression of the activation function layer is as follows:
where f (i) is the LeakyReLU activation function, and i represents the profile input value.
8. An electrolyte superheat degree identification method, characterized by comprising:
obtaining an electrolyte morphology image;
preprocessing the electrolyte morphology image to obtain a characteristic diagram;
training according to the electrolyte superheat degree recognition model training method of any one of claims 1 to 7 to obtain an electrolyte superheat degree recognition model;
and identifying the position and the superheat degree of a fire hole of each characteristic diagram by adopting the electrolyte superheat degree identification model.
9. An electronic device, comprising:
at least one processor; and the number of the first and second groups,
a memory communicatively coupled to the at least one processor; wherein,
the memory stores instructions executable by the at least one processor, the instructions, when executed by the at least one processor, causing the at least one processor to perform the steps of the electrolyte superheat recognition model training method of any one of claims 1 to 7, or to perform the steps of the electrolyte superheat recognition method of claim 8.
10. A non-transitory computer readable storage medium storing computer instructions which, when executed by at least one processor, cause the at least one processor to perform the steps in the electrolyte superheat recognition model training method of any one of claims 1 to 7, or to perform the steps in the electrolyte superheat recognition method of claim 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211377092.6A CN115908843A (en) | 2022-11-04 | 2022-11-04 | Superheat degree recognition model training method, recognition method, equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211377092.6A CN115908843A (en) | 2022-11-04 | 2022-11-04 | Superheat degree recognition model training method, recognition method, equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115908843A true CN115908843A (en) | 2023-04-04 |
Family
ID=86494741
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211377092.6A Pending CN115908843A (en) | 2022-11-04 | 2022-11-04 | Superheat degree recognition model training method, recognition method, equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115908843A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173200A (en) * | 2023-11-03 | 2023-12-05 | 成都数之联科技股份有限公司 | Image segmentation method, device, equipment and medium |
-
2022
- 2022-11-04 CN CN202211377092.6A patent/CN115908843A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173200A (en) * | 2023-11-03 | 2023-12-05 | 成都数之联科技股份有限公司 | Image segmentation method, device, equipment and medium |
| CN117173200B (en) * | 2023-11-03 | 2024-02-02 | 成都数之联科技股份有限公司 | Image segmentation method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598736B (en) | Power equipment infrared image fault positioning, identifying and predicting method | |
| WO2022012110A1 (en) | Method and system for recognizing cells in embryo light microscope image, and device and storage medium | |
| CN113592828B (en) | Nondestructive testing method and system based on industrial endoscope | |
| CN111860106B (en) | Unsupervised bridge crack identification method | |
| CN111429425B (en) | Rolling mill entrance abnormity identification method based on deep learning | |
| CN112258490A (en) | Low-emissivity coating intelligent damage detection method based on optical and infrared image fusion | |
| CN114749342A (en) | Method, device and medium for identifying coating defects of lithium battery pole piece | |
| CN115527234A (en) | Infrared image cage dead chicken identification method based on improved YOLOv5 model | |
| CN111144749A (en) | Intelligent labeling crowdsourcing platform for power grid images and working method | |
| CN111259733A (en) | Point cloud image-based ship identification method and device | |
| CN115908843A (en) | Superheat degree recognition model training method, recognition method, equipment and storage medium | |
| CN113052295A (en) | Neural network training method, object detection method, device and equipment | |
| CN112949510A (en) | Human detection method based on fast R-CNN thermal infrared image | |
| CN115240119A (en) | Pedestrian small target detection method in video monitoring based on deep learning | |
| CN115171045A (en) | YOLO-based power grid operation field violation identification method and terminal | |
| CN116295022A (en) | Pig body ruler measurement method based on deep learning multi-parameter fusion | |
| CN113706496B (en) | Aircraft structure crack detection method based on deep learning model | |
| CN115035082A (en) | YOLOv4 improved algorithm-based aircraft transparency defect detection method | |
| CN114881286A (en) | Short-time rainfall prediction method based on deep learning | |
| CN113313678A (en) | Automatic sperm morphology analysis method based on multi-scale feature fusion | |
| CN117253192A (en) | Intelligent system and method for silkworm breeding | |
| CN112270370A (en) | Vehicle apparent damage assessment method | |
| CN116363532A (en) | Unmanned aerial vehicle image traffic target detection method based on attention mechanism and re-parameterization | |
| CN110738208A (en) | efficient scale-normalized target detection training method | |
| CN114022831A (en) | Binocular vision-based livestock body condition monitoring method and 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 |