Superheat degree recognition model training method, recognition method, device and storage medium

Abstract

The application discloses a superheat degree recognition model training method, a recognition method, equipment and a storage medium, the training method comprises the following steps: obtaining an electrolyte topographic image, marking a fire eye opening position, and obtaining a fire eye sample data set; performing background removal processing and category labeling on the electrolyte topographic image, and obtaining a category sample data set; preprocessing the fire eye sample data set to obtain a first feature map; the constructed YOLO-V5 network model comprises a Backbone part, a Neck part and a prediction layer part, the Backbone part comprises a Focus layer, a MobileNet-V2 module and a first CSP module; the YOLO-V5 network model is trained by using the fire eye sample data set; each sample in the category sample data set is preprocessed to obtain a second feature map; the MobileNet-V2 module of the YOLO-V5 network model is trained by using the category sample data set, and a final YOLO-V5 network model is obtained. The application can not only automatically locate the fire eye opening position, but also accurately recognize the superheat degree state.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aluminum electrolysis, and particularly relates to an electrolyte superheat state recognition model training method and recognition method for an aluminum electrolysis cell based on improved YOLO-V5, an equipment, and a storage medium, which are applied to recognition of the electrolyte superheat state in an aluminum electrolysis production process. BACKGROUND

[0002] The superheat of the electrolyte of the aluminum electrolysis cell refers to the difference between the electrolyte temperature and the primary crystallization temperature. At present, in the aluminum electrolysis production process, the superheat is one of the key indicators reflecting the current production efficiency and quality of the aluminum electrolysis cell. Maintaining an appropriate superheat can improve the current efficiency, stabilize the production process, and reduce the production energy consumption.

[0003] At present, the superheat measurement method needs to measure the electrolyte temperature and the primary crystallization temperature respectively. The electrolyte temperature is generally measured online by using a thermocouple or an infrared thermometer, while there are currently three methods to obtain the primary crystallization temperature:

[0004] One is to sample the electrolyte, understand the composition through chemical analysis, and then calculate and obtain by using an empirical formula.

[0005] The second is to sample the electrolyte and test and obtain by using a step cooling curve method. Whether it is the first method or the second method, they are both offline, the measurement process is relatively complex, the measurement time is relatively long, and they cannot be measured in real time, and there is a large hysteresis.

[0006] The third is to design a reference probe to be inserted into the electrolyte melt, and analyze the temperature difference potential between the reference and the melt to measure the superheat. Although the third method can realize real-time measurement, the reference probe is disposable, and a new probe must be used for each measurement. Therefore, for a long time, the superheat is basically judged by experienced workers observing the fire eye state information of the aluminum electrolysis cell. However, since the workers observe the fire eye state information of the aluminum electrolysis cell with their naked eyes, there is a large subjectivity and randomness, the measurement level is uneven, and the manpower has spatial and temporal limitations, which not only wastes a large amount of human resources, but also is difficult to realize 24h long-time monitoring and observation.

[0007] In recent years, some experts and scholars hope to replace manual work with computer vision technology to avoid the above-mentioned shortcomings, but the automation degree and model inference speed of the current technology still have room for further improvement.

[0008] At present, there are two difficulties in automatically identifying the aluminum electrolysis superheat state by using computer vision technology:

[0009] First, the degree of automation is insufficient. The prior art needs to manually position the "fire eye" opening to collect electrolyte topography images, and the collected images have a lot of invalid background information. How to automatically position the "fire eye" opening and only collect effective electrolyte topography images is a key problem.

[0010] Second, the model parameter quantity is too large, and the inference speed is too slow. It is necessary to improve the network inference speed to improve the single identification work efficiency. SUMMARY

[0011] The purpose of the present application is to provide an overheat identification model training method, an identification method, an equipment and a storage medium, so as to solve the problems of low automation degree caused by the inability of traditional technology to automatically position the fire eye opening, and low single identification work efficiency.

[0012] The present application solves the above technical problems through the following technical solutions: an electrolyte overheat identification model training method, comprising the following steps:

[0013] An electrolyte topography image is obtained, and the electrolyte topography image is labeled with a fire eye opening position to obtain a fire eye sample data set. The electrolyte topography image is processed and labeled with a class to obtain a class sample data set.

[0014] Each sample in the fire eye sample data set is preprocessed to obtain a first feature map.

[0015] A YOLO-V5 network model is constructed, which includes a Backbone part, a Neck part and a prediction layer part connected in sequence. The Backbone part includes a Focus layer, a MobileNet-V2 module and a first CSP module connected in sequence.

[0016] The YOLO-V5 network model is trained using the fire eye sample data set composed of the first feature map to obtain a trained YOLO-V5 network model.

[0017] Each sample in the class sample data set is preprocessed to obtain a second feature map.

[0018] The MobileNet-V2 module of the trained YOLO-V5 network model is trained using the class sample data set composed of the second feature map, and the parameters of the Focus layer, the first CSP module, the Neck part and the prediction layer part are unchanged to obtain a final YOLO-V5 network model.

[0019] Further, the specific implementation process of training the YOLO-V5 network model by using the fire eye sample data set composed of the first feature map is as follows:

[0020] The first feature map is extracted by using the Focus layer to obtain a 304*304 size feature map;

[0021] The 304*304 size feature map output by the Focus layer is extracted by using the MobileNet-V2 module to obtain 152*152, 76*76 and 38*38 size feature maps;

[0022] The 152*152, 76*76 and 38*38 size feature maps output by the MobileNet-V2 module are segmented by using the first CSP module, and are merged by cross-stage hierarchical operation to obtain merged 152*152, 76*76 and 38*38 size feature maps;

[0023] The 152*152, 76*76 and 38*38 size feature maps output by the first CSP module are up-sampled and feature-fused by using the Neck part to obtain 76*76, 38*38 and 19*19 size feature maps;

[0024] The 76*76, 38*38 and 19*19 size feature maps output by the Neck part are optimized to remove large detection heads and adaptively adjust anchor boxes by using the prediction layer part to obtain the target detection frame of the fire eye opening.

[0025] Further, the Neck part includes a second CSP module, a CBL module and an SPP module;

[0026] The specific implementation process of up-sampling and feature-fusing the 152*152, 76*76 and 38*38 size feature maps output by the first CSP module by using the Neck part is as follows:

[0027] The 38*38 size feature map output by the first CSP module is segmented and fused by using the second CSP module, and then is extracted by the CBL module to obtain a 38*38 size feature map;

[0028] The 38*38 size feature map output by the CBL module is down-sampled by using the SPP module to obtain a 19*19 size feature map;

[0029] The 19*19 size feature map output by the SPP module is upsampled to obtain a 38*38 size feature map, and the 38*38 size feature map obtained by upsampling is subjected to segmentation fusion processing by the second CSP module, and then subjected to feature extraction by the CBL module to obtain a 38*38 size feature map;

[0030] The 19*19 size feature map output by the SPP module is upsampled to obtain a 76*76 size feature map, and the 76*76 size feature map obtained by upsampling is subjected to segmentation fusion processing by the second CSP module, and then subjected to feature extraction by the CBL module to obtain a 76*76 size feature map.

[0031] Further, the prediction layer part includes a convolution layer and three detection heads with sizes of 76x76x255, 38x38x255 and 19x19x255 respectively;

[0032] In the prediction layer part, the 76x76x255 detection head for large targets is first optimized and removed, then the original anchor frame is adaptively adjusted to [10, 14, 23, 27, 37, 58], [81, 82, 135, 169, 344, 319], and finally the target detection frame of the fire eye is output based on the loss function and back propagation.

[0033] Further, the MobileNet-V2 module includes an avgpool module, seven bottleneck modules, a batch normalization layer and an activation function layer, wherein the bottleneck module includes a channel-wise convolution layer and a point-wise convolution layer.

[0034] Further, the expression of the normalization loss function of the batch normalization layer is:

[0035]

[0036] In the formula, is the normalization loss function, x (k) is the loss function value after linear transformation by the batch normalization layer, E[] represents the mean value of the loss function value, and Var is the mean square difference operator.

[0037] Further, the activation function layer adopts a LeakyReLU activation function, and the specific expression is:

[0038]

[0039] In the formula, f(i) is the LeakyReLU activation function, and i represents the feature map input value.

[0040] Based on the same inventive concept, the present application further provides an electrolyte overheating degree identification method, the identification method comprising:

[0041] obtaining an electrolyte topographic image;

[0042] preprocessing the electrolyte topographic image to obtain a feature map;

[0043] training an electrolyte overheating degree identification model according to the electrolyte overheating degree identification model training method described above;

[0044] identifying the fire eye position and the overheating degree of each feature map using the electrolyte overheating degree identification model.

[0045] Based on the same inventive concept, the present application further provides an electronic device, which comprises:

[0046] at least one processor; and,

[0047] a memory in communication connection with the at least one processor; wherein,

[0048] the memory stores instructions executable by the at least one processor, and the instructions, when executed by the at least one processor, cause the at least one processor to perform the steps in the electrolyte overheating degree identification model training method described above, or perform the steps in the electrolyte overheating degree identification method described above.

[0049] Based on the same inventive concept, the present application further provides 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 overheating degree identification model training method described above, or perform the steps in the electrolyte overheating degree identification method described above.

[0050] Advantages

[0051] Compared with the prior art, the present application has the following advantages:

[0052] The overheat recognition model training method, recognition method, device and storage medium provided by the application replace the DarkNet with the MobileNet-V2 lightweight network structure on the basis of YOLO-V5; in the training process, the improved YOLO-V5 network model is trained for fire eye opening position recognition by using a fire eye sample data set, and the MobileNet-V2 in the improved YOLO-V5 network model is separately trained for overheat state recognition by using a category sample data set, so that the accuracy of model overheat state recognition is greatly improved, and the final YOLO-V5 network model can not only automatically locate the fire eye opening position, but also accurately recognize the overheat state.

[0053] The application replaces the DarkNet with the MobileNet-V2 lightweight network structure, improves the inference speed while ensuring the recognition accuracy, automatically removes a large amount of invalid background information from the image, and only retains the effective electrolyte melt morphology image, so that the aluminum electrolysis overheat state is quickly and accurately recognized. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only one embodiment of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Figure 1 It is an electrolyte overheat recognition model training method flowchart in the embodiment of the application;

[0056] Figure 2 It is a backbone network structure diagram of the YOLO-V5 network model in the embodiment of the application;

[0057] Figure 3 It is a structure diagram of the Neck part and the prediction layer part of the YOLO-V5 network model in the embodiment of the application;

[0058] Figure 4 It is a MobileNet-V2 module structure schematic diagram in the embodiment of the application;

[0059] Figure 5 It is a general 3*3*4 convolution layer operation schematic diagram in the embodiment of the application;

[0060] Figure 6 It is a per-channel convolution operation schematic diagram in the embodiment of the application;

[0061] Figure 7 It is a per-point convolution operation schematic diagram in the embodiment of the application;

[0062] Figure 8 is the MobileNet-V2 model overheat recognition accuracy curve in the embodiment of the present application;

[0063] Figure 9 is a GIoU curve schematic diagram of the YOLO-V5 network model in the embodiment of the present application;

[0064] Figure 10 is the average precision parameter curve of the YOLO-V5 network model in the embodiment of the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the present application will be described clearly and completely in the embodiments of the present application in combination with the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0066] The technical solutions of the present application will be described in detail in the following specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0067] As shown in the embodiment, the electrolyte overheat recognition model training method provided by the embodiment comprises the following steps: Figure 1

[0068] Step 1: Making fire eye sample data set and category sample data set

[0069] Step 1.1: Making fire eye sample data set: using an industrial camera to shoot to obtain electrolyte topography images, and marking the position of the fire eye opening on the electrolyte topography images to obtain the position labels of the left upper coordinates and the right lower coordinates of the fire eye opening area, and then obtaining the fire eye sample data set.

[0070] In the embodiment, 24000 electrolyte topography images are obtained, and the fire eye sample data set obtained contains 24000 fire eye samples. According to the ratio of 8:2, the fire eye sample data set is divided into a training set and a test set, i.e. 19200 fire eye samples constitute the training set, and 4800 fire eye samples constitute the test set, and the size of each fire eye sample is 1920x1080.

[0071] Step 1.2: Making category sample data set: using an industrial camera to shoot to obtain electrolyte topography images, and performing background processing and category labeling on the electrolyte topography images to obtain the category sample data set.

[0072] ​The background seriously affects the extraction of the electrolyte morphology feature by the YOLO-V5 network model. In the process of making the category sample dataset, by performing the cutting and segmentation processing on the image, only the effective electrolyte melt morphology image is reserved, so that the overheating state recognition can be accelerated. In the embodiment, the category labeling includes overhigh, high, normal, low and overlow. According to the 8:2 ratio, the training set and the test set of the category sample dataset are divided, that is, 19200 category samples constitute the training set, and 4800 category samples constitute the test set. The size of each category sample is 1920*1080.

[0073] In the embodiment, the fire eye sample dataset is used for fire eye opening position recognition training of the YOLO-V5 network model, and the category sample dataset is only used for single overheating state recognition training of the MobileNet-V2 module in the YOLO-V5 network model. The other part parameters of the YOLO-V5 network model remain unchanged.

[0074] Step 2: Preprocessing is performed on each sample in the fire eye sample dataset to obtain a first feature map.

[0075] The size of each fire eye sample is 1920*1080. Preprocessing is performed on each fire eye sample to obtain a first feature map with a size of 608*608.

[0076] Step 3: Construction of the YOLO-V5 network model

[0077] The YOLO-V5 network model is constructed, as shown in Figures 1-3 The YOLO-V5 network model includes a Backbone part, a Neck part and a prediction layer part connected in sequence. The Backbone part includes a Focus layer, a MobileNet-V2 module and a first CSP module (i.e. a cross-stage local network) connected in sequence. The YOLO-V5 network model of the application is improved on the basis of the traditional YOLO-V5 network. In the backbone network part, the lightweight MobileNet-V2 is used to replace the DarkNet, that is, the depth separable convolution is used to replace the ordinary convolution operation, so that the inference speed can be improved while the accuracy is guaranteed.

[0078] Step 4: Training the YOLO-V5 network model constructed in step 2 by using the fire eye sample dataset.

[0079] In the embodiment, the specific training process of the YOLO-V5 network model is as follows:

[0080] Step 4.1: The first feature map is extracted by using the Focus layer to obtain a feature map with a size of 304*304;

[0081] Step 4.2: Use the MobileNet-V2 module to extract features from the 304*304 feature map output by the Focus layer, and obtain feature maps of sizes 152*152, 76*76, and 38*38.

[0082] Step 4.3: Use the first CSP module to segment the 152*152, 76*76, and 38*38 feature maps output by the MobileNet-V2 module, and merge them through cross-stage layering operations to obtain merged 152*152, 76*76, and 38*38 feature maps.

[0083] Step 4.4: Use the Neck part to upsample and fuse 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.

[0084] Step 4.5: Optimize the 76*76, 38*38, and 19*19 feature maps output by the Neck part using the prediction layer to remove large detection heads and adaptively adjust the anchor boxes to obtain the target detection boxes for the fire-eye opening.

[0085] The MobileNet-V2 module consists of an avgpool module, seven bottleneck modules, batch normalization layers, and activation function layers, as follows: Figure 4 As shown, the bottleneck module includes a 3x3 channel-wise convolutional layer (e.g., Figure 6 (as shown) and two 1*1 pointwise convolutional layers (as shown) Figure 7 (As shown). Replacing ordinary convolution operations with depthwise separable convolutions can reduce the number of convolution parameters by 8 / 9 while maintaining the same performance. In the convolution operation process of the MobileNet-V2 module, a 1*1 pointwise convolutional layer is first used for dimensionality increase. The dimensionality-incremented feature map is then fed into a 3*3 channelwise convolutional layer. The feature map obtained from the 3*3 channelwise convolutional layer is then fed into a 1*1 pointwise convolutional layer for dimensionality reduction. Ultimately, this not only achieves better model capacity but also further reduces the computational cost of convolution operations and improves inference speed.

[0086] The expression for the normalized loss function of the batch normalization layer is:

[0087]

[0088] In the formula, For the normalized loss function, x (k) E[] represents the loss function value after the linear transformation of this layer, and Var is the mean of the loss function value.

[0089] The activation function layer adopts a LeakyReLU activation function, and the specific expression is as follows:

[0090]

[0091] In the formula, f(i) is a LeakyReLU activation function, and i represents a feature map input value.

[0092] The Neck part obtains tensor feature maps of different scales by upsampling, feature fusion, etc. on the tensor feature maps of different scales output by the Backbone part. As shown in FIGS. 1 and 2, the Neck part includes a second CSP module (cross-stage partial network), a CBL module (concatenated convolutional layer module), and an SPP module (spatial pyramid pooling module). The specific implementation process of upsampling and feature fusion of the 152*152, 76*76, and 38*38 size feature maps output by the first CSP module by using the Neck part is as follows: Figure 2 3 As shown in FIGS. 1 and 2, the Neck part includes a second CSP module (cross-stage partial network), a CBL module (concatenated convolutional layer module), and an SPP module (spatial pyramid pooling module). The specific implementation process of upsampling and feature fusion of the 152*152, 76*76, and 38*38 size feature maps output by the first CSP module by using the Neck part is as follows:

[0093] Step 4.41: The 38*38 size feature map output by the first CSP module is subjected to segmentation fusion processing by using the second CSP module, and then feature extraction is performed by the CBL module to obtain a 38*38 size feature map.

[0094] Step 4.42: The 38*38 size feature map output by the CBL module in step 4.21 is subjected to down-sampling by using the SPP module to obtain a 19*19 size feature map.

[0095] Step 4.43: The 19*19 size feature map output by the SPP module in step 4.22 is subjected to up-sampling to obtain a 38*38 size feature map, and then the 38*38 size feature map obtained by up-sampling is subjected to segmentation fusion processing by using the second CSP module, and then feature extraction is performed by the CBL module to obtain a 38*38 size feature map.

[0096] Step 4.44: The 19*19 size feature map output by the SPP module in step 4.22 is subjected to up-sampling to obtain a 76*76 size feature map, and then the 76*76 size feature map obtained by up-sampling is subjected to segmentation fusion processing by using the second CSP module, and then feature extraction is performed by the CBL module to obtain a 76*76 size feature map.

[0097] ​The cross-stage local network structure of the deep network with enhanced learning ability (CSPnet) is adopted in the Neck part to enhance the network feature fusion capability and reduce the network model size. Specifically, the feature map output by the backbone network is further extracted through the cross-stage local network and the convolution layer module, and then the SPP module processes the feature map in three scales, which can better utilize the image feature matrix output by the backbone network and obtain more scale electrolyte topography image feature maps. Finally, the tensor feature maps of different scales obtained by processing are transmitted to the prediction layer part.

[0098] As shown in Figure 3 The prediction layer part includes a convolution layer and three detection heads with sizes of 76*76*255, 38*38*255 and 19*19*255, respectively. To solve the problem of target stray distribution and too small target pixel ratio caused by the industrial camera view angle, in the prediction layer part, the 76*76*255 detection head for large targets is optimized and removed, and the original anchor box is adaptively adjusted to [10, 14, 23, 27, 37, 58], [81, 82, 135, 169, 344, 319], and finally the target detection box of the fire eye is output based on the loss function and back propagation.

[0099] Step 5: Preprocess each sample in the category sample data set to obtain a second feature map.

[0100] Each category sample has a size of 1920*1080, and each category sample is preprocessed to obtain a second feature map with a size of 608*608.

[0101] Step 6: Use the category sample data set to separately train the MobileNet-V2 module in the YOLO-V5 network model trained in step 4.

[0102] The MobileNet-V2 module in the trained YOLO-V5 network model is trained using the category sample data set composed of the second feature map, and the parameters (such as weight values) of the Focus layer, the first CSP module, the Neck part and the prediction layer part remain unchanged, i.e. the Focus layer, the first CSP module, the Neck part and the prediction layer part do not change with the training of the MobileNet-V2 module, and the final YOLO-V5 network model is obtained.

[0103] The output results of the YOLO-V5 network model in this embodiment are evaluated by using a generalized intersection over union (referred to as GIOU), a mean average precision (referred to as MAP) and an inference speed. The GIOU is a loss function of a regression target box. When the GIOU is used as an evaluation index, the GIOU has the following characteristics: non-negativity, symmetry and scale invariance. The smaller the GIOU value is, the higher the output precision of the target box is. The calculation formula of the GIOU is:

[0104]

[0105]

[0106] In the formula, GIoU represents the GIOU, IoU represents the intersection over union, A and B represent two target detection boxes of any two fire eyes, C represents a minimum square box that can cover 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 union area of the A box and the B box, and |A∩B| represents the intersection area of the A box and the B box.

[0107] The mean average precision is an index for measuring the detection precision of multiple-labeled images. In a multiple-labeled image detection task, the image is labeled more than once, and the mean average precision (MAP) similar to the method used in information retrieval is used to calculate the precision. The greater the mean average precision value is, the higher the target detection precision is. The mean average precision is calculated by drawing a PR curve, that is, a two-dimensional curve with precision and recall as the vertical and horizontal coordinates. Generally, the precision is the accuracy, and the recall is the recall rate.

[0108] The inference speed is defined as the number of images that can be detected per second. The faster the inference speed is, the better the real-time performance of the target detection network is.

[0109] Simulation experiment: The platform of the embodiment is a Windows 10 operating system, and the development environment is PyCharm Community Edition 2021.1.1x64. The experimental model is on a Pytorch 1.10.0 framework, and a MobileNet-V2 learning network is used. The model training is completed in an Nvdia 3060Ti (12G of video memory) GPU and a CUDA 11.2 experimental environment.

[0110] The training parameters of the YOLO-V5 network model are set as follows: the number of training rounds is 200 rounds, the bitch size is 16, that is, 16 images are put in one time, and the initial learning rate is 0.001.

[0111] Figure 8For the MobileNet-V2 overheating state recognition accuracy curve, the model can reach 99.6% in the recognition accuracy of the overheating state when iterating to about 160 rounds, which is greatly improved compared with the previous highest recognition accuracy of 95%. Figure 9 In the figure, the ordinate is the generalization intersection over union value, and the abscissa is the training round number. When iterating about 200 times, the generalization intersection over union value of the YOLO-V5 network model is smaller than that of the traditional YOLO-V5, and the target frame output precision is higher. Figure 10 For the average precision parameter curve of the YOLO V5 network model, the ordinate is the average precision value, and the abscissa is the training round number. When iterating about 200 times, the average precision value of the YOLO V5 network model is 45.4, and the average precision value of the traditional YOLO V5 is 37.2, and the target detection precision is higher. In terms of inference speed, the YOLO V5 network model of the present application can detect 158 pictures per second, and the traditional YOLO V5 network can detect 96 pictures per second. The detection speed of the YOLO V5 network model of the present application is improved by 22%, and the real-time performance is better. Table 1 is the experimental comparison data of the parameter amount, inference speed and average precision of the traditional YOLO-V5 and the YOLO V5 network model of the present application. The parameter amount of the YOLO-V5 is greatly reduced from 7.2M to 1.7M by changing the backbone network of the YOLO-V5 to the lightweight MobileNet-V2 network.

[0112] Table 1 Experimental comparison data

[0113] Model Parameter amount (M) Inference speed (ms) Average accuracy (MAP) Traditional YOLO-V5 7.2 0.9 37.2 YOLO-v5 of the present application 1.7 0.63 45.4

[0114] Based on the same inventive concept, the embodiments of the present application also provide an electrolyte overheating recognition method, the recognition method comprising:

[0115] Step 1: Obtain an electrolyte topographic image;

[0116] Step 2: Preprocess the electrolyte topographic image to obtain a feature map;

[0117] Step 3: Train an electrolyte overheating recognition model according to the electrolyte overheating recognition model training method described above;

[0118] Step 4: Use the electrolyte overheating recognition model to recognize the fire eye position and overheating degree of each feature map.

[0119] The overheat recognition model training method, recognition method, device and storage medium provided by the application replace the DarkNet with the MobileNet-V2 lightweight network structure on the basis of YOLO-V5; in the training process, the improved YOLO-V5 network model is first trained for fire eye opening position recognition by using a fire eye sample data set, and then the MobileNet-V2 in the improved YOLO-V5 network model is separately trained for overheat state recognition by using a category sample data set, so that the accuracy of model overheat state recognition is greatly improved, and the final YOLO-V5 network model can not only automatically locate the fire eye opening position, but also accurately recognize the overheat state. The application replaces the DarkNet with the MobileNet-V2 lightweight network structure, improves the inference speed while ensuring the recognition accuracy, automatically removes a large amount of invalid background information from the image, and only retains the effective electrolyte melt morphology image, so that the aluminum electrolysis overheat state is quickly and accurately recognized.

[0120] The above only discloses specific embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or modifications within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for training an electrolyte superheat recognition model, characterized in that, Includes the following steps: Electrolyte morphology images are acquired, and the locations of the fire-eye openings are marked on the electrolyte morphology images to obtain a fire-eye sample dataset; background removal and category labeling are performed on the electrolyte morphology images to obtain a category sample dataset; Each sample in the Huoyan sample dataset is preprocessed to obtain a first feature map; Construct a YOLO-V5 network model, which includes a Backbone part, a Neck part, and a prediction layer part connected in sequence. The Backbone part includes a Focus layer, a MobileNet-V2 module, and a first CSP module connected in sequence. The YOLO-V5 network model is trained using the FireEye sample dataset composed of the first feature maps to obtain the trained YOLO-V5 network model. Each sample in the aforementioned category sample dataset is preprocessed to obtain a second feature map; The MobileNet-V2 module of the trained YOLO-V5 network model is trained using a class sample dataset composed of the second feature map. The parameters of the Focus layer, the first CSP module, the Neck part, and the prediction layer part remain unchanged to obtain the final YOLO-V5 network model.

2. The electrolyte superheat identification model training method according to claim 1, characterized in that, The specific implementation process of training the YOLO-V5 network model using the FireEye sample dataset composed of the first feature maps is as follows: The Focus layer is used to extract features from the first feature map to obtain a 304*304 feature map. The MobileNet-V2 module is used to extract features from the 304*304 feature map output by the Focus layer, resulting in feature maps of sizes 152*152, 76*76, and 38*38. The first CSP module is used to segment the 152*152, 76*76, and 38*38 feature maps output by the MobileNet-V2 module, and a merging operation is performed through a cross-stage layering operation to obtain merged 152*152, 76*76, and 38*38 feature maps. The Neck section is used to upsample and fuse 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. The prediction layer is used to optimize the 76*76, 38*38, and 19*19 feature maps output by the Neck part to remove large detection heads and adaptively adjust the anchor boxes to obtain the target detection boxes for the fire eye opening.

3. The electrolyte superheat identification model training method according to claim 2, characterized in that, The Neck section includes a second CSP module, a CBL module, and an SPP module; The specific implementation process of upsampling and feature fusion of the 152*152, 76*76, and 38*38 feature maps output by the first CSP module using the Neck part is as follows: The second CSP module is used to segment and fuse the 38*38 feature map output by the first CSP module, and then the CBL module is used to extract features to obtain a 38*38 feature map. The SPP module is used to downsample the 38*38 feature map output by the CBL module to obtain a 19*19 feature map; The 19*19 feature map output by the SPP module is upsampled to obtain a 38*38 feature map. The second CSP module then performs segmentation and fusion processing on the upsampled 38*38 feature map, and the CBL module performs feature extraction to obtain a 38*38 feature map. The 19*19 feature map output by the SPP module is upsampled to obtain a 76*76 feature map. The second CSP module then performs segmentation and fusion processing on the upsampled 76*76 feature map, and the CBL module performs feature extraction to obtain a 76*76 feature map.

4. The electrolyte superheat identification model training method according to claim 2, characterized in that, The prediction layer includes a convolutional layer and three detection heads with sizes of 76×76×255, 38×38×255, and 19×19×255, respectively. In the prediction layer, the 76×76×255 detection head targeting large targets is first optimized and removed. Then, the original anchor boxes are adaptively adjusted to [10,14,23,27,37,58] and [81,82,135,169,344,319]. Finally, the target detection box of the fire eye is output based on the loss function and backpropagation.

5. The method for training an electrolyte superheat recognition model according to any one of claims 1 to 4, characterized in that, The MobileNet-V2 module consists of an avgpool module, seven bottleneck modules, a batch normalization layer, and an activation function layer. The bottleneck module includes channel-wise convolutional layers and point-wise convolutional layers.

6. The electrolyte superheat identification model training method according to claim 5, characterized in that, The expression for the normalization loss function of the batch normalization layer is: In the formula, For the normalized loss function, x (k) E[] represents the loss function value after linear transformation by the batch normalization layer, and Var is the mean of the loss function value.

7. The electrolyte superheat identification model training method according to claim 5, characterized in that, The activation function layer uses the LeakyReLU activation function, the specific expression of which is: In the formula, f(i) is the LeakyReLU activation function, and i represents the input value of the feature map.

8. A method for identifying electrolyte superheat, characterized in that, The identification method includes: Acquire electrolyte morphology images; The electrolyte morphology image is preprocessed to obtain a feature map; The electrolyte superheat identification model is trained according to any one of claims 1 to 7; The electrolyte superheat identification model is used to identify the location of the fire hole and the degree of superheat in each feature map.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor. When the instructions are executed by the at least one processor, the at least one processor performs the steps in the electrolyte overheating identification model training method according to any one of claims 1 to 7, or performs the steps in the electrolyte overheating identification method according to claim 8.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions, which, when executed by at least one processor, cause the at least one processor to perform the steps in the electrolyte overheating identification model training method of any one of claims 1 to 7, or to perform the steps in the electrolyte overheating identification method of claim 8.

Images