Steel cold rolling annealing furnace component detection method based on semi-supervised learning

Abstract

The present invention discloses a method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning, and relates to the field of industrial production technology. A high-precision FasterR‑CNN detection model is adopted, and the backbone network adopts the ResNet50 network, and a feature pyramid is added to capture feature information of different scales, thereby improving the accuracy of detection. In view of the problem of excessive manpower and material costs for labeling steelmaking equipment and the problem of biased prediction caused by imbalanced categories of components during semi-supervised learning, the present invention adopts a semi-supervised learning target detection method-an unbiased teacher method, firstly training the teacher model separately with labeled data, and then the teacher generates pseudo labels to train the student model, and the student model gradually updates the teacher model through the EMA algorithm. Through the unbiased teacher target detection method of semi-supervised learning, accurate detection of components of a steel cold rolling annealing furnace is achieved with only a small amount of labeled data training.

Description

Technical Field

[0001] The invention relates to the technical field of industrial production, and in particular to a component detection method for a steel cold rolling annealing furnace based on semi-supervised learning. Background Art

[0002] The steel industry is an important material foundation. Intelligent monitoring and management of steelmaking equipment components in the steel industry production process is particularly important. The existing monitoring of steelmaking equipment operation is generally carried out manually, which has the following problems: heavy workload, dangerous and harsh working environment, low efficiency, and delayed fault judgment. Therefore, manual detection methods are increasingly incompatible with the requirements of enterprises for energy conservation and consumption reduction and improving production efficiency. The development of an automatic monitoring system based on artificial intelligence technology has become an important way to solve the problem of manual detection.

[0003] Deep neural network models have performed well in target detection tasks and achieved good results. However, training these networks requires a large number of people to annotate the data. The cost of manual annotation of steel cold rolling annealing furnace component data is too high, and a large amount of collected image data is not annotated. At present, the methods of using semi-supervised learning to detect steel components are mainly considered from two directions: consistency regularization and pseudo-annotation. Consistency regularization is to regularize the model to be invariant and robust to the input data enhancement, and keep the output of the original input data and the enhanced input data consistent for learning. Pseudo-annotation is to let the model generate pseudo-annotations on unlabeled data, so as to train the randomly enhanced unlabeled data to match the corresponding pseudo-annotations for learning. However, due to some unique characteristics of steel components, such as imbalanced component categories and imbalanced foreground and background, the above methods will be too biased towards the categories with the dominant number of detections when detecting, and ignore the secondary categories, which makes the trained model prone to biased predictions and introduces serious overfitting. In view of this, we propose a steel cold rolling annealing furnace component detection method based on semi-supervised learning. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning, which solves the problems mentioned in the above background technology.

[0005] To achieve the above objectives, the present invention is implemented by the following technical scheme: a method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning, the detection method comprising the following steps:

[0006] S1: Data preprocessing: use annotation tools to manually annotate a small amount of training data, build an annealing furnace component data set, and perform data expansion and enhancement;

[0007] S2: Extract feature maps containing different scales of the image, use the first five convolutional networks of the ResNet50 network for feature extraction, and then build a feature pyramid network between several convolutional networks to capture features of different scales:

[0008] S3: Input the feature map in the feature pyramid network into the region generation network to generate candidate region frames based on feature maps of different scales, then splice the candidate regions corresponding to feature maps of different scales and pass them into the subsequent network;

[0009] S4: Build a feature pyramid network, perform Roi Pooling of different scales on the candidate region boxes, then concatenate the feature maps and input them into the fully connected layer to predict the target category and regress the position of the target bounding box;

[0010] S5: Teacher model training stage: A semi-supervised learning method is used to train the Faster R-CNN target detection model. First, a small number of labeled samples are used to train the model, which is called the teacher model.

[0011] S6: In the mutual learning stage between teacher and student models, the parameters of the teacher model are fixed and the student model is trained. For unlabeled data, on the one hand, it is input into the teacher model after weak enhancement to generate pseudo labels, and on the other hand, it is input into the student model after strong enhancement to output labels, and then the loss function is calculated with the pseudo labels of the teacher model, and then the student model is updated through back propagation. The student model trained with unlabeled data is optimized with the teacher model through the EMA algorithm to obtain the final network model.

[0012] S7: The model is deployed to automatically detect the components of the steel cold rolling annealing furnace, and the category and location coordinates can be obtained.

[0013] Optionally, the S1 further includes:

[0014] S11: Use a robot to capture images from different angles under different lighting conditions, collect and organize image data, a total of 1,000 images, unify the image size, and independently build a dataset of steel cold rolling annealing furnace components;

[0015] S12: Use annotation tools to manually annotate fixed-scale image data and annotate three types of steel components: motors, bearings, and reducers;

[0016] S13: Data enhancement, which performs rotation, Gaussian blur, color transformation and other processing on image data to expand the data set and increase diversity;

[0017] S14: Divide the dataset into training set, test set, and validation set, and read the data in the COCO dataset format.

[0018] Optionally, the S2 further includes:

[0019] S21: The data-enhanced image is input into the Conv1 layer with a convolution kernel size of 7×7 and a step size of 2 for convolution operation, then connected to the ReLU activation function, and then passed through a maximum pooling layer with a convolution kernel size of 3×3 and a step size of 2 for pooling operation;

[0020] S22: The convolutional layers Conv2, Conv3, Conv4, and Conv5 are all composed of blocks, each of which consists of three layers of convolution. The sizes of the convolution kernels are 1×1, 3×3, and 1×1, respectively. The number of convolution kernels in Conv2 is 64, 64, and 256, respectively. The number of convolution kernels in the subsequent convolutional networks doubles.

[0021] S23: For the features of steel components of different scales, a feature pyramid network is used to extract features. The feature map obtained by the Conv5 layer is upsampled, and the feature map of the Conv4 layer is convolved by 1×1, so that the two feature maps have the same resolution. The processed Conv4 low-level features and the processed Conv5 high-level features are then accumulated element by element to construct a feature pyramid of one layer, called p4.

[0022] S24: Finally, we obtain feature pyramids p2, p3, p4 and p5 features with richer information and different scales, which are then input into the region generation network to detect steel components with different scale features.

[0023] Optionally, the S3 further includes:

[0024] S31: The feature maps of the four scales in the feature pyramid network obtained in step S2 are respectively passed into the region generation network to generate several anchor frames corresponding to the feature maps of the four scales, and then divided into positive sample anchor frames and negative sample anchor frames through a softmax classifier, and the positive sample anchor frames are used as candidate region frames;

[0025] S32: Use the bounding box regression method to correct the position of the positive sample anchor box, and then use the non-maximum suppression method to reduce the overlap rate of the candidate boxes to obtain the candidate area boxes of the motor, bearing and reducer;

[0026] S4: Build a feature pyramid network, perform Roi Pooling of different scales on the candidate region boxes, then splice the feature maps and input them into the fully connected layer to predict the target category and regress the position of the target bounding box.

[0027] Optionally, the S4 further includes:

[0028] S41: splicing the candidate region frames of different scales obtained in step S3, and then connecting them to the subsequent feature pyramid pooling layer to perform Roi Pooling, thereby obtaining pooling features with richer feature information;

[0029] S42: The pooled features are concatenated and then connected to a fully connected layer. The category information of the target box is predicted through the softmax classifier, and the position information of the target box is predicted. Then, the non-maximum suppression method is used to reduce the overlap rate of the candidate boxes.

[0030] Optionally, the S5 further includes:

[0031] S51: The teacher model uses a small amount of labeled data, with a ratio of 25%, that is, 200 images with labeled information for supervised learning. The detection model uses the Faster R-CNN model with the backbone network ResNet50_FPN constructed in steps S2-S4;

[0032] S52: Use a small number of labeled samples to supervise the training of the Faster R-CNN teacher model, loss function as follows:

[0033]

[0034] S53: Iterate the training until the model converges. After the teacher model training is completed, the image is passed through the teacher model to generate pseudo labels.

[0035] Optionally, the S6 further includes:

[0036] S61: The student model is the same as the teacher model, so the Faster R-CNN model with ResNet50_FPN as the backbone network is used;

[0037] S62: Fix the parameters of the teacher model, and use the teacher model to generate pseudo labels for the unlabeled data to train the student model;

[0038] S63: The student model trained with unlabeled data is fine-tuned by the EMA algorithm to the teacher model, and the teacher model is continuously optimized to obtain the final network model. The parameter update formula of the teacher model is as follows:

[0039]

[0040] S64: Iterate the training until the teacher model converges and save the training weight parameters. After the training is completed, the teacher model can be used to detect the components of the steel cold rolling annealing furnace.

[0041] Optionally, the S7 further includes:

[0042] S71: In actual scenarios, use a high-definition camera to obtain the original image, and perform pre-processing such as denoising and image enhancement on the image;

[0043] S72: Input the image into the network model, use the trained Faster R-CNN model to extract image features, remove redundant target area frames through the non-maximum suppression method, and finally detect the target area frames corresponding to the three categories of steel components: motors, bearings, and reducers, and complete the automatic detection of steel components.

[0044] The present invention provides a method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning. It has the following beneficial effects:

[0045] The steel cold rolling annealing furnace component detection method based on semi-supervised learning adopts an independently constructed steel cold rolling annealing furnace component data set, fully explores the visual features of steel components, and enhances the diversity of data samples through image enhancement, thereby ensuring that the detection model is robust. The present invention adopts a high-precision Faster R-CNN detection model, the backbone network adopts a ResNet50 network, and adds a feature pyramid to capture feature information of different scales, thereby improving the accuracy of detection. In view of the problem of excessive human and material costs for steelmaking equipment labeling and the problem of biased prediction caused by the imbalance of component categories during semi-supervised learning, the present invention adopts a semi-supervised learning target detection method-unbiased teacher method, first trains the teacher model separately with labeled data, and then the teacher generates pseudo labels to train the student model, and the student model gradually updates the teacher model through the EMA algorithm. Through the semi-supervised learning unbiased teacher target detection method, accurate detection of steel cold rolling annealing furnace components is achieved with only a small amount of labeled data training. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is the backbone network ResNet50_FPN framework diagram of the model of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1 The present invention provides a technical solution: a method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning, the detection method comprising the following steps:

[0049] S1: Data preprocessing: use annotation tools to manually annotate a small amount of training data, build an annealing furnace component data set, and perform data expansion and enhancement.

[0050] Specifically, the S1 data preprocessing process also includes:

[0051] S11: Use a robot to capture images from different angles under different lighting conditions, collect and organize image data, a total of 1,000 images, unify the image size, and independently build a dataset of steel cold rolling annealing furnace components;

[0052] S12: Use annotation tools to manually annotate fixed-scale image data, annotate three types of steel components: motors, bearings, and reducers. Each image corresponds to a json-formatted annotation file, and then convert all annotation files into a total annotation file according to the COCO dataset format;

[0053] S13: Data enhancement: rotating, Gaussian blurring, color transformation, etc. the image data to expand the data set and increase diversity, thereby improving the generalization ability of the model for steel component detection;

[0054] S14: Divide the dataset into a training set (800 images), a test set (100 images), and a validation set (100 images), and read the data in the COCO dataset format.

[0055] S2: Extract feature maps containing different scales of the image, use the first five convolutional networks of the ResNet50 network for feature extraction, and then build a feature pyramid network between several convolutional networks to capture features of different scales.

[0056] Specifically, the process of extracting image information by the convolutional layer in step S2 includes the following steps:

[0057] S21: The data-enhanced image is input into the Conv1 layer with a convolution kernel size of 7×7 and a step size of 2 for convolution operation, then connected to the ReLU activation function, and then passed through a maximum pooling layer with a convolution kernel size of 3×3 and a step size of 2 for pooling operation;

[0058] S22: The convolutional layers Conv2, Conv3, Conv4, and Conv5 are all composed of blocks, each of which consists of three layers of convolution. The convolution kernel sizes are 1×1, 3×3, and 1×1, respectively. The number of convolution kernels in Conv2 is 64, 64, and 256, respectively. The number of convolution kernels in the following convolutional networks doubles. Conv2 is implemented by 3 blocks, Conv3 is implemented by 4 blocks, Conv4 is implemented by 23 blocks, and Conv5 is implemented by 3 blocks.

[0059] S23: A feature pyramid network is used to extract features for steel components of different scales. The feature map obtained from the Conv5 layer is upsampled, and the feature map from the Conv4 layer is convolved 1×1 so that the two feature maps have the same resolution. The processed Conv4 low-level features and the processed Conv5 high-level features are then accumulated element by element to construct a layer of feature pyramid, called p4. By analogy, the Conv4 layer and the Conv3 layer network generate p3 according to the above steps, and the Conv3 layer and the Conv2 layer network generate p2. The feature map obtained from the Conv5 layer alone is called p5;

[0060] S24: Finally, we obtain feature pyramids p2, p3, p4 and p5 features with richer information and different scales, which are then input into the region generation network to detect steel components with different scale features.

[0061] S3: Input the feature map in the feature pyramid network into the region generation network to generate candidate region boxes based on feature maps of different scales, and then splice the candidate regions corresponding to feature maps of different scales and pass them into the subsequent network.

[0062] Specifically, S3 includes the following steps:

[0063] S31: The feature maps of the four scales in the feature pyramid network obtained in step S2 are respectively passed into the region generation network to generate several anchor frames corresponding to the feature maps of the four scales, and then divided into positive sample anchor frames (overlap rate greater than 0.7) and negative sample anchor frames (overlap rate less than 0.3) through a softmax classifier, and the positive sample anchor frames are used as candidate region frames;

[0064] S32: Use the bounding box regression method to correct the position of the positive sample anchor box, and then use the non-maximum suppression method to reduce the overlap rate of the candidate boxes to obtain the candidate area boxes of the motor, bearing and reducer;

[0065] S4: Build a feature pyramid network, perform Roi Pooling of different scales on the candidate region boxes, then splice the feature maps and input them into the fully connected layer to predict the target category and regress the position of the target bounding box.

[0066] Specifically, the process of step S4 Roi Pooling operation and network classification prediction also includes:

[0067] S41: splicing the candidate region frames of different scales obtained in step S3, and then connecting them to the subsequent feature pyramid pooling layer to perform Roi Pooling, thereby obtaining pooling features with richer feature information;

[0068] S42: The pooled features are concatenated and then connected to a fully connected layer. The category information of the target box is predicted through the softmax classifier, and the location information of the target box is predicted. The non-maximum suppression method is then used to reduce the overlap rate of the candidate boxes. Finally, the category and location of the target box are obtained, and the automatic detection of the three steel components of motors, bearings and reducers is completed.

[0069] S5: Teacher model training stage: A semi-supervised learning method is used to train the Faster R-CNN target detection model. First, a small number of labeled samples are used to train the model, which is called the teacher model.

[0070] Specifically, the process of step S5 teacher model training stage also includes:

[0071] S51: The teacher model uses a small amount of labeled data, with a ratio of 25%, that is, 200 images with labeled information for supervised learning. The detection model uses the Faster R-CNN model with the backbone network ResNet50_FPN constructed in steps S2-S4;

[0072] S52: Use a small number of labeled samples to supervise the training of the Faster R-CNN teacher model, loss function as follows:

[0073]

[0074] The labeled dataset is in For input image data, is the data label, N s is the total number of labeled data; is the classification loss function of the region generation network, It is the regression loss function of the region generation network, which is used to calculate the category information of the region generation network and the location information of the candidate box; is the classification loss function for the region of interest; is the regression loss function of the region of interest, which is used to calculate the category information of the final model and the location information of the candidate box;

[0075] S53: Iterate the training until the model converges. After the teacher model training is completed, the image is passed through the teacher model to generate pseudo labels.

[0076] S6: Mutual learning stage between teacher and student models: fix the parameters of the teacher model and train the student model. For unlabeled data, on the one hand, weak enhancement is used to input the teacher model to generate pseudo labels, and on the other hand, strong enhancement is used to input the student model to output labels, and then the loss function is calculated with the pseudo labels of the teacher model, and then back-propagation is used to update the student model. The student model trained with unlabeled data is optimized with the teacher model through the EMA algorithm to obtain the final network model.

[0077] Specifically, the process of step S6, the mutual learning stage of the teacher and student models, also includes:

[0078] S61: The student model is the same as the teacher model, so the Faster R-CNN model with ResNet50_FPN as the backbone network is used. The student model uses unlabeled data to assist the teacher model in learning, thereby improving the accuracy of the teacher model;

[0079] S62: First, fix the parameters of the teacher model, and use the teacher model to generate pseudo labels for the unlabeled data to train the student model. Specifically, the unlabeled data is fed into the teacher model after weak enhancement to generate pseudo labels, and then fed into the student model after strong enhancement to output labels, and then the loss function is calculated with the pseudo labels of the teacher model, and then back-propagation is used to update the student model parameters;

[0080] S63: The student model trained with unlabeled data is fine-tuned by the EMA algorithm to the teacher model, and the teacher model is continuously optimized to obtain the final network model. The parameter update formula of the teacher model is as follows:

[0081]

[0082] in, are the model parameters trained in the teacher model training phase, are the parameters of the teacher model after the i-th iteration of the mutual learning phase between the teacher and student models, is the parameter of the student model after the kth iteration, γ is the learning rate, α is the coefficient of the exponential moving average method, and λ u is the balancing factor, balancing the supervised loss function and the unsupervised loss function values;

[0083] S64: Iterate the training until the teacher model converges and save the training weight parameters. After the training is completed, the teacher model can be used to detect the components of the steel cold rolling annealing furnace.

[0084] S7: The model is deployed to automatically detect the components of steel cold rolling annealing furnaces, and the categories and location coordinates can be obtained, achieving good detection results with only a small number of labeled samples.

[0085] Specifically, the process of detecting the annealing furnace components in the actual scenario in step S7 also includes:

[0086] S71: In actual scenarios, use a high-definition camera to obtain the original image, and perform pre-processing such as denoising and image enhancement on the image;

[0087] S72: Input the image into the network model, use the trained Faster R-CNN model to extract image features, remove redundant target area frames through the non-maximum suppression method, and finally detect the target area frames corresponding to the three categories of steel components: motors, bearings, and reducers, and complete the automatic detection of steel components.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning, characterized in that: The detection method comprises the following steps: S1: Data preprocessing: use annotation tools to manually annotate a small amount of training data, build an annealing furnace component data set, and perform data expansion and enhancement; S2: Extract feature maps containing different scales of the image, use the first five convolutional networks of the ResNet50 network for feature extraction, and then build a feature pyramid network between several convolutional networks to capture features of different scales: S3: Input the feature map in the feature pyramid network into the region generation network to generate candidate region frames based on feature maps of different scales, then splice the candidate regions corresponding to feature maps of different scales and pass them into the subsequent network; S4: Build a feature pyramid network, perform Roi Pooling of different scales on the candidate region boxes, then concatenate the feature maps and input them into the fully connected layer to predict the target category and regress the position of the target bounding box; S5: Teacher model training stage: A semi-supervised learning method is used to train the Faster R-CNN target detection model. First, a small number of labeled samples are used to train the model, which is called the teacher model. S6: In the mutual learning stage between teacher and student models, the parameters of the teacher model are fixed and the student model is trained. For unlabeled data, on the one hand, it is input into the teacher model after weak enhancement to generate pseudo labels, and on the other hand, it is input into the student model after strong enhancement to output labels, and then the loss function is calculated with the pseudo labels of the teacher model, and then the student model is updated through back propagation. The student model trained with unlabeled data is optimized with the teacher model through the EMA algorithm to obtain the final network model. S7: Deploy the model to automatically detect the components of the steel cold rolling annealing furnace and obtain the category and location coordinates.

2. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: Said S1 further comprises: S11: Use a robot to capture images from different angles under different lighting conditions, collect and organize image data, a total of 1,000 images, unify the images into uniform sizes, and independently build a dataset of steel cold rolling annealing furnace components; S12: Use annotation tools to manually annotate fixed-scale image data and annotate three types of steel components: motors, bearings, and reducers; S13: Data enhancement, which performs rotation, Gaussian blur, and color transformation on image data to expand the data set and increase diversity; S14: Divide the dataset into training set, test set, and validation set, and read the data in the COCO dataset format.

3. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: The S2 further comprises: S21: The data-enhanced image is input into the Conv1 layer with a convolution kernel size of 7×7 and a step size of 2 for convolution operation, then connected to the ReLU activation function, and then passed through a maximum pooling layer with a convolution kernel size of 3×3 and a step size of 2 for pooling operation; S22: The convolutional layers Conv2, Conv3, Conv4, and Conv5 are all composed of blocks, each of which consists of three layers of convolution. The sizes of the convolution kernels are 1×1, 3×3, and 1×1, respectively. The number of convolution kernels in Conv2 is 64, 64, and 256, respectively. The number of convolution kernels in the subsequent convolutional networks doubles. S23: For the features of steel components of different scales, a feature pyramid network is used to extract features. The feature map obtained by the Conv5 layer is upsampled, and the feature map of the Conv4 layer is convolved by 1×1, so that the two feature maps have the same resolution. The processed Conv4 low-level features and the processed Conv5 high-level features are then accumulated element by element to construct a feature pyramid of one layer, called p4. S24: Finally, we obtain feature pyramids p2, p3, p4 and p5 features with richer information and different scales, which are then input into the region generation network to detect steel components with different scale features.

4. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: The S3 further comprises: S31: The feature maps of the four scales in the feature pyramid network obtained in step S2 are respectively passed into the region generation network to generate several anchor frames corresponding to the feature maps of the four scales, and then divided into positive sample anchor frames and negative sample anchor frames through a softmax classifier, and the positive sample anchor frames are used as candidate region frames; S32: Use the bounding box regression method to correct the position of the positive sample anchor box, and then use the non-maximum suppression method to reduce the overlap rate of the candidate boxes to obtain the candidate area boxes of the motor, bearing and reducer; S4: Build a feature pyramid network, perform Roi Pooling of different scales on the candidate region boxes, then splice the feature maps and input them into the fully connected layer to predict the target category and regress the position of the target bounding box.

5. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: The S4 further comprises: S41: splicing the candidate region frames of different scales obtained in step S3, and then connecting them to the subsequent feature pyramid pooling layer to perform Roi Pooling, thereby obtaining pooling features with richer feature information; S42: The pooled features are concatenated and then connected to a fully connected layer. The category information of the target box is predicted through the softmax classifier, and the position information of the target box is predicted.

6. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: The S5 further comprises: S51: The teacher model uses a small amount of labeled data, with a ratio of 25%, that is, 200 images with labeled information for supervised learning. The detection model uses the Faster R-CNN model with the backbone network ResNet50_FPN constructed in steps S2-S4; S52: Use a small number of labeled samples to supervise the training of the Faster R-CNN teacher model; S53: Iterate the training until the model converges. After the teacher model training is completed, the image is passed through the teacher model to generate pseudo labels.

7. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: The S6 further comprises: S61: The student model is the same as the teacher model, so the Faster R-CNN model with ResNet50_FPN as the backbone network is used; S62: Fix the parameters of the teacher model, and use the teacher model to generate pseudo labels for the unlabeled data to train the student model; S63: The student model trained with unlabeled data is fine-tuned by the EMA algorithm to the teacher model, and the teacher model is continuously optimized to obtain the final network model; S64: Iterative training is performed until the teacher model converges, and the training weight parameters are saved. The teacher model after training is used to detect components of steel cold rolling annealing furnace.

8. The method for detecting components of a steel cold rolling annealing furnace based on semi-supervised learning according to claim 1, characterized in that: The S7 further comprises: S71: In actual scenarios, use a high-definition camera to obtain the original image, and perform image denoising and image enhancement preprocessing; S72: Input the image into the network model, use the trained Faster R-CNN model to extract image features, remove redundant target area frames through the non-maximum suppression method, and finally detect the target area frames corresponding to the three categories of steel components: motors, bearings, and reducers, and complete the automatic detection of steel components.

Images