Bubble identification method for lead-based reactor SGTR accident simulation research

Abstract

The invention discloses a bubble identification method for lead-based reactor SGTR accident simulation research, and relates to the technical field of target detection, and the method comprises the steps: building a visual lead-based reactor SGTR accident simulation experiment platform, collecting a bubble image in an accident simulation experiment, and carrying out the preprocessing; constructing a lead-based reactor SGTR accident simulation experiment bubble image data set; an SE attention module and a CBAM module are sequentially added between the last layer C3 network and the SPP network of the backbone network of the traditional yolov5 model to improve the traditional yolov5 model; extracting lead-based reactor SGTR accident simulation experiment bubble image features in the backbone network, performing feature fusion in the neck network, and finally performing regression prediction in the head network; according to the method, the two-dimensional enhancement of the bubble features is realized, the key features of the bubble image in channel and space dimensions are captured more accurately, and the feature extraction capability of the complex bubble image is effectively improved, so that the detection precision and speed are remarkably improved in a target detection task.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, in particular to a bubble identification method for lead-based reactor SGTR accident simulation research. Background Art

[0002] As a fourth-generation advanced nuclear energy system, the lead-bismuth reactor (LBE) has attracted much attention due to its high safety, stability and cooling performance. When the lead-bismuth reactor is operating normally, the lead-bismuth alloy coolant in the primary circuit carries a large amount of heat generated by the core and exchanges heat with the water in the secondary circuit in the steam generator. Under high-temperature flow conditions, the lead-bismuth alloy is highly corrosive to the heat transfer tube material. After long-term operation, the inner wall of the heat transfer tube may experience problems such as corrosion thinning, pitting, and stress corrosion cracking. Secondly, the large temperature and pressure differences between the primary and secondary circuits cause the transfer tube to be subjected to high thermal stress during operation. Frequent startup, shutdown, and load changes will cause repeated thermal stress, resulting in fatigue cracks in weak parts of the heat transfer tube (such as welds and bends). Therefore, a Steam Generator Tube Rupture (SGTR) accident may occur. After the SGTR accident occurs, the secondary circuit water or steam enters the primary circuit high-temperature LBE. Due to the high temperature of the LBE, the water in contact with it will quickly evaporate, generating a large amount of water vapor. These water vapors emerge from the contact interface in the form of bubbles, forming an obvious vapor-liquid two-phase flow phenomenon. These bubbles float and gather in the LBE, changing the flow characteristics of the primary circuit coolant. These bubbles are carried into the core by the LBE, which may cause local reactivity fluctuations in the core, deterioration of heat transfer, system damage and other problems, which will have a serious impact on the normal operation of the reactor. Therefore, accurately and quickly identifying the bubbles generated in the SGTR accident simulation experiment is of great significance to the future research on lead-bismuth reactors.

[0003] Traditional bubble recognition methods mainly rely on manual observation or simple image processing technology, which has low accuracy and poor efficiency. Summary of the Invention

[0004] In order to overcome the defect in the prior art that bubbles cannot be quickly and accurately identified in lead-based reactor SGTR accident simulation research experiments, the present invention proposes a bubble identification method for lead-based reactor SGTR accident simulation research.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a bubble identification method for lead-based reactor SGTR accident simulation research, comprising:

[0006] S1: Build a visual platform for lead-based reactor SGTR accident simulation experiment and collect bubble images in accident simulation experiments;

[0007] S2: preprocessing the collected bubble image to obtain a processed bubble image;

[0008] S3: Annotate the bubble positions in the processed bubble images, construct a bubble image dataset for the lead-based reactor SGTR accident simulation experiment, and divide it into training set and validation set;

[0009] S4: Improve the traditional yolov5 model to obtain the improved yolov5 model;

[0010] S5: Use the training set and validation set to train and validate the improved yolov5 model to obtain the bubble recognition model;

[0011] S6: Input the bubble image into the bubble recognition model and output the bubble recognition result.

[0012] Preferably, in step S2, the preprocessing includes performing denoising processing on the bubble image in the experiment using Gaussian filtering and performing enhancement processing on the bubble image in the experiment using a histogram equalization method.

[0013] Preferably, in step S4, the traditional yolov5 model is improved to obtain an improved yolov5 model, including:

[0014] S41: Add the SE attention module to the next layer of the last C3 network in the traditional Yolov5 model backbone network to enhance the channel features of the bubble features in the early stage of the network without changing the size of the input features;

[0015] S42: Add a CBAM module between the SE attention module and the SPP network to mine bubble features from both channel and spatial dimensions.

[0016] Preferably, in step S41, the implementation process of the SE attention module includes:

[0017] S411: The input feature f of the SE attention module first passes through the global average pooling layer for spatial feature compression, and global average pooling is implemented in the spatial dimension to obtain channel description features;

[0018] S412: The channel description features are passed through the FC fully connected layer to perform channel feature learning to obtain channel attention information;

[0019] S413: Multiply the channel attention information and the input feature f of the SE attention module channel by channel to obtain the target attention bubble feature.

[0020] Preferably, in step S412, the FC fully connected layer is composed of two fully connected layers connected in series, which perform nonlinear transformation on the channel description features; the first fully connected layer reduces the vector dimension of the channel description features and introduces the nonlinear activation function ReLU; the second fully connected layer restores the reduced vector dimension to be consistent with the original number of channels, and then uses the Sigmoid activation function to output a weight coefficient ranging from 0 to 1, that is, the channel attention information.

[0021] Preferably, in step S42, the CBAM module includes a channel attention module and a spatial attention module connected in series, and the implementation process of the CBAM module includes:

[0022] S421: Input the input features of the CBAM module to the channel attention module for global pooling to obtain channel features;

[0023] S422: Multiply the input features of the CBAM module and the channel features element-wise to obtain comprehensive channel features;

[0024] S423: Input the comprehensive channel features into the spatial attention module for pooling and convolution operations to obtain spatial attention features;

[0025] S424: Multiply the comprehensive channel feature and the spatial attention feature element-wise to obtain the channel space comprehensive feature.

[0026] Preferably, step S5 includes first inputting the training set into the improved yolov5 model, extracting the bubble image features of the lead-based reactor SGTR accident simulation experiment in the backbone network of the improved yolov5 model, then performing feature fusion in the neck network, and finally performing regression prediction in the head network; wherein the head network is composed of three output layers, each output layer is responsible for detecting bubbles of different scales, and the improved yolov5 model is trained and verified by using the training set and the validation set to finally obtain a bubble recognition model for the lead-based reactor SGTR accident simulation experiment.

[0027] A readable storage medium stores a computer program, which, when executed, implements a bubble identification method for lead-based reactor SGTR accident simulation research.

[0028] A bubble identification system for lead-based reactor SGTR accident simulation research includes a memory and a processor. The memory stores a computer program. The processor is connected to the memory and is used to execute the computer program to implement a bubble identification method for lead-based reactor SGTR accident simulation research.

[0029] The advantages of the present invention are:

[0030] (1) This paper improves the YOLOv5 model by adding the SE attention module and the CBAM module, achieving a dual-dimensional enhancement of bubble features, more accurately capturing the key features of bubble images in the channel and spatial dimensions, and effectively improving the feature extraction capability of complex bubble images, thereby significantly improving the detection accuracy and speed in target detection tasks;

[0031] (2) The present invention adaptively adjusts the feature channel weights through the SE module channel attention mechanism to highlight key bubble features and suppress background interference. At the same time, the CBAM module combines channel and spatial attention to accurately locate the spatial distribution of bubbles, solving the problem of missed detection caused by the variable bubble morphology in traditional methods.

[0032] (3) The present invention uses Gaussian filtering and histogram equalization to pre-process the bubble image, effectively overcoming the image noise and uneven illumination problems caused by the high temperature environment of LBE, and improving the clarity of the bubble outline; the introduction of the attention mechanism module significantly reduces the false detection rate caused by lead-bismuth alloy flow artifacts and steam occlusion;

[0033] (4) By constructing a dedicated bubble data set and an automated recognition model, the present invention solves the pain points of strong subjectivity in manual observation and difficulty in quantifying bubble dynamic parameters (such as rising speed and coalescence frequency), providing high-precision data support for SGTR accident analysis;

[0034] (5) The present invention studies bubble identification in lead-based reactor SGTR accident simulation experiments, quickly and accurately identifying tiny bubbles, and laying the foundation for rapid response to real lead-bismuth reactor SGTR accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the lead-based reactor SGTR accident simulation experimental device of the present invention;

[0036] Figure 2 This is the original yolov5 model diagram;

[0037] Figure 3 This is the improved yolov5 model diagram of the present invention;

[0038] Figure 4 This is the SE module architecture diagram of the present invention;

[0039] Figure 5 This is a diagram of the CBAM module architecture of the present invention;

[0040] Figure 6 This is a bubble detection result diagram of the present invention;

[0041] Figure 7 The figure is a flow chart of the steps of the method of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0043] like Figure 7 As shown, the present invention proposes a bubble identification method for lead-based reactor SGTR accident simulation research, comprising:

[0044] S1: Build a visual experimental platform for lead-based reactor SGTR accident simulation and collect bubble images in accident simulation experiments.

[0045] A visual simulation experiment platform was built to carry out a simulation experiment on the movement of bubbles in a water medium, simulating the accident condition of bubbles being injected vertically downward after an SGTR accident. The experiment was conducted under room temperature, using a high-frame-rate camera to collect the bubble image sequence in the SGTR accident simulation experiment in real time. The high-frame-rate camera used had a maximum resolution of 1920×1080 and a maximum shooting speed of 90260fps. In order to observe the dynamic behavior of bubbles in a water medium, the high-frame-rate camera supporting software was used to realize camera control and data transmission, and record the diffusion and movement behavior of the bubbles. The present invention has developed an experimental platform for the movement of bubbles in a water medium, Figure 1 The diagram below is a schematic diagram of the experimental setup. It mainly includes an air source, a glass experimental water tank, a computer, etc.

[0046] S2: preprocessing the collected bubble image to obtain a processed bubble image;

[0047] The preprocessing includes using Gaussian filtering to perform denoising on the bubble image in the experiment and using histogram equalization or contrast-first adaptive histogram equalization method to perform enhancement processing on the experimental image.

[0048] S3: Annotate the bubble positions in the processed bubble images, construct a bubble image dataset for the lead-based reactor SGTR accident simulation experiment, and divide it into training set and validation set;

[0049] In this example, a total of 1017 bubble images were processed. These images were renamed to facilitate bubble location annotation. A bubble image dataset for the lead-based reactor SGTR accident simulation experiment was then constructed. The training dataset was split into a training set and a validation set in an 8:2 ratio. The data was annotated using Labelimg software, which supports multiple annotation formats, including PascalVOC, YOLO, and CreateML. Labelimg's simple interface and intuitive operation make it suitable for annotating target objects in images.

[0050] S4: Improve the traditional yolov5 model to obtain the improved yolov5 model;

[0051] The YOLO series models have fast detection capabilities, which is crucial for real-time detection of bubbles. As the new generation target detection network of the YOLO series, YOLOv5 is the result of continuous integrated innovation based on YOLOv3 and YOLOv4. Although YOLOv5 is slightly inferior to YOLOv4 in terms of performance, it far exceeds the latter in flexibility and speed, and has significant advantages in the rapid deployment of the model. In view of this, the present invention uses the YOLOv5 model to carry out bubble recognition.

[0052] Yolov5 breaks through traditional detection methods and can simultaneously achieve target detection and localization without relying on traditional methods such as sliding windows and region proposals. It cleverly combines the two key elements of speed and accuracy, demonstrating excellent performance in real-time target detection tasks. Yolov5 uses an end-to-end training strategy based on deep learning, and with the help of a single neural network model, it can accurately predict bounding boxes and class probabilities directly from the entire image. These remarkable features have enabled Yolov5 to be widely used in the field of industrial inspection and can effectively meet the stringent requirements of various practical application scenarios.

[0053] like Figure 2 As shown in the figure, the traditional yolov5 model includes: backbone network, neck network and head network. The backbone network is used to extract the features of the input image, gradually reducing the size of the feature map while increasing the number of channels. The neck network is a network module between the backbone network and the head network. It is used to further perform feature fusion and upsampling operations on the basis of the features extracted by the backbone network to provide more advanced semantic information and the ability to adapt to pictures of different scales. The head network further processes the extracted features and generates the final output result.

[0054] The backbone network includes the first Conv network, the second Conv network, the first C3 network, the third Conv network, the second C3 network, the fourth Conv network, the third C3 network, the fifth Conv network, the fourth C3 network and the SPP network, which are connected in sequence. The Conv network represents a convolutional network, and the C3 network consists of three standard convolutional layers and multiple Bottleneck modules to enhance the feature extraction capability of the network.

[0055] The neck network includes the sixth Conv network, the first Upsample network, the first Concat network, the fifth C3 network, the seventh Conv network, the second Upsample network, the second Concat network, the sixth C3 network, the eighth Conv network, the third Concat network, the seventh C3 network, the ninth Conv network, the fourth Concat network, and the eighth C3 network connected in sequence; at the same time, the second C3 network is connected to the second Concat network, the third C3 network is connected to the first Concat network, the spp network is connected to the sixth Conv network, the second Concat network is connected to the sixth C3 network, the seventh Conv network is connected to the first Concat network The three Concat networks are connected, the sixth Conv network is connected to the fourth Concat network; and the first Concat network is used to splice the output of the third C3 network and the output of the first Upsample network, the second Concat network is used to splice the output of the second C3 network and the output of the second Upsample network, the third Concat network is used to splice the output of the seventh Conv network and the output of the eighth Conv network, and the fourth Concat network is used to splice the output of the sixth Conv network and the output of the ninth Conv network; wherein the Upsample network is an upsampling network, and the Concat network is a splicing network.

[0056] The head network is a detection network. The Detect network is provided with three detection layers, which respectively detect the output of the sixth C3 network, the output of the seventh C3 network, and the output of the eighth C3 network.

[0057] The yolov5 model is an existing model structure in this field, in which the backbone network, neck network, and head network are clearly defined in the yolov5 model and are technical common sense in this field.

[0058] In step S4, the traditional yolov5 model is improved to obtain the improved yolov5 model, such as Figure 3 Shown, including:

[0059] S41: Add the SE (Squeeze-and-Excitation) module to the next layer of the fourth C3 network in the traditional Yolov5 model backbone network to enhance the channel features of the bubble features in the early stage of the network without changing the size of the input features;

[0060] Dynamically adjust the weights of different channels, enhance the response of important feature channels, and suppress unimportant channels; enhance feature selection capabilities to improve model accuracy;

[0061] In step S41, the SE attention module is added to the next layer of the fourth C3 network in the traditional yolov5 model backbone network. The implementation process of the SE attention module is as follows: Figure 4 Shown, including:

[0062] S411: The input feature f of the SE attention module first passes through the global average pooling layer for spatial feature compression, and global average pooling is implemented in the spatial dimension to obtain channel description features;

[0063] The dimension of the input feature f is W×H×C, and the dimension of the channel description feature is 1×1×C;

[0064] S412: The channel description features are passed through the FC fully connected layer to perform channel feature learning to obtain channel attention information;

[0065] The dimension of the channel attention information is 1×1×C;

[0066] The FC fully connected layer consists of two fully connected layers connected in series, which performs a nonlinear transformation on the channel description features. The first fully connected layer reduces the vector dimension of the channel description features to reduce the amount of computation, and introduces the nonlinear activation function ReLU to enhance the model's expressiveness. The second fully connected layer restores the reduced vector dimension to the same number of original channels, and then uses the Sigmoid activation function to output a weight coefficient ranging from 0 to 1, i.e., the channel attention information. In this way, the SE attention mechanism can focus on channel features that are more discriminative for target detection and classification, suppress useless or interfering channel information, thereby improving the model's ability to extract bubble image features and thus improving detection accuracy.

[0067] S413: Multiply the channel attention information and the input feature f of the SE attention module channel by channel to obtain the target attention bubble feature.

[0068] The dimension of the target attention bubble feature is W×L×C.

[0069] S42: Add a CBAM (Convolutional Block Attention Module) module between the SE attention module and the SPP network to mine bubble features from two dimensions: channel and space.

[0070] The CBAM module includes a channel attention module and a spatial attention module connected in series.

[0071] It combines the channel attention module and the spatial attention module to optimize the channel dimension first and then the spatial dimension, so that the network pays more attention to important areas and features.

[0072] In step S42, a CBAM module is added between the SE attention module and the SPP network in the traditional yolov5 model backbone network. The implementation process of the CBAM module is as follows: Figure 5 Shown, including:

[0073] S421: Input the input features of the CBAM module to the channel attention module for global pooling to obtain channel features;

[0074] S422: Multiply the input features of the CBAM module and the channel features element-wise to obtain comprehensive channel features;

[0075] S423: Input the comprehensive channel features into the spatial attention module for pooling and convolution operations to obtain spatial attention features;

[0076] S424: Multiply the comprehensive channel feature and the spatial attention feature element-wise to obtain the channel space comprehensive feature.

[0077] By integrating the CBAM attention mechanism into the yolov5 backbone network, the model can more accurately capture the key features of bubble images in the channel and spatial dimensions, enhance the feature extraction capability of small-scale bubbles, and effectively improve the feature extraction capability of complex bubble images, thereby significantly improving the detection accuracy in target detection tasks and better coping with various challenges in bubble image detection.

[0078] In the overall framework of the improved yolov5 model, SE and CBAM attention mechanisms work together to play an important role in improving model performance.

[0079] S5: The improved yolov5 model is trained and verified using the training set and validation set to obtain the bubble recognition model for the lead-based reactor SGTR accident simulation experiment;

[0080] In step S5, the improved yolov5 model is trained and verified using the training set and the validation set to obtain a bubble recognition model for the lead-based reactor SGTR accident simulation experiment. First, the training set is input into the improved yolov5 model, and the bubble image features of the lead-based reactor SGTR accident simulation experiment are extracted in the backbone network of the improved yolov5 model. Then, feature fusion is performed in the neck network to better utilize the features extracted by the backbone network, and finally, regression prediction is performed in the head network; wherein the head network is composed of three output layers, each output layer is responsible for detecting bubbles of different scales, and the head network is provided with three detection layers. The detection layer corresponding to the output of the sixth C3 network is used to detect small bubbles, which is derived from the shallow features of the bubbles, retains edge details at high resolution, and matches tiny bubbles through small anchor frames; the detection layer corresponding to the output of the seventh C3 network is used to detect medium bubbles, fuses the mid-layer features of the bubbles, balances semantics and position information, and uses medium anchor frames to detect aggregated bubbles; the detection layer corresponding to the output of the eighth C3 network is used to detect large bubbles or bubble groups, fuses the deep features of the bubbles to extract the global context, and the large anchor frame covers the complete area of ​​the large bubbles. By using the training set and validation set to train and validate the improved yolov5 model, a bubble recognition model for the lead-based reactor SGTR accident simulation experiment was finally obtained.

[0081] The physical properties of bubbles in SGTR accidents (such as initial generation, aggregation, and fragmentation) will cause significant differences in bubble size, ranging from isolated bubbles as small as millimeters to clusters of aggregated bubbles as large as centimeters. A single-scale detection layer is difficult to simultaneously capture target features in a wide size range; for traditional convolutional neural networks, there is a contradiction between the semantic information of the deep network and the detailed information of the shallow network: the deep feature map (after multiple downsampling) has stronger semantic information (such as the overall shape of the bubble), but the spatial resolution is low, and it is difficult to accurately locate small bubbles; the shallow feature map retains high-resolution details (such as the edge of the bubble), but lacks high-level semantic understanding; there is a problem of matching the receptive field with the target size: a large receptive field is suitable for detecting large targets, but will lose the local features of small targets; a small receptive field is sensitive to small targets, but cannot cover the global information of large targets.

[0082] Therefore, lead-based reactor accident simulation experiments require real-time monitoring, and multi-scale parallel detection (rather than cascade detection) can meet real-time requirements while ensuring accuracy.

[0083] S6: Input the bubble image into the bubble recognition model of the lead-based reactor SGTR accident simulation experiment, and output the bubble recognition result, which includes the bubble and the corresponding position information.

[0084] like Figure 6As shown in the figure, the bubble detection results of the lead-based reactor SGTR accident simulation experiment of the present invention are shown. The figure not only shows the detected bubbles, but also the location information of the bubbles and the confidence level of the bubbles. For example, the figure shows three possible bubbles with confidence levels of 0.87, 0.86 and 0.62 respectively.

[0085] In order to verify the effectiveness of the improved model, the present invention conducts ablation experiments based on the original yolov5 model. The comparison results of the improved model are shown in Table 1.

[0086] Table 1 Comparison results of the improved models

[0087]

[0088] The results in Table 1 show that with the simultaneous introduction of the SE and CBAM attention mechanisms, the Yolov5+SE+CBAM model achieves significant improvements over the original Yolov5 model, with precision, recall, and average precision increasing by 14.8%, 8%, and 5%, respectively. Experimental verification further confirms the effectiveness of the improved algorithm.

[0089] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0090] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0091] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A bubble identification method for lead-based reactor SGTR accident simulation research, characterized in that: include: S1: Build a visual platform for lead-based reactor SGTR accident simulation experiment and collect bubble images in accident simulation experiments; S2: preprocessing the collected bubble image to obtain a processed bubble image; S3: Annotate the bubble positions in the processed bubble images, construct a bubble image dataset for the lead-based reactor SGTR accident simulation experiment, and divide it into training set and validation set; S4: Improve the traditional yolov5 model to obtain the improved yolov5 model; S5: Use the training set and validation set to train and validate the improved yolov5 model to obtain the bubble recognition model; S6: Input the bubble image into the bubble recognition model and output the bubble recognition result.

2. The bubble identification method for lead-based reactor SGTR accident simulation research according to claim 1, characterized in that: In step S2, the preprocessing includes performing denoising processing on the bubble image in the experiment using Gaussian filtering and performing enhancement processing on the bubble image in the experiment using a histogram equalization method.

3. The bubble identification method for lead-based reactor SGTR accident simulation research according to claim 1, characterized in that: In step S4, the traditional yolov5 model is improved to obtain an improved yolov5 model, including: S41: Add the SE attention module to the next layer of the last C3 network in the traditional Yolov5 model backbone network to enhance the channel features of the bubble features in the early stage of the network without changing the size of the input features; S42: Add a CBAM module between the SE attention module and the SPP network to mine bubble features from both channel and spatial dimensions.

4. The bubble identification method for lead-based reactor SGTR accident simulation research according to claim 3 is characterized in that: In step S41, the implementation process of the SE attention module includes: S411: The input feature f of the SE attention module first passes through the global average pooling layer for spatial feature compression, and global average pooling is implemented in the spatial dimension to obtain channel description features; S412: The channel description features are passed through the FC fully connected layer to perform channel feature learning to obtain channel attention information; S413: Multiply the channel attention information and the input feature f of the SE attention module channel by channel to obtain the target attention bubble feature.

5. The bubble identification method for lead-based reactor SGTR accident simulation research according to claim 4, characterized in that: In step S412, the FC fully connected layer is composed of two fully connected layers connected in series, and a nonlinear transformation is performed on the channel description features; The first fully connected layer reduces the vector dimension of the channel description feature and introduces the nonlinear activation function ReLU; the second fully connected layer restores the reduced vector dimension to the same as the original number of channels, and then uses the Sigmoid activation function to output a weight coefficient ranging from 0 to 1, that is, the channel attention information.

6. The bubble identification method for lead-based reactor SGTR accident simulation research according to claim 3, characterized in that: In step S42, the CBAM module includes a channel attention module and a spatial attention module connected in series. The implementation process of the CBAM module includes: S421: Input the input features of the CBAM module to the channel attention module for global pooling to obtain channel features; S422: Multiply the input features of the CBAM module and the channel features element-wise to obtain comprehensive channel features; S423: Input the comprehensive channel features into the spatial attention module for pooling and convolution operations to obtain spatial attention features; S424: Multiply the comprehensive channel feature and the spatial attention feature element-wise to obtain the channel space comprehensive feature.

7. The bubble identification method for lead-based reactor SGTR accident simulation research according to claim 1, characterized in that: Step S5 includes first inputting the training set into the improved yolov5 model, extracting the bubble image features of the lead-based reactor SGTR accident simulation experiment in the backbone network of the improved yolov5 model, then performing feature fusion in the neck network, and finally performing regression prediction in the head network; The head network consists of three output layers, each of which is responsible for detecting bubbles of different scales. The improved yolov5 model is trained and verified using the training set and validation set, and finally a bubble recognition model for the lead-based reactor SGTR accident simulation experiment is obtained.

8. A readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the bubble identification method for lead-based reactor SGTR accident simulation research as described in any one of claims 1 to 7 is implemented.

9. A bubble identification system for lead-based reactor SGTR accident simulation research, characterized in that: The invention comprises a memory and a processor, wherein a computer program is stored in the memory, the processor is connected to the memory, and the processor is used to execute the computer program to implement a bubble identification method for lead-based reactor SGTR accident simulation research as described in any one of claims 1 to 7.

Images