A nuclear containment crack detection method and system based on deep learning

Abstract

The present invention provides a deep learning-based nuclear containment crack detection method and system. The method comprises automatically collecting and preprocessing image data using a remote acquisition system, selecting and annotating images containing cracks, and establishing a containment crack recognition dataset and a crack segmentation dataset. A deep learning model is constructed, including a containment crack recognition network model and a containment crack segmentation network model, each of which incorporates an attention mechanism module and a void space pyramid pooling module. The containment crack recognition network model and the containment crack segmentation network model are trained using the containment crack recognition dataset and the crack segmentation dataset, respectively. The trained containment crack recognition network model and the containment crack segmentation network model are then used to sequentially identify and segment cracks in containment surface images. The present invention employs a two-stage detection process to achieve efficient and high-precision crack segmentation of nuclear containment surface images.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear power plant containment detection, and in particular relates to a nuclear containment crack detection solution based on deep learning. Background Art

[0002] Traditional fossil energy sources are becoming increasingly scarce, while the pollution caused by the extensive use of fossil fuels is increasingly negatively impacting the ecological environment. Nuclear energy is an indispensable alternative energy source, both economically and environmentally. Currently, nuclear energy has become the world's second-largest low-carbon energy source, with approximately 450 reactors providing approximately 10% of the world's electricity. The nuclear power industry is of great significance to controlling air pollution, adjusting the energy structure, and transforming the economic development model.

[0003] The nuclear containment vessel is the last barrier in a nuclear power plant to prevent the release of radioactive materials from the reactor. Although concrete-constructed containments offer excellent sealing and stability, their concrete walls are exposed to sunlight, chloride-rich rain, and sea breezes, which can lead to carbonization of the concrete surface. Under the load of the containment vessel itself, cracking can also occur to some extent on the shear wall surfaces. To verify structural strength and compressive strength, nuclear power plants regularly conduct pressure tests on the containment vessel. During these tests, a significant pressure differential exists between the inside and outside of the containment vessel, which can also affect the walls. Therefore, regular inspection and maintenance of the containment vessel are essential for the safe operation of nuclear power plants. Cracks are a common aging defect in containment vessels, and detection of these cracks is a key indicator during visual inspections.

[0004] Early crack detection in containment vessels mostly relied on manual visual inspection, a method that is time-consuming, risky, and inefficient. To improve detection efficiency, digital image acquisition technology has been applied to structural defect detection. Among the more successful applications are the SCANITE remote defect detection system developed by the French company SITES and the CEDIS system of Wuhan University. The CEDIS system not only enables remote imaging but also automated route planning and image capture. The CEDIS system is in operation at several nuclear power plants in my country. While remote acquisition of containment surface images is possible, automated processing of these images has not yet been achieved.

[0005] With the development of science and the update of technology, digital image processing technology has been widely used in life and work. Compared with manual annotation methods, this method has the characteristics of good reproducibility, high accuracy and high flexibility.

[0006] Detection technology based on digital image processing offers advantages such as speed and convenience, and has gradually replaced manual inspection methods. However, crack detection using traditional digital image processing methods relies heavily on the algorithm designer's experience. Detection results are significantly affected by noise generated by factors such as lighting, angle, and deformation, resulting in low detection accuracy and poor robustness.

[0007] In summary, there is an urgent need for an efficient and accurate detection method for containment crack detection. Summary of the Invention

[0008] The present invention aims to provide a deep learning-based nuclear containment crack detection solution to address efficiency and accuracy issues in nuclear containment crack detection. Directly processing observed images to segment cracks is labor-intensive and can also be affected by noise. Therefore, the present invention proposes a two-stage containment crack segmentation method based on deep learning. First, images containing cracks are pre-screened using a classification network. Then, a segmentation network is used to extract cracks, achieving efficient and high-precision containment crack segmentation.

[0009] In order to achieve the above objectives, the technical solution proposed in the present invention includes a nuclear containment crack detection method based on deep learning, comprising the following steps:

[0010] Step 1: Use a remote acquisition system to automatically collect and preprocess nuclear containment appearance image data, select and annotate images containing cracks, and establish a containment crack identification dataset and a crack segmentation dataset;

[0011] Step 2: Construct a deep learning model, including building a containment crack identification network model and a containment crack segmentation network model. The containment crack segmentation model is based on the U-net convolutional neural network, and adds an attention mechanism module and a void space pyramid pooling module. The attention mechanism is used to highlight areas that need to be focused on, while suppressing the activation of features in irrelevant areas to improve the model sensitivity and accuracy. The void space pyramid pooling module is added to optimize the extraction of local and global information, enhancing the ability of images to obtain contextual information and multi-scale information.

[0012] Step 3: using the containment crack identification dataset and the crack segmentation dataset, respectively train the containment crack identification network model and the containment crack segmentation network model;

[0013] Step 4: Use the trained containment crack recognition network model and containment crack segmentation network model to identify and segment the cracks in the containment appearance image in turn; the implementation method is as follows:

[0014] First, the nuclear containment image to be detected is input into the containment crack recognition network model to screen out images with cracks.

[0015] The filtered images containing cracks are then input into the containment crack segmentation network model for processing, and the nuclear containment cracks in the image are marked.

[0016] Moreover, step 1 is implemented as follows,

[0017] (1) First, a control network is deployed on the ground near the containment, and a shooting plan is planned according to the terrain. A collection device consisting of a single-lens reflex camera and a pan-tilt device is used to quickly and automatically collect high-resolution images of the exterior from a long distance;

[0018] (2) After obtaining the original collected images, the oblique images are corrected, spliced, resampled and cropped according to the acquisition point coordinates and internal and external orientation elements to obtain orthographic images of uniform size and resolution, forming a panoramic view of the containment shell appearance;

[0019] (3) The containment appearance images are divided into two categories: crack presence and crack absence to construct a crack recognition dataset, and the cracks in the images with cracks are marked to construct a crack segmentation dataset.

[0020] Moreover, the containment crack identification network is implemented using Inception V3.

[0021] Moreover, the containment crack identification network adopts a decoding-encoding structure, which includes four Encode layers and four Decode layers. The processing results of each Encode layer are transmitted to the next Encode layer and the AG+ module at the same level respectively, and the processing results of each Decode layer are transmitted to the next Decode layer and the AG+ module at the next level respectively. The last Encode layer is connected to a convolution module and an ASPP module in sequence, and then connected to the AG+ module at the same level and the first Decode layer respectively.

[0022] Moreover, the ASPP module consists of a 1×1 convolution, three dilated convolutions with different dilated rates, and an image pooling module. Image Pooling includes an average pooling, a 1×1 convolution Conv, and an upsampling layer Upsample. The dilated convolution uses dilated rates of 6, 12, and 18, respectively. The ASPP module finally merges the results of multiple dilated convolutions and pooling processes and inputs them into a 1×1 convolution layer to obtain the output.

[0023] Moreover, the AG+ module adopts a parallel structure with two inputs and one output, one of which is the output feature x from the same encoding layer. l, the second input is the input feature g from the decoding layer,

[0024] One branch is to combine the input feature g of the decoding layer and the output feature x of the encoding layer l Input a 1×1 convolution respectively, then add the results and input them into the ReLU activation function, and then pass through a 1×1 convolution layer and a Sigmoid activation function to obtain the additive attention coefficient;

[0025] The other branch is to combine the input feature g of the decoding layer and the output feature x of the encoding layer l Input a 1×1 convolution respectively, then multiply the results and input the result into the ReLU activation function, and then pass through a 1×1 convolution layer and a Sigmoid activation function to obtain the dot product attention coefficient;

[0026] Finally, the results of the two branches are added together and then added to the output feature x of the encoding layer. l Multiply to enhance the crack information of interest.

[0027] Moreover, the training of the containment crack identification network model is implemented by transfer learning; the training process of the containment crack segmentation network model adopts the Adam optimization algorithm, and the polynomial StepLR is used as the learning rate scheduler to prevent model overfitting.

[0028] On the other hand, the present invention also provides a nuclear containment crack detection system based on deep learning, which is used to implement the nuclear containment crack detection method based on deep learning as described above.

[0029] Moreover, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the nuclear containment crack detection method based on deep learning as described above.

[0030] Alternatively, it includes a readable storage medium having a computer program stored thereon, and when the computer program is executed, it implements the nuclear containment crack detection method based on deep learning as described above.

[0031] The present invention provides a novel containment crack segmentation method based on deep learning, which adopts a two-stage detection method for cracks, and can achieve high-efficiency and high-precision segmentation of cracks on the surface image of the nuclear containment, reducing the workload of personnel in the detection process and improving detection efficiency and detection accuracy.

[0032] The solution of the present invention is simple and convenient to implement and has strong practicality. It solves the problems of low practicality and inconvenience in actual application existing in related technologies, can improve user experience, and has important market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is an overall flow chart of containment crack detection according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of transfer learning according to an embodiment of the present invention;

[0035] Figure 3 This is a structural diagram of the ASPP module according to an embodiment of the present invention;

[0036] Figure 4 This is a structural diagram of a security crack segmentation network model according to an embodiment of the present invention;

[0037] Figure 5 This is a structural diagram of the Attention Gate+ according to an embodiment of the present invention;

[0038] Figure 6 This is an example diagram of the detection results of an embodiment of the present invention, where part (a) is the original image, part (b) is the label image, and part (d) is the prediction result. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0040] This invention discloses a deep learning-based containment crack detection solution. First, a camera is used to remotely capture images of the containment exterior. These images are processed, and tilted images are corrected, spliced, and resampled to produce orthophotos of uniform size and resolution. Selected images are then cropped to construct a containment crack recognition dataset. A deep learning recognition model is then built to filter out images with cracks from a large number of containment images. These filtered images with cracks are then annotated to construct a crack segmentation dataset. Subsequently, an ASPP module and an attention mechanism are added to a U-Net network to construct a containment crack segmentation model, improving the model's accuracy in identifying small cracks. The model is then trained using the crack segmentation dataset. The trained recognition and segmentation networks are then used to process the images to be detected. The recognition model first determines whether cracks are present in the image. If so, the image is fed into the segmentation model to segment the cracks. This solution improves crack detection efficiency while achieving good detection accuracy.

[0041] In the embodiment of the present invention, a comprehensive process of containment crack detection based on deep learning is proposed. Figure 1 The specific steps are as follows:

[0042] S1. Use a remote acquisition system to automatically collect nuclear containment appearance image data, preprocess the images, select and annotate images containing cracks, and establish a containment crack identification dataset and a crack segmentation dataset;

[0043] (1) First, a control network is deployed on the ground near the containment, and a shooting plan is planned based on the terrain. It is recommended that the acquisition equipment consisting of a single-lens reflex camera and a pan / tilt device independently developed by the inventor team (patent publication CN105606627A) be used to quickly and automatically capture high-resolution images of the exterior from a distance, ensuring that most of the surface area of ​​the containment is covered.

[0044] (2) After obtaining the original acquired image, the oblique image is corrected, spliced, resampled, and cropped based on the acquisition point coordinates and internal and external orientation elements using the principles of photogrammetry to obtain an orthophoto with uniform size and resolution, forming a panoramic image of the containment shell's exterior. Image correction is to perform differential correction on the acquired oblique image to obtain an orthophoto of the containment shell's exterior. In this embodiment, the image size after resampling and cropping is 4593×3675, and the actual range size corresponding to each image is 750mm×600mm.

[0045] (3) Use a collection device consisting of a single-lens reflex camera and a pan-tilt device to collect high-resolution images of the appearance. The collected images are then corrected and spliced. In the embodiment, the image of size 4593×3675 is cropped into small images of size 512*512, divided into two categories: crack presence and crack absence, to construct a crack recognition dataset, and then photos with cracks are cropped from the image and annotated. In specific implementation, the 4593×3675 image can be cropped using a sliding window method with a step size of 512. The last column and the last row that are less than 512 are padded forward. The annotation can be implemented using existing software or can be pre-annotated and the results are used as input data. The crack area is set to white and the non-crack area is set to black to obtain the containment crack segmentation dataset. Furthermore, the ratios of the training set, validation set, and test set are preferably recommended to be set to 0.7, 0.2, and 0.1.

[0046] S2. Build a deep learning model: Build a containment crack identification network model and a containment crack segmentation network model;

[0047] (1) Containment Crack Identification Network: This embodiment preferably uses Inception V3 as the identification network. Inception V3 is an existing technology with 11 Inception modules, including 96 convolutional layers, and is an improvement on the Inception V1 network. It mainly introduces batch normalization (BN) and optimizes the Inception module.

[0048] (2) The containment crack segmentation model is improved based on the U-Net network. The U-Net model will continuously reduce the image size during the downsampling calculation process. The cracks on the containment have complex shapes and different sizes, resulting in poor detection of small-scale cracks in the crack segmentation task. Based on the U-net convolutional neural network, an attention mechanism module (AG+) and an atrous space pyramid pooling module (ASPP) are added. The attention mechanism is used to highlight the areas that need to be focused on, while suppressing the activation of features in irrelevant areas to improve the sensitivity and accuracy of the model. The addition of the atrous space pyramid pooling module optimizes the extraction of local and global information and enhances the image's ability to obtain contextual information and multi-scale information.

[0049] Furthermore, the containment crack segmentation model structure is constructed as follows Figure 4 As shown in the figure, it is divided into two parts: encoding and decoding. The encoding part mainly consists of a convolution module, a pooling layer, and an ASPP module. The convolution module mainly includes a convolution layer, a batch normalization layer, and an activation function. Since the cracks account for a small proportion in the image, the convolution in each convolution unit is a 3×3 convolution (Conv) with a stride of 1. The activation layer uses the ReLU function (Rectified Linear Unit), which is used to improve the nonlinear expression ability of the model. To prevent overfitting, a batch normalization (BN) layer is added after the convolution. A maximum pooling layer is added between the convolution modules to reduce the dimensionality of the feature map. The ASPP module is added at the end of the encoding part.

[0050] Furthermore, the decoding part needs to reduce the high-dimensional feature map obtained by the encoding part to the same size as the input image. It mainly consists of an upsampling module and an attention module. The upsampling module contains an upsampling operation and the same convolution module as the encoding process. The upsampling operation can effectively expand the size of the high-dimensional feature map. The attention mechanism, combined with contextual information, highlights the areas that need to be focused on, while suppressing the activation of features in irrelevant areas to improve the sensitivity and accuracy of the model. At the end of the network, 1×1 convolution and sigmoid function are used to map the 64-channel features to two types of ("cracks" and "non-cracks") pixel-level segmentation maps.

[0051] For ease of implementation and reference, the following is a detailed description of the network implementation of the embodiment of the present invention:

[0052] 1) Construct a containment crack segmentation network. It is recommended to implement it based on the Inception V3 model in the existing technology and use the transfer learning method for training. The transfer learning method and the structure of the classification network are as follows: Figure 2As shown in the figure, Input represents the input layer, Conv represents the convolutional layer, Max Pool represents the maximum pooling layer, FC represents the fully connected layer, Inception represents the Inception module group, and softmax represents the softmax function. The Inception V3 model has three Inception modules, denoted as Inception A, Inception B, and Inception C. Only the last Inception module group and the output layer are trained; the rest use pre-trained parameters to reduce the model's requirement for the number of dataset images.

[0053] 2) Using the parameters obtained through training, a containment crack recognition model is constructed, and the image to be detected is input into the recognition model to filter out images with cracks.

[0054] 3) Construct a containment crack segmentation model based on deep learning. The constructed model is improved based on Unet. The model structure is as follows Figure 4 As shown, the overall structure is still a decoding-encoding structure. The input image is a 512*512 three-channel color image. The convolution layer parameters, pooling, and upsampling output feature map sizes of each stage are shown in the following table. The model has four Encode layers (respectively recorded as Encode1, Encode2, Encode3, and Encode4. The results of each Encode processing are passed to the next Encode layer and the AG+ module at the same level) and four Decode layers (respectively recorded as Decode1, Decode2, Decode3, and Decode4. The results of each Decode layer processing are passed to the next Decode layer and the AG+ module at the next level). For example, the result of the input image processed by Encode1 is transmitted to Encode2 and the AG+ of the same level respectively, the result of Encode2 processing is transmitted to Encode3 and the AG+ of the same level respectively, the result of Encode3 processing is transmitted to Encode4 and the AG+ of the same level respectively, and the last Encode layer Encode4 is connected to a convolution module and an ASPP module in sequence and then connected to the AG+ of the same level and the first Decode layer Decode4 respectively. At the same time, an AG+ (Attention Gate+) is passed between the Encode layer and the Decode layer.

[0055] Module connections are used to enhance the information of interest to the model, that is, the result of Encode4 processing is passed to Decode4 via the same level AG+, the result of Encode3 processing is passed to Decode3 via the same level AG+, the result of Encode2 processing is passed to Decode2 via the same level AG+, and the result of Encode1 processing is passed to Decode4 via the same level AG+.

[0056] The results of Decode4 processing are transmitted to Decode3 and the AG+ module of the previous level respectively. The results of Decode3 processing are transmitted to Decode2 and the AG+ module of the previous level respectively. The results of Decode2 processing are transmitted to Decode1 and the AG+ module of the previous level respectively. Decode1 obtains the output result.

[0057]

[0058] 4) Each Encode module consists of a 3×3 convolution and a 2×2 maximum pooling. The ASPP module is added at the end of the encoding part. Its structure is as follows Figure 3 As shown, it consists of a 1×1 convolution and three dilated convolutions with different dilated rates, as well as an image pooling module Image Pooling. Image Pooling includes an average pooling Pooling, a 1×1 convolution Conv, and an upsampling layer Upsample. The dilated convolutions use dilated rates of 6, 12, and 18, respectively. The ASPP module finally merges the results of multiple dilated convolutions and pooling processes and concats them into a 1×1 convolution layer to obtain the output Output. ASPP obtains different receptive fields through parallel dilated convolutions, which is beneficial for segmenting objects of different scales. Each Decode module consists of a 2×2 upsampling layer and a 3×3 convolution layer. In order to combine contextual information, an AG+ (Attention Gate+) module is constructed, and its structure is as follows Figure 5 As shown in Figure 2, fusing the output of the encoding layer with the features of the decoding layer can effectively suppress background features and highlight crack features. At the end of the network, a 1×1 convolution and a sigmoid function are used to map the 64-channel features to pixel-level segmentation maps of two classes (cracks and non-cracks).

[0059] 5) Further, Figure 5 The AG+ (Attention Gate+) module shown in the figure uses a parallel structure with two inputs and one output. One of the inputs is the output feature x from the same encoding layer. l , the second input is the input feature g from the decoding layer. The two inputs have different feature attributes. Since the output feature x of the encoder l Coming from a shallow network, it contains more spatial information. l Compared with g, which comes from a deeper network, it has better feature representation. One of the branches of the AG+ (Attention Gate+) module is to combine the input feature g of the decoding layer and the output feature x of the encoding layer. lInput a 1×1 convolution respectively, then add the results and input them into the ReLU activation function, and then pass through a 1×1 convolution layer and Sigmoid activation function to obtain the additive attention coefficient. The other branch is to add the input feature g of the decoding layer and the output feature x of the encoding layer. l Input a 1×1 convolution respectively, then multiply the results and input the result into the ReLU activation function, and then pass through a 1×1 convolution layer and Sigmoid activation function to get the dot product attention coefficient. Finally, add the results of the two branches and add them to the output feature x of the encoding layer. l Multiply to enhance the crack information of interest. Figure 5 Conv is a 1×1 convolution. Relu and Sigmoid are nonlinear activation functions. In addition, since x l Since g and g come from networks of different depths, their sizes are different, so an upsampling operation is required to make them the same size as x. l same.

[0060] S3 uses the data set constructed in S1 to train the crack recognition model and the crack segmentation model.

[0061] (1) The training of the crack recognition network model adopts the transfer learning method, which directly uses the parameters obtained by training on a large sample data set. During the training process, some areas are frozen and only the last layer of the Inception module and the output layer are trained. This reduces data requirements and speeds up training.

[0062] (2) In the crack segmentation network model training process, the Adam optimization algorithm is adopted, and the polynomial StepLR is used as the learning rate scheduler to prevent the model from overfitting.

[0063] Furthermore, the parameters for training the containment crack segmentation model are set as follows: a batch size of 6, and training using the Adam optimizer. The loss function uses a combination of BCE and Dice loss. This effectively mitigates the impact of dataset class imbalance while ensuring normal gradient changes. The formula is as follows:

[0064]

[0065]

[0066] L=0.5L BcE +0.5L Dice loss

[0067] Where n is the total number of images, y i is the true value of sample i, is the predicted result. X is the original value, and Y is the predicted value. The segmentation model is trained using the crack segmentation dataset, and the Adam optimization algorithm is used. Adam combines the advantages of Adagrad (a gradient descent algorithm with adaptive learning rate) and momentum gradient descent. It can adapt to sparse gradients and alleviate gradient oscillation. The Adam formula is as follows:

[0068] m t =beta1*m t-1 +(1-beta1)*g

[0069] v t =beta2*v t-1 +(1-beta2)*g*g

[0070]

[0071] Formula 1 is used to calculate the first-order exponentially smoothed value of the historical gradient, which is used to obtain the gradient value with momentum. Where mt is the first-order moment estimate of the gradient, t is the time step, initialized to 0, g is the gradient at time step t, and beta1 represents the exponential decay rate of the first-order moment estimate, with an initial value of 0.9.

[0072] Formula 2: Calculate the second-order exponential smoothing value of the historical gradient square to obtain the learning rate weight parameter of each weight parameter. t is the second-order moment estimate of the gradient, t is the time step, initialized to 0, g is the gradient when the time step is t, beta2 represents the exponential decay rate of the second-order moment estimate, and the initial value is 0.999.

[0073] Formula 3: Calculate the variable update value, which is proportional to the first-order exponential smoothing value of the historical gradient and inversely proportional to the first-order exponential smoothing value of the square of the historical gradient. t is the learning rate, and ε is a constant added to maintain numerical stability

[0074] For ease of implementation reference, the torch.optim.Adam and parameter settings used in the training process are provided as follows:

[0075] torch.optim.Adam(params, lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0)

[0076] Params: Parameters used for iterative optimization

[0077] lr (float): learning rate (1e-4)

[0078] betas (Tuple[float, float]): coefficients (0.9, 0.999) used to calculate the gradient and the square of the gradient

[0079] eps (float): A term added to the denominator to increase numerical stability (1e-8)

[0080] weight_decay (float): weight decay (L2 penalty) (0)

[0081] To prevent overfitting during model training, we use the polynomial StepLR as the learning rate scheduler. After every 20 epochs, the learning rate is reduced by a factor of 0.5.

[0082] S4. Using the trained crack recognition model and crack segmentation model, cracks in the containment shell appearance image are identified and segmented in sequence;

[0083] (1) First, the nuclear containment image to be detected is input into the crack recognition network to filter out images with cracks and improve the detection efficiency.

[0084] (2) The image containing cracks filtered out in the previous step is input into the crack segmentation network for processing, and the cracks in the nuclear containment shell in the image are marked.

[0085] (3) After the image is processed by the segmentation model, there are still many noises in the segmented cracks. Post-processing is performed to connect the broken cracks in the image. A noise elimination method is constructed to eliminate the independent noise in the predicted image. The detection results are as follows Figure 6 As shown, part (a) is the original image, part (b) is the label image, and part (d) is the prediction result. Figure 1 、 2 Comparing the processing results, it can be seen that the cracks in the figure can be accurately identified, and the noise of false detection is small.

[0086] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0087] In some possible embodiments, a deep learning-based nuclear containment crack detection system is provided, including a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a deep learning-based nuclear containment crack detection method as described above.

[0088] In some possible embodiments, a deep learning-based nuclear containment crack detection system is provided, comprising a readable storage medium having a computer program stored thereon. When the computer program is executed, the deep learning-based nuclear containment crack detection method as described above is implemented.

[0089] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A nuclear containment crack detection method based on deep learning, characterized by: The following steps are included: Step 1: Use a remote acquisition system to automatically collect and preprocess nuclear containment appearance image data, select and annotate images containing cracks, and establish a containment crack identification dataset and a crack segmentation dataset; Step 2: Constructing a deep learning model, including building a containment crack identification network model and a containment crack segmentation network model. The containment crack segmentation network model is based on the U-net convolutional neural network, adding an attention mechanism module and a void space pyramid pooling module. The attention mechanism is used to highlight areas that need to be focused on, while suppressing the activation of features in irrelevant areas to improve the model sensitivity and accuracy. The void space pyramid pooling module is added to optimize the extraction of local and global information, enhancing the ability of images to obtain contextual information and multi-scale information. Step 3: using the containment crack identification dataset and the crack segmentation dataset, respectively train the containment crack identification network model and the containment crack segmentation network model; Step 4: using the trained containment crack recognition network model and containment crack segmentation network model, cracks in the containment appearance image are identified and segmented in sequence; The implementation is as follows, First, the nuclear containment image to be detected is input into the containment crack recognition network model to screen out images with cracks. The filtered images containing cracks are then input into the containment crack segmentation network model for processing, and the nuclear containment cracks in the image are marked.

2. The nuclear containment crack detection method based on deep learning according to claim 1, characterized in that: Step 1 is implemented as follows, (1) First, a control network is deployed on the ground near the containment, and a shooting plan is planned according to the terrain. A collection device consisting of a single-lens reflex camera and a pan-tilt device is used to quickly and automatically collect high-resolution images of the exterior from a long distance; (2) After obtaining the original collected images, the oblique images are corrected, spliced, resampled and cropped according to the acquisition point coordinates and internal and external orientation elements to obtain orthographic images of uniform size and resolution, forming a panoramic view of the containment shell appearance; (3) The containment appearance images are divided into two categories: crack presence and crack absence to construct a crack recognition dataset, and the cracks in the images with cracks are marked to construct a crack segmentation dataset.

3. The nuclear containment crack detection method based on deep learning according to claim 1, characterized in that: The containment crack identification network is implemented using Inception V3.

4. The nuclear containment crack detection method based on deep learning according to claim 1, characterized in that: The containment crack identification network adopts a decoding-encoding structure, which includes four Encode layers and four Decode layers. The results processed by each Encode layer are respectively transmitted to the next Encode layer and the AG+ module at the same level, and the results processed by each Decode layer are respectively transmitted to the next Decode layer and the AG+ module at the next level. The last Encode layer is connected to a convolution module and an ASPP module in sequence, and then connected to the AG+ module at the same level and the first Decode layer.

5. The nuclear containment crack detection method based on deep learning according to claim 4 is characterized in that: The ASPP module consists of a 1×1 convolution, three dilated convolutions with different dilation rates, and an image pooling module. Image Pooling includes an average pooling, a 1×1 convolution Conv, and an upsampling layer Upsample. The dilated convolution uses dilated rates of 6, 12, and 18, respectively. The ASPP module finally merges the results of multiple dilated convolutions and pooling processes and inputs them into a 1×1 convolution layer to obtain the output.

6. The nuclear containment crack detection method based on deep learning according to claim 4, characterized in that: The AG+ module adopts a parallel structure with two inputs and one output, one of which is the output feature x from the same encoding layer. l , the second input is the input feature g from the decoding layer, one The branch is the input feature g of the decoding layer and the output feature x of the encoding layer l Input a 1×1 convolution respectively, then add the results and input them into the ReLU activation function, and then pass through a 1×1 convolution layer and a Sigmoid activation function to obtain the additive attention coefficient; The other branch is to combine the input feature g of the decoding layer and the output feature x of the encoding layer l Input a 1×1 convolution respectively, then multiply the results and input the result into the ReLU activation function, and then pass through a 1×1 convolution layer and a Sigmoid activation function to obtain the dot product attention coefficient; Finally, the results of the two branches are added together and then added to the output feature x of the encoding layer. l Multiply to enhance the crack information of interest.

7. The nuclear containment crack detection method based on deep learning according to claim 1, 2, 3, 4, 5, or 6, characterized in that: The training of the containment crack identification network model is implemented by transfer learning. The training process of the containment crack segmentation network model adopts the Adam optimization algorithm, and the polynomial StepLR is used as the learning rate scheduler to prevent model overfitting.

8. A nuclear containment crack detection system based on deep learning, characterized by: Used to implement a nuclear containment crack detection method based on deep learning as described in any one of claims 1-7.

9. The deep learning-based nuclear containment crack detection system according to claim 8, characterized in that: It includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a nuclear containment crack detection method based on deep learning as described in any one of claims 1 to 7.

10. The nuclear containment crack detection system based on deep learning according to claim 8, characterized in that: It includes a readable storage medium, on which a computer program is stored. When the computer program is executed, a nuclear containment crack detection method based on deep learning as described in any one of claims 1 to 7 is implemented.

Images