Nuclear power equipment defect detection method based on Mixed-FFM model

Abstract

A nuclear power equipment defect detection method based on a Mixed-FFM model belongs to the technical field of nuclear power equipment fault detection, and comprises the following steps: firstly, based on a self-attention and convolution feature fusion module, constructing a hybrid encoder, and extracting three scale feature quantities of input image data; constructing a decoder to carry out fusion, deconvolution, batch normalization and linearization processing on the three scale characteristic quantities to obtain an initial Mixed-FFM model; constructing a training set and a verification set based on the X-ray image data set of the nuclear power equipment, and training and calibrating the Mixed-FFM model to obtain an optimal Mixed-FFM model; and inputting the X-ray image data set of the nuclear power equipment to be detected into the optimal Mixed-FFM model to obtain a defect diagnosis result. Compared with the prior art, the method based on deep learning can efficiently and accurately extract features from a large amount of data, has strong adaptability and generalization ability, has better robustness for complex environments and noise, realizes rapid detection of equipment defect areas, improves defect detection efficiency, and reduces equipment maintenance cost.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear power equipment fault detection, and particularly relates to a method for detecting defects in nuclear power equipment based on a Mixed-FFM model. Background Art

[0002] Nuclear power equipment is a highly complex technical system, and its normal operation is crucial for energy supply. However, due to long-term operation and harsh environmental conditions such as high temperature and high pressure, nuclear power equipment is prone to being affected by various external factors, resulting in potential defects and failures. If these defects cannot be detected and repaired in time, they may have a serious impact on the safety and reliability of the equipment, and even lead to serious accidents.

[0003] In order to improve the safety and reliability of nuclear power equipment, fast and effective defect detection technology is crucial. In the field of nuclear power equipment defect detection, traditional algorithms are mainly used for detection at present. These traditional methods often have low accuracy and are very sensitive to noise and interference when dealing with complex and changeable environments, and it is difficult to adapt to the challenges of diverse defects. Summary of the Invention

[0004] In order to solve the limitations of traditional algorithms in extracting fault features, the present invention provides a method for detecting defects in nuclear power equipment based on a Mixed-FFM model.

[0005] A method for detecting defects in nuclear power equipment based on a Mixed-FFM model, the method for detecting defects in nuclear power equipment includes:

[0006] Based on a self-attention and convolutional feature fusion module, construct a hybrid encoder to extract three-scale feature quantities of the input image data; construct a decoder to perform fusion, deconvolution, batch normalization, and linearization processing on the three-scale feature quantities to obtain an initial Mixed-FFM model;

[0007] Based on the X-ray image dataset of nuclear power equipment, construct a training set and a validation set, and train and calibrate the Mixed-FFM model to obtain an optimal Mixed-FFM model;

[0008] Input the X-ray image dataset of the nuclear power equipment to be detected into the optimal Mixed-FFM model to obtain a defect diagnosis result.

[0009] Preferably, the step of constructing a hybrid encoder based on a self-attention and convolutional feature fusion module to extract three-scale feature quantities of the input image data includes:

[0010] Based on a self-attention and convolutional feature fusion module, construct a large-scale encoding branch to perform feature extraction on the whole of the input image to obtain large feature quantities;

[0011] Based on the self-attention and convolution feature fusion module, a medium-scale encoding branch is constructed to extract features from any sub-region of 50% of the width and height of the input image, obtaining medium feature quantities;

[0012] Based on the self-attention and convolution feature fusion module, a small-scale encoding branch is constructed to extract features from any sub-region of 25% of the width and height of the input image, obtaining small feature quantities.

[0013] Preferably, the large-scale encoding branch is constructed based on the self-attention and convolution feature fusion module to extract features from the whole of the input image, obtaining large feature quantities, including:

[0014] Stack a set number of the self-attention and convolution feature fusion modules, use the output of the previous self-attention and convolution feature fusion module as the input of the current self-attention and convolution feature fusion module, and the stacked self-attention and convolution feature fusion modules perform feature extraction through a single multi-layer perceptron, and construct a large-scale encoding branch;

[0015] The self-attention sub-model in each self-attention and convolution feature fusion module performs global feature extraction on the whole of the input image; the convolution sub-model in each self-attention and convolution feature fusion module performs local feature extraction on the whole of the input image and performs feature extraction through the multi-layer perceptron, obtaining large feature quantities.

[0016] Preferably, the large feature quantity is obtained by the following formula:

[0017] X4 = MLP(X3)

[0018] In the formula, X4 is the large feature quantity, MLP(·) represents the multi-layer perceptron, and X3 is the self-attention and convolution fusion feature;

[0019] The self-attention and convolution fusion feature X3 is obtained by the following formula:

[0020] X3 = Linear([X1, X2])

[0021] In the formula, Linear(·) represents the linear layer, X1 is the convolution feature obtained by the convolution sub-model, and X2 is the weighted fusion feature obtained by the self-attention sub-model;

[0022] The convolution feature X1 is obtained by the following formula:

[0023] X1 = Convs(Q)

[0024] In the formula, Convs(·) is the stacked convolution layer, and Q is the Q-dimensional feature mapped from the input image;

[0025] The weighted fusion feature X2 is obtained by the following formula:

[0026]

[0027] where K is the K-dimensional feature mapped from the input image; V is the V-dimensional feature mapped from the input image; d is the number of channels of the Q-dimensional feature; and T represents the transpose operation.

[0028] Preferably, before stacking a set number of the self-attention and convolution feature fusion modules, it further includes:

[0029] Construct three parallel linear layers to map the input feature map into three sets of features;

[0030] Add a self-attention module and a convolution module after the linear layer to obtain the self-attention and convolution feature fusion module;

[0031] The self-attention module is a single self-attention layer, and the convolution module is stacked by three convolution layers with a convolution kernel size of 3, a stride of 1, and a padding number of 1.

[0032] Preferably, the construction of the decoder for fusing, deconvolving, batch normalizing, and linearizing three-scale feature quantities includes:

[0033] Based on a single linear layer, a single convolution layer, and a single multi-layer perceptron, construct a multi-scale feature fusion module;

[0034] The three-scale feature quantities are fused in the channel dimension through the linear layer, fused in the spatial scale through the convolution layer, and feature extraction is performed through the multi-layer perceptron to obtain the fused feature quantity;

[0035] Based on the fused feature quantity, perform deconvolution, batch normalization, and linearization processing.

[0036] Preferably, the fused feature quantity is obtained by the following formula:

[0037] X e = MLP(Conv(Linear([X l , X m , X s )))

[0038] where MLP(·) represents the multi-layer perceptron, Linear(·) represents the linear layer, and Conv(·) represents the single convolution layer.

[0039] Preferably, based on the fused feature quantity, the deconvolution, batch normalization, and linearization processing are performed as shown in the following formula:

[0040]

[0041] Among them, Linear(·) represents the linear layer, Norm(·) represents the batch normalization layer, and DConv(·) represents the single-layer transposed convolution layer.

[0042] Preferably, based on the nuclear power equipment X-ray image dataset, a training set and a validation set are constructed, and the Mixed-FFM model is trained and calibrated to obtain the optimal Mixed-FFM model, including:

[0043] Adjust the pixels and grayscale of the nuclear power equipment X-ray image dataset to the set values, and generate paired label images through manual annotation, and construct the training set and the validation set according to the set ratio;

[0044] Based on the training set, train the Mixed-FFM model by the cross-entropy and intersection over union loss methods, and input the validation set to test the Mixed-FFM model.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] The technical solution provided by the present invention includes: First, based on the self-attention and convolutional feature fusion module, a hybrid encoder is constructed to extract three-scale feature quantities of the input image data; a decoder is constructed to fuse, transpose-convolve, batch-normalize, and linearize the three-scale feature quantities to obtain the initial Mixed-FFM model; based on the nuclear power equipment X-ray image dataset, a training set and a validation set are constructed, and the Mixed-FFM model is trained and calibrated to obtain the optimal Mixed-FFM model; the X-ray image dataset of the nuclear power equipment to be detected is input into the optimal Mixed-FFM model to obtain the defect diagnosis result. In contrast, the deep learning-based method can efficiently and accurately extract features from a large amount of data, has strong self-adaptability and generalization ability, has better robustness to complex environments and noises, realizes the rapid detection of the defective areas of the equipment, improves the defect detection efficiency, and reduces the equipment maintenance cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is the flowchart of the nuclear power equipment defect detection method based on the Mixed-FFM model in the present invention;

[0048] Figure 2 is the schematic diagram of the Mixed-FFM model framework in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To better understand the present invention, the content of the present invention will be further described below in conjunction with the specification drawings and examples.

[0050] Embodiment 1

[0051] This embodiment provides a nuclear power equipment defect detection method based on the Mixed-FFM model. The nuclear power equipment defect detection method includes:

[0052] Step 1: Based on the self-attention and convolutional feature fusion module, construct a hybrid encoder to extract three-scale feature quantities of the input image data; construct a decoder to fuse, deconvolve, batch-normalize, and linearize the three-scale feature quantities to obtain an initial Mixed-FFM model;

[0053] Step 2: Based on the nuclear power equipment X-ray image dataset, construct a training set and a validation set, and train and calibrate the Mixed-FFM model to obtain an optimal Mixed-FFM model;

[0054] Step 3: Input the X-ray image dataset of the nuclear power equipment to be detected into the optimal Mixed-FFM model to obtain a defect diagnosis result.

[0055] In Step 1, the construction of the hybrid encoder based on the self-attention and convolutional feature fusion module to extract three-scale feature quantities of the input image data includes:

[0056] Based on the self-attention and convolutional feature fusion module, construct a large-scale encoding branch to extract features of the whole input image to obtain large feature quantities;

[0057] Based on the self-attention and convolutional feature fusion module, construct a medium-scale encoding branch to extract features of any sub-region with 50% of the width and height of the input image to obtain medium feature quantities;

[0058] Based on the self-attention and convolutional feature fusion module, construct a small-scale encoding branch to extract features of any sub-region with 25% of the width and height of the input image to obtain small feature quantities.

[0059] Taking the extraction of large feature quantities as an example, the construction of the large-scale encoding branch based on the self-attention and convolutional feature fusion module to extract features of the whole input image to obtain large feature quantities includes:

[0060] Stack a set number of the self-attention and convolutional feature fusion modules, use the output of the previous self-attention and convolutional feature fusion module as the input of the current self-attention and convolutional feature fusion module, and the stacked self-attention and convolutional feature fusion modules perform feature extraction through a single multi-layer perceptron and construct a large-scale encoding branch;

[0061] Each self-attention sub-model in the self-attention and convolution feature fusion module extracts global features of the entire input image; each convolution sub-model in the self-attention and convolution feature fusion module extracts local features of the entire input image and performs feature extraction through the multi-layer perceptron to obtain a large amount of features.

[0062] The large amount of features is obtained through the following formula:

[0063] X4 = MLP(X3)

[0064] In the formula, X4 is the large amount of features, MLP(·) represents the multi-layer perceptron, and X3 is the self-attention and convolution fusion feature;

[0065] The self-attention and convolution fusion feature X3 is obtained through the following formula:

[0066] X3 = Linear([X1, X2])

[0067] In the formula, Linear(·) represents the linear layer, X1 is the convolution feature obtained by the convolution sub-model, and X2 is the weighted fusion feature obtained by the self-attention sub-model;

[0068] The convolution feature X1 is obtained through the following formula:

[0069] X1 = Convs(Q)

[0070] In the formula, Convs(·) is a stacked convolution layer, and Q is the Q-dimensional feature mapped from the input image;

[0071] The weighted fusion feature X2 is obtained through the following formula:

[0072]

[0073] In the formula, K is the K-dimensional feature mapped from the input image; V is the V-dimensional feature mapped from the input image; d is the number of channels of the Q-dimensional feature; T is the transpose operation.

[0074] In addition, before stacking a set number of the self-attention and convolution feature fusion modules, it further includes:

[0075] Construct three parallel linear layers to map the input feature map into three groups of features;

[0076] Add a self-attention module and a convolution module after the linear layer to obtain the self-attention and convolution feature fusion module;

[0077] The self-attention module is a single self-attention layer, and the convolution module is stacked by three convolution layers with a convolution kernel size of 3, a stride of 1, and a padding number of 1.

[0078] In step one, the constructed decoder performs fusion, deconvolution, batch normalization, and linearization processing on three scale feature quantities, including:

[0079] Based on a single linear layer, a single convolutional layer, and a single multi-layer perceptron, a multi-scale feature fusion module is constructed;

[0080] The three scale feature quantities are fused in the channel dimension through the linear layer, fused in the spatial scale through the convolutional layer, and feature extraction is performed through the multi-layer perceptron to obtain a fused feature quantity;

[0081] Based on the fused feature quantity, deconvolution, batch normalization, and linearization processing are performed.

[0082] The fused feature quantity is obtained through the following formula:

[0083] X e = MLP(Conv(Linear([X l , X m , C s )))

[0084] where MLP(·) represents a multi-layer perceptron, Linear(·) represents a linear layer, and Conv(·) represents a single convolutional layer.

[0085] Based on the fused feature quantity, deconvolution, batch normalization, and linearization processing are performed, as shown in the following formula:

[0086]

[0087] where Linear(·) represents a linear layer, Norm(·) represents a batch normalization layer, and DConv(·) represents a single deconvolution layer.

[0088] In step two, based on the nuclear power equipment X-ray image dataset, a training set and a validation set are constructed, and the Mixed-FFM model is trained and calibrated to obtain an optimal Mixed-FFM model, including:

[0089] The nuclear power equipment X-ray image dataset is adjusted in pixels and grayscale to set values, and paired label images are generated through manual annotation, and a training set and a validation set are constructed according to a set ratio;

[0090] Based on the training set, the Mixed-FFM model is trained by the cross-entropy and intersection over union loss methods, and the validation set is input to test the Mixed-FFM model.

[0091] Embodiment Two

[0092] Based on the same inventive concept, this embodiment provides a method for defect detection of nuclear power equipment based on the Mixed-FFM model. The specific implementation steps are as follows:

[0093] As Figure 1 shown, the present invention first processes the collected X-ray images of nuclear power equipment into a common image format through image preprocessing, and forms a training set and a validation set after annotation. Subsequently, the Mixed-FFM model is trained on the formed X-ray image dataset of the defect area of nuclear power equipment. Finally, the trained defect detection model provides assistance for defect detection of nuclear power equipment.

[0094] The method for defect detection of nuclear power equipment based on the Mixed-FFM model of the present invention includes the following steps:

[0095] 1) Process the collected X-ray images of nuclear power equipment into a common image format, that is, perform gray-scale adjustment on the X-ray image I 1 and use linear transformation to map the numerical range of the X-ray image to 0 to 255 to make full use of the information contained in the image:

[0096]

[0097] wherein, I 1 is the original X-ray image, I 2 is the converted X-ray gray-scale image, min(·) is the operation of taking the global minimum value, and max(·) is the operation of taking the global maximum value.

[0098] 2) Manually annotate the defect areas of the preprocessed X-ray gray-scale image I 2 by professional inspectors, generate the corresponding label image Y according to the manual annotation, and randomly distribute all X-ray images and the corresponding label images to the training set (X train , Y train ) and the validation set (X eval , Y eval ) in a ratio of 8:2 to form the X-ray image dataset required for this task.

[0099] 3) Construction of the Mixed-FFM model. The schematic diagram of the model framework is as Figure 2As shown in the figure, the Mixed-FFM model constructed by the present invention mainly consists of two parts: a hybrid encoder and a decoder. The hybrid encoder is composed of three branches focusing on different feature scales, and each branch is stacked by a self-attention and convolutional feature fusion module. The hybrid encoder extracts global features through self-attention, extracts local features through convolution, and obtains image representations with different granularities through three branches of different feature scales. The feature outputs of the three branches of the hybrid encoder are used as the input of the decoder. The decoder fuses and deconvolves the rich features extracted by the hybrid encoder through a multi-scale feature fusion module and a stacked feature decoding module, continuously restores the spatial information of the image, and finally obtains the detection result of the defect area in the X-ray image.

[0100] 3-1) The described hybrid encoder is composed of three parallel branches of large, medium, and small feature scales. Among the three branches, the large feature scale branch processes the complete image, and the medium and small feature scale branches receive the intercepted local images. Among them, the medium feature scale branch uses any sub-region with 50% of the original image width and height, and the small feature scale branch uses any sub-region with 25% of the original image width and height. Each of the three branches is stacked by a self-attention and convolutional feature fusion module. The features X l 、X m 、X S extracted at multiple feature scales by the three branches will be fused and decoded by the decoder. The three branches contain extensive feature connections, including residual connections within the same branch (X3 = X2 + X1), where X1 is generally the input feature of the current module, X2 is the output feature of the current module, and X3 is the output feature after adding the input feature, that is, the result of the residual connection), and feature fusion between different branches. The residual connection rule within the same branch is to connect the head and tail of every three modules once to alleviate the problem of abnormal gradients during training. The feature fusion between different branches is to more fully fuse the features of different feature scales. The connection rule is that the information of the large feature scale branch flows to the medium and small feature scale branches, and the information of the medium feature scale branch flows to the small feature scale branch. Among them, the feature information is circulated once every 6 modules. The large, medium, and small feature scale branches respectively contain 18, 12, and 6 self-attention and convolutional feature fusion modules.

[0101] 3-2) The self-attention and convolution feature fusion module described above is a fusion of a conventional self-attention module and a conventional convolution module. The first layer of the module is three parallel linear layers that unify the number of channels of the feature map and simultaneously map the input feature map into three groups of features, namely Q, K, and V. Subsequently, there are a convolution branch and a self-attention branch. The convolution branch is composed of three stacked convolution layers with a convolution kernel size of 3, a stride of 1, and a padding number of 1. The convolution layer processes the feature Q to obtain the feature X1. The self-attention branch consists of a single conventional self-attention layer, that is, calculates the similarity matrix of Q and K, and weights the feature V after scaling by the number of channel dimensions to obtain the weighted fusion feature X2. Subsequently, the convolution feature X1 and the self-attention feature X2 are concatenated in the channel dimension and fused through a single linear layer to obtain the self-attention and convolution fusion feature X3. Finally, a single multi-layer perceptron module extracts features from the fusion feature X3 to obtain the output feature X4 of the module. Among them, the format of the feature needs to be adjusted when entering and leaving the self-attention branch. When entering, it needs to be adjusted from D×H×W to L×D, and when leaving, it needs to be adjusted back from L×D to D×H×W, where L = H×W. The above process can be described by the formula as follows:

[0102] X4 = MLP(X3),

[0103] X3 = Linear([X1, X2]),

[0104] X1 = Convs(Q),

[0105]

[0106] Among them, MLP(·) represents the multi-layer perceptron, Linear(·) represents the linear layer, and Convs(·) represents the stacked convolution layer.

[0107] 3-3) The decoder described above is composed of a single multi-scale feature fusion module and four stacked feature decoding modules. The decoder receives the output features from the three branches of the hybrid encoder. First, it fuses the features of different feature scales from the three branches through a single multi-scale feature module, and then restores the spatial information of the image through the stacked four feature decoding modules to obtain the segmentation result of the defect area.

[0108] 3-4) The multi-scale feature fusion module includes a single linear layer, a single convolution layer, and a single multi-layer perceptron. The features X l , X m , X s output by the three feature scale branches of the hybrid encoder are concatenated in the channel dimension and fused in the channel dimension through the linear layer, then fused in the spatial scale through a convolution layer with a convolution kernel size of 5, a stride of 1, and a padding number of 2, and finally processed by the multi-layer perceptron to obtain the feature X eThis process can be expressed by the formula:

[0109] X e = MLP(Conv(Linear([X l , X m , X s )))

[0110] where MLP(·) represents a multi-layer perceptron, Linear(·) represents a linear layer, and Conv(·) represents a single-layer convolutional layer.

[0111] The feature decoding module described in 3-5) consists of a transposed convolutional layer, a batch normalization layer, and a linear layer. The transposed convolutional layer reconstructs the spatial information of the feature map, the batch normalization layer is used to improve the convergence speed, and the linear layer weights the features in the channel dimension. This process can be described by the formula:

[0112]

[0113] where Linear(·) represents a linear layer, Norm(·) represents a batch normalization layer, and DConv(·) represents a single-layer transposed convolutional layer.

[0114] 4) Train the Mixed-FFM model using cross-entropy and intersection over union (IoU) losses. Cross-entropy is the most commonly used deep learning loss function and is used as the main loss to ensure that the model can converge within the expected time. The IoU loss is suitable for segmentation tasks such as defect area detection and can improve the convergence speed of the model. The final loss function adopted is the combination of cross-entropy and IoU, and the formula is expressed as:

[0115]

[0116] where CE(·) and IoU(·) represent cross-entropy and IoU losses respectively, and α and β are the weight coefficients when the two losses are added. During training, α can be set to 1 and β can also be set to 1.

[0117] 5) When the trained Mixed-FFM model is actually applied to the task of detecting defects in X-ray images of nuclear power equipment, first perform the same data preprocessing operations on the X-ray images to be processed as in step 1), then input the processed images into the trained Mixed-FFM model, and then obtain the probability distribution map for defect area detection. Take the channel layer where the maximum value is located at each pixel position in the defect area probability map output by the decoder to obtain a binary map indicating the defect area. This binary map can be used to assist professional defect detection personnel in implementing defect detection tasks in actual nuclear power defect detection scenarios. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

Claims

1. A method for defect detection of nuclear power equipment based on the Mixed-FFM model, characterized in that, The nuclear power equipment defect detection method includes: Based on the self-attention and convolutional feature fusion module, a hybrid encoder is constructed to extract three-scale feature quantities of the input image data; a decoder is constructed to fuse, deconvolve, batch-normalize, and linearize the three-scale feature quantities to obtain the initial Mixed-FFM model; Based on the nuclear power equipment X-ray image dataset, a training set and a validation set are constructed, and the Mixed-FFM model is trained and calibrated to obtain the optimal Mixed-FFM model; The X-ray image dataset of the nuclear power equipment to be detected is input into the optimal Mixed-FFM model to obtain the defect diagnosis result.

2. The method for defect detection of nuclear power equipment based on the Mixed-FFM model according to claim 1, wherein The construction of a hybrid encoder based on the self-attention and convolutional feature fusion module to extract three-scale feature quantities of the input image data includes: Based on the self-attention and convolutional feature fusion module, a large-scale encoding branch is constructed to extract features of the whole input image to obtain large feature quantities; Based on the self-attention and convolutional feature fusion module, a medium-scale encoding branch is constructed to extract features of any sub-region with 50% of the width and height of the input image to obtain medium feature quantities; Based on the self-attention and convolutional feature fusion module, a small-scale encoding branch is constructed to extract features of any sub-region with 25% of the width and height of the input image to obtain small feature quantities.

3. The method for defect detection of nuclear power equipment based on the Mixed-FFM model according to claim 2, characterized in that, The construction of a large-scale encoding branch based on the self-attention and convolutional feature fusion module to extract features of the whole input image to obtain large feature quantities includes: A set number of the self-attention and convolutional feature fusion modules are stacked, and the output of the previous self-attention and convolutional feature fusion module is used as the input of the current self-attention and convolutional feature fusion module. The stacked self-attention and convolutional feature fusion modules perform feature extraction through a single multi-layer perceptron, and a large-scale encoding branch is constructed; The self-attention sub-model in each self-attention and convolutional feature fusion module performs global feature extraction on the whole input image; the convolutional sub-model in each self-attention and convolutional feature fusion module performs local feature extraction on the whole input image and performs feature extraction through the multi-layer perceptron to obtain large feature quantities.

4. The method for detecting defects of nuclear power equipment based on the Mixed-FFM model according to claim 3, characterized in that, The large feature quantities are obtained through the following formula: X4 = MLP(X3) In the formula, X4 is the large feature quantity, MLP(·) represents the multi-layer perceptron, and X3 is the self-attention and convolutional fusion feature; The self-attention and convolutional fusion feature X3 is obtained through the following formula: X3 = Linear([X1, X2]) In the formula, Linear(·) represents the linear layer, X1 is the convolutional feature obtained by the convolutional sub-model, and X2 is the weighted fusion feature obtained by the self-attention sub-model; The convolutional feature X1 is obtained through the following formula: X1 = Convs(Q) In the formula, Convs(·) is the stacked convolutional layer, and Q is the Q-dimension feature mapped from the input image; The weighted fusion feature X2 is obtained through the following formula: Where K is the K - dimensional feature mapped from the input image; V is the V - dimensional feature mapped from the input image; d is the number of channels of the Q - dimensional feature; and T is the transpose operation.

5. The method for detecting defects in nuclear power equipment based on the Mixed-FFM model according to claim 3, wherein, Before stacking a set number of the self - attention and convolution feature fusion modules, it further includes: Construct three parallel linear layers to map the input feature map into three groups of features; Add a self - attention module and a convolution module after the linear layer to obtain the self - attention and convolution feature fusion module; The self - attention module is a single self - attention layer, and the convolution module is stacked by three convolution layers with a convolution kernel size of 3, a stride of 1, and a padding number of 1.

6. The method for detecting defects in nuclear power equipment based on the Mixed-FFM model according to claim 1, characterized in that, The construction of the decoder for fusing, de - convolving, batch - normalizing, and linearizing three - scale feature quantities includes: Based on a single linear layer, a single convolution layer, and a single multi - layer perceptron, construct a multi - scale feature fusion module; The three - scale feature quantities are fused in the channel dimension through the linear layer, fused in the spatial scale through the convolution layer, and feature extraction is performed through the multi - layer perceptron to obtain the fused feature quantity; Based on the fused feature quantity, perform de - convolution, batch - normalization, and linearization processing.

7. The method for defect detection of nuclear power equipment based on the Mixed-FFM model according to claim 6, wherein The fused feature quantity is obtained by the following formula: X e = MLP(Conv(Linear([X l , X m , X s ))) Where MLP(·) represents the multi - layer perceptron, Linear(·) represents the linear layer, and Conv(·) represents the single - layer convolution layer.

8. The method for detecting defects in nuclear power equipment based on the Mixed-FFM model according to claim 6, characterized in that Based on the fused feature quantity, perform de - convolution, batch - normalization, and linearization processing, as shown in the following formula: Where Linear(·) represents the linear layer, Norm(·) represents the batch - normalization layer, and DConv(·) represents the single - layer de - convolution layer.

9. The method for defect detection of nuclear power equipment based on the Mixed-FFM model according to claim 1, wherein, Based on the nuclear power equipment X - ray image dataset, construct a training set and a validation set, and train and calibrate the Mixed - FFM model to obtain the optimal Mixed - FFM model, including: Adjust the pixels and grayscale of the nuclear power equipment X - ray image dataset to set values, and generate paired label images through manual annotation, and construct a training set and a validation set according to a set ratio; Based on the training set, train the Mixed - FFM model by the cross - entropy and intersection - over - union loss methods, and input the validation set to test the Mixed - FFM model.

Images