A nuclear power plant equipment appearance defect self-learning identification method based on feature probability distribution

Abstract

The application belongs to the technical field of target detection and self-learning, and particularly relates to a nuclear power plant equipment appearance defect self-learning identification method based on feature probability distribution. The method comprises the following steps: collecting defect appearance images of key equipment of a nuclear power plant; constructing a nuclear power plant equipment defect detection data set; extracting potential features of equipment appearance defects; capturing global features of defect appearances; generating feature queries; continuously adapting to new defect features in the training process; after reasoning is completed, a target classification head and a regression head will output category information and corresponding boundary box coordinates of each detection target; and for category prediction, learned category probability is multiplied by classification probability to generate final category prediction. The method has the beneficial effects that by combining a regional attention mechanism and a dynamic multivariate Gaussian distribution, the method effectively solves the problems of category imbalance and complex defect appearances of existing target detection algorithms in complex background environments.

Description

Technical Field

[0001] This invention belongs to the field of target detection and self-learning technology, specifically relating to a self-learning identification method for appearance defects in nuclear power plant equipment based on feature probability distribution. Background Technology

[0002] With the continued growth of global energy demand, nuclear power, as an efficient and low-carbon energy source, is gaining increasing attention from more and more countries. In particular, the safety requirements for nuclear power plant equipment are crucial; the reliability, operating status, and appearance quality of nuclear power plant equipment directly affect the safe operation and long-term stability of the plant. Therefore, how to effectively inspect nuclear power plant equipment, especially identifying appearance defects, has become an important research topic in nuclear power plant maintenance and repair.

[0003] Nuclear power plants contain a wide variety of equipment, including reactors, pressure vessels, steam generators, pumps, valves, and piping systems. This equipment operates under harsh environments with high temperatures, high pressures, and strong radiation, making it prone to surface defects such as damage, corrosion, and cracks. Failure to detect and repair these defects in a timely manner can lead to equipment failure or even major safety accidents. Therefore, the ability to quickly and accurately identify surface defects in nuclear power plant equipment, especially during visual inspections, has become a crucial issue for improving the safety of nuclear power plant equipment and ensuring the safe operation of nuclear power plants.

[0004] Traditional manual inspection methods have limitations such as low efficiency, high cost, and strong subjectivity. They also cannot be used for real-time monitoring in harsh environments, and there is a risk of missed detections and misjudgments. They are difficult to meet the requirements of large-scale and high-frequency equipment monitoring in modern power plants.

[0005] In summary, current image data on external defects in nuclear power plant equipment suffers from limited quantity and coverage of defect types, severely restricting the performance and practicality of defect detection algorithms. Furthermore, during data collection at nuclear power plant sites, the diverse types of equipment and faults make it difficult to obtain high-quality and sufficient defect data, further impacting model training effectiveness and convergence speed. Additionally, the following issues also exist: The defect data for nuclear power plant equipment suffers from incomplete coverage of defect types and an imbalance in categories.

[0006] Nuclear power equipment typically consists of multiple interdependent systems, such as the reactor, cooling system, and electrical system. A failure in any of these systems can trigger a chain reaction. Furthermore, nuclear power plants operate in complex environments characterized by high temperatures, high pressures, and strong radiation, resulting in a wide variety of defect types that are difficult to predict accurately, and new defects are constantly emerging.

[0007] Unlike object detection for known objects, unknown defect categories are not labeled. Therefore, when training object detection models, features of objects including unknown objects will be incorrectly penalized as background, leading to poor performance of existing detection methods. Summary of the Invention

[0008] The purpose of this invention is to provide a self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution. This method can solve the problems of data imbalance, diverse defect types, and insufficient self-learning ability in the detection of appearance defects of nuclear power plant equipment. By combining the regional attention mechanism with self-learning ability, it significantly improves the detection accuracy and adaptability to new types of defects.

[0009] The technical solution of the present invention is as follows: A self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution, comprising the following steps: Step 1: Collect images of defects in key equipment of the nuclear power plant; Step 2: Annotate the target bounding boxes of the acquired images, annotate the outer bounding boxes of the defect areas, and construct a nuclear power equipment defect detection dataset; Step 3: Use image data augmentation methods to process the images of nuclear power equipment. The augmented image data is then used for feature extraction through the backbone network of the network to extract potential features of equipment appearance defects. Step 4: Use the backbone network and neck network of YOLOv12 as the backbone network and neck network part of the model, and use the region attention mechanism of YOLOv12 to optimize the computational efficiency and capture the global features of the defect appearance. Step 5: Using a ViT-based encoder-decoder structure, generate feature queries from the feature maps extracted by YOLOv12; Step 6: The class-aware detection head and the decoder part of the ViT network were jointly optimized along the probability density model as image features. By transforming the features of the target object into a multivariate Gaussian distribution in the query embedding space and dynamically updating the mean and variance of the distribution, the model can continuously adapt to new defect features during training. Step 7: After inference is completed, the target classification head and regression head will output the category information of each detected target and its corresponding bounding box coordinates. For category prediction, the learned category probability is multiplied by the classification probability to produce the final category prediction.

[0010] In step 2, the image data is labeled. For each image, the bounding box is first labeled to delineate the maximum boundary of the defect area and label the defect category. The information corresponding to each labeled box includes: the position of the box and the target class information. Each labeled box corresponds to a category label. The model is trained by comparing the original image with the labeled image, so that the model can learn the spatial distribution and features of the defect area.

[0011] In step 3, various image enhancement strategies are adopted to improve data quality. By transforming the original image in different ways, data diversity is enhanced, overfitting is reduced, and the model's adaptability to unknown defect types is improved.

[0012] In step 4, a paradigm shift is achieved by integrating a region attention-centric mechanism, a streamlined architecture design, and an optimized training pipeline. This leverages the remaining efficient layers to aggregate the network R-ELAN, FlashAttention, and 7×7 separable convolutions to provide throughput and accuracy.

[0013] Step 5, which involves converting the feature map into a query, is achieved through a linear transformation of the feature map. Each position in the feature map is mapped to a specific query vector, and then positional encoding is performed to generate the query vector. These query vectors are input to the encoder part of ViT, becoming the input of the ViT model, used to aggregate important information from the feature map. Each query vector corresponds to a specific region in the input image. The ViT encoder-decoder structure, through an iterative interaction process, is defined as follows: , in, Q It is a query vector. K It is a key vector. V It is a value vector. It is the dimension of the key vector.

[0014] In step 7, during the prediction of the target category, a decoupled prediction method for object content and object category is adopted. By learning the content and category of the defect separately, the model's adaptability to new defect types is improved. In each layer of the ViT encoder, the encoder aggregates information through a self-attention mechanism. In the decoder, the information processed by the encoder is decoded through multi-layer interactive iteration to gradually refine the aggregated feature representation. In each layer, the decoder interacts with the information received from the encoder and its own initial query vector to further optimize and extract the features of the defect region.

[0015] Step 6 employs a self-learning target detection method based on feature probability distribution, which includes combining a head based on feature probability distribution with ViT to improve the YOLOv12 model. The self-learning detection method based on feature probability distribution first extracts features through the backbone network and neck network, then generates a query containing potential target defect regions. Next, the encoder-decoder structure of ViT performs deep iterative interaction on the feature query. Finally, the target class detection result is obtained through the collaboration of the target classification head and the target regression head, while retaining the detection results of unknown classes. The unknown class detection head retains the detection results of unknown classes, and the unknown target bounding box optimization part processes redundant bounding boxes to obtain the final prediction result of the unknown class.

[0016] Step 6 employs decoupled prediction of object content and object category, by learning the probability of defect targets and defect categories separately. , in, It belongs to a certain category when all possible defects exist. l The probability, This indicates that given a certain query q, it belongs to the defect category. l The probability, It represents the probability that the target defect exists in query q. When the object does not exist, the conditional probability of the object category is 0. That is, when the model determines that the current query area is background, the classification head will not assign any target defect category. The model's prediction process is defined as follows: , in, This indicates the determination of the target category in the query. When the query represents the background, the category-aware detection head can predict a very low probability and will not assign any target defect category. When the query contains a target defect, the category-aware detection head will predict a probability value close to 1, and then the classification head will classify and identify the target object. This represents the probability of a specific category.

[0017] In step 6, the query embedding is generated through multi-layer interactive iteration. The generated query embedding is used to estimate the target probability and category of each target region. Through the category-aware head, the model predicts whether each query embedding contains an object, that is, whether the region is a target defect region. The category prediction determines whether the region needs further classification by calculating the probability of each query. For each query embedding predicted as a target defect, the category-aware head is used to classify and predict whether it belongs to a known category or an unknown category.

[0018] The beneficial effects of this invention are as follows: it can efficiently and accurately detect different defects in nuclear power plant equipment and has a certain degree of adaptability to unknown defects. By combining a region attention mechanism and a dynamic multivariate Gaussian distribution, this invention effectively solves the problems of class imbalance and complex defect appearances that occur in existing target detection algorithms in complex background environments.

[0019] Regarding data acquisition, this invention employs diverse image acquisition enhancement techniques, taking into account various factors such as different lighting conditions, viewing angles, distances, and defect types, making the dataset more representative and diverse. To further enhance data diversity, this invention utilizes multiple data augmentation strategies, thereby improving the quality of the training data.

[0020] Furthermore, this invention employs an improved YOLOv12 model based on an attention mechanism, innovatively integrating the ViT module and category-aware head into the YOLOv12 network structure. This allows the model to automatically focus on defect regions during training, thereby improving its ability to extract detailed features. During the model training phase, by comparing the improved model with traditional methods, this invention verifies its advantages in self-learning detection of unknown defects. Experimental results show that the improved model achieves a significant improvement in the average accuracy of detecting both known and unknown defects. Compared to traditional methods, it is more accurate in detecting known categories and has stronger adaptability to unknown categories. Attached Figure Description

[0021] Figure 1 A flowchart of the overall process for defect detection in nuclear power equipment; Figure 2 This is a diagram of a self-learning network structure. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] A self-learning method for identifying appearance defects in nuclear power plant equipment based on feature probability distribution includes the following steps: Step 1: Collect visual images of defects in key equipment of the nuclear power plant (including pipes, valves, containers, pumps, etc.). During the data acquisition process, various lighting conditions, shooting angles, material types, and defect types should be covered to ensure data diversity and representativeness. Step 2: Annotate the target bounding boxes of the acquired images, annotate the outer bounding boxes of the defect areas, and construct a nuclear power equipment defect detection dataset; In step 2, the image data is labeled, and the quality of the labeled data directly affects the training effect of the model. For each image, bounding boxes need to be labeled first to delineate the maximum boundary of the defect region and label the defect category. The information corresponding to each labeled box includes: the position of the box and the target class information. Each labeled box corresponds to a label (category label), and the model is trained by comparing the original image with the labeled image, so that the model can learn the spatial distribution and features of the defect region.

[0024] Step 3: Image data augmentation methods (such as rotation, scaling, translation, affine transformation, etc.) are used to process the images of nuclear power equipment to increase sample diversity and improve the model's generalization ability. The augmented image data will be used for feature extraction through the backbone network of the network to extract potential features of equipment appearance defects; In step 3, various image enhancement strategies are employed to improve data quality. Since image data on external defects of nuclear power plant equipment is relatively scarce and the types of defects vary greatly depending on the environment, data enhancement becomes a necessary step to improve the robustness and generalization ability of the model. By transforming the original images in different ways, data diversity is enhanced, overfitting is reduced, and the model's adaptability to unknown defect types is improved.

[0025] Step 4: Use the backbone network and neck network of YOLOv12 as the backbone network and neck network part of the model, and use the region attention mechanism of YOLOv12 to optimize the computational efficiency and capture the global features of the defect appearance. In step 4, a paradigm shift was achieved by integrating a region-attention-centric mechanism, a streamlined architecture, and an optimized training pipeline. Building upon the foundations laid by its predecessors, YOLOv12 introduces a series of enhancements designed to maximize accuracy and computational efficiency. At its core is a redesigned feature extraction strategy that leverages Residual Efficient Layer Aggregation Network (R-ELAN), FlashAttention, and 7×7 separable convolutions to deliver superior throughput and accuracy. By fusing these strategies, YOLOv12 improves the performance of object detection tasks, ensuring it can skillfully handle complex visual scenes with varying levels of detail and appearance.

[0026] Step 5: Employing a ViT (VisionTransformer)-based encoder-decoder structure, feature queries are generated from the features extracted by YOLOv12. These queries are then fed into the ViT encoder, where multi-level interactive iterations refine the extraction of local detail features in the defect region. This process effectively enhances the ability to identify defects across multiple scales and complex backgrounds. In step 5, the process of converting the feature map into a query is first achieved by performing a linear transformation on the feature map. Specifically, each location (or each local region) of the feature map is mapped to a specific query vector, and then positional encoding is performed to generate the query vector.

[0027] Secondly, these query vectors are fed into the encoder part of ViT, becoming the input to the ViT model, used to aggregate important information from the feature maps. Each query vector corresponds to a specific region in the input image, with particular attention paid to defect regions. The ViT encoder-decoder structure refines and enhances the understanding of defect features through an iterative interaction process. This process utilizes ViT's self-attention mechanism to capture global dependencies in the image, defined as: , in, Q It is a query vector. K It is a key vector. V It is a value vector. It is the dimension of the key vector.

[0028] In the prediction of target categories, this invention proposes a decoupled prediction method for object content and object category. By learning the content and category of defects separately, without relying on pre-set unknown category labels, the model's adaptability to new defect types is improved. In each layer of the ViT encoder, the encoder aggregates information through a self-attention mechanism. Through this mechanism, the model can combine global information (e.g., background information of defects) with local information (e.g., the edges of defect appearance), enhancing its ability to perceive target features.

[0029] In the decoder section, the information processed by the encoder is decoded through multiple layers of interactive iteration, gradually refining the aggregated feature representation. At each layer, the decoder interacts with the information received from the encoder and its own initial query vector to further optimize and extract features from the defective regions.

[0030] Step 6: To address the shortcomings of traditional detection methods in detecting unknown categories, this invention designs an innovative category-aware detection head capable of recognizing unknown categories. The category-aware detection head and the decoder part of the ViT network are jointly optimized along the probability density model that serves as image features. Furthermore, by transforming the features of the target object into a multivariate Gaussian distribution in the query embedding space and dynamically updating the mean and variance of the distribution, the model can continuously adapt to new defect features during training. This strategy effectively solves the problem of constantly increasing and changing defect types, enabling the model to learn and adjust in real time in a constantly changing environment.

[0031] In step 6, a self-learning target detection method based on feature probability distribution is proposed, such as... Figure 1 As shown, the YOLOv12 model is improved by combining a feature probability distribution-based head and ViT. The self-learning detection method based on feature probability distribution first extracts features through a backbone and neck network, then generates a query containing potential target defect regions. Next, the ViT encoder-decoder structure performs deep iterative interaction on the feature query. Finally, the target class detection result is obtained through collaboration between the target classification head and the target regression head, while retaining the detection results for unknown classes. Secondly, the unknown class detection head retains the detection results for unknown classes. Finally, the unknown target bounding box optimization part processes redundant bounding boxes to obtain the final prediction result for the unknown class.

[0032] To avoid directly identifying unknown target objects, this invention proposes decoupled prediction of object content and object category. By learning object content and object category separately, it is not necessary to identify known objects during the training process. The proposed method does not rely on unknown category labels to detect unknown classes; instead, it starts from the perspective of different defect target features and determines whether a candidate region contains a potential defect target by learning the similarity of features of different target categories, thereby achieving the purpose of detecting unknown categories. Specifically, this invention proposes decoupled prediction of object content and object category. By learning the probability of defect targets and defect categories separately, , in, It belongs to a certain category when all possible defects exist. l The probability, This indicates that given a certain query q, it belongs to the defect category. l The probability, It represents the probability that the target defect exists in query q. When , it represents the background probability unrelated to the defect target. When the object does not exist, the conditional probability of the object category is 0; that is, when the model determines that the current query area is background, the classification head will not assign any target defect category. Ultimately, the model's prediction process can be defined as: , in, This indicates the determination of the target category in the query. When the given query represents the background, the category-aware detection head can predict a very low probability and will not assign any target defect category. When the query contains a target defect, the category-aware detection head will predict a probability value close to 1, and then the classification head will classify and identify the target object. The probability of a specific category is used to classify and identify the target object through a classification head. The classification head first assumes that the area has a defect, and then determines which category it belongs to. When the value is close to 0, it is judged as background and no defect is assigned. Conversely, it outputs a high probability and then identifies the specific defect category.

[0033] To address the issue of constantly emerging defects, this invention employs a probabilistic model. Specifically, the probability of the target object in the sample is transformed into a multivariate Gaussian distribution in the query embedding space, i.e. The likelihood probability of a target is used to predict whether a target object exists in a query. The calculation method is defined as follows: , Here, dist() represents the semantic distance of the query embedding. This invention uses a class-independent Gaussian distribution to model the embedded query, which allows the model to learn cross-class features.

[0034] In real-world scenarios, the distribution of target defects can change, and there may be distribution discrepancies between different target defects or even between the same target defect. Therefore, using a fixed Gaussian distribution to model query embeddings cannot continuously adapt to target features. This invention employs dynamically weighted updates of distribution parameters, allowing the model to continuously learn and adapt to target features. Specifically, the dynamically updated mean can be defined as: , in, This represents the number of samples currently processed, and q represents the currently processed embedded query. As the number of data samples increases, the impact of new samples decreases, and the update process becomes smoother. When updating the mean, the weight of the current number of samples is taken into account, allowing the model to gradually adapt to new samples without having an excessive impact on old samples.

[0035] Dynamic variance update can be defined as: , in, This represents the difference between the new sample and the existing samples. The number of data samples also plays a balancing role when updating the variance matrix.

[0036] The model is trained using a two-stage architecture. First, the distribution parameters are estimated using a dynamic Gaussian distribution. Second, the likelihood estimate of the matching query embedding is maximized. To ensure the matching query embedding is as close as possible to the center of the Gaussian distribution, thus improving the similarity between the query embedding and the target category in the high-dimensional semantic space, the Mahalanobis distance is used to penalize the matching query to maximize its likelihood. This penalty can be defined as: , in, Let be the Mahalanobis distance between the query embedding and the batch mean, and Q be the sum of the indexes of the query embeddings. By penalizing query embeddings that are far from the batch mean, the model can make the query embeddings of the target more clustered around the distribution center, thereby improving the accuracy of the prediction of the target class.

[0037] During model training, a composite loss function is used, which includes target regression head loss, target classification head loss, and category-aware head loss, to optimize the joint performance of defect regions and target categories.

[0038] The classification loss is used to train the target classifier head. It determines whether a query corresponds to a known target category or an unknown background by performing a classification prediction on each query embedding. The classification loss formula is: , in, Indicates the first Each query embedding represents the probability that it belongs to the target class. A sigmoid focus loss is used to address class imbalance and improve focus on hard-to-classify targets.

[0039] This is used to train the bounding box regression head, calculating the difference between the predicted and ground truth bounding boxes. Commonly used regression losses include L1 loss and generalized intersection-union loss (gIoULoss). The regression head loss formula is as follows: , The class loss of the class-aware head is: , The final total loss function is a weighted sum of the classification loss, bounding box regression loss, and object recognition loss, and can be defined as: , in, , , These are the weights of each loss term. The model learns the defect detection target by optimizing these loss functions.

[0040] Step 7: After inference is complete, the target classification head and regression head will output the category information of each detected target and its corresponding bounding box coordinates (i.e., defect location information). For category prediction, the learned category probability is multiplied by the classification probability to produce the final category prediction.

[0041] Step 7 shows that the query embedding refines the local detail features of the defect region through multi-layer interactive iteration. The final generated query embedding is used to estimate the target probability and category of each target region. Using a category-aware head, the model predicts whether each query embedding contains an object, i.e., whether the region is a target defect region. Category prediction determines whether the region needs further classification by calculating the probability of each query (i.e., whether it contains an object). For each query embedding predicted as a target defect, the model then uses a category-aware head to perform classification prediction, identifying whether it belongs to a known category or an unknown category.

[0042] In addition to category prediction, the model also performs bounding box regression for each target region, calculating the target's localization accuracy. Using the regression head, the model outputs the precise location of each detection box. During inference, such as... Figure 2 As shown, the model also performs Hungarian matching, avoiding complex nonmaximum suppression operations.

[0043] Finally, the model outputs a category label and corresponding bounding box coordinates for each detected target. If some target defects belong to an unknown category, the model will label them as "unknown" and assign them a corresponding probability value. The inference results will be visualized in practical applications, typically displaying the detected target boxes as image annotations and indicating their categories.

[0044] Example: The self-learning identification algorithm for appearance defects in nuclear power plant equipment proposed in this invention, based on feature probability distribution, has strong application value and can play an important role in actual nuclear power plant equipment inspection. The specific implementation steps are as follows: Step 1: Data Acquisition and Augmentation In implementing this invention, data acquisition, particularly image data of nuclear power plant equipment, is crucial. The goal of this step is to construct a diverse dataset encompassing different types of equipment, varying lighting conditions, different shooting angles, different materials, and varying degrees of corrosion. Specifically, critical equipment in nuclear power plants, such as pipes, valves, containers, and pumps, may exhibit defects such as rust and cracks. To ensure the comprehensiveness and representativeness of the dataset, image acquisition must cover equipment images under diverse environments, including various lighting conditions, different viewing angles, different surface materials, and defects with varying degrees of corrosion.

[0045] Step 2: Target Labeling and Dataset Construction After data acquisition, the images are labeled. This step involves labeling the bounding boxes of defect areas in each acquired image. Specifically, the outer bounding box of each defect is drawn, and the corresponding defect type is labeled. Each label includes the defect category label, the location of the bounding box (such as center coordinates and length and width dimensions), and other information.

[0046] To further improve the quality of the dataset and the accuracy of the annotation, this invention employs a pixel-level mask annotation method to ensure high precision of the annotation information. This annotation method provides the model with more detailed and accurate training data, helping the model better understand the spatial distribution and detailed features of defect areas.

[0047] Step 3: Image Data Enhancement and Feature Extraction After data labeling, data augmentation was performed on the three pairs of images, and features were extracted using a network. Data augmentation includes not only image rotation, scaling, cropping, and affine transformations, but also adjustments to parameters such as hue and brightness to simulate the appearance of devices under different environments. This approach increases sample diversity, reduces the model's dependence on the training set, and thus improves its adaptability and robustness to unknown defects.

[0048] After image enhancement, the next stage is feature extraction. This invention uses the backbone and neck network of YOLOv12 as the foundation for feature extraction. YOLOv12 has efficient feature extraction capabilities and can complete a comprehensive analysis of device images in a short time. During this process, the YOLOv12 network extracts local features of the image through convolutional layers and enhances the acquisition of global information through subsequent multi-scale fusion methods. YOLOv12 incorporates a region attention mechanism to improve its ability to focus on defect regions. This mechanism allows the model to adaptively adjust the attention region, automatically enhancing the model's focus on defect regions and suppressing attention to background regions. In this way, the model not only improves the accuracy of defect localization but also reduces the waste of computational resources, making the entire detection process more efficient.

[0049] Step 4: Introduce a ViT-based encoder-decoder architecture To refine the extraction of local detail features in defect areas, this invention employs a ViT-based encoder-decoder structure. First, feature queries are generated using feature maps extracted by YOLOv12. These feature queries are then embedded at specific locations, and the resulting query vectors are fed into the ViT encoder. Within the encoder, the query vectors undergo multi-level interactive iterations, progressively refining the features of the defect area at each level, thereby further improving the accuracy of defect identification.

[0050] Step 5: Category-aware detection and self-learning In traditional object detection methods, models often fail to accurately identify defects of unknown categories. However, this invention introduces a category-aware detection head and combines it with self-learning capabilities, enabling the model to continuously adapt to the emergence of new defect types. During training, the model learns how to transform target features into a multivariate Gaussian distribution in the query embedding space. During each training iteration, the model dynamically adjusts its mean and variance to continuously adapt to new defect features. This self-learning mechanism allows the model to update in real time within a constantly changing environment, ensuring its ability to detect unknown defect categories.

[0051] Step 6: Multi-task loss function and optimization To ensure the convergence and efficiency of the model during training, this invention designs a multi-task loss function that combines target classification loss, bounding box regression loss, and category-aware head loss. These loss functions work together to optimize the model's joint performance between defect regions and target categories, ensuring that the model can simultaneously maintain both localization and classification accuracy when facing different types of defects.

[0052] Step 7: Output the reasoning results After inference is complete, the target classification head and regression head will output the category information of each detected target and its corresponding bounding box coordinates. For category prediction, the learned category probability is multiplied by the classification probability to produce the final category prediction.

Claims

1. A self-learning method for identifying appearance defects in nuclear power plant equipment based on feature probability distribution, characterized in that, Includes the following steps: Step 1: Collect images of defects in key equipment of the nuclear power plant; Step 2: Annotate the target bounding boxes of the acquired images, annotate the outer bounding boxes of the defect areas, and construct a nuclear power equipment defect detection dataset; Step 3: Use image data augmentation methods to process the images of nuclear power equipment. The augmented image data is then used for feature extraction through the backbone network of the network to extract potential features of equipment appearance defects. Step 4: Use the backbone network and neck network of YOLOv12 as the backbone network and neck network part of the model, and use the region attention mechanism of YOLOv12 to optimize the computational efficiency and capture the global features of the defect appearance. Step 5: Using a ViT-based encoder-decoder structure, generate feature queries from the feature maps extracted by YOLOv12; Step 6: The class-aware detection head and the decoder part of the ViT network were jointly optimized along the probability density model as image features. By transforming the features of the target object into a multivariate Gaussian distribution in the query embedding space and dynamically updating the mean and variance of the distribution, the model can continuously adapt to new defect features during training. Step 7: After inference is completed, the target classification head and regression head will output the category information of each detected target and its corresponding bounding box coordinates. For category prediction, the learned category probability is multiplied by the classification probability to produce the final category prediction.

2. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 1, characterized in that: In step 2, the image data is labeled. For each image, the bounding box is first labeled to delineate the maximum boundary of the defect area and label the defect category. The information corresponding to each labeled box includes: the position of the box and the target class information. Each labeled box corresponds to a category label. The model is trained by comparing the original image with the labeled image, so that the model can learn the spatial distribution and features of the defect area.

3. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 1, characterized in that: In step 3, various image enhancement strategies are adopted to improve data quality. By transforming the original image in different ways, data diversity is enhanced, overfitting is reduced, and the model's adaptability to unknown defect types is improved.

4. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 1, characterized in that: In step 4, a paradigm shift is achieved by integrating a region attention-centric mechanism, a streamlined architecture design, and an optimized training pipeline. This leverages the remaining efficient layers to aggregate the network R-ELAN, FlashAttention, and 7×7 separable convolutions to provide throughput and accuracy.

5. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 1, characterized in that: The process of converting the feature map into a query in step 5 is achieved by performing a linear transformation on the feature map. Each position of the feature map is mapped to a specific query vector, and then position encoding is performed to generate the query vector. These query vectors will be fed into the encoder part of ViT, becoming the input to the ViT model, used to aggregate important information from the feature maps. Each query vector corresponds to a specific region in the input image. The ViT encoder-decoder structure is defined through an iterative interaction process: ， in, Q It is a query vector. K It is a key vector. V It is a value vector. It is the dimension of the key vector.

6. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 5, characterized in that: In step 7, during the prediction of the target category, a decoupled prediction method for object content and object category is adopted. By learning the content and category of the defect separately, the model's adaptability to new defect types is improved. In each layer of the ViT encoder, the encoder aggregates information through a self-attention mechanism. In the decoder, the information processed by the encoder is decoded through multi-layer interactive iteration to gradually refine the aggregated feature representation. In each layer, the decoder interacts with the information received from the encoder and its own initial query vector to further optimize and extract the features of the defect region.

7. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 1, characterized in that: Step 6 employs a self-learning target detection method based on feature probability distribution, which includes combining a head based on feature probability distribution with ViT to improve the YOLOv12 model. The self-learning detection method based on feature probability distribution first extracts features through the backbone network and neck network, then generates a query containing potential target defect regions. Next, the encoder-decoder structure of ViT performs deep iterative interaction on the feature query. Finally, the target class detection result is obtained through the collaboration of the target classification head and the target regression head, while retaining the detection results of unknown classes. The unknown class detection head retains the detection results of unknown classes, and the unknown target bounding box optimization part processes redundant bounding boxes to obtain the final prediction result of the unknown class.

8. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 7, characterized in that: Step 6 employs decoupled prediction of object content and object category, by learning the probability of defect targets and defect categories separately. ， in, It belongs to a certain category when all possible defects exist. l The probability, This indicates that given a certain query q, it belongs to the defect category. l The probability, It represents the probability that the target defect exists in query q. When the object does not exist, the conditional probability of the object category is 0. That is, when the model determines that the current query area is background, the classification head will not assign any target defect category. The model's prediction process is defined as follows: ， in, This indicates the determination of the target category in the query. When the query represents the background, the category-aware detection head can predict a very low probability and will not assign any target defect category. When the query contains a target defect, the category-aware detection head will predict a probability value close to 1, and then the classification head will classify and identify the target object. This represents the probability of a specific category.

9. The self-learning identification method for appearance defects of nuclear power plant equipment based on feature probability distribution as described in claim 1, characterized in that: In step 6, the query embedding is generated through multi-layer interactive iteration. The generated query embedding is used to estimate the target probability and category of each target region. Through the category-aware head, the model predicts whether each query embedding contains an object, that is, whether the region is a target defect region. The category prediction determines whether the region needs further classification by calculating the probability of each query. For each query embedding predicted as a target defect, the category-aware head is used to classify and predict whether it belongs to a known category or an unknown category.

Images