Real-time defect detection method for nuclear fuel element

Abstract

The invention discloses a real-time defect detection method for a nuclear fuel element, and relates to the technical field of nuclear fuel element detection, and the method comprises the steps: inputting a to-be-detected nuclear fuel element image into a pre-trained NCSN-based defect detection model, carrying out the calculation to obtain NCSN multistage loss, and calculating the NCSN multistage loss; the NCSN multi-level loss is the fractional loss of the nuclear fuel element image under different levels of noise of the NCSN; screening a defect image from the nuclear fuel element image to be detected according to NCSN multi-stage loss by using a defect screening network based on a K-means algorithm; and carrying out image reconstruction on the defect image by using an NCSN network to obtain a normal image based on defect image defect removal, calculating a residual image of the defect image and the normal image, and segmenting and positioning a defect position by using an Otsu threshold. According to the invention, higher efficiency, stability and accuracy can be provided in real-time defect detection of the nuclear fuel element, and intelligent and efficient development of an industrial detection technology is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear fuel element detection, and in particular to a real-time defect detection method for nuclear fuel elements. Background Art

[0002] Nuclear fuel elements are the most core and critical components of nuclear reactors, and their structural integrity is one of the keys that directly affects the technical status of nuclear reactors. The structural integrity of fuel elements has a significant impact on the corrosion resistance, impact resistance, and heat transfer performance of nuclear fuels, and can also cause problems such as fuel element defect migration, breakage, and leakage, leading to the leakage of radioactive substances. Therefore, structural integrity testing of fuel elements is an indispensable step to ensure the safe operation of reactors.

[0003] For fuel element defect detection, the goal is to identify defect information through image features. Defect information has the following characteristics: unknown: defect information has unknown burst behavior, data structure and distribution; heterogeneity: defects are usually irregular; category imbalance: normal data will account for the vast majority of data, while data containing defects is relatively small.

[0004] At present, conventional manual methods are mainly used for defect identification, and a large amount of test data will be generated in batch testing, which has problems such as low efficiency, easy omissions, and subjective factors. With the development of computer technology, methods for automatically detecting fuel elements using intelligent algorithms have emerged. For example, there are many supervised learning methods that have been applied to anomaly detection and have good results. However, it is difficult to obtain a large number of labeled data sets in fuel element defect detection, which will affect the detection results. Summary of the invention

[0005] The present invention provides a real-time defect detection method for nuclear fuel elements, which can solve the above problems.

[0006] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] The present invention provides a real-time defect detection method for nuclear fuel elements, comprising the following steps:

[0008] S1. Input the image of the nuclear fuel element to be detected into the pre-trained defect detection model based on the NCSN network, and calculate the NCSN multi-level loss, which is the fractional loss of the nuclear fuel element image under different levels of noise of the NCSN network;

[0009] S2, using a defect screening network based on the K-means algorithm to screen out defect images from the nuclear fuel element images to be inspected according to the NCSN multi-level loss;

[0010] S3. Reconstruct the defect image using the NCSN network to obtain a reconstructed image based on the defect image with defects removed, calculate the residual image between the defect image and the reconstructed image, and locate the defect position using Otsu threshold segmentation.

[0011] As a further description of the above technical solution: by calculating the fractional loss of nuclear fuel element images under different levels of noise, the defect detection model based on the NCSN network is trained; the loss function used is:

[0012]

[0013] Among them, the first term is the estimated score function output by the model sθ(x), the second term is the true score function, and the loss function is the squared Euclidean distance between the score output by the model and the true score.

[0014] As a further description of the above technical solution: the network noise level of the defect detection model based on the NCSN network has 10 levels.

[0015] As a further description of the above technical solution: in step S2, losses with noise levels of 5, 6, 7, and 8 are selected from the NCSN multi-level losses as inputs to the defect screening network.

[0016] As a further description of the above technical solution: in the process of the K-means algorithm in step S2, the number of clustering categories is set to 2, representing the normal image set and the defective image set respectively, and the number of nuclear fuel element images to be inspected for clustering each time is set to 1.

[0017] As a further description of the above technical solution: in the process of the K-means algorithm in step S2, when the number of categories in one of the two categories is 0, the centroid is initialized C(c1, c2), where c1 is the centroid of the normal image set and c2 is the centroid of the abnormal image set; the centroid initialization process includes inputting the loss value Loss of the nuclear fuel element images to be clustered, if c1≥Loss, then c1=Loss, otherwise no adjustment is made, if c2≤Loss, then c2=Loss, otherwise no adjustment is made.

[0018] As a further description of the above technical solution: the K-means algorithm process in step S2 also includes calculating the distance between the centroid and the input image. When the number of one of the two categories is 0, an inter-class distance judgment is performed. When the inter-class distance is greater than the threshold λ, the input image will be classified into a new category; when the number of two categories is both greater than 0, the distance between the input image and the two centroids is compared, and the image is divided into a category with a smaller distance.

[0019] As a further description of the above technical solution: the K-means algorithm process in step S2 also includes updating the centroid coordinates after classifying the new image.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1) A defect image recognition method based on NCSN network with different noise level losses is proposed. Through the training of multi-level noise loss, the recognition accuracy and efficiency of defect images by unsupervised learning methods can be effectively improved, providing a more reliable foundation for subsequent real-time defect detection;

[0022] 2) Combined with the K-means clustering algorithm, by performing real-time clustering of different noise level losses in the NCSN network, combined with class spacing control and historical clustering data memory, clustering while detecting is effectively achieved, thereby greatly improving the efficiency of defect detection and the robustness of the model, avoiding the computational cost of each image reconstruction, and improving real-time detection capabilities.

[0023] 3) For defective images, the annealed Langevin dynamic sampling method is used to remove defects, and Otsu threshold segmentation is performed based on the residual between the image after defect removal and the original image, so as to accurately locate the defect position, further improving the accuracy of defect detection and positioning precision.

[0024] 4) The present invention overcomes the problems of low efficiency and instability of manual detection methods, realizes intelligent detection through deep learning, avoids the dependence of traditional supervised learning methods on large-scale labeled data, and adopts the unsupervised learning NCSN generation network method for anomaly detection, which improves the level of automation of detection and adapts to the challenge of data scarcity in industrial environments.

[0025] 5) The real-time detection method combined with K-means clustering proposed in the present invention can effectively improve the detection speed without increasing the additional computing burden, ensure the real-time performance of generating network anomaly detection, and optimize the defect detection efficiency in production practice.

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 It is the overall framework diagram of the model in the embodiment;

[0029] Figure 2 is a K-means clustering algorithm framework diagram in the embodiment;

[0030] Figure 3 is an AUROC curve diagram of different noise levels in the embodiment;

[0031] Figure 4 1 is the defect image and positioning result in the embodiment. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0033] Figure 1 : is a main flow chart of the technical solution of the embodiment of the present invention. A real-time defect detection method for nuclear fuel elements proposed in the embodiment of the present invention comprises the following steps:

[0034] Step (1): Construct a dataset of 1460 X-ray images of fuel element substitutes. These images are cross-sectional images of CT images of fuel element substitutes, including normal image samples and defect image samples. The dataset includes a training set and a test set. The training set consists of 1260 normal image samples to simulate the scarcity of defect samples in a natural industrial environment. The test set includes 80 normal images and 120 defect images for a more comprehensive evaluation.

[0035] Step (2): Build a defect detection model based on the NCSN network and perform model training. Input the training set data into the network, and set the network noise level to 10. Train by calculating the score loss of images under different levels of noise. The final loss function is as follows:

[0036]

[0037] where s θ (x) represents model s θ (x) Output estimated score function, s θ is a defect detection model based on the NCSN network model with parameter θ; The image after adding different noises to the input image is defined as Where x is the input image, σ i is the noise level, ∈ represents the standard Gaussian distribution. Approximately, p(x) represents the probability density function of the input image x, then Represents the true score function, that is, the score function of the input image. In summary, the loss function is the squared Euclidean distance between the score output by the model and the true score.

[0038] Step (3): Build a defect screening network based on the K-means algorithm. Based on the network trained in step (2), the network test mode can output the loss of different images at different noise levels. The AUROC curve of the test set image at different noise levels is as follows: Figure 3 As shown, in order to increase the detection efficiency and ensure the robustness of the network, the output loss with noise levels of 5, 6, 7, and 8 are selected as the input of the K-means clustering network.

[0039] like Figure 2 As shown, the process of K-means clustering algorithm is as follows:

[0040] 1) The number of clustering categories in the clustering algorithm is 2, representing the normal image set and the defect image set respectively. A is the number of images in the normal image set, and B is the number of images in the defect image set.

[0041] 2) Input an image F of a nuclear fuel element to be detected. In order to achieve synchronization of detection and identification, the number of images to be clustered each time is 1;

[0042] 3) When the number of one of the two categories is 0, the centroid will be initialized to C(c1,c2), where c1 is the centroid of the normal image set and c2 is the centroid of the abnormal image set;

[0043] 4) Centroid initialization process: Input the Loss of the image to be clustered F. Since the Loss of the defective image is greater than that of the normal image, if c1≥Loss, then let c1=Loss, otherwise no adjustment is made; if c2≤Loss, then let c2=Loss, otherwise no adjustment is made;

[0044] 5) Calculate the distance between the centroid and the input image. When the number of one of the two categories is 0, the inter-class distance judgment will be performed. Only when the inter-class distance is greater than the threshold λ, the image will be classified into the new category. When the number of both categories is greater than 0, compare the distance between the input image and the two centroids, and divide the image into the category with the smaller distance.

[0045] 6) After dividing the image F into categories, the centroid coordinates are updated as follows (taking c1 as an example), where is the k+1 generation centroid cluster, is the k-generation centroid cluster, n k is the number of samples in the c1 cluster of generation k, n k =n k-1 +m k-1, α is the forgetting factor representing the degree of memory of the historical centroid clusters, and its value range is [0,1], represents the centroid of the new image of the current generation, m k is the number of samples added to cluster c1 in the new image h.

[0046]

[0047] 7) Repeat steps 5)-6) until the distance between the centroid and the image is less than the threshold θ;

[0048] 8) Repeat steps 2)-7) until all images are clustered.

[0049] Step (4): For the defective image selected in step (3), the NCSN network is used to reconstruct the image, obtain the reconstructed image based on the defective image, and calculate the residual image of the two images. The Otsu threshold segmentation is used to locate the defect position, such as Figure 4 shown.

[0050] In the embodiment of the present invention, the area under the receiver operating characteristic curve (ROC) (AUC) is selected as a quantitative evaluation index of model performance. The ROC curve is an image drawn with the false positive rate (FPR) as the horizontal coordinate and the true positive rate (TPR) as the vertical coordinate. The ROC curve describes the performance of the classification model under various threshold conditions in detail. Specifically, the definitions of FPR and TPR are as follows:

[0051]

[0052] Among them, TP, TN, FN and FP represent true positive, true negative, false negative and false positive, respectively.

[0053] The AUROC curve obtained in step (3) is as follows: Figure 3 As shown in Table 1, the AUC values ​​of different curves are selected as the input of the K-means clustering network after comparison. The output loss of noise levels 5, 6, 7, and 8 are selected as the input of the K-means clustering network. The defect recognition accuracy in the final clustering result of the test set is 100.00% and the error is 0. For the screened defective images, the NCSN network is used for image reconstruction to obtain a normal image based on the defective image to remove the defect, and the residual map of the two images is calculated. The Otsu threshold segmentation is used to locate the defect position, as shown in Figure 4 shown.

[0054] Table 1 AUC at different noise levels

[0055]

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A real-time defect detection method for nuclear fuel elements, characterized in that: The following steps are involved: S1. Input the image of the nuclear fuel element to be detected into the pre-trained defect detection model based on the NCSN network, and calculate the NCSN multi-level loss, which is the fractional loss of the nuclear fuel element image under different levels of noise of the NCSN network; S2, using a defect screening network based on the K-means algorithm to screen out defect images from the nuclear fuel element images to be inspected according to the NCSN multi-level loss; S3. Reconstruct the defect image using the NCSN network to obtain a reconstructed image based on the defect image with defects removed, calculate the residual image between the defect image and the reconstructed image, and locate the defect position using Otsu threshold segmentation.

2. The real-time defect detection method for nuclear fuel elements according to claim 1, characterized in that: By calculating the fractional loss of nuclear fuel element images under different levels of noise, the defect detection model based on the NCSN network is trained; the loss function used is: Among them, the first term is the estimated score function output by the model sθ(x), the second term is the true score function, and the loss function is the squared Euclidean distance between the score output by the model and the true score.

3. The real-time defect detection method for nuclear fuel elements according to claim 1, characterized in that: There are 10 levels of network noise in the defect detection model based on the NCSN network.

4. The real-time defect detection method for nuclear fuel elements according to claim 3, characterized in that: In step S2, the noise level losses of 5, 6, 7, and 8 are selected from the NCSN multi-level loss as the input of the defect screening network.

5. The real-time defect detection method for nuclear fuel elements according to claim 1, characterized in that: In the process of the K-means algorithm in step S2, the number of clustering categories is set to 2, representing a normal image set and a defective image set, respectively, and the number of nuclear fuel element images to be detected for clustering each time is set to 1.

6. The real-time defect detection method for nuclear fuel elements according to claim 5, characterized in that: In the process of the K-means algorithm in step S2, when the number of one of the two categories is 0, the centroid is initialized C(c1,c2), where c1 is the centroid of the normal image set and c2 is the centroid of the abnormal image set; The initial process of centroid includes inputting the loss value Loss of the nuclear fuel element image to be clustered, if c1≥Loss, then c1=Loss, otherwise no adjustment is made, if c2≤Loss, then c2=Loss, otherwise no adjustment is made.

7. The real-time defect detection method for nuclear fuel elements according to claim 6, characterized in that: The K-means algorithm process in step S2 also includes calculating the distance between the centroid and the input image. When the number of one of the two categories is 0, the inter-class distance is judged. When the inter-class distance is greater than the threshold λ, the input image will be classified into a new category. When the number of both categories is greater than 0, the distance between the input image and the two centroids is compared, and the image is divided into a category with a smaller distance.

8. The real-time defect detection method for nuclear fuel elements according to claim 7, characterized in that: The K-means algorithm process in step S2 also includes updating the centroid coordinates after classifying the new image.

Images