Residual core detection and grading method for sub-image of aviation turbine blade

The method enhances residual core detection in aviation turbine blades by using a Patch SVDD model with DRN and ECA for neutron images, improving accuracy and efficiency in detecting and grading residual cores.

CN119090847AActive Publication Date: 2024-12-06SICHUAN UNIVERSITY OF SCIENCE AND ENGINEERING

Patent Information

Application Number
CN202411198744.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-06
Estimated Expiration
2044-08-29

AI Technical Summary

Technical Problem

Existing non-destructive testing methods for detecting residual cores in aviation turbine blades are prone to human misjudgment, inefficiency, and are not suitable for complex internal structures, particularly those with apertures less than 0.5 mm, necessitating improved automated detection methods.

Method used

A residual core detection and classification method for neutron images using a pre-trained model based on the Patch SVDD model with a dilated residual network (DRN) and efficient channel attention (ECA) module, enhancing image quality and feature extraction, followed by quantitative grading of residual cores.

Benefits of technology

The method significantly improves the accuracy and efficiency of residual core detection and classification, reducing misjudgment and enhancing production efficiency by accurately identifying and grading residual cores without relying on annotation information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090847A_ABST
    Figure CN119090847A_ABST
Patent Text Reader

Abstract

The invention discloses a residual core detection and grading method for a neutron image of an aviation turbine blade, and the method comprises the steps: obtaining the neutron image of the aviation turbine blade to be detected, and carrying out the enhancement processing of the neutron image of the blade; the enhanced blade sub-image is input into a pre-trained residual core detection model, a residual core detection result image is output, the residual core detection result image comprises a normal image without a residual core and an abnormal image with a residual core, and the residual core detection model is obtained based on training of an improved Patch SVDD model; and carrying out residual core area identification on the residual core detection result image, and classifying abnormal images with residual cores according to the size of the identified residual core area and a set grade to realize residual core grading of the aviation turbine blade. According to the method, the detection accuracy of the residual core of the sub-image of the aviation turbine blade is improved, the detection result is quantitatively evaluated, and the improvement of the core removal efficiency in industrial production practice is facilitated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of aeroengine blade detection, and in particular to a residual core detection and grading method for a neutron image of an aeroengine blade. Background Art

[0002] As the main power source of aircraft, engine blades operate in extreme environments of high temperature, high stress, and high speed. Because the internal channel structure of the blades has features such as bends, blind holes, and microcavities, it is easy to produce residual cores that are difficult to detect. The presence of residual cores will block the heat dissipation channels of the blades, thereby forming hot spots on the blades, which will cause chemical corrosion under the influence of high temperature and stress, and then seriously damage the blades; in addition, in a high temperature environment, the flow of residual cores may destroy the balance of the engine and even cause accidents. Therefore, all produced engine turbine blades must undergo strict testing before being used in aircraft engines to confirm that there are no residual cores inside them to ensure engine quality and aviation flight safety.

[0003] In recent years, a variety of non-destructive testing technologies (NDT) have been applied to the residual core detection of turbine blades, aiming to accurately identify the residual core inside the turbine blades of aircraft engines without damaging the structural integrity of the turbine blades. These technologies include eddy current testing, endoscopic testing, industrial CT testing, X-ray testing, and neutron imaging testing. The above technologies mainly rely on experienced inspectors to judge the later imaging results, but the manual judgment process is inevitably prone to misjudgment due to human fatigue, such as false positive phenomenon (that is, there is a residual core in the aircraft turbine blade, but it is misjudged as normal by the inspector), and manual judgment also has problems such as low efficiency and long time consumption.

[0004] Due to the rapid development of deep learning, many researchers have begun to apply deep learning-based methods to residual core detection tasks. Some studies have used the YOLO algorithm to detect residual cores in endoscopic images of aviation turbine blades, which has improved detection efficiency. However, this supervised learning method relies on a large amount of labeled information, and endoscopic measurement is not suitable for blades with complex internal structures, tortuous channels, and apertures less than 0.5 mm. Therefore, it is necessary to develop new residual core detection methods. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a residual core detection and grading method for neutron images of aviation turbine blades, which can realize the detection, positioning and quantitative grading of residual cores without relying on annotation information, thereby improving the core removal efficiency in production practice.

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

[0007] A residual core detection and classification method for neutron images of aviation turbine blades, comprising the following steps:

[0008] S10, acquiring a neutron image of the aviation turbine blade to be tested, and performing enhancement processing on the blade neutron image;

[0009] S20, inputting the enhanced blade neutron image into a pre-trained residual core detection model, and outputting residual core detection result images, including two types of normal images without residual cores and abnormal images with residual cores, wherein the residual core detection model is built based on the Patch SVDD model, and the dilated residual network DRN is used instead as the feature extraction network, and an efficient channel attention ECA module is added at the output position of the dilated residual network DRN;

[0010] S30, performing residual core area recognition on the residual core detection result image, and classifying abnormal images with residual cores according to the size of the recognized residual core area and the set level, so as to achieve residual core classification of aviation turbine blades.

[0011] Specifically, in S10, a neutron imaging detection method is used to obtain a neutron image of the aviation turbine blade to be tested.

[0012] Specifically, the process of enhancing the leaf neutron image in S10 includes:

[0013] S11, using the Gamma transformation method to enhance the contrast of the leaf neutron image, and using the Gaussian filtering method based on the spatial mean filter to smooth the leaf neutron image;

[0014] S12, using the Sobel operator to extract the horizontal edge and the vertical edge of the blade neutron image after smoothing, to obtain high-frequency information of the blade neutron image;

[0015] S13. Linearly superimposing the extracted high-frequency information with the leaf neutron image enhanced by the Gamma transform method to improve the quality of the leaf neutron image.

[0016] Specifically, the process of constructing the residual core detection model in S20 includes:

[0017] The Patch SVDD model is selected as the basic model. The input image is divided into blocks through a sliding window method to form a large number of partially overlapping blocks. These blocks are input into the feature extraction network for feature extraction, so that the features extracted from each block correspond to a single point in the feature space. The encoder is trained to find the center of blocks with similar semantics. At the same time, an encoder and a classifier are trained to predict the relative position of two adjacent blocks in space to verify whether the encoder has sufficient semantic information. Then, the degree of abnormality of the block is obtained by calculating the Euclidean distance between the block to be tested and the normal block in the feature space.

[0018] Using the dilated residual network DRN to replace the original feature extraction network of the Patch SVDD model can capture more complex spatial relationships and shape features and enhance the recognition of residual core shape features. The dilated residual network DRN is composed of dilated convolution introduced on the basis of the ResNet network structure.

[0019] Then, an efficient channel attention ECA module is added to the output position of the dilated residual network DRN to filter out irrelevant features and reduce the impact of noise. The ECA module uses one-dimensional convolution to achieve cross-channel dimensional interaction, avoid channel dimensionality reduction, and improve classification accuracy.

[0020] Specifically, the final loss function L of the residual core detection model is PatchSVDD It is expressed as:

[0021] L PatchSVDD =λL SVDD ′+L SSL

[0022]

[0023] L SSL = Cross-entropy(y,C φ (f θ (p1),f θ (p2)))

[0024] Where λ represents the scaling hyperparameter, p i′ Represents block p i nearby blocks, f is the encoder, f θ (p i ) represents the encoder calculation block p i The obtained eigenvector, f θ (p i′ ) represents the encoder calculation block p i′ The obtained feature vector; p1 is a randomly sampled block, p2 is a block in the neighborhood of p1, y is the relative position between the two blocks, C φ For the classifier, Cross-entropy represents the cross entropy loss.

[0025] Specifically, the training process of the residual core detection model in S20 includes:

[0026] S21a, collect existing aviation turbine blade neutron images to build a data set;

[0027] S21b, using the Labelme tool to accurately label the residual core area in the image with residual core;

[0028] S21c, the dataset is divided into a training set, a test set and a defect annotation set, where the training set contains only normal image samples without residual cores to simulate the scarcity of residual core samples in natural industrial environments, and the test set contains normal image samples without residual cores and abnormal image samples with residual cores for comprehensive evaluation;

[0029] S21d, performing enhancement processing on the leaf neutron image in the data set to form an enhanced data set;

[0030] S21e. Use the enhanced data set to train the constructed residual core detection model to obtain a pre-trained residual core detection model.

[0031] Furthermore, in S21a, when there are different types of aviation turbine blades in the existing aviation turbine blade neutron images, the blade neutron images are first classified according to the types of the aviation turbine blades, so that each type of aviation turbine blade neutron image is independently a data set of its own type;

[0032] In S21b-S21e, data processing and model training are performed on each type of data set to obtain a pre-trained residual core detection model corresponding to each type of aviation turbine blade;

[0033] In S20, the type of the aviation turbine blade to which the enhanced blade neutron image belongs is first determined, and then a pre-trained residual core detection model of the corresponding type is selected to perform residual core detection on the enhanced blade neutron image.

[0034] Specifically, the specific process of S30 includes:

[0035] S31, converting the blade image of the input residual core detection result image into a grayscale image, converting the grayscale image into a binary image by a binarization method, separating the outline of the aviation turbine blade from the background, and calculating the total number of pixels of the aviation turbine blade in the binary image;

[0036] S32, binarizing the abnormal image of the input residual core detection result image, calculating the number of residual core pixel points in the image, calculating the ratio of the total number of aviation turbine blade pixel points to the number of residual core pixel points, and dividing the residual core detection result image into a normal blade group and an abnormal blade group;

[0037] S33, normalizing the abnormal image of the abnormal leaf group, and then using Gaussian smoothing method to process the normalized image to reduce the influence of noise and soften the edge, and then binarizing the smoothed image to calculate the number of pixels in the residual core area to determine the size of the residual core area;

[0038] S34. Calculate the ratio of the number of pixels in the residual core area to the total number of pixels in the aviation turbine blade to obtain the proportion of the residual core area in the aviation turbine blade, and determine the residual core grading level of the aviation turbine blade corresponding to the abnormal image with the residual core in combination with a pre-set proportion grade classification.

[0039] In the step S31, morphological erosion processing is performed on the binary image to reduce the influence of noise and artifacts, thereby clearly retaining the edge and shape of the aviation turbine blade.

[0040] Specifically, the residual core grading levels of the aviation turbine blades in S34 include two categories: Major and Critical, wherein Major refers to blades that need to be de-cored twice or less, and Critical refers to blades that need to be de-cored more than twice.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) The present invention enhances the neutron image of the aviation turbine blade to be tested, thereby improving the image contrast and detail clarity, providing convenience for subsequent residual core detection. The residual core detection model is obtained by integrating DRN and ECA into the Patch SVDD model for training, which significantly enhances the model's ability to recognize residual core features and improves the accuracy of residual core detection. The residual core grading process is added to perform a detailed quantitative evaluation of the residual core identified in the image, thereby improving the efficiency of core removal in industrial production practice.

[0043] (2) The present invention adopts an improved gamma transformation algorithm in image enhancement processing, which effectively improves the contrast of the sub-image in the blade and the clarity of its details, thereby improving the accuracy of subsequent residual core detection.

[0044] (3) The present invention improves the Patch SVDD model by using the dilated residual network (DRN) and the efficient channel attention (ECA) module. The residual core detection model trained in this way can more effectively capture and learn rich and accurate semantic information, significantly improving the model's ability to recognize residual core features, thereby achieving more accurate residual core detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention.

[0046] Figure 2 It is an overall framework diagram of the residual core detection model in the embodiment of the present invention.

[0047] Figure 3 It is a schematic diagram of the process of leaf neutron image enhancement processing in an embodiment of the present invention.

[0048] Figure 4 It is a schematic diagram of the process of residual core classification in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below in conjunction with the accompanying drawings and examples. The embodiments of the present invention include but are not limited to the following examples.

[0050] Example

[0051] like Figures 1 to 4 As shown, the residual core detection and classification method for neutron images of aviation turbine blades includes the following steps:

[0052] S10. A neutron imaging detection method is used to obtain a neutron image of the aviation turbine blade to be tested. For example, a parallel cold neutron beam is used to irradiate the aviation turbine blade to obtain a neutron image of the aviation turbine blade. The quality of the neutron image is subject to the influence of the neutron imaging equipment, the neutron source intensity, the accuracy of the imaging system and physical factors, resulting in low contrast of the imaging result, noise, artifacts, blurred edges and other degradation phenomena. The existence of these phenomena greatly interferes with the identification of the residual core of the turbine blade. Therefore, in order to reduce the occurrence of misjudgment, missed judgment and inaccurate positioning in the detection, the present invention also enhances the neutron image of the blade.

[0053] Specifically, the process of enhancing the leaf neutron image in S10 includes:

[0054] S11. The Gamma transformation method is used to enhance the contrast of the leaf sub-image. At the same time, the Gaussian filtering method based on the spatial mean filter is used to smooth the leaf sub-image to reduce the impact of noise on the subsequent edge extraction results.

[0055] S12. Using the Sobel operator to extract the horizontal edge and the vertical edge of the smoothed blade sub-image, high-frequency information of the blade sub-image is obtained.

[0056] S13. Linearly superimpose the extracted high-frequency information with the blade neutron image enhanced by the Gamma transform method to further improve the quality of the blade neutron image and achieve aviation turbine blade neutron image enhancement.

[0057] S20, inputting the enhanced blade neutron image into a pre-trained residual core detection model, and outputting residual core detection result images, including two types: normal images without residual cores and abnormal images with residual cores.

[0058] The residual core detection model is built based on the Patch SVDD model, specifically:

[0059] The Patch SVDD model is selected as the basic model. The input image is divided into blocks through a sliding window method to form a large number of partially overlapping blocks. These blocks are input into the feature extraction network for feature extraction, so that the features extracted from each block correspond to a separate point in the feature space. The encoder is trained to find the center of blocks with similar semantics. At the same time, an encoder and a classifier are trained to predict the relative positions of two adjacent blocks in the space to verify whether the encoder has sufficient semantic information. Then, the Euclidean distance between the block to be tested and the normal block in the feature space is calculated to obtain the abnormality of the block.

[0060] Using the dilated residual network DRN to replace the original feature extraction network of the Patch SVDD model can capture more complex spatial relationships and shape features and enhance the recognition of residual core shape features. The dilated residual network DRN is composed of dilated convolutions introduced on the basis of the ResNet network structure. The introduction of dilated convolutions allows the DRN network to achieve the same receptive field as the original network while keeping the output size consistent with the input, without the need for pooling operations, making it particularly suitable for residual core detection and positioning tasks. The residual structure can effectively propagate low-level features to high-level layers, thereby making full use of contextual details and solving problems such as gradient disappearance and model degradation. The introduction of DRN means that the model no longer focuses only on texture features and ignores shape features. The deepening of the network depth enables the model to capture more abstract semantic information and has a stronger ability to learn complex and fine feature expressions.

[0061] Then, an efficient channel attention ECA module is added to the output position of the dilated residual network DRN to effectively filter out irrelevant features and reduce the impact of noise, so as to clearly highlight the target elements and improve the generalization ability and performance of the model; the ECA module uses one-dimensional convolution to achieve cross-channel dimensional interaction, avoid channel dimensionality reduction, and improve classification accuracy. ECA is a lightweight plug-and-play module that allows the network to focus on the target to be detected, correspondingly achieve the outstanding performance of important features, and thus improve the accuracy of the network model in target detection.

[0062] Specifically, the final loss function L of the residual core detection model is PatchSVDD It is expressed as:

[0063] L PatchSVDD =λL SVDD ′+L SSL

[0064]

[0065] L SSL = Cross-entropy(y,C φ (f θ (p1),f θ (p2)))

[0066] Where λ represents the scaling hyperparameter, p i′ Represents block p i nearby blocks, f is the encoder, f θ (p i ) represents the encoder calculation block p i The obtained eigenvector, f θ (p i′ ) represents the encoder calculation block p i′ The obtained feature vector; p1 is a randomly sampled block, p2 is a block in the neighborhood of p1, y is the relative position between the two blocks, C φ For the classifier, Cross-entropy represents the cross entropy loss.

[0067] Specifically, the training process of the residual core detection model includes:

[0068] S21a. Collect existing aviation turbine blade neutron images to construct a dataset.

[0069] S21b, use the Labelme tool to accurately mark the residual core area in the image with residual core for subsequent analysis.

[0070] S21c. The dataset format is set with reference to the public dataset MVTec-AD. The dataset is divided into a training set, a test set, and a groundtruth defect annotation set. The training set only contains normal image samples without residual cores to simulate the scarcity of residual core samples in natural industrial environments. The test set contains normal image samples without residual cores and abnormal image samples with residual cores for comprehensive evaluation.

[0071] S21d, performing enhancement processing on the blade neutron image in the data set to form an enhanced data set; the enhancement processing adopts the enhancement processing method in S10, which can be achieved by building a neutron image enhancement module.

[0072] S21e. Use the enhanced data set to train the constructed residual core detection model to obtain a pre-trained residual core detection model.

[0073] Furthermore, when collecting existing neutron images of aviation turbine blades in S21a, there are usually many different types of aviation turbine blades. In order to avoid different types of aviation turbine blades affecting each other in residual core detection, we first classify the neutron images of the blades according to the type of aviation turbine blades, so that each type of neutron image of aviation turbine blades is independent as a data set of its own type. For example, 1,272 neutron images of aviation turbine blades were collected in the experiment. These images contain three different types of aviation turbine blades. Based on the different characteristics of these three types of blades, they are divided into three categories of A, B, and C in advance to facilitate subsequent accurate model training and residual core detection.

[0074] Therefore, it can be achieved through the process of S21b-S21e: data processing and model training are performed on each type of data set respectively to obtain a pre-trained residual core detection model corresponding to each type of aviation turbine blade.

[0075] During the model training phase, patch svdd divides the input image into many overlapping patches by sliding windows, and then sends each patch to the encoder for feature extraction. During extraction, the input image block p is subdivided into 2×2 sub-patches and passed through a hierarchical encoder. The smaller encoder f is used first. small Extract the features of each sub-block. Then, the aggregate encoder f big To aggregate the overall features of the image block p. The features of each extracted patch are mapped to the feature space (the feature extraction and mapping process will be continuously optimized through the loss function, while minimizing the distance of the normal patch in the feature space while ensuring the performance of the model), forming multiple minimum hyperspheres, and obtaining a feature library containing only normal features, which is stored locally.

[0076] In the model detection stage, patch svdd divides the input image into many overlapping patches by sliding windows, and then sends each patch to the encoder for feature extraction. When extracting, the input image block p is subdivided into 2×2 sub-patches and passed through a hierarchical encoder. The smaller encoder f is used first. small Extract the features of each sub-block. Then, the aggregate encoder f big To aggregate the overall features of the image block p. The features of each extracted patch are mapped to the feature space, and the nearest neighbor distance of each feature in the feature library is calculated using the KNN algorithm. Since the model has an aggregation effect on similar features, the nearest neighbor distance of normal patch features in the memory library is closer than that of abnormal patch features. Therefore, when detecting each patch, its nearest neighbor distance in the feature library can be used as the abnormality score of the patch.

[0077] For each image, the maximum value of the anomaly scores in the patches it contains can be used as the anomaly score of the image. For each pixel in the image, the average score of all the patches it contains can be used as the anomaly score of the pixel. Based on the anomaly scores of the entire image and each pixel, not only can the abnormal image be identified, but also the abnormal area in the image can be located.

[0078] In S20, the type of the aviation turbine blade to which the enhanced blade neutron image belongs is first determined, and then a pre-trained residual core detection model of the corresponding type is selected to perform residual core detection on the enhanced blade neutron image.

[0079] S30, performing residual core area recognition on the residual core detection result image, and classifying abnormal images with residual cores according to the size of the recognized residual core area and the set level, so as to achieve residual core classification of aviation turbine blades.

[0080] The residual core is usually located in the corner of the inner cavity, and the complex and narrow inner cavity further hinders the leaching of the core. In a specific channel, the residual core may be quite large, which requires multiple effective core leaching during the core removal process. However, if the core leaching time is too long, the casting may be corroded by the core leaching liquid, resulting in an excessively large interface layer or grain boundary corrosion, which will cause the casting to be abandoned. Therefore, quantifying the residual core in the blade cavity to improve the core leaching efficiency is of great significance to production practice. Unlike the defect classification of general industrial products, the residual core has a huge impact on the safe use of aviation turbine blades, and the technical conditions of the blades usually require that there is no residual core. In this regard, according to the size of the residual core in the blade, we mainly divide it into two categories: Major and Critical, where Major refers to blades that require two or less core removals, and Critical refers to blades that require more than two core removals. In this case, the blade core removal efficiency can be significantly improved, thus contributing to industrial production practice.

[0081] The residual core classification can be achieved by building a residual core classification module. The specific implementation process is as follows:

[0082] S31. Convert the input residual core detection result image blade image into a grayscale image, convert the grayscale image into a binary image through a binarization method, and separate the outline of the aviation turbine blade from the background. Due to the inherent noise and artifacts in the neutron image, there is a risk of mistaking these noise and artifacts for the target object during the binarization process. To this end, we perform morphological erosion processing on the obtained binary image to reduce the influence of noise and artifacts, thereby clearly retaining the edge and shape of the aviation turbine blade. Finally, calculate the total number of pixels of the aviation turbine blade in the binary image.

[0083] S32, binarize the abnormal image of the input residual core detection result image, calculate the number of residual core pixel points in the image, calculate the ratio of the total number of aviation turbine blade pixel points to the number of residual core pixel points, and divide the residual core detection result image into a normal blade group and an abnormal blade group. The normal blade group can be marked as Good.

[0084] S33, normalize the abnormal image of the abnormal leaf group, and then use Gaussian smoothing method to process the normalized image to reduce the impact of noise and soften the edge, and then binarize the smoothed image to calculate the number of pixels in the residual core area to determine the size of the residual core area.

[0085] S34. Calculate the ratio of the number of pixels in the residual core area to the total number of pixels in the aviation turbine blade to obtain the proportion of the residual core area in the aviation turbine blade, and determine the residual core grading level of the aviation turbine blade corresponding to the abnormal image with residual core, such as "Major" blade and "Critical" blade, in combination with a pre-set proportion grade classification.

[0086] The thresholds in the above process are determined through a series of parameter tuning experiments, aiming to maximize the segmentation effect and accuracy. During the tuning process, we adopted an iterative approach to continuously adjust the threshold parameters and guide the selection of parameters by evaluating the quality of the segmented image. This process is based on specific evaluation indicators, including but not limited to segmentation accuracy and retention of target features.

[0087] In the experiment, the built model can be used to detect the defective test sample image to determine the position and size of the residual core on the image. The output of the patch svdd model will show the position of the residual core in the image, such as Figure 2 As shown in the output of , the grading module will determine the size of the residual core in the image and display the corresponding grade in the upper right corner of the image. For example, if there is a small piece of residual core in the image, the word "Major" will be displayed in the upper right corner of the image.

[0088] The effects of the present invention are verified by experiments as follows.

[0089] In the experiment, 1272 neutron images of aviation turbine blades were collected to construct a data set. Since these images contain three different types of turbine blades, the original data were divided into three independent data sets, A, B, and C, according to the different characteristics of the three blades, for subsequent training and testing. The processing of image data and data sets and the construction of the model are all carried out according to the aforementioned method of the present invention.

[0090] The evaluation index configuration is as follows: 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 axis and the true positive rate (TPR) as the vertical axis. The ROC curve describes in detail the performance of the classification model under various threshold conditions. Specifically, the definitions of FPR and TPR are as follows:

[0091]

[0092] Among them, TP, TN, FN and FP represent true positive, true negative, false negative and false positive, respectively.

[0093] Ablation experiments were conducted on the aforementioned dataset to verify the effects of neutron image enhancement and encoder model improvement on the accuracy of the residual core detection model. The results are shown in Tables 1 and 2, where the bold fonts are the optimal values ​​in each column. The results show that the detection performance of the model has improved after the introduction of the neutron image enhancement module, with the A, B, and C types improving by an average of 1.3% in both Image-AUC and Pixel-AUC. At the same time, after the encoder model improvement was introduced, the three blades had a significant improvement in Image-AUC and Pixel-AUC compared to the basic model, with an average increase of 4.5% and 3.3% respectively. The effectiveness of the improvement has been fully verified.

[0094]

[0095] Table 1 Image-AUC of ablation experiments

[0096]

[0097] Table 2 Pixel-AUC of ablation experiments

[0098] In the above table, Patch SVDD means that the conventional Patch SVDD model is used as the basic model. ImprovedGamma means that the neutron image enhancement module is introduced, that is, the process of enhancing the neutron image of the blade is added. DRN&ECA means that the encoder model improvement is introduced, that is, the residual core detection model of the improved Patch SVDD model is adopted. Average means the average value of the three data sets ABC.

[0099] SPADE PaDiM Patch SVDD Ours A 98.4% 96.7% 94% 99.1% B 90.3% 87.9% 89.1% 93.2% C 84.1% 73.4% 87.8% 92.1% Average 90.93% 86.0% 90.3% 94.8%

[0100] Table 3 Image-AUC of different models

[0101] SPADE PaDiM Patch SVDD Ours A 64.9% 96% 93.4% 97.3% B 66% 93.5% 90.8% 94.6% C 70.4% 88.3% 92.6% 94.8% Average 67.1% 93.2% 92.3% 95.6%

[0102] Table 4 Pixel-AUC of different models

[0103] Further comparative experiments were conducted to verify the effectiveness and advantages of the residual core detection model of the present invention compared with other anomaly detection models. Tables 3 and 4 list the detection results of different models on the aforementioned established data sets. Among them, Semantic Pyramid Anomaly Detection (SPADE) and Patch Distribution Modeling (PaDiM) both adopted the same statistical-based method as the present invention. It can be seen from Tables 3 and 4 that the method of the present invention is superior to other methods in detection accuracy, and achieves the optimal values ​​in Image-AUC and Pixel-AUC of the three categories. As far as we know, this is because the quality of the neutron image is improved by applying the neutron image enhancement module, so that the model can more accurately capture the details and features in the image. The DRN network effectively retains the subtle features in the image by reducing the downsampling operation, and the residual structure can make full use of shallow and deep information. The ECA attention module further enhances the model's ability to capture key information by focusing on important channel features, making the feature representation richer and more accurate. Therefore, we have achieved advanced performance in residual core detection in neutron images of aviation blades.

[0104] In order to verify the performance of the residual core classification module, the present invention compares the model classification with the manual classification and obtains the following results, as shown in Table 5. Good represents normal blades, Major represents blades that need two or fewer core removals, and Critical represents blades that need more than two core removals. It can be seen from the table that the residual core classification module designed by the present invention can achieve an accuracy rate of 80.8%, which has a high use value.

[0105] grade Total number of samples Number of correct classifications Accuracy Good 214 177 82.7% Major 103 83 80.5% Critical 137 107 78.1% total 454 367 80.8%

[0106] Table 5 Accuracy of residual core classification module

[0107] The above embodiments are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any changes made by adopting the design principles of the present invention and performing non-creative work on this basis should fall within the protection scope of the present invention.

Claims

1. A method for detecting and grading residual cores in neutron images of aviation turbine blades, characterized in that: The following steps are involved: S10, acquiring a neutron image of the aviation turbine blade to be tested, and performing enhancement processing on the blade neutron image; S20, inputting the enhanced blade neutron image into a pre-trained residual core detection model, and outputting residual core detection result images, including two types of normal images without residual cores and abnormal images with residual cores, wherein the residual core detection model is built based on the Patch SVDD model, and the dilated residual network DRN is used instead as the feature extraction network, and an efficient channel attention ECA module is added at the output position of the dilated residual network DRN; S30, performing residual core area recognition on the residual core detection result image, and classifying abnormal images with residual cores according to the size of the recognized residual core area and the set level, so as to achieve residual core classification of aviation turbine blades.

2. The method for detecting and grading residual cores in neutron images of aviation turbine blades according to claim 1, characterized in that: In the S10, a neutron imaging detection method is used to obtain a neutron image of the aviation turbine blade to be tested.

3. The method for detecting and grading residual cores in neutron images of aviation turbine blades according to claim 1, characterized in that: The process of enhancing the leaf neutron image in S10 includes: S11, using the Gamma transformation method to enhance the contrast of the leaf neutron image, and using the Gaussian filtering method based on the spatial mean filter to smooth the leaf neutron image; S12, using the Sobel operator to extract the horizontal edge and the vertical edge of the blade neutron image after smoothing, to obtain high-frequency information of the blade neutron image; S13. Linearly superimposing the extracted high-frequency information with the leaf neutron image enhanced by the Gamma transform method to improve the quality of the leaf neutron image.

4. The method for detecting and grading residual cores in neutron images of aviation turbine blades according to claim 1, characterized in that: The process of constructing the residual core detection model in S20 includes: The Patch SVDD model is selected as the basic model. The input image is divided into blocks through a sliding window method to form a large number of partially overlapping blocks. These blocks are input into the feature extraction network for feature extraction, so that the features extracted from each block correspond to a single point in the feature space. The encoder is trained to find the center of blocks with similar semantics. At the same time, an encoder and a classifier are trained to predict the relative position of two adjacent blocks in space to verify whether the encoder has sufficient semantic information. Then, the degree of abnormality of the block is obtained by calculating the Euclidean distance between the block to be tested and the normal block in the feature space. Using the dilated residual network DRN to replace the original feature extraction network of the Patch SVDD model can capture more complex spatial relationships and shape features and enhance the recognition of residual core shape features. The dilated residual network DRN is composed of dilated convolution introduced on the basis of the ResNet network structure. Then, an efficient channel attention ECA module is added to the output position of the dilated residual network DRN to filter out irrelevant features and reduce the impact of noise. The ECA module uses one-dimensional convolution to achieve cross-channel dimensional interaction, avoid channel dimensionality reduction, and improve classification accuracy.

5. The method for detecting and grading residual cores in neutron images of aviation turbine blades according to claim 4, characterized in that: The final loss function L of the residual core detection model is PatchSVDD It is expressed as: THE PatchSVDD =λL SVDD′ +L SSL L SSL =Cross-entropy(y,C φ (f θ (p1),f θ (p2))) Where λ represents the scaling hyperparameter, p i′ Represents block p i nearby blocks, f is the encoder, f θ (p i ) represents the encoder calculation block p i The obtained eigenvector, f θ (p i′ ) represents the encoder calculation block p i′ The obtained feature vector; p1 is a randomly sampled block, p2 is a block in the neighborhood of p1, y is the relative position between the two blocks, C φ For the classifier, Cross-entropy represents the cross entropy loss.

6. The residual core detection and classification method for neutron images of aviation turbine blades according to claim 5 is characterized in that: The training process of the residual core detection model in S20 includes: S21a, collect existing aviation turbine blade neutron images to build a data set; S21b, using the Labelme tool to accurately label the residual core area in the image with residual core; S21c, the dataset is divided into a training set, a test set and a defect annotation set, where the training set contains only normal image samples without residual cores to simulate the scarcity of residual core samples in natural industrial environments, and the test set contains normal image samples without residual cores and abnormal image samples with residual cores for comprehensive evaluation; S21d, performing enhancement processing on the leaf neutron image in the data set to form an enhanced data set; S21e. Use the enhanced data set to train the constructed residual core detection model to obtain a pre-trained residual core detection model.

7. The residual core detection and classification method for neutron images of aviation turbine blades according to claim 6 is characterized in that: In S21a, when there are different types of aviation turbine blades in the existing aviation turbine blade neutron images, the blade neutron images are first classified according to the types of the aviation turbine blades, so that each type of aviation turbine blade neutron image is independently a data set of its own type; In S21b-S21e, data processing and model training are performed on each type of data set to obtain a pre-trained residual core detection model corresponding to each type of aviation turbine blade; In S20, the type of the aviation turbine blade to which the enhanced blade neutron image belongs is first determined, and then a pre-trained residual core detection model of the corresponding type is selected to perform residual core detection on the enhanced blade neutron image.

8. A method for detecting and grading residual cores in neutron images of aviation turbine blades according to any one of claims 1 to 7, characterized in that: The specific process of S30 includes: S31, converting the blade image of the input residual core detection result image into a grayscale image, converting the grayscale image into a binary image by a binarization method, separating the outline of the aviation turbine blade from the background, and calculating the total number of pixels of the aviation turbine blade in the binary image; S32, binarizing the abnormal image of the input residual core detection result image, calculating the number of residual core pixel points in the image, calculating the ratio of the total number of aviation turbine blade pixel points to the number of residual core pixel points, and dividing the residual core detection result image into a normal blade group and an abnormal blade group; S33, normalizing the abnormal image of the abnormal leaf group, and then using Gaussian smoothing method to process the normalized image to reduce the influence of noise and soften the edge, and then binarizing the smoothed image to calculate the number of pixels in the residual core area to determine the size of the residual core area; S34. Calculate the ratio of the number of pixels in the residual core area to the total number of pixels in the aviation turbine blade to obtain the proportion of the residual core area in the aviation turbine blade, and determine the residual core grading level of the aviation turbine blade corresponding to the abnormal image with the residual core in combination with a pre-set proportion grade classification.

9. The method for detecting and grading residual cores in neutron images of aviation turbine blades according to claim 8, characterized in that: In the S31, morphological erosion processing is performed on the binary image to reduce the influence of noise and artifacts, so as to clearly retain the edge and shape of the aviation turbine blade.

10. The residual core detection and classification method for neutron images of aviation turbine blades according to claim 8, characterized in that: The residual core grading levels of the aviation turbine blades in S34 include two categories: Major and Critical, where Major refers to blades that need to be de-cored twice or less, and Critical refers to blades that need to be de-cored more than twice.

Citation Information

Patent Citations

  • Reference line determination method for neutron detection sensitivity of hollow blade residual core

    CN112525933A

Cited By

  • Turbine blade CT image segmentation and point detection method and system

    CN120852369A