A method for detecting and grading residual cores in neutron images of aero turbine blades

Abstract

The present invention discloses a method for detecting and grading residual cores in neutron images of aero-turbine blades, which includes obtaining the neutron image of the aero-turbine blade to be measured and performing enhancement processing on the neutron image of the blade; inputting the enhanced neutron image of the blade into a pre-trained residual core detection model to output a residual core detection result image, including two types: a normal image without a residual core and an abnormal image with a residual core, wherein the residual core detection model is obtained by training using an improved Patch SVDD model; identifying the residual core region in the residual core detection result image, and classifying the abnormal images with residual cores according to the size of the identified residual core region according to the set grades, so as to realize the grading of the residual cores of the aero-turbine blades. The present invention improves the accuracy of detecting residual cores in neutron images of aero-turbine blades and quantitatively evaluates the detection results, which helps to improve the core removal efficiency in industrial production practice.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engine blade detection. Specifically, it relates to a method for detecting and grading residual cores in neutron images of aero-turbine blades. Background Art

[0002] As the main power source of an aircraft, engine blades operate in extreme environments of high temperature, high stress, and high rotational speed. Due to the presence of features such as bends, blind holes, and micro-cavities in the internal channel structure of the blades, it is easy to generate 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 disrupt the engine balance and even cause accidents. Therefore, all produced engine turbine blades must be strictly inspected before being used in aero-engines to confirm that there are no residual cores inside to ensure the engine quality and aviation flight safety.

[0003] In recent years, a variety of non-destructive testing techniques (NDT) have been applied to the detection of residual cores in turbine blades, aiming to accurately identify the internal residual cores without damaging the structural integrity of aero-engine turbine blades. These techniques include eddy current testing, endoscope testing, industrial CT testing, X-ray testing, and neutron imaging testing methods. The above techniques mainly rely on experienced inspectors to judge the later imaging results, but it is inevitable to have misjudgments due to human fatigue during the manual judgment process, such as false positive phenomena (that is, there are residual cores in aero-turbine blades, but they are misjudged as normal by the inspector), and there are also problems such as low efficiency and long time consumption in manual judgment.

[0004] Due to the rapid development of deep learning, many researchers have begun to apply deep learning-based methods to the task of residual core detection. Some studies use the YOLO algorithm to detect residual cores in endoscope images of aero-turbine blades, which improves the detection efficiency. However, this supervised learning method relies on a large amount of labeled information, and endoscope testing is not applicable to blades with complex internal structures, tortuous channels, and pore diameters less than 0.5 mm. Therefore, new methods for detecting residual cores need to be developed. Summary of the Invention

[0005] According to the problems existing in the prior art, the present invention provides a method for detecting and grading residual cores in neutron images of aero-turbine blades, which can realize the detection, positioning, and quantitative grading of residual cores without relying on labeled information, and improve the core removal efficiency in production practice.

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

[0007] A method for detecting and grading residual cores in neutron images of aero-turbine blades, comprising the following steps:

[0008] S10. Obtain the neutron image of the aero-turbine blade to be measured, and perform enhancement processing on the neutron image of the blade;

[0009] S20. Input the enhanced neutron image of the blade into a pre-trained residual core detection model, and output a residual core detection result image, including two types: a normal image without a residual core and an abnormal image with a residual core. The residual core detection model is constructed based on the Patch SVDD model, uses the dilated residual network DRN as the feature extraction network instead, and adds an efficient channel attention ECA module at the output position of the dilated residual network DRN;

[0010] S30. Identify the residual core region in the residual core detection result image, and classify the abnormal images with residual cores according to the size of the identified residual core region according to the set grades, so as to realize the grading of the residual cores of the aero-turbine blades.

[0011] Specifically, in S10, the neutron imaging detection method is used to obtain the neutron image of the aero-turbine blade to be measured.

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

[0013] S11. Use the Gamma transformation method to enhance the contrast of the neutron image of the blade, and at the same time use the Gaussian filtering method based on the spatial mean filter to smooth the neutron image of the blade;

[0014] S12. Use the Sobel operator to extract the horizontal edge and vertical edge of the smoothed neutron image of the blade to obtain the high-frequency information of the neutron image of the blade;

[0015] S13. Linearly superimpose the extracted high-frequency information and the neutron image of the blade enhanced by the Gamma transformation method to improve the quality of the neutron image of the blade.

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

[0017] Select the Patch SVDD model as the basic model. Through the sliding window method, the input image is divided into blocks, forming many 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. Train the encoder to find the center of the blocks with similar semantics, and at the same time train an encoder and a classifier to predict the relative positions of two adjacent blocks in space to verify whether the encoder has mastered sufficient semantic information. Then, by calculating the Euclidean distance between the block to be measured and the normal block in the feature space, the degree of abnormality of the block is obtained;

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

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

[0020] Specifically, the final loss function L of the residual core detection model PatchSVDD 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 the blocks near block p i , f is the encoder, f θ (p i ) represents the feature vector calculated by the encoder for block p i , f θ (p i′ ) represents the feature vector calculated by the encoder for block p i′ ; p1 is a randomly sampled block, p2 is a block within the neighborhood of p1, y is the relative position between the two blocks, C φ is the classifier, and 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 neutron images of aero-engine turbine blades to construct a dataset;

[0027] S21b. Use the Labelme tool to accurately label the residual core area in the images with residual cores;

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

[0029] S21d. Enhance the neutron images of the blades in the dataset to form an enhanced dataset;

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

[0031] Further, in S21a, when there are different types of aero-engine turbine blades in the existing neutron images of aero-engine turbine blades, classify the neutron images of the blades according to the type of aero-engine turbine blade first, so that the neutron images of each type of aero-engine turbine blade are independent datasets of their respective types;

[0032] In S21b - S21e, perform data processing and model training on each type of dataset respectively to obtain a pre-trained residual core detection model corresponding to each type of aero-engine turbine blade;

[0033] In S20, first determine the type of aero-engine turbine blade to which the enhanced neutron image of the blade belongs, and then select the pre-trained residual core detection model of the corresponding type to perform residual core detection on the enhanced neutron image of the blade.

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

[0035] S31. Convert the blade image of the input residual core detection result image into a grayscale image, convert the grayscale image into a binary image through a binarization method, separate the aero-engine turbine blade contour from the background, and calculate the total number of pixel points of the aero-engine turbine blade in the binary image;

[0036] S32. Binarize the abnormal images in the input residual core detection result images, calculate the number of residual core pixel points in the images, calculate the ratio of the total number of aviation turbine blade pixel points to the number of the residual core pixel points, and divide the residual core detection result images into a normal blade group and an abnormal blade group;

[0037] S33. Normalize the abnormal images in the abnormal blade group, then process the normalized images using the Gaussian smoothing method to reduce the influence of noise and soften the edges, then binarize the smoothed images, calculate the number of pixel points in the residual core area, and determine the size of the residual core area;

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

[0039] Among them, in S31, morphological erosion processing is performed on the binary image to reduce the influence of noise and artifacts, so as to clearly retain the edges and shape of the aviation turbine blade.

[0040] Specifically, the residual core classification levels of the aviation turbine blade in S34 include two categories: Major and Critical. Among them, Major refers to the blade that needs to be removed from the core two times or less, and Critical refers to the blade that needs to be removed from the core more than two times.

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

[0042] (1) The present invention enhances the sub-images of the aviation turbine blade to be measured, improves the image contrast and its detail clarity, provides convenience for subsequent residual core detection, and trains a residual core detection model by fusing DRN and ECA into the Patch SVDD model, significantly strengthening the model's recognition ability of residual core features, improving the accuracy of residual core detection, and also improving the core removal efficiency in industrial production practice by adding a process of residual core classification to quantitatively evaluate the residual cores identified in the images in detail.

[0043] (2) The present invention adopts an improved gamma transformation algorithm in the image enhancement processing, effectively improving the contrast of the sub-images of 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 resulting core detection model can more effectively capture and learn rich and accurate semantic information, significantly improving the model's ability to identify core features, thereby achieving more accurate core detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a schematic flowchart of an embodiment of the present invention.

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

[0047] Figure 3 is a schematic flowchart of the sub-image enhancement process of the blade in an embodiment of the present invention.

[0048] Figure 4 is a schematic flowchart of the core grading process in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

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

[0050] Embodiment

[0051] As Figures 1 to 4 shown, the method for core detection and grading of sub-images of aero-engine turbine blades includes the following steps:

[0052] S10. Obtain the neutron image of the aero-engine turbine blade to be measured by using the neutron imaging detection method. For example, irradiate the aero-engine turbine blade with a parallel cold neutron beam to obtain the neutron image of the aero-engine turbine blade. The quality of the neutron image is subject to the influence of neutron imaging equipment, neutron source intensity, imaging system accuracy, and physical factors, resulting in degraded phenomena such as low contrast, noise, artifacts, and blurred edges in the imaging result. The existence of these phenomena brings great interference to 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 detection, the present invention also performs enhancement processing on the neutron image of the blade.

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

[0054] S11. Use the Gamma transformation method to enhance the contrast of the neutron image of the blade, and at the same time use the Gaussian filtering method based on the spatial mean filter to smooth the neutron image of the blade, reducing the influence of noise on the subsequent edge extraction result.

[0055] S12. Use the Sobel operator to extract the horizontal and vertical edges of the neutron image of the blade after smoothing processing, and obtain the high-frequency information of the neutron image of the blade.

[0056] S13. Linearly superimpose the extracted high-frequency information on the neutron image of the blade enhanced by the Gamma transformation method to further improve the quality of the neutron image of the blade and realize the enhancement of the neutron image of the aero-turbine blade.

[0057] S20. Input the enhanced neutron image of the blade into a pre-trained core defect detection model, and output a core defect detection result image, including two types: a normal image without core defects and an abnormal image with core defects.

[0058] The core defect detection model is constructed based on the Patch SVDD model. Specifically:

[0059] Select the Patch SVDD model as the basic model. Through the sliding window method, the input image is divided into blocks, forming many 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. Train the encoder to find the center of the blocks with similar semantics, and at the same time train an encoder and a classifier to predict the relative positions of two adjacent blocks in the space, verify whether the encoder has mastered sufficient semantic information, and then calculate the Euclidean distance between the block to be tested and the normal block in the feature space to obtain the degree of abnormality of the block.

[0060] Use the dilated residual network DRN to replace the original feature extraction network of the Patch SVDD model, which can capture more complex spatial relationships and shape features, and enhance the recognition of core defect shape features. The dilated residual network DRN is composed of introducing dilated convolutions on the basis of the ResNet network structure. The introduction of dilated convolutions enables the DRN network to maintain the output size consistent with the input while achieving the same receptive field as the original network, and there is no need for pooling operations, making it particularly suitable for the detection and localization tasks of core defects. The residual structure therein can effectively propagate low-level features to high-level, thereby realizing the full utilization of context details and solving problems such as gradient disappearance and model degradation. The introduction of DRN makes the model no longer only focus on texture features and ignore shape features. The deepening of the network depth enables the model to capture more abstract semantic information and has a stronger ability to learn the expression of complex and fine features.

[0061] Then, an Efficient Channel Attention (ECA) module is added at the output position of the Dilated Residual Network (DRN) to effectively filter out irrelevant features, reduce the impact of noise, 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, avoiding channel dimensionality reduction and improving the classification accuracy. ECA is a lightweight plug-and-play module that allows the network to focus on the target to be detected, correspondingly highlighting the important features, and thus improving the accuracy of the network model for target detection.

[0062] Specifically, the final loss function L of the residual core detection model PatchSVDD 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 the block near block p i , f is the encoder, f θ (p i ) represents the feature vector calculated by the encoder for block p i , f θ (p i′ ) represents the feature vector calculated by the encoder for block p i′ ; p1 is a randomly sampled block, p2 is a block within the neighborhood of p1, y is the relative position between the two blocks, C φ is the classifier, and Cross-entropy represents the cross-entropy loss.

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

[0068] S21a. Collect existing neutron images of aero turbine blades to construct a dataset.

[0069] S21b. Use the Labelme tool to accurately label the residual core area in the images with residual cores for subsequent analysis.

[0070] S21c. The data set format refers to the public data set MVTec-AD. The data set is divided into a training set, a test set, and a ground truth defect annotation set. The training set only contains normal image samples without core remnants to simulate the scarcity of core remnant samples in the natural industrial environment. The test set contains normal image samples without core remnants and abnormal image samples with core remnants for comprehensive evaluation.

[0071] S21d. Enhance the neutron images of the blades in the data set to form an enhanced data set. The enhancement process uses the enhancement method in S10 and can be achieved by building a neutron image enhancement module.

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

[0073] Furthermore, when collecting existing neutron images of aero-engine turbine blades in S21a, there are usually various types of aero-engine turbine blades. To avoid the mutual influence of different types of aero-engine turbine blades on the core remnant detection results, we first classify the neutron images of the blades according to the types of aero-engine turbine blades, so that the neutron images of each type of aero-engine turbine blade are independent as data sets of their respective types. For example, 1272 neutron images of aero-engine turbine blades are collected in the experiment. These images contain three different types of aero-engine turbine blades. Based on the different characteristics of these three types of blades, they are divided into three categories, A, B, and C in advance, for subsequent accurate model training and core remnant detection.

[0074] Thus, through the process of S21b - S21e, it can be achieved that data processing and model training are respectively performed on data sets of each type, and a pre-trained core remnant detection model corresponding to each type of aero-engine turbine blade is obtained.

[0075] In the model training stage, patch svdd divides the input image into many overlapping patches by means of a sliding window, and then sends each patch into the encoder for feature extraction. When extracting, the input image patch p is subdivided into 2×2 sub-patches through a hierarchical encoder. First, a smaller encoder f small is used to extract the features of each sub-block. Then, an aggregation encoder f big is used to aggregate the overall features of the image patch p. The features of each extracted patch are mapped into the feature space (the process of feature extraction mapping will be continuously optimized through the loss function, while keeping the distance of normal patches in the feature space as close as possible and ensuring the performance of the model), forming multiple smallest hyperspheres, obtaining a feature library containing only normal features, and storing it locally.

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

[0077] For each image, the maximum value among the anomaly scores of the patches it contains can be used as the anomaly score of the image. For each pixel in the image, the average value of the scores of all the patches it is included in can be used as the anomaly score of the pixel. According to the anomaly scores of the entire image and each pixel, not only can abnormal images be identified, but also the abnormal regions in the images can be located.

[0078] In S20, first determine the type of the aero-turbine blade to which the sub-image in the enhanced blade belongs, and then select the pre-trained core detection model of the corresponding type to perform core detection on the sub-image in the enhanced blade.

[0079] S30. Identify the core region in the core detection result image, and classify the abnormal images with cores according to the size of the identified core region according to the set levels, so as to realize the core grading of the aero-turbine blades.

[0080] The remaining cores are usually located in the corners of the inner cavity, and the complex and narrow inner cavity further hinders the leaching of the cores. In specific channels, there may be quite a lot of remaining cores, 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 solution, resulting in an excessive interfacial layer or intergranular corrosion, thus leading to the rejection of the casting. Therefore, quantifying the remaining cores in the blade inner cavity to improve the core leaching efficiency is of great significance to production practice. Different from the defect grading of general industrial products, the remaining cores have a huge impact on the safe use of aero-turbine blades, and the technical conditions of the blades usually require no remaining cores. In this regard, according to the size of the remaining cores in the blades, we mainly classify them into two categories: Major and Critical, where Major refers to the blades that require core removal two times or less, and Critical refers to the blades that require core removal more than two times. In this case, the core removal efficiency of the blades can be significantly improved, thus contributing to industrial production practice.

[0081] The remaining core grading can be achieved by building a remaining core grading module, and its specific implementation process is as follows:

[0082] S31. Convert the input image of the remaining core detection result (blade image) into a grayscale image, and then convert the grayscale image into a binary image through a binarization method to separate the aero-turbine blade contour from the background. Due to the inherent noise and artifacts in the neutron image, there is a risk of misidentifying these noises and artifacts as target objects during the binarization process. In this regard, we perform morphological erosion processing on the obtained binary image to reduce the influence of noise and artifacts, so as to clearly retain the edges and shape of the aero-turbine blade. Finally, calculate the total number of pixel points of the aero-turbine blade in the binary image.

[0083] S32. Perform binarization processing on the abnormal image of the input remaining core detection result image, calculate the number of pixel points of the remaining core in the image, calculate the ratio of the total number of pixel points of the aero-turbine blade to the number of pixel points of the remaining core, and divide the remaining core detection result image into a normal blade group and an abnormal blade group. Among them, the normal blade group can be marked as Good.

[0084] S33. Perform normalization processing on the abnormal image of the abnormal blade group, then use the Gaussian smoothing method to process the normalized image to reduce the influence of noise and soften the edges, then perform binarization processing on the smoothed image, calculate the number of pixel points in the remaining core area, and determine the size of the remaining core area.

[0085] S34. Calculate the ratio of the number of pixel points in the remaining core area to the total number of pixel points of the aero-turbine blade, obtain the proportion of the remaining core area in the aero-turbine blade, and combine the pre-set proportion grade classification to determine the remaining core grading level of the aero-turbine blade corresponding to the abnormal image with remaining cores, such as "Major" blades and "Critical" blades.

[0086] The threshold in the above process is determined through a series of parameter tuning experiments aimed at maximizing the segmentation effect and accuracy. During the tuning process, we adopted an iterative method, continuously adjusted the threshold parameters, and guided the parameter selection by evaluating the quality of the segmented images. This process is based on specific evaluation metrics, including but not limited to segmentation accuracy and the retention of target features, etc.

[0087] In the experiment, the established model can be used to detect defective test sample images to determine the position and size of the residual core in the image. The output of the patch svdd model will show the position of the residual core in the image, as Figure 2 shown in the output. The grading module will judge 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 residual core in the image, the word "Major" will be displayed in the upper right corner of the image.

[0088] The following verifies the effect of the present invention through experiments.

[0089] In the experiment, 1272 sub-images of aero turbine blades were collected to construct a data set. Since these images contain three different types of turbine blades, according to the different characteristics of the three types of blades, the original data was divided into three independent data sets, namely A, B, and C, for subsequent separate training and detection. The processing of the image data and data sets and the construction of the model were all carried out according to the foregoing method process of the present invention.

[0090] The evaluation metric is configured as: the area under the receiver operating characteristic curve (ROC) (AUC) is selected as the quantitative evaluation metric for the model performance. The ROC curve is an image plotted with the false positive rate (FPR) as the abscissa and the true positive rate (TPR) as the ordinate. The ROC curve depicts 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] An ablation experiment was carried out on the established dataset above to verify the influence 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 font represents the optimal value in each column. The results show that the detection performance of the model has been improved after introducing the neutron image enhancement module. Among them, the types A, B, and C have an average increase of 1.3% in both Image-AUC and Pixel-AUC. At the same time, after introducing the encoder model improvement, the three types of blades have a greater improvement in Image-AUC and Pixel-AUC compared with the basic model, with an average increase of 4.5% and 3.3%. This fully verifies the effectiveness of the improvement.

[0094]

[0095] Table 1 Image-AUC of the ablation experiment

[0096]

[0097] Table 2 Pixel-AUC of the ablation experiment

[0098] In the above table, Patch SVDD represents the use of the conventional Patch SVDD model as the basic model. ImprovedGamma represents the introduction of the neutron image enhancement module, that is, adding a process of enhancing the neutron image of the blade. DRN&ECA represents the introduction of the encoder model improvement, that is, using the residual core detection model that improves the Patch SVDD model. Average represents the average value of the three datasets A, B, and C.

[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 carried out 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 previously established data set. Among them, both Semantic Pyramid Anomaly Detection (SPADE) and Patch Distribution Modeling (PaDiM) adopt the same statistics-based method as the present invention. It can be seen from Tables 3 and 4 that the method of the present invention is more excellent than other methods in terms of detection accuracy, and reaches the optimal value in Image-AUC and Pixel-AUC of 3 categories. As far as we know, this is because by applying the neutron image enhancement module to improve the quality of neutron images, the model can capture the details and features in the images more accurately. 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 the information of the shallow layer and the deep layer. 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 obtained advanced performance in the detection of residual cores in neutron images of aviation blades.

[0104] To verify the performance of the residual core grading module, the present invention will compare the grading results of the model with manual grading and obtain the following results, as shown in Table 5. Among them, Good represents normal blades, Major represents blades that require core removal twice or less, and Critical represents blades that require core removal more than twice. It can be seen from the table that the residual core grading module designed by the present invention can achieve an accuracy rate of 80.8%, which has high application value.

[0105] Level Total number of samples Number of correctly classified Accuracy rate 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 the residual core grading module

[0107] The above embodiments are only the preferred embodiments of the present invention, and do not limit the protection scope of the present invention. Any changes made by using the design principle of the present invention and non-creative labor on this basis shall fall within the protection scope of the present invention.

Claims

1. A method for detecting and grading residual cores in neutron images of aero turbine blades, characterized in that It includes the following steps: S10. Obtain the neutron image of the aviation turbine blade to be measured, and enhance the neutron image of the blade: S11. Use the Gamma transformation method to enhance the contrast of the neutron image of the blade, and at the same time use the Gaussian filtering method based on the spatial mean filter to smooth the neutron image of the blade; S12. Use the Sobel operator to extract the horizontal edge and vertical edge of the smoothed neutron image of the blade to obtain the high-frequency information of the neutron image of the blade; S13. Linearly superimpose the extracted high-frequency information and the neutron image of the blade enhanced by the Gamma transformation method to improve the quality of the neutron image of the blade; S20. Input the enhanced neutron image of the blade into the pre-trained core residue detection model, and output the core residue detection result image, including two types: a normal image without core residue and an abnormal image with core residue. The core residue detection model is constructed based on the Patch SVDD model, uses the dilated residual network DRN instead as the feature extraction network, and adds an efficient channel attention ECA module at the output position of the dilated residual network DRN; The construction process of the core residue detection model includes: Select the Patch SVDD model as the basic model. Through the sliding window method, divide the input image into blocks to form many partially overlapping blocks. Input these blocks 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. Train the encoder to find the center of the blocks with similar semantics, and at the same time train an encoder and a classifier to predict the relative positions of two adjacent blocks in space to verify whether the encoder has mastered sufficient semantic information. Then, calculate the Euclidean distance between the block to be measured and the normal block in the feature space to obtain the degree of abnormality of the block; Use the dilated residual network DRN to replace the original feature extraction network of the Patch SVDD model, which can capture more complex spatial relationships and shape features, and enhance the recognition of the core residue shape features. The dilated residual network DRN is composed of introducing dilated convolution on the basis of the ResNet network structure; Then add an efficient channel attention ECA module at the output position of the dilated residual network DRN to filter out irrelevant features and reduce the influence of noise. The ECA module uses one-dimensional convolution to achieve cross-channel dimensional interaction, avoids channel dimensionality reduction, and improves the classification accuracy; S30. Identify the core residue area in the core residue detection result image, and classify the abnormal images with core residue according to the size of the identified core residue area according to the set levels to achieve the core residue grading of the aviation turbine blade.

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

3. A method for detecting and grading residual cores in neutron images of aero turbine blades according to claim 1, characterized in that, The final loss function L of the residual core detection model PatchSVDD is expressed as: L 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 the block near block p i , f is the encoder, and fθ(p i ) represents the feature vector calculated by the encoder for block p i ; f θ (p i′ ) represents the feature vector calculated by the encoder for block p i′ ; p1 is a randomly sampled block, p2 is a block within the neighborhood of p1, y is the relative position between the two blocks, and C φ is the classifier, and Cross - entropy represents the cross - entropy loss.

4. A method for detecting and grading residual cores in neutron images of aero turbine blades according to claim 3, characterized in that, The training process of the core residue detection model in S20 includes: S21a. Collect the existing neutron images of aviation turbine blades to construct a data set; S21b. Use the Labelme tool to accurately label the residual core area in the image with core residue; S21c. Divide the dataset into a training set, a test set, and a defect annotation set, where the training set only contains normal image samples without residual cores to simulate the scarcity of residual core samples in the natural industrial environment, and the test set contains normal image samples without residual cores and abnormal image samples with residual cores for comprehensive evaluation; S21d. Enhance the blade neutron images in the dataset to form an enhanced dataset; S21e. Use the enhanced dataset to train the constructed residual core detection model to obtain a pre-trained residual core detection model.

5. A method for detecting and grading residual cores in neutron images of aero turbine blades according to claim 4, characterized in that, In S21a, when there are different types of aero-engine turbine blade neutron images in the existing ones, classify the blade neutron images according to the type of aero-engine turbine blade first, so that the neutron images of each type of aero-engine turbine blade are independent as datasets of their respective types; In S21b - S21e, perform data processing and model training on each type of dataset respectively to obtain a pre-trained residual core detection model corresponding to each type of aero-engine turbine blade; In S20, first determine the type of aero-engine turbine blade to which the enhanced blade neutron image belongs, and then select the pre-trained residual core detection model of the corresponding type to perform residual core detection on the enhanced blade neutron image.

6. A method for detecting and grading residual cores in neutron images of aero turbine blades according to any one of claims 1-5, characterized in that, The specific process of S30 includes: S31. Convert the blade image of the input residual core detection result image into a grayscale image, convert the grayscale image into a binary image through a binarization method, separate the aero-engine turbine blade contour from the background, and calculate the total number of pixel points of the aero-engine turbine blade in the binary image; S32. Perform binarization processing on the abnormal image of the input residual core detection result image, calculate the number of pixel points of the residual core in the image, calculate the ratio of the total number of pixel points of the aero-engine turbine blade to the number of pixel points of the residual core, and divide the residual core detection result image into a normal blade group and an abnormal blade group; S33. Normalize the abnormal image of the abnormal blade group, then use the Gaussian smoothing method to process the normalized image to reduce the influence of noise and soften the edges, then perform binarization processing on the smoothed image, calculate the number of pixel points of the residual core area, and determine the size of the residual core area; S34. Calculate the ratio of the number of pixel points of the residual core area to the total number of pixel points of the aero-engine turbine blade to obtain the proportion of the residual core area in the aero-engine turbine blade, and combine the pre-set proportion level classification to determine the residual core grading level of the aero-engine turbine blade corresponding to the abnormal image with a residual core.

7. A method for detecting and grading residual cores in neutron images of aero turbine blades according to claim 6, characterized in that, In S31, perform morphological erosion processing on the binary image to reduce the influence of noise and artifacts, so as to clearly retain the edge and shape of the aero-engine turbine blade.

8. A method for detecting and grading residual cores in neutron images of aero turbine blades according to claim 6, characterized in that, The residual core grading levels of the aero-engine turbine blade in S34 include two categories: Major and Critical. Among them, Major refers to the blade that needs to be removed from the core two times or less, and Critical refers to the blade that needs to be removed from the core more than two times.

Images