Aero-engine blade defect detection method in few-sample scene

Abstract

The invention belongs to the technical field of computer vision and industrial detection, and particularly discloses an aero-engine blade defect detection method in a few-sample scene. The method comprises five steps of data enhancement, image down-sampling, feature learning, model training and model reasoning. Wherein the data enhancement comprises three sub-steps of defect region extraction, graph self-adaptive adjustment and defect sample synthesis; the image down-sampling comprises two sub-steps of wave basis adaptive selection and wavelet transform down-sampling; the feature learning comprises two sub-steps of weighted feature fusion and distance loss normalization. According to the method, the realization difficulty of a visual detection method in an actual industrial scene is deeply analyzed, and the complete aero-engine blade defect detection method facing the sample scarcity problem is provided. The method is suitable for automatic defect detection in the blade production process, the blade quality detection precision and efficiency can be remarkably improved, the manual reinspection cost is reduced, the automation and digitization level in the production process is expected to be further improved, and the production efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and industrial inspection, and particularly relates to a method for detecting defects of aero-engine blades in a few-shot scenario. Background Art

[0002] Aero-engine blades are crucial components in an aero propulsion system, and their performance is directly related to the efficiency and operation safety of the engine. The working environment of the blades is extremely harsh, and factors such as high temperature, high pressure, and strong vibration may cause various defects in the blades during operation. If these defects are not detected in time, they may seriously affect the overall performance and safety of the engine, and even lead to catastrophic failures. Therefore, defect detection in the early stage is crucial for ensuring manufacturing quality, which can not only effectively extend the service life of the engine but also reduce the maintenance cost. Traditional manual visual inspection methods are widely used in blade defect detection, but they have significant limitations, mainly manifested as low efficiency, poor accuracy, and susceptibility to subjective factors in manual inspection. In addition, it is difficult for manual inspection to effectively identify subtle defects in a complex environment, unable to meet the requirements of high-efficiency and precise detection in modern aero-engine blade manufacturing.

[0003] In recent years, vision detection methods based on deep learning have become a research hotspot in aero-engine blade defect detection. Deep learning technology, especially convolutional neural networks (CNNs), can automatically extract complex features in images and perform efficient defect recognition. Compared with traditional manual visual inspection methods, deep learning networks can automatically learn key features in images through training based on a large amount of labeled data, improving the accuracy and efficiency of detection. Using deep learning for surface defect detection of blades can not only automatically and real-time identify multiple defect types but also accurately judge subtle defects in a complex environment. The application of this method not only improves the detection accuracy but also significantly reduces the labor intensity of manual inspection, promoting the improvement of the quality of aero-engine blades and the increase of production efficiency.

[0004] However, despite the significant progress made by deep learning in blade surface defect detection, there are still many challenges in practical industrial applications. Among them, the scarcity of defect samples and the problem of class imbalance are particularly prominent. Due to the relatively low production volume of aeroengine blades and the strong customization characteristics of each blade, the available defect samples are very limited. At the same time, due to the low frequency of defect occurrence, the ratio of positive samples to negative samples is seriously imbalanced, resulting in the imbalance of training data and further affecting the performance of the model. In addition, deep learning models usually require a large amount of labeled data for effective training, while in the actual production environment, the cost of obtaining a large amount of labeled data is very high. Facing these problems, it is particularly important to study how to perform efficient blade defect detection in the few-shot scenario. Therefore, conducting research on the method of aeroengine blade defect detection for few-shot learning can not only effectively solve the problems of insufficient defect samples and class imbalance, but also promote the wide application of deep learning technology in actual industrial scenarios. Summary of the Invention

[0005] The purpose of the present invention is to: in the process of aeroengine quality inspection, in order to reduce the impact brought by sample shortage and class imbalance, and improve the visual detection accuracy through means such as data augmentation and network structure optimization, the present invention provides a method for aeroengine blade defect detection in a few-shot scenario.

[0006] The present invention is mainly realized through the following technical solutions: a method for aeroengine blade defect detection in a few-shot scenario, characterized by including the following steps:

[0007] 1. Data augmentation. It mainly includes three sub-steps: defect area extraction, graphic adaptive adjustment, and defect sample synthesis.

[0008] 1.1 Defect area extraction. Defect area extraction is to copy and extract the key area information of defect samples. The key area information includes the vertex coordinates of the circumscribed polygon of the blade defect mask and the color information of each pixel inside the polygon.

[0009] 1.2 Graphic adaptive adjustment. Graphic adaptive adjustment is to adjust the shape and color of the extracted defect area graphic. By applying color space transformation, a high degree of consistency between the defect area and the background color tone of the target normal sample image is achieved.

[0010] 1.3 Defect sample synthesis. Defect sample synthesis completes the determination of the pasting position and posture of the defect graphic and the pasting synthesis of the defect graphic. According to the edge direction of the blade contour in the normal blade image, the curve fitting algorithm is used to calculate the matching degree between the blade and the defect edge, ensuring the natural fusion between the defect and the blade contour.

[0011] 2. Image downsampling. It mainly includes two sub-steps: wavelet basis adaptive selection and wavelet transform downsampling.

[0012] 2.1 Wavelet basis adaptive selection. Wavelet basis adaptive selection automatically selects a wavelet basis for subsequent processing according to the characteristics of the input image. By analyzing the frequency characteristics and texture characteristics of the image through frequency domain analysis and statistical analysis, and then obtaining a comprehensive score through weighted summation, this score is used to guide the selection of the wavelet basis that best suits the current input image tensor characteristics.

[0013] 2.2 Wavelet transform downsampling. Wavelet transform downsampling performs wavelet transform on the input image, reducing the model's computational amount while retaining the detailed information of the image. After selecting the wavelet basis, the input feature map is decomposed into four sub-images using wavelet transform, and each sub-image retains one-fourth of the original dimension. The decomposition process retains the key spatial-frequency relationship and achieves lossless feature encoding at the same time.

[0014] 3. Feature learning. It mainly includes two sub-steps: weighted feature fusion and distance loss normalization.

[0015] 3.1 Weighted feature fusion. Weighted feature fusion adaptively adds the input components with weights. According to the task-specific requirements and dataset characteristics, the feature fusion is adjusted adaptively, so as to amplify the expression of key features and enhance the feature quality and effect.

[0016] 3.2 Distance loss normalization. A new loss function is introduced to replace the original loss function of the model. By modeling the bounding box as a two-dimensional Gaussian distribution and using the Wasserstein distance to calculate the similarity between these distributions.

[0017] 4. Model training. Model training iteratively updates the network weights through an optimization algorithm to minimize the loss function. Calculate the predicted output of each round through forward propagation, then compare it with the true label, calculate the error, and adjust the network weights through the backpropagation algorithm. The batch gradient descent method is used during the training process, and the performance on the validation set is calculated regularly to avoid overfitting and perform early stopping strategies.

[0018] 5. Model inference. Model inference is the process of predicting new input data using the trained weights. In the inference stage, the input data is calculated through forward propagation to obtain the final output. During this process, the weights of the network remain fixed, only calculations are performed without updates. During the inference process, the input features are processed layer by layer, and after a series of operations such as activation functions, convolutional layers, and pooling layers, the prediction result of the model and the blade defect detection result are finally obtained.

[0019] Advantages of the present invention: The present invention deeply analyzes the difficulties in the defect detection of aero-engine blades in actual industrial scenarios, and proposes a complete defect detection method for aero-engine blades in a few-shot scenario, which can effectively address the problems of scarce samples and class imbalance, and significantly improve the detection accuracy in the defect detection of aero-engine blades. This method is applicable to the high-precision surface quality detection of aero-engine blades, improves the detection automation level and work efficiency, and provides strong support for quality assurance in the aviation field. Brief Description of the Drawings

[0020] Figure 1 is a flowchart of the aero-engine blade defect detection method of the present invention.

[0021] Figure 2 is a schematic diagram of the aero-engine blade defect sample synthesis method of the present invention.

[0022] Figure 3 is a schematic diagram of the detection result of the corrosion defect of the aero-engine blade of the present invention.

[0023] Figure 4 is a flowchart of the aero-engine blade defect detection feature fusion algorithm of the present invention.

[0024] Figure 5 is a schematic diagram of the detection result of the breakage defect of the aero-engine blade of the present invention.

[0025] Figure 6 is a schematic diagram of the detection result of the fracture defect of the aero-engine blade of the present invention. Detailed Embodiment

[0026] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0027] The equipment used is a six-degree-of-freedom collaborative robotic arm with a maximum working radius of 886.5 mm, and a camera with a field of view angle of 15 degrees, a resolution of 1000*1000, and a depth of field range of 10 - 1200 mm. The implementation of the present invention includes the following steps ( Figure 1 ):

[0028] 1.1 Defect area extraction. With the help of OpenCV, the defect area of the original defect sample image is extracted, the defect area in the image is extracted using the blade defect mask, and the defect area is represented by the vertex coordinates of the circumscribed polygon of the polygon. This polygon includes all defect pixel information, and the pixel color information inside each polygon is recorded and extracted one by one to provide the required data for the subsequent steps of processing.

[0029] 1.2 Graphic Adaptive Adjustment. Adaptive adjustment is performed on defective graphics. The color space transformation method is adopted to adjust the color and shape of the defective area to make the hue between the defective area and the background image more matched, so as to ensure the naturalness and consistency of the final image.

[0030] 1.3 Defective Sample Synthesis. As Figure 2 shown, through the curve fitting algorithm (such as B-spline curve), the appropriate defective pasting position is determined, and the morphology of the defective area is ensured to be consistent with the natural contour of the blade. Then, using the pasting algorithm, the defective area is seamlessly synthesized with the blade image, and the posture and angle of the defect are adjusted to make it naturally integrated with the structure of the blade, avoiding splicing marks and obvious visual incoordination.

[0031] 2.1 Wavelet Basis Adaptive Selection. The wavelet selection module performs frequency domain analysis on the spectral data after Fourier transform, and systematically evaluates the energy distribution of high-frequency and low-frequency spectral components. This process quantifies the spectral characteristics of the input tensor by calculating the high-frequency to low-frequency energy ratio, and this ratio is used as the frequency feature score. The quantitative formula for energy distribution analysis is expressed as:

[0032]

[0033]

[0034] where I(x, y) is the pixel value in the image spatial domain, and M and N are the width and height of the image respectively.

[0035] By using the metrics derived from the gray-level co-occurrence matrix (GLCM), including contrast, energy, uniformity, and entropy, statistical analysis is performed on local image regions to evaluate texture features simultaneously. The calculation methods of these quantitative metrics are as follows:

[0036] S Con =∑ i,j (i - j) 2 ·P(i, j)

[0037] S Ene =∑ i,j P(i, j) 2

[0038]

[0039] S Ent =-∑ i,j P(i, j)·log(P(i, j))

[0040]

[0041] where the composite texture score Score texObtained by the weighted summation of the scores of each texture index, where W represents the weight coefficient of each statistical parameter.

[0042] 2.2 Wavelet transform downsampling. After comprehensively quantifying the frequency features and texture features in the input tensor, the optimal wavelet basis is adaptively selected according to the signal features. During the wavelet selection process, the comprehensive frequency-texture score γ is used as a key hyperparameter and is systematically optimized through iterative experiments.

[0043] Wavelet transform decomposes the signal into high-frequency and low-frequency components. The process of decomposing an image into four sub-image components: HL, LH, HH, and yL. Each sub-image retains one-fourth of the original size, where the yL sub-image retains the low-frequency information, while the HL and LH sub-images capture the horizontal and vertical high-frequency details respectively, and the HH sub-image encodes the diagonal high-frequency features. The combination of wavelet matching and wavelet transform constitutes a lossless feature encoding module. This module uses wavelet transform to decompose the input feature map into four sub-images with half of the original resolution, increasing the channel dimension to four times the original while maintaining the key spatial-frequency relationship.

[0044] 3.1 Weighted feature fusion. As Figure 2 shown, BiFPN performs feature fusion through weighted summation, where the features at each level are adaptively weighted by learnable parameters and then normalized and integrated into the fusion process. The detailed calculation process is as follows:

[0045]

[0046] where O represents the output feature, I i is the input feature, w is the node weight, and ε is the learning rate, set to 0.0001.

[0047] The BiFPN module addresses this limitation by promoting bidirectional lateral connections that facilitate cross-resolution information flow, allowing for deeper interactions between high-level semantic features and low-level spatial details. By eliminating nodes with single input edges and redundant connections, BiFPN significantly simplifies the bidirectional feature network topology.

[0048] 3.1 Distance loss normalization. NWD-Loss models the bounding box as a two-dimensional Gaussian distribution, which can better describe the weights of different pixels within the bounding box. In this distribution, the pixel at the center of the bounding box has the highest weight, and the importance of pixels decreases from the center to the boundary. Specifically, for a horizontal bounding box R=(cx,cy,w,h), where (cx,cy) represents the center coordinates, and w and h represent the width and height respectively. The equation of the inscribed ellipse can be expressed as:

[0049]

[0050] where (μ x , μ y ) is the center coordinate of the ellipse, and σx and σy are the semi-axis lengths along the x-axis and y-axis respectively. Therefore, the probability density function of the two-dimensional Gaussian distribution is given by the following equation:

[0051]

[0052]

[0053] For two two-dimensional Gaussian distributions μ1 = N1(m, Σ) and μ2 = N2(m, Σ), their second-order Wasserstein distance is defined as:

[0054]

[0055] Since the Wasserstein distance is a distance metric, it cannot be directly used as a similarity metric (i.e., a value between 0 and 1). Therefore, its exponential form is used for normalization to obtain a new metric standard - the Normalized Wasserstein Distance (NWD):

[0056]

[0057] 4. Model training. The model training stage mainly optimizes the model continuously through forward propagation, error calculation, and backpropagation. The model network is as Figure 4 shown. In each round of training, the input data is passed through forward propagation to calculate the prediction result, which is then compared with the actual label to calculate the error value. Then, the weights of the neural network are updated through the backpropagation algorithm, and the batch gradient descent method is used to optimize the model. During the training process, a validation set is also used to evaluate the performance of the model to avoid overfitting, and the training is stopped through an early stopping strategy to ensure the balance between the generalization ability and accuracy of the model.

[0058] 5. Model inference. In the inference stage, the trained model will perform defect detection on new blade images. First, the input image is passed through forward propagation to calculate the prediction result of the model. Through layer-by-layer processing, the key information of the image is gradually extracted, and finally, the output of defect detection is generated. The entire inference process is efficient and accurate, capable of quickly identifying the defects on the blade and providing reliable detection results for practical applications.

[0059] In the detection effect diagram of this embodiment, Figure 3 , Figure 5 , Figure 6 respectively show the detection results of three types of aero-engine blade defects (corrosion, breakage, fracture). It can be seen that the defect detection method described in the present invention can handle various complex situations and has high robustness.

[0060] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for detecting defects in aircraft engine blades in a few-sample scenario, characterized in that: The method comprises the following steps: Step S100: defect area extraction. Defect area extraction is to copy and extract the key area information of the defect sample. The key area information includes the vertex coordinates of the circumscribed polygon of the blade defect mask and the color information of each pixel in the polygon. Step S200: Graphics adaptive adjustment. Graphics adaptive adjustment is to adjust the shape and color of the extracted defect area graphics. By applying color space transformation, a high degree of consistency between the defect area and the target normal sample image background color tone is achieved. Step S300: Defect sample synthesis. Defect sample synthesis completes the determination of the pasting position and posture of the defect pattern and the pasting synthesis of the defect pattern. According to the edge direction of the blade contour in the normal blade image, a curve fitting algorithm is used to calculate the matching degree between the blade and the defect edge to ensure the natural fusion between the defect and the blade contour. Step S400: Adaptive wave basis selection. Adaptive wave basis selection is to automatically select a wave basis for subsequent processing based on the input image features. The frequency characteristics and texture characteristics of the image are analyzed by frequency domain analysis and statistical analysis, and then a comprehensive score is obtained by weighted summation. The score is used to guide the selection of the wave basis that best suits the tensor characteristics of the current input image. Step S500: Wavelet transform downsampling. Wavelet transform downsampling is to perform wavelet transform on the input image, which retains the image detail information while reducing the amount of model calculation. After selecting the wavelet basis, the input feature map is decomposed into four sub-images using wavelet transform, each sub-image retains a quarter of the original dimension. The decomposition process retains the key space-frequency relationship and achieves lossless feature encoding. Step S600: weighted feature fusion. Weighted feature fusion is to perform adaptive weighted addition on each input component. Feature fusion is adaptively adjusted according to task-specific requirements and data set characteristics, thereby amplifying the expression of key features and enhancing feature quality and effect. Step S700: Distance loss normalization. A new loss function is introduced to replace the original loss function of the model, by modeling the bounding box as a two-dimensional Gaussian distribution and using the Wasserstein distance to calculate the similarity between these distributions. Step S800: Model training. Model training is to iteratively update the network weights through an optimization algorithm to minimize the loss function. The predicted output of each round is calculated through forward propagation, and then compared with the true label, the error is calculated and the network weights are adjusted through the back-propagation algorithm. The batch gradient descent method is used during the training process, and the performance on the validation set is regularly calculated to avoid overfitting and perform an early stopping strategy. Step S900: Model inference. Model inference is the process of using the trained weights to predict new input data. In the inference stage, the input data is calculated through forward propagation to obtain the final output. In this process, the weights of the network remain fixed and are only calculated without being updated. In the inference process, the input features are processed layer by layer, and after a series of operations such as activation functions, convolution layers, and pooling layers, the prediction results of the model and the blade defect detection results are finally obtained.

2. The method for detecting defects in aircraft engine blades in a few-sample scenario according to claim 1, characterized in that: In step S400, the frequency domain analysis and statistical analysis of the image tensor are performed as follows: the high and low frequency energy distribution is evaluated through the spectrum data after Fourier transformation, and the frequency feature score is obtained as follows: In the formula, the numerator represents the high-frequency energy in the spectral data, and the denominator represents the low-frequency energy in the spectral data. Texture feature analysis is performed on the local image region of the input tensor, and the indicators derived from the gray-level co-occurrence matrix, including contrast, energy, uniformity, and entropy, are used to calculate the texture feature score as follows: Score tex =W Con ·S con +W Ene ·S Ene +W Hom ·S Hom +W Ent ·S Ent The composite texture score is obtained by weighted summation of the texture index scores S, and W represents the weight coefficient of each statistical parameter.

3. The method for detecting defects in aircraft engine blades in a few-sample scenario according to claim 1, characterized in that: In step S500, the feature representation learning part includes a convolution layer, batch normalization and activation function, which are used to refine the feature extraction process and adjust the channel dimension of the subsequent network layer.

4. The method for detecting defects in aircraft engine blades in a few-sample scenario according to claim 1, characterized in that: In step S600, the deep learning network adaptively adjusts the weights corresponding to the input features and calculates the output features.

5. The method for detecting defects in aircraft engine blades in a few-sample scenario according to claim 1, characterized in that: In step S700, the two-dimensional Wasserstein distance in the image is: The normalized Wasserstein distance is: In the formula, cx, cy, are the horizontal coordinate, vertical coordinate, length and width of the bounding box respectively.

6. The method for detecting defects in aircraft engine blades in a few-sample scenario according to claim 1, characterized in that: In step S800, the model is trained using transfer learning and data enhancement techniques. The model pre-trained on a large-scale dataset is used as the initial weight and fine-tuned with a small amount of aircraft engine blade defect data, thereby improving the generalization ability of the model under limited samples.

7. The method for detecting defects in aircraft engine blades in a few-sample scenario according to claim 1, characterized in that: In step S900, the model is inferred in real time through the inference algorithm, and the Wasserstein distance optimization result of the bounding box is combined to quickly and accurately detect blade defects. The optimized model weights are used to perform forward propagation on the input blade image, and the normalized distance loss function and bounding box prediction results are combined to achieve high-precision defect detection.

Images