Aero-engine blade defect segmentation method based on edge detection

Abstract

The application discloses an aero-engine blade defect segmentation method based on edge detection. A defect edge segmentation model used in the method is divided into four prediction branches. Output feature maps of a feature sampling module of a previous prediction branch are input into a next prediction branch after being subjected to maximum pooling. Each prediction branch comprises a feature sampling module and a multi-scale feature enhancement attention module. Features extracted by the feature sampling module are input into the multi-scale feature enhancement attention module after being subjected to convolution and activation, so that feature information enhancement and channel enhancement are performed. Output feature maps of the multi-scale feature enhancement attention module are subjected to difference processing, and then subjected to full connection and activation, so that output feature maps of the prediction branch are obtained. The output feature maps of the four prediction branches are spliced, and then subjected to full connection and activation, so that output of the defect segmentation model is obtained. The trained defect edge segmentation model is used for defect segmentation, and defect edges are extracted. Features are extracted from different scales, feature information and channel enhancement are used to avoid loss of tiny and weak defect edge information, and segmentation precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine blade defect detection technology, specifically a method for segmenting aero-engine blade defects based on edge detection. Background Technology

[0002] Blades are a crucial component of aero-engines, and defects such as cracks and edge damage are unavoidable during the manufacturing process. These defects accelerate engine failure and, in severe cases, force engine shutdown, posing significant safety hazards. Therefore, defect detection is a critical step in the aero-engine blade manufacturing process, playing a vital role in maintaining the normal operation of both the aero-engine and the aircraft.

[0003] Non-destructive testing (NDT) refers to methods that, without damaging or affecting the performance of the tested mechanical material or harming its internal structure, utilize changes in thermal, acoustic, optical, electrical, and magnetic responses caused by abnormalities or defects in the material's internal structure to detect and test the structure, state, and defects of the specimen's interior and surface. Common NDT methods include radiographic testing, ultrasonic testing, magnetic particle testing, and eddy current testing, all of which have achieved good results in detecting internal defects in aero-engine blades. For surface defects on aero-engine blades, the mainstream deep learning-based defect detection focuses on defect classification and localization, i.e., identifying the type and location of defects. Defect segmentation, on the other hand, involves extracting defect edges and completely segmenting them. Defect segmentation is significantly more challenging than defect identification.

[0004] However, surface defects on aero-engine blades are inherently tiny and difficult to detect with the naked eye, belonging to the micrometer-scale. During deep learning, as the network deepens, operations such as convolution and downsampling can lead to the loss of these minute defect features, preventing the model from learning this information and thus affecting segmentation performance. To address this issue, this application proposes an edge detection-based aero-engine blade defect segmentation method that improves segmentation performance and accuracy while avoiding the loss of edge information for these minute defects. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a method for segmenting defects in aero-engine blades based on edge detection.

[0006] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0007] A method for defect segmentation of aero-engine blades based on edge detection, characterized by the following steps:

[0008] Step 1: Acquire images of defects in aero-engine blades and perform data augmentation;

[0009] Step 2: Constructing the defect edge segmentation model; The defect edge segmentation model consists of four prediction branches, which extract defect edge features from different scales and perform segmentation. The output feature map of the feature sampling module of the previous prediction branch is max-pooled and then input into the next prediction branch. Each prediction branch includes a feature sampling module and a multi-scale feature enhancement attention module. The features extracted by the feature sampling module are convolved and activated before being input into the multi-scale feature enhancement attention module for feature information enhancement and channel enhancement. The output feature map of the multi-scale feature enhancement attention module is interpolated and then fully connected and activated to obtain the output feature map of the prediction branch. The output feature maps of the four prediction branches are concatenated and then fully connected and activated to obtain the output of the defect segmentation model.

[0010] Step 3: Use the aero-engine blade defect images from Step 1 to train the defect edge segmentation model. After normalizing the aero-engine blade defect images to be segmented, input them into the trained defect edge segmentation model to segment defects and extract defect edges.

[0011] Furthermore, the input feature map of the multi-scale feature enhancement attention module is processed by convolution and activation, and then by dilated convolution at four different scales. The feature maps obtained from all dilated convolutions are concatenated together and then sequentially processed by global max pooling, convolution, and activation to generate channel attention maps. Finally, the channel attention maps are multiplied by the concatenated feature maps and then convolved to obtain the output feature map of the multi-scale feature enhancement attention module.

[0012] Furthermore, the feature sampling module of the first prediction branch of the defect edge segmentation model includes a convolution for changing the number of channels and three cascaded block structures, while the feature sampling modules of the remaining three prediction branches each include four cascaded block structures, which are used to extract features.

[0013] Furthermore, the input feature map of the block structure is sequentially processed through depthwise separable convolution, activation, and 1×1 convolution, and then concatenated with the input feature map to obtain the output feature map of the block structure.

[0014] Furthermore, the defects in the aero-engine blades include cracks and edge defects.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] (1) In the process of deep learning, as the network deepens, operations such as convolution and downsampling will cause the disappearance of small and weak defect features. In order to improve the expressive power of defect features, the defect edge segmentation model of the present invention is divided into four prediction branches. It extracts feature maps of different scales from the input image, and then extracts defect edge features of different scales from the feature maps of different scales to avoid the loss of small and weak defect edge information. In a single prediction branch, the multi-scale feature enhancement attention module uses dilated convolution of different scales to increase the receptive field of the model, so that the model can learn more defect edge features from different receptive fields, thereby enhancing the feature information and further avoiding the loss of small and weak defect edge information, thus improving the segmentation accuracy. The feature map obtained by dilated convolution is then refined by global max pooling, convolution and sigmoid activation function to achieve channel enhancement, further enhance the representation of defect edge features and suppress background noise.

[0017] (2) The defect edge segmentation model of the present invention can accommodate more network capacity to improve accuracy without increasing too much computational complexity, thus better balancing the accuracy and efficiency of defect segmentation of aero-engine blades. It has a good segmentation effect on small defects such as cracks and edge defects that are not easily identified by the naked eye in aero-engine blades, effectively improving the defect segmentation accuracy of aero-engine blades in multi-scale complex backgrounds. Attached Figure Description

[0018] Figure 1 This is a structural diagram of the defect edge segmentation model;

[0019] Figure 2 This is an architecture diagram of the block structure;

[0020] Figure 3 Architecture diagram of the attention module for multi-scale features;

[0021] Figure 4 The image shows the defect edge segmentation results for different models. Detailed Implementation

[0022] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to illustrate the technical solutions of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0023] This invention relates to a method for defect segmentation of aero-engine blades based on edge detection (hereinafter referred to as the method, see [link]). Figures 1-4 The process includes the following steps:

[0024] Step 1: Acquire images of defects in aero-engine blades and create a dataset;

[0025] This embodiment acquires a total of 100 images of defects in aero-engine blades, including the main defect categories: cracks and edge defects. Data augmentation is performed using methods such as rotation and flipping to obtain a dataset. The dataset is then divided into a training set and a test set to ensure that both the training set and the test set contain all defect categories.

[0026] Step 2: Construct a defect edge segmentation model; such as... Figure 1 As shown, the defect edge segmentation model uses a block structure as the basic unit and is divided into four prediction branches. The output feature map of the feature sampling module of the previous prediction branch is input into the next prediction branch after max pooling. Max pooling doubles the number of channels of the feature map and halves the height and width. Each prediction branch includes a feature sampling module and a multiscale feature enhanced attention (MFEA) module. The output feature map of the feature sampling module is fed into the multiscale feature enhanced attention module after a 1×1 convolution operation and a GELU (Gaussian Error Linear Unit) activation function for feature information enhancement and channel enhancement. The output feature map of the multiscale feature enhanced attention module is interpolated and then passed through a fully connected (Conv1×1-1) and sigmoid activation function to obtain the output feature map of the prediction branch. The four prediction branches produce four feature maps of different scales. The output feature maps of the four prediction branches are concatenated and then passed through a fully connected and activated module to obtain the output of the defect segmentation model, i.e., the defect segmentation result.

[0027] The first prediction branch's feature sampling module includes a 3×3-C convolution and three cascaded block structures. The feature sampling modules of the other three prediction branches each include four cascaded block structures. The input image size for the defect segmentation model is 3×H×W. The input image enters the first prediction branch and undergoes a 3×3-C convolution, expanding its size to C×H×W, where C represents the number of channels, and H and W represent the height and width. Figure 2 As shown, the input feature map of the block structure is sequentially passed through depthwise separable convolution (DWConv), GELU activation function and 1×1-C convolution, and then concatenated with the input feature map to obtain the output feature map of the block structure. The depthwise separable convolution uses a convolution in each channel of the feature map, and the outputs of all convolutions are concatenated to obtain the output of the depthwise separable convolution, which helps to improve computational efficiency.

[0028] like Figure 3As shown, the input feature map of the multi-scale feature enhancement attention module has a size of nC×H×W, where n is a positive integer. After the input feature map undergoes a 1×1-M (M is the number of channels) convolution and the GELU activation function, it is then subjected to four dilated convolutions at different scales. This effectively increases the receptive field size of the model, avoids losing too much defect edge information, and enables the model to learn more defect edge features at different scales from the receptive fields, thereby enhancing feature information and improving defect segmentation accuracy. The feature maps obtained from all dilated convolutions are concatenated together. The concatenated feature map is then subjected to global max pooling, a 1×1-1 convolution, and a Sigmoid activation function to generate a channel attention map, thereby enhancing channels and suppressing background noise. Global max pooling is used to generate channel descriptors to describe the global distribution of defect edge features in the channel-level feature response. The channel attention map is multiplied with the concatenated feature map to filter out feature maps that meet the conditions, avoiding a large amount of redundant information. The multiplied feature map is then subjected to a 1×1-M convolution to obtain the output feature map of the multi-scale feature enhancement attention module, with a size of M×H×W.

[0029] Input feature map X∈R of dilated convolution C×H×W After passing through the convolution filter w∈R H×W The output feature at position (i,j) in the input feature map is y. ij The formula for calculating dilated convolution is:

[0030]

[0031] Where, x [·] ∈X represents a submap of the input feature map X, and r represents the dilation rate, i.e., the sampling step size;

[0032] The formula for channel enhancement operation is:

[0033] Y2=σ(Conv(GMP(Y1)))·Y1(2)

[0034] Where Y1 represents the feature map obtained by concatenating all the feature maps after dilated convolution, GMP(·) represents the global max pooling operation, Conv(·) represents the convolution operation, σ represents the sigmoid activation function, and Y2 represents the feature map obtained by multiplying the channel attention map and the feature map Y1.

[0035] Step 3: Train and test the defect edge segmentation model, and use the trained defect edge segmentation model for defect segmentation of aero-engine blades to obtain defect edges;

[0036] Set the parameters and initialize the model: the maximum number of iterations (epochs) is 30, the learning rate is 0.001, the weight decay of the regularization term is 0.0005, and the maximum tolerance for the correct matching of the predicted value and the true value is 0.0075.

[0037] The training set is input into the initialized defect edge segmentation model for training, and the trained defect edge segmentation model is validated using a validation set. Deep supervision is employed, and the model parameters are optimized through backpropagation. Deep supervision can alleviate the gradient vanishing problem and improve the model's convergence speed and output accuracy. For predicted values ​​of p... u s The prediction loss l for the u-th pixel in the s-th prediction image u s for:

[0038]

[0039] Among them, y u Let η be the base ground truth edge probability of the u-th pixel, η be the predefined probability threshold, α = λ·(1-β) be the percentage of positive pixel samples, β be the percentage of negative pixel samples, and λ be the scaling factor.

[0040] In summary, the total prediction loss of the defect edge segmentation model is:

[0041]

[0042] After normalization, the defect image of the aero-engine blade to be segmented is input into the trained defect edge segmentation model for defect segmentation and extraction of defect edges.

[0043] To verify the effectiveness of the method of this invention, the defect edge segmentation model of this invention was compared with common defect segmentation models on a test set. The segmentation results of each model for cracks and edge defects are shown in [reference]. Figure 4 The performance of the models was comprehensively evaluated by selecting ODS (Optimal Dataset Scale, global best), OIS (single-graph best), Param(M) (number of model parameters), M-Adds (total number of multiplication and addition operations in the model) and FPS (frames per second). The statistical results of the indicators for each model are shown in Table 1.

[0044] Table 1 Statistical results of indicators for each model

[0045] Model ODS OIS Param(M) M-Adds FPS HED 0.706 0.714 14.718 160.46 90 RCF 0.609 0.616 14.804 205.26 90 BDCN 0.814 0.820 16.302 289.93 52 DexiNed 0.808 0.811 35.083 261.08 70 PiDiNet 0.832 0.850 0.710 39.06 102 EEDN-30 0.830 0.845 0.184 10.9 125 EEDN-60 0.859 0.880 0.712 41.3 97

[0046] In Table 1, "EEDN-30" and "EEDN-60" indicate that the defect edge segmentation model of this invention performs defect segmentation on images with 30 and 60 channels, respectively. As can be seen from the table, the defect edge segmentation model of this invention significantly outperforms other models in all indicators. Therefore, the defect edge segmentation effect of this invention is better and the segmentation accuracy is higher. Increasing the number of channels can improve the segmentation effect, but it increases the number of parameters and reduces the segmentation speed. Figure 4 It can be seen that other models suffer from problems such as blurred defect edge segmentation, unclear boundaries, and excessive interference (many white spots). In contrast, the model of this invention can clearly and completely segment defect edges. This is because the multi-scale feature enhancement attention module of this invention uses global max pooling, convolution, and the sigmoid function to refine the channel size, further enhancing the defect feature representation and suppressing background noise. At the same time, the multi-scale feature enhancement attention module uses dilated convolutions of different scales to effectively enhance the receptive field size of network neurons, thereby improving the defect segmentation accuracy. Compared with other models, the defect edge segmentation model of this invention can learn multi-scale feature information and feature channel enhancement at the same time, thus achieving good segmentation results.

[0047] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for defect segmentation of aero-engine blades based on edge detection, characterized in that, The method includes the following steps: Step 1: Acquire images of defects in aero-engine blades and perform data augmentation; Step 2: Constructing the defect edge segmentation model. The defect edge segmentation model consists of four prediction branches, which extract defect edge features from different scales and perform segmentation. The output feature map of the feature sampling module of the previous prediction branch is max-pooled and then input into the next prediction branch. Each prediction branch includes a feature sampling module and a multi-scale feature enhancement attention module. The features extracted by the feature sampling module are convolved and activated before being input into the multi-scale feature enhancement attention module for feature information enhancement and channel enhancement. The output feature map of the multi-scale feature enhancement attention module is interpolated, then fully connected and activated to obtain the output feature map of the prediction branch. The output feature maps of the four prediction branches are concatenated and then fully connected and activated to obtain the output of the defect segmentation model. Step 3: Use the aero-engine blade defect images from Step 1 to train the defect edge segmentation model. After normalizing the aero-engine blade defect images to be segmented, input them into the trained defect edge segmentation model to segment defects and extract defect edges.

2. The method for segmenting aero-engine blade defects based on edge detection according to claim 1, characterized in that, The input feature map of the multi-scale feature enhancement attention module is processed by convolution and activation, and then by dilated convolution at four different scales. The feature maps obtained by all dilated convolutions are concatenated together and then processed by global max pooling, convolution and activation in sequence to generate channel attention maps. Finally, the channel attention maps are multiplied by the concatenated feature maps and then convolved to obtain the output feature map of the multi-scale feature enhancement attention module.

3. The method for segmenting aero-engine blade defects based on edge detection according to claim 1, characterized in that, The feature sampling module of the first prediction branch of the defect edge segmentation model includes a convolution for changing the number of channels and three cascaded block structures. The feature sampling modules of the other three prediction branches each include four cascaded block structures, which are used to extract features.

4. The method for segmenting aero-engine blade defects based on edge detection according to claim 3, characterized in that, The input feature map of the block structure is sequentially processed through depthwise separable convolution, activation, and 1×1 convolution, and then concatenated with the input feature map to obtain the output feature map of the block structure.

5. The method for segmenting aero-engine blade defects based on edge detection according to any one of claims 1 to 4, characterized in that, The defects in the aero-engine blades include cracks and edge defects.

Images