Aero-engine small-scale damage detection method and system

Abstract

The invention discloses an aero-engine small-scale damage detection method and system, and belongs to the technical field of image processing. An aero-engine small-scale damage detection model is constructed; inputting the acquired damage sample data of the aero-engine into the trained aero-engine small-scale damage detection model, extracting information features in a space direction through coordinate attention, segmenting the information features into a horizontal direction and a vertical direction, and generating attention weights in the horizontal direction and the vertical direction; shallow boundary information and deep semantic information are extracted in the feature fusion layer through a selective boundary aggregation structure, and fusion is carried out through attention weight; the fusion features are detected through an output layer, and an aero-engine damage detection result is output. According to the method, the detection precision of small-scale damage can be effectively improved, and the intelligent detection level of the damage of the aero-engine is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to a method and system for detecting small-scale damages of aeroengines. Background Art

[0002] As the power source of an aircraft, the operating state of an aeroengine directly affects the flight safety of the aircraft and the lives of the crew on board. The core part of an aeroengine works in an environment of high temperature, high pressure and high rotational speed for a long time, and the internal blades are prone to damages such as cracks, dents, ablation, etc. Therefore, it is very necessary to regularly inspect the interior of an aeroengine to ensure its safe and stable operation. The traditional damage detection of aeroengines mainly uses borescope detection technology to detect internal damages of aeroengines, but it highly depends on the skills and experience of inspectors, cannot achieve automatic detection and identification, and has low detection efficiency. Automatic damage detection methods can improve the detection efficiency and avoid the interference of human factors. Therefore, it is of great significance to achieve accurate automatic detection of aeroengine damages.

[0003] Deep Learning is a branch of machine learning that focuses on using multi-layer neural networks to process and understand complex data. Its core idea is to construct an artificial neural network composed of multiple levels by imitating the neural network structure of the human brain, so as to automatically learn features from a large amount of data and make predictions. In recent years, deep learning has gradually been applied to the automatic detection of aeroengine damages.

[0004] Currently, due to the existence of a large number of small-scale damages in aeroengine damages, when constructing the existing aeroengine damage detection models, the problem of detecting and identifying small-scale damages in aeroengine damages is not considered, resulting in low detection accuracy of small-scale damages and inability to meet the actual use requirements. Summary of the Invention

[0005] In view of the problems existing in the above field, the present invention proposes a method and system for detecting small-scale damages of aeroengines. Through the constructed small-scale damage detection model of aeroengines, the method can effectively improve the detection accuracy of aeroengine damages and enhance the intelligent detection level of aeroengine damages.

[0006] To solve the above technical problems, the present invention discloses a method for detecting small-scale damages of aeroengines, including: Obtaining damage sample data of the aeroengine to be detected; Build a network model; introduce the Coordinate Attention mechanism CA into the original backbone network of YOLOv8, and fuse it with the C2f structure to form the CA-C2f structure; based on the original neck network of YOLOv8, use the features of 4x downsampling to construct a new feature fusion layer, and use the Selective Border Aggregation structure and the C2f structure to reconstruct the neck network; add an output layer composed of detection heads of multiple different scales after the feature fusion layer; train the network model to obtain a trained small-scale damage detection model for aeroengines; Input the damage sample data into the trained small-scale damage detection model for aeroengines, extract the information features in the spatial direction through the Coordinate Attention mechanism, divide the information features into horizontal and vertical directions, and generate the attention weights in the horizontal and vertical directions; extract the shallow boundary information and deep semantic information through the Selective Border Aggregation structure in the feature fusion layer, and fuse them through the attention weights; the fused features are detected through the output layer to output the aeroengine damage detection results.

[0007] Preferably, introducing the Coordinate Attention mechanism CA into the YOLOv8 backbone network and fusing it with the C2f structure to form the CA-C2f structure specifically includes: In the YOLOv8 backbone network, add the Coordinate Attention mechanism to form the backbone network of the small-scale damage detection model for aeroengines; The Coordinate Attention mechanism combines spatial information on the basis of the traditional attention mechanism, including coordinate information embedding and coordinate attention generation, where: Coordinate information embedding aggregates features of the input feature map along the horizontal and vertical directions, retaining the spatial position information of the input feature map; Coordinate attention generation enhances the feature response of the input feature map by generating direction-sensitive attention weights.

[0008] Preferably, generating the attention weights in the horizontal and vertical directions specifically includes: Through coordinate information embedding for the input X , use pooling kernels of two spatial ranges ([[]] H , 1) and (1, [[[]] W ), perform global pooling on each channel along the horizontal and vertical directions to generate a pair of direction-aware feature maps; Extract the information features in the spatial direction through coordinate attention generation, generate a feature that fuses the spatial information in the horizontal and vertical directions, divide the information features into horizontal and vertical directions, and process the features in the horizontal and vertical directions respectively to generate the attention weights in the horizontal and vertical directions.

[0009] Preferably, in the feature fusion layer, shallow boundary information and deep semantic information are extracted through a selective boundary aggregation structure and fused through attention weights, which specifically includes: Through the selective boundary aggregation structure, the boundary information of the shallow layer and the semantic information of the deep layer are selectively aggregated and fused to depict the damage contour and recalibrate the damage position; The selective boundary aggregation structure consists of recalibration attention units, which adaptively select and fuse features from different levels; The low-level and high-level features are input into two recalibration attention units in different ways, the outputs of the two recalibration attention units are concatenated, and processed by a 3×3 convolution to optimize the feature combination; By multiplying, the attention weights in the horizontal and vertical directions are respectively applied to the feature maps of the optimized feature combination. The attention weight in the horizontal direction is used to adjust the feature response of each row, and the attention weight in the vertical direction is used to adjust the feature response of each column.

[0010] Preferably, an output layer composed of multiple detection heads with different scales is added after the feature fusion layer, which specifically includes: The added new output layer is used as the detection head; The detection head consists of four output layers with different scales, and each output layer is responsible for detecting target objects of different sizes, including target position, type, and confidence information.

[0011] Preferably, the damage sample data of the aero-engine to be detected is 2710 aero-engine damage images of ablation damage, crack damage, dent damage, and material loss damage collected from the aero-engine borescope inspection report, including 834 ablation damage images, 436 crack damage images, 526 dent damage images, and 914 material loss damage images, to construct a damage sample data set.

[0012] Preferably, the obtaining of the damage sample data of the aero-engine to be detected further includes: Using the Labelimg software to annotate the sample images of the ablation damage, crack damage, dent damage, and material loss damage of the aero-engine collected; Save the labels of the labeled images in txt format, where each line represents a damaged object. The data in the first column represents the damage type, 0 represents ablation damage, 1 represents crack damage, 2 represents dent damage, and 3 represents material loss damage; the data in the second column represents the X coordinate of the center of the damage after normalization, the data in the third column represents the Y coordinate of the center of the damage after normalization, the data in the fourth column represents the width of the damage after normalization, and the data in the fifth column represents the height of the damage after normalization; the data in the second column and the fourth column are normalized using the width of the original image, and the data in the third column and the fifth column are normalized using the height of the original image.

[0013] Preferably, the obtaining of the damage sample data of the aero-engine to be detected further includes: Perform data augmentation on the aero-engine damage image using methods of rotation, scaling, and adding noise to obtain the augmented damage data; Divide the augmented damage data into a training set and a validation set in a ratio of 8:2, and train the network model through the training set to obtain a trained aero-engine small-scale damage detection model.

[0014] Preferably, there is also provided an aero-engine small-scale damage detection system, including: A data acquisition module for obtaining the damage sample data of the aero-engine to be detected; A damage detection model construction module for constructing a network model; in the original backbone network of YOLOv8, the coordinate attention mechanism CA is introduced and fused with the C2f structure to form a CA-C2f structure; on the basis of the original neck network of YOLOv8, features with 4 times downsampling are used to construct a new feature fusion layer, and the neck network is reconstructed using the selective boundary aggregation structure and the C2f structure; an output layer composed of detection heads of multiple different scales is added after the feature fusion layer; the network model is trained to obtain a trained aero-engine small-scale damage detection model; A damage detection module for inputting the damage sample data into the trained aero-engine small-scale damage detection model, extracting information features in the spatial direction through the coordinate attention mechanism, splitting the information features into horizontal and vertical directions, and generating attention weights in the horizontal and vertical directions; extracting shallow boundary information and deep semantic information through the selective boundary aggregation structure in the feature fusion layer, and fusing them through the attention weights; the fused features are detected through the output layer to output the aero-engine damage detection result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The small-scale damage detection method for aero-engines proposed by the present invention constructs a small-scale damage detection model for aero-engines. By adding an attention mechanism to the backbone network of the YOLOv8 model, it can help the model focus more precisely on the damage area, capture both long-range spatial dependencies and retain the relationships between channels, enabling the model to more accurately locate and focus on important target areas in the image, thereby improving the model's feature expression ability. The neck network of the traditional YOLOv8 model fuses the features of 8 times, 16 times, and 32 times downsampling of the backbone network, and has good detection effects for large-scale and medium-scale damages. However, due to the small proportion of small-scale damages in the image, the small amount of information carried, and the unclear features, and during the feature extraction of the backbone network, multiple downsampling operations will cause serious loss of the detailed information of small-scale damages, resulting in low detection accuracy for small-scale damages. However, shallower features can well capture the details in the image and retain more information about small-scale damages. Therefore, the present invention uses the features of 4 times downsampling on the basis of the neck network of the YOLOv8 model. The constructed new feature fusion layer can fuse more information about small-scale damages, well capture the details in the image, and retain more information about small-scale damages. Through the selective boundary aggregation structure, it selectively aggregates the boundary information of the shallow layer and the semantic information of the deep layer for fusion, realizing the effective combination of features at different levels, thereby improving the detection accuracy of the aero-engine damage detection results output by the fusion features through the output layer. Description of the Drawings

[0016] Figure 1 It is a flowchart of the small-scale damage detection method for aero-engines proposed by the present invention; Figure 2 It is an aero-engine damage image collected by the present invention; Figure 3 It is an example of a label file for labeling the aero-engine damage image collected by the present invention; Figure 4 It is a network architecture diagram of the traditional YOLOv8 model; Figure 5 It is a C2f structure diagram of the traditional YOLOv8 model; Figure 6 It is an SPPF structure diagram of the traditional YOLOv8 model; Figure 7 It is a coordinate attention CA structure diagram introduced in the backbone network of the small-scale damage detection model for aero-engines constructed by the present invention; Figure 8 It is a CA-C2f structure diagram of the small-scale damage detection model for aero-engines constructed by the present invention; Figure 9 It is a backbone network structure diagram of the small-scale damage detection model for aero-engines constructed by the present invention; Figure 10 Neck network structure diagram of the small-scale damage detection model for aero-engines constructed according to the present invention; Figure 11 Architecture of the small-scale damage detection model for aero-engines constructed according to the present invention; Figure 12 Comparison of detection results of different models provided by the embodiments of the present invention. Detailed implementation manners

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the attached Figures 1 - 12 drawings in the embodiments of the present invention. It should be understood that the terms described in the present invention are only used to describe specific embodiments, and are not intended to limit the present invention.

[0018] As Figure 1 shown, the present invention proposes a small-scale damage detection method for aero-engines, including the following steps: S1: Obtain damage sample data of the aero-engine to be detected; S2: Construct a network model; in the original backbone network of YOLOv8, a coordinate attention mechanism CA is introduced and fused with the C2f structure to form a CA-C2f structure; on the basis of the original neck network of YOLOv8, features of 4-fold downsampling are used to construct a new feature fusion layer, and a selective boundary aggregation structure and the C2f structure are used to reconstruct the neck network; an output layer composed of multiple detection heads with different scales is added after the feature fusion layer; the network model is trained to obtain a trained small-scale damage detection model for aero-engines; S3: Input the damage sample data into the trained small-scale damage detection model for aero-engines, extract information features in the spatial direction through the coordinate attention mechanism, divide the information features into horizontal and vertical directions, and generate attention weights in the horizontal and vertical directions; extract shallow boundary information and deep semantic information through the selective boundary aggregation structure in the feature fusion layer, and fuse them through the attention weights; the fused features are detected through the output layer, and the aero-engine damage detection results are output.

[0019] In step S1, a damage data set is established The damage detection model based on deep learning requires high-quality data sets, but there is currently no publicly available aero-engine damage data set. Therefore, it is necessary to collect and sort out relevant damage samples, annotate them, and establish a data set for aero-engine damage detection to provide a data basis for subsequent research.

[0020] (1) Image sample collection The long-term inspection reports of aero-engines have recorded the damage conditions of the engines and accumulated a large amount of typical damage images and video data. Therefore, damage sample data is selected from these inspection reports. The common damage types of aero-engines include chip, curl, material loss, crack, deformation, ablation, corrosion, and dent, etc. However, since some damage types occur less frequently, it is difficult to collect enough samples for model training. Therefore, four relatively common damage types, namely ablation damage, crack damage, dent damage, and material loss damage, are selected to construct the dataset. Among them, dent damage belongs to a type of small-scale damage. A total of 2,710 aero-engine damage images are collected from the aero-engine inspection reports, including 834 ablation damage images, 436 crack damage images, 526 dent damage images, and 914 material loss damage images. Some sample data is as Figure 2 shown.

[0021] (2) Label the selected aero-engine damage image samples When training the object detection network, the damage images are used as the input of the network, and corresponding label information needs to be provided to adjust the network weights, so as to continuously improve the detection accuracy of the network. The original damage images do not have label information. Therefore, the Labelimg software is used to label the selected aero-engine damage sample images.

[0022] The labels of the images are stored in txt format. The specific damage label data is as Figure 3 shown. Each row represents a damage object. The data in the first column represents the damage type, 0 represents ablation damage, 1 represents crack damage and dent damage, 2 represents dent damage, and 3 represents material loss damage. The data in the second column represents the X coordinate of the damage center after normalization, the data in the third column represents the Y coordinate of the damage center after normalization, the data in the fourth column represents the width of the damage after normalization, and the data in the fifth column represents the height of the damage after normalization. The data in the second column and the fourth column are normalized using the width of the original image, and the data in the third column and the fifth column are normalized using the height of the original image.

[0023] (3) Divide the aero-engine damage image samples into a training set, a validation set, and a test set according to a set ratio The aero-engine damage sample dataset is divided into a training set, a validation set, and a test set according to a set ratio. The dataset consists of two folders, images and labels. The images folder is used to store the damage images, and the labels folder is used to store the label files. Both folders contain three sub-folders, train, val, and test. The train folder is used to store the training data, the val folder is used to store the validation data, and the test folder is used to store the test data.

[0024] The present invention uses YOLOv8 as the basic model. YOLOv8 is an efficient and fast one-stage object detection model, mainly composed of three parts: the backbone network Backbone, the neck network Neck, and the detection head Head. The specific structure is as Figure 4 shown.

[0025] Among them, the backbone network mainly includes CBS, C2f, and (Spatial Pyramid Pooling–Fast, SPPF). CBS consists of a common convolution, batch normalization, and SiLU activation function; the Shortcut parameter of the C2f structure in the backbone network is True. It divides the input into two parts. One part passes through the residual edge, and the other part passes through a path containing N Bottleneck structures for convolution. Then, the output of each Bottleneck structure is Concat combined with the residual part. Finally, it passes through a CBS structure to exchange and fuse information between feature maps of different stages, increasing the nonlinear ability and representation ability of the model. The specific structure of C2f is as Figure 5 shown.

[0026] SPPF is improved from SPP (Spatial Pyramid Pooling, SPP). The SPP structure applies max pooling of different scales to the input feature map in parallel, which can obtain feature information from different scales and enhance the model's perception ability of objects of different scales. While SPPF uses a serial method, connecting three max poolings of the same scale in series, which improves the detection speed while keeping the output result consistent with that of SPP. The SPPF structure is as Figure 6 shown.

[0027] The neck network mainly includes CBS, C2f, upsampling, and Concat structures. CBS consists of a common convolution, batch normalization, and SiLU activation function; the C2f structure in the neck network is as Figure 5 shown, where the Shortcut parameter is False; upsampling enlarges the feature map with a lower resolution to align it with the feature map with a higher resolution; Concat is used to fuse feature maps from different scales. These structures are interconnected to form (Feature pyramid networks, FPN) structure and (Path aggregation network, PAN) structure. FPN constructs a feature pyramid through top-down feature fusion, fusing low-level features with high-level features to obtain richer semantic information and improve the detection performance of the model; different from FPN, PAN constructs a feature pyramid in a bottom-up manner, aggregating low-level features with high-level features through lateral connections and cascading operations to make up for and strengthen the position information.

[0028] The detection head is the last part of the model, responsible for generating the output of object detection, including object location, category, and confidence information. The detection head consists of three output layers with different scales, and each output layer is responsible for detecting target objects of different sizes to improve the model's detection ability for targets of different scales.

[0029] In step S2, a small-scale damage detection model for aeroengines is constructed. The overall structure of the deep learning-based small-scale damage detection model for aeroengines constructed based on the YOLOv8 model consists of three parts: a backbone network, a neck network, and a detection head connected in series in sequence.

[0030] In the original backbone network of the YOLOv8 model, an attention mechanism is added to form the backbone network of the small-scale damage detection model for aeroengines.

[0031] The attention mechanism can help the model focus more precisely on the damage area, enabling the model to capture the key features of the damage, ignore background noise, enhance the model's expression ability, and thus improve the detection accuracy for small-scale damage. Coordinate Attention (CA), based on the traditional attention mechanism, combines spatial information, enabling the model to better capture the correlation in the spatial and channel dimensions. The specific structure of CA is as Figure 7 shown.

[0032] The implementation process of CA mainly includes two steps: coordinate information embedding and coordinate attention generation.

[0033] The core idea of CA is to decompose the traditional channel attention mechanism and embed the spatial direction information into the attention mechanism, capturing both long-range spatial dependencies and retaining the relationship between channels, enabling the model to more precisely locate and focus on important target areas in the image.

[0034] The main objective of coordinate information embedding is to perform feature aggregation on the input feature map along the horizontal direction ( X direction) and the vertical direction ( Y direction) to retain spatial position information. Specifically, for a given input X , two pooling kernels with different spatial ranges ( H , 1) and (1, W ) are used to perform global pooling on each channel along the horizontal and vertical directions, generating a pair of direction-aware feature maps. This enables the attention module to capture long-range dependencies in one spatial direction and retain precise position information in the other spatial direction, thereby helping the model more accurately locate the target of interest.

[0035] The core task of coordinate attention generation is to enhance the input features by generating direction-sensitive attention weights. Specifically, first, the information in the horizontal and vertical directions is integrated to generate a feature that fuses the spatial information in the horizontal and vertical directions. Then, the feature is divided into two parts in the horizontal and vertical directions, and these two parts are processed separately to generate the attention weights in the horizontal and vertical directions. Finally, by multiplying, the attention weights in the horizontal and vertical directions are applied to the feature map respectively. The attention weight in the horizontal direction is used to adjust the feature response of each row, and the attention weight in the vertical direction is used to adjust the feature response of each column. In this way, the model's attention to important regions can be enhanced, while the response to unimportant regions can be weakened, enabling the model to more accurately locate and identify the targets in the image in the spatial dimension, thereby improving the model's feature expression ability.

[0036] CA is introduced into the YOLOv8 backbone network, and CA is fused with the C2f structure to form the CA-C2f structure. The CA-C2f structure is as Figure 8 shown. The backbone network of the small-scale damage detection model for aero-engines constructed in the present invention is as Figure 9 shown.

[0037] The neck network is mainly responsible for fusing the features of different scales extracted by the backbone network, thereby improving the model's detection ability for different-scale damages. The neck network of the YOLOv8 model fuses the features of 8 times, 16 times, and 32 times downsampling of the backbone network, and has good detection effects on large-scale and medium-scale damages. However, since small-scale damages account for a small proportion in the image, carry little information, and have unclear features, and during the feature extraction of the backbone network, multiple downsampling operations will cause serious loss of the detailed information of small-scale damages, resulting in low detection accuracy for small-scale damages. Shallower features can well capture the details in the image and retain more information about small-scale damages.

[0038] Therefore, on the basis of the original neck network of the YOLOv8 model in the present invention, a new feature fusion layer is constructed by using the features of 4 times downsampling to fuse more information about small-scale damages and improve the detection accuracy of the model for small-scale damages. The schematic diagram of the neck network of the small-scale damage detection model for aero-engines constructed in the present invention is as Figure 10 shown in Figure (a) of

[0039] The neck network of the small-scale damage detection model for aero-engines mainly consists of a C2f structure and a Selective Boundary Aggregation (SBA) structure. Among them, the C2f structure is as Figure 5 shown, and the Shortcut parameter is False.

[0040] Low-level features are rich in details, with more obvious boundaries and less distortion. High-level features, on the other hand, contain rich semantic information. Directly fusing low-level features with high-level features may lead to redundancy and inconsistency. The SBA structure selectively aggregates the boundary information in the shallow layer and the semantic information in the deep layer to depict a more accurate damage contour and recalibrate the damage location. The SBA structure is mainly composed of a Re-calibration Attention Unit (RAU). By adaptively selecting and fusing features from different levels, the RAU can fully utilize high-level semantic features while retaining boundary information, significantly reducing information redundancy and inconsistency. The SBA structure is as shown in Figure 10 Figure (b) in it. The CBR in it consists of ordinary convolution, batch normalization, and the ReLU activation function. ⊙ represents element-wise multiplication, and ⊗ is the reverse operation by subtracting the feature T1', which can refine the inaccurate and rough output into an accurate and complete prediction map. represents element-wise addition.

[0041] As Figure 10 shown in Figure (b) in it, low-level and high-level features are input into two RAUs in different ways to make up for the missing spatial boundary information in high-level semantic features and the missing semantic information in low-level features. Finally, the outputs of the two RAUs are concatenated and processed with a 3×3 convolution to effectively enhance the expressiveness of the feature map and optimize the feature combination. This feature fusion strategy realizes the effective combination of features at different levels and refines the rough features, which can significantly improve the model performance.

[0042] After adding a new output layer to the newly constructed feature fusion layer, it is specifically used to detect small-scale damages and improve the detection accuracy of small-scale damages.

[0043] The detection head consists of four output layers with different scales. Each output layer is responsible for detecting target objects of different sizes, including target location, category, and confidence information.

[0044] The deep learning aero-engine small-scale damage detection model constructed based on the YOLOv8 model is as shown in Figure 11 it, and it is named (Coordinate Attention Selective Boundary Aggregation-You Only Look Once, CS-YOLO). Inside the dashed box A is the improved CA-C2f structure, inside the dashed box B is the improved neck network, and inside the dashed box C is the newly added small-scale damage detection head.

[0045] In step S3, damage detection is performed on the aero-engine. To enable the model to fully learn the damage features, methods of rotation, scaling, and adding noise are used to perform data augmentation on the aero-engine damage images. First, 872 samples are selected as the test set and data augmentation is performed. After augmentation, the test set has a total of 1783 images. Second, data augmentation is performed on the remaining 1838 samples, and the sample data is amplified from 1838 to 6221. Then, the 6221 damage data are divided into a training set and a validation set in an 8:2 ratio.

[0046] The constructed small-scale damage detection model of the aero-engine is trained using the training set, the training effect of each round is verified using the validation set, and the final detection effect of the model is tested using the test set.

[0047] The trained small-scale damage detection model of the aero-engine is used to perform damage detection on the collected aero-engine damage images, which can realize the automatic detection of aero-engine damage and accurately detect the categories and locations of different-scale damages in images and videos.

[0048] The present invention also proposes a small-scale damage detection system for aero-engines, including: A data acquisition module for obtaining damage sample data of the aero-engine to be detected; A damage detection model construction module for constructing a network model; in the original backbone network of YOLOv8, the coordinate attention mechanism CA is introduced and fused with the C2f structure to form a CA-C2f structure; on the basis of the original neck network of YOLOv8, features with 4-fold downsampling are used to construct a new feature fusion layer, and a selective boundary aggregation structure and the C2f structure are used to reconstruct the neck network; an output layer composed of multiple detection heads with different scales is added after the feature fusion layer; the network model is trained to obtain a trained small-scale damage detection model of the aero-engine; A damage detection module for inputting the damage sample data into the trained small-scale damage detection model of the aero-engine, extracting information features in the spatial direction through the coordinate attention mechanism, dividing the information features into horizontal and vertical directions, and generating attention weights in the horizontal and vertical directions; in the feature fusion layer, shallow boundary information and deep semantic information are extracted through the selective boundary aggregation structure and fused through the attention weights; the fused features are detected through the output layer to output the aero-engine damage detection results.

[0049] The small-scale damage detection method for aero-engines proposed by the present invention can effectively improve the detection accuracy of small-scale damages.

[0050] Embodiment To verify the feasibility of the method proposed in the present invention, taking the small-scale damage detection method of aero-engines based on deep learning as an example, the small-scale damage detection method of aero-engines proposed in the present invention is verified.

[0051] The implementation steps of this embodiment include: Step 1: Collect, organize, and label damage sample data to establish a damage dataset; Step 2: Aero-engine small-scale damage detection model constructed based on the YOLOv8 model: Introduce an attention mechanism into the original backbone network part of the YOLOv8 model, fuse coordinate attention with the C2f structure to form a CA-C2f structure; Based on the original neck network of the YOLOv8 model, use the features of 4x downsampling to construct a new feature fusion layer, and use the selective boundary aggregation structure SBA and the C2f structure to reconstruct the neck network; Add an output layer composed of detection heads of multiple different scales after the feature fusion layer, specifically for detecting small-scale damage; Step 3: Train the aero-engine small-scale damage detection model; Use the damage dataset to train the aero-engine small-scale damage detection model; Step 4: Use the trained weight file to test with the test set to test the detection performance of the constructed aero-engine small-scale damage detection model; Step 5: Use the trained aero-engine small-scale damage detection model to detect aero-engine damage.

[0052] Select the number of parameters, average precision (Average Precision, AP ), mean average precision (mean Average Precision, mAP ) as the model evaluation indicators. Experiments are carried out on a device with a processor of Intel Core i7-13700F 2.10GHz, a graphics card of NVIDIA GeForce RTX 4060Ti, and 8GB of memory. AP Denotes the area under the PR curve formed with recall (Recall, R ) and precision (Precision, P ) as the coordinate axes. mAP Denotes the AP average value of all damage categories.

[0053] R , P , AP and mAP The calculation formulas are as follows: The present invention designs the following ablation experiments to evaluate the improved method. The comparison of the ablation experiment results is shown in Table 1. Among them, Experiment a is that the YOLOv8 model is trained and tested on the aero-engine damage dataset; Experiment b is to improve the backbone network, and the CA is introduced into the backbone network to improve the C2f structure; Experiment c is the finally improved model, on the basis of Experiment b, the neck network is improved, a new feature fusion layer is constructed, and the SBA structure is introduced to reconstruct the neck network.

[0054] Table 1 Comparison of Detection Results of Ablation Experiments From the comparison between Experiment a and Experiment b, it can be seen that after introducing CA into the backbone network, the AP of small-scale damage dents has increased by 2%, mAP has increased by 0.3%, but the AP of cracks and material loss damage have both decreased slightly.

[0055] From the comparison between Experiment b and Experiment c, it can be seen that the AP of small-scale damage dents has increased by 0.6%, the AP of ablation damage has increased by 1.1%, the AP of material loss damage has increased by 1.2%, the AP of crack damage has increased by 0.1%, mAP has increased by 0.7.

[0056] From the comparison between Experiment a and Experiment c, it can be seen that the AP of small-scale damage dents has increased by 2.6, the AP of ablation damage has increased by 1.1%, the AP of material loss damage has increased by 0.4%, the AP of crack damage is the same as that of the original model, mAP has increased by 1%, and the number of parameters has increased by 25.9%.

[0057] It can be seen that the number of parameters of the finally improved model has increased slightly, but AP and mAP both have relatively large improvements. Especially for small-scale damage such as dents, AP has increased by 2.6%. Compared with the original model, the improved model has better comprehensive detection performance.

[0058] To further verify the effectiveness of the model improvement, under the same experimental conditions, a comparative experiment was conducted on the small-scale damage detection model CS-YOLO of aero-engine proposed in the present invention with the traditional YOLOv3, YOLOv5, and YOLOv6. The results of detection using different models are shown in Table 2.

[0059] Table 2 Comparison of Detection Results of Different Models Compared with YOLOv3, for ablation and crack damage, the AP of CS-YOLO is 0.1% lower than that of YOLOv3, but for dent and material loss damage, the AP of CS-YOLO is 0.7% and 0.8% higher than that of YOLOv3 respectively, and the mAP of CS-YOLO is 0.3% higher than that of YOLOv3, and the number of parameters of CS-YOLO is about 1 / 7 of that of YOLOv3.

[0060] Compared with YOLOv5, for ablation, crack, dent, and material loss damage, the AP of CS-YOLO is 1%, 1.7%, 3%, and 0.3% higher than that of YOLOv5 respectively, and the mAP of CS-YOLO is 1.5% higher than that of YOLOv3.

[0061] Compared with YOLOv6, for ablation, crack, dent, and material loss damage, the AP of CS-YOLO is 0.8%, 2.2%, 3.3%, and 0.3% higher than that of YOLOv5 respectively, and the mAP of CS-YOLO is 1.6% higher than that of YOLOv3.

[0062] Compared with other models given in Table 2, the small-scale damage detection model CS-YOLO of aero-engine proposed in the present invention has higher detection accuracy for various types of damage. The detection accuracy for small-scale damage dent damage reaches 90.8%, and the comprehensive performance is better than other models.

[0063] As Figure 12 shown, the detection results of different models are given, and the detection effects of different models are more intuitively compared. The first column to the fourth column are the detection results of YOLOv3, YOLOv5, YOLOv6, and CS-YOLO respectively.

[0064] From Figure 12 Figure (a) in it, it can be seen that for ablation damage detection, YOLOv5 and YOLOv6 had missed detections and failed to detect the damage marked by the ellipse in the figure, while YOLOv3 and CS-YOLO did not have missed detections. From Figure 12As can be seen from Figure (b) in [reference], for crack damage detection, YOLOv3, YOLOv5, and YOLOv6 all missed detections and failed to detect the damage marked by the ellipse in the figure, while CS-YOLO did not have missed detections. From Figure 12 As can be seen from Figure (c) in [reference], for dent damage detection, YOLOv3, YOLOv5, and YOLOv6 all missed detections and failed to detect the damage marked by the circle in the figure. The CS-YOLO proposed in the present invention has been improved for small-scale damage detection and can accurately detect dent damage. From Figure 12 As can be seen from Figure (d) in [reference], for material loss damage detection, since the damage size on the far right in the image is small, YOLOv3, YOLOv5, and YOLOv6 all missed detections and failed to detect the damage marked by the ellipse in the figure, while CS-YOLO accurately detected all the damage in the image.

[0065] In summary, compared with other models, the small-scale damage detection model of aero-engine proposed in the present invention has higher detection accuracy, especially for small-scale damage.

[0066] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

[0067] In addition, unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the art to which the present invention belongs. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In case of conflict with any incorporated document, the content of this specification shall prevail.

Claims

1. A method for detecting small-scale damage of an aircraft engine, characterized in that: The following steps are involved: Obtain damage sample data of the aircraft engine to be tested; Construct a network model; this network model introduces the coordinate attention mechanism CA into the original YOLOv8 backbone network and fuses it with the C2f structure to form a CA-C2f structure; based on the original YOLOv8 neck network, a new feature fusion layer is constructed using 4x down-sampling features, and the neck network is reconstructed using the selective boundary aggregation structure and the C2f structure; An output layer consisting of multiple detection heads of different scales is added after the feature fusion layer; the network model is trained to obtain a trained aircraft engine small-scale damage detection model; The damage sample data is input into the trained aircraft engine small-scale damage detection model, and the information features of the spatial direction are extracted through the coordinate attention mechanism. The information features are divided into horizontal and vertical directions, and the attention weights of the horizontal and vertical directions are generated. In the feature fusion layer, the shallow boundary information and deep semantic information are extracted through the selective boundary aggregation structure and fused through the attention weight; The fused features are detected through the output layer to output the aircraft engine damage detection results.

2. The method for detecting small-scale damage to an aircraft engine according to claim 1, characterized in that: The coordinate attention mechanism CA is introduced into the YOLOv8 backbone network and integrated with the C2f structure to form a CA-C2f structure, which specifically includes: In the YOLOv8 backbone network, a coordinate attention mechanism is added to form the backbone network of the aircraft engine small-scale damage detection model; The coordinate attention mechanism combines spatial information on the basis of the traditional attention mechanism, including coordinate information embedding and coordinate attention generation, where: Coordinate information embedding is to aggregate the input feature map along the horizontal and vertical directions, and retain the spatial position information of the input feature map; Coordinate attention generation enhances the feature response of the input feature map by generating direction-sensitive attention weights.

3. The method for detecting small-scale damage to an aircraft engine according to claim 2, characterized in that: The generating of attention weights in the horizontal direction and the vertical direction specifically includes: By embedding the coordinate information into the input X , using two pooling kernels of different spatial ranges ( H , 1) and (1, W ), globally pool each channel along the horizontal and vertical directions to generate a pair of direction-aware feature maps; The information features of the spatial direction are extracted through coordinate attention generation, and a feature that integrates the horizontal and vertical spatial information is generated. The information feature is divided into horizontal and vertical directions, and the horizontal and vertical features are processed separately to generate horizontal and vertical attention weights.

4. The method for detecting small-scale damage to an aircraft engine according to claim 3, characterized in that: The feature fusion layer extracts shallow boundary information and deep semantic information through a selective boundary aggregation structure, and fuses them through attention weights, specifically including: Through the selective boundary aggregation structure, the shallow boundary information and deep semantic information are selectively aggregated and fused to depict the damage contour and recalibrate the damage location; The selective boundary aggregation structure consists of recalibrated attention units, which adaptively select and fuse features from different levels; The low-level and high-level features are input into two recalibrated attention units in different ways. The outputs of the two recalibrated attention units are concatenated and processed using 3×3 convolution to optimize the feature combination. The horizontal and vertical attention weights are applied to the feature map of the optimized feature combination by multiplication. The horizontal attention weight is used to adjust the feature response of each row, and the vertical attention weight is used to adjust the feature response of each column.

5. The method for detecting small-scale damage to an aircraft engine according to claim 1, characterized in that: The output layer consisting of multiple detection heads of different scales is added after the feature fusion layer, specifically including: The newly added output layer is used as the detection head; The detection head consists of four output layers of different scales, each of which is responsible for detecting target objects of different sizes, including target location, type and confidence information.

6. The method for detecting small-scale damage to an aircraft engine according to claim 1, characterized in that: The damage sample data of the aircraft engine to be inspected are 2710 images of four types of aircraft engine damage, namely, ablation damage, crack damage, dent damage and material loss damage, collected from the aircraft engine borescope inspection report, including 834 ablation damage images, 436 crack damage images, 526 dent damage images and 914 material loss damage images, to construct a damage sample data set.

7. The method for detecting small-scale damage to an aircraft engine according to claim 1, characterized in that: The step of obtaining damage sample data of the aircraft engine to be tested further includes: Labelimg software was used to annotate the sample images of ablation damage, crack damage, dent damage, and material loss damage of aircraft engines. The labels of the annotated images are saved in txt format, where each row represents a damage object. The first column of data represents the damage type, 0 represents ablation damage, 1 represents crack damage, 2 represents dent damage, and 3 represents material missing damage; the second column of data represents the X coordinate of the normalized damage center, the third column of data represents the Y coordinate of the normalized damage center, the fourth column of data represents the normalized damage width, and the fifth column of data represents the normalized damage height; the second and fourth columns of data are normalized using the width of the original image, and the third and fifth columns of data are normalized using the height of the original image.

8. The method for detecting small-scale damage to an aircraft engine according to claim 7, characterized in that: The step of obtaining damage sample data of the aircraft engine to be tested further includes: The aircraft engine damage image is enhanced by using the methods of rotation, scaling and adding noise to obtain enhanced damage data; The damage data after data enhancement is divided into a training set and a validation set in a ratio of 8:

2. The network model is trained with the training set to obtain the trained aero-engine small-scale damage detection model.

9. An aircraft engine small-scale damage detection system, characterized in that: include: A data acquisition module, used to obtain damage sample data of the aircraft engine to be tested; Damage detection model construction module, used to build a network model; this network model introduces the coordinate attention mechanism CA into the original YOLOv8 backbone network, and fuses it with the C2f structure to form a CA-C2f structure; based on the original YOLOv8 neck network, it uses 4 times down-sampling features to build a new feature fusion layer, and uses the selective boundary aggregation structure and C2f structure to reconstruct the neck network; An output layer consisting of multiple detection heads of different scales is added after the feature fusion layer; the network model is trained to obtain a trained aircraft engine small-scale damage detection model; The damage detection module is used to input the damage sample data into the trained aircraft engine small-scale damage detection model, extract the information features of the spatial direction through the coordinate attention mechanism, divide the information features into horizontal and vertical directions, and generate attention weights in the horizontal and vertical directions; In the feature fusion layer, the shallow boundary information and deep semantic information are extracted through the selective boundary aggregation structure and fused through the attention weight; The fused features are detected through the output layer to output the aircraft engine damage detection results.

Images