Visible light and ultraviolet fusion detection method for aero-engine blade microcracks

Abstract

The invention belongs to the field of blade microcrack detection, and particularly discloses an aero-engine blade microcrack visible light and ultraviolet light fusion detection method which comprises the following steps: S1, treating the surface of a blade by adopting a chromic acid anodic oxidation technology; s2, the treated leaves are irradiated through a visible light source and an ultraviolet light source respectively; s3, constructing a visible light and ultraviolet fusion network; s3, the two backbone networks extract features of the visible light image and the ultraviolet image respectively; s4, the neck network generates fusion features with richer semantic information; s5, outputting category and position information of the microcracks with different sizes through the head network; according to the method, chromic acid anodic oxidation is adopted to process the aero-engine compressor blade, microcrack characteristics are improved, meanwhile, data are collected under visible light and ultraviolet light, visible light and ultraviolet light blade microcrack data sets with different characteristics are constructed, and efficient identification of blade microcracks is achieved through a visible light and ultraviolet light fusion network.

Description

Technical Field

[0001] This invention belongs to the field of blade microcrack detection, specifically relating to a visible light and ultraviolet light fusion detection method for microcracks in aero-engine blades. Background Technology

[0002] As a core aerodynamic component of the power system, the structural integrity of aero-engine blades directly determines the engine's thrust-to-weight ratio, fuel efficiency, and safety margin. During manufacturing, inherent material defects, process deviations, and structural design flaws can all induce the initiation of surface microcracks, severely impacting blade yield. During engine service, foreign objects such as sand and metal particles inevitably enter the blade surface. These hard objects impact the blade surface at high speed, creating a complex multiaxial stress field in the impact area, accompanied by micron-level cracks or notches. These initial damages, under the combined effects of harsh service environments and cyclic loads, easily induce the initiation and propagation of microcracks, seriously affecting engine performance and flight safety. Therefore, detecting microcracks in aero-engine blades and determining whether they meet relevant standards is extremely important. This measure has immeasurable value in preventing potential failures, extending engine life, and ensuring flight safety.

[0003] Traditional methods for detecting microcracks in blades mainly include visual inspection, vibration testing, ultrasonic testing, penetrant testing, and radiographic testing. Visual inspection, using simple tools like magnifying glasses for macroscopic observation of the blade surface, is convenient and inexpensive, but its resolution is typically limited to millimeters, making it difficult to effectively identify micron-sized cracks. Vibration testing, based on acoustic principles, identifies damage by analyzing the vibration spectrum characteristics generated when the blade is struck; however, this method is highly dependent on operator experience and struggles to accurately locate defects. Ultrasonic testing uses piezoelectric transducers to emit high-frequency sound waves, identifying defects by analyzing the propagation characteristics and echo signals within the material. While it offers high sensitivity, it requires specialized equipment and skilled technicians. Penetrant testing utilizes capillary action, allowing a colored penetrant to penetrate open defects on the surface, which are then revealed by a developer. This method provides good visualization of surface cracks but cannot detect internal defects, and its effectiveness is significantly affected by the properties of the penetrant. X-ray inspection, based on the attenuation characteristics of materials to X-rays, can directly reveal the internal structural features of blades. However, its equipment is expensive and poses radiation safety hazards, requiring strict protective measures. Therefore, various traditional inspection methods each have their own characteristics in terms of detection accuracy, applicability, ease of operation, and economy. Exploring a detection method that combines high efficiency and high accuracy for blade microcrack detection has significant engineering application value.

[0004] Traditionally, even after microcracks are detected using various microcrack detection methods, human interpretation of the results is still required. This work presents two main problems: First, it demands a certain level of responsibility, patience, and eyesight from the personnel conducting the microcrack detection, requiring extensive training to become proficient. Second, the repetitive nature of the work, coupled with the large number of blades involved, can easily lead to fatigue and boredom among the personnel, affecting the efficiency and accuracy of microcrack inspection. Summary of the Invention

[0005] The purpose of this invention is to provide a visible light and ultraviolet light fusion detection method for microcracks in aero-engine blades, so as to solve the problem that the traditional detection methods mentioned in the background art have significant limitations in terms of accuracy, efficiency and applicability.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A visible-ultraviolet fusion detection method for microcracks in aero-engine blades includes:

[0008] S1. Chromic acid anodizing technology is used to treat the blade surface;

[0009] S2. Irradiate the treated blades with visible light and ultraviolet light sources respectively, collect corresponding images, and construct a multimodal microcrack dataset;

[0010] S3. Construct a visible light ultraviolet fusion network, wherein the network architecture of the visible light ultraviolet fusion network includes two backbone networks, a neck network, and a head network;

[0011] S4. The two backbone networks extract features from the visible light image and ultraviolet image in the multimodal microcrack dataset, respectively. The backbone network uses the C2STR module to optimize the feature extraction capability. The C2STR module is composed of CSPBottleneck and Swing Transformer.

[0012] S5. The neck network utilizes a feature pyramid network to fuse the visible light image features and the ultraviolet image features from different stages of the backbone network to generate fused features with richer semantic information.

[0013] S6. Multi-scale target detection is performed on the fusion features with richer semantic information through the head network, and the category and location information of microcracks of different sizes are output.

[0014] Preferably, the treatment of the blade surface using chromic acid anodizing technology includes:

[0015] In the presence of H2CrO4 / CrO4 2－In the electrolyte, by applying direct current, the blade is connected to the positive terminal of the power supply as the anode, and the graphite is connected to the negative terminal of the power supply as the cathode.

[0016] Preferably, the construction of the multimodal microcrack dataset includes:

[0017] Visible light and ultraviolet light images were collected from the same location on the leaf to form raw data;

[0018] The original data was enhanced using translation, flipping, rotation, cutout, brightness adjustment, and mosaic enhancement methods.

[0019] The data was manually annotated using Labelimg, with the annotation type set to microcrack. The annotation was based on the ultraviolet image, and the annotation area covered the characteristic region of the microcrack.

[0020] Preferably, the structure of the dual backbone network includes five downsampling modules, each module containing five CBS (Conv-BN-SiLU), four C2STR, and one SPPF (Spatial Pyramid Pooling-Fast). The C2STR module combines with the Swin Transformer through branch processing to achieve interaction between local and global features.

[0021] Preferably, the Swing Transformer employs window attention (W-MSA) and shifted window attention (SW-MSA) mechanisms, combined with residual connections and multilayer perceptrons (MLP), to reduce computational complexity and enhance global feature extraction capabilities.

[0022] Preferably, the neck network is based on a PANet structure and includes seven Concats, six C2STRs, three Upsamples, and three DWConvs.

[0023] Preferably, the head network comprises four detection layers, corresponding to feature maps of sizes 20×20, 40×40, 80×80, and 160×160, respectively, for detecting large, medium, small, and tiny targets. Each detection layer consists of four CBS layers with a kernel of 3 and two CBS layers with a kernel of 1. The CBS layers with a kernel of 3 are used to further integrate the extracted features, and the CBS layers with a kernel of 1 are used to convert these features into prediction results of VUFNet. The bounding box regression and classification tasks are jointly optimized through mean squared error (MSE) and cross-entropy loss function.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] This invention enhances the microcrack characteristics of aero-engine compressor blades by treating them with chromic acid anodizing. Simultaneously, data is collected under visible and ultraviolet light to construct distinct visible and ultraviolet microcrack datasets. A visible-ultraviolet fusion network employs a dual-backbone network to extract visible and ultraviolet features separately, and a C2STR algorithm is designed to optimize feature extraction capabilities. Furthermore, DWConv is used to enhance computational efficiency. Experiments show that the detection speed is significantly superior to single-mode and other mainstream models, achieving efficient identification of blade microcracks and providing an efficient and reliable solution for the intelligent inspection of aero-engine blades. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0027] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0028] Figure 2 This is a schematic diagram of the chromic acid anodizing principle of the present invention;

[0029] Figure 3 This is a diagram showing the microcrack results of the blades of the present invention after chromic acid anodizing treatment;

[0030] Figure 4 This is a partial image of blade crack data collected according to the present invention;

[0031] Figure 5 This is a diagram illustrating the data enhancement effect of the present invention;

[0032] Figure 6 This is a diagram of the VUFNet network architecture of the present invention;

[0033] Figure 7 This is a structural diagram of the various modules in the main body of the present invention;

[0034] Figure 8 This is a structural diagram of the Neck of the present invention;

[0035] Figure 9 This is a structural diagram of Conv and DWConv of the present invention;

[0036] Figure 10 This is a diagram of the Detect structure of the present invention;

[0037] Figure 11 This is a graph showing the changes in the evaluation indicators of the present invention.

[0038] Figure 12 This is a diagram showing the detection results of microcracks in the blades according to the present invention. Detailed Implementation

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0042] As attached Figure 1 To be continued Figure 12 As shown:

[0043] One hundred compressor blades of a certain type of in-service engine were collected from a factory. These blades had been damaged under complex service conditions.

[0044] Example 1: This example provides a visible-ultraviolet fusion detection method for microcracks in aero-engine blades, including:

[0045] S1. Chromic acid anodizing technology is used to treat the blade surface;

[0046] The treatment of blade surfaces using chromic acid anodizing technology includes:

[0047] In the presence of H2CrO4 / CrO4 2－ In the electrolyte, by applying direct current, the blade is connected to the positive terminal of the power supply as the anode, and the graphite is connected to the negative terminal as the cathode, thereby forming a dense oxide film on the surface of the metal blade. The schematic diagram is attached. Figure 2 (Leftmost) As shown in the attached document; Figure 2 (The blade shown on the right) During the chromic acid anodizing process, the surface of the blade will open up the microcracks that were bonded together due to the anodic corrosion, revealing its defects. Because the chromic acid solution has high permeability and low viscosity, it can quickly penetrate into the microcracks. Under capillary action, after the chromic acid anodized product has been washed with water and left for one hour, the chromic acid in the microcracks seeps out, making the area appear yellow.

[0048] S2. Irradiate the treated blades with visible light and ultraviolet light sources respectively, collect corresponding images, and construct a multimodal microcrack dataset;

[0049] Appendix Figure 3 (Leftmost) This image was acquired under visible light. Faint yellow dots can be observed on the overall image of the blade. After high-magnification magnification, the crack length is clearly visible to be in the millimeter range, but the width is only in the micrometer range. Due to chromic acid leaching, a large area of ​​yellow appears around the crack. This yellow area is roughly the same length as the crack, but it is significantly expanded in width. Taking the magnified image at the bottom as an example, the crack width is 5 μm, and the widest point of the yellow area reaches 0.51 mm, an increase of 102 times. This demonstrates that chromic acid anodizing treatment can significantly improve the characteristics of microcracks in the blade and is an effective treatment method.

[0050] Because chromic acid has the property of absorbing ultraviolet light, ultraviolet light is used to irradiate the leaves. Figure 3 (Rightmost) is an image taken under ultraviolet light. Obvious black spots can be found in the overall image of the leaf. After being magnified with a high magnification lens, a black area can be clearly seen. By comparison, this black area is consistent with the yellow area taken under fluorescent light, but its contrast with the surrounding area is significantly improved.

[0051] In summary, after chromic acid anodizing, the microcracks on the blade appear yellow under visible light, and the microcrack characteristics are well preserved. Under ultraviolet light, the microcracks appear black, with higher contrast than under visible light, but the microcrack characteristics are lost, and the image is more cluttered than under visible light.

[0052] The construction of the multimodal microcrack dataset includes:

[0053] Visible light and ultraviolet light images were collected from the same location on the leaf to form raw data;

[0054] The original data was enhanced using translation, flipping, rotation, cutout, brightness adjustment, and mosaic enhancement methods.

[0055] The data was manually annotated using Labelimg, with the annotation type set to microcrack. The annotation was based on the ultraviolet image, and the annotation area covered the characteristic region of the microcrack.

[0056] The blades after chromate anodizing were photographed using a high-precision industrial camera, specifically a Huarui Technology MK1628M lens and A7500CG20 camera. Data was collected under different shooting distances, angles, and light sources to increase data diversity and enhance the final model's recognition capabilities. To ensure data consistency, photographs were taken at the same location under both visible and ultraviolet light, resulting in a total of 500 sets of data. Figure 4 The data is presented in two parts: (a) is an image acquired under visible light illumination. In this type of data, the microcracks are lighter in color, making them difficult to distinguish, but the background is clean and free of interfering targets; (b) is an image acquired under 365nm ultraviolet light illumination. In this type of data, the microcrack features are more obvious and easier to distinguish, but the cluttered background can easily lead to misjudgment by the model. The corresponding data at the top and bottom of the image are taken from the same location under different light sources and are considered as one group. The data was manually labeled using Labelimg, with the label type set to "microcrack" because the microcrack features are more obvious and the area is larger under ultraviolet light illumination; therefore, the labeled area is based on the image acquired under ultraviolet light. Figure 4 (b) The box indicates the labeled area;

[0057] These data augmentation methods can be used individually or in combination for images. (See attached image.) Figure 5 The image is augmented with the following effects: (leftmost) translation; (second from left) a combination of translation, flipping, and rotation; (third from left) a combination of translation, rotation, cutout, and brightness adjustment; and (fourth from left) mosaic. Ultimately, the data was augmented from 500 sets to 2000 sets, divided into training, validation, and test sets in an 8:1:1 ratio. Rich data is fundamental to deep learning and is used to achieve better training results.

[0058] S3. Construct the visible light and ultraviolet fusion network VUFNet. The network architecture of the visible light and ultraviolet fusion network includes two backbone networks, a neck network, and a head network.

[0059] S4. Two backbone networks extract features from the visible light and ultraviolet light images in the multimodal microcrack dataset, respectively. The backbone network uses the C2STR module to optimize the feature extraction capability. The C2STR module is composed of CSP Bottleneck and SwingTransformer.

[0060] The dual-backbone network structure includes five downsampling modules. Each module contains five CBS (Conv-BN-SiLU), four C2STR, and one SPPF (Spatial Pyramid Pooling-Fast). The C2STR module combines with the Swin Transformer through branch processing to achieve interaction between local and global features.

[0061] The backbone is responsible for transforming the input image into rich feature maps, which can be used as input for subsequent tasks to improve the model's performance and accuracy. Figure 6 It can be seen that Backbone downsampled the input image 5 times, obtaining feature maps at 5 scales;

[0062] Table 1. Main Structure

[0063]

[0064] As shown in the table, the input image size is 640×640×3. After feature extraction in Backbone, the output size of each module changes, with the spatial dimension gradually decreasing and the channel dimension gradually increasing.

[0065] like Figure 7 As shown in the left image, CBS is a convolutional module that includes a Conv2d, a BatchNormal2d, and a SiLU activation function. In Conv2d, k represents the kernel size and s represents the stride size, used for feature extraction; BatchNormal2d represents the batch normalization layer, used to accelerate convergence and improve model accuracy and robustness; SiLU is an activation function that introduces a smooth non-linear transformation, enabling the model to capture complex relationships in the data while maintaining the effectiveness of gradient propagation.

[0066] like Figure 7(Second from left) shows the structure diagram of the C2STR module. It improves upon the C2f module (Faster Implementation of CSP Bottleneck with 2 convolutions) by adding a Swin Transformer. Through feature transformation, branching, and feature fusion, it extracts and transforms the features of the input data, generating a more representative feature map. The C2STR module first processes the input data through a CBS, then divides the output into two equal parts. One part is passed directly forward, while the other part is processed by multiple Bottleneck modules (each Bottleneck module consists of two CBS modules and a Concat module). Finally, the results of the two parts are concatenated along the channel dimension and then processed by a CBS and a Swin Transformer to obtain the final output. The calculation formula for the C2STR module is:

[0067] F out =SwinTransformer(CBS) 1,1 (Concat(F1,F2…,F n )))

[0068]

[0069] Where F in and F out These are the input features and output features, respectively; F1, F2, ..., F n Features extracted for the 1st, 2nd, ..., nth branches; CBS k,s This is a convolution with a kernel size of k×k and a stride of s;

[0070] Swin Transformer employs window attention (W-MSA) and shifted window attention (SW-MSA) mechanisms, combined with residual connections and multilayer perceptrons (MLP), to reduce computational complexity and enhance global feature extraction capabilities.

[0071] like Figure 7(Second from left) The gray area on the right shows the Swin Transformer architecture. It combines Window Attention (W-MSA) and Shifted Window Attention (SW-MSA). The W-MSA mechanism divides the input into local windows and performs self-attention computation within these windows to facilitate the interaction between local and global information. However, due to the limitation of window size, W-MSA alone sometimes cannot capture global features in the image. Therefore, SW-MSA is introduced, shifting the window within each layer by a fixed size to help the model better understand the entire image and capture global features more effectively. Then, the MLP module uses fully connected layers to adjust and reorganize the features interacted with by the WTA module, helping the model better utilize the features obtained through the attention mechanism. Finally, the features processed by the first residual connection are combined with the features extracted by the MLP module through the second residual connection, further increasing the network depth. The calculation formula for the Swin Transformer is:

[0072] F1=W-MSA(LN(F in ))+F in

[0073] F2 = MLP(LN(F1)) + F1

[0074] F3 = SW - MSA(LN(F2)) + F2

[0075] F out =MLP(LN(F3))+F3 (2) The main feature of the Swin Transformer is the introduction of a hierarchical block self-attention mechanism, which can greatly reduce computational complexity and make the model more efficient in processing multi-scale images. At the same time, the self-attention mechanism of the Swin Transformer can help the model capture global contextual information, which is very important for understanding microcracks in blades, especially microcracks that are difficult to distinguish visually. In addition, the Swin Transformer is designed to have strong generalization ability, which means that even with limited training data or changes in data distribution, the model can maintain good performance and effectively detect microcracks in blades.

[0076] like Figure 7 As shown on the right, SPPF is an improved pooling layer structure. It consists of two CBS layers and multiple parallel max-pooling layers. These pooling layers divide the input feature map into sub-regions of different sizes, perform pooling operations within each sub-region, and then concatenate the pooling results. Its function is to enhance feature representation by introducing multi-scale spatial context information while maintaining network computational efficiency, enabling it to handle variable-sized input data without adjusting the network structure.

[0077] Fout =CBS 1,1 (Concat(F1,F2,F3,F4))

[0078] F1 = CBS 1,1 (F in )

[0079] F2 = Max 5，1 (CBS 1,1 (F in ))

[0080] F3 = Max 9，1 (Max 5，1 (CBS 1,1 (F in ))

[0081] F4 = Max 13，1 (Max 9，1 (Max 5，1 (CBS 1,1 (F in (3)

[0082] Max k，s This is a max pooling algorithm with a kernel size of k×k and a step size of s.

[0083] S5. The Neck network utilizes a feature pyramid network to fuse visible light image features and ultraviolet image features from different stages of the backbone network, generating fused features with richer semantic information.

[0084] The neck network is based on the PANet architecture and contains seven Concats, six C2STRs, three Upsamples, and three DWConvs.

[0085] The Neck section is responsible for multi-scale feature fusion, enhancing feature representation capabilities by fusing feature maps from different stages of the Backbone. For example... Figure 8 As shown, the Neck design in this application is based on PANet, using DWConv instead of ordinary convolutional layers, and adding a feature map of one scale, containing 7 Concat, 6 C2STR, 3 Upsample, and 3 DWConv. Our Neck part integrates feature maps from four different scales of visible light and ultraviolet images, and then integrates the Vis-Ult feature maps of different levels through top-down and bottom-up paths. Finally, it outputs four fused feature maps of different scales. This process can be calculated as follows:

[0086] F1 = Concat(F Vis,5 ,F Ult,5 )

[0087] F2 = C2STR(Concat(F Vis,4 ,F Ult,4 Upsample(F1)))

[0088] F3 = C2STR(Concat(F Vis,3 ,F Ult,3 Upsample(F2)))

[0089] F4 = C2STR(Concat(F Vis,2 ,F Ult,2 Upsample(F3)))

[0090] F out1 =C2STR(F4)

[0091] F out2 =C2STR(Concat(F3,DWConv(F out1 )))

[0092] F out3 =C2STR(Concat(F2,DWConv(F out2 )))

[0093] F out4 =C2STR(Concat(F1,DWConv(F)) out3 (4)

[0094] Where F Vis,n F represents the visible light feature map input to the nth module. Ult,n This represents the ultraviolet feature map input to the nth module; all other parameters are the same as... Figure 8 The corresponding annotations in the text.

[0095] like Figure 9As shown, Conv processes input multi-channel images using several multi-channel convolutional kernels, and the output feature map extracts both channel features and spatial features. The DWConv technique used in this application decomposes traditional convolution operations into two steps: depthwise convolution and pointwise convolution. It includes a 3x3 depthwise convolution, followed by BN and ReLU layers, then a 1x1 pointwise convolution, and finally another BN and ReLU layer. First, the depthwise convolution independently convolves each channel of the input data, generating feature map channels with the same number of input channels. Then, the pointwise convolution uses a 1x1 convolutional kernel to linearly combine the feature maps generated by the depthwise convolution to reduce the number of parameters and computational complexity. The number of output channels of the pointwise convolution can be adjusted as needed to control the depth of the output feature map. This structure not only effectively reduces the computational load of the model but also maintains its expressive power. The ratio of the number of parameters to the computational cost of the two convolutions can be calculated as follows:

[0096] Parameter ratio:

[0097] Computational complexity ratio:

[0098] The kernel size is k, the input channels are M, the output channels are N, and the output feature map size is D×D. Generally, N is large enough to be negligible. Figure 9 If the convolution kernel is 3, then the number of parameters and computational cost of using depthwise separable convolution will be reduced to about one-ninth of the original.

[0099] S6. Multi-scale target detection is performed using fusion features with richer semantic information through the Head network, outputting the category and location information of microcracks of different sizes.

[0100] The head network contains four detection layers, corresponding to feature maps of sizes 20×20, 40×40, 80×80, and 160×160, respectively, used to detect large, medium, small, and tiny targets. Each detection layer consists of four CBS layers with a kernel of 3 and two CBS layers with a kernel of 1. The CBS layers with a kernel of 3 are used to further integrate the extracted features, while the CBS layers with a kernel of 1 are used to transform these features into prediction results of VUFNet. The bounding box regression and classification tasks are jointly optimized through mean squared error (MSE) and cross-entropy loss functions.

[0101] The test results can be calculated as follows:

[0102] D=HNet{NNet[Concat(BNet(Visible),BNet(Ultraviolet))]} (1)

[0103] Where D represents the target detection result, including the type and location of the detected target, BNet(·), NNet[·] and HNet{·} represent the trunk, neck and head respectively, Concat(A, B) means stitching, which means that the input feature maps A and B are superimposed according to the channel size, Visible is the visible light image, and Ultraviolet is the ultraviolet image.

[0104] The Head section is responsible for the final object detection and classification tasks, including a detection head and a classification head. The detection head generates the detection results, and the classification head outputs the probability distribution for each category. The Head designed in this application is as follows: Figure 10 As shown, it consists of four CBSs with a kernel of 3 and two CBSs with a kernel of 1. The CBSs with a kernel of 3 further integrate the extracted features, while the CBSs with a kernel of 1 transform these features into prediction results from VUFNet.

[0105] Bbox Loss (Bounding Box Regression Loss) is used to calculate the difference between the predicted bounding box and the ground truth bounding box. Mean Squared Error (MSE) is a commonly used loss function that imposes a higher penalty for larger errors, which helps the model quickly correct large prediction mistakes. The model uses MSE as the loss function, and its formula is as follows:

[0106]

[0107] Where, x i Represents the coordinates of the actual bounding box. This represents the coordinates of the predicted bounding box. This loss function serves as the optimization objective, guiding the model to reduce the discrepancy between the predicted and ground truth bounding boxes during training.

[0108] Classification loss (CLS loss) measures the difference between the model's predicted class distribution and the true label distribution. Cross-entropy loss is a commonly used loss function in classification tasks, penalizing incorrect predictions significantly. Therefore, CLS loss helps the model optimize its predictions in classification problems, making the predicted probability distribution as close as possible to the true label distribution. Its calculation formula is:

[0109]

[0110] Among them, y o,c It is an indicator. It is 1 if sample o belongs to class c, and 0 otherwise. o It is the probability that the model predicts that sample o belongs to category c.

[0111] The head's input consists of four scales: 20×20, 40×40, 80×80, and 160×160. The receptive field sizes corresponding to the visible light and ultraviolet light images are 32×32, 16×16, 8×8, and 4×4, respectively, corresponding to large, medium, small, and micro-sized targets. Therefore, even when the target size varies significantly, the head can effectively detect microcracks in blades of different sizes.

[0112] Experimental example:

[0113] Evaluation indicators:

[0114] In the blade microcrack detection task, we use mAP0.5, mAP0.5:0.95, and detection speed as evaluation metrics. The following parameters are used in the formulas for some of the above evaluation metrics: True positives (TP), False positives (FP), False negatives (FN), and True negatives (TN). Intersection over Union (IoU) represents the ratio of intersection and concatenation between the bounding box and the true box.

[0115] Accuracy is a metric for evaluating the overall classification performance of a model; it represents the proportion of samples correctly predicted by the model out of all samples. The calculation formula is:

[0116]

[0117] Precision represents the proportion of samples that the model predicts to be positive, but which are actually positive. The formula is:

[0118]

[0119] Recall represents the proportion of positive samples that the model can identify out of all actual positive samples. The calculation formula is:

[0120]

[0121] Mean precision (AP) is equal to the area under the precision-recall curve. The formula is:

[0122]

[0123] mAP (Mean Average Precision) is a weighted average of the AP values ​​for all sample classes, used to measure the model's detection performance across all classes. The calculation formula is:

[0124]

[0125] mAP0.5 represents the average accuracy when the detection model IoU is set to 0.5, and mAP0.5:0.95 represents the average accuracy when the detection model IoU is set to 0.5-0.95 (with values ​​in 0.05 intervals).

[0126] Detection speed is measured in frames per second (fps) and refers to the number of images the model can recognize per second.

[0127] Model training:

[0128] In this paper, we use Windows 10 operating system, the compiler is Python 3.10.14, Pytorch 2.1.1, CUDA 11.8. All models are trained, validated, and reasoned on NVIDIA RTX3080.

[0129] Figure 11 The graph shows the changes in four evaluation metrics during training: precision, recall, mAP0.5, and mAP0.5:0.95. As can be seen from the graph, with the increase in training epochs, the values ​​of each metric show a steady upward trend, eventually stabilizing. This phenomenon indicates that the model training process is efficient and stable, capable of fully learning and extracting data features, exhibiting good convergence, and without overfitting. Specifically, the model performs excellently on the validation set: precision reaches 99.5%, indicating that a very high proportion of samples predicted as positive are actually positive; recall reaches 91.2%, indicating that the model can effectively identify most positive samples; mAP0.5 reaches 96.4%, reflecting the excellent overall detection performance of the model at an IoU threshold of 0.5; and mAP0.5:0.95 reaches 78.6%, further demonstrating that the model maintains high detection accuracy at different IoU thresholds. These results demonstrate that the model has excellent performance and strong generalization ability in object detection tasks.

[0130] The last trained model was used for testing, and the results are as follows: Figure 12 As shown (for easier observation, we display the detection results on the ultraviolet image).

[0131] Experimental results show that the proposed method can accurately identify defect types and precisely locate defect areas. Further analysis reveals that the detection results demonstrate good consistency with manual judgment in defect type identification and region segmentation, fully validating the algorithm's reliability. These results indicate that the proposed method exhibits excellent performance in blade microcrack identification and localization tasks, and has practical application value.

[0132] It is important to note that the constructions and arrangements of this application shown in several different exemplary embodiments are merely illustrative. Although only a few embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible (e.g., changes in the size, dimensions, structure, shape, and proportions of various elements, as well as parameter values ​​(e.g., temperature, pressure, etc.), mounting arrangements, use of materials, color, orientation, etc.) without substantially departing from the novel teachings and advantages of the subject matter described in this application). For example, an element shown as integrally formed may be composed of multiple parts or elements, the position of elements may be inverted or otherwise altered, and the nature or number or position of discrete elements may be changed or altered. Therefore, all such modifications are intended to be included within the scope of the invention. The order or sequence of any process or method steps may be changed or rearranged according to alternative embodiments. In the claims, any "device plus function" clause is intended to cover the structure described herein that performs the function, and not only structurally equivalent but also equivalent in structure. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of the invention. Therefore, the present invention is not limited to the specific embodiments, but extends to various modifications that still fall within the scope of the appended claims.

[0133] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of carrying out the invention as currently considered, or those features that are not relevant to implementing the invention) may be omitted.

[0134] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine work of design, manufacturing, and production without requiring much experimentation.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting microcracks in aero-engine blades using visible light and ultraviolet light fusion, characterized in that, include: S1. Chromate anodizing technology is used to treat the blade surface; S2. Irradiate the treated blades with visible light and ultraviolet light sources respectively, collect corresponding images, and construct a multimodal microcrack dataset; S3. Construct a visible light ultraviolet fusion network, wherein the network architecture of the visible light ultraviolet fusion network includes two backbone networks, a neck network, and a head network; S4. The two backbone networks extract features from the visible light image and ultraviolet image in the multimodal microcrack dataset, respectively. The backbone network uses a C2STR module to optimize the feature extraction capability. The C2STR module is composed of a combination of CSPBottleneck and Swing Transformer. S5. The neck network utilizes a feature pyramid network to fuse visible light image features and ultraviolet image features from different stages of the backbone network to generate fused features with richer semantic information. S6. Multi-scale target detection is performed on the fusion features with richer semantic information through the head network, and the category and location information of microcracks of different sizes are output.

2. The visible-ultraviolet fusion detection method for microcracks in aero-engine blades according to claim 1, characterized in that, The process of treating the blade surface using chromic acid anodizing technology includes: In the presence of H2CrO4 / CrO4 2－ In the electrolyte, by applying direct current, the blade is connected to the positive terminal of the power supply as the anode, and the graphite is connected to the negative terminal of the power supply as the cathode.

3. The visible-ultraviolet fusion detection method for microcracks in aero-engine blades according to claim 1, characterized in that, The construction of the multimodal microcrack dataset includes: Visible light and ultraviolet light images were collected from the same location on the leaf to form raw data; The original data was enhanced using translation, flipping, rotation, cutout, brightness adjustment, and mosaic enhancement methods. The data was manually annotated using Labelimg, with the annotation type set to microcrack. The annotation was based on the ultraviolet image, and the annotation area covered the characteristic region of the microcrack.

4. The visible-ultraviolet fusion detection method for microcracks in aero-engine blades according to claim 1, characterized in that, The structure of the dual backbone network includes five downsampling modules, each containing five CBS, four C2STR, and one SPPF. The C2STR module is combined with the Swing Transformer through branch processing.

5. The visible-ultraviolet fusion detection method for microcracks in aero-engine blades according to claim 1, characterized in that, The Swin Transformer employs window attention and shifted window attention mechanisms, combined with residual connections and a multilayer perceptron.

6. The visible-ultraviolet fusion detection method for microcracks in aero-engine blades according to claim 1, characterized in that, The neck network is based on the PANet architecture and includes seven Concats, six C2STRs, three Upsamples, and three DWConvs.

7. The visible-ultraviolet fusion detection method for microcracks in aero-engine blades according to claim 1, characterized in that, The head network comprises four detection layers, corresponding to feature maps of sizes 20×20, 40×40, 80×80, and 160×160, respectively, for detecting large, medium, small, and tiny targets. Each detection layer consists of four CBS layers with a kernel of 3 and two CBS layers with a kernel of 1. The CBS layers with a kernel of 3 are used to further integrate the extracted features, while the CBS layers with a kernel of 1 are used to transform these features into prediction results from VUFNet. The bounding box regression and classification tasks are jointly optimized using mean squared error and cross-entropy loss functions.

Images