Aero-engine abnormal borescope image screening method based on multi-stage generative adversarial

Abstract

The invention discloses an aero-engine abnormal borescope image screening method based on multi-stage generative adversarial, and relates to the field of intelligent detection and machine vision, in particular to the aero-engine abnormal borescope image screening method. The invention aims to solve the problem of low screening accuracy of the abnormal borescope image of the aero-engine due to the fact that the characteristics of weak texture, small abnormal area, high background redundancy and the like in the borescope image are difficult to process in the prior art. The method comprises the following steps: 1, acquiring a healthy aero-engine borescope image data set; 2, constructing a BiGenAD network model, and obtaining a trained BiGenAD network model based on the data set; the BiGenAD network model comprises a random simulator, a pixel restorer and a restoration discriminator; and 3, performing category prediction on the aero-engine borescope image data to be detected based on the trained BiGenAD network model to obtain whether the aero-engine borescope image data to be detected is a healthy image or an abnormal image.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and machine vision, specifically to a method for screening abnormal borehole images of aero-engines. Background Technology

[0002] With the continuous development of the aero-engine lifecycle management system, borescope technology has become an important means of monitoring engine structural health. Borescope inspection typically acquires image information of internal engine components in the form of video streams, resulting in massive amounts of data. In practical engineering applications, efficient and accurate anomaly image screening technology has become a key fundamental capability in the field of aero-engine maintenance.

[0003] Currently, deep learning-based anomaly detection methods are mainly divided into two categories: supervised learning and unsupervised learning.

[0004] 1. Supervised learning methods: These methods rely on a large number of manually labeled anomaly samples for model training. However, in the scenario of borehole inspection of aero-engines, the number of real damage samples is extremely limited (severe damage such as tearing and curling is extremely rare), and the distribution of each category is significantly imbalanced. This data imbalance makes it difficult for supervised models to exhaust all anomaly patterns, thus limiting their generalization ability.

[0005] 2. Unsupervised learning methods: This is a current research hotspot, aiming to learn distribution characteristics from healthy samples. Existing unsupervised methods mainly include three categories, but all have shortcomings in the borehole exploration scenario:

[0006] Reconstruction-based methods rely on healthy samples to learn reconstruction mappings and identify anomalies through reconstruction errors. However, the interior space of aero-engines is narrow and enclosed, and borescope images exhibit characteristics such as uneven illumination, high information redundancy, and weak texture, making it difficult for reconstruction models to accurately restore structural details and resulting in a high false detection rate.

[0007] Feature-based learning methods typically employ models pre-trained on natural images (such as ImageNet) to extract features. However, due to the significant differences in statistical distribution between borehole images and natural images, directly transferring pre-trained features can introduce inherent biases, resulting in insufficient representation of subtle damage.

[0008] Self-supervised learning-based methods: Although proxy tasks improve the understanding of data structures, they often only focus on global representations and cannot finely model the offset of local damaged regions at the structural and texture levels.

[0009] 3. Current status of Generative Adversarial Networks (GANs): Although GANs have been explored in industrial inspection, most existing methods lack a unified framework that can simultaneously simulate abnormal distributions and learn health features in detail, making it difficult to solve the problem of "small abnormal regions and weak structural details" in borehole images.

[0010] In summary, existing technologies struggle to address issues such as weak texture, small anomalous regions, and high background redundancy in borehole images. Therefore, there is an urgent need for a screening method that does not rely on anomalous samples, can construct a high-precision healthy feature distribution, and enhances sensitivity to local structural anomalies through generative adversarial mechanisms. Summary of the Invention

[0011] The purpose of this invention is to address the problem that existing technologies struggle to handle the characteristics of borehole images, such as weak texture, small abnormal regions, and high background redundancy, which leads to low accuracy in screening aero-engine anomalous borehole images. Therefore, this invention proposes a multi-stage generative adversarial method for screening aero-engine anomalous borehole images.

[0012] The specific process of the aero-engine anomaly borehole image screening method based on multi-stage generative adversarial methods is as follows:

[0013] Step 1: Obtain a healthy dataset of borehole images of aircraft engines;

[0014] Step 2: Construct the BiGenAD network model. Train the BiGenAD network model based on a healthy aero-engine borehole image dataset to obtain a trained BiGenAD network model.

[0015] The BiGenAD network model includes a random simulator, a pixel restorer, and a restore discriminator;

[0016] Step 3: Using the trained BiGenAD network model, analyze the borehole image data of the aircraft engine under test. Perform category prediction to obtain borehole image data of the aircraft engine under test. Is it a healthy image or an abnormal image?

[0017] The beneficial effects of this invention are as follows:

[0018] This invention proposes a multi-stage generative adversarial learning-based borehole image anomaly screening method (BiGenAD), which addresses the problems of scarce anomaly samples, difficulty in reconstructing weakly textured images, significant feature transfer bias, and unstable detection in existing technologies. It achieves high-precision automatic screening of aero-engine borehole images under unsupervised conditions. Compared with existing technologies, this invention has the following significant advantages:

[0019] 1. This invention eliminates the reliance on anomalous samples, addressing the industry pain points of sample scarcity and uneven distribution. Traditional supervised learning methods heavily depend on a large number of labeled anomalous samples, while in the actual environment of aero-engine borehole inspection, damage occurs very infrequently (severe damage such as "tears" is extremely rare), and the distribution of various types of damage is highly uneven. This invention adopts an unsupervised learning paradigm, eliminating the need for pre-collection of anomalous samples and performing feature mining solely through widely available healthy images. By generating diverse pseudo-anomaly images through a random simulator, the model can be exposed to various types and complexities of anomaly patterns during the training phase. This design not only eliminates the reliance on anomalous samples but also eliminates the need to exhaustively list all anomaly types beforehand, thus enabling flexible handling of various unknown damage patterns and significantly improving the model's applicability and generalization ability in practical engineering applications.

[0020] 2. This invention overcomes the detection bottleneck in weak texture and complex backgrounds, achieving fine-grained structural modeling. Addressing the reconstruction difficulties caused by the narrow interior space, uneven illumination, weak texture, and high background redundancy of aero-engines, this invention achieves deep mining of the distribution of health status features through the collaborative work of a pixel restorer and a random simulator. Under the constraint of adversarial training, the pixel restorer is forced to learn the inherent distribution of healthy samples in terms of spatial texture, illumination pattern, and structural consistency. Experiments demonstrate that even for weak texture regions commonly found in borehole images, this invention maintains stable reconstruction capabilities, significantly improving the separability of anomalous regions. This overcomes the problems of detail loss and high false detection rates in weak texture images of traditional reconstruction methods (such as VAE and Autoencoder), improving the accuracy of aero-engine anomalous borehole image screening.

[0021] 3. Because the pixel restorer only learns the healthy distribution, it cannot effectively restore abnormal regions (such as cracks and scratches) in the input image, resulting in significant structural differences between the restored image and the input image. To accurately capture these differences, this invention introduces multi-scale structural similarity (MS-SSIM) as an evaluation metric. Compared to single-scale evaluation, the multi-scale metric can better balance the similarity between the global and local parts of the image, effectively improving the stability of abnormal image screening in complex scenarios such as illumination changes, background noise, and local occlusion.

[0022] 4. A ternary composite adversarial loss function was constructed to balance generation diversity and restoration accuracy. This invention designs a ternary composite loss function comprising adversarial loss, L1 pixel loss, and structural similarity loss. The adversarial loss ensures the realism and distribution consistency of the generated images; the L1 loss ensures the pixel-level accuracy of the restored images, reducing the loss of edge details; and the structural similarity loss macroscopically constrains the perceptual consistency of the images in terms of brightness, contrast, and structure. Ablation experiments show that these three components are significantly complementary, and their synergistic effect ensures both the diversity of generated images and the structural integrity of the restored images, thereby greatly improving the overall detection performance of the model.

[0023] 5. The detection accuracy and performance indicators reach industry-leading levels. In actual testing on the Turbo19 aero-engine borescope dataset, this invention demonstrated excellent anomaly screening capabilities. Quantitative evaluation results show that the AUC of this invention's method reached 97.33, verifying its strong discriminative ability in distinguishing between healthy and abnormal samples; the accuracy reached 99.12%, significantly outperforming existing mainstream models such as MSTUnet and DDPM. Simultaneously, the high recall rate of 96.57% reflects the model's extremely low false negative rate, which is crucial for aviation safety maintenance.

[0024] 6. Possessing excellent cross-domain generalization ability, this invention is applicable to various industrial visual inspection tasks. Besides aerospace borehole inspection, it maintains extremely high detection accuracy on multiple publicly available industrial datasets, including magnetic tile surface inspection, hot-rolled steel strip defect detection (NEU-Seg), and texture defect detection (DAGM 2007). Even under challenges such as large intra-class differences in damage, high inter-class similarity, and complex textures, the AUC of this invention still reaches 96.52% to 98.08%. This fully demonstrates the good versatility of this technical solution and its wide applicability to various scenarios such as industrial defect detection and quality control.

[0025] 7. High computational efficiency, meeting the real-time requirements of engineering deployment. Computational complexity testing shows that the total number of parameters in this invention's model is moderate (23.35 M). Under mainstream GPU hardware conditions, the average inference time per image is only 0.095 seconds. This means the model possesses excellent real-time processing capabilities, meeting the needs of rapid and automated screening of large-scale video frames in aviation maintenance scenarios, ensuring high accuracy while also considering the timeliness of engineering applications.

[0026] 8. This invention provides visualized interpretability, enhancing the reliability of detection results. Through Class Activation Mapping (CAM) technology, it can intuitively highlight abnormal areas (such as blade cracks and surface scratches) that the model focuses on during the decision-making process. This visualization not only verifies the accuracy of the model's detection but also makes the detection process transparent and interpretable, facilitating engineers to quickly locate damage and providing an intuitive basis for subsequent maintenance decisions. Attached Figure Description

[0027] Figure 1 The BiGenAD model framework diagram designed for this invention;

[0028] The BiGenAD model effectively combines the dual tasks of image generation and restoration through the collaborative work of two generators. It eliminates the need for pre-collection of anomalous samples, relying solely on widely available healthy images for feature mining. The first generator (a random simulator) generates diverse and random local anomalies to simulate potential damage outside the healthy distribution. The second generator (a pixel restorer) performs global pixel restoration on the generated image samples, enhancing the model's understanding of local anomaly representations. A restoration discriminator supervises the generator output, improving the anomaly simulation capability of the first generator and the restoration capability of the second generator through adversarial learning. During training, the BiGenAD model, based on a designed ternary adversarial loss function, ensures the structural realism and pixel-level restoration quality of the generator output, establishing a comprehensive healthy feature distribution mapping. During testing, a proposed multi-scale structural similarity assessment evaluates the distribution deviation between the input and restored images, and, combined with the "irreversibility" of healthy samples, effectively identifies anomalous images.

[0029] Figure 2 This is a visualization of the class activation mapping (CAM) effect of the method of this invention on the Turbo19 dataset; the red highlighted area in the figure intuitively shows the BiGenAD model's ability to accurately locate abnormal areas such as scratches and cracks during the decision-making process, verifying the model's sensitivity to local minor damage;

[0030] Figure 3 This is a visualization of the class activation mapping (CAM) effect of the method of this invention on the Magnetic-Tile dataset; the red highlighted area in the figure intuitively shows the BiGenAD model's ability to accurately locate abnormal areas such as scratches and cracks during the decision-making process, verifying the model's sensitivity to local minor damage;

[0031] Figure 4This is a visualization of the class activation mapping (CAM) effect of the method of the present invention on the NEU-Seg dataset; the red highlighted area in the figure intuitively shows the BiGenAD model's ability to accurately locate abnormal areas such as scratches and cracks during the decision-making process, verifying the model's sensitivity to local minor damage;

[0032] Figure 5 This is a visualization of the class activation mapping (CAM) effect of the method of this invention on the DAGM 2007 dataset; the red highlighted area in the figure intuitively shows the BiGenAD model's ability to accurately locate abnormal areas such as scratches and cracks during the decision-making process, verifying the model's sensitivity to local minor damage;

[0033] To further elucidate the training performance and decision-making process of the BiGenAD model, this invention employs Class Activation Mapping (CAM) for visualization. This map is an efficient visualization tool in the field of deep learning, capable of intuitively highlighting the image regions activated by the convolutional neural network during the decision-making process. Figure 2 and Figure 3 The visualizations show activation mapping examples of BiGenAD on four datasets (some with magnified views), clearly displaying the anomalous regions activated by the BiGenAD model during the decision-making process (highlighted in red). These results not only validate BiGenAD's accuracy in anomaly detection but also intuitively demonstrate its ability to handle diverse data. This visualization is of significant guiding importance for understanding the decision-making mechanism of the BiGenAD model and assessing its reliability in practical applications. Detailed Implementation

[0034] Specific Implementation Method 1: The specific process of this implementation method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods is as follows:

[0035] Step 1: Obtain a healthy dataset of borehole images of aircraft engines;

[0036] Step 2: Construct the BiGenAD network model. Train the BiGenAD network model based on a healthy aero-engine borehole image dataset to obtain a trained BiGenAD network model.

[0037] The BiGenAD network model includes a random simulator (first generator), a pixel restorer (second generator), and a restore discriminator;

[0038] Step 3: Using the trained BiGenAD network model to analyze the borehole image data of the aircraft engine under test. Perform category prediction to obtain borehole image data of the aircraft engine under test. Is it a healthy image or an abnormal image?

[0039] This invention relates to the fields of intelligent inspection and machine vision, and particularly to an unsupervised anomaly screening technology for borehole images of aero-engines. Specifically, it involves a rapid screening method and system based on multi-stage generative adversarial learning, anomaly representation modeling, health feature distribution construction, and complex industrial visual data. In the digital borehole inspection process, the first step is to perform preliminary screening of massive borehole image data. In the aero-engine lifecycle management system, the borehole inspection process is recorded in dynamic, continuous video format, typically containing thousands to tens of thousands of frames. Applying complex damage detection algorithms directly to each video frame would not only significantly increase computational resource consumption but also significantly prolong the inspection cycle, affecting maintenance efficiency. Therefore, this invention aims to rapidly screen potential anomaly samples from massive amounts of data, so as to concentrate limited computational resources on images that may contain damage.

[0040] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step two, a BiGenAD network model is constructed, and the BiGenAD network model is trained based on a healthy aero-engine anomaly borehole image dataset to obtain a trained BiGenAD network model.

[0041] The BiGenAD network model includes a random simulator (first generator), a pixel restorer (second generator), and a restore discriminator;

[0042] The specific process is as follows:

[0043] Step 21: Centralize the healthy air-to-ground engine borehole image data. Input a random simulator, output a random simulator Corresponding abnormal image ;

[0044] Step 22: The output of the random simulator Corresponding abnormal image Input pixel restorer, pixel restorer output Corresponding restored image ;

[0045] Steps two and three: The output of the restorer... Corresponding restored image The input is a discriminator, which outputs a probability value in the interval [0 -1]. This probability value represents the restored image. Hole-view images of healthy aircraft engines Confidence level;

[0046] The discriminator uses a ResNet34 network structure;

[0047] Step 24: Repeat steps 21 to 23 until the ternary composite loss function converges to obtain the trained BiGenAD network model.

[0048] The BiGenAD model effectively combines the dual tasks of image generation and restoration through the collaborative work of two generators, without requiring pre-collection of anomalous samples, and only using widely available healthy images for feature mining. The first generator (a random simulator) generates random and diverse local anomalies to simulate potential damage outside the healthy distribution. The second generator (a pixel restorer) enhances the model's understanding of local anomaly representations by performing global pixel restoration on the generated image samples. A restoration discriminator supervises the generator output, improving the anomaly simulation ability of the first generator and the restoration ability of the second generator through adversarial learning. During training, the model uses a designed ternary adversarial loss function to ensure the structural realism and pixel-level restoration quality of the generator output, and establishes a comprehensive healthy feature distribution mapping. During testing, the proposed Multi-Scale Structural Similarity (MS-SSIM) is used to evaluate the distribution deviation between the input image and the restored image, and then, combined with the "irreversibility" of healthy samples, to effectively determine anomalous images.

[0049] The reconstruction discriminator is a crucial supervisory module in the entire generative adversarial architecture. Its main task is to evaluate the realism of the output images from the random simulator and the pixel reconstruction generator. In standard generative adversarial modes, the discriminator improves the generator's reconstruction ability by continuously learning to distinguish between real and reconstructed images. It outputs probability values ​​indicating whether the input image is real or generated. This adversarial training mechanism forces the generator to continuously improve itself to generate more realistic random images.

[0050] In the BiGenAD architecture designed in this invention, the reconstruction discriminator adopts the classic ResNet34 classification network architecture. It progressively extracts local and global features of the image through multiple convolutional layers and maps the high-dimensional feature vectors to binary classification outputs through fully connected layers. The probability value within the interval [0-1] represents the confidence level that the input image is a real healthy sample. In adversarial training, the pixel reconstruction unit attempts to reconstruct a restored image consistent with the real healthy sample, while the reconstruction discriminator strives to distinguish the restored image from the real healthy image. This adversarial process is designed to continuously improve the reconstruction capability of the pixel reconstruction unit, making its output restored image more consistent with the structural distribution of healthy samples.

[0051] The other steps and parameters are the same as in Specific Implementation Method 1.

[0052] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step two-one, the healthy aero-engine borehole image data is collected... Input a random simulator, output a random simulator Corresponding abnormal image The specific process is as follows:

[0053] Step 211: Healthy borehole images of aircraft engines Input to the encoder, the encoder outputs high-dimensional features;

[0054] The encoder is an encoder in a self-encoder;

[0055] The first part of the simulator is similar to the encoder structure of an autoencoder, which is responsible for extracting high-dimensional feature representations of the input image;

[0056] Step 212: Add random noise to the high-dimensional features output by the encoder to obtain the high-dimensional features after adding random noise;

[0057] Noise embedding: Random noise is introduced into the intermediate layer of the random simulator to simulate different anomaly types;

[0058] Steps 2-3: Input the high-dimensional features with added random noise into the decoder; the decoder outputs the abnormal image. ;

[0059] Decoder outputs abnormal image As the output of a random simulator;

[0060] Abnormal images With healthy aircraft engine borescope images They have the same dimensions;

[0061] The decoder is the decoder in the autoencoder.

[0062] The latter part of the random simulator is the decoder, which is responsible for remapping the high-dimensional feature maps extracted by the encoder back to the image space.

[0063] The main task of the random simulator is to generate synthetic images with random anomalies from healthy input images. This process borrows design principles from Conditional Generative Adversarial Networks (CGNs), simulating various damages in the real environment by using features from healthy input images. This allows the training process to provide diverse anomaly samples, thereby improving the model's ability to understand complex local anomalies in real-world scenarios.

[0064] The core characteristic of the random simulator lies in the diversity of generated anomalous samples. This characteristic is achieved by combining a conditional generative model with random noise. The convolutional neural network is not only used to extract high-dimensional features from the input image but also serves as a conditional input for generating anomalous images, ensuring that the generated anomalies are correlated with the original data. Compared to traditional generative adversarial structures, this design ensures that the generated images maintain global structural consistency with the original image, while local features generate specific anomalies based on the input noise.

[0065] Furthermore, to enhance the diversity of random anomalies, BiGenAD employs random noise injection. Small-scale pixel perturbations are introduced into random regions to simulate localized damage in real-world scenes; shape transformations and intensity perturbations are used to simulate more complex anomalies, such as large-area blurring, color distortion, and texture breakage. These generation methods ensure that BiGenAD is exposed to anomaly images of varying types and complexities during training, thereby strengthening its understanding of anomaly distribution representations.

[0066] Finally, the decoder remaps the perturbed feature map back to an image space of the same size as the original input through upsampling and deconvolution operations, generating the final anomalous image. In this process, BiGenAD gradually restores the global structure of the input image while preserving the anomalous features introduced in the encoder stage. Therefore, the output of the random simulator retains both the overall features of the input image and applies randomly generated anomalous regions locally or globally.

[0067] Other steps and parameters are the same as in specific implementation method one or two.

[0068] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: in step two-two, the output of the random simulator is... Corresponding abnormal image Input pixel restorer, pixel restorer output Corresponding restored image ;

[0069] The specific process is as follows:

[0070] Abnormal images output by the random simulator Input encoder, encoder output image latent feature space;

[0071] The latent feature space of the encoder output image is input into the decoder, and the decoder outputs a borehole image that is as close as possible to that of a healthy aircraft engine. The restored image ;

[0072] The encoder is an encoder in a self-encoder;

[0073] The decoder is the decoder in the autoencoder.

[0074] The task of the pixel restorer is to extract anomalous images generated by a random simulator. The image is restored to be as close as possible to the original input image. The restored image This restorer is based on an autoencoder architecture and consists of two parts: an encoder and a decoder. The encoder is responsible for processing the input aberration image. The image is converted into a low-dimensional feature representation, and the decoder reconstructs the original image based on these features, attempting to eliminate local anomalies introduced by the random simulator, so that the output image is as close as possible to the original healthy image.

[0075] The pixel restorer's restoration process relies on the powerful feature learning capabilities of the autoencoder structure. By encoding the aberration image, the pixel restorer can extract the image's latent feature space and, during the decoding stage, gradually reconstruct the aberration regions in the image based on these features. Because the autoencoder preserves the global structural information of the image in its latent feature representation, the pixel restorer can effectively reconstruct the overall shape and details of the original image during image restoration. In this process, the core objective of the pixel restorer is not only to generate a restored healthy image but also to ensure that the restored image maintains consistency with the original image in both local details and global structure.

[0076] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0077] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the process of obtaining the ternary composite loss function in steps Two and Four is as follows:

[0078] In generative adversarial learning frameworks, the design of the loss function is crucial for the effective training of the generator and discriminator. To ensure that the random simulator and pixel restorer can generate high-quality anomalous images and restore near-realistic healthy images, this invention designs a special ternary composite loss function.

[0079] 1) Constructing adversarial loss terms ;

[0080] 2) Construct the L1 loss term ;

[0081] 3) Constructing a structural similarity loss term ;

[0082] 4) Based on adversarial loss terms L1 loss term Structural similarity loss term Construct a ternary composite loss function .

[0083] The other steps and parameters are the same as in any of the specific implementation methods one to four.

[0084] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One through Five in that: in step 1), the adversarial loss term is constructed. ;

[0085] This loss term employs standard adversarial loss to enhance the realism of the generated images, prompting an adversarial game between the generator and the discriminator; the mathematical expression is shown below:

[0086] (1)

[0087] in,

[0088] Indicates the loss against; Indicates a random simulator; Indicates pixel restorer;

[0089] Hole-view images indicating a healthy aircraft engine; Indicates the discriminator; Hole-view images indicating the health of an aircraft engine The probability value output by the input discriminator;

[0090] This indicates a demand for expectation;

[0091] express An abnormal image output after inputting into a random simulator; Representing abnormal images The probability value output by the input discriminator;

[0092] This represents the restored image output by the pixel restorer. ; Represents the restored image The probability value output by the input discriminator;

[0093] Minimizing this loss term makes the reconstructed discriminator... It is impossible to distinguish between generated and real images. Therefore, the adversarial loss term ensures that the generator can gradually approximate the distribution of healthy samples when generating images, thereby improving the generation and restoration quality of abnormal images.

[0094] The other steps and parameters are the same as those in any of the specific implementation methods one to five.

[0095] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: in step 2), the L1 loss term is constructed. The mathematical expression is shown below:

[0096] (2)

[0097] in,

[0098] Represents the L1 loss term;

[0099] Hole-view images indicating a healthy aircraft engine;

[0100] express Anomaly images output after inputting into a random simulator;

[0101] Representing abnormal images The restored image output after inputting the pixel restorer ;

[0102] By minimizing the L1 loss term, also known as the absolute error loss, which measures the model's error by calculating the absolute value of the deviation between the predicted and true values, this loss term is used to evaluate the restored image output by the pixel restorer. Compared with the original health image The L1 loss method minimizes pixel-level differences between the original and restored images, ensuring the accuracy of the restored image at the pixel level. Compared to the common L2 loss, the L1 loss method is more suitable for image restoration tasks because it penalizes anomalous pixels more directly and can minimize pixel deviations between the restored and original images to preserve image edge details, thereby improving restoration accuracy.

[0103] The other steps and parameters are the same as those in any of the specific implementation methods one to six.

[0104] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that: in step 3), the structural similarity loss term is constructed. The specific process is as follows:

[0105] 31) Calculate structural similarity The mathematical expression is shown below:

[0106] (3)

[0107] in,

[0108] Indicates structural similarity;

[0109] Hole-view images indicating the health of an aircraft engine The average brightness;

[0110] Represents the restored image The average brightness;

[0111] express and covariance;

[0112] Hole-view images indicating the health of an aircraft engine Standard deviation;

[0113] Represents the restored image Standard deviation;

[0114] Represents a constant; Represents a constant;

[0115] Compared to pixel-level L1 loss, structural similarity index can reflect the differences between images macroscopically and holistically. Therefore, this invention incorporates it into the damage function as an effective supplement to adversarial loss and L1 loss.

[0116] 32) Based on structural similarity Calculate the structural similarity loss term; the mathematical expression is shown in the following formula:

[0117] (4)

[0118] in,

[0119] Represents the structural similarity loss term;

[0120] Representing abnormal images The restored image output after inputting the pixel restorer ;

[0121] By minimizing similarity loss, the pixel restorer can ensure the restored image... With the original image Maintaining structural consistency improves the overall quality of the restoration.

[0122] This loss term is used to evaluate the restored image output by the pixel restorer. With the original image Similarity in global structure. This invention uses the Structural Similarity (SSIM) metric to achieve this goal. This metric is commonly used to measure the similarity between two images in terms of brightness, contrast, and structure, and is also suitable for evaluating the visual perceptual consistency of images.

[0123] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0124] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that: in 4) the adversarial loss term... L1 loss term Structural similarity loss term Construct a ternary composite loss function The mathematical expression is shown below:

[0125] (5)

[0126] in, , , These are the counter-loss terms. L1 loss term With structural similarity loss term The weighting coefficients are used to balance the weights of the various loss terms.

[0127] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0128] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that: in step three, the borehole image data of the aircraft engine under test is based on the trained BiGenAD network model. Perform category prediction to obtain borehole image data of the aircraft engine under test. Is it a healthy image or an abnormal image?

[0129] The specific process is as follows:

[0130] Step 31: Transfer the borehole image data of the aircraft engine to be tested. The pixel restorer (second generator) in the pre-trained BiGenAD network model is input, and the pixel restorer (second generator) outputs the restored image. ;

[0131] Evaluating the deviation between the generated restored image and the input image is crucial. In order to accurately measure the restoration deviation, this invention designs a multi-scale structural similarity as an evaluation index based on the structural similarity mentioned above, thereby providing a more comprehensive deviation evaluation.

[0132] Step 3.2: Transfer the borehole image data of the aircraft engine to be tested. Corresponding restored image Perform separately The next downsampling yields At different resolutions Corresponding restored image ;

[0133] Step 33: Calculate at different resolutions Corresponding restored image The structural similarity between them is expressed as:

[0134] (6)

[0135] in,

[0136] Indicates the first At each resolution and Structural similarity values ​​between them; ;

[0137] Indicates the first At a resolution of [resolution value], the borehole image data of the aircraft engine under test The average brightness;

[0138] Indicates the first At each resolution Corresponding restored image The average brightness;

[0139] Indicates the first At each resolution and covariance;

[0140] Indicates the first At a resolution of [resolution value], the borehole image data of the aircraft engine under test Standard deviation;

[0141] Indicates the first At a resolution of [resolution value], the borehole image data of the aircraft engine under test Corresponding restored image Standard deviation;

[0142] Represents a constant; Represents a constant;

[0143] Steps 3 and 4: Process the structural similarity at different resolutions to obtain multi-scale similarity. This is used to measure the difference in local detail and global structure between two images; the calculation formula is shown below:

[0144] (7)

[0145] in, Indicates the first Weights at each resolution; Indicates the total number of scales;

[0146] In the calculation results, the closer the multi-scale similarity value is to 1, the better the restored image. With the borehole image data of the aircraft engine under test The higher the similarity, that is, the smaller the difference between the two, the better the borehole image data of the tested aero-engine. It should be a healthy sample;

[0147] Conversely, the closer the multi-scale similarity value is to 0, the better the restored image. With the borehole image data of the aircraft engine under test The lower the similarity, that is, the greater the difference between the two, the more likely the borehole image data of the aircraft engine under test is to be of poor quality. There may be damage.

[0148] The core motivation behind this design is that, through adversarial training, the pixel restorer can reconstruct abnormal samples as closely as possible to a healthy state. When the input is a healthy sample, the distribution deviation between the restored image and the original image is minimal, thus enabling effective screening. Compared to traditional anomaly detection methods, this design cleverly utilizes the irreducibility of healthy samples during the restoration process to distinguish between healthy and abnormal samples, providing an innovative and efficient solution for anomaly screening in borehole images.

[0149] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0150] The beneficial effects of the present invention are verified using the following embodiments:

[0151] Example 1: Implementation plan for anomaly screening in borehole images of aero-engines based on the BiGenAD model; the specific process is as follows:

[0152] This embodiment is built based on real aero-engine borehole video stream data (Turbo19 dataset). First, data preprocessing and environment deployment are performed. The BiGenAD model is built on the PyTorch deep learning framework and deployed on a high-performance workstation equipped with an NVIDIA RTX 3080 Ti GPU and an AMD Ryzen 9 5900X CPU.1 During the data input stage, the resolution of all borehole images is uniformly adjusted to [resolution value missing]. Pixel values ​​are normalized to eliminate the effects of uneven lighting and accelerate model convergence. A multi-stage generative adversarial network (GAN) is then constructed, comprising a random simulator, a pixel restorer, and a restoration discriminator. During training, the Adam optimizer is used to minimize a ternary composite loss function comprising adversarial loss, L1 pixel loss, and structural similarity loss. Specific hyperparameters are set as follows: initial learning rate of 0.0002, with a 10-fold decay strategy every 200 iterations; batch size of 32; and a total of 5000 iterations. To balance generative diversity and restoration accuracy, the weighting coefficients of the loss function are set as follows: adversarial loss weight... L1 loss weights Structural similarity loss weights During the testing and screening phase, the image to be detected is input into a trained pixel restorer to obtain a restored image, and the multi-scale structural similarity (MS-SSIM) score between the two is calculated. If the score is lower than a preset threshold, the image is determined to have potential damage; otherwise, it is determined to be healthy. Experimental results show that this implementation achieves an AUC of 97.33 and an accuracy of 99.12% on the Turbo19 dataset, with an average inference time of only 0.095 seconds per image, which can meet the real-time detection requirements in aviation maintenance scenarios.

[0153] Example 2: A generalized detection solution applied to different industrial scenarios; the specific process is as follows:

[0154] This embodiment demonstrates the application of this method to other industrial texture backgrounds, taking magnetic tile surface defect detection (Magnetic-Tile dataset) as an example. This embodiment maintains the same network structure and core hyperparameters (such as learning rate, batch size, etc.) as Embodiment 1, only replacing the training data with healthy magnetic tile surface images. The model learns the specific texture distribution of the magnetic tile surface and uses a pixel reconstructor to detect the "irreconstructibility" of defects such as cracks and pores. Experiments show that the scheme achieves an AUC of 96.52 and an accuracy of 95.33% on the magnetic tile dataset. This proves that the present invention is not only applicable to aerial borehole images, but also has good generalization ability and engineering application value in processing industrial product defect detection tasks with different lighting conditions and texture features.

[0155] Example 3: Implementation plan for surface defect detection of hot-rolled steel strip; the specific process is as follows:

[0156] This embodiment further demonstrates the application of the method of the present invention in the detection of surface defects in complex industries, specifically for the identification of surface defects in hot-rolled steel strip (based on the NEU-Seg dataset). Unlike borehole images of aero-engines, the damage types on the surface of hot-rolled steel strip are characterized by "large intra-class differences and high inter-class similarities," posing a greater challenge to the model's feature extraction and distribution modeling capabilities.

[0157] In this embodiment, the BiGenAD model architecture remains consistent with that of Embodiment 1, still including three core components: a random simulator, a pixel restorer, and a restoration discriminator. The training data is replaced with images of healthy hot-rolled steel strip surfaces. The model primarily performs generalization learning for six typical surface damage types: rolling marks, indentations, shrinkage cavities, scratches, and pits. During training, the random simulator introduces perturbations into the latent space to generate pseudo-abnormal images containing textured defects similar to those on the steel strip surface; the pixel restorer is forced to learn the consistent texture distribution of the steel strip surface to achieve the repair of abnormal areas.

[0158] Experimental results demonstrate that the method of this invention exhibits extremely high robustness in this scenario. Quantitative testing on the NEU-Seg dataset shows that the model achieves an AUC of 98.08, an accuracy of 98.20%, and a precision of 97.62%. This indicates that the BiGenAD model can not only adapt to the weak texture features of borehole images, but also accurately utilize the "irreducibility" difference to achieve anomaly screening in industrial scenarios such as hot-rolled steel strip with complex textures and varied defect morphologies by learning the distribution characteristics of healthy samples. This fully verifies the effectiveness and generalization ability of the technical solution of this invention in a wide range of industrial visual inspection tasks.

[0159] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for screening anomalous borehole images of aero-engines based on multi-stage generative adversarial methods, characterized in that: The specific process of the method is as follows: Step 1: Obtain a healthy dataset of borehole images of aircraft engines; Step 2: Construct the BiGenAD network model. Train the BiGenAD network model based on a healthy aero-engine borehole image dataset to obtain a trained BiGenAD network model. The BiGenAD network model includes a random simulator, a pixel restorer, and a restore discriminator; Step 3: Using the trained BiGenAD network model to analyze the borehole image data of the aircraft engine under test. Perform category prediction to obtain borehole image data of the aircraft engine under test. Is it a healthy image or an abnormal image? 2. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 1, characterized in that: In step two, a BiGenAD network model is constructed. The BiGenAD network model is trained based on a healthy aero-engine anomaly borehole image dataset to obtain a trained BiGenAD network model. The BiGenAD network model includes a random simulator, a pixel restorer, and a restore discriminator; The specific process is as follows: Step 21: Centralize the healthy air-to-ground engine borehole image data. Input a random simulator, output a random simulator Corresponding abnormal image ; Step 22: The output of the random simulator Corresponding abnormal image Input pixel restorer, pixel restorer output Corresponding restored image ; Steps two and three: The output of the restorer... Corresponding restored image The input is a discriminator, which outputs a probability value in the interval [0 -1]. This probability value represents the restored image. Hole-view images of healthy aircraft engines Confidence level; The discriminator uses a ResNet34 network structure; Step 24: Repeat steps 21 to 23 until the ternary composite loss function converges to obtain the trained BiGenAD network model.

3. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 2, characterized in that: In step two, the healthy aircraft engine borehole image dataset is collected. Input a random simulator, output a random simulator Corresponding abnormal image ; The specific process is as follows: Step 211: Healthy borehole images of aircraft engines Input encoder, encoder output feature; The encoder is an encoder in a self-encoder; Step 212: Add random noise to the features output by the encoder to obtain the features after adding random noise; Steps 2-3: Input the features with added random noise into the decoder; the decoder outputs the abnormal image. ; Decoder outputs abnormal image As the output of a random simulator; Abnormal images With healthy aircraft engine borescope images They have the same dimensions; The decoder is the decoder in the autoencoder.

4. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 3, characterized in that: In step two, the output of the random simulator will be... Corresponding abnormal image Input pixel restorer, pixel restorer output Corresponding restored image ; The specific process is as follows: Abnormal images output by the random simulator Input encoder, encoder output image latent feature space; The latent feature space of the encoder output image is input into the decoder, and the decoder outputs a healthy air-to-air engine borehole image. The restored image ; The encoder is an encoder in a self-encoder; The decoder is the decoder in the autoencoder.

5. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 4, characterized in that: The process of obtaining the ternary composite loss function in step two of the above steps is as follows: 1) Constructing adversarial loss terms ; 2) Construct the L1 loss term ; 3) Constructing a structural similarity loss term ; 4) Based on adversarial loss terms L1 loss term Structural similarity loss term Construct a ternary composite loss function .

6. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 5, characterized in that: In step 1), an adversarial loss term is constructed. ; The mathematical expression is as follows: (1) in, Indicates the loss against; Indicates a random simulator; Indicates pixel restorer; Hole-view images indicating a healthy aircraft engine; Indicates the discriminator; Hole-view images indicating the health of an aircraft engine The probability value output by the input discriminator; This indicates a demand for expectation; express An abnormal image output after inputting into a random simulator; Representing abnormal images The probability value output by the input discriminator; This represents the restored image output by the pixel restorer. ; Represents the restored image The probability value output after inputting into the discriminator.

7. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 6, characterized in that: In step 2), the L1 loss term is constructed. The mathematical expression is shown below: (2) in, Represents the L1 loss term; Hole-view images indicating a healthy aircraft engine; express An abnormal image output after inputting into a random simulator; Representing abnormal images The restored image output after inputting the pixel restorer .

8. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 7, characterized in that: In step 3), a structural similarity loss term is constructed. ; The specific process is as follows: 31) Calculate structural similarity The mathematical expression is shown below: (3) in, Indicates structural similarity; Hole-view images indicating the health of an aircraft engine The average brightness; Represents the restored image The average brightness; express and covariance; Hole-view images indicating the health of an aircraft engine Standard deviation; Represents the restored image Standard deviation; Represents a constant; Represents a constant; 32) Based on structural similarity Calculate the structural similarity loss term; the mathematical expression is shown in the following formula: (4) in, Represents the structural similarity loss term; Representing abnormal images The restored image output after inputting the pixel restorer .

9. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 8, characterized in that: The adversarial loss term mentioned in 4) L1 loss term Structural similarity loss term Construct a ternary composite loss function The mathematical expression is shown below: (5) in, , , These are the counter-loss terms. L1 loss term With structural similarity loss term The weighting coefficients.

10. The method for screening aero-engine anomaly borehole images based on multi-stage generative adversarial methods according to claim 9, characterized in that: In step three, the borehole image data of the aircraft engine under test is based on the trained BiGenAD network model. Perform category prediction to obtain borehole image data of the aircraft engine under test. Is it a healthy image or an abnormal image? The specific process is as follows: Step 31: Transfer the borehole image data of the aircraft engine to be tested. The pixel restorer in the trained BiGenAD network model is input, and the pixel restorer outputs the restored image. ; Step 3.2: Transfer the borehole image data of the aircraft engine to be tested. Corresponding restored image Perform separately The next downsampling yields... At different resolutions Corresponding restored image ; Step 33: Calculate at different resolutions Corresponding restored image The structural similarity between them is expressed as: (6) in, Indicates the first At each resolution and Structural similarity values ​​between them; ; Indicates the first At a resolution of [resolution value], the borehole image data of the aircraft engine under test The average brightness; Indicates the first At each resolution Corresponding restored image The average brightness; Indicates the first At each resolution and covariance; Indicates the first At a resolution of [resolution value], the borehole image data of the aircraft engine under test Standard deviation; Indicates the first At a resolution of [resolution value], the borehole image data of the aircraft engine under test Corresponding restored image Standard deviation; Represents a constant; Represents a constant; Steps 3 and 4: Process the structural similarity at different resolutions to obtain multi-scale similarity. The calculation formula is shown below: (7) in, Indicates the first Weights at each resolution; Indicates the total number of scales; The closer the multi-scale similarity value is to 1, the better the restored image. With the borehole image data of the aircraft engine under test The higher the similarity, that is, the smaller the difference between the two, the better the borehole image data of the aircraft engine under test. It should be a healthy sample; Conversely, the closer the multi-scale similarity value is to 0, the better the restored image. With the borehole image data of the aircraft engine under test The lower the similarity, that is, the greater the difference between the two, the more likely the borehole image data of the aircraft engine under test is to be of poor quality. Damage exists.

Images