An aero-engine abnormality detection method based on self-driven hierarchical learning
The aero-engine anomaly detection method based on self-driven hierarchical learning utilizes normal images for preprocessing and various self-supervised learning tasks, solving the problems of missed detections and resource waste in aero-engine inspection, and achieving efficient and stable anomaly detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEAST FORESTRY UNIV
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for detecting anomalies in aero-engines are prone to missing detections, and there is a waste of resources during model training, especially the difficulty in obtaining defect samples and the high cost of manual annotation.
A self-driven hierarchical learning approach is adopted. Through unsupervised pre-training and downstream fine-tuning, normal images of aero-engines are pre-processed to construct a backbone network and various self-supervised learning tasks, including local magnification discrimination, viewpoint change adaptation, and instance push-pull, to achieve the detection of aero-engine anomalies.
Model training can be completed without relying on defect samples and manual annotation, significantly reducing data preparation and system deployment costs, improving the ability to detect minor anomalies and complex structural anomalies, avoiding missed detections, and making the detection process simple and stable.
Smart Images
Figure CN122115969A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aero-engine anomaly detection, and in particular to an aero-engine anomaly detection method based on self-driven hierarchical learning. Background Technology
[0002] As a critical component of aircraft, aero engines play an indispensable role in ensuring safe flight operations. In routine maintenance, timely detection and management of potential defects within these engines are crucial to preventing safety accidents. Given the complex and non-removable structure of engines, endoscopic technology has become the primary method for inspecting critical internal components such as impellers and blades without disassembling the engine casing. While this technology allows for detailed inspections, it involves complex procedures and is prone to subjective judgment. Therefore, achieving objective judgment in engine anomaly assessment has become a major research focus in this field.
[0003] Currently, the maintenance and inspection of aero-engines primarily employ deep learning-based defect detection methods. These methods train the model based on a specific category distribution assumption, predicting that the sum of probabilities for all categories within the training sample distribution is 1. However, in actual inspections, aero-engine internal components are highly durable, and damage is rare. Therefore, the model training set does not contain a wide range of defect categories. Even for known engine defect categories, internal heterogeneity exists (such as tears of different shapes and sizes). Consequently, the training set cannot encompass all defect scenarios. During actual prediction, if a defect not included in the training set appears, existing aero-engine defect detection methods may treat it as a normal or unfamiliar image, leading to missed detections. Furthermore, the training set for existing aero-engine defect detection models is manually labeled, requiring a large number of professionals with specialized knowledge, resulting in resource waste during the training process. Summary of the Invention
[0004] This invention addresses the problems of missed detections and resource waste during model training in existing aero-engine anomaly detection methods by proposing a self-driven hierarchical learning-based aero-engine anomaly detection method.
[0005] An aero-engine anomaly detection method based on self-driven hierarchical learning includes the following steps:
[0006] A1. Acquire normal images of the aircraft engine and preprocess them to obtain preprocessed normal images of the aircraft engine.
[0007] A2. Use the pre-processed normal images of the aero-engine to perform unsupervised pre-training on the aero-engine anomaly detection model to obtain a pre-trained aero-engine anomaly detection model.
[0008] A3. Construct a downstream fine-tuning dataset, utilize the backbone network of the pre-trained aero-engine anomaly detection model to build the final aero-engine anomaly detection model, and train the final aero-engine anomaly detection model using the downstream fine-tuning dataset to obtain the trained final aero-engine anomaly detection model.
[0009] A4. Obtain the image of the aircraft engine to be tested, input the image of the aircraft engine to be tested into the trained final anomaly detection model of the aircraft engine, and obtain whether the image of the aircraft engine to be tested is a normal image or an abnormal image.
[0010] Furthermore, the preprocessing of the normal image of the aircraft engine in A1 specifically includes:
[0011] A1.1 Normal image of an aircraft engine Rotate the image at different angles to obtain a rotated image. ;
[0012] A1.2, Under normal aircraft engine image A preset area of size is randomly selected as the magnification area, and the magnification area is magnified at different magnifications to obtain a magnified image. ;
[0013] in, It's the rotation angle. It's the magnification factor. Indicates rotation angle A normal image of the engine. Indicates magnification factor A normal image of the engine.
[0014] Furthermore, the aero-engine anomaly detection model includes: a backbone network, a feature dimensionality reduction module, a feature decoupling module, a magnification determination module, a viewing angle change adaptation module, and an instance push-pull module;
[0015] The backbone network is used to extract features from magnified or rotated images to obtain image features. ,Will Input feature dimensionality reduction module and instance push-pull module;
[0016] in, Pick or , It is a feature of magnified images. It is a feature of rotated images;
[0017] The feature dimensionality reduction module is used to... Reduced to the preset dimension, obtaining the dimensionality reduction After dimensional reduction Input is fed into the feature decoupling module;
[0018] The feature decoupling module is used to decouple the dimensionality-reduced... The decoupled magnification feature or decoupled rotation feature is obtained by average decoupling along the channel dimension. The decoupled magnification feature is sent to the magnification determination module, and the decoupled rotation feature is sent to the viewpoint change adaptation module.
[0019] The magnification determination module is used to obtain a normal image of the aero-engine based on the decoupled magnification features. The probability of each magnification;
[0020] The perspective change adaptation module is used to obtain a normal image of the aircraft engine based on the decoupled rotation features. The probability of each rotation angle;
[0021] The push-pull method described above is used to determine the probability that different image features belong to the same normal image of an aircraft engine.
[0022] Furthermore, the feature dimensionality reduction module sequentially includes: a max pooling layer, a feature flattening layer, a fully connected layer, a batch normalization layer, and a ReLU activation function layer.
[0023] Furthermore, the magnification determination module includes a fully connected layer and a softmax layer in sequence.
[0024] Furthermore, the viewpoint change adaptation module includes a fully connected layer and a softmax layer in sequence.
[0025] Furthermore, the instance push-pull module includes: a generalization unit and a category determination unit;
[0026] The generalization unit is used for image features. Perform L2 generalization to obtain the generalized image features;
[0027] The category determination unit is used to obtain the probability that two generalized image features come from the same normal image of an aero-engine.
[0028] Furthermore, the category determination unit is used to obtain the probability that two generalized image features come from the same normal image of an aero-engine, specifically:
[0029]
[0030] in, It is an image go through Generalized image features after view transformation It is an image go through Generalized image features after view transformation yes The corresponding view belongs to the image category. The probability, , , It is an image label. It is an image go through Generalized image features after view transformation It is the total number of images. , All or .
[0031] Furthermore, the aero-engine anomaly detection model is trained unsupervised using the following loss function:
[0032]
[0033]
[0034]
[0035]
[0036] in, It is the overall loss function. , and These are weighting coefficients. It is a magnification category label. It represents the total number of magnification categories. It is the category of the true magnification of the magnified area of the image. It is the predicted probability of the true magnification category of the magnified area of the image. It is a balancing term. The loss is determined by the magnification. It is a rotation angle category label. This represents the total number of rotation angle categories. This is the actual rotation category in the image. It is a regulating item. It is the predicted probability of the image rotation angle category. It is perspective distortion loss. It represents the structural similarity of features between two images in a positive sample pair. It is the structural similarity of two image features in a negative sample pair. It is a temperature parameter;
[0037] The positive sample pair is either a normal image of the same aero-engine with two rotated images at different rotation angles or a normal image of the same aero-engine with two magnified images at different magnifications.
[0038] The negative sample pair is either a rotated image corresponding to two normal images of two different aero engines, or a magnified image corresponding to two normal images of two different aero engines.
[0039] Furthermore, in A3, the downstream fine-tuning dataset is constructed, and the final aero-engine anomaly detection model is built using the backbone network of the pre-trained aero-engine anomaly detection model. The final aero-engine anomaly detection model is then trained using the downstream fine-tuning dataset to obtain the trained final aero-engine anomaly detection model. Specifically:
[0040] A3.1. Obtain aircraft engine image samples and assign labels to each aircraft engine sample. Use the aircraft engine image samples and labels to form a downstream fine-tuning dataset.
[0041] The labels include: normal and abnormal;
[0042] A3.2 Freeze the backbone network parameters in the pre-trained aero-engine anomaly detection model, and splice the pre-trained backbone network with the binary classification head to obtain the final aero-engine anomaly detection model.
[0043] A3.3. Use the downstream fine-tuning dataset to train the final anomaly detection model for aero-engines and fine-tune the binary classifier head to obtain the trained final anomaly detection model for aero-engines.
[0044] The beneficial effects of this invention are as follows:
[0045] This invention proposes a self-driven hierarchical learning-based method for detecting aero-engine anomalies. It eliminates the need for any defect sample data or manual annotation, utilizing only unlabeled healthy samples for model training and aero-engine anomaly identification. This effectively solves the problems of difficult-to-obtain defect samples, high manual annotation costs, and limited engineering applications, significantly reducing data preparation and system deployment costs and avoiding waste of human resources. By designing a series of progressively advanced self-supervised learning tasks from local to global and from single-viewpoint to multi-viewpoint, this invention enables the final aero-engine anomaly detection model to fully learn the feature representations of the aero-engine's internal structure at different scales and perspectives. This effectively improves the perception and representation capabilities of minute and complex structural anomalies, overcoming the difficulty of identifying unknown anomalies outside the training distribution in traditional supervised learning methods and avoiding missed detections. Furthermore, this invention eliminates the need for additional manual intervention or complex post-processing steps in the prediction stage, directly utilizing the trained backbone network for anomaly discrimination. It boasts advantages such as a simple implementation process, high system stability, and ease of engineering deployment. Attached Figure Description
[0046] Figure 1 Diagram showing the operation of an aircraft engine endoscopic inspection.
[0047] Figure 2 This is an overall framework diagram of the present invention;
[0048] Figure 3 A hierarchical framework diagram for self-supervised learning tasks;
[0049] Figure 4 This is a schematic diagram of an endoscope inspection of an aircraft engine.
[0050] Figure 5 This is a flowchart of the present invention;
[0051] Figure 6 This is a schematic diagram of the hierarchical structure of self-supervised tasks;
[0052] Figure 7 This is a sample image from a dataset of endoscopic images of aircraft engines. Detailed Implementation
[0053] Unsupervised anomaly detection techniques offer significant advantages. Primarily, they excel at identifying unfamiliar or unlabeled anomaly categories, crucial for addressing novel or rare defects that may arise in real-world scenarios. Furthermore, unsupervised algorithms do not require high-quality manual annotation of training samples, providing a more flexible and efficient detection method. Currently, unsupervised anomaly detection methods can be categorized into three types: reconstruction-based methods, feature-learning-based methods, and self-supervised learning methods. Reconstruction-based methods identify anomalies by minimizing reconstruction errors in normal data. Feature-learning-based methods function by learning latent features within the data distribution. On the other hand, self-supervised learning methods enable models to identify inherent data features and detect anomalies by creating and solving pre-tasks. This invention analyzes the complex characteristics of endoscopic images. Figure 1 As shown, due to the narrow and complex internal structure of aero-engines, endoscopic images used for this purpose often contain impellers and blades at different angles and focal lengths, resulting in high image heterogeneity and strong viewpoint dependence. Furthermore, the precision manufacturing characteristics of aero-engines mean that even minute defects can have serious consequences; therefore, understanding and characterizing subtle abnormal areas is crucial. A schematic diagram of the structure for endoscopic inspection of an aero-engine is shown below. Figure 4 The diagram illustrates the engine body, the industrial endoscope probe, and its insertion path, serving to demonstrate a typical application scenario for this invention. This invention incorporates three progressive self-supervised learning tasks. The invention will now be described in conjunction with specific implementation methods.
[0054] Specific implementation method one: as follows Figures 5-6 As shown, the specific process of an aero-engine anomaly detection method based on self-driven hierarchical learning in this embodiment is as follows:
[0055] A1. Acquire normal images of the aircraft engine and preprocess them to obtain preprocessed normal images of the aircraft engine, specifically as follows:
[0056] A1.1 Normal image of an aircraft engine Rotate the image at different angles to obtain a rotated image. ;
[0057] A1.2 Randomly select normal images of aircraft engines 5% to 10% of the total area is designated as the magnification area, and then this area is magnified at different magnifications to obtain a magnified image. ;
[0058] in, , It's the rotation angle. , It's the magnification factor. Indicates rotation angle A normal image of the engine. Indicates magnification factor A normal image of the engine.
[0059] Among them, two rotated images with different rotation angles corresponding to the normal image of the same aero-engine or two magnified images with different magnifications corresponding to the normal image of the same aero-engine are positive samples;
[0060] Rotated images corresponding to normal images of two different aero engines, or magnified images corresponding to normal images of two different aero engines, are negative samples;
[0061] A2. Use the pre-processed normal images of the aero-engine to perform unsupervised pre-training on the aero-engine anomaly detection model to obtain a pre-trained aero-engine anomaly detection model.
[0062] like Figure 2 As shown, the aero-engine anomaly detection model includes: a backbone network, a feature dimensionality reduction module, a feature decoupling module, a magnification determination module, a viewing angle change adaptation module, and an instance push-pull module;
[0063] The backbone network uses Inception v3 to extract features from magnified or rotated images, thereby obtaining image features. ,Will Input feature dimensionality reduction module and instance push-pull module;
[0064] in, Pick or , It is a feature of magnified images. It is a feature of rotated images;
[0065] The feature dimensionality reduction module includes, in sequence: a max pooling layer, a feature flattening layer, a fully connected layer, a batch normalization layer, and a ReLU activation function layer;
[0066] The feature dimensionality reduction module is used for... Perform dimensionality reduction to obtain the dimensionality-reduced result. After dimensional reduction Input is fed into the feature decoupling module;
[0067] Dimensional reduction It has 256 dimensions;
[0068] The feature decoupling module is used to decouple the dimensionality-reduced... The decoupled magnification feature or decoupled rotation feature is obtained by average decoupling along the channel dimension. The decoupled magnification feature is sent to the magnification determination module, and the decoupled rotation feature is sent to the viewpoint change adaptation module.
[0069] The magnification determination module is used to obtain a normal image of the aero-engine based on the decoupled magnification features. The probability of each magnification;
[0070] The magnification determination module includes a fully connected layer and a softmax layer in sequence.
[0071] The perspective change adaptation module is used to obtain a normal image of the aircraft engine based on the decoupled rotation features. The probability of each rotation angle;
[0072] The viewpoint change adaptation module includes a fully connected layer and a softmax layer in sequence.
[0073] The instance push-pull module includes: a generalization unit and a category determination unit;
[0074] The generalization unit is used for image features. Perform L2 generalization to obtain the generalized image features;
[0075] The features Pick or ;
[0076] The category determination unit outputs the probability that two image features come from the same normal image of an aero-engine, i.e., the probability of a positive sample, based on the similarity index between the images. Specifically:
[0077]
[0078] in, It is an image go through Generalized image features after view transformation It is an image go through Generalized image features after view transformation yes The corresponding view belongs to the image category. The probability, , It is an image label. It is an image go through Generalized image features after view transformation It is the total number of images. , All or .
[0079] A3. Construct a downstream fine-tuning dataset, use the backbone network in the pre-trained aero-engine anomaly detection model to construct the final aero-engine anomaly detection model, and train the final aero-engine anomaly detection model using the downstream fine-tuning dataset to obtain the trained final aero-engine anomaly detection model. Specifically:
[0080] A3.1. Obtain aircraft engine image samples and assign labels to each aircraft engine sample. Use the aircraft engine image samples and labels to form a downstream fine-tuning dataset.
[0081] The labels include: normal and abnormal;
[0082] A3.2 Freeze the backbone network parameters in the pre-trained aero-engine anomaly detection model, and splice the pre-trained backbone network with the binary classification head to obtain the final aero-engine anomaly detection model.
[0083] A3.3. Train the final anomaly detection model for aero-engines using the downstream fine-tuning dataset, fine-tune the binary classifier head, and obtain the trained final anomaly detection model for aero-engines.
[0084] A4. Acquire the image of the aircraft engine to be tested, input the image of the aircraft engine to be tested into the trained final anomaly detection model of the aircraft engine for forward inference, the model outputs anomaly score, and the determination is made based on the anomaly score and the threshold to obtain whether the image of the aircraft engine to be tested is a normal image or an abnormal image.
[0085] The overall framework of the method proposed in this invention is as follows: Figure 2 As shown. For each image, we perform two key transformations: random magnification (magnification factor of 1.0, 1.2, 1.5, or 2.0), and rotation transformation. (Rotation angles are 0°, 90°, 180°, or 270°). Subsequently, the backbone network maps the input to higher-order feature vectors, and then decouples these higher-order feature vectors. The decoupled features are optimized end-to-end through three progressive learning tasks (magnification determination, perspective transformation adaptation, and instance push-pull), a process achieved by a comprehensive objective function. The overall framework of this invention consists of three self-supervised tasks: magnification determination, perspective transformation adaptation, and instance push-pull, such as... Figure 6 As shown, Figure 6This paper demonstrates a hierarchical structure from "local magnification discrimination" to "viewpoint change adaptation" and then to "instance push-pull," illustrating the invention's layered learning approach from local to global and from single-view to multi-view. The first task is to magnify random regions within an image to accommodate the detail sensitivity of the internal structure of an aero-engine. The second task is to learn rotation-invariant features of anomalous regions to adapt to the diversity of blade arrangements. The third task addresses the diversity of endoscopic viewpoints and objects, providing robust transformation-invariant representations. These tasks, from local to global and from single-view learning to multi-view association, comprehensively enhance the model's understanding of potential anomalies. Through a customized objective function, these learning tasks are integrated into a unified framework for end-to-end collaborative optimization.
[0086] Specific Implementation Method Two: The loss function of the magnification determination module is as follows:
[0087]
[0088] in, It is a magnification category label. It represents the total number of magnification categories. It is the category of the true magnification of the magnified area of the image. It is the predicted probability of the true magnification category of the magnified area of the image. It is a balancing term. The loss is determined by the magnification.
[0089] In this step, the magnification is labeled 1, 2, 3, and 4, which correspond to magnifications of 1.0, 1.2, 1.5, and 2.0, respectively.
[0090] The magnification judgment task of this invention is designed based on the sensitivity to the details of the internal structure of aero-engines. In actual inspection, identifying potential anomalies often requires in-depth analysis of the texture details of specific regions. This task forces the model to learn to extract key information from subtle differences in local regions, thereby enhancing its ability to identify subtle anomaly regions. Existing research has proposed various image anomaly region learning tasks, such as random distortion, local blurring, and image completion. However, these methods often alter the original pixel structure of the image, making it difficult to accurately understand subtle anomaly regions. Therefore, this invention introduces a local magnification judgment task. This task is inspired by the principle of using a magnifying glass—observing details by magnifying a specific region (…). Figure 3 Specifically, from the original image A rectangular region is randomly selected as the magnification area, with its size ranging from 5% to 10% of the total area of the original image. This design is based on the characteristic that the high precision of blade manufacturing leads to potentially extremely subtle details: a smaller dynamic magnification area can capture sufficient structural details while preserving necessary contextual information. Crucially, this magnification method does not alter the pixel structure, maintaining texture consistency within the region, enabling the model to effectively learn subtle local changes.
[0091] Specific implementation method three: The loss function of the viewpoint change adaptation module is as follows:
[0092]
[0093] in, It is a rotation angle category label. This represents the total number of rotation angle categories. This is the actual rotation category in the image. It is a regulating item. It is the predicted probability of the image rotation angle category. It is perspective distortion loss.
[0094] In this step, the rotation angle categories are labeled 1, 2, 3, and 4, which correspond to magnifications of 0°, 90°, 180°, and 270°, respectively.
[0095] The perspective change adaptation module is based on a deep understanding of the diversity of how internal components of an aero-engine appear from different perspectives. Due to the variety of endoscope insertion paths and angles, the same component may appear in different orientations at different angles. This orientation diversity requires the detection system to understand and adapt to different perspectives. By training a model to predict the rotation state of the image, it can recognize and adapt to these changes in orientation, thereby improving the ability to understand and identify complex structures.
[0096] Specific implementation method four: The loss function of the push-pull module is as follows:
[0097]
[0098] in, It represents the structural similarity of features between two images in a positive sample pair. It is the structural similarity of two image features in a negative sample pair. It is a temperature parameter.
[0099] In this embodiment, although magnification determination and perspective variation adaptation can enhance the model's ability to grasp the details and orientation of internal components of aero-engines, these methods mainly focus on feature representation at the single-frame image level, which is insufficient to cope with the diverse detection perspectives and photographic instances in endoscopic images. Therefore, an instance push-pull task is designed to provide robust transformation-invariant feature representations for endoscopic images. This task is based on the key assumption that even after different data transformations, images captured from the same perspective should maintain high consistency in the feature space; conversely, images from different perspectives should be clearly distinguishable. Existing research shows that contrastive analysis can achieve similar goals by learning feature representations through contrastive loss, achieving significant results in natural image processing. However, the closed and narrow environment of endoscopic examination results in image differences that are much smaller than those in natural images of open scenes. Therefore, common contrastive losses such as Sim-CLR cannot be directly applied to this task and may even lead to model convergence failure. To solve this problem, this invention designs a refined instance push-pull mechanism (… Figure 3 First, L2 norm normalization is applied to the output features of the backbone network to ensure that the norm of each feature vector is 1. This step not only helps stabilize the training process but also improves the consistency of the model when handling different scaling and rotation transformations. The goal of this invention is to maximize similarity in positive sample pairs while minimizing similarity in negative sample pairs, i.e., "pushing away" the negative class and "pulling closer" the positive class. This process is achieved by minimizing the sum of the negative log-likelihood values of all negative sample pairs within a batch.
[0100] Specific Implementation Method Five: The total loss function of the aero-engine anomaly detection model is as follows:
[0101]
[0102] in, It is the overall loss function. , and It is the weighting coefficient.
[0103] The total loss function in this step promotes task-based collaborative learning, enabling the model to be trained and optimized end-to-end.
[0104] Compared with the prior art, the present invention has the following advantages:
[0105] (1) It does not rely on defect samples and manual annotation, and is suitable for rare defect scenarios:
[0106] This invention uses only unlabeled healthy endoscopic images during the training phase and drives feature learning through a self-supervised task. It eliminates the need to collect and label a large number of defect samples, significantly reducing data preparation costs. It is particularly suitable for aero-engine scenarios where actual defect samples are extremely scarce.
[0107] (2) Able to identify novel defects not found in the training set:
[0108] By learning the feature distribution of healthy samples, this invention identifies samples that deviate from the healthy distribution as abnormal, thereby detecting potential new or morphologically different defects, overcoming the limitation of traditional multi-class classification models that "only recognize defects that have been seen".
[0109] (3) Learning hierarchical features from local to global perspectives to enhance sensitivity to minor defects:
[0110] The local magnification discrimination task simulates the process of using a "magnifying glass" in actual maintenance, guiding the model to focus on local texture changes and minute geometric deformations on the blade surface, thus enabling it to better identify minute defects such as scratches, tears, and dents. Experimental results show that on an aero-engine endoscopic image dataset, the method of this invention significantly outperforms various existing industrial anomaly detection methods in terms of AUC, accuracy, precision, and recall.
[0111] (4) It has good robustness to changes in viewing angle and association with multiple views:
[0112] The viewpoint change adaptive task explicitly models the structural invariance under different rotation angles. Combined with the instance push-pull task, which brings the same instance closer to multiple views and pushes away different instances in the feature space, the model can maintain stable detection performance in scenarios with changes in endoscope insertion angle, lighting changes, and different component combinations.
[0113] (5) Simple structure, end-to-end optimization, and easy to implement in engineering:
[0114] This invention integrates three self-supervised tasks and a feature extraction backbone network within a unified framework, enabling end-to-end training and inference using mainstream deep learning frameworks.
[0115] (6) The method of the present invention has been validated on multiple public industrial datasets such as the magnetic tile surface defect data MT, the texture defect dataset DAGM2007, and the hot rolled steel strip surface defect dataset NEU-Seg. The method has achieved better detection performance than the comparison algorithm, indicating that the method has good versatility and can be extended to other industrial surface defect detection tasks.
[0116] Example 1: To verify the beneficial effects of the present invention, the following experiments were conducted:
[0117] System and Environment Configuration
[0118] This embodiment uses an industrial endoscope to image components such as the compressor of an aero-engine. The endoscope has a light source and an imaging module at its front end. The images are transmitted to a computer equipped with a GPU through a data acquisition module, where a deep learning algorithm program is run, and the detection results and alarm information are output on the display terminal.
[0119] Dataset Construction and Preprocessing
[0120] Several endoscopic images of the engine's internal components, both healthy and defective, were collected. Only the healthy images were used for model training, while the defective images were used for testing and evaluation. All images were uniformly scaled to a predetermined size and subjected to basic data augmentation such as normalization and random flipping. The models were then divided into training, validation, and test sets. In this embodiment, the Turbo19 dataset was used for model training, validation, and testing. The Turbo19 dataset is shown below. Figure 7 As shown, Figure 7 It displays typical images of defect-free blades as well as blades with different defect types such as curling, dents, scratches, and tears.
[0121] Model training process
[0122] (1) Input the healthy images in the training set into the random local magnification module and the rotation transformation module to generate views with different magnification and different angles;
[0123] (2) The above view is fed into a convolutional neural network with Inception v3 as the backbone, and extracted and compressed into a low-dimensional feature vector;
[0124] (3) For the local magnified view, the magnification classification training is carried out through the local magnification discrimination branch; for the rotated view, the rotation angle classification training is carried out through the viewpoint adaptation branch; at the same time, positive sample pairs of multiple views of the same image and negative sample pairs between different images are constructed, and comparative learning is carried out through the instance push-pull branch.
[0125] (4) The network is optimized end-to-end using a total loss function that includes three self-supervised losses. The weights of each loss and the training hyperparameters are adjusted using the validation set to obtain a well-trained model.
[0126] Anomaly Detection Process
[0127] During the detection phase, the endoscopic image to be inspected is directly input into the trained network to extract feature vectors, which are then compared with a feature distribution model built based on healthy samples to calculate the probability of anomalies. When the probability of anomalies exceeds a preset threshold, the image is determined to be abnormal and an alarm is issued. If necessary, a heatmap can be generated based on class activation mapping to visually indicate suspected defect areas.
[0128] Application effect
[0129] On actual aero-engine endoscopic image data, the method in this embodiment outperforms various unsupervised or supervised anomaly detection algorithms in terms of AUC, accuracy, and recall. It can effectively detect small and novel defects without relying on defect annotation, making it suitable for engineering deployment in engine endoscopic maintenance scenarios.
[0130] Example 2:
[0131] Building upon Example 1, the method of this invention is directly applied to publicly available industrial datasets such as those for magnetic tile surface defects, textured surface defects, and hot-rolled steel strip surface defects. By replacing the training data and appropriately adjusting the hyperparameters, without modifying the overall network structure or self-supervised task design, it still achieves superior detection performance compared to the comparative methods. Experimental results show that this invention achieves superior detection performance compared to existing anomaly detection methods on a self-built aero-engine endoscopic inspection dataset and multiple publicly available industrial inspection datasets. This verifies the significant advantages of this invention in terms of detection accuracy, robustness, and generalization ability, and demonstrates its applicability to various industrial inspection scenarios, possessing good engineering application value and industrial applicability.
Claims
1. A method for detecting anomalies in aero-engines based on self-driven hierarchical learning, characterized in that... The specific process of the method is as follows: A1. Acquire normal images of the aircraft engine and preprocess them to obtain preprocessed normal images of the aircraft engine. A2. Use the pre-processed normal images of the aero-engine to perform unsupervised pre-training on the aero-engine anomaly detection model to obtain a pre-trained aero-engine anomaly detection model. A3. Construct a downstream fine-tuning dataset, use the backbone network in the pre-trained aero-engine anomaly detection model to construct the aero-engine final anomaly detection model, use the downstream fine-tuning dataset to train the aero-engine final anomaly detection model, and obtain the trained aero-engine final anomaly detection model. A4. Obtain the image of the aircraft engine to be tested, input the image of the aircraft engine to be tested into the trained final anomaly detection model of the aircraft engine, and obtain whether the image of the aircraft engine to be tested is a normal image or an abnormal image.
2. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 1, characterized in that: The preprocessing of the normal image of the aircraft engine in A1 specifically includes: A1.1 Normal image of an aircraft engine Rotate the image at different angles to obtain a rotated image. ; A1.2, Under normal aircraft engine image A preset area of size is randomly selected as the magnification area, and the magnification area is magnified at different magnifications to obtain a magnified image. ; in, It is the rotation angle. It's the magnification factor. Indicates rotation angle A normal image of the engine. Indicates magnification factor A normal image of the engine.
3. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 2, characterized in that: The aero-engine anomaly detection model includes: a backbone network, a feature dimensionality reduction module, a feature decoupling module, a magnification determination module, a viewing angle change adaptation module, and an instance push-pull module; The backbone network is used to extract features from magnified or rotated images to obtain image features. ,Will Input feature dimensionality reduction module and instance push-pull module; in, Pick or , It is a feature of magnified images. It is a feature of rotated images; The feature dimensionality reduction module is used to... Reduced to the preset dimension, obtaining the dimensionality-reduced result. After dimensional reduction Input is fed into the feature decoupling module; The feature decoupling module is used to decouple the dimensionality-reduced... The decoupled magnification feature or decoupled rotation feature is obtained by average decoupling along the channel dimension. The decoupled magnification feature is sent to the magnification determination module, and the decoupled rotation feature is sent to the viewpoint change adaptation module. The magnification determination module is used to obtain a normal image of the aero-engine based on the decoupled magnification features. The probability of each magnification; The perspective change adaptation module is used to obtain a normal image of the aircraft engine based on the decoupled rotation features. The probability of each rotation angle; The push-pull method described above is used to determine the probability that different image features belong to the same normal image of an aircraft engine.
4. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 3, characterized in that: The feature dimensionality reduction module includes, in sequence: a max pooling layer, a feature flattening layer, a fully connected layer, a batch normalization layer, and a ReLU activation function layer.
5. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 4, characterized in that: The magnification determination module includes a fully connected layer and a softmax layer in sequence.
6. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 5, characterized in that: The viewpoint change adaptation module includes a fully connected layer and a softmax layer in sequence.
7. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 6, characterized in that: The instance push-pull module includes: a generalization unit and a category determination unit; The generalization unit is used for image features. Perform L2 generalization to obtain the generalized image features; The category determination unit is used to obtain the probability that two generalized image features come from the same normal image of an aero-engine.
8. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 7, characterized in that: The category determination unit is used to obtain the probability that two generalized image features come from the same normal image of an aero-engine, specifically: in, It is an image go through Generalized image features after view transformation It is an image go through Generalized image features after view transformation yes The corresponding view belongs to the image category. The probability, , , It is an image label. It is an image go through Generalized image features after view transformation It is the total number of images. , All or .
9. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 8, characterized in that: The unsupervised training of the aero-engine anomaly detection model uses the following loss function: in, It is the overall loss function. , and These are weighting coefficients. It is a magnification category label. It represents the total number of magnification categories. It is the category of the true magnification of the magnified area of the image. It is the predicted probability of the true magnification category of the magnified area of the image. It is a balancing term. The loss is determined by the magnification. It is a rotation angle category label. This represents the total number of rotation angle categories. This is the actual rotation category in the image. It is a regulating item. It is the predicted probability of the image rotation angle category. It is perspective distortion loss. It represents the structural similarity of features between two images in a positive sample pair. It is the structural similarity of two image features in a negative sample pair. It is a temperature parameter; The positive sample pair is either a normal image of the same aero-engine with two rotated images at different rotation angles or a normal image of the same aero-engine with two magnified images at different magnifications. The negative sample pair is either a rotated image corresponding to two normal images of two different aero engines, or a magnified image corresponding to two normal images of two different aero engines.
10. The aero-engine anomaly detection method based on self-driven hierarchical learning according to claim 9, characterized in that: The process described in A3 involves constructing a downstream fine-tuning dataset, building a final aero-engine anomaly detection model using the backbone network of a pre-trained aero-engine anomaly detection model, and training the final aero-engine anomaly detection model using the downstream fine-tuning dataset. The resulting trained final aero-engine anomaly detection model is as follows: A3.
1. Obtain aircraft engine image samples and assign labels to each aircraft engine sample. Use the aircraft engine image samples and labels to form a downstream fine-tuning dataset. The labels include: normal and abnormal; A3.2 Freeze the backbone network parameters in the pre-trained aero-engine anomaly detection model, and splice the pre-trained backbone network with the binary classification head to obtain the final aero-engine anomaly detection model. A3.
3. Use the downstream fine-tuning dataset to train the final anomaly detection model for aero-engines and fine-tune the binary classifier head to obtain the trained final anomaly detection model for aero-engines.