Engine turbine blade defect detection system based on deep learning

Abstract

The invention relates to the technical field of aerospace equipment detection, and particularly discloses an engine turbine blade defect detection method and system based on deep learning. Blade static images, dynamic videos and real-time image data are collected through an industrial camera, a video recorder and an external camera; after the image is subjected to denoising, normalization and enhancement preprocessing, an improved YOLOv8 deep learning algorithm is utilized to construct a defect detection model, and four types of defects including scratches, oil stains, rust stains and damage are accurately recognized; establishing a defect evaluation model based on multi-dimensional indexes such as an accuracy rate, a recall rate, F1-Score, mAP and the like, dividing the defect severity into four levels of slight, moderate, serious and fatal, and generating a detection report containing visual annotation and processing suggestions; and results are fed back to related departments in real time through the display screen, the data interface and the mobile terminal. The method is high in detection precision, high in efficiency and high in scene adaptability, a scientific basis is provided for blade manufacturing quality control and maintenance decision making, and operation safety of an aero-engine is effectively guaranteed.

Description

Technical Field

[0001] This invention relates to the field of aerospace equipment inspection technology, specifically to a deep learning-based engine turbine blade defect detection system. Background Technology

[0002] As a core component of aero-engines, turbine blades operate under extreme conditions of high temperature, high pressure, and high speed for extended periods, making their surfaces prone to defects such as scratches, oil stains, rust, and breakage. These defects not only reduce blade aerodynamic performance and exacerbate thermal stress concentration but can also lead to fatigue fracture, seriously threatening the operational safety and service life of aero-engines. Therefore, accurate and rapid defect detection is crucial in blade manufacturing, inspection, and maintenance.

[0003] Current traditional methods for turbine blade defect detection have many limitations. Manual inspection relies on visual observation and experience-based judgment by technicians, resulting in low efficiency, difficulty in meeting the demands of large-scale inspections, and significant susceptibility to subjective factors, easily missing even minor defects. Furthermore, inconsistent judgment standards among different inspectors lead to a lack of uniformity in results. Traditional image processing-based detection algorithms, such as threshold segmentation and edge detection, are prone to misjudgment or missed detection when faced with complex blade surface reflections, surface texture interference, and defects that are similar to background features, failing to meet high-precision inspection requirements. In addition, existing inspection technologies are mostly limited to static image detection, lacking support for dynamic scenes and real-time inspection, making it difficult to adapt to the efficient collaborative needs of modern production and maintenance.

[0004] Therefore, developing an intelligent, high-precision, and multi-scenario adaptable engine turbine blade defect detection system is crucial to addressing the shortcomings of existing technologies and ensuring the safe operation of aero engines. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a deep learning-based engine turbine blade defect detection system. Through an improved deep learning algorithm and a multi-dimensional evaluation model, the system can accurately identify four types of blade defects, classify their severity, and provide real-time feedback, thus providing a scientific basis for blade quality control and maintenance decisions.

[0006] To solve the above-mentioned technical problems, the technical solution provided by the present invention is: a method for detecting engine turbine blade defects based on deep learning, comprising the following steps:

[0007] S1. Image Data Acquisition: Using industrial cameras, video recorders, and external cameras, static high-definition images, dynamic video frames, and real-time image data of the engine turbine blades are acquired. During acquisition, the image resolution is maintained at a uniform 640 pixels × 640 pixels, with a static image acquisition frequency of no less than 1 frame / second and a dynamic video acquisition frame rate of no less than 30 frames / second. The acquired image data is transmitted to the data processing module via wired or wireless transmission.

[0008] S2. Data Preprocessing and Defect Identification: The acquired image data is preprocessed. Gaussian filtering is used to remove noise interference, and image size normalization is achieved through interpolation. Data augmentation is performed using flipping, rotation, and brightness adjustment to improve the model's generalization ability. An improved YOLOv8 defect detection model is loaded. This model, through optimizing the backbone network feature extraction module, adjusting the multi-scale feature fusion strategy for the neck, and optimizing the detection head parameters, can accurately capture the unique features of four types of defects: scratches, oil stains, rust stains, and damage. The model is trained on the AeroengineDataset dataset, divided into 239 training images, 52 validation images, and 21 test images. Parameters such as batch size 4 and training epochs 300 are set, and training is completed in the PyTorch framework. Through model inference, the defect location coordinates, category, and confidence score are output.

[0009] S3. Defect Assessment and Report Generation: Establish a multi-dimensional defect assessment model, using precision, recall, F1-Score, and mAP (IoU=0.5) as core assessment indicators. Combined with the defect size ratio, structural impact, and development risk, the severity of defects is classified into four levels: minor, moderate, severe, and fatal. A detection report is generated, including basic blade information, collected parameters, defect details, visualized annotated images, and processing suggestions.

[0010] S4. Results Feedback and Application: Test reports are fed back to the quality control department, maintenance management department, and relevant personnel in real time through display screens, data interface transmission, and mobile terminal push notifications. The quality control department can optimize production processes based on the reports, the maintenance management department can develop targeted maintenance plans, and relevant personnel can monitor the blade quality status in real time.

[0011] Furthermore, the improved YOLOv8 algorithm is specifically optimized for different defect features: for scratches, it focuses on extracting the features of thin lines and grayscale differences; for oil stains, it focuses on irregular boundaries and reflective areas; for rust stains, it highlights color abnormalities and texture changes; and for damage, it strengthens the features of edge breakage and structural missing parts, ensuring detection accuracy in complex environments.

[0012] Furthermore, the grading criteria of the multi-dimensional defect assessment model are as follows: minor defects refer to defects whose size accounts for less than 1% of the blade surface area, do not affect the basic function of the blade, and have no obvious risk of development; moderate defects refer to defects whose size accounts for less than 5% of the blade, have a slight impact on aerodynamic performance, and require regular monitoring; severe defects refer to defects whose size accounts for less than 10% of the blade, affect structural strength, and require timely repair; and fatal defects refer to defects whose size accounts for ≥10% of the blade or are located in critical stress areas, directly threatening operational safety, and requiring immediate replacement of the blade.

[0013] This invention also provides a deep learning-based engine turbine blade defect detection system, comprising:

[0014] Image acquisition module: It consists of an industrial camera, a video recorder, and an external camera. The industrial camera is used to acquire static high-definition images, while the video recorder and external camera support dynamic video and real-time image acquisition to meet the needs of different detection scenarios.

[0015] The data processing module includes an image preprocessing unit, a model inference unit, and a data storage unit. The image preprocessing unit performs image denoising, normalization, and enhancement operations; the model inference unit runs an improved YOLOv8 defect detection model to achieve defect identification and analysis; and the data storage unit stores raw data, preprocessed data, model files, and detection results, supporting data traceability and secondary analysis.

[0016] Defect assessment module: Includes a multi-dimensional defect assessment model, which, based on defect identification results and assessment indicators, completes the classification of defect severity and generates a standardized inspection report.

[0017] Results feedback module: Visualizes test results and labeled images on a display screen, supports report export (Excel, PDF format) through data interface, and enables real-time alerts and information synchronization via mobile terminal push.

[0018] Furthermore, the system operates on a Windows operating system, uses the Python 3.10.0 programming language, is based on the PyTorch 2.6.0+126 deep learning framework, is compatible with CUDA 12.6, and supports NVIDIA GeForce RTX 4060 and above GPUs (with ≥8GB of video memory) to ensure the real-time performance and efficiency of model inference.

[0019] 1. High detection accuracy: Through the improved YOLOv8 algorithm, the feature extraction and fusion strategy is optimized, and special optimization is carried out for the unique features of four types of defects. Combined with multi-dimensional evaluation indicators, the accurate identification and classification of defects is achieved, with an mAP value of 0.965, which is significantly better than traditional detection methods.

[0020] 2. High detection efficiency: Supports multiple data input formats such as static images, dynamic videos, and real-time acquisition. The model inference speed is fast, with single image preprocessing taking 0.3ms, inference taking 4.6ms, and post-processing taking 2.8ms, meeting the needs of large-scale detection and real-time monitoring, and greatly improving detection efficiency.

[0021] 3. Strong scene adaptability: It can adapt to complex inspection environments such as complex curved surface reflection of blades, changes in light, and mixed defects, and supports multiple application scenarios such as quality control in the manufacturing process and fault diagnosis in the maintenance process. It is also compatible with static and dynamic inspection requirements.

[0022] 4. Sufficient decision support: Through a multi-dimensional defect assessment model, the severity level is classified, and detailed reports containing defect visualization annotations and handling suggestions are generated, providing comprehensive data support for production optimization and maintenance decisions, forming a closed-loop management of detection-decision-handling. Attached Figure Description

[0023] Figure 1 is a flowchart of a deep learning-based engine turbine blade defect detection method according to the present invention.

[0024] Figure 2 is a system block diagram of an engine turbine blade defect detection system based on deep learning according to the present invention.

[0025] Figure 3 is a schematic diagram of the improved YOLOv8 network structure.

[0026] Figure 4 is a visual representation of the defect detection results. Detailed Implementation

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] The present invention will be described in detail with reference to Figures 1-4.

[0029] A deep learning-based method for detecting defects in engine turbine blades includes the following steps:

[0030] 1. Image data acquisition

[0031] An image acquisition module consists of an industrial camera (1920×1080 resolution), a video recorder (30fps), and an external high-definition camera (30fps). For a specific type of aero-engine turbine blade, static high-definition images, dynamic rotating videos, and real-time detection images were acquired. Data was transmitted to the data processing module via USB 3.0 wired transmission, and all images were uniformly resized to 640 pixels × 640 pixels. During the acquisition process, comprehensive imaging of the blade from different angles and areas was performed to ensure no blind spots in detection.

[0032] 2. Data Preprocessing and Defect Identification

[0033] Image preprocessing: Gaussian filtering (standard deviation σ=0.8) was used to remove image noise. The image size was normalized to 640×640 by bilinear interpolation. Data augmentation was performed by random horizontal flipping, rotation (0-90°), and brightness adjustment (±15%).

[0034] Model Training and Inference: The AeroengineDataset dataset was used, divided into a training set of 239 images, a validation set of 52 images, and a test set of 21 images. The batch size was set to 4, with 300 training epochs. The initial learning rate was 0.01, momentum was 0.937, and weight decay was 0.0005. Model training was performed using the PyTorch 2.6.0 framework. The improved YOLOv8 algorithm enhanced defect feature capture capabilities by increasing the number of convolutional layers in the backbone network, optimizing the neck SPPF module, and adjusting the head anchor box size. After training, the model weight file size was 22.5MB, with a computational cost of 28.4 GFLOPs.

[0035] Defect identification results: Detection was performed on 21 images in the test set. The overall precision was 0.937, recall was 0.962, F1-Score was 0.965, and mAP was 0.784. Specifically, scratch defects had a precision of 0.862 and a recall of 0.995, oil stain defects had a precision of 0.976 and a recall of 0.995, rust defects had a precision of 0.977 and a recall of 1.0, and broken defects had a precision of 0.995 and a recall of 0.892. All types of defects were accurately identified.

[0036] 3. Defect Assessment and Report Generation

[0037] Multi-dimensional assessment: Based on the detection results, the precision, recall, F1-Score and mAP of various defects are calculated. Combined with defect size measurement (actual size is converted by the image pixel ratio), it is determined that the defect accounts for 3% of the blade surface area, which is a moderate defect.

[0038] Inspection report generation: The report includes basic blade information (model: XX-789, production batch: 202506, inspection date: XX / XX / 2025), acquisition parameters (image resolution: 640×640, acquisition device: industrial camera), defect detection results (defect type: scratches + rust stains, location coordinates: (120,350)-(180,370), (250,180)-(290,220), confidence level: 0.92, 0.88, severity: moderate defect), defect visualization and annotation image, and handling suggestions (regularly monitor defect development and perform surface treatment during the next maintenance).

[0039] 4. Results Feedback and Application

[0040] The test report is displayed in real time on the system screen and simultaneously pushed to the quality control department's terminal and the maintenance management department's mobile app, supporting PDF export for archiving. Based on the report, the quality control department optimizes the blade surface treatment process, and the maintenance management department includes the blade in its periodic monitoring list, scheduling targeted treatment during the next overhaul to ensure engine operational safety.

[0041] To implement the above method, the present invention also provides a deep learning-based engine turbine blade defect detection system, comprising:

[0042] Image acquisition module: Composed of an industrial camera (model: Basler acA2500-14uc), a video recorder (model: DS-7200HQHI-K1 series), and an external high-definition camera (model: Logitech C922 Pro), supporting static, dynamic, and real-time image acquisition.

[0043] The data processing module includes an image preprocessing unit, a model inference unit, and a data storage unit. The preprocessing unit uses OpenCV to perform denoising, normalization, and enhancement operations; the model inference unit loads an improved YOLOv8 model to achieve defect identification; and the data storage unit uses a MySQL database to store raw data, processed data, model files, and inspection reports, supporting data querying and traceability.

[0044] Defect assessment module: Includes a multi-dimensional assessment model, integrating precision, recall, F1-Score, and mAP calculation functions, combined with defect size conversion and risk assessment algorithms, to complete defect severity classification and generate standardized reports.

[0045] Results feedback module: Equipped with a 15-inch touchscreen display to show test results and visualized annotated images; provides USB and network interfaces to support report export and data transfer; integrates mobile push service to push test results and warning information in real time via APP.

[0046] In practical application at an aero-engine manufacturing enterprise, this system has cumulatively inspected over 500 turbine blades, achieving an accuracy rate of over 96%. The inspection efficiency is 8 times higher than that of manual inspection, effectively reducing missed and false inspections. It provides reliable technical support for blade quality control and maintenance decisions, and significantly improves the operational safety and reliability of aero-engines.

[0047] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for detecting defects in engine turbine blades based on deep learning, characterized in that, Includes the following steps: S1. Acquire image data of engine turbine blades through an image acquisition device, wherein the image data includes static images, dynamic video frames, and real-time acquired images; The acquired image data is transmitted to the data processing module; S2. Preprocess the image data, including image denoising, size normalization, and data augmentation. A defect detection model is constructed using an improved YOLOv8 deep learning algorithm. Feature extraction and analysis are performed on preprocessed image data to identify four types of defects on the blade surface: scratch, dirty, rust, and damage. The defect location, category, and confidence level are output. S3. A multi-dimensional defect assessment model is established, which combines the defect type, size, location, and confidence level parameters to classify the severity of blade defects. A defect detection report is generated based on the detection and evaluation results; S4. Feedback the test report to the quality control department, maintenance management department and relevant terminal equipment in real time.

2. The method for detecting engine turbine blade defects based on deep learning according to claim 1, characterized in that: In step S1, the image acquisition device includes an industrial camera, a video recorder, and an external camera. The industrial camera is used to acquire static high-definition images of the blades, while the video recorder and the external camera are used to acquire dynamic images and video data of the blades in real time. The image data is transmitted to the data processing module through a data transmission interface, and the transmission methods include wired transmission and wireless transmission.

3. The method for detecting engine turbine blade defects based on deep learning according to claim 2, characterized in that: The image data acquisition parameters include an image resolution of 640 pixels × 640 pixels, and an acquisition frequency adjusted according to the detection scenario, with a static image acquisition frequency of no less than 1 frame / second and a dynamic video acquisition frame rate of no less than 30 frames / second.

4. The method for detecting engine turbine blade defects based on deep learning according to claim 1, characterized in that: In step S2, the image preprocessing specifically includes: removing image noise using a Gaussian filtering algorithm, normalizing the image size to 640 pixels × 640 pixels using interpolation, and performing data enhancement using image flipping, rotation, and brightness adjustment methods; the improved YOLOv8 algorithm enhances the defect feature capture capability by optimizing the feature extraction module of the backbone network, adjusting the multi-scale feature fusion strategy of the neck, and optimizing the parameter configuration of the detection head.

5. The method for detecting engine turbine blade defects based on deep learning according to claim 4, characterized in that: The training process of the defect detection model includes: using the AeroengineDataset dataset, dividing it into 239 images as the training set, 52 images as the validation set, and 21 images as the test set; setting the batch size to 4, the total number of training rounds to 300, the initial learning rate to 0.01, the momentum parameter factor to 0.937, the weight decay coefficient to 0.0005, and training the model based on the PyTorch framework.

6. The method for detecting engine turbine blade defects based on deep learning according to claim 1, characterized in that: The evaluation metrics of the multi-dimensional defect assessment model in step S3 include precision, recall, F1-Score, and mean precision (mAP), where mAP is the calculation result when the IoU threshold is 0.5; the defect severity classification includes four levels: minor defect, moderate defect, severe defect, and fatal defect, and the classification is based on the proportion of defect size to blade surface area, the degree of impact of defect on blade structural strength, and defect development risk.

7. The method for detecting engine turbine blade defects based on deep learning according to claim 6, characterized in that: The defect detection report includes basic information about the blade, image acquisition parameters, defect detection results (including defect location coordinates, category, confidence level, and severity level), a visually labeled image of the defect, and processing suggestions.

8. A deep learning-based engine turbine blade defect detection system, used to implement the detection method according to any one of claims 1-7, characterized in that, include: Image acquisition module: includes an industrial camera, video recorder, and external camera, used to acquire static images, dynamic video frames, and real-time image data of engine turbine blades; Data processing module: includes an image preprocessing unit, a model inference unit, and a data storage unit. The image preprocessing unit performs noise reduction, normalization, and enhancement processing on the acquired image data; the model inference unit loads an improved YOLOv8 defect detection model to identify and analyze defects in the images; the data storage unit stores raw image data, preprocessed data, model files, and detection results; Defect assessment module: includes a multi-dimensional defect assessment model. Based on the defect identification results from the data processing module, and combined with assessment indicators, the severity of defects is graded, and an inspection report is generated; Result feedback module: the inspection report is fed back to the quality control department, maintenance management department, and relevant personnel in real time through displays, data interfaces, mobile terminal push notifications, etc.

9. The engine turbine blade defect detection system based on deep learning according to claim 8, characterized in that: The system operating environment includes a Windows operating system, Python 3.10.0 programming language, CUDA version 12.6, and an NVIDIA GeForce RTX 4060 GPU (8GB of video memory), supporting blade defect detection from static images, video files, and real-time camera data.

10. The engine turbine blade defect detection system based on deep learning according to claim 8, characterized in that: The result feedback module supports visual annotation of defect images, including defect bounding boxes, defect categories, and confidence scores. It also provides a function to export detection data, supporting the export of detection reports in Excel and PDF formats.

Images