Aero-engine turbine blade micro-defect detection and grading method

Abstract

The application discloses an aero-engine turbine blade micro-defect detection and grading method, which is realized based on a multiple mechanism optimization network and comprises the following steps: collecting aero-engine turbine blade videos or images and marking defect features offline; constructing an aero-engine turbine blade micro-defect detection network, optimizing the convolution modules of a feature extraction network and a feature fusion network in the network, adding an attention mechanism, and optimizing a model boundary loss function; setting aero-engine turbine blade micro-defect detection network training parameters and training an aero-engine turbine blade micro-defect detection network; inputting the aero-engine turbine blade videos or images into the trained aero-engine turbine blade micro-defect detection network, and outputting defect confidence and a boundary box; and grading the aero-engine turbine blade micro-defects according to the recognized defect parameters.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection technology, and in particular to a method for detecting and rating tiny defects in turbine blades of an aero-engine. Background Art

[0002] Aircraft engine turbine blades are core components of aircraft propulsion systems, and their performance and reliability are crucial to flight safety. Operating in high-temperature, high-pressure, and high-speed environments, turbine blades are prone to minor defects. If not detected and addressed promptly, these defects can lead to blade breakage and, in turn, serious accidents. A method for detecting and rating minor defects in aircraft engine turbine blades, based on a multi-mechanism optimization network, is crucial for improving the accuracy and efficiency of blade inspections and ensuring flight safety, providing strong technical support for aircraft engine maintenance and overhaul.

[0003] Among the existing related technologies for aircraft engine turbine blade defect detection before the present invention, there are several comparative patents and documents: CN202311193610.3, CN202311069680.8, and CN202211433839.5.

[0004] The above-mentioned specific patent reference documents are:

[0005] 1) "A Miniature Wirelessly Controllable Vision Device for Aircraft Engine Turbine Blade Inspection," Patent No. CN202311193610.3. This invention discloses a miniature wirelessly controllable vision device for aircraft engine turbine blade inspection, relating to the field of industrial robot vision technology. From top to bottom, it comprises: a camera and lighting module, a rotating frame, an electrothermal drive module, an electronic component module, a battery, and a support frame. The camera and lighting module performs image acquisition, lighting, and sensing functions. The electrothermal drive module rotates the camera and lighting module based on the different thermal expansion coefficients of different materials. The electronic component module performs data processing and transmission functions. The battery serves as a power source, and the support frame allows the vision system to be mounted on live insects or microrobots. The camera and lighting module can rotate vertically and horizontally at specified angles, thereby capturing images in different directions. The system of the present invention has the advantages of small size, light weight, a wide image acquisition range, and wireless controllability. It can be mounted on small organisms or microrobots and placed in aircraft engines to perform turbine blade inspection tasks. The detection method of the present invention is different from the above inventions. It improves the detection method for minor defects of aircraft engine turbine blades and quantitatively grades and classifies four types of defects.

[0006] 2) "A hollow turbine blade detection method, system and equipment", patent number CN202311069680.8. This invention discloses a hollow turbine blade detection method, system and equipment, which relates to the field of X-ray detection technology. Obtain the standard CAD model data of the ceramic core, the wax mold standard CAD model data, the mold shell standard CAD model data and the blade casting standard CAD model data of the hollow turbine blade; obtain N two-dimensional images of the ceramic core, M two-dimensional images of the wax mold, Q two-dimensional images of the mold shell and W two-dimensional images of the blade casting; obtain the three-dimensional model data of the ceramic core, the three-dimensional model data of the wax mold, the three-dimensional model data of the mold shell and the three-dimensional model data of the blade casting; obtain the comparison result of the hollow turbine blade to be detected; perform difference calculation between the comparison result and the preset threshold to obtain the difference; obtain the updated difference; determine whether the updated difference is greater than the first threshold; if so, output that the full-process detection of the hollow turbine blade is unqualified; otherwise, output that the full-process detection of the hollow turbine blade is qualified. The present invention improves the accuracy of the full-process detection of the hollow turbine blade. The detection object of the present invention is different from that of the above inventions. The defect detection algorithm and improvement of the present invention are studied on specific aircraft engine turbine blades and can detect tiny defects.

[0007] 3) "A Rapid Inspection Device for Aircraft Engine Turbine Blades," Patent No. CN202211433839.5. This invention discloses a rapid inspection device for aircraft engine turbine blades, relating to the field of turbine blade inspection technology. The device comprises a transmission disc for transporting turbine blades, with a plurality of support rings arranged in an array on the upper end of the transmission disc for supporting the turbine blades. The upper end surfaces of the support rings are provided with rolling elements for facilitating fine-tuning of the turbine blades. A detection component for inspecting the turbine blades is provided along the transmission path of the transmission disc. The detection component comprises a base having a transmission slot formed therein that cooperates with the transmission disc. The inspection method herein does not require the blades to be disassembled for inspection, effectively ensuring processing efficiency. Furthermore, the inspection method herein does not require immersion, eliminating the problem of a slow immersion process and effectively ensuring inspection effectiveness. Furthermore, the flow-through inspection method herein facilitates loading and unloading, is highly safe, and has strong practicality. The turbine blade inspection of this invention differs from the aforementioned invention in that the aforementioned invention focuses on rapid inspection, while the present invention focuses on accurate detection of minor defects. Summary of the Invention

[0008] In order to solve the above technical problems, the purpose of the present invention is to provide a method for detecting and rating small defects of aircraft engine turbine blades, which is implemented based on a multi-mechanism optimization network.

[0009] The purpose of the present invention is achieved through the following technical solutions:

[0010] A method for detecting and rating minor defects in aircraft engine turbine blades is implemented based on a multi-mechanism optimization network, including:

[0011] A. Collect videos or images of aircraft engine turbine blades and annotate defect features offline;

[0012] B. Build a network for detecting minor defects in aircraft engine turbine blades. Optimize the convolutional modules of the feature extraction network and feature fusion network, add an attention mechanism, and optimize the model's boundary loss function.

[0013] C. Setting training parameters for an aircraft engine turbine blade micro-defect detection network and training the aircraft engine turbine blade micro-defect detection network, inputting an aircraft engine turbine blade video or image into the trained aircraft engine turbine blade micro-defect detection network, and outputting a defect confidence score and a bounding box;

[0014] D. Rating minor defects in aircraft engine turbine blades based on identified defect parameters.

[0015] Compared with the prior art, one or more embodiments of the present invention may have the following advantages:

[0016] This method uses video or image-based detection to detect minor defects in aircraft engine turbine blades and grading their severity, effectively applying it to safe maintenance of aircraft engine turbine blades. This method has significant practical engineering value in improving the accuracy and efficiency of blade inspections and ensuring flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for detecting and rating minor defects in aircraft engine turbine blades based on a multi-mechanism optimization network;

[0018] Figure 2 This is the LFD-YOLO network diagram based on multiple mechanism optimization networks;

[0019] Figure 3 It is the LDConv structure diagram;

[0020] Figure 4 This is a diagram of the DAT structure;

[0021] Figure 5 This is the effect diagram of detecting tiny defects in aircraft engine turbine blades. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to embodiments and accompanying drawings.

[0023] like Figure 1As shown in FIG, the process of the method for detecting and rating minor defects of aircraft engine turbine blades based on a multi-mechanism optimization network is shown, which includes the following steps:

[0024] Step 10: Capture videos or images of aircraft engine turbine blades and annotate defect features offline;

[0025] Step 20: Build an aircraft engine turbine blade micro-defect detection network, optimize the convolution modules of the feature extraction network and feature fusion network in the network, add an attention mechanism, and optimize the model boundary loss function.

[0026] Step 30 sets training parameters for an aircraft engine turbine blade micro-defect detection network and trains the aircraft engine turbine blade micro-defect detection network, inputs an aircraft engine turbine blade video or image into the trained aircraft engine turbine blade micro-defect detection network, and outputs a defect confidence score and a bounding box.

[0027] Step 40 is to rate the minor defects of the aircraft engine turbine blades using the identified defect parameters.

[0028] The above-mentioned step 10 specifically includes: collecting videos or images of tiny defects in aircraft engine turbine blades through an industrial endoscope; and using the open source tool labelme to annotate various defects such as cracks, burns, notches, material loss, and dents.

[0029] The above step 20 specifically includes: the aircraft engine turbine blade micro-defect detection network is an LFD-YOLO network, the LFD-YOLO network uses YOLOv8 as the basic network, and uses LDConv to optimize the CBS module of the initial network; uses the DAT deformable attention mechanism to optimize the feature extraction and feature fusion network; uses Focaler-IoU to optimize the loss function module of the initial network, and completes the automatic detection of micro-defects in aircraft engine turbine blades, such as Figure 2 Shown is the LFD-YOLO network diagram.

[0030] like Figure 3 The structure diagram is shown in Figure 2. LDConv replaces some CBS modules in the YOLOv8 backbone network with LDConv. LDConv uses irregular convolution kernels for convolution operations. First, the initial sampling coordinates P of the irregular convolution kernel need to be generated. n , Table 1 is P n Generate algorithm pseudocode:

[0031] Table 1

[0032]

[0033]

[0034] After defining the irregular convolution initial coordinate P n After that, the corresponding convolution operation at position P0 can be defined as follows:

[0035]

[0036] Where R and w are sampling grid and convolution parameters respectively.

[0037] First, use Algorithm 1 to generate the initial coordinates P n , generates the corresponding sampling coordinates (P n +P0); secondly, use Conv2d to perform a convolution operation on the input image, obtain the corresponding convolution kernel offset through the convolution operation, and then add the offset to the original coordinates to adjust the initial sampling shape. This step realizes the dynamic adjustment of the convolution kernel shape to adapt to the image characteristics; finally, the feature map is resampled according to the adjusted sampling shape. The resampled feature map is reshaped, convolved again, and normalized, and finally the feature map is output through the activation function SiLU.

[0038] The original Yolov8 network uses CIoU as the bounding box regression loss. The CIoU calculation formula is as follows:

[0039]

[0040] Among them, ρ 2 (B,B gt ) represents the Euclidean distance between the center points of the prediction box and the ground-truth box, c represents the B box, B gt The minimum diagonal distance between the bounding boxes, w, h, w gt 、h gt Represent the width and height of the prediction box and ground-truth box respectively. Then the CIoU loss function L CIoU The calculation formula is:

[0041] L CIoU =1-CIOU (5)

[0042] This paper uses Focaler-CIoU loss to replace the CIoU loss in Yolov8, and improves the performance of the detector in different detection tasks by focusing on different regression samples. Focaler Use the linear interval mapping method to reconstruct the IoU loss, IoU Focaler The calculation formula is as follows:

[0043]

[0044] Among them, [d,u]∈[0,1], by adjusting the d and u values, IoU Focaler More suitable for different regression samples. Combined with the above CIoU calculation process, the Focaler-CIoU loss function L Focaler-CIoU The calculation formula is as follows:

[0045] L Focaler-CIoU =L CIoU +IoU-IoU Focaler (7)

[0046] Optimized network bounding box regression loss function L B_box for:

[0047] L B_box =L DFL +L Focaler-CIoU (8)

[0048] like Figure 4 The following is the DAT structure diagram. In the feature extraction network, DAT is added to the last layer of Backbone, which enables the network to adaptively adjust the attention weight according to the input image features, flexibly control the intensity and range of attention, and enhance the network's ability to express small features. In the feature fusion network, DAT is added after the C2f module in the downsampling stage, so that the network can maintain high-quality perception of deep discriminative features and high-level semantic information during the feature fusion stage. The DAT module process includes: input is a feature map x of size H×W×C, from the dimension H G ×W G Select reference points from the ×2 uniform grid to achieve effective downsampling of the original grid.

[0049] H G =H / r,W G =W / r (9)

[0050] r is a pre-set scaling factor. The reference point value is a linear interval coordinate {(0,0),…,(H G -1,W G -1)}, according to the grid shape H G ×W G Normalize the reference points to [-1, +1], (-1, -1) represents the upper left corner, and (+1, +1) represents the lower right corner. To obtain the offset of each reference point, linearly project the feature map onto the query tokens:

[0051] q=xW q (10)

[0052] Among them, W q is the query projection weight. Then q passes through the offset generation subnetwork to generate the offset Δp:

[0053] Δp=θ offset (q) (11)

[0054] Combine the reference point and offset information to obtain the deformed reference point, and use bilinear interpolation on the deformed reference point to obtain the sampling of the feature map x:

[0055]

[0056] right Perform linear projection, W k 、W v Projection weights for key and value:

[0057]

[0058] Finally, integrating the above results and applying the multi-head attention mechanism, we can get the output z of the deformable attention module. (m) :

[0059]

[0060] Where σ(·) represents the softmax function, d = C / M is the size of each head, q (m) 、 represents the query, key, and value vectors output by the mth attention head, is the bilinear interpolation of the relative position deviation.

[0061] The above step 30 specifically includes: randomly dividing the dataset of 2000 images of tiny defects in aircraft engine turbine blades into training set, validation set, and test set according to the ratio of 7:2:1; the network training parameter settings mainly include: the optimizer is SGD, the initial learning rate is 0.01, the batch size is 8, and the number of training rounds is 500.

[0062] like Figure 5 The figure shows the effect of detecting small defects in aircraft engine turbine blades; the above step 40 specifically includes: based on the identified defects, the following rating parameters are used for cracks, burns, notches, material loss, etc.

[0063] 1) Crack defect rating of aircraft engine turbine blades (the parameters used for judgment are the maximum crack length L and maximum width W), as shown in Table 2:

[0064] Table 2

[0065]

[0066] 2) Burn defect rating of aircraft engine turbine blades (the parameter used for judgment is the burn area ratio P), as shown in Table 3:

[0067] Table 3

[0068]

[0069] 3) Rating of notch defects in aircraft engine turbine blades (the parameter used for determination is radial depth d), as shown in Table 4:

[0070] Table 4

[0071]

[0072]

[0073] 4) Aero-engine turbine blade material loss defect rating (the parameter used for judgment is the contour length C of the loss area), as shown in Table 5:

[0074] Table 5

[0075]

[0076] Although the embodiments disclosed herein are as described above, the contents described herein are merely embodiments for facilitating understanding of the present invention and are not intended to limit the present invention. Any person skilled in the art may make any modifications and variations in the form and details of the embodiments without departing from the spirit and scope of the present invention. However, the scope of patent protection of the present invention shall remain subject to the scope defined by the appended claims.

Claims

1. A method for detecting and rating minor defects in aircraft engine turbine blades, characterized in that: The method is based on multiple mechanisms to optimize network implementation, including: A. Collect videos or images of aircraft engine turbine blades and annotate defect features offline; B. Build a network for detecting minor defects in aircraft engine turbine blades. Optimize the convolutional modules of the feature extraction network and feature fusion network, add an attention mechanism, and optimize the model's boundary loss function. C. Setting training parameters for an aircraft engine turbine blade micro-defect detection network and training the aircraft engine turbine blade micro-defect detection network, inputting an aircraft engine turbine blade video or image into the trained aircraft engine turbine blade micro-defect detection network, and outputting a defect confidence score and a bounding box; D. Rating minor defects in aircraft engine turbine blades based on identified defect parameters; The aircraft engine turbine blade micro-defect detection network in the B is the LFD-YOLO network. The LFD-YOLO network adopts Yolov8 as the basic network, and adopts LDConv to optimize the CBS module of the initial network; adopts the DAT deformable attention mechanism to optimize the feature extraction network and the feature fusion network; adopts Focaler-IoU to optimize the loss function module of the initial network, thereby completing the automatic detection of micro-defects in aircraft engine turbine blades.

2. The method for detecting and rating minor defects of aircraft engine turbine blades according to claim 1, characterized in that: The defect characteristics in A include cracks, burns, notches, material loss and indentations.

3. The method for detecting and rating minor defects of aircraft engine turbine blades according to claim 1, characterized in that: Replace some CBS modules in the Yolov8 backbone network with LDConv, which uses irregular convolution kernels for convolution operations, including: Define the initial sampling coordinates P of the irregular convolution kernel n , generates the corresponding sampling coordinates (P n +P0); use Conv2d to perform a convolution operation on the input image, obtain the corresponding convolution kernel offset through the convolution operation, and then add the convolution kernel offset to the original sampling coordinates to adjust the initial sampling shape; resample the feature map according to the adjusted initial sampling shape, and the resampled feature map is reshaped, convolved again, and normalized before outputting the feature map through the activation function SiLU.

4. The method for detecting and rating minor defects in aircraft engine turbine blades according to claim 3, characterized in that: The corresponding convolution operation at coordinate P0 is defined as follows: Among them, R and w are the sampling grid and convolution parameters respectively.

5. The method for detecting and rating minor defects of aircraft engine turbine blades according to claim 1, characterized in that: The CIoU is used as the bounding box regression loss in the Yolov8 network. The CIoU calculation formula is as follows: Among them, ρ 2 (B,B gt ) represents the Euclidean distance between the center points of the prediction box and the ground-truth box, α represents the weight parameter, v represents the aspect ratio correction factor, and c represents the B box, B gt The minimum diagonal distance between the bounding boxes, w, h, w gt 、h gt Represent the width of the prediction box, the height of the prediction box, the width of the ground-truth box, and the height of the ground-truth box respectively; then the CIoU loss function L CIoU The calculation formula is: L CIoU =1-CIoU (5) Use Focaler-CIoU loss to replace CIoU loss, IoU in Yolov8 Focaler Use the linear interval mapping method to reconstruct the IoU loss, IOU Focaler The calculation formula is as follows: Among them, [d,u]∈[0,1], by adjusting the d and u values, IoU Focaler Applicable to different regression samples; combined with the calculation process of CIoU, the loss function L of Focaler-CIoU Focaler-CIoU The calculation formula is as follows: L Focaler-CIoU =L CIoU +IoU-IoU Focaler (7) Combined with DFL bounding box loss L DFL , the optimized network bounding box regression loss function L B_box for: L B_box =L DFL +L Focaler-CIoU (8) 6. The method for detecting and rating minor defects of aircraft engine turbine blades according to claim 1, characterized in that: In the feature extraction network, DAT is added to the last layer of Backbone, so that the network can adaptively adjust the attention weight according to the input image features, control the intensity and range of attention, and enhance the network's ability to express tiny features; in the feature fusion network, DAT is added to the C2f module in the downsampling stage, so that the network can maintain high-quality perception of deep discriminative features and high-level semantic information in the feature fusion stage.

7. The method for detecting and rating minor defects in aircraft engine turbine blades according to claim 6, characterized in that: The input of the DAT deformable attention mechanism module is a feature map x of size H×W×C, which is obtained from the dimension H G ×W G Select reference points from a ×2 uniform grid to achieve effective downsampling of the original grid: H G =H / r,W G =W / r (9) Where H represents the height of the input feature map, W represents the width of the input feature map, and C represents the number of channels of the input feature map; r is a pre-set scaling factor; the reference point value is a linear interval coordinate {(0,0),…,(H G -1,W G -1)}, according to the grid shape H G ×W G Normalize the reference points to [-1, +1], (-1, -1) represents the upper left corner of the grid shape, and (+1, +1) represents the lower right corner of the grid shape; to obtain the offset of each reference point, linearly project the feature map to the query tokens: q=xW q (10) Among them, W q The query projection weight is then passed through the offset generation subnetwork to generate the offset Δp: Δp=θ offset (q) (11) Combine the reference point and offset information to obtain the deformed reference point, and use bilinear interpolation on the deformed reference point to obtain the sampling of the feature map x: Perform linear projection on x, W k 、W v Projection weights for key and value: Finally, the multi-head attention mechanism is applied to obtain the deformable attention module output z (m) : Where σ(·) represents the softmax function, d = C / M is the size of each head, C is the number of channels of the input feature map, M is the number of heads of the multi-head attention, and q (m) 、 represents the query, key, and value vectors output by the mth attention head, is the bilinear interpolation of the relative position deviation.

8. The method for detecting and rating minor defects in aircraft engine turbine blades according to claim 1, characterized in that: In C, a dataset of images of minute defects in aircraft engine turbine blades is obtained through videos or images of aircraft engine turbine blades, and the dataset is divided into a training set, a validation set, and a test set; the network training parameters include an optimizer SGD, an initial learning rate of 0.01, a batch size of 8, and a number of training rounds of 500.

9. The method for detecting and rating minor defects in aircraft engine turbine blades according to claim 2, characterized in that: In the above D, the rating of identified cracks, burns, notches, and material loss defects includes: The crack defect level of aero-engine turbine blades is determined by the crack length L and width W: Cracked leaf surface: 0mm<L≤1mm, defect level is Level I; 1mm<L≤3mm and 0mm<W≤2mm, defect level is Level II; 1mm<L≤3mm and 2mm<W≤3mm, defect level is Level III; L>3mm, defect level is Level IV; Cracked blade root: cracks exist, defect level is IV; The burn defect level of aero-engine turbine blades is determined by the parameter burn area ratio P: Burned rotor: 0<P≤10%, defect level is Grade I; 10%<P≤15%, defect level is Grade II; 15%<P≤20%, defect level is Grade III; P>20%, defect level is Grade IV; Burned static: 0<P≤15%, defect level is Grade I; 15%<P≤20%, defect level is Grade II; 20%<P≤25%, defect level is Grade III; P>25%, defect level is Grade IV; By parameter radial depth d radial Determine the level of notch defects in aircraft engine turbine blades: Notch stator: 0mm<d radial ≤2mm, the defect level is level I; 2mm<d radial ≤3mm, the defect level is level II; 3mm<d radial ≤4mm, defect level is level III; d radial >3mm, the defect level is IV; Notched rotor: If there is a notch on the blade, the defect level is Level III; if there is a notch on the blade root, the defect level is Level IV; The parameter is the contour length C of the lost area lost Determine the material loss defect level of aircraft engine turbine blades: Material loss stator blade: 0mm<C lost ≤2mm, the defect level is Grade I; 2mm<C lost ≤4mm, the defect level is II; 4mm<C lost ≤5mm, defect level is III; C lost >3mm, the defect level is IV; Material loss low pressure blade: 0mm<C lost ≤2mm, the defect level is Grade I; 2mm<C lost ≤3mm, the defect level is II; 3mm<C lost ≤4mm, defect level is III; C lost >4mm, the defect level is Level III; Material loss high-pressure blade: There is a material loss defect, and the defect level is Level IV.

Images