Uniformity detection method for aero-engine fan blade preform

Abstract

The invention relates to the technical field of three-dimensional woven composite material detection, solves the technical problem that a traditional method cannot carry out yarn uniformity online detection in the weaving process of a three-dimensional woven preform, and particularly relates to a uniformity detection method for an aero-engine fan blade preform, which comprises the following steps: correcting an original image; feature extraction and fusion are carried out to obtain fusion features; an interleaving point detection task is decoupled into a central point prediction sub-task and a size estimation sub-task; determining the position distribution of effective interlacing points in the effective detection area; and accurate evaluation of the yarn uniformity of the preform is realized. According to the method, automatic detection of the yarn uniformity of the aero-engine fan blade preform can be accurately and efficiently completed, the engineering detection problems of large scale span, high density, variable angles and the like of yarn interweaving points are effectively solved, and the quality evaluation capability of the blade preform in the aviation industrial production process is remarkably improved; and the actual requirements of high-standard production in the aviation industry are met.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional braided composite material detection, and particularly relates to a method for detecting the uniformity of a preform of an aeroengine fan blade. Background Art

[0002] As a power core component of an aircraft, the performance of an aeroengine directly determines the overall performance level and service life of the aircraft. Among the many core components of an aeroengine, the fan blade plays a key role, not only directly determining the thrust output and operating efficiency of the engine, but also its reliability and durability being an important basis for ensuring flight safety. With the continuous improvement of the requirements for the overall performance and safety of engines in the modern aviation industry, traditional metal materials have gradually become difficult to meet the requirements in terms of high thrust-to-weight ratio, long life, and high reliability. Therefore, three-dimensional braided composite fan blades with excellent specific strength, specific stiffness, fatigue resistance, damage tolerance, and thermal stability have gradually become a research and application hotspot for aeroengine components.

[0003] However, the manufacturing process of three-dimensional braided composite fan blades is relatively complex. Especially in the weaving stage of the blade preform, the uniformity of the yarns is directly related to the overall mechanical properties, structural integrity of the composite material after the blade is formed, and its reliability under harsh working conditions. Current production practices show that factors such as design defects in the weaving process plan, unreasonable setting of weaving process parameters, and human operation errors can significantly affect the spatial distribution uniformity of the yarns, and then cause error accumulation in the blade forming stage, resulting in a reduction in the mechanical properties of the final product, weakening of the damage tolerance and impact resistance, and even directly shortening the service life of the blade. Thus, it can be seen that timely and accurately detecting and feedbacking the uniformity status of the yarns during the preform weaving process is of crucial significance for ensuring the overall quality of the composite fan blade.

[0004] At present, the online detection methods for the yarn uniformity of three-dimensional braided composite preforms in the aviation industry still face a series of severe challenges, which are mainly reflected in the following aspects: First, due to the tiny size, extremely high density, and complex arrangement of the yarn intersection points, traditional detection technologies are difficult to accurately identify and evaluate the yarn uniformity on an efficient and real-time industrial production line; second, the traditional off-line manual sampling detection method has low efficiency and large randomness, and cannot provide timely and effective production process feedback, thus it is difficult to achieve comprehensive coverage of the uniformity detection throughout the weaving cycle; in addition, the existing detection methods based on traditional image processing usually target two-dimensional simple fabric samples, without fully considering the actual situation of complex three-dimensional braided structures, high fabric density, and coexistence of multiple fiber bundle interweaving modes in actual industrial production, resulting in a significant decrease in the reliability and accuracy of the existing technologies. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a method for detecting the uniformity of an aero-engine fan blade preform, which solves the technical problem that the traditional method cannot perform on-line detection of yarn uniformity during the weaving process of a three-dimensional braided preform.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: A method for detecting the uniformity of an aero-engine fan blade preform, the method comprising the following steps: S1. Obtain the original image collected in real time from the production line, and correct the intersection points of the carbon fiber horizontal and vertical yarns in the original image to a standard angle presented from a unified perspective through an image angle normalization module; S2. Based on the global path and the local path of the dual-path feature extraction network, perform feature extraction on the corrected original image respectively to obtain dual-path features; S3. Fusion the dual-path features through a dual-path feature fusion network adopting a cross-attention mechanism and a multi-scale self-attention mechanism to obtain fusion features; S4. Input the fusion features into the context decoupled detection head module, decouple the intersection point detection task into two subtasks of center point prediction and size estimation, and obtain the center position of the intersection point located in the detection frame and the width and height of the detection frame; S5. Define the image edge region threshold, delimit the effective detection region based on the detection frame and eliminate the incomplete intersection points existing at the image edge, and obtain the position distribution of the effective intersection points within the effective detection region; S6. Calculate the evaluation index for evaluating the yarn distribution uniformity in the fan blade preform according to the position distribution of the effective intersection points to achieve accurate evaluation of the preform yarn uniformity.

[0007] Further, in step S1, the specific process includes the following steps: S11. Taking the ResNet50 network model as the basic feature extraction network, extract the initial feature map from the original image, and then use the feature pyramid network to generate multi-scale feature maps layer by layer to capture the visual features of the intersection points at different scales; S12. Set rotation-sensitive anchor boxes according to the different sizes and inclination angles of the yarn intersection points , and the expression is: ; wherein, is the center coordinate of the rotation-sensitive anchor box; is the width and height of the rotation-sensitive anchor box; is the rotation angle of the rotation-sensitive anchor box; S13. Based on the rotation-sensitive anchor box Generate rotated candidate regions using the region proposal network head and obtain the angle prediction values of the rotated multi-scale feature maps. And through the ROIAlign operation, unify the multi-scale feature maps of different sizes into a feature map of a fixed size and containing intersection points. S14. Develop an abnormal angle filtering mechanism, and through the angle threshold eliminate the abnormal angles in the angle prediction values . The expression is: ; Among them, is the reserved angle prediction value; is the angle prediction value of the rotated intersection point feature map ; is the predicted mean value of the rotation angle; S15. After averaging the angle prediction values , obtain the corrected angle , and correct the original image through the rotation transformation matrix. The rotation transformation matrix is: ; Among them, represents the pixel coordinates of the intersection point in the original image; is the coordinate of the intersection point after the rotation transformation matrix.

[0008] Furthermore, a rotation angle regression branch is introduced in the region proposal network head, that is: First, assign a series of discrete initial angles to each initial rotation-sensitive anchor box to cover the possible tilting angles of the weaving texture; Then, the region proposal network head regresses and predicts the angle offset of each rotation-sensitive anchor box through the convolutional features, and obtains the correction value of this rotation-sensitive anchor box relative to its initial angle. The finally predicted rotation angle of the rotation-sensitive anchor box is denoted as , and its calculation formula is: ; In the formula, is the initial angle of the rotation-sensitive anchor box ; is the angle increment regressed and output by the region proposal network head.

[0009] Further, in step S2, the dual-path feature includes a global semantic feature for characterizing the overall spatial distribution and context semantic feature of the intersection points, and a local spatial detail feature including the detailed texture and local fine spatial information of the intersection points. The specific process includes the following steps: S21. The global path adopts a global feature extraction architecture based on Vision Transformer, and captures the overall spatial distribution and context semantic feature of the preform intersection points through the self-attention mechanism; S22. The local path adopts a classical convolutional neural network structure, and effectively extracts the detailed texture and local fine spatial information of the intersection points through multi-layer convolutional operations; S23. Define the feature outputs of the global path and the local path as and respectively, that is: ; wherein, represents the mapping function for extracting the global semantic feature through Vision Transformer; represents the mapping function for extracting the local spatial detail feature through the classical convolutional neural network.

[0010] Further, in step S3, the specific process includes the following steps: S31. Use the cross-attention mechanism to take the global semantic feature output by the global path as the query, and the local spatial detail features output by the local path as the key and value respectively, to realize the deep interaction and fusion between the two path features to obtain the fusion feature , and the expression is: ; wherein, the query ; the key ; the value is the feature transformation matrix; is the learnable parameter; is the feature dimension; S32. Use the multi-scale self-attention mechanism to perform weighted integration on the fusion feature at multiple scales respectively, and capture the cross-scale dependence relationship between intersection points at different scales to obtain the comprehensive feature , and the expression is: ; wherein, represents the scale set; represents the attention weight coefficient corresponding to the scale ; represents the comprehensive feature after being fused by the multi-scale self-attention mechanism; represents the selected scale set; Indicates the fused features obtained at scale .

[0011] Furthermore, in step S4, the context decoupling detection head module includes a center point attention module, a size attention module, and an offset attention module. The specific process includes the following steps: S41. Center point attention module: Convert the position prediction in the intersection point detection task into the prediction of the center point heat map, and optimize the network training to output the center position of the intersection point through the loss function . The expression is: ; In the formula, is the predicted value of the network predicted heat map at the position (i, j); is the label value of the true heat map at the position (i, j); is the modulation factor of the focal loss; is the number of true intersection points in the image; S42. Size attention module: Independently predict the detection box size including width and height at the center position of each intersection point, and use the loss function to optimize the prediction error. The expression is: ; In the formula, and are the predicted width and height of the detection box of the i-th intersection point; and are the true width and height of the detection box of the i-th intersection point; S43. Offset attention module: Predict the slight offset of the actual center position of the intersection point relative to the center position of the predicted heat map, and use the loss function to reduce the position prediction error and further improve the fine degree of intersection point positioning. The expression of the loss function is: ; In the formula, and are the offsets of the center of the i-th intersection point predicted by the network relative to the grid; and are the corresponding true offsets; is the smooth absolute value loss function, which is the loss function used to measure the prediction offset error.

[0012] Furthermore, in step S5, the specific process includes the following steps: S51. Define the threshold of the image edge region to accurately delimit the effective detection region and eliminate the possible incomplete edge intersection points; S52. Among all the detected intersection point sets , identify and mark the incomplete intersection points located at the image edge, defined as follows: ; where is the center point of the intersection point, and is the center coordinate; , , , respectively represent the preset left, right, upper, and lower boundary thresholds; is the horizontal coordinate of the i-th valid intersection point in the image; is the vertical coordinate of the i-th valid intersection point in the image; S53. Calculate the average value of the right boundaries of the incomplete intersection points located at the left edge in the detection box, denoted as the average boundary , that is: ; In the formula, is the number of intersection points at the left edge; is the -th horizontal coordinate position of the right boundary of the detection box of the incomplete intersection point at the left edge; S54. According to the average boundary , establish the effective detection boundary standard, and remove all detection boxes whose center points of all intersection points are located on the left side of this boundary, that is: ; In the formula, represents the set of valid intersection points; represents the j-th intersection point after boundary screening; are the horizontal and vertical coordinates of the intersection point center; S55. Process the edge regions including the upper, lower, and right boundaries in the image respectively according to the above standard, so as to obtain the set of effective detection points after removing the incomplete intersection points at the edges.

[0013] Furthermore, in step S54, the specific process of establishing the effective detection boundary standard includes the following steps: S541. Statistically calculate the average value of the horizontal coordinates of the right boundaries of all detection boxes of the incomplete intersection points located at the left edge of the image, which is used to determine the left boundary of the effective detection area, and the expression is: ; In the formula, is the horizontal coordinate value of the right boundary of the i-th detection box of the incomplete intersection point located at the left edge of the image; N is the total number of incomplete intersection points located at the left edge of the image; S542. Define the effective detection area of the left boundary of the image with the average value of the abscissa as the reference. If the abscissa value of the center point of the intersection point is , then the intersection point is considered to be complete and valid; if the abscissa value of the center point of the intersection point is , then the intersection point is considered to be located in the edge area and belongs to an incomplete intersection point, which should be excluded from the subsequent evaluation.

[0014] Further, in step S6, the specific process includes the following steps: S61. Calculate the actual area of the effective detection area , and the calculation formula is: ; In the formula, , are the boundary coordinates of the effective detection area in the image; S62. Count the number of intersection points in the effective area , and use the number of intersection points per unit area as the evaluation index for the yarn uniformity , accurately quantify and evaluate the weaving quality and uniformity of the fan blade preform, that is: ; If the yarn uniformity evaluation index is close to the reference value , and within the allowable error range, then the yarn distribution uniformity of the fan blade preform is considered to be good and meets the quality requirements; If the yarn uniformity evaluation index is significantly lower than the reference value , it indicates that there are too few yarn intersection points per unit area, the yarn density may be too small, and there may be sparse weaving or holes; on the contrary, if the yarn uniformity evaluation index is higher than the reference value , it means that the local yarn is crowded or repeated, indicating uneven weaving.

[0015] By means of the above technical solutions, the present invention provides a method for detecting the uniformity of an aeroengine fan blade preform, which has at least the following beneficial effects: 1. The present invention can accurately and efficiently complete the automated detection of the uniformity of the yarn of the preform of the fan blade of an aircraft engine, effectively solving the engineering detection problems such as the large scale span, high density and variable angle of the yarn interlacing points, significantly improving the quality assessment ability of the blade preform in the production process of the aviation industry, and meeting the actual needs of the high-standard production of the aviation industry.

[0016] 2. The present invention realizes the accurate, efficient and automated detection of the uniformity of the yarn of the preform of the fan blade of an aircraft engine through the design of multi-scale feature interactive fusion and context decoupling attention mechanism, which significantly improves the quality assessment and control level on the industrial production line and meets the strict requirements of the aviation industry for production quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 It is a flow chart of the method for detecting uniformity of a preform in the present invention; Figure 2 This is a network structure diagram for preform uniformity detection in the present invention. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0019] In view of the series of severe challenges faced by the online detection method of yarn uniformity of three-dimensional woven composite preforms, how to achieve accurate and online detection of yarn uniformity during the weaving process of three-dimensional woven preforms of aircraft engine fan blades under the premise of high precision, high real-time performance and low interference has become a key technical problem that needs to be solved urgently.

[0020] An advanced online uniformity detection method is proposed, which has important theoretical significance and engineering application value for further improving the manufacturing quality and production efficiency of aircraft engine fan blade preforms and ensuring the overall performance stability and reliability of aircraft engines. Figure 1 - Figure 2 This embodiment proposes a uniformity detection method for aircraft engine fan blade preforms. Through the design of multi-scale feature interactive fusion and context decoupling attention mechanism, accurate, efficient and automated detection of the uniformity of the yarn of aircraft engine fan blade preforms is achieved, which significantly improves the quality assessment and control level on the industrial production line and meets the strict requirements of the aviation industry for production quality. The method includes the following steps: S1. Obtain the original images collected in real time on the production line, and use the image angle normalization module to correct the intersection points of the carbon fiber horizontal and vertical yarns in the original images to a standard angle presented from a unified perspective. In view of the obvious angle differences that usually exist in the original images of the preforms of aero-engine fan blades collected in the actual production environment, which affect the accurate positioning and detection accuracy of the subsequent carbon fiber yarn intersection points, in this embodiment, an image angle normalization module is first proposed to uniformly correct the collected original images to the standard angle to ensure that the features of the intersection points are presented from a unified perspective and improve the accuracy of subsequent feature detection. The specific process includes the following steps: S11. Use the ResNet50 network model as the basic feature extraction network to extract the initial feature map from the original image, and then use the Feature Pyramid Network (FPN) to generate multi-scale feature maps layer by layer to capture the visual features of the intersection points at different scales; S12. Set rotation-sensitive anchor boxes according to the different sizes and inclination angles of the yarn intersection points , and the expression is: ; wherein, is the center coordinate of the rotation-sensitive anchor box; is the width and height of the rotation-sensitive anchor box; is the rotation angle of the rotation-sensitive anchor box; S13. Based on the rotation-sensitive anchor box use the Region Proposal Network head (RPN) to generate rotation candidate regions (ROIs), obtain the angle prediction value of the rotation multi-scale feature map , and through the ROIAlign operation, unify the multi-scale feature maps of different sizes to a fixed size and a feature map containing the intersection points.

[0021] In this embodiment, a Region Proposal Network (RPN) based on rotation-sensitive anchor boxes is adopted to predict the rotation angle of the intersection point image. Specifically, a rotation angle regression branch is introduced into the Region Proposal Network head, that is: First, a series of discrete initial angles, such as 0°, 5°, 10°, 15°, 20°, 25°, 30°, are assigned to each initial rotation-sensitive anchor box to cover the possible inclination angles of the weaving texture. These angles are all from actual observations. Generally, the inclination angles of the weaving texture on the surface of the preforms of aero-engine fan blades are within 30°.

[0022] Then, the Region Proposal Network head performs regression prediction on the angle offset of each rotation-sensitive anchor box through the convolutional features, and obtains the correction value of this rotation-sensitive anchor box , rotation-sensitive anchor box The finally predicted rotation angle is denoted as , and its calculation formula is: ; In the formula, is the initial angle of the rotation-sensitive anchor box ; is the angle increment output by the region proposal network head regression.

[0023] Finally, the trained region proposal network head can output the angle prediction value of each rotation candidate region for any input image .

[0024] S14. Develop an abnormal angle filtering mechanism to eliminate abnormal angles in the angle prediction value through the angle threshold . The expression is: ; where, is the reserved angle prediction value; is the angle prediction value of the rotation intersection feature map ; is the predicted mean value of the rotation angle.

[0025] In this embodiment, the angle threshold is used to determine the size of abnormal values in the predicted angles. In this embodiment, the angle threshold is set to 5° (about 0.087 radians). That is, when the absolute value of the difference between the angle prediction value of a certain feature map rotation and the predicted mean value exceeds 5°, it is considered to deviate from the normal range, and this prediction is regarded as abnormal and eliminated. This value is based on the observed yarn knitting inclination fluctuation range in actual production. As the default value, 5° can well balance the strictness of abnormal elimination and the sensitivity of retaining normal changes through experiments.

[0026] S15. Take the average of the angle prediction value to obtain the corrected angle , and correct the original image through the rotation transformation matrix. The rotation transformation matrix is: ; where, represents the pixel coordinates of the intersection point in the original image, that is, before the correction rotation, the position coordinates of this point in the original image; is the coordinate of the intersection point after the rotation transformation matrix.

[0027] To horizontally align the yarn interweaving texture in the image, the original image is rotated around the origin to the correction angle After that, the position of this point will move to the new coordinates , through this transformation, the original image can be rotated by the correction angle as a whole , so that the originally inclined yarn texture can be rotated to a unified angle. In this embodiment, the angle normalization correction process of the original image is completed through the above method, and the achieved technical effects include: Weaving texture alignment: By calculating the correction angle of the rotation-sensitive anchor box set And rotating and correcting the image, this step rotates the randomly oriented carbon fiber yarn interweaving texture to a unified perspective, making the overall arrangement direction of the yarn interweaving points in the image consistent, and eliminating the inclination difference caused by different shooting and acquisition angles.

[0028] Abnormal angle elimination: An abnormal angle filtering mechanism is introduced during the rotation correction process to eliminate outliers that deviate too much from the average angle, avoiding the interference of extremely few abnormal rotation detection values on the correction result, and improving the stability and accuracy of the rotation correction.

[0029] Improve the subsequent detection accuracy: After the above unified angle correction, the spatial distribution law of the interweaving points in the image is more unified, providing a standardized input for the feature extraction network. This directly improves the reliability of the subsequent interweaving point detection algorithm, eliminates the problem of inconsistent feature distribution caused by different image rotation angles, and reduces the difficulty of interweaving point detection.

[0030] Generally speaking, through the angle normalization process, the yarn interweaving texture is aligned in the image, significantly reducing the impact of the inclination diversity of the interweaving points on the detection, and improving the accuracy and consistency of feature extraction and interweaving point detection.

[0031] S2. Based on the global path and local path of the dual-path feature extraction network, the corrected original image is respectively subjected to feature extraction to obtain dual-path features. The dual-path features include global semantic features used to characterize the overall spatial distribution and context semantic features of the interweaving points, and local spatial detail features containing the detailed texture and local fine spatial information of the interweaving points. After the original image undergoes angle normalization, the spatial features of the interweaving points of the carbon fiber horizontal and vertical yarns become more regular and unified, but there are still problems such as large scale differences and dense distributions. To accurately capture the rich detailed features of the interweaving points in the corrected image, this embodiment constructs a multi-scale dual-path feature extraction module to respectively extract the global semantic features and local spatial detail features of the original image, providing sufficient feature support for subsequent fine interweaving point detection. The specific process includes the following steps: S21, the global path adopts the global feature extraction architecture based on visual Transformer (ViT), which captures the overall spatial distribution and contextual semantic features of the preform interweaving points through the self-attention mechanism; S22, the local path adopts the classic convolutional neural network (CNN) structure, which effectively extracts the detailed texture and local fine spatial information of the interweaving points through multi-layer convolution operations, solving the detection problem of large scale changes of interweaving points; S23, define the feature outputs of the global path and the local path respectively and ,Right now: ; in, Represents the mapping function of extracting global semantic features through visual Transformer, and its output It mainly includes the overall spatial distribution and contextual information of the intersection points; Represents the mapping function of extracting local spatial detail features through the classic convolutional neural network, and its output It mainly contains the detailed texture and local fine spatial information of each interlaced point. Using this representation method, the functions and outputs of the two different feature extraction branches can be clearly distinguished, thereby providing complementary information for subsequent feature fusion.

[0032] In this embodiment, the dual-path feature extraction network adopts a dual-branch design in architecture: on the one hand, the global feature extraction architecture (ViT) based on visual Transformer is used to realize global semantic feature extraction, which can capture the long-distance dependency and overall arrangement of the preform interweaving points; on the other hand, the classic convolutional neural network (CNN) structure is used to deeply extract local detail information to obtain the texture features of each interweaving point. The combination of the two can take into account both the global and the local, solving the problem that the traditional single path is difficult to take into account.

[0033] By introducing multi-scale feature extraction and feature pyramid structure in the local path, the network can adapt to the feature changes caused by scale differences at the intersection points, thus reflecting multi-scale enhancement. The outputs of the dual paths not only exist independently, but also achieve deep fusion through cross-attention in the subsequent process, so that global semantics and local details can complement each other to improve detection accuracy, thus reflecting cross-scale information fusion.

[0034] Therefore, compared with the existing scheme that only uses a single path feature extraction or a simple parallel two-branch scheme, the present invention combines Transformer and CNN and optimizes parameters and structures according to the characteristics of yarn interlacing points. It has the advantages of multi-scale enhancement and cross-scale information fusion, which significantly improves the accuracy, robustness and adaptability of detection, thus having obvious technical innovation.

[0035] S3. By using a dual-path feature fusion network with a cross-attention mechanism and a multi-scale self-attention mechanism to fuse the dual-path features and obtain the fused features, the robustness of the intersection point features is significantly improved. The specific process includes the following steps: S31. Using the cross-attention mechanism, taking the global semantic features output by the global path as the query (Query), and the local spatial detail features output by the local path as the key (Key) and value (Value) respectively, to achieve deep interaction and fusion between the two-path features and obtain the fused features , the expression is: ; Among them, the query ; the key ; the value is the feature transformation matrix; is the learnable parameter; is the feature dimension; S32. To further improve the representation ability of the fused features for intersection points at different scales, use the multi-scale self-attention mechanism to weighted integrate the fused features at multiple scales respectively, capture the cross-scale dependence relationship between intersection points at different scales, and obtain the comprehensive features , the expression is: ; Among them, represents the scale set; represents the attention weight coefficient corresponding to the scale , which is learned and used to adjust the contribution of features at different scales in the final fused representation; represents the comprehensive features after being fused by the multi-scale self-attention mechanism; represents the selected scale set, for example, the index set of different downsampling ratios or feature layer levels; represents the fused features obtained at the scale , and the superscript is used to indicate that this feature corresponds to the th scale.

[0036] In this embodiment, by weighted summing the fused features at all scales, a comprehensive feature that combines the information of each scale is finally obtained , so as to better capture the details and global semantics of the intersection points at different scales. By introducing a fusion scheme that combines the cross-attention mechanism and the multi-scale self-attention mechanism. Its specific innovation is reflected in: First, the cross-attention mechanism (using global features as queries and local features as keys and values) is utilized to achieve in-depth interaction of dual-path features, enabling global information to actively guide the weighted extraction of local details and generating more discriminative fusion features.

[0037] Secondly, on this basis, self-attention calculations are performed on the fusion features at different scales, and then weighted summation is carried out using learnable scale weights to capture cross-scale dependencies, making the entire detection method still stable when facing changes in the sizes of preform intersection points.

[0038] This fusion method is more accurate and robust than traditional simple feature splicing or element-wise addition methods, and this structure is proposed for the first time in the yarn uniformity detection of aero-engine fan blade preforms (which can be extended to 3D braided / woven preforms), thus demonstrating obvious innovation.

[0039] S4. Input the comprehensive feature into the context decoupled detection head module, decouple the intersection point detection task into two subtasks of center point prediction and size estimation, and obtain the center position of the intersection point located in the detection box and the width and height of the detection box. The context decoupled detection head module designed in this embodiment consists of a center point attention module, a size attention module, and an offset attention module. The specific process includes the following steps: S41. The center point attention module converts the position prediction in the intersection point detection task into the prediction of the center point heat map, and optimizes the network training to output the center position of the intersection point through the loss function to effectively improve the accuracy of center point prediction and reduce the missed detection of small-scale targets. In this embodiment, the center point heat map prediction method is adopted to solve the problem of insufficient detection accuracy of traditional position regression methods for small targets (preform intersection points). The specific process includes: Generate the output heat map of the true center point: For the true center position of each intersection point, generate a two-dimensional Gaussian distribution on the output heat map to make the response value of the output heat map at the true center position the highest, that is, the target value , and the response in the remaining regions decays according to the distance from the center. The expression is: ; In the formula, is the true center coordinate position of the intersection point; is the standard deviation parameter that controls the size of the Gaussian response region, usually determined according to the size of the intersection point.

[0040] Output of the network prediction heat map: The network outputs a prediction heat map with the same size as the input image ​, the predicted value at each position in the heat map represents the probability or confidence that there is an intersection center at that position; Optimizing the loss function for heat map prediction: To optimize the network's prediction of the heat map, the focal loss function is adopted to effectively alleviate the problem of imbalance between positive and negative samples. Its definition is: ; In the formula, is the predicted value of the network's predicted heat map at the position (i,j); is the label value of the true heat map at the position (i,j); is the modulation factor of the focal loss, usually set to 2; is the number of true intersections in the image, used to normalize the loss.

[0041] During the training process, by minimizing the optimization function , the response value of the output heat map at the center of the true intersection is close to 1, while the response value in the background area is close to 0; In the inference stage, by performing local maximum detection (peak detection) on the output heat map , the rough center position of each intersection is determined, and then the center point position is sub-pixel corrected in combination with the offset prediction to obtain the final accurate center point position of the intersection.

[0042] S42. Size attention module, independently predicting the detection box size including width and height for each intersection center position, and using the loss function to optimize the prediction error, effectively ensuring the accuracy and stability of the detection box size prediction. Since the loss function has strong robustness to outliers and can avoid excessive loss caused by individual abnormal samples, thus ensuring the stability of training.

[0043] Among them, in this embodiment, a special size attention module is used to predict the detection box size of each intersection. For each intersection center position, the network outputs two values in parallel on the same feature map, which are the predicted width and the predicted height respectively. During training, with the width and height of the true detection box of the intersection as the supervised annotation, the loss function is used for regression optimization, and the expression is: ; In the formula, is the number of true intersections in the image; and are the predicted width and height of the detection box of the i-th intersection; and are the true width and height of the detection box of the i-th intersection.

[0044] In this embodiment, by minimizing this loss, the network is driven to output accurate size predictions. After training, the size attention module learns to give relatively accurate size estimates in terms of pixels. In the inference stage, for the center points detected in each heatmap, the and output by the network are directly read as the sizes of the detection boxes. In this way, through the cooperation of the center point heatmap and size regression, the center and size of the positioning box are determined for each intersection point. The independent size prediction branch can avoid interference with the center positioning task and improve the independent optimization effect of the two tasks, thereby making the size prediction more accurate and stable.

[0045] S43. Offset attention module, which predicts the slight offset of the actual center position of the intersection point relative to the center position of the predicted heatmap, and uses the loss function to reduce the position prediction error and further improve the refinement of the intersection point positioning. The expression of the loss function is: ; In the formula, is the number of true intersection points in the image; and are the offsets of the center of the i-th intersection point predicted by the network relative to the grid; and are the corresponding true offsets; is the smooth absolute value loss function, which is a loss function used to measure the prediction offset error.

[0046] When this loss is minimized, it means that the prediction of the intersection point center is very close to the true position. In this embodiment, is defined as: 0.5e when |e| < 1 2 , and |e| - 0.5 when |e| ≥ 1. This piecewise function gives a quadratic penalty when the error is small, thereby improving the accuracy; and gives a linear penalty when the error is large, thereby ensuring stability.

[0047] In this embodiment, by optimizing the loss function , it is ensured that the offsets and output by the network are highly consistent with the true values, and the positioning accuracy of the intersection point center reaches the sub-pixel level. In the above method proposed in this embodiment: First, the detection task is decoupled into two subtasks: center positioning and size estimation. Each subtask is processed by a dedicated module, which reduces the task complexity and avoids the instability that may be brought about by directly regressing the entire bounding box. The center point attention module effectively improves the detection rate of dense small targets (intersection points) and significantly reduces the missed detection rate; the size attention module independently predicts and uses the loss function Optimization is carried out to ensure the accuracy of the detection box size and avoid the problem of abnormal dimensions caused by positioning deviation; the offset attention module further refines the center position, keeping the positioning error within the sub-pixel range.

[0048] Secondly, the introduction of the attention mechanism (center point attention, size attention, offset attention) means that the network allocates focus resources to these three branches, thus improving the prediction quality of key elements (position, size).

[0049] Thirdly, through the reasonable design of the loss function, the training process is more stable, correct predictions with high confidence are strengthened, and incorrect predictions with low confidence are suppressed.

[0050] The method of the present invention reduces false detections and missed detections, and is particularly more accurate in locating small-sized intersection points; in terms of evaluation indicators, both the average false detection rate and the missed detection rate are significantly reduced. Therefore, the design of the context decoupling detection head module significantly improves the accuracy, stability and robustness of intersection point detection, providing reliable data support for subsequent uniformity evaluation.

[0051] S5. Define the threshold of the image edge region, delimit the effective detection region based on the detection box, and eliminate the incomplete intersection points existing at the image edge to obtain the position distribution of effective intersection points within the effective detection region. During the detection and positioning process of intersection points, due to the limitation of the imaging field of view, the intersection points located at the image boundary are often incomplete, and these incomplete intersection points may bring errors to the subsequent calculation of yarn uniformity, affecting the accuracy and stability of the detection results. Therefore, this embodiment proposes an edge intersection point removal module to effectively exclude the incomplete intersection points at the image edge, significantly improving the accuracy and reliability of uniformity evaluation. In step S5, the specific process includes the following steps: S51. According to the production practice experience and the boundary characteristics of intersection point detection, define the threshold of the image edge region for accurately delimiting the effective detection region and eliminating possible incomplete edge intersection points. The threshold of the image edge region is to delimit the reliable detection region in the image and exclude those regions where the intersection points are incomplete due to insufficient imaging field of view. The specific method is as follows: According to the average detection box size of intersection points in the preform image, let the average width of the intersection point detection box be and the average height be , taking and as the benchmarks, define the left and right edge thresholds and of the image, as well as the upper and lower edge thresholds and , that is: ; Among them, and are the total width and height of the image, respectively.

[0052] Any intersection point center located at 、 、 or is regarded as an incomplete point and must be removed. This method uses the average size of the detection box as a reference, which can better reflect the possible defect problems in the edge area of the image, so that the subsequent uniformity evaluation is only carried out within the complete area.

[0053] S52. In all the detected intersection point sets , identify and mark the incomplete intersection points located at the image edge, which are defined as follows: ; where is the center point of the intersection point, and is the center coordinate; 、 、 、 represent the preset left, right, upper, and lower boundary thresholds respectively; is the horizontal coordinate of the i-th valid intersection point in the image; is the vertical coordinate of the i-th valid intersection point in the image; S53. Calculate the average value of the right boundary of the incomplete intersection points located at the left edge in the detection box, denoted as the average boundary , that is: ; In the formula, represents the right boundary position of the detection box of the incomplete intersection points at the left edge, and the superscript represents taking the average value, that is, average; is the number of intersection points at the left edge; is the -th horizontal coordinate position of the right boundary of the detection box of the incomplete intersection points at the left edge; S54. Establish an effective detection boundary standard according to the average boundary . All the detection boxes whose center points of the intersection points are located on the left side of this boundary are removed, that is: ; In the formula, represents the set of valid intersection points, that is, the set of the remaining intersection point centers after removing the incomplete points at the edge. All the points in this set are located within the effective detection area of the image and are the point set finally used for uniformity evaluation; represents the j-th intersection point after boundary screening; are the abscissa and ordinate of the center of the intersection point, used to determine whether the position of the point is on the left or right side of a certain threshold; In step S54, "j" represents the points selected from all the intersection point sets, emphasizing that this set has been processed through the previous steps before differentiation, to avoid symbol confusion and highlight the independence and clarity of the screening process.

[0054] Specifically, the specific process of establishing the effective detection boundary standard includes the following steps: S541. Statistically calculate the average value of the abscissas of the right boundaries of all the incomplete intersection point detection frames located on the left edge of the image , which is used to determine the left boundary of the effective detection area, and the expression is: ; In the formula, is the abscissa value of the right boundary of the i-th incomplete intersection point detection frame located on the left edge of the image; N is the total number of incomplete intersection points located on the left edge of the image; S542. Taking the average abscissa value as the benchmark, define the effective detection area of the left boundary of the image; If the abscissa value of the center point of the intersection point , then it is considered that this intersection point is complete and effective; if the abscissa value of the center point of the intersection point , then it is considered that this intersection point is located in the edge area and belongs to an incomplete intersection point, and should be excluded from the subsequent evaluation; the expression is: ; S543. Repeat steps S541 - S542 to process the right, upper, and lower edge areas of the image respectively, to establish an effective detection boundary standard with four boundary directions, and finally determine the set of intersection points in the overall effective detection area. Through this method, the interference of incomplete intersection points located on the edge of the image to the uniformity evaluation can be effectively excluded, and the evaluation accuracy and reliability can be improved.

[0055] S55. Process the other edge areas of the image (i.e., the remaining upper, lower, and right boundaries) according to the above standards, so as to obtain an effective detection point set after removing the incomplete intersection points at the edges , and the expression is: ; In this embodiment, the processing of the left edge in steps S53 - S54 is equally applicable to other edge regions of the image, namely the remaining upper, lower, and right boundaries. In other words, for each edge, the following process is "repeated": finding the incomplete points of the edge, calculating the average boundary, establishing a threshold criterion, and removing the invalid points of the edge. Step S55 precisely illustrates that the upper, lower, and right edges are processed by a similar method to ensure that all incomplete intersection points of the four surrounding boundaries are excluded. Finally, a set of effective detection points after processing all edges is obtained. 。

[0056] Through the above edge intersection point removal steps, it is possible to effectively avoid the calculation deviation of the yarn uniformity index caused by incomplete edge intersection points, improve the accuracy of the uniformity evaluation result, and ensure the industrial reliability and practical application value of the detection result.

[0057] S6. Calculate the evaluation index for evaluating the yarn distribution uniformity in the fan blade preform according to the position distribution of effective intersection points. to achieve an accurate evaluation of the yarn uniformity of the preform. The specific process includes the following steps: S61. Calculate the actual area of the effective detection region according to the position distribution of the effective intersection points located within the effective detection region after being processed by step S5. to provide an accurate spatial calculation basis for the uniformity evaluation. After the edge processing in step S5, the boundary coordinates of the effective detection region in the image are determined. The boundary coordinates of the effective detection region in the image are: and , where (left boundary), (right boundary), (upper boundary), (lower boundary). These boundaries together enclose a rectangular region, which is the effective detection region. The actual area The calculation formula is: ; The actual area is the area of the region in the image that is truly used for uniformity calculation after removing the incomplete edge intersection points.

[0058] S62. Count the number of intersection points within the effective region after edge processing. Take the number of intersection points per unit area as the yarn uniformity evaluation index. to accurately quantify and evaluate the weaving quality and uniformity of the fan blade preform, that is: ; In this embodiment, the yarn uniformity evaluation index It can provide references for process designers and production quality inspectors, and compare the yarn uniformity evaluation index calculated in actual production with the design value of each preform component to check whether the number of intersection points per unit area is too high or too low, so as to judge the uniformity.

[0059] During the evaluation, compare the calculated yarn uniformity evaluation index with the expected ideal uniformity value or this value of the product during the design stage to judge the uniformity of the yarn distribution of the preform. The evaluation basis usually comes from design or experience standards: for example, according to the preform process requirements, the ideal number of intersection points per unit area is .

[0060] If the yarn uniformity evaluation index is close to the reference value , and within the allowable error range (the empirical value according to the on-site process requirements is generally 0.2%), it is considered that the yarn distribution uniformity of the preform of the fan blade is good and meets the quality requirements. At this time, a conclusion can be given: the weaving quality of the preform is reliable, and the yarn density in each area is balanced.

[0061] If the yarn uniformity evaluation index is significantly lower than the reference value , it indicates that the number of yarn intersection points per unit area is too small, the yarn density may be too small, and there may be sparse or loopholes in weaving; on the contrary, if the yarn uniformity evaluation index is much higher than the reference value , it may also mean that the local yarn is crowded or repeated, indicating uneven weaving. Therefore, the real-time calculation of uniformity will provide real-time uniformity information reference for the on-site production department and production personnel.

[0062] Those of ordinary skill in the art can understand that all or part of the steps in implementing the above embodiments can be completed by instructing relevant hardware through a program. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between each embodiment can be referred to each other. For the above embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0064] The above embodiments have introduced the present invention in detail. Specific examples are used in this article to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for detecting uniformity of an aircraft engine fan blade preform, characterized in that: The method comprises the following steps: S1. Acquire the original image collected in real time from the production line, and correct the interweaving points of the carbon fiber horizontal and vertical yarns in the original image to a standard angle presented at a uniform viewing angle through an image angle standardization module; S2, based on the global path and the local path of the dual-path feature extraction network, extract features from the corrected original image to obtain dual-path features; S3, fusing dual-path features to obtain fused features through a dual-path feature fusion network using a cross-attention mechanism and a multi-scale self-attention mechanism; S4, input the fused features into the context decoupling detection head module, decouple the interweaving point detection task into two subtasks: center point prediction and size estimation, and obtain the center position of the interweaving point in the detection frame and the width and height of the detection frame; S5, defining an image edge region threshold, demarcating an effective detection region based on a detection frame and removing incomplete interlaced points existing at the edge of the image, and obtaining a position distribution of effective interlaced points within the effective detection region; S6. Calculate the evaluation index for evaluating the uniformity of yarn distribution in the fan blade preform based on the effective interlacing point position distribution , achieving accurate evaluation of preform yarn uniformity.

2. The uniformity detection method according to claim 1, characterized in that: In step S1, the specific process includes the following steps: S11. Using the ResNet50 network model as the basic feature extraction network, the initial feature map is extracted from the original image, and then the feature pyramid network is used to generate multi-scale feature maps layer by layer to capture the visual features of interweaving points at different scales. S12. Setting a rotation sensitive anchor frame according to different sizes and tilt angles of the yarn interlacing points , the expression is: ； in, is the center coordinate of the rotation-sensitive anchor box; is the width and height of the rotation-sensitive anchor box; is the rotation angle of the rotation-sensitive anchor box; S13, based on rotation-sensitive anchor box Use the region proposal network head to generate rotation candidate regions and obtain the angle prediction value of the rotation multi-scale feature map , and unify the multi-scale feature maps of different sizes into a feature map of fixed size containing interlaced points through ROIAlign operation; S14. Develop an abnormal angle filtering mechanism, through the angle threshold Angle prediction value The abnormal angles in are eliminated, and the expression is: ； in, is the predicted angle value that has been retained; is the rotation interlaced point feature map The angle prediction value of is the predicted mean of the rotation angle; S15, angle prediction value The correction angle is obtained by averaging , the original image is corrected by the rotation transformation matrix, and the rotation transformation matrix is: ； in, represents the pixel coordinates of the interlaced points in the original image; is the coordinate of the interlaced point after the rotation transformation matrix.

3. The uniformity detection method according to claim 2, characterized in that: The rotation angle regression branch is introduced into the region proposal network head, namely: First, for each initial rotation-sensitive anchor box A series of discrete initial angles are assigned to cover possible weave texture tilt angles; Then, the region proposal network head uses convolutional features to represent each rotation-sensitive anchor box. The angle offset is used for regression prediction to obtain the rotation-sensitive anchor frame. The correction value relative to its initial angle , rotation sensitive anchor box The final predicted rotation angle is recorded as , and its calculation formula is: ； In the formula, is a rotation-sensitive anchor box The initial angle of is the angle increment of the regression output of the region proposal network head.

4. The uniformity detection method according to claim 1, characterized in that: In step S2, the dual-path features include global semantic features for characterizing the overall spatial distribution and contextual semantic features of the interweaving points, and local spatial detail features including detail textures of the interweaving points and local fine spatial information. The specific process includes the following steps: S21, the global path adopts a global feature extraction architecture based on visual Transformer, and captures the overall spatial distribution and contextual semantic features of the preform interweaving points through the self-attention mechanism; S22, the local path adopts the classic convolutional neural network structure, and effectively extracts the detailed texture and local fine spatial information of the interweaving points through multi-layer convolution operations; S23, define the feature outputs of the global path and the local path respectively and ,Right now: ； in, Represents the mapping function of extracting global semantic features through visual Transformer; Represents the mapping function of extracting local spatial detail features through a classic convolutional neural network.

5. The uniformity detection method according to claim 1, characterized in that: In step S3, the specific process includes the following steps: S31. Use the cross-attention mechanism to take the global semantic features output by the global path as the query, and the local spatial detail features output by the local path as the key and value respectively, to achieve deep interactive fusion between the two path features to obtain the fused features. , the expression is: ； Among them, query ;key ;value is the feature transformation matrix; is a learnable parameter; is the feature dimension; S32, using the multi-scale self-attention mechanism to fusion features at multiple scales Perform weighted integration to capture the cross-scale dependencies between intertwined points of different scales to obtain comprehensive features , the expression is: ； in, represents a set of scales; Indicates the corresponding scale The attention weight coefficient under ; Represents the comprehensive features after fusion of multi-scale self-attention mechanism; represents the selected scale set; Indicated in scale The fusion features obtained below.

6. The uniformity detection method according to claim 1, characterized in that: In step S4, the context decoupling detection head module includes a center point attention module, a size attention module and an offset attention module. The specific process includes the following steps: S41, center point attention module, converts the position prediction in the intersection point detection task into the center point heat map prediction through the loss function Optimize the center position of the network training output interlacing point, the expression is: ； In the formula, Predict the predicted value of the heat map at position (i, j) for the network; is the label value of the real heat map at position (i, j); is the modulation factor of focus loss; is the number of real interlaced points in the image; S42, size attention module, independently predicts the detection box size including width and height of each intersection center position, and uses the loss function Optimize the prediction error, the expression is: ； In the formula, and is the predicted width and height of the i-th intersection point detection box; and is the actual width and height of the i-th intersection point detection box; S43, offset attention module, predicts the slight offset of the actual center position of the interweaving point relative to the predicted center position of the heat map, and uses the loss function Reduce the position prediction error and further improve the precision of the intersection point positioning. The loss function The expression is: ； In the formula, and is the offset of the center of the i-th interlaced point predicted by the network relative to the grid; and is the corresponding true offset; It is a smoothed absolute value loss function, which is a loss function used to measure the prediction offset error.

7. The uniformity detection method according to claim 1, characterized in that: In step S5, the specific process includes the following steps: S51, defining an image edge region threshold to accurately define an effective detection region and remove possible incomplete edge interlacing points; S52, a set of all detected interweaving points In the above example, the incomplete interlaced points at the edge of the image are identified and marked, which are defined as follows: ； in, is the center point of the interlaced points, and is the center coordinate; , , , Respectively represent the preset left, right, upper and lower boundary thresholds; is the horizontal coordinate of the i-th valid interlacing point in the image; is the vertical coordinate of the i-th valid interlacing point in the image; S53, calculating the average value of the right boundary of the incomplete interweaving point at the left edge of the detection frame, and recording it as the average boundary ,Right now: ； In the formula, is the number of interlacing points on the left edge; For the The horizontal coordinate position of the right boundary of the left edge incomplete interweaving point detection frame; S54, according to the average boundary The effective detection boundary standard is established, and the detection boxes whose center points of all interlaced points are located on the left side of the boundary are removed, that is: ； In the formula, represents the set of valid interleaving points; represents the jth interlaced point after boundary screening; are the abscissa and ordinate of the center of the interlacing point; S55: Process the edge regions including the upper, lower and right borders of the image according to the above standards, so as to obtain a set of valid detection points after removing the incomplete interlaced points at the edges. .

8. The uniformity detection method according to claim 7, characterized in that: In step S54, the specific process of establishing the effective detection boundary standard includes the following steps: S541, counting the average value of the horizontal coordinates of the right borders of all incomplete interlaced point detection boxes located at the left edge of the image , used to determine the left boundary of the effective detection area, the expression is: ； In the formula, is the right boundary abscissa value of the i-th incomplete interlaced point detection frame located at the left edge of the image; N is the total number of incomplete interlaced points located at the left edge of the image; S542, average value of horizontal axis As a benchmark, define the effective detection area on the left edge of the image; If the horizontal coordinate value of the center point of the intersection point , then the intersection point is considered complete and valid; if the horizontal coordinate value of the center point of the intersection point is , then the intersection is considered to be located in the edge area and is an incomplete intersection, and should be excluded from subsequent evaluation; S543, repeating steps S541-S542, processing the right, upper and lower edge regions of the image respectively, to establish an effective detection boundary standard with four boundary directions.

9. The uniformity detection method according to claim 1, characterized in that: In step S6, the specific process includes the following steps: S61. Calculate the actual area of ​​the effective detection area , the calculation formula is: ； In the formula, , is the boundary coordinate of the effective detection area in the image; S62. Count the number of intersection points in the effective area The number of interlacing points per unit area is used as the evaluation index of yarn uniformity. , accurately and quantitatively evaluate the weaving quality and uniformity of fan blade preforms, namely: ； If the yarn uniformity evaluation index Close to reference value , and it is within the allowable error range, then it is considered that the yarn distribution uniformity of the fan blade preform is good and meets the quality requirements; If the yarn uniformity evaluation index Significantly lower than the reference value , indicating that there are too few yarn interlacing points per unit area, the yarn density may be too small, and there are sparse weaving or holes; on the contrary, if the yarn uniformity evaluation index Higher than reference value , meaning localized yarn crowding or duplication, indicating uneven weaving.

Images