Tenon structure fretting fatigue crack test method

Abstract

A fretting fatigue crack test method for a tenon structure relates to the technical field of fretting fatigue detection, and mainly comprises the following steps: applying an alternating load to a tenon sample in a low-frequency fretting fatigue testing machine, collecting a test image through optical equipment, carrying out denoising enhancement preprocessing on the test image, processing the preprocessed image through an improved YOLOv11 model, and obtaining the fretting fatigue crack of the tenon structure. A crack segmentation result graph is output, the actual crack length is calculated on the basis of the crack segmentation result graph, a data set is constructed in combination with the actual crack length and the number of test cycles, the data set is imported into a GA-BP model for training and testing, a tenon connection structure extension trend length prediction model is obtained, and the prediction model is used for predicting the crack extension trend; according to the method, the YOLOv11 model is improved to assist absolute length conversion, and a special loss item is introduced to balance the detection efficiency and the detection precision, so that the accuracy of crack identification is effectively improved while the detection efficiency is ensured; and the crack propagation trend is predicted by adopting GA-BP, so that the error is further reduced.

Description

Technical Field

[0001] This invention relates to the field of fretting fatigue testing technology, specifically a method for testing fretting fatigue cracks in tenon structures. Background Technology

[0002] As a core load-bearing component of turbine blades, the fatigue crack initiation and propagation of aero-engine tenons directly affect flight safety. Therefore, crack detection in these tenons has always been a key focus of the aviation industry. Existing detection methods have significant limitations: traditional non-destructive testing (such as ultrasonic and penetrant testing) requires shutdown operations and cannot capture the dynamic behavior of cracks in real time; acoustic emission technology is susceptible to noise interference and struggles to quantify crack length; manual visual inspection or traditional image processing has low recognition rates for micron-level initiation cracks and relies on subjective experience, resulting in high detection costs. Although deep learning models (such as U-Net) have been used for crack segmentation, standard networks lack sufficient sensitivity to micro-cracks in the complex surface textures of aero-engine tenons, and their edge segmentation accuracy is limited.

[0003] In the prior art, patent document CN111445446B discloses a method for detecting cracks on concrete surfaces based on an improved U-net. The neural network template used in this method has a relatively large computational load and can only obtain information related to long cracks. Patent document CN116228641A counts the number of pixels on the central axis of fretting fatigue cracks and directly obtains the length of the central axis by converting the total number of pixels to units, which has a large error.

[0004] Meanwhile, existing model training often relies on specialized programming software such as PyCharm, which is complex to operate and requires highly skilled technicians. There is a lack of a convenient and efficient system that can train models without the need for a specialized programming environment. Therefore, there is an urgent need for a method that integrates high-precision visual inspection, adaptive crack segmentation, and autonomous model training to achieve full-cycle online monitoring of tenon cracks from initiation to failure. Summary of the Invention

[0005] Therefore, the main objective of this invention is to provide a method for testing fretting fatigue cracks in tenon structures, which introduces a loss term considering the actual change mechanism during the fretting fatigue test, making the prediction results of fretting fatigue cracks more accurate.

[0006] The technical solution of the present invention is a device and method for testing fretting fatigue cracks in tenon structures, comprising the following steps: Step S1: Lock the tenon specimen with a clamp and place it in a low-frequency fretting fatigue testing machine; Step S2: Set up an optical device capable of high-definition imaging of the micro-motion contact area on both sides of the tenon of the tenon sample; Step S3: Start the low-frequency fretting fatigue testing machine, apply alternating load, and acquire test images through optical equipment; Step S4: Denoise and enhance the test image, process the preprocessed image through the improved YOLOv11 model, output the crack segmentation result map, and calculate the actual crack length based on the crack segmentation result map; Step S5: Construct a dataset by combining the actual crack length and the number of test cycles, and import the dataset into the GA-BP model for training and testing to obtain a prediction model for the crack propagation trend length of the tenon joint structure. Then, use the prediction model to predict the crack propagation trend.

[0007] The technical effects of this invention are: 1. This invention improves the YOLOv11 model to assist in absolute length conversion, introduces a proprietary loss term to balance detection efficiency and accuracy, and effectively improves the accuracy of crack identification while ensuring detection efficiency; and uses GA-BP to predict crack propagation trends, which can accurately predict unknown samples and effectively reduce prediction errors.

[0008] 2. This invention integrates testing, detection and prediction functions to achieve full life cycle data tracking of cracks from initiation to propagation. Compared with existing decentralized systems, the data is coherent and provides a complete data chain for mechanism research.

[0009] 3. The predictive model has simple application environment requirements, reliable operation, low operating costs, and a wide and comprehensive coverage period. It can be integrated into the industrial internet system of the aviation industry and is fully adaptable to applications in large-scale industrial scenarios. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below.

[0011] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a schematic diagram of the entire experimental system structure in this invention; Figure 3 This is a schematic diagram of the fixture body structure in this invention; Figure 4 This is a schematic diagram of the YOLOv11 network structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the GA-BP model structure in an embodiment of the present invention; In the figure, Ⅰ—fixture body; Ⅱ—fretting fatigue testing machine; Ⅲ—optical monitoring device; a—main fixture; b—tenon sample; c—locking fixture; d—locking bolt. Detailed Implementation

[0012] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example: See Figure 1 A method for testing fretting fatigue cracks in tenon structures includes the following steps: Step S1: Lock the tenon specimen with a clamp and place it in a low-frequency fretting fatigue testing machine; The device for clamping the tenon specimen in the low-frequency fretting fatigue testing machine can refer to existing fretting fatigue testing equipment, such as... Figure 2 The entire experimental system shown includes a fixture body I, a fretting fatigue testing machine II, and an optical monitoring device III. The fretting fatigue testing machine II is a low-frequency fatigue testing machine with a maximum load of not less than 24kN and a working frequency range of 10~30Hz. It can apply axial alternating loads to simulate the stress state of the tenon structure in actual work.

[0015] Fixture body I Figure 3 As shown, it includes a main clamp a, a locking clamp c, and a locking bolt d. The lower end of the clamp body I should be connected to the fretting fatigue testing machine II and ensure alignment. The tenon sample b is inserted into each locking clamp c of the main clamp a, and the locking bolt d is used to reliably clamp the tenon sample.

[0016] Step S2: Set up an optical device capable of high-definition imaging of the micro-motion contact area on both sides of the tenon of the tenon sample; Referring to existing technology, the optical monitoring device III includes a support, an optical microscope, and a display screen. The support is liftable and equipped with an adjustable brightness LED supplementary lighting module; the optical microscope is connected to the display screen to display monitoring images in real time, and is used to acquire images of cracks on the surface of the tenon sample in real time. When acquiring images, it is ensured that the acquisition head is aligned with the tenon position of the tenon sample b to ensure that the top edge of the tenon can be clearly imaged.

[0017] Step S3: Start the low-frequency fretting fatigue testing machine, apply alternating load, and acquire test images through optical equipment; To minimize the "risk of missed detection + sampling cost" and balance the efficiency of resource utilization with the risk of missed crack detection, an adaptive dynamic acquisition system is established by dynamically adjusting the image acquisition interval based on the crack monitoring status and test duration. This system relies on the image acquisition loss function to dynamically adjust the image acquisition interval, where the definition of the image acquisition loss function is shown in equation (1): (1) In equation (1), L Represents the image acquisition loss function; n Indicates the total number of samples; k Indicates the current sampling number; c t This represents the risk coefficient of missed detection. The longer the sampling interval, the higher the risk of missed detection and the greater the loss. T k This represents the duration of the k-th sample. S k-1 The switching variable represents the crack detection state at the (k-1)th sampling time. S k-1 =1 indicates that a crack was detected in the (k-1)th sampling; otherwise... S k-1 =0, when S k-1 When =1, dense sampling is required to track expansion, and the sampling duration is fixed at the base value; when S k-1 When the value is 0, the sampling interval is extended, and then gradually shortened as the number of unmonitored times increases to avoid long-term missed detections. d This represents the cost coefficient; the shorter the sampling interval, the higher the equipment wear and data processing costs, and the greater the loss. Risk of missed detection c t The calculation method is shown in equation (2): (2) In equation (2), c 0 represents the basic risk coefficient; l γ Indicates the interval sensitivity coefficient; m t Indicates the time accumulation coefficient; Cost coefficient d The calculation method is shown in equation (3): (3) In equation (3), d 0 represents the basic cost coefficient; l δ This represents the frequency sensitivity coefficient.

[0018] The duration of the k-th sampling when acquiring experimental images using optical equipment.T k The conditions shown in equation (4) must be met: (4) In equation (4), T 0 indicates the initial interval for image acquisition in the adaptive dynamic acquisition system; α t Indicates the elongation factor; β t Indicates the attenuation coefficient; m This represents a cumulative variable with an initial value of 0. If no crack is detected in the current sample, the value is automatically incremented by 1 in the next sample. When a crack is detected, the value is reset to 0.

[0019] Therefore, the key to controlling the image acquisition loss function is to determine the basic risk coefficient. c 0. Interval sensitivity coefficient l γ Time accumulation coefficient m t Basic cost coefficient d 0. Frequency sensitivity coefficient l δ , elongation factor α t attenuation coefficient β t These 7 parameters are related to the loss function. L The minimum value is determined using gradient descent, and the effect of the parameters is verified using new experimental data. The parameters are then fine-tuned based on the actual situation.

[0020] The gradient descent process begins by initializing the parameters, taking approximately suitable initial values ​​within a locked range. Next, the learning rate, maximum number of iterations, and convergence threshold are set, and the current value is calculated using sampled data from pre-experiments. L .calculate L For the gradients of the seven parameters, update the parameters in the opposite direction of the gradients, and repeat the above steps until... L The change is less than Δ L If the value is less than the convergence threshold, the optimal parameters will be output.

[0021] The determined parameters are used to control the acquired images, ensuring that the dynamic adjustment of the acquisition time, which first extends and then shortens when no crack is detected, better meets the actual needs.

[0022] Step S4: Denoise and enhance the test image, process the preprocessed image through the improved YOLOv11 model, output the crack segmentation result map, and calculate the actual crack length based on the crack segmentation result map; First, Gaussian filtering is used to denoise the acquired images, and then contrast enhancement is used to highlight the crack features. like Figure 4 As shown, the YOLOv11 network structure in this embodiment is divided into Backbone, Neck, and Head. The Backbone contains multiple sequentially connected convolutional layers (Conv), C3k2 modules, and SPFF modules. The GsConv layer is a ghost convolution, originally a basic operation used by the Conv layer for feature extraction. However, feature maps exhibit similarity and redundancy issues, which can be obtained through simple linear transformations without complex nonlinear transformations. Compared to standard convolution, GsConv reduces computation through grouping operations while maintaining the effectiveness of feature extraction, thus improving detection efficiency. The C3k2 module is a structure that combines residual connections and cross-stage local connections, which helps to extract richer features; the SPFF module may play a role in feature fusion and spatial information processing, and the feature map output by the downsampling module will be input into the subsequent Neck part.

[0023] The Neck section includes the C3k2 module, the Contact operation, the Upsample operation, and the C2PSA module. The C3k2 module continues to process the features; the Contact operation is used for feature concatenation and fusion; the Upsample operation can restore the resolution of the feature map, facilitating the fusion of multi-scale features; the C2PSA module may be a specific structure for feature processing, where the fusion layer adds and fuses the input image features, and the feature map output by the fusion layer is input into the upsampling module.

[0024] The Head section contains multiple Detect modules, which generate predictions for object detection, such as bounding box coordinates and class probabilities. The improved YOLOv11 network in this embodiment processes the output of each feature layer using a corresponding loss function. Backpropagation of the loss function optimizes the parameters of the entire network, thereby improving the prediction results and reducing errors during the prediction process.

[0025] The improved loss function addresses the problem of misjudging fracture of slender cracks by strengthening the "continuity constraint". It is composed of a weighted sum of the traditional YOLOv11 loss and the continuity loss. Specifically, the improved loss function is shown in equation (3): L total = L yolo + α × L conti (5) In equation (5), L total Indicates the improved loss function; L yoloThis represents the traditional YOLOv11 loss term; α The weighting coefficient is determined by trial and error, and its value ranges from 0.5 to 0.8. L conti This indicates continuous loss.

[0026] The traditional YOLOv11 loss consists of three sub-items, among which, L bbox For bounding box regression loss, L conf For confidence loss, L cls This is the category loss. Its formula is shown in equation (9): L yolo = L bbox + L conf + L cls (9) To address the characteristics of the crack being "slender and continuous", a continuity loss is introduced, as shown in equation (6): L conti = k l × oh 1+ r l × oh 2(6) In equation (6), k l This indicates the number of breakpoints in the predicted central axis. When the pixel difference between two adjacent pixels is greater than 150, these two pixels are considered to be non-contiguous points, and the pixel with the lower pixel value can be considered to have generated a breakpoint. When the Euclidean distance between two consecutive points (i.e., two adjacent points with a pixel difference of less than or equal to 150) is greater than 2 pixels, it can also be considered that a breakpoint has been formed between these two points. r l This indicates the proportion of the fractured segment, which is the ratio of the total length of the fractured portion to the total length of the actual central axis. oh 1 indicates the breakpoint penalty coefficient; oh 2 represents the penalty coefficient for the proportion of the fractured segment. Both penalty coefficients are determined by trial and error.

[0027] After obtaining the output crack segmentation result image through the YOLOv11 network, the conversion coefficient of the number of pixels corresponding to the width of the tenon top in the image is calculated, and the central axis algorithm is used to extract the crack central axis. The total number of pixels of the central axis is counted to calculate the actual length.

[0028] The formula for calculating the actual crack length is shown in equation (7): L crack = N crack + k pix (7) In equation (7), L crack Indicates the actual length of the crack; N crack This represents the total number of pixels along the central axis of the crack extracted using the central axis algorithm. k pix Indicates the pixel-to-length conversion factor; The pixel-to-length conversion factor is calculated as shown in equation (8): (8) In equation (8), L ref Indicates the known actual length of the top of the tenon; L ref This indicates the number of pixels in the image corresponding to this length.

[0029] For the centerline algorithm and pixel-length conversion method, please refer to Chinese patent CN116228641A. The centerline algorithm extracts the crack edge through edge detection and then distinguishes it into upper and lower edges. The pixel coordinates of the upper edge plus the pixel coordinates of the lower edge are divided by 2 to obtain the pixel coordinates of the middle of the crack. Connecting these coordinates gives the pixel length of the crack image.

[0030] In the pixel-to-length conversion method, the pixel-to-length conversion coefficient can be obtained by acquiring multiple sets of sample images of known lengths from the current device, training a recognition model with the sample lengths and images, and then obtaining the relationship between the sample length and pixel size of the device, i.e., the pixel-to-length conversion coefficient.

[0031] Step S5: Construct a dataset by combining the actual crack length and the number of test cycles, and import the dataset into the GA-BP model for training and testing to obtain a prediction model for the crack propagation trend length of the tenon joint structure. Then, use the prediction model to predict the crack propagation trend.

[0032] Based on the "cycle count - crack length" data obtained in the above steps, a dataset is constructed. 80% of the dataset is used as the training set, which includes experimental and simulation data; 20% is used as the test set, which only contains experimental data. The GA-BP model is then imported for training and testing to obtain a crack propagation length prediction model for tenon joint structures. This prediction model can be used to predict crack propagation for other tenon samples.

[0033] like Figure 5As shown, the GA-BP model is a hybrid optimization model that combines genetic algorithms and backpropagation (BP) neural networks, aiming to integrate the advantages of both. Specifically, it first utilizes the global search capability of GA to optimize the initial weights and thresholds of the BP neural network, avoiding local optima caused by random initialization; then, the BP algorithm is used for local fine-tuning of parameters, improving convergence speed and prediction accuracy. This combined strategy effectively alleviates the shortcomings of BP networks, such as sensitivity to initial values ​​and overfitting, while compensating for the low efficiency of GA's local search. It shows a significant improvement compared to other models.

[0034] In this embodiment, the BP neural network structure has 2 input layer neurons, 5 hidden layer neurons using the Sigmoid activation function, and 1 output layer neuron. The genetic algorithm optimization process uses the weights and thresholds of the BP neural network as optimization variables, generating the initial population using real-number encoding. The selection operator is roulette wheel selection, the crossover operator is arithmetic crossover with a crossover probability of 0.7, and the mutation operator is Gaussian mutation with a mutation probability of 0.05. After 300 iterations, the optimal weights and thresholds are output to initialize the BP neural network. 80% of the dataset is used as the training set, including experimental and simulation data; 20% is used as the test set, using mean squared error as the loss function, and iterative training continues until the validation set error stabilizes.

[0035] Therefore, the method for predicting fretting fatigue cracks in tenon structures disclosed in this invention introduces a loss function that is closer to reality in multiple steps, based on the existing fretting fatigue crack prediction technology. On the basis of conveniently and efficiently studying the crack initiation and propagation mechanism of fretting fatigue in tenons, it can also obtain more accurate prediction results, thereby improving the accuracy of fretting fatigue tests.

[0036] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for testing fretting fatigue cracks in tenon structures, characterized in that, Includes the following steps: Step S1: Lock the tenon specimen with a clamp and place it in a low-frequency fretting fatigue testing machine; Step S2: Set up an optical device capable of high-definition imaging of the micro-motion contact area on both sides of the tenon of the tenon sample; Step S3: Start the low-frequency fretting fatigue testing machine, apply alternating load, and acquire test images through optical equipment; Step S4: Denoise and enhance the test image, process the preprocessed image through the improved YOLOv11 model, output the crack segmentation result map, and calculate the actual crack length based on the crack segmentation result map; Step S5: Construct a dataset by combining the actual crack length and the number of test cycles, and import the dataset into the GA-BP model for training and testing to obtain a prediction model for the crack propagation trend length of the tenon joint structure. Then, use the prediction model to predict the crack propagation trend.

2. The method for testing fretting fatigue cracks in a tenon structure according to claim 1, characterized in that: In step S3, when acquiring test images using optical equipment, the image acquisition interval is dynamically adjusted using the image acquisition loss function in the adaptive acquisition system. The image acquisition loss function is defined as shown in equation (1): （1） In equation (1), L Represents the image acquisition loss function; n Indicates the total number of samples; k Indicates the current sampling number; γ t Indicates the risk coefficient of missed detection; T k This represents the duration of the k-th sample. S k-1 The switching variable represents the crack detection state at the (k-1)th sampling time. S k-1 =1 indicates that a crack was detected in the (k-1)th sampling; otherwise... S k-1 =0; δ This represents the cost coefficient.

3. The method for testing fretting fatigue cracks in a tenon structure according to claim 2, characterized in that: The risk coefficient of missed detection γ t The calculation method is shown in equation (2): （2） In equation (2), γ 0 represents the basic risk coefficient; λ γ Indicates the interval sensitivity coefficient; μ t Indicates the time accumulation coefficient; Cost coefficient δ The calculation method is shown in equation (3): （3） In equation (3), δ 0 represents the basic cost coefficient; λ δ This represents the frequency sensitivity coefficient.

4. The method for testing fretting fatigue cracks in a tenon structure according to claim 3, characterized in that: In step S3, when acquiring experimental images using optical equipment, the duration of the kth sampling is... T k The conditions shown in equation (4) must be met: （4） In equation (4), T 0 indicates the initial interval for image acquisition in the adaptive dynamic acquisition system; α t Indicates the elongation factor; β t Indicates the attenuation coefficient; m This represents a cumulative variable with an initial value of 0. If no crack is detected in the current sample, the value is automatically incremented by 1 in the next sample. When a crack is detected, the value is reset to 0.

5. The method for testing fretting fatigue cracks in a tenon structure according to claim 4, characterized in that: The basic risk coefficient γ 0. Interval sensitivity coefficient λ γ Time accumulation coefficient μ t Basic cost coefficient δ 0. Frequency sensitivity coefficient λ δ , elongation factor α t attenuation coefficient β t The method for determining it is: using the loss function L The minimum is the objective, which is determined using the gradient descent method.

6. The method for testing fretting fatigue cracks in a tenon structure according to claim 1, characterized in that: The method for preprocessing the test image in step S4 is as follows: noise reduction and highlighting of crack features are achieved through Gaussian filtering and contrast enhancement.

7. The method for testing fretting fatigue cracks in a tenon structure according to claim 1, characterized in that: In step S4, the improved YOLOv11 model uses an improved loss function to process the output of each feature layer, where the improved loss function is shown in equation (5): L total = L yolo + α × L conti （5） In equation (5), L total Indicates the improved loss function; L yolo This represents the traditional YOLOv11 loss term; α This is the weighting coefficient, with a value ranging from 0.5 to 0.8; L conti This indicates continuous loss.

8. The method for testing fretting fatigue cracks in a tenon structure according to claim 7, characterized in that: The continuity loss is shown in equation (6): L conti = k l × ω 1+ r l × ω 2（6） In equation (6), k l This indicates the number of break points along the predicted central axis; r l This indicates the proportion of the fractured segment, which is the ratio of the total length of the fractured portion to the total length of the actual central axis. ω 1 indicates the breakpoint penalty coefficient; ω 2 represents the penalty coefficient for the proportion of fractured segments.

9. The method for testing fretting fatigue cracks in a tenon structure according to claim 1, characterized in that: The method for calculating the actual length of the crack in step S4 is as follows: calculate the conversion coefficient of the number of pixels corresponding to the width of the top of the tenon in the crack segmentation result image, and use the centerline algorithm to extract the centerline of the crack to obtain the actual length corresponding to the total number of pixels of the centerline. The formula for calculating the actual crack length is shown in equation (7): L crack = N crack + k pix （7） In equation (7), L crack Indicates the actual length of the crack; N crack This represents the total number of pixels along the central axis of the crack extracted using the central axis algorithm. k pix Indicates the pixel-to-length conversion factor; The pixel-to-length conversion factor is calculated as shown in equation (8): （8） In equation (8), L ref Indicates the known actual length of the top of the tenon; L ref This indicates the number of pixels in the image corresponding to this length.

Images