Tensile Testing System for Carbon Fiber Reinforced Ethylene Propylene Diene Monomer Rubber Material and Its Testing Method

Abstract

The present invention relates to the technical field of material performance detection, and discloses a tensile detection system for a carbon fiber composite ethylene propylene diene monomer rubber material and a detection method thereof. The system includes a real-time acquisition module, a multi-dimensional feature analysis module, a dynamic fusion module, a defect detection module, and a real-time feedback module. The real-time acquisition module uses multi-modal sensors to synchronously obtain strain distribution data, infrared thermal image data, and acoustic emission signal streams; the multi-dimensional feature analysis module processes the data to generate corresponding features; the dynamic fusion module fuses the features to generate an enhanced feature matrix; the defect detection module inputs the matrix into a pre-trained model to output a defect probability distribution map; the real-time feedback module evaluates the tensile state according to the distribution map and issues a graded warning. The invention can comprehensively and accurately detect internal defects of the material during tension, give early warnings in a timely manner, improve the accuracy and reliability of material detection, and ensure the safety of related projects.

Description

Technical Field

[0001] The present invention relates to the technical field of material property detection, and particularly to a tensile detection system for carbon fiber composite ethylene propylene diene monomer (EPDM) rubber material and its detection method. Background Art

[0002] In the modern industrial field, carbon fiber composite EPDM rubber material is widely used in many key fields such as aerospace, automobile manufacturing, and construction engineering due to its excellent comprehensive properties, such as high strength, aging resistance, corrosion resistance, etc. In aerospace, it is used to manufacture components such as aircraft wings and fuselages. When bearing various stresses during flight, the integrity of the material is crucial; in automobile manufacturing, it is used for seals, shock-absorbing components, etc. to ensure the safety performance and comfort of vehicles; in construction engineering, it is used for bridge expansion joints, building waterproofing, etc. to ensure the stability and durability of building structures.

[0003] However, during the actual use of such materials, especially when subjected to tensile forces, various defects are prone to occur inside. During the production process of the materials, due to the quality differences of raw materials and the imperfect processing technology, the internal structure may be uneven, forming potential defects. During long-term use, under the repeated action of environmental factors (such as temperature and humidity changes) and mechanical stresses, the generation and development of internal defects will also be triggered. These defects will seriously affect the mechanical properties and service life of the materials, and further threaten the safety and reliability of related projects.

[0004] Currently, there are many limitations in traditional material tensile detection methods. Common visual inspection methods can only detect obvious defects on the surface of materials and are powerless for tiny or hidden defects inside the materials. Although ultrasonic detection can detect internal defects of materials, the positioning and quantitative analysis of defects are not accurate enough, and it is easily interfered by the complexity of the internal structure of the materials. Although X-ray detection has high precision, the equipment cost is expensive, the detection process is cumbersome, the professional requirements for operators are high, and there is radiation hazard, which is not suitable for on-site rapid detection.

[0005] Some existing detection systems based on single sensors can only obtain one kind of information during the tensile process of materials, such as only detecting strain or only detecting temperature changes, and cannot comprehensively reflect the tensile state of materials. Moreover, in terms of data processing and analysis, these systems often adopt simple algorithms, cannot fully explore the potential information in the data, and are difficult to accurately judge and early warn of internal defects of materials. In practical engineering applications, there is an urgent need for a system and method that can detect internal defects of carbon fiber composite EPDM rubber materials in real time, comprehensively, and accurately under the tensile state to meet the strict requirements of modern industry for material quality and safety. Summary of the Invention

[0006] The purpose of the present invention is to provide a tensile detection system and a detection method for carbon fiber composite EPDM rubber materials to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: A tensile detection system for carbon fiber composite EPDM rubber materials, the system includes:

[0008] A real-time acquisition module, used to synchronously obtain strain distribution data, infrared thermal image data, and acoustic emission signal flow during the tensile process of the material through a multi-modal sensor;

[0009] A multi-dimensional feature analysis module, which performs spatial grid discretization processing on the strain distribution data to generate a strain gradient feature set, and at the same time performs dynamic segmentation of the temperature field on the infrared thermal image data to extract heat conduction anomaly indicators;

[0010] A dynamic fusion module, based on an adaptive weight allocation algorithm, non-linearly fuses the strain gradient feature set, heat conduction anomaly indicators, and frequency domain energy features of the acoustic emission signal flow to generate an enhanced feature matrix;

[0011] A defect detection module, which inputs the enhanced feature matrix into a pre-trained first deep learning model. The model is based on a parallel architecture of a three-dimensional convolutional neural network and an attention mechanism, and outputs a probability distribution map of internal defects of the material;

[0012] A real-time feedback module, which generates a tensile state evaluation result according to the probability distribution map of defects and triggers a hierarchical warning instruction.

[0013] Preferably, the multi-modal sensor includes:

[0014] A high-resolution fiber Bragg grating sensor array, embedded on the surface of the material according to a preset topological structure, for collecting strain distribution data;

[0015] An infrared thermal imager and an acoustic emission sensor group, which respectively obtain thermal image data and acoustic emission signals synchronously at a fixed sampling frequency.

[0016] Preferably, the steps of the spatial grid discretization processing include:

[0017] Dividing the strain distribution data into three-dimensional voxel units according to the geometric parameters of the material; calculating the strain gradient amplitude of each voxel unit using the Laplace operator;

[0018] Performing a difference operation on adjacent voxel units to generate a strain gradient direction consistency index.

[0019] Preferably, the steps of the temperature field dynamic segmentation include:

[0020] Perform morphological opening operation on the infrared thermal image data to eliminate noise interference;

[0021] Identify the boundary of the heat conduction abnormal area based on the region growing algorithm;

[0022] Calculate the time-varying standard deviation of the temperature difference between the abnormal area and the background, and generate a heat conduction anomaly index.

[0023] Preferably, the implementation of the adaptive weight allocation algorithm includes:

[0024] Model the temporal correlation of the strain gradient feature through a gated recurrent unit network;

[0025] Calculate the static weight of the heat conduction anomaly index by using the entropy weight method;

[0026] Analyze the dynamic coupling relationship between the acoustic emission frequency domain energy feature and the strain gradient by using the covariance matrix;

[0027] Fuse the temporal correlation, static weight and dynamic coupling relationship to generate a feature fusion weight vector.

[0028] Preferably, input the enhanced feature matrix into a parallel three-dimensional convolutional layer and a multi-head self-attention layer;

[0029] Extract local spatial features through a three-dimensional convolutional kernel, and model the cross-channel global dependence through an attention mechanism;

[0030] Adopt a channel attention gating mechanism to fuse local and global features, and output a defect probability distribution map.

[0031] Preferably, the trigger logic of the real-time feedback module includes:

[0032] If the probability value of a continuous area in the defect probability distribution map exceeds the first threshold, initiate a first-level warning and reduce the stretching rate;

[0033] If the probability value exceeds the second threshold and the area of the region expands, trigger a second-level warning and terminate the stretching experiment;

[0034] Synchronize the warning instruction and the defect distribution data to the distributed monitoring platform.

[0035] Preferably, the division strategy of the three-dimensional voxel unit includes:

[0036] Dynamically adjust the voxel unit size according to the number of layers in the material thickness direction;

[0037] Adopt an octree structure for local refinement segmentation in the strain mutation area;

[0038] Perform bilinear interpolation on the voxel unit boundary to eliminate the aliasing effect.

[0039] Preferably, the optimization steps of the region growth algorithm include: setting a dynamic growth threshold based on the gradient magnitude of the thermal image data, smoothly expanding the seed point region using an anisotropic diffusion equation, and avoiding overgrowth of the region through contour curvature constraints.

[0040] Preferably, the present invention further includes a tensile detection method for a carbon fiber composite EPDM rubber material, and the method includes:

[0041] Performing real-time detection and early warning of the tensile state of the material through the above-mentioned tensile detection system for the carbon fiber composite EPDM rubber material.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] From the perspective of obtaining detection data, the system uses multi-modal sensors, including a high-resolution fiber Bragg grating sensor array, an infrared thermal imager, and an acoustic emission sensor group. The high-resolution fiber Bragg grating sensor array is embedded on the surface of the material according to a preset topological structure, and can accurately collect strain distribution data, providing key information for analyzing the internal stress changes of the material. The infrared thermal imager and the acoustic emission sensor group work synchronously to obtain thermal image data and acoustic emission signals respectively. The collaborative work of multiple sensors realizes comprehensive monitoring of the tensile process of the material from multiple dimensions, avoids the limitations of single-sensor information acquisition, makes the detection data richer and more comprehensive, and lays a solid foundation for accurately analyzing the state of the material subsequently.

[0044] In terms of data processing and feature analysis, the operation of the multi-dimensional feature analysis module is highly innovative. The strain distribution data is discretized by spatial grid processing. By dividing three-dimensional voxel units, calculating the strain gradient magnitude and the strain gradient direction consistency index, the internal change characteristics of the strain in the material can be deeply explored, and the strain mutation region can be accurately captured, which is of great significance for discovering the location and degree of potential defects. The temperature field of the infrared thermal image data is dynamically segmented, the noise interference is eliminated through morphological opening operation, the boundary of the heat conduction abnormal region is identified using the region growth algorithm, and the heat conduction abnormal index is calculated, which can effectively detect the heat conduction abnormal situation caused by defects inside the material, providing another important basis for defect detection.

[0045] Based on the adaptive weight assignment algorithm, the dynamic fusion module non-linearly fuses the strain gradient feature set, the thermal conduction anomaly index, and the frequency-domain energy features of the acoustic emission signal stream to generate an enhanced feature matrix. This fusion method fully considers the importance and mutual relationship of different features. By modeling the temporal correlation of the strain gradient features through a gated recurrent unit network, calculating the static weight of the thermal conduction anomaly index using the entropy weight method, and analyzing the dynamic coupling relationship between the acoustic emission frequency-domain energy features and the strain gradient through the covariance matrix, the fused feature matrix can more accurately reflect the true state inside the material, greatly improving the accuracy and reliability of defect detection.

[0046] The defect detection module adopts a first deep learning model based on the parallel architecture of a three-dimensional convolutional neural network and an attention mechanism. The three-dimensional convolutional kernel extracts local spatial features, the attention mechanism models cross-channel global dependencies, and then the local and global features are fused through a channel attention gating mechanism to output the probability distribution map of internal defects in the material. This deep learning model can automatically learn the complex features during the tensile process of the material. Compared with traditional detection algorithms, it has stronger feature extraction ability and pattern recognition ability, and can more accurately locate and identify the micro-defects and potential defects inside the material, effectively improving the accuracy and sensitivity of detection.

[0047] The real-time feedback module generates a tensile state evaluation result based on the probability distribution map of defects and triggers a hierarchical early warning instruction. According to different defect probability thresholds and regional change situations, different levels of early warnings are initiated and corresponding measures are taken, such as reducing the tensile rate or terminating the tensile experiment. At the same time, the early warning instruction and the defect distribution data are synchronized to the distributed monitoring platform. This function realizes the real-time monitoring and dynamic adjustment of the tensile process of the material, can timely detect the abnormal state of the material, avoid material failure caused by the development of defects, and ensure the safety and reliability of related projects. In actual production and engineering applications, it can greatly reduce safety accidents caused by material defects, reduce economic losses, improve production efficiency and product quality. At the same time, the application of the distributed monitoring platform facilitates multi-department collaborative work and improves the practicality and operability of the detection system. Description of the Drawings

[0048] Figure 1 It is the working principle diagram of the tensile detection system for the carbon fiber composite ethylene propylene diene monomer rubber material described in the present invention;

[0049] Figure 2 It is the strategy diagram of spatial grid discretization processing and voxel division;

[0050] Figure 3 It is the implementation diagram of the adaptive weight assignment algorithm;

[0051] Figure 4 It is the construction diagram of the deep learning model. Specific Embodiments

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] Please refer to Figures 1 - 4 , the present invention provides a technical solution: a tensile detection system for carbon fiber composite EPDM rubber material, aiming to achieve precise detection and real-time warning of internal defects during the tensile process of the material. The following will elaborate on the specific embodiments of the present invention in detail:

[0054] The real-time acquisition module synchronously obtains key data during the tensile process of the material through multimodal sensors. At the beginning of the tensile experiment of the material, the multimodal sensors start to work. The high-resolution fiber Bragg grating sensor array is embedded in the surface of the material according to the preset topological structure. As the material deforms under the action of tensile force, this sensor array can accurately collect strain distribution data; the infrared thermal imager and the acoustic emission sensor group respectively obtain thermal image data and acoustic emission signals synchronously at a fixed sampling frequency. These data comprehensively reflect the physical state changes of the material during tensile.

[0055] The multi-dimensional feature analysis module deeply processes the collected data. For the strain distribution data, first divide it into three-dimensional voxel units according to the geometric parameters of the material, then calculate the strain gradient amplitude of each voxel unit using the Laplace operator, perform differential operations on adjacent voxel units, and generate a strain gradient direction consistency index, thereby obtaining a strain gradient feature set. For the infrared thermal image data, first perform morphological opening operation on it to eliminate noise interference, identify the boundary of the heat conduction abnormal area based on the region growing algorithm, calculate the time-varying standard deviation of the temperature difference between the abnormal area and the background, and generate a heat conduction abnormality index.

[0056] The dynamic fusion module non-linearly fuses the strain gradient feature set, the heat conduction abnormality index, and the frequency domain energy feature of the acoustic emission signal stream based on the adaptive weight allocation algorithm. Model the temporal correlation of the strain gradient feature through a gated recurrent unit network; calculate the static weight of the heat conduction abnormality index using the entropy weight method; analyze the dynamic coupling relationship between the acoustic emission frequency domain energy feature and the strain gradient using the covariance matrix; fuse these factors to generate a feature fusion weight vector, and finally generate an enhanced feature matrix.

[0057] The defect detection module inputs the enhanced feature matrix into a pre-trained first deep learning model based on a parallel architecture of a three-dimensional convolutional neural network and an attention mechanism. The enhanced feature matrix is input into a parallel three-dimensional convolutional layer and a multi-head self-attention layer. The three-dimensional convolutional kernel extracts local spatial features, and the attention mechanism models cross-channel global dependencies. Then, a channel attention gating mechanism is used to fuse local and global features, and a probability distribution map of internal defects of the material is output.

[0058] The real-time feedback module generates a tensile state evaluation result based on the probability distribution map of defects and triggers a hierarchical early warning instruction. If the probability value in a continuous area in the probability distribution map of defects exceeds the first threshold, a first-level early warning is initiated and the stretching rate is reduced; if the probability value exceeds the second threshold and the area of the region expands, a second-level early warning is triggered and the stretching experiment is terminated; finally, the early warning instruction and the defect distribution data are synchronized to the distributed monitoring platform.

[0059] The present invention will be further described below in conjunction with Embodiments 1 to 6:

[0060] Embodiment 1:

[0061] In this embodiment, the specific composition and working principle of the multi-modal sensor are described in detail. The multi-modal sensor plays a key role in data acquisition in the entire detection system, and its performance directly affects the accuracy of subsequent detection results.

[0062] The high-resolution fiber Bragg grating sensor array is embedded in the material surface according to a preset topological structure. In actual operation, the design of the preset topological structure needs to be determined according to the shape, size of the material and the expected stress conditions. For example, for a rectangular carbon fiber composite ethylene propylene diene monomer (EPDM) rubber material specimen, the sensor array can be distributed in a regular grid pattern on the material surface to ensure full coverage of the areas where strain changes may occur. When the material is under tension, the strain of the fiber Bragg grating will be caused by the deformation of the material. According to the strain-wavelength relationship of the fiber Bragg grating, the high-resolution fiber Bragg grating sensor array can convert the strain change on the material surface into a wavelength change, and then accurately collect strain distribution data. This kind of sensor has the advantages of high precision and strong anti-interference ability, and can reflect the strain situation on the material surface in real time and accurately.

[0063] The infrared thermal imager and the acoustic emission sensor group respectively obtain thermal image data and acoustic emission signals synchronously at a fixed sampling frequency. The infrared thermal imager generates a thermal image by detecting the infrared radiation energy on the surface of an object. During the tensile process of the material, the internal defects or stress concentration areas in the material will cause abnormal heat transfer. The infrared thermal imager takes pictures of the material surface at a fixed sampling frequency (such as 25 frames per second) to obtain thermal image data at different times. These thermal image data record the distribution and change of the surface temperature of the material, providing a basis for subsequent analysis of abnormal heat conduction.

[0064] The acoustic emission sensor group is used to monitor the acoustic emission signals generated by the internal structural changes of the material during the tensile process. When phenomena such as microcrack propagation and friction occur inside the material, elastic waves, that is, acoustic emission signals, will be generated. The acoustic emission sensor group collects these elastic waves by converting them into electrical signals. Its fixed sampling frequency (such as 1 MHz) can ensure that tiny internal structural change signals of the material are captured. By analyzing the acoustic emission signals, the generation and development of internal defects of the material can be understood.

[0065] Through the collaborative work of the high-resolution fiber Bragg grating sensor array, the infrared thermal imager and the acoustic emission sensor group, the multi-modal sensor can comprehensively and accurately obtain the strain distribution data, infrared thermal image data and acoustic emission signal flow during the tensile process of the material, laying a solid foundation for subsequent feature analysis and defect detection.

[0066] Example 2:

[0067] This example elaborates in detail the specific implementation process of spatial grid discretization and three-dimensional voxel unit division strategy.

[0068] When performing spatial grid discretization, first divide the strain distribution data into three-dimensional voxel units according to the geometric parameters of the material. The geometric parameters of the material include the length, width, thickness, etc. of the material. For example, for a carbon fiber composite ethylene propylene diene monomer rubber material with a length of 100 mm, a width of 50 mm, and a thickness of 10 mm, assuming it is divided into 100 units in the length direction, 50 units in the width direction, and the voxel unit size in the thickness direction is dynamically adjusted according to the number of layers. If it is set to be divided into 5 layers in the thickness direction, the size of each voxel unit in the thickness direction is 2 mm, and remains unchanged in the length and width directions. In this way, tiny three-dimensional voxel units are constructed, and the continuous strain distribution data is discretized into these units.

[0069] The Laplace operator is used to calculate the strain gradient amplitude of each voxel unit. The Laplace operator is a commonly used image processing operator, which is used to calculate the change rate of the strain distribution in the present invention. For each voxel unit, its strain gradient amplitude reflects the severity of the strain change within the unit. Assume that the strains of a certain voxel unit in the x, y, and z directions are 、 、 , respectively. The formula for calculating the strain gradient amplitude by the Laplace operator can be expressed as: , where, 、 、 represent the second-order partial derivatives of the strain in the x, y, and z directions respectively. By calculating these partial derivatives and substituting them into the formula, the strain gradient amplitude of each voxel unit can be obtained.

[0070] Differential operations are performed on adjacent voxel units to generate a strain gradient direction consistency index. The difference in the strain gradient direction between adjacent voxel units can reflect the uniformity of the internal structure of the material. Differential operations are carried out by comparing the included angle or direction cosine of the strain gradient vectors of adjacent voxel units. Assume that the strain gradient vectors of two adjacent voxel units are respectively and , and the strain gradient direction consistency index can be represented by calculating the cosine value of their included angle, that is , The closer the value is to 1, the more consistent the strain gradient directions of the two voxel units are, and the more uniform the internal structure of the material is in this area; on the contrary, the smaller the value, the greater the difference in the strain gradient direction, indicating that there may be structural changes or defects.

[0071] In terms of the division strategy of three-dimensional voxel units, the voxel unit size is dynamically adjusted according to the number of layers in the thickness direction of the material. For materials with a larger thickness, the number of layers is appropriately increased, and the size of the voxel units in the thickness direction is reduced to more accurately capture the strain changes in the thickness direction; for materials with a smaller thickness, the number of layers can be reduced, and the size of the voxel units in the thickness direction is increased to improve the calculation efficiency.

[0072] In the strain mutation region, an octree structure is used for local refinement segmentation. The strain mutation region is often the area where internal defects or stress concentration of the material appears. Using the octree structure can divide these regions more carefully. The octree structure recursively divides the spatial region into eight sub-regions. In the strain mutation region, the octree is continuously subdivided so that the voxel unit size is smaller, more accurately reflecting the strain changes in this area. For example, when it is detected that the strain change rate in a certain area exceeds the set threshold, it is determined as the strain mutation region, and the octree local refinement segmentation is started to further divide this area into smaller voxel units to obtain more accurate strain data.

[0073] Bilinear interpolation is performed on the voxel unit boundary to eliminate the jagged effect. Since when dividing voxel units, the boundary may be discontinuous or jagged, affecting the accuracy of the data and subsequent analysis. Bilinear interpolation calculates the strain value of the boundary point by performing linear interpolation on the four adjacent nodes of the voxel unit boundary, making the boundary smoother. Assume that a certain point on the voxel unit boundary is located between the four adjacent nodes , , , , and the formula for calculating the strain value of point by bilinear interpolation is: , where , , and are the strain values of four adjacent nodes A, B, C, and D respectively. u and v are the interpolation coefficients of point P in the AB and AD directions, and their values range from 0 to 1, which are determined according to the position of point P. Through bilinear interpolation, the sawtooth effect at the voxel unit boundary is effectively eliminated, and the quality of the strain data is improved.

[0074] Example 3:

[0075] This example focuses on the specific implementation details of the dynamic segmentation of the temperature field and the optimization of the region growing algorithm.

[0076] During the dynamic segmentation of the temperature field, first perform morphological opening operation on the infrared thermal image data to eliminate noise interference. The thermal image data obtained by the infrared thermal imager may contain various noises, such as electronic noise, environmental interference, etc. The morphological opening operation can remove the tiny noise points and burrs in the thermal image through the operations of erosion first and then dilation. Specifically, when operating, select a suitable structuring element, such as a circular or square structuring element, and perform erosion operation on the thermal image data, that is, compare each pixel point in the thermal image with the structuring element. If all pixel values within the area covered by the structuring element are greater than or equal to the current pixel value, then keep the current pixel value, otherwise set it to 0, so that some isolated noise points can be removed; then perform dilation operation. Contrary to the erosion operation, if there is a pixel value greater than the current pixel value within the area covered by the structuring element, then update the current pixel value to this maximum value, so as to restore the useful area that has been eroded, and at the same time further smooth the image boundary. Through the morphological opening operation, clearer and more accurate thermal image data is obtained, providing a good basis for subsequent analysis.

[0077] Identify the boundary of the heat conduction abnormal area based on the region growing algorithm. The region growing algorithm starts with seed points and gradually merges adjacent pixels according to certain growth criteria to form the heat conduction abnormal area. In this example, a dynamic growth threshold is set based on the gradient magnitude of the thermal image data. The gradient magnitude of the thermal image data reflects the severity of the temperature change. For areas with large temperature changes, they are more likely to be heat conduction abnormal areas. By calculating the gradient magnitude G pixrl of each pixel in the thermal image, set a dynamic threshold T, and T can be dynamically adjusted according to the overall statistical characteristics (such as mean, standard deviation) of the thermal image data. For example , where is the mean of the gradient magnitude of the thermal image data, is the standard deviation, and k is an empirical coefficient, usually taking values between 1 and 3. When the gradient magnitude G pixrl of a pixel is greater than T, this pixel is considered a potential point in the heat conduction abnormal area.

[0078] The anisotropic diffusion equation is used to smoothly expand the seed point region. The anisotropic diffusion equation can smooth the region while protecting the edges. Different from traditional isotropic diffusion (such as Gaussian filtering), anisotropic diffusion adjusts the diffusion coefficient according to the gradient information of the image. In the thermal image, for the edge region (i.e., the region with large temperature changes), the diffusion coefficient is small, suppressing diffusion and protecting the edges; for the flat region, the diffusion coefficient is large, performing smoothing. Its basic formula is: , where represents the thermal image data, represents time, is the gradient operator, is the diffusion coefficient function, which is a function of the gradient magnitude of the thermal image data , usually taking , , is a parameter that controls the diffusion intensity. By iteratively solving this equation, the seed point region gradually expands while maintaining the edges.

[0079] Avoid overgrowth of the region through contour curvature constraints. The contour curvature reflects the degree of bending of the region boundary. When the region grows to a certain extent, if the contour curvature is too large, it indicates that the region may overgrow into the normal region. Calculate the contour curvature K of the region boundary. When K exceeds the set threshold K max , stop the region growth. For example, for a region boundary curve composed of discrete points, approximate the contour curvature by calculating the change rate of the tangent angle between adjacent points. In this way, the boundary of the heat conduction anomaly region can be accurately identified, overgrowth of the region can be avoided, and the accuracy of heat conduction anomaly region identification can be improved.

[0080] Calculate the time-varying standard deviation of the temperature difference between the abnormal region and the background to generate a heat conduction anomaly index. After determining the heat conduction anomaly region, calculate the difference between the temperature in this region and the background temperature, and calculate the standard deviation of these temperature differences over a period of time as the heat conduction anomaly index. Assume that the temperature differences between the abnormal region and the background collected at times are respectively , then the formula for the time-varying standard deviation is: , where is the average value of the temperature differences. This index can quantitatively reflect the degree and stability of heat conduction anomalies, providing an important basis for subsequent feature fusion and defect detection.

[0081] Example 4:

[0082] This example details the specific implementation process of the adaptive weight allocation algorithm.

[0083] The adaptive weight allocation algorithm aims to reasonably integrate the strain gradient feature set, the thermal conduction anomaly index, and the frequency-domain energy features of the acoustic emission signal flow to generate a more representative enhanced feature matrix.

[0084] Model the temporal correlation of strain gradient features through a gated recurrent unit network. The gated recurrent unit (GRU) is a special recurrent neural network that can effectively handle the long-term dependence problem in time series data. During the tensile process of the material, the strain gradient features change continuously over time and have a certain temporal correlation. Input the strain gradient feature sequence into the GRU network. The reset gate r t and the update gate z t in the GRU network can adaptively control the flow of information. The reset gate r t determines how to combine the new input information with the past memory, and its calculation formula is: , where is the sigmoid function, W r is the weight matrix, x t is the strain gradient feature input at the current moment, and h t-1 is the hidden state at the previous moment. The update gate z t determines how much of the past memory to retain and how much new information to add, and the formula is: . Through the coordinated action of the reset gate and the update gate, the GRU network can learn the temporal correlation of the strain gradient features and output a feature representation containing temporal information , providing a basis for subsequent weight allocation.

[0085] Calculate the static weight of the thermal conduction anomaly index using the entropy weight method. The entropy weight method is an objective weighting method that determines weights based on the degree of variation of index data. For the thermal conduction anomaly index data, first calculate its information entropy . Assume that there are m samples and n indices for the thermal conduction anomaly index data. For the th index, its information entropy calculation formula is: , where , , is the th value of the th thermal conduction anomaly index for the th sample. The information entropy reflects the degree of dispersion of the index data. The greater the degree of dispersion, the smaller the information entropy, and the greater the weight of the index. Calculate the weight of each thermal conduction anomaly index according to the information entropy, and the formula is:

[0086] Analyze the dynamic coupling relationship between the acoustic emission frequency-domain energy characteristics and the strain gradient using the covariance matrix. The covariance matrix can measure the linear correlation between two variables. Combine the acoustic emission frequency-domain energy characteristics and the strain gradient characteristics into a vector , where S i represents the i-th component of the acoustic emission frequency-domain energy characteristics, and G j represents the j-th component of the strain gradient characteristics. Calculate the covariance matrix , and its element represents the covariance between the i-th and j-th characteristic components, and the formula is: , where is the number of samples, and are respectively the -th and -th and -th characteristic component values of the -th sample, and

[0087] are the averages of the corresponding characteristic components. By analyzing the magnitudes and signs of the corresponding elements of the acoustic emission frequency-domain energy characteristics and the strain gradient characteristics in the covariance matrix, the dynamic coupling relationship between them can be understood, that is, positive correlation, negative correlation or no correlation, and the strength of the correlation. t of the temporal correlation of the strain gradient characteristics output by the GRU network is normalized to the same order of magnitude as other weights. Assume that the normalized temporal correlation is represented as .

[0088] Combine the previously calculated static weight of the heat conduction anomaly index, and the dynamic coupling relationship between the acoustic emission frequency-domain energy characteristics and the strain gradient (reflected by the covariance matrix). Different weights are assigned to the three parts of the strain gradient characteristic set, the heat conduction anomaly index, and the frequency-domain energy characteristics of the acoustic emission signal flow according to their respective characteristics and importance.

[0089] Let the weight vector of the strain gradient characteristics be , the weight vector of the heat conduction anomaly index be , and the weight vector of the frequency-domain energy characteristics of the acoustic emission signal flow be . Determine these weight vectors through comprehensive calculations of , and the analysis results of the covariance matrix. For example, the weighted summation method can be used, and can be set as a function related to , directly use , Adjust according to the relevant elements in the covariance matrix. Specifically, let ( is an adjustment coefficient used to balance the weight size), , for , if the weight corresponding to the component with a strong positive correlation between the acoustic emission frequency domain energy feature and the strain gradient in the covariance matrix increases, and the weight corresponding to the component with a strong negative correlation decreases, after a series of calculations and adjustments, finally , , are combined into a feature fusion weight vector . This weight vector can reasonably weight and fuse each feature according to the importance and mutual relationship of different features during the tensile process of the material, so as to better highlight key information when generating the enhanced feature matrix in the dynamic fusion module and improve the accuracy of subsequent defect detection.

[0090] Example 5:

[0091] This example mainly elaborates on the construction and working process of the first deep learning model.

[0092] When constructing the first deep learning model, the enhanced feature matrix is input into the parallel three-dimensional convolutional layer and the multi-head self-attention layer. The enhanced feature matrix contains multi-faceted information such as the strain gradient feature set, heat conduction anomaly index, and frequency domain energy feature of the acoustic emission signal flow processed by the dynamic fusion module.

[0093] The three-dimensional convolutional layer consists of multiple three-dimensional convolutional kernels. The three-dimensional convolutional kernels slide in the three-dimensional space (assumed to be the length, width, and height directions) of the enhanced feature matrix for convolutional operations. Each three-dimensional convolutional kernel has its own weight parameters. By learning these weight parameters, the convolutional kernel can extract local spatial features in the enhanced feature matrix. For example, a three-dimensional convolutional kernel of size , when traversing the enhanced feature matrix, will perform a weighted sum of the feature values within the area it covers, and add a bias term to obtain a new feature value, thereby generating a new feature map. By using multiple three-dimensional convolutional kernels with different weights, different types of local spatial features can be extracted, such as local strain concentration features near micro-defects inside the material and local temperature change features in the heat conduction anomaly area.

[0094] The multi-head self-attention layer is used to model cross-channel global dependencies. In the enhanced feature matrix, there may be correlations between features in different channels. The multi-head self-attention mechanism maps the input features to multiple low-dimensional subspaces, calculates the attention weights separately in each subspace, and then concatenates the results. The specific process is as follows: First, the enhanced feature matrix Through three linear transformations respectively , , the query vector , the key vector , and the value vector are obtained. Then, the attention scores are calculated in each subspace, where and are the -th row and the -th row of the query vector and the key vector respectively, is the dimension of the key vector, is the number of key vectors. The attention score represents the degree of association between the query vector and the key vector . Then, the value vectors are weighted and summed according to the attention scores to obtain the output of each subspace ( represents different subspaces). Finally, the outputs of multiple subspaces are concatenated and passed through a linear transformation to obtain the final output of the multi-head self-attention layer. In this way, the multi-head self-attention layer can capture the global dependencies across channels in the enhanced feature matrix, such as the co-variation relationship between different types of features.

[0095] The channel attention gating mechanism is used to fuse local and global features. The channel attention gating mechanism weights the channels according to the importance of the features. First, the local spatial features output by the 3D convolutional layer and the global dependency features output by the multi-head self-attention layer are fused to obtain the fused feature F. Then, the fused feature F is compressed in the spatial dimension through global average pooling operation to obtain the channel description vector . Next, the channel description vector is passed through two fully connected layers. The first fully connected layer has fewer neurons and plays a role in dimensionality reduction, and the second fully connected layer has the same number of neurons as the number of channels and plays a role in dimensionality increase to obtain the attention weight vector . Finally, the attention weight vector is multiplied element-wise with the fused feature to obtain the feature processed by the channel attention gating mechanism, which is the finally output defect probability distribution map. In this way, the model can highlight the features more valuable for defect detection, suppress irrelevant features, and thus more accurately output the defect probability distribution map inside the material.

[0096] Example 6:

[0097] This embodiment focuses on the triggering logic and specific operations of the real-time feedback module.

[0098] The real-time feedback module in the entire detection system is responsible for generating the tensile state evaluation results based on the defect probability distribution map and triggering the hierarchical early warning instructions to ensure the safety and effectiveness of the detection process.

[0099] After the defect probability distribution map is generated, the real-time feedback module starts to analyze it. If there are continuous regions in the defect probability distribution map where the probability values exceed the first threshold, a first-level early warning is triggered and the stretching rate is reduced at this time. The first threshold is preset based on a large amount of experimental data and practical experience. For example, the first threshold is set to 0.5. When the system detects that there is a continuous region in the defect probability distribution map where the defect probability value corresponding to each pixel point is greater than 0.5, it is determined that the first-level early warning condition is met. When the first-level early warning is triggered, the system will send an instruction through the controller that controls the stretching device to reduce the stretching rate. The reduction of the stretching rate can reduce the deformation speed of the material, slow down the development of defects, and gain time for further detection and analysis. For example, the original stretching rate is 1 mm per second, and after the first-level early warning is triggered, the stretching rate is reduced to 0.5 mm per second. At the same time, the system will record the first-level early warning information, including the warning time, warning location (i.e., the location of the continuous region with high defect probability in the material), and other information.

[0100] If the probability value exceeds the second threshold and the area of the region expands, a second-level early warning is triggered and the stretching experiment is terminated. The second threshold is usually greater than the first threshold, such as set to 0.8. When there is a region with a probability value greater than 0.8 in the defect probability distribution map and the area of this region has significantly expanded compared with before, the system triggers a second-level early warning. After the second-level early warning is triggered, the system immediately sends a stop instruction to the stretching device to stop the stretching experiment, avoiding further damage to the material and preventing more serious safety problems. At the same time, the system synchronizes the second-level early warning information and detailed defect distribution data to the distributed monitoring platform. The defect distribution data includes the probability value of each pixel point in the defect probability distribution map, the shape, size, location, and other information of the defect region. The distributed monitoring platform can be a network-based multi-terminal data sharing platform. Relevant staff can log in to the platform through terminal devices such as computers and mobile phones to view this information, timely understand the abnormal situation during the tensile process of the material, and make subsequent processing decisions, such as repairing or replacing the material, etc.

[0101] Through such a hierarchical early warning mechanism, the real-time feedback module can respond to the defect situation during the tensile process of the material in a timely and effective manner, ensure the safety and reliability of the detection process, and at the same time provide important data support for subsequent material research and quality control.

[0102] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0103] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A tensile testing system for a carbon fiber composite EPDM rubber material, characterized in that, The system includes: A real-time acquisition module, which is used to synchronously obtain strain distribution data, infrared thermal image data, and acoustic emission signal streams during the tensile process of the material through multimodal sensors; A multi-dimensional feature analysis module, which performs spatial grid discretization processing on the strain distribution data to generate a strain gradient feature set, and at the same time performs dynamic segmentation of the temperature field on the infrared thermal image data to extract heat conduction anomaly indicators; A dynamic fusion module, which based on an adaptive weight allocation algorithm, non-linearly fuses the strain gradient feature set, heat conduction anomaly indicators, and frequency domain energy features of the acoustic emission signal stream to generate an enhanced feature matrix; A defect detection module, which inputs the enhanced feature matrix into a pre-trained first deep learning model. The model is based on a parallel architecture of a three-dimensional convolutional neural network and an attention mechanism, and outputs a probability distribution map of internal defects of the material; A real-time feedback module, which generates a tensile state evaluation result according to the defect probability distribution map and triggers a hierarchical warning instruction; The steps of the spatial grid discretization processing include: Dividing the strain distribution data into three-dimensional voxel units according to the geometric parameters of the material; calculating the strain gradient amplitude of each voxel unit using the Laplace operator; Performing a difference operation on adjacent voxel units to generate a strain gradient direction consistency index; The steps of the temperature field dynamic segmentation include: Performing a morphological opening operation on the infrared thermal image data to eliminate noise interference; Identifying the boundary of the heat conduction anomaly region based on the region growing algorithm; Calculating the time-varying standard deviation of the temperature difference between the anomaly region and the background to generate a heat conduction anomaly indicator; The implementation of the adaptive weight allocation algorithm includes: Modeling the temporal correlation of strain gradient features through a gated recurrent unit network; Calculating the static weight of the heat conduction anomaly indicator using the entropy weight method; Analyzing the dynamic coupling relationship between the acoustic emission frequency domain energy feature and the strain gradient using the covariance matrix; Fusing the temporal correlation, static weight, and dynamic coupling relationship to generate a feature fusion weight vector.

2. The tensile detection system according to claim 1, characterized in that, The multimodal sensors include: A high-resolution fiber Bragg grating sensor array, which is embedded on the surface of the material according to a preset topological structure for collecting strain distribution data; An infrared thermal imager and an acoustic emission sensor group, which respectively obtain thermal image data and acoustic emission signals synchronously at a fixed sampling frequency.

3. The tensile detection system according to claim 1, characterized in that The construction steps of the first deep learning model include: Inputting the enhanced feature matrix into a parallel three-dimensional convolutional layer and a multi-head self-attention layer; Extracting local spatial features through a three-dimensional convolutional kernel and modeling cross-channel global dependencies through the attention mechanism; Fusing local and global features using a channel attention gating mechanism to output a defect probability distribution map.

4. The tensile detection system according to claim 1, wherein The triggering logic of the real-time feedback module includes: If the probability value of a continuous region in the defect probability distribution map exceeds the first threshold, start a first-level warning and reduce the tensile rate; If the probability value exceeds the second threshold and the area of the region expands, trigger a second-level warning and terminate the tensile experiment; Synchronize the warning instruction and the defect distribution data to the distributed monitoring platform.

5. The tensile detection system according to claim 1, characterized in that The division strategy of the three-dimensional voxel units includes: Dynamically adjusting the voxel unit size according to the number of layers in the thickness direction of the material; Using an octree structure for local refinement segmentation in the strain mutation region; Perform bilinear interpolation on the voxel unit boundary to eliminate the aliasing effect.

6. The tensile detection system according to claim 1, characterized in that, The optimization steps of the region growing algorithm include: setting a dynamic growth threshold based on the gradient magnitude of the thermal image data, using an anisotropic diffusion equation to smoothly expand the seed point region, and avoiding overgrowth of the region through contour curvature constraints.

7. A tensile testing method for a carbon fiber composite EPDM rubber material, characterized in that, The method includes: Perform real-time detection and warning of the tensile state of the material through the system described in any one of claims 1-6.

Images