Quantitative identification method for ultrahigh-speed impact damage of UHMWPE (Ultra High Molecular Weight Polyethylene) fiber reinforced material
Through deep learning model and variational autoencoder technology, combined with ultrasonic detection signals and appearance data, a three-dimensional damage model is generated, which solves the problems of high cost and low accuracy in ultra-high-speed impact damage detection of UHMWPE fiber reinforced materials, and achieves high accuracy, Low-cost quantitative identification of damage.
Patent Information
- Application Number
- CN202411040105.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-19
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art has problems of high cost, low accuracy or general lack of generality in ultra-high-speed impact damage detection of UHMWPE fiber reinforced materials, resulting in inaccurate detection results of quantitative damage degree or high cost.
Using deep learning model, combining ultrasonic detection signals and appearance data, through variational autoencoder and convolutional neural network architecture, damage section images and three-dimensional models are generated to quantitatively identify damage, replacing high-precision and high-cost equipment, and using low-precision and low-precision Cost detection method obtains data.
It realizes high-precision quantitative damage recognition, reduces detection costs, improves the accuracy of detection results, can more accurately characterize internal damage, and reduces the use of high-cost equipment.
Smart Images

Figure CN118980750A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of material detection, and in particular relates to a quantitative identification method for ultrahigh-speed impact damage of a UHMWPE fiber-reinforced material. Background Art
[0002] UHMWPE (ultra-high molecular weight polyethylene) fiber-reinforced materials are widely used in aerospace, military protection and other fields due to their excellent properties, such as high impact resistance, high wear resistance and low friction coefficient. However, under ultra-high-speed impact, the material may also be damaged, posing a threat to the integrity and safety of the structure. Therefore, the development of an effective ultra-high-speed impact damage quantitative identification method is crucial to ensure the safety of UHMWPE fiber-reinforced material structures. However, in damage identification, internal defects can be detected by ultrasonic testing, X-ray testing, magnetic particle testing, industrial CT, acoustic emission testing and other methods, but the above methods all have the disadvantages of insufficient universality, high equipment cost or low detection accuracy, resulting in the final quantitative damage degree detection results being high in cost or insufficient in accuracy. Therefore, there is an urgent need for a damage degree identification method that can reduce costs and improve damage degree detection results. Summary of the invention
[0003] In order to solve the above technical problems, the present invention proposes a quantitative identification method for ultra-high velocity impact damage of UHMWPE fiber-reinforced materials to solve the above problems in the prior art.
[0004] To achieve the above object, the present invention provides a method for quantitatively identifying ultrahigh-speed impact damage of UHMWPE fiber-reinforced materials, comprising:
[0005] Conduct ultra-high-speed impact tests on UHMWPE fiber-reinforced materials and collect ultrasonic detection signals and appearance data after the test;
[0006] Constructing and training a deep learning model, wherein the output of the deep learning model is a section matrix of a damaged UHMWPE fiber-reinforced material;
[0007] The section matrix iteratively generated by the trained deep learning model is obtained, the section matrix is judged and screened according to the ultrasonic detection signal to generate a damage section image, a damage three-dimensional model is generated according to the damage section matrix, the damage three-dimensional model is quantitatively identified, and the damage degree of the UHMWPE fiber reinforced material is generated.
[0008] Optionally, the process for ultra-high velocity impact testing of UHMWPE fiber-reinforced materials includes:
[0009] The UHMWPE fiber reinforced material is visually inspected and placed still after the inspection. The UHMWPE fiber reinforced material is fixed in an ultra-high speed impact testing machine for an ultra-high speed impact test. After the test, an ultrasonic detection signal and appearance data are obtained through an ultrasonic detection signal and a 3D scanner.
[0010] Optionally, the deep learning model adopts a variational autoencoder, wherein the variational autoencoder adopts a convolutional neural network architecture.
[0011] Optionally, the variational autoencoder includes an encoder and a decoder, wherein the output of the decoder is a section matrix of the damaged UHMWPE fiber-reinforced material, and the input is a latent parameter in a latent space.
[0012] Optionally, the process of training the deep learning model includes:
[0013] Obtain historical detection data of UHMWPE fiber-reinforced materials using X-ray detection and industrial CT detection methods, generate a corresponding historical damage three-dimensional model based on the historical detection data, slice the historical damage three-dimensional model to generate a training set; train the deep learning model using the training set to obtain a trained deep learning model.
[0014] Optionally, the process of acquiring the injury section image includes:
[0015] Iteratively adjusting the latent parameters, processing the latent parameters in the iterative process through a trained decoder, and generating a section matrix in the iterative process;
[0016] generating an initial three-dimensional damage model according to the ultrasonic detection signal, and slicing the initial three-dimensional damage model to generate an initial section image;
[0017] The section matrix in the iterative process is converted into an iterative section image, and the iterative section image is judged and screened according to the initial section image to obtain a damaged section image.
[0018] Optionally, the process of quantitatively identifying the damage three-dimensional model includes:
[0019] The surface damage area is obtained according to the appearance data statistics, and the internal damage area is obtained according to the damage three-dimensional model statistics. The proportion of the surface damage area and the internal damage area to the UHMWPE fiber reinforced material is weighted and calculated to obtain the damage degree of the UHMWPE fiber reinforced material.
[0020] Optionally, a damage cross-section image is generated by an optimization algorithm, wherein an objective function, fitness and population update method are set based on hidden parameters as the initial population, wherein the objective function is to minimize the sum of the difference between the cross-section image generated by the decoder and each pixel in the corresponding feature area of the ultrasound cross-section image, and the fitness is the inverse of the objective function; based on the objective function, fitness and population update method, the optimization algorithm is used to search for the best result until the maximum number of iterations is reached or the objective function meets certain requirements, thereby obtaining a damage boundary cross-section image.
[0021] Optionally, the process of generating the initial cross-sectional image further includes replacing the ultrasonic detection method with acoustic emission detection or magnetic particle detection, generating a detection signal according to the replaced method, and generating the cross-sectional image according to the detection signal.
[0022] Compared with the prior art, the present invention has the following advantages and technical effects:
[0023] Through the above technical scheme, the present invention trains the deep learning model through high-precision training data, and its deep learning model can generate data with the same precision as the training data, and at the same time obtains the detection data through a low-precision detection method, selects the high-precision data generated by the deep learning model similar to the detection data, constructs a three-dimensional damage model according to the high-precision data, and performs quantitative damage calculation on the three-dimensional damage model to obtain the final quantitative damage identification data. The present invention replaces the high-cost and high-precision equipment with a deep learning model, and detects the detection data of the current UHMWPE fiber reinforced material specimens through a low-precision detection method. The deep learning model converts the low-precision detection data into high-precision detection data that can characterize small or detailed damage. The three-dimensional damage model established in this way is more accurate for internal damage performance and more accurate for the calculation of the final quantitative damage identification result. At the same time, only low-cost equipment can be used to complete detection and identification, reducing the use of high-cost equipment, and has strong practicality. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0025] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION
[0026] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0027] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0028] like Figure 1 As shown, in this embodiment, a method for quantitatively identifying ultra-high-speed impact damage of UHMWPE fiber-reinforced materials is provided. A deep learning model is constructed by using equipment with high accuracy but low equipment cost or low universality, and the detection features with low detection accuracy but low cost and high universality are used to map a more accurate three-dimensional damage model through the learning model, and quantitative identification of damage is performed through the three-dimensional damage model.
[0029] The detection methods of internal damage of UHMWPE fiber-reinforced materials have their own advantages and disadvantages, which are summarized as follows: Ultrasonic detection: Principle: Use the penetration performance of ultrasound to determine the internal structure, defects and performance of objects. Advantages: High efficiency, accuracy, low cost and high reliability. Disadvantages: The application range is relatively narrow and may not be suitable for some materials with complex structures or large curvatures. Scope of application: Widely applicable to the internal defect detection of fiber-reinforced materials.
[0030] X-ray detection: Principle: Detect the internal structure and defects of materials through the different absorption degrees of X-rays by the materials. Advantages: large penetration depth, wide detection range, strong applicability to different materials. Disadvantages: high equipment cost, difficult operation, and the presence of harmful radioactive substances. Scope of application: Applicable to the detection of hidden defects inside fiber-reinforced materials.
[0031] Magnetic particle testing: Principle: Use magnetic field to perform nondestructive testing on materials, and judge whether the material has defects by observing the distribution and color change of magnetic particles. Advantages: Easy operation, low cost, and little damage to materials. Disadvantages: May be affected by the complex internal structure of the material. Scope of application: Applicable to surface and near-surface defect detection.
[0032] Industrial CT (Computed Tomography) detection method: Principle: Non-destructive testing technology that reconstructs the internal structure image of the tested object based on projection data. Advantages: Non-contact, non-destructive, high resolution, high reliability, and can accurately provide three-dimensional position data of the internal details of the tested object. Disadvantages: High equipment cost. Scope of application: Applicable to accurate detection of complex internal structures of fiber-reinforced materials.
[0033] Acoustic emission detection: Principle: When a material is deformed or fractured under stress, the energy generated is released in the form of stress waves, and the damage is judged by detecting stress waves. Advantages: Real-time monitoring is possible without causing additional damage to the material. Disadvantages: It may be affected by the complex internal structure of the material and has high requirements for equipment. Scope of application: Applicable to monitoring the damage of fiber-reinforced materials under dynamic loads.
[0034] Among the above detection methods, the X-ray detection or industrial CT detection method has high accuracy and can accurately give internal details or tiny defects, but the above equipment does not have the advantages of universality and low cost. Ultrasonic detection, magnetic particle detection, and acoustic emission detection can detect internal defects, but the response to some complex structures and tiny details is not complete. Therefore, the present invention hopes to reflect the above-mentioned high-precision detection results through a deep learning model, and at the same time select the corresponding internal defect structure through the characteristic performance reflected by the detection method with lower accuracy, and construct a three-dimensional damage model based on the screened internal defect structure, and determine the degree of damage of the final UHMWPE fiber reinforced material according to the range and size of the internal damage and the range and size of the appearance damage.
[0035] Detailed explanation of the above:
[0036] In the present embodiment, a quantitative identification method for ultra-high-speed impact damage of UHMWPE fiber-reinforced materials is provided. Taking a low-precision ultrasonic detection method as a basic detection method as an example, an ultra-high-speed impact damage test is firstly performed on the UHMWPE fiber-reinforced material. When the UHMWPE fiber-reinforced material is damaged, the ultrasonic detection signal during the test and the appearance data after the test are collected; a decoder is constructed and trained by a damage data set measured by the above-mentioned high-precision detection method, wherein the input of the decoder is a hidden parameter, and the output is a section matrix of the damaged UHMWPE fiber-reinforced material; the hidden parameters are iteratively adjusted, and according to the section matrix corresponding to the hidden parameters in the iterative adjustment process, the section matrix is judged and screened according to the ultrasonic detection signal and the appearance data to generate a damage section matrix, and a damage three-dimensional model is generated according to the damage section matrix, and the damage three-dimensional model is quantitatively identified to generate the damage degree of the UHMWPE fiber-reinforced material.
[0037] As some embodiments, the present invention uses an ultra-high speed impact testing machine for testing. Before the test, a UHMWPE fiber-reinforced material sample that meets the standards is prepared according to the test requirements. The sample should undergo necessary appearance inspection to ensure that there are no defects and parts that do not meet the requirements. The sample should be placed under standard environmental conditions for at least 24 hours before the test to ensure its stable performance. The sample is fixed in an appropriate position on the testing machine to ensure that the sample is aligned with the impact object. The parameters of the testing machine are adjusted to meet the test requirements. An impact test is performed, and the sample after the impact is ultrasonically inspected by an ultrasonic detection device, and the appearance data, i.e., surface data, is scanned by a 3D scanner. Ultrasonic data and appearance data after the test are collected, wherein the appearance data includes the damage range and size of the surface of the sample after the UHMWPE fiber-reinforced material test;
[0038] As some embodiments, the deep learning model used in this embodiment is a variational autoencoder, wherein the variational autoencoder adopts an encoder-decoder structure. At the same time, in order to facilitate the use of the deep learning model, a two-dimensional deep learning model is adopted to convert the transverse section of the three-dimensional model of UHMWPE fiber reinforced material obtained by the high-precision detection method into a matrix format, and the section matrix is used as the input of the encoder. There is a latent space between the encoder and the decoder, wherein the latent space contains latent variables, and the encoder outputs latent variables and latent variable distribution parameters. At the same time, the latent variables are used as the input of the decoder so that the decoder can output the corresponding section matrix, wherein the encoder characterizes the mapping relationship between the section matrix and the latent variables, and the section matrix can be mapped to the latent variables. Through training, the decoder learns this mapping relationship, and when used alone, different section matrices can be generated through different latent variables, wherein the potential distribution of the latent variables in the latent space is common knowledge in the field, and the structure of the variational autoencoder can be set through the conventional structure in the field. At the same time, the variational autoencoder can also be used for the generation scheme of 3D models, and the generated section data needs to be replaced with model data, which will not be repeated here.
[0039] As some embodiments, the variational autoencoder used in the present invention adopts a convolutional neural network architecture, and the structure of the encoder includes: four sequentially connected convolution modules, Flatten layer, Linear layer, LReLU layer and Linear layer, wherein the convolution module includes a sequentially connected Conv layer, BN layer and LReLU layer, and the decoding structure includes an Unflatten layer, a deconvolution Deconv layer, a BN layer, an LReLU layer, two sequentially connected upsampling modules, a convolution module, a convolution layer and an output layer, wherein the upsampling module includes a sequentially connected Upsample layer, a Conv layer, a BN layer and an LReLU layer. Among the above contents, the Conv layer is a two-dimensional convolution layer; the Deconv layer is a two-dimensional deconvolution layer; the Linear layer is a fully connected layer; the BN layer is a two-dimensional batch normalization layer; the LReLU layer indicates the use of the LeakyReLU activation function; Upsample indicates a two-dimensional bilinear interpolation upsampling operation; Flatten indicates expanding the tensor format data into a vector, Unflatten is the reverse process of the above operation, and the output layer uses the tanh function. The above model parameters can be determined according to actual needs and will not be described in detail here.
[0040] As some embodiments, a high-precision detection method is obtained through relevant databases or other laboratory data, such as the original information or three-dimensional model of UHMWPE fiber-reinforced materials obtained by X-ray detection, and the original information is converted into a three-dimensional model, and the three-dimensional model is processed with equal intervals to generate a section image, and a corresponding section matrix is generated according to the section image, wherein each element of the section matrix represents a pixel in the section image or the internal damage shown, and the content is used as a training set to train a variational autoencoder, i.e., a deep learning model, and the deep learning model is trained using an Adam optimizer. The loss function of the VAE model training is expressed as follows:
[0041] min L=λ 1 NMSE(x)+λ 2 KL(x)
[0042]
[0043] Where NMSE is the normalized root mean square error, which is used to measure the difference between the slice matrix generated by the decoder, i.e., the reconstructed sample x, and the original slice matrix, i.e., the original sample x. 0 The gap between them; KL represents the Kullback-Leibler divergence, which is used as a regular term in the loss function to ensure the generation ability of the VAE model. 1 and 2 represents the corresponding item weight coefficient, μ and σ represent the mean and standard deviation of each component of the latent variable in the latent space, Represents the expected function of the encoder input data operation result, represents the expected function of the encoder output data operation result. For the present invention, λ is usually taken 1 =2000~40000,λ 2 =1.
[0044] As some embodiments, after training, the deep learning model uses a decoder to generate a section matrix, wherein hidden variables are randomly adjusted and selected in the latent space, and the corresponding section matrix is generated by the decoder using the hidden variables, and the corresponding section image is generated by the section matrix. By measuring the feature gap between the section image generated by the decoder and the section image acquired by ultrasound, the hidden variables are optimized and selected with the goal of minimizing the gap, and the final section image is generated as the regenerated section image. In the above content, the decoder can store the relationship between the hidden variables and the section image acquired by the high-precision detection method, and the section image finally generated by the decoder can generate corresponding minor damage like the high-precision detection method. The low-precision detection method cannot detect minor or detailed damage, but its larger damage can be detected like the high-precision detection.
[0045] Specifically, analysis is performed on ultrasonic data to generate a three-dimensional damage model of the UHMWPE fiber-reinforced material obtained by ultrasonic testing, and the three-dimensional damage model is sliced to generate an ultrasonic section image. In the ultrasonic section image, areas with obvious damage features are screened out through manual annotation or deep recognition models, such as convolutional neural networks, and the pixels in the area are used as a basis for comparison with the pixels in the corresponding area of the section image generated by the decoder, with minimizing the difference as an optimization goal.
[0046] As some embodiments, the hidden parameters are used as initial parameters, and iterative adjustments are made, and the corresponding slice matrix is generated by the decoder, and the slice matrix is converted into a slice image, and the difference between the slice image generated by the decoder and the corresponding feature area of the ultrasound slice image is measured until the maximum number of iterations is reached or the difference is within the required range, and then the iteration is stopped to generate the final slice image as the regenerated slice image. Different slice images are integrated and assembled, and the damaged area is smoothed by a smoothing method to generate a corresponding three-dimensional model.
[0047] Specifically, the hidden parameters are iteratively adjusted through an optimization algorithm, wherein the optimization algorithm may use a taboo algorithm, a genetic algorithm, or a moth-to-flame algorithm, etc.
[0048] It mainly includes: according to the hidden parameters set as the initial population, construct and initialize the population, wherein the population includes a number of individuals, wherein the individual is a set of hidden parameters; setting the objective function, fitness and population update method, wherein the objective function is to minimize the sum of the difference between the section image generated by the decoder and each pixel of the corresponding feature area of the ultrasound section image, and the fitness is the inverse of the objective function; according to the objective function, fitness and population update method, the optimization algorithm is used to search for the best until the maximum number of iterations is reached or the objective function meets certain requirements, and the damage boundary section image is obtained. The specific implementation content of the optimization algorithm population update method is well known to those skilled in the art and will not be repeated here.
[0049] As some embodiments, after the final three-dimensional damage model is constructed, the surface damage and internal damage in the three-dimensional damage model are counted, and then the proportion of the surface damage range to the complete UHMWPE fiber reinforced material specimen and the proportion of the internal damage range to the complete UHMWPE fiber reinforced material specimen are calculated. The weighted sum of the surface damage ratio and the internal damage ratio is used to obtain the degree of damage of the UHMWPE fiber reinforced material specimen, that is, the final quantitative identification data, wherein the weights in the weighting are determined according to manual experience.
[0050] Through the above technical scheme, the present invention trains the deep learning model through high-precision training data, and its deep learning model can generate data with the same precision as the training data, and at the same time obtains the detection data through a low-precision detection method, selects the high-precision data generated by the deep learning model similar to the detection data, constructs a three-dimensional damage model according to the high-precision data, and performs quantitative damage calculation on the three-dimensional damage model to obtain the final quantitative damage identification data. The present invention replaces the high-cost and high-precision equipment with a deep learning model, and detects the detection data of the current UHMWPE fiber reinforced material specimens through a low-precision detection method. The deep learning model converts the low-precision detection data into high-precision detection data that can characterize small or detailed damage. The three-dimensional damage model established in this way is more accurate for internal damage performance and more accurate for the calculation of the final quantitative damage identification result. At the same time, only low-cost equipment can be used to complete detection and identification, reducing the use of high-cost equipment, and has strong practicality.
[0051] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A quantitative identification method for ultrahigh-velocity impact damage of UHMWPE fiber-reinforced materials. It is characterized in that include: Conduct ultra-high-speed impact tests on UHMWPE fiber-reinforced materials and collect ultrasonic detection signals and appearance data after the test; Constructing and training a deep learning model, wherein the output of the deep learning model is a section matrix of a damaged UHMWPE fiber-reinforced material; The section matrix iteratively generated by the trained deep learning model is obtained, the section matrix is judged and screened according to the ultrasonic detection signal to generate a damage section image, a damage three-dimensional model is generated according to the damage section matrix, the damage three-dimensional model is quantitatively identified, and the damage degree of the UHMWPE fiber reinforced material is generated.
2. The method according to claim 1, It is characterized in that The process of ultra-high velocity impact testing of UHMWPE fiber-reinforced materials includes: The UHMWPE fiber reinforced material is visually inspected and placed still after the inspection. The UHMWPE fiber reinforced material is fixed in an ultra-high speed impact testing machine for an ultra-high speed impact test. After the test, an ultrasonic detection signal and appearance data are obtained through an ultrasonic detection signal and a 3D scanner.
3. The method according to claim 1, It is characterized in that The deep learning model adopts a variational autoencoder, wherein the variational autoencoder adopts a convolutional neural network architecture.
4. The method according to claim 3, It is characterized in that The variational autoencoder includes an encoder and a decoder, wherein the output of the decoder is a section matrix of a damaged UHMWPE fiber-reinforced material, and the input is a latent parameter in a latent space.
5. The method according to claim 1, It is characterized in that The process of training the deep learning model includes: Obtain historical detection data of UHMWPE fiber-reinforced materials using X-ray detection and industrial CT detection methods, generate a corresponding historical damage three-dimensional model based on the historical detection data, slice the historical damage three-dimensional model to generate a training set; train the deep learning model using the training set to obtain a trained deep learning model.
6. The method according to claim 4, It is characterized in that The process of acquiring the injury section image includes: Iteratively adjusting the latent parameters, processing the latent parameters in the iterative process through a trained decoder, and generating a section matrix in the iterative process; generating an initial three-dimensional damage model according to the ultrasonic detection signal, and slicing the initial three-dimensional damage model to generate an initial section image; The section matrix in the iterative process is converted into an iterative section image, and the iterative section image is judged and screened according to the initial section image to obtain a damaged section image.
7. The method according to claim 1, It is characterized in that The process of quantitatively identifying the three-dimensional damage model includes: The surface damage area is obtained according to the appearance data statistics, and the internal damage area is obtained according to the damage three-dimensional model statistics. The proportion of the surface damage area and the internal damage area to the UHMWPE fiber reinforced material is weighted and calculated to obtain the damage degree of the UHMWPE fiber reinforced material.
8. The method according to claim 1, It is characterized in that The damage section image is generated by an optimization algorithm, wherein the hidden parameters are used as the initial population, and the objective function, fitness and population update method are set, wherein the objective function is to minimize the sum of the difference between the section image generated by the decoder and each pixel of the corresponding feature area of the ultrasound section image, and the fitness is the inverse of the objective function; according to the objective function, fitness and population update method, the optimization algorithm is used to search for the best result until the maximum number of iterations is reached or the objective function meets certain requirements, thereby obtaining the damage boundary section image.
9. The method according to claim 6, It is characterized in that The process of generating the initial cross-sectional image also includes replacing the ultrasonic detection method with acoustic emission detection or magnetic particle detection, generating a detection signal according to the replaced method, and generating the cross-sectional image according to the detection signal.
