Damage Location Method and Device for Carbon Fiber Composite Materials Based on Convolutional Neural Network and Single-Channel Carbon Nanotube Sensors

Abstract

This application discloses a method and device for locating damage in carbon fiber composite materials based on a convolutional neural network and a single-channel carbon nanotube sensor. The method includes: preparing a single-channel carbon nanotube sensor; preparing a carbon fiber composite laminate for damage location data acquisition; equally dividing the measured area according to a coordinate system to form a plurality of grid coordinates; establishing a convolutional neural network model; using the acquired data labels as input and output data sets, and randomly dividing the input and output data sets into a training set and a test set for the convolutional neural network model; using the training set and the test set to train and verify the convolutional neural network model until the convolutional neural network model reaches convergence; and using the convolutional neural network model to locate structural damage. The technical solution provided by this application avoids the data complexity problem caused by multi-channel data acquisition and realizes end-to-end damage location of carbon fiber reinforced composite materials.

Description

Technical Field

[0001] The present invention relates to the field of structural damage location, and in particular to a damage location method and device for carbon fiber composite materials based on a convolutional neural network and a single-channel carbon nanotube sensor. Background Art

[0002] With the development trend of global carbon neutrality and energy efficiency policies, the market share of CFRP (Carbon Fiber-reinforced Polymer) has continued to increase in recent years. CFRP structures have the characteristics of high stiffness and light weight, and are increasingly used in rail transportation, aerospace, wind turbine blades and other fields. However, large composite structures are affected by loads, environment and other factors during service, which can easily cause structural failure and lead to safety accidents. Therefore, there is an increasing need for structural health monitoring technology (SHM) to monitor the health status of composite structures.

[0003] Traditional damage localization methods include ultrasonic scanning, radiographic testing, and thermal imaging. However, these methods are difficult to implement for real-time inspections and are limited by the inspection cycle. Real-time monitoring technologies such as fiber Bragg grating (FBG), piezoelectric ceramics, and acoustic emission (AE) overcome the shortcomings of offline monitoring techniques and have been widely studied for CFRP damage localization. Researchers generally employ multi-channel data acquisition schemes to conduct CFRP damage localization research. For example, Sikdar et al. constructed a monitoring network using eight AE sensors and used a convolutional neural network model to classify CFRP damage sources based on the AE data. Jung et al. introduced four piezoelectric films between CFRP layers. They applied discrete wavelet transforms to the impact signals collected during impact tests, converting them into image data for training and testing a deep learning model. The trained model was then used to predict the impact location. Alsaadi et al. solidified a polyimide film with 16 electrode arrays onto the surface of a CFRP structure. During an impact test, they simultaneously collected resistance data at 16 locations to generate a 2D damage contour map for damage location prediction. The above damage localization methods all have the following problems to varying degrees: multi-channel data acquisition increases data dimensionality, making data processing more difficult and computationally intensive; serpentine CNT sensors with concentration gradients can achieve damage localization in CFRP structures using single-channel data acquisition. However, the monitoring data typically contains noise and nonlinear relationships, making it difficult to process using traditional mathematical models. Summary of the Invention

[0004] The present invention provides a method and device for locating damage in carbon fiber composite materials based on a convolutional neural network and a single-channel carbon nanotube sensor. CFRP damage localization can be achieved through single-channel data acquisition only, avoiding the data complexity problem caused by multi-channel data acquisition. In addition, end-to-end CFRP damage localization is achieved by combining with a convolutional neural network.

[0005] According to a first aspect of the present invention, a method for localizing damage in a carbon fiber composite material using a convolutional neural network and a single-channel carbon nanotube sensor is provided, comprising the following steps:

[0006] The carbon nanotube solution is sprayed onto the template according to a preset concentration gradient to prepare a single-channel carbon nanotube sensor;

[0007] The prepared single-channel carbon nanotube sensor is introduced into the middle interface of the carbon fiber composite material to prepare a carbon fiber composite laminate for damage location data acquisition;

[0008] Selecting a tested area of the carbon fiber composite material laminate, establishing a coordinate system in the tested area, and equally dividing the tested area according to the coordinate system to form a plurality of grid coordinates;

[0009] Build a convolutional neural network model;

[0010] sequentially loading at least some of the plurality of grid coordinates to obtain single-channel carbon nanotube sensor signals, labeling them according to the loading position coordinates, using the obtained data labels as input and output data sets, and randomly dividing the input and output data sets into a training set and a test set for a convolutional neural network model;

[0011] The convolutional neural network model is trained using the training set, and the trained convolutional neural network model is verified using the test set until the convolutional neural network model reaches convergence;

[0012] The converged convolutional neural network model is used to locate structural damage and output the damage coordinates.

[0013] In addition to or as an alternative to one or more of the features disclosed above, the concentration of the carbon nanotube solution is 2-2.5% by gravity.

[0014] In addition to or as an alternative to one or more features disclosed above, the carbon nanotube solution is sprayed into the template according to a preset concentration gradient, and the carbon nanotube solution is sprayed into at least one of a bent shape, a winding shape, an S shape, a bow shape and a W shape.

[0015] In addition to or as an alternative to one or more of the features disclosed above, spraying the carbon nanotube solution into the template according to a predetermined concentration gradient includes:

[0016] Set the unit area of concentration gradient change to 80-120 mm 2 In each spraying process, a unit area is added from one end of the template along the template path, and this is repeated 3 to 5 times until the carbon nanotubes completely cover the template surface.

[0017] In addition to or as an alternative to one or more of the features disclosed above, the template is made of one of copper, aluminum, stainless steel, plastic and a composite material.

[0018] In addition to one or more of the features disclosed above, or as an alternative, after spraying, the template is immersed in an etching solution and etched for 6 to 10 hours, and finally washed and dried to obtain the single-channel carbon nanotube sensor with a concentration gradient.

[0019] In addition to one or more of the features disclosed above, or as an alternative, the ply of the carbon fiber composite laminate is [0 / 90 / 0 / 90 / 0 / CNT 1 / 2 ] S , where s indicates that the layer design is symmetrical; CNT indicates a single-channel carbon nanotube sensor, CNT 1 / 2 It indicates that a single-channel carbon nanotube sensor is used as the middle layer of the carbon fiber composite material laminate; and insulating layers are laid on the upper and lower surfaces of the single-channel carbon nanotube sensor.

[0020] In addition to or instead of one or more of the features disclosed above, during the process of sequentially loading at least a portion of the plurality of grid coordinates, the loading rate of the indenter is 0.5 to 1.5 mm / min, and after the indenter is pressed down a unit distance, the displacement is maintained for 1 to 3 minutes. The obtained one-dimensional resistance signal of the single-channel carbon nanotube sensor is preprocessed, and the Gram angle sum field method is used to encode the resistance data into a two-dimensional image for training and testing of a convolutional neural network model.

[0021] In addition to or as an alternative to one or more of the features disclosed above, establishing a convolutional neural network model includes the following steps:

[0022] Step S1: Network initialization. The VGG-16 model is used as the network architecture, which contains 13 convolutional layers and 3 fully connected layers. The number of network input layer nodes n, the number of output layer nodes m, the learning efficiency η, and the activation function f are determined according to the input and output data sets; the convolutional layer weights W and the fully connected layer weights W are initialized. fc ; Initialize the convolutional layer and fully connected layer bias b;

[0023] Step S2: Perform a series of convolution operations on the input image, calculate the output of each convolution layer, and apply the ReLU excitation function. Let X be the input image, the weight of the convolution layer is W, the bias is b, the output of the hidden layer is H, and the ReLU excitation function is f. The convolution layer output calculation formula is as follows: H = f(W*X+b), where * represents the convolution operation;

[0024] Step S3: Flatten the final output of the convolutional layer into a one-dimensional vector. Let the flattened input vector be Z and the weight of the fully connected layer be W. fc , bias is b fc , the output of the fully connected layer is O, O = W fc □Z+b fc , the last layer outputs the predicted value O k ;

[0025] Step S4: Error calculation, output O according to network prediction k and the expected output Y k , calculate the network prediction error e k , using mean square error as the loss function,

[0026] Step S5: Weight update, based on the network prediction error e k Calculate the gradient and update the weights of the convolutional layer and the fully connected layer through the back-propagation algorithm.

[0027] Step S6: Threshold update, based on the error e k Update the bias items of the convolutional layer and the fully connected layer,

[0028] Step S7: Determine whether the algorithm iteration is completed. If the iteration is not completed, re-execute step S2. If the iteration is completed, output the convolutional neural network prediction model.

[0029] In addition to or as an alternative to one or more features disclosed above, the training of the convolutional neural network model using the training set comprises the following steps:

[0030] Step T1: Randomly provide the input parameters of the training set samples to the input layer of the initialized convolutional neural network, and then forward the signal layer by layer until the output result is generated;

[0031] Step T2: Calculate the error by comparing the output result with the actual value, propagate the error back to the fully connected layer and convolutional layer, and update the weights and thresholds;

[0032] Step T3: Pass the error signal to the convolutional layer again and update the weights between the convolutional layer and the fully connected layer;

[0033] Step T4: Repeat steps T1 to T3 until the sum of the training errors of the training set samples reaches a preset maximum number of iterations, and the training stops. At this time, the trained convolutional neural network model is obtained.

[0034] In addition to or as an alternative to one or more features disclosed above, the verifying the trained convolutional neural network model using the test set includes the following steps:

[0035] Step T1: Keeping all parameters of the trained convolutional neural network model unchanged, each test sample in the test set is applied to the trained convolutional neural network model for verification;

[0036] Step T2: Calculate the test error and evaluate the generalization ability of the model based on the test error. Use the mean square error as the error metric.

[0037] In addition to or as an alternative to one or more features disclosed above, sequentially loading at least part of the plurality of grid coordinates includes:

[0038] At least part of the plurality of grid coordinates is loaded in sequence through a static indentation test, wherein the grid coordinates corresponding to the loading are the loading position coordinates.

[0039] According to a second aspect of the present invention, the present invention further provides a damage localization device for carbon fiber composite materials based on a convolutional neural network and a single-channel carbon nanotube sensor, comprising:

[0040] A sensor preparation unit is used to spray a carbon nanotube solution onto a template according to a preset concentration gradient to prepare a single-channel carbon nanotube sensor;

[0041] A laminate preparation unit is used to introduce the prepared single-channel carbon nanotube sensor into the middle interface of the carbon fiber composite material to prepare a carbon fiber composite laminate for damage location data acquisition;

[0042] a grid division unit, configured to select a tested area of the carbon fiber composite material laminate, establish a coordinate system in the tested area, and equally divide the tested area according to the coordinate system to form a plurality of grid coordinates;

[0043] a data grouping unit, configured to sequentially load at least a portion of the plurality of grid coordinates through a static indentation test, label acquired single-channel carbon nanotube sensor signals according to the loading position coordinates, use the acquired data labels as input and output data sets, and randomly divide the input and output data sets into a training set and a test set for a convolutional neural network model;

[0044] Neural network building unit, used to build convolutional neural network models;

[0045] A neural network training and verification unit, configured to train the convolutional neural network model using the training set and verify the trained convolutional neural network model using the test set until the convolutional neural network model reaches convergence;

[0046] The damage localization unit is used to locate structural damage using the converged convolutional neural network model and output the damage coordinates.

[0047] One of the above technical solutions has the following advantages or beneficial effects:

[0048] The present invention realizes damage location of CFRP through single-channel data acquisition, effectively simplifying the complexity and hardware cost brought by traditional multi-channel data acquisition, and highlighting the simplicity of system design. On this basis, combined with the end-to-end learning capability of convolutional neural network (CNN), damage-related features are automatically extracted from the original data, avoiding the tediousness of manual feature extraction, and realizing the automation of the whole process from data input to damage location. This technology not only solves the balance between data complexity and positioning accuracy, but also demonstrates the innovative ability to maintain high-precision damage location under the premise of simplifying hardware design. In addition, due to the reduced hardware requirements and improved degree of automation, this technology has shown stronger scalability and adaptability in practical applications. It is not only suitable for CFRP structure monitoring, but also has the potential to be extended to other composite material structure monitoring, further highlighting the technical level and broad application prospects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The technical solutions and other beneficial effects of the present invention will be made apparent by describing in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0050] Figure 1 is a flow chart for preparing a serpentine CNT sensor according to an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of sample preparation and experiment according to Example 1 of the present invention;

[0052] Figure 3 is a flowchart of CFRP damage location according to an embodiment of the present invention;

[0053] Figure 4 is a comparison between the predicted coordinates and the actual coordinates according to the first embodiment of the present invention;

[0054] Figure 5 It is a verification of the generalization ability of the predicted coordinates according to the first embodiment of the present invention;

[0055] Figure 6 is a flow chart of a damage localization method provided according to an embodiment of the present invention;

[0056] Figure 7 It is a virtual structural diagram of a damage localization device provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0058] In the description of the present invention, it should be noted that, unless otherwise specified or limited, the term "and / or" herein is merely a description of an association relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, and B exists alone. Furthermore, the character " / " herein, unless otherwise specified, generally indicates that the associated objects are in an "or" relationship.

[0059] For descriptive purposes, spatially relative terms such as "below," "beneath," "below," "above," "on," etc. may be used herein to describe the relationship of one element or feature illustrated in the accompanying drawings to another element or feature. Spatially relative terms are intended to encompass different orientations of the device when in use, in operation, and / or in manufacture other than the orientation depicted in the accompanying drawings. For example, if the device in the accompanying drawings is turned over, elements described as being "below" or "beneath" other elements or features will be "above" the other elements or features. Thus, the exemplary term "below" may encompass both above and below orientations. Furthermore, the device may be oriented in other directions (e.g., rotated 90 degrees or in other orientations), and therefore the spatially relative descriptors used herein should be interpreted accordingly.

[0060] Reference Figure 6 In one embodiment of the present invention, a method for locating damage in a carbon fiber composite material based on a convolutional neural network and a single-channel carbon nanotube sensor is disclosed, which may include the following steps:

[0061] Step S1, spraying a carbon nanotube solution onto a template according to a preset concentration gradient to prepare a single-channel carbon nanotube sensor;

[0062] Step S2, introducing the prepared single-channel carbon nanotube sensor into the middle interface of the carbon fiber composite material to prepare a carbon fiber composite material laminate for damage location data acquisition;

[0063] Step S3, selecting a tested area of the carbon fiber composite material laminate, establishing a coordinate system in the tested area, and equally dividing the tested area according to the coordinate system to form a plurality of grid coordinates;

[0064] Step S4, establishing a convolutional neural network model;

[0065] Step S5, sequentially loading at least some of the plurality of grid coordinates, labeling the acquired single-channel carbon nanotube sensor signals according to the loading position coordinates, using the acquired data labels as input and output data sets, and randomly dividing the input and output data sets into a training set and a test set for a convolutional neural network model;

[0066] Step S6, training the convolutional neural network model using the training set, and verifying the trained convolutional neural network model using the test set until the convolutional neural network model reaches convergence;

[0067] In step S7, the converged convolutional neural network model is used to locate structural damage and output the damage coordinates.

[0068] It can be understood that the order of step S4 and step S5 can be changed. Step S4 can be executed first and then step S5, or step S5 can be executed first and then step S4. The two steps can also be performed simultaneously. The specific execution order can be set according to actual conditions and is not limited here.

[0069] Furthermore, the concentration of the carbon nanotube solution is 2-2.5% in terms of gravity fraction, which does not cause clogging of the airbrush during spraying, allowing for smooth spraying.

[0070] Furthermore, in the step of spraying the carbon nanotube solution onto the template according to a preset concentration gradient, the carbon nanotube solution is sprayed into at least one of a bent shape, a circuitous shape, an S shape, a bow shape, and a W shape. Spraying the carbon nanotube solution into at least one of a bent shape, a circuitous shape, an S shape, a bow shape, and a W shape according to a preset concentration gradient can significantly enhance the sensitive capture capability of stress waves. These complex shapes increase the effective coverage area of the material surface, enabling the sensor to more comprehensively capture and transmit stress waves in different directions, thereby improving the accuracy of damage location. The non-linear shape design enables the sensor to have multi-directional response capabilities, which is particularly prominent under complex stress fields. In addition, the circuitous and bent shapes improve the stability and consistency of the signal by dispersing the propagation path of the stress wave, further enhancing the positioning accuracy. More importantly, the flexible design of these shapes can be optimized and adjusted according to different CFRP structures and monitoring requirements to achieve customized sensor design, thereby improving the overall performance and adaptability of the sensor.

[0071] Furthermore, the step of spraying the carbon nanotube solution onto the template according to a preset concentration gradient includes:

[0072] Set the unit area of concentration gradient change to 80-120 mm 2 In each spraying process, a unit area is added from one end of the template along the template path, and this is repeated 3 to 5 times until the carbon nanotubes completely cover the template surface.

[0073] First, this method of gradually increasing the concentration ensures the uniform distribution of carbon nanotubes on the template surface, effectively reducing performance fluctuations caused by uneven concentration. Second, by increasing the unit area in each spraying and repeating the operation multiple times, a gradual concentration gradient structure can be formed on the template surface, making the sensor more sensitive to stress waves in different areas, thereby improving detection accuracy and reliability. In addition, the layer-by-layer progressive concentration setting can also enhance the overall structural stability of the sensor, ensure the adhesion and durability of the carbon nanotube coating, and thus extend the service life of the sensor. Overall, this setting method not only optimizes the performance of the sensor, but also improves its adaptability to complex stress fields.

[0074] Optionally, the template is made of one of copper, aluminum, stainless steel, plastic and composite materials.

[0075] Furthermore, after spraying, the template is immersed in an etching solution and etched for 6 to 10 hours, and finally washed and dried to obtain the single-channel carbon nanotube sensor with a concentration gradient.

[0076] Optionally, the ply of the carbon fiber composite material laminate is [0 / 90 / 0 / 90 / 0 / CNT 1 / 2 ]S , where s indicates that the layer design is symmetrical; CNT indicates a single-channel carbon nanotube sensor, CNT 1 / 2 It indicates that a single-channel carbon nanotube sensor is used as the middle layer of the carbon fiber composite material laminate; and insulating layers are laid on the upper and lower surfaces of the single-channel carbon nanotube sensor.

[0077] Optionally, in the process of loading at least part of the several grid coordinates in sequence, the loading rate of the indenter is 0.5 to 1.5 mm / min, and the displacement of the indenter is maintained for 1 to 3 minutes after the indenter is pressed down by 1 unit distance. The one-dimensional resistance signal of the obtained single-channel carbon nanotube sensor is preprocessed, and the Gram angle sum field method is used to encode the resistance data into a two-dimensional image for training and testing the convolutional neural network model.

[0078] Furthermore, the establishment of the convolutional neural network model includes the following steps:

[0079] Step T1: Network initialization. The VGG-16 model is used as the network architecture, which contains 13 convolutional layers and 3 fully connected layers. The number of network input layer nodes n, the number of output layer nodes m, the learning efficiency η, and the activation function f are determined according to the input and output data sets; the convolutional layer weights W and the fully connected layer weights W are initialized. fc ; Initialize the convolutional layer and fully connected layer bias b;

[0080] Among them, the VGG16 model is a deep convolutional neural network developed by the research team of Oxford University (Visual Geometry Group) in 2014. The VGG16 model is well suited for classification and positioning tasks. The convolution layers of VGG16 all use 3×3 convolution kernels. This design is a notable feature of VGG16. A 3×3 convolution kernel contains the smallest unit of the top, bottom, left and right of a pixel. Continuous layers of 3×3 convolution kernels can fit more complex features and increase the depth of the network. Two 3×3 convolution kernels can replace a 5×5 convolution kernel, and three 3×3 convolution kernels can replace a 7×7 convolution kernel. In this way, a convolution layer with multiple small convolution kernels replaces a convolution layer with a larger convolution kernel. On the one hand, the number of parameters is reduced, and on the other hand, the number of nonlinearities is increased, the learning ability is better, and the expression ability of the network is also improved.

[0081] Step T2: Perform a series of convolution operations on the input image, calculate the output of each convolution layer, and apply the ReLU excitation function. Let X be the input image, the weight of the convolution layer is W, the bias is b, the output of the hidden layer is H, and the ReLU excitation function is f. The convolution layer output calculation formula is as follows: H = f(W*X+b), where * represents the convolution operation;

[0082] Step T3: Flatten the final output of the convolutional layer into a one-dimensional vector. Let the flattened input vector be Z and the weight of the fully connected layer be W. fc , bias is b fc , the output of the fully connected layer is O, O = W fc □Z+b fc , the last layer outputs the predicted value O k ;

[0083] Step T4: Error calculation, output O according to network prediction k and the expected output Y k , calculate the network prediction error e k , using mean square error as the loss function,

[0084] Step T5: Weight update, based on the network prediction error e k Calculate the gradient and update the weights of the convolutional layer and the fully connected layer through the back-propagation algorithm.

[0085] Step T6: Threshold update, based on the error e k Update the bias items of the convolutional layer and the fully connected layer,

[0086]

[0087] Step T7: Determine whether the algorithm iteration is completed. If the iteration is not completed, re-execute step S2. If the iteration is completed, output the convolutional neural network prediction model.

[0088] Modifying the VGG16 model to predict coordinates through a regression layer and then training and testing it on image datasets offers significant technical advantages. First, this modification enables the model to directly map image features to spatial locations with high precision, enabling more accurate object localization. Leveraging VGG16's powerful feature extraction capabilities, the model not only performs well in classification tasks but also directly outputs coordinate predictions through the regression layer, enabling efficient end-to-end learning and reducing intermediate steps and computational overhead. This architecture makes the model more adaptable and can be extended to tasks such as pose estimation and keypoint detection, expanding its application range. Furthermore, direct coordinate prediction simplifies post-processing, improving the system's real-time performance and responsiveness, while also enhancing the model's generalization capabilities, making it more stable and reliable across diverse scenarios. In summary, this modification improves the model's positioning accuracy, adaptability, and performance, making it more practical in a variety of applications.

[0089] Furthermore, the training of the convolutional neural network model using the training set includes the following steps:

[0090] Step P1: Randomly provide the input parameters of the training set samples to the input layer of the initialized convolutional neural network, and then forward the signal layer by layer until the output result is generated;

[0091] Step P2: Calculate the error by comparing the output result with the actual value, propagate the error back to the fully connected layer and convolutional layer, and update the weights and thresholds;

[0092] Step P3: Pass the error signal to the convolutional layer again and update the weights between the convolutional layer and the fully connected layer;

[0093] Step P4: Repeat steps T1 to T3 until the sum of the training errors of the training set samples reaches a preset maximum number of iterations, and the training stops. At this time, the trained convolutional neural network model is obtained.

[0094] Using a training set to train a convolutional neural network model offers several technical advantages. First, by continuously iteratively optimizing model parameters, the accuracy of feature extraction and the model's generalization capabilities are improved, enabling it to perform well even on unseen data. During training, the model gradually learns and extracts more representative high-dimensional features, enhancing its recognition and localization capabilities. Furthermore, by diversifying the training set and using regularization techniques, the model avoids overfitting, improves robustness, and ensures stable performance in complex scenarios. The multi-layered structure of a convolutional neural network enables it to grasp complex nonlinear relationships, providing greater flexibility and accuracy when handling delicate tasks. Finally, as the amount of training data increases, the model is continuously optimized through adaptive learning, continuously improving its predictive accuracy and performance. In summary, this training process significantly improves the model's accuracy, generalization, and reliability in practical applications.

[0095] Furthermore, the verification of the trained convolutional neural network model using the test set includes the following steps:

[0096] Step R1: Keeping all parameters of the trained convolutional neural network model unchanged, apply each test sample in the test set to the trained convolutional neural network model for verification;

[0097] Step R2: Calculate the test error and evaluate the generalization ability of the model based on the test error. Use the mean square error as the error metric.

[0098] Through the above implementation method, using the test set to verify the convolutional neural network model can bring the following beneficial effects at the technical level. First, through the verification of the test set, the performance of the model on unseen data can be comprehensively evaluated, and the generalization ability and prediction accuracy of the model can be accurately measured. Test set verification helps to discover overfitting problems that may occur in the model during the training process, thereby guiding further optimization and adjustment of the model to ensure that the model not only performs well on the training set, but also maintains high accuracy in actual applications. In addition, through the feedback of the test set, model developers can identify the weaknesses of the model in different scenarios and conditions, and then perform targeted optimization to improve the robustness and reliability of the model. Overall, using the test set to verify the convolutional neural network model can effectively improve the practicality and stability of the model, laying a solid foundation for its application in complex real-world environments.

[0099] Furthermore, the sequentially loading at least part of the plurality of grid coordinates includes:

[0100] At least part of the plurality of grid coordinates is loaded in sequence through a static indentation test, wherein the grid coordinates corresponding to the loading are the loading position coordinates.

[0101] Reference Figure 7 The present application also discloses a damage localization device for carbon fiber composite materials based on a convolutional neural network and a single-channel carbon nanotube sensor, comprising:

[0102] The sensor preparation unit 101 is used to spray a carbon nanotube solution onto a template according to a preset concentration gradient to prepare a single-channel carbon nanotube sensor;

[0103] A laminate preparation unit 102 is used to introduce the prepared single-channel carbon nanotube sensor into the middle interface of the carbon fiber composite material to prepare a carbon fiber composite laminate for damage location data acquisition;

[0104] A grid division unit 103 is configured to select a tested area of the carbon fiber composite material laminate, establish a coordinate system in the tested area, and equally divide the tested area according to the coordinate system to form a plurality of grid coordinates;

[0105] a data grouping unit 104 configured to sequentially load at least a portion of the plurality of grid coordinates through a static indentation test, label the acquired single-channel carbon nanotube sensor signals according to the loading position coordinates, use the acquired data labels as input and output data sets, and randomly divide the input and output data sets into a training set and a test set for a convolutional neural network model;

[0106] A neural network construction unit 105 is used to establish a convolutional neural network model;

[0107] A neural network training and verification unit 106 is configured to train the convolutional neural network model using the training set and verify the trained convolutional neural network model using the test set until the convolutional neural network model reaches convergence;

[0108] The damage localization unit 107 is used to locate structural damage using the converged convolutional neural network model and output damage coordinates.

[0109] The methods and approaches involved in each unit have been described in detail in the aforementioned damage localization method and will not be repeated here.

[0110] The following describes in detail a method and device for predicting dynamic stress at a structural defect based on a BP neural network through several specific embodiments.

[0111] Example 1:

[0112] Figure 1 Demonstrates the fabrication of serpentine CNT sensors with concentration gradients by spray coating using an etched copper mesh template method.

[0113] Specifically, the preparation method of the serpentine CNT sensor with a concentration gradient includes:

[0114] Step F1, such as Figure 1 As shown in Figure a, a copper mesh was cut into a 100mm*100mm serpentine structure using a laser cutting method, with a width and spacing of 10mm. The serpentine copper mesh structure was then ultrasonically treated to remove surface contaminants and oxide layers.

[0115] Step F2, further diluting a CNTs (10 wt %) N,N-dimethylformamide (DMF) solution and a DMF solution in a ratio of 1:2 to prepare a CNTs dispersion for spraying;

[0116] Step F3, ultrasonic treatment of the CNTs dispersion was performed using an ultrasonic probe (Ymnl-1000Y, Nanjing Emanuel Instrument Equipment Co., Ltd.) at a constant temperature of 11°C, with a probe diameter of 15 mm, a power of 80%, an operating time of 3 s, a pause time of 5 s, and a total dispersion time of 20 min;

[0117] Step F4, such as Figure 1 As shown in b, the CNTs dispersion is sprayed onto the surface of the serpentine copper mesh structure using an airbrush. The entire spraying process is carried out on an electric heating plate with a temperature set at 90°C to accelerate the volatilization of the DMF solution.

[0118] Step F5: To achieve a concentration gradient effect on CNTs, set the unit area of the concentration gradient change to 100 mm 2During each spraying process, the area of the copper mesh is increased one unit at a time along the copper mesh path from one end, and this process is repeated three times until the CNTs are fully deposited on the copper wire surface under an optical microscope. During the spraying process, the distance between the copper mesh and the spray gun is controlled (10 cm). In other embodiments, the spraying can be repeated four or five times, which is not particularly limited, as long as the CNTs are fully deposited on the copper wire surface.

[0119] Step F6, such as Figure 1 As shown in c, after spraying, the copper mesh was etched with a dilute nitric acid solution and dried to obtain a serpentine CNT sensor with a concentration gradient. The drying process was carried out in an air drying oven at 65°C.

[0120] In step F7, the prepared serpentine CNT sensor is introduced between the layers of the CFRP structure to prepare a specimen for damage location research.

[0121] Furthermore, the preparation method of the experimental specimen includes:

[0122] Step M1, such as Figure 2 As shown in a, the CFRP laminate is [0 / 90 / 0 / 90 / 0 / CNT 1 / 2 ]s, the size of the CFRP laminate is 110mm*110mm*2mm;

[0123] In step M2, the prepared serpentine CNT sensor is introduced into the middle of the plywood to form a sensor network. To insulate the introduced serpentine CNT sensor from the CFRP, a 20 μm thick epoxy resin film is laid on the upper and lower surfaces of the serpentine CNT sensor.

[0124] In step M3, copper tapes are bonded as electrodes at both ends of the serpentine CNT sensor for resistance measurement, and conductive silver paste is used to minimize contact resistance.

[0125] Step M4: The laminated prepregs are put into a CFRP laminate by a hot press. The hot press pressure is set to 1 MPa, and the temperature is maintained at 90°C for 30 minutes. The temperature is then increased to 120°C and maintained for 90 minutes, and finally cooled to room temperature at a rate of 1°C per minute.

[0126] Step M5, such as Figure 2 As shown in b, a tested area of CFRP is selected, a coordinate system xoy is established in the tested area, and the tested area is equally divided according to the coordinate system xoy, with a grid coordinate spacing of 10 mm;

[0127] Step M6, such as Figure 2As shown in Fig. 2a, nine coordinate points (1,0), (1,4), (1,8), (5,0), (5,4), (5,8), (8,0), (8,4), and (8,8) were selected to conduct static indentation experiments in sequence. In the static indentation experiments, the indenter loading rate was 1 mm / min. After the indenter was pressed down 1 mm, the displacement was maintained for 2 minutes. The resistance signal of the serpentine CNT sensor was collected by a Keithley source meter.

[0128] Step M7, as Figure 3 As shown in ab, the resistance signal of the obtained serpentine CNT sensor was preprocessed and encoded into a two-dimensional image every 3 seconds using the Gramian summation angular field (GASF) encoding method. The dataset label is the grid coordinate corresponding to the indenter loading.

[0129] In step M8, these images are divided into data according to a ratio of 7:3 and used as input for the training set and test set respectively for training and testing the convolutional neural network model.

[0130] Furthermore, the training of the convolutional neural network model using the training set includes the following steps:

[0131] Step N1, use Figure 3 The two-dimensional image dataset obtained in ab is used as input data ( Figure 2 c), image dataset (input feature X d ) is composed of the convolution kernel (W P ) is convolved and activated by ReLU to obtain an output (Y P ), which is halved to 112*112*64 after max pooling. The two-dimensional image is convolved, ReLU activated and pooled, and the dimension is finally converted from 112*112*128 to 7*7*512 (see Figure 3 d) The features extracted from the 2D image are represented as flat vectors and the size is converted from 1*1*4096 to 1*1*1000 after the fully connected layer. Figure 3 As shown in Figure 5, the output value of a fully connected layer is finally transmitted to the regression layer to predict the coordinates (x, y) of the center of the damage area.

[0132] In step N2, the positioning accuracy of the model is evaluated using the mean absolute error.

[0133] Where m is the number of samples, y i is the distance between the actual coordinate and the origin, is the distance between the predicted coordinates and the origin.

[0134] Step N3, Figure 4A comparison of the predicted positions of the selected 8 points (except point (8, 8)) by ultrasound scanning and CNN model is given, and the average absolute error is less than 0.52 mm.

[0135] In step N4, the untrained coordinate point data of point (8, 8) is input into the established CNN model to evaluate its generalization ability. Figure 5 ac shows that at Δt2, Δt 27 and Δt 60 The positioning situation at three moments, MAE is as follows Figure 5 d, they are 5.09mm, 5.51mm and 5.23mm respectively, and the average MAE is less than 5.5mm. Figure 5 AD shows that the CNN model can well learn the features of 2D images from large-scale datasets and exhibit good generalization ability, thus being able to predict new damage locations.

[0136] The technical solution of this application has the following advantages:

[0137] 1. The sensor pattern of this application is designable, and the sensor size can be personalized according to the actual application requirements.

[0138] 2. Traditional CFRP damage localization methods typically rely on multi-channel data acquisition, which increases data processing complexity and hardware costs. By using single-channel data acquisition, this invention significantly reduces data processing complexity and costs while maintaining accurate damage localization. This advantage demonstrates technological innovation that reduces system complexity while maintaining performance.

[0139] 3. While reducing the number of data channels, this invention combines the use of CNNs to extract sufficient information from single-channel data, achieving high-precision damage localization. In this way, this invention resolves the contradiction between data complexity and positioning accuracy, demonstrating the innovative ability to achieve efficient and precise positioning while simplifying hardware design.

[0140] The above description is only part of the implementation methods of the embodiments of the present invention and does not impose any form of limitation on the application. The protection scope of the embodiments of the present invention is not limited to this. Any simple modifications, equivalent changes and modifications that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in the embodiments of the present invention should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A damage localization method for carbon fiber composite materials based on convolutional neural networks and single-channel carbon nanotube sensors, characterized in that: The following steps are involved: The carbon nanotube solution is sprayed onto the template according to a preset concentration gradient to prepare a single-channel carbon nanotube sensor; The prepared single-channel carbon nanotube sensor is introduced into the middle interface of the carbon fiber composite material to prepare a carbon fiber composite laminate for damage location data acquisition; Selecting a tested area of the carbon fiber composite material laminate, establishing a coordinate system in the tested area, and equally dividing the tested area according to the coordinate system to form a plurality of grid coordinates; Build a convolutional neural network model; sequentially loading at least some of the plurality of grid coordinates to obtain single-channel carbon nanotube sensor signals, labeling them according to the loading position coordinates, using the obtained data labels as input and output data sets, and randomly dividing the input and output data sets into a training set and a test set for a convolutional neural network model; The convolutional neural network model is trained using the training set, and the trained convolutional neural network model is verified using the test set until the convolutional neural network model reaches convergence; Use the converged convolutional neural network model to locate structural damage and output the damage coordinates; In the process of sequentially loading at least a portion of the plurality of grid coordinates, the loading rate of the indenter is 0.5 to 1.5 mm / min, and after the indenter is pressed down a unit distance, the displacement is maintained for 1 to 3 minutes. The obtained one-dimensional resistance signal of the single-channel carbon nanotube sensor is preprocessed, and the resistance data is encoded into a two-dimensional image using the Gram angle sum field method for training and testing the convolutional neural network model; The establishment of the convolutional neural network model includes the following steps: Step S1: Network initialization, using the VGG-16 model as the network architecture, which contains 13 convolutional layers and 3 fully connected layers. The number of network input layer nodes is determined according to the input and output data sets. n , the number of output layer nodes m , learning efficiency η , activation function f ; Initialize the convolutional layer weights and the fully connected layer weights ; Initialize convolutional layer and fully connected layer bias ; Step S2: Perform a series of convolution operations on the input image, calculate the output of each convolution layer, and apply the ReLU excitation function. X is the input image, and the weight of the convolutional layer is W , the bias is b , the hidden layer output is H , the ReLU activation function is f , then the convolutional layer output calculation formula is as follows: , where * represents the convolution operation; Step S3: Flatten the final output of the convolutional layer into a one-dimensional vector. Let the flattened input vector be , the weight of the fully connected layer is , the bias is , the output of the fully connected layer is , , the last layer outputs the predicted value ; Step S4: Error calculation, output based on network prediction and expected output , calculate the network prediction error , using mean square error as the loss function, ; Step S5: Weight update, based on the network prediction error Calculate the gradient and update the weights of the convolutional layer and the fully connected layer through the back-propagation algorithm. , ; Step S6: Threshold update, based on the error Update the bias items of the convolutional layer and the fully connected layer, ; Step S7: Determine whether the algorithm iteration is completed. If the iteration is not completed, re-execute step S2. If the iteration is completed, output the convolutional neural network prediction model.

2. The damage localization method according to claim 1, wherein: Calculated by gravity, the concentration of the carbon nanotube solution is 2-2.5%.

3. The damage localization method according to claim 1, wherein: The carbon nanotube solution is sprayed into the template according to a preset concentration gradient, and the carbon nanotube solution is sprayed into at least one of a bent shape, a winding shape, a snake shape, a bow shape and a W shape.

4. The damage localization method according to claim 3, wherein: The step of spraying the carbon nanotube solution onto the template according to a preset concentration gradient comprises: Set the unit area of concentration gradient change to 80~120mm 2 In each spraying process, a unit area is added from one end of the template along the template path, and this is repeated 3 to 5 times until the carbon nanotubes completely cover the template surface.

5. The damage localization method according to claim 4, wherein: The material of the template is one of copper, aluminum, stainless steel, plastic and composite material.

6. The damage localization method according to claim 4, wherein: After the spraying is completed, the template is immersed in an etching solution and etched for 6 to 10 hours. Finally, the template is cleaned and dried to obtain the single-channel carbon nanotube sensor with a concentration gradient.

7. The damage localization method according to claim 1, wherein: The ply of the carbon fiber composite material laminate is [0 / 90 / 0 / 90 / 0 / CNT 1 / 2 ] S， Among them, s means that the layer design is symmetrical; CNT means single-channel carbon nanotube sensor, CNT 1 / 2 It indicates that a single-channel carbon nanotube sensor is used as the middle layer of the carbon fiber composite material laminate; and insulating layers are laid on the upper and lower surfaces of the single-channel carbon nanotube sensor.

8. The damage localization method according to claim 1, wherein: The method of training the convolutional neural network model using the training set comprises the following steps: Step T1: Randomly provide the input parameters of the training set samples to the input layer of the initialized convolutional neural network, and then forward the signal layer by layer until the output result is generated; Step T2: Calculate the error by comparing the output result with the actual value, propagate the error back to the fully connected layer and convolutional layer, and update the weights and thresholds; Step T3: Pass the error signal to the convolutional layer again and update the weights between the convolutional layer and the fully connected layer; Step T4: Repeat steps T1 to T3 until the sum of the training errors of the training set samples reaches the preset maximum number of iterations. The training stops and a trained convolutional neural network model is obtained.

9. The damage localization method according to claim 1, wherein: The method of verifying the trained convolutional neural network model using the test set includes the following steps: Step T1: Keeping all parameters of the trained convolutional neural network model unchanged, each test sample in the test set is applied to the trained convolutional neural network model for verification; Step T2: Calculate the test error and evaluate the generalization ability of the model based on the test error. Use the mean square error as the error metric. .

10. The damage localization method according to claim 1, wherein: The sequentially loading at least part of the plurality of grid coordinates includes: At least part of the plurality of grid coordinates is loaded in sequence through a static indentation test, wherein the grid coordinates corresponding to the loading are the loading position coordinates.

Images