Electromagnetic tomography flaw detection method for carbon fiber inhaul cable

Abstract

The invention relates to an electromagnetic tomography flaw detection method for a carbon fiber inhaul cable, which comprises the following steps: S1, constructing a simulation model of the carbon fiber composite inhaul cable, and obtaining an initial sensitivity matrix and an initial voltage; s2, defect shapes and positions are randomly set for the simulation model, excitation is applied to the simulation model with defects, detection voltage is obtained, and a defect image is constructed; s3, repeating the step S2 for at least 600 times to obtain a defect image set and a detection voltage set; s4, iteratively updating the initial sensitivity matrix according to the initial voltage, the preprocessed defect image and the detection voltage; s5, constructing a defect positioning model; s6, applying voltage to the detected carbon fiber inhaul cable to obtain detection voltage of the detected carbon fiber inhaul cable; and S7, constructing a cross-section image of the detected carbon fiber inhaul cable, inputting the cross-section image into the trained defect positioning model, and determining the defect position according to the interface image with the defect. According to the method, the sensitivity matrix is updated, so that the influence of micro-vibration noise on imaging quality is reduced, and defect imaging is clearer.

Description

Technical Field

[0001] The invention relates to an electromagnetic tomography detection method, in particular to an electromagnetic tomography flaw detection method for a carbon fiber cable. Background Art

[0002] Carbon fiber composite cables overcome the problems of heavy weight and easy surface corrosion of traditional steel cables, and meet the load-bearing requirements of long-span cables. They are an important direction for the lightweight development of super-large amusement facilities and aerial passenger ropeways. Due to the influence of manufacturing process and service environment, carbon fiber composite cables will inevitably have damage defects such as debonding and broken strands. The shape and location of the defects will directly affect the load-bearing capacity and service life of carbon fiber composite cables, and even lead to safety accidents due to cable fracture and failure. Therefore, the electromagnetic tomography of carbon fiber composite cable defects is of great significance for its safe and reliable service.

[0003] Electromagnetic tomography (EMT) is a non-contact, non-invasive non-destructive testing technology. Based on the principle of electromagnetic induction, it can visualize the conductivity distribution of the measured area and is widely used in industrial flaw detection and multiphase flow detection. In order to improve the imaging quality of electromagnetic tomography systems, some scholars have conducted a lot of research in the selection of excitation modes, sensor array optimization, image reconstruction algorithms, etc. Image reconstruction algorithms can be divided into fast reconstruction algorithms and iterative reconstruction algorithms. Both require the prior sensitivity matrix to be obtained in advance and participate in the image reconstruction calculation. Therefore, the sensitivity matrix plays a vital role in the object field image reconstruction process. The sensitivity matrix is ​​mainly obtained through perturbation methods and field quantity extraction methods. Among them, the field quantity extraction method is superior in terms of acquisition speed and imaging effect. At present, relevant studies have shown that electromagnetic tomography technology has a good imaging effect in defect detection of high conductivity materials such as metals. However, it is difficult to accurately identify the defect position and shape for defect detection of low conductivity materials such as carbon fiber cables.

[0004] Since carbon fiber cables have the characteristics of electrical anisotropy and low conductivity, the occurrence of defects has little effect on the change of the detection voltage value of the coil sensor, and the detection voltage is easily affected by micro-vibration and mechanical noise. In addition, the prior sensitivity matrix obtained by the traditional field quantity extraction method, the distribution of the conductivity of the object field obtained is seriously affected by edge artifacts, so that the position and shape information of the defect image is easily submerged in the artifact shape, making it difficult to detect and identify defects in carbon fiber cables. Summary of the invention

[0005] The purpose of the present invention is to provide an electromagnetic tomography flaw detection method for carbon fiber cables to solve the problem of inaccurate recognition of the position and shape of defects in carbon fiber composite cables.

[0006] The object of the present invention is achieved in that:

[0007] A method for electromagnetic tomography flaw detection of a carbon fiber cable comprises the following steps:

[0008] S1. Construct an electromagnetic simulation model of a carbon fiber cable according to the specifications of the tested carbon fiber cable, apply current excitation to the three-dimensional electromagnetic simulation model of the carbon fiber cable to obtain an initial voltage, and establish an initial sensitivity matrix of the material field region of the cross-section of the carbon fiber cable by a field quantity extraction method;

[0009] S2. randomly set the shape and position of the defect in the simulation model, apply excitation to the simulation model with the defect, and obtain a detection voltage; draw a cross-sectional image of the simulation model where the defect is located according to the defect shape and position set in the simulation model to obtain a defect image; perform edge gradient filling on the defect image, and add Gaussian noise to the detection voltage;

[0010] S3. Repeat step S2 at least 600 times to obtain a defect image set and a detection voltage set of the electromagnetic simulation model of the carbon fiber cable;

[0011] S4. Iteratively update the initial sensitivity matrix according to the initial voltage, the defect image set and the detection voltage set;

[0012] S5. Construct a defect location model for detecting intrinsic defects in carbon fiber cable;

[0013] S6. Setting a detection probe, the structure of which is that a plurality of annular coils pointing to the center of the circle are set on the inner ring surface of the annular carrier, which are divided into detection coils and excitation coils. The number of detection coils is the same as the number of excitation coils, and they are evenly distributed in a double-layer ring around the object field. The inner coil is set as the detection coil, and the outer coil is set as the excitation coil; the detection probe moves from one end of the carbon fiber cable to be tested to the other end at a preset speed, and completes the collection of detection voltage every 300 milliseconds to obtain a collection of detection voltages of the carbon fiber cable to be tested;

[0014] S7. According to the set of detection voltages of the tested carbon fiber cable and the iteratively updated sensitivity matrix, the grayscale value of each pixel in the cross section of the tested carbon fiber cable is calculated, and a cross-sectional image of the tested carbon fiber cable is generated according to the grayscale value of each pixel. The cross-sectional image of the tested carbon fiber cable is input into the trained defect location model, and a clear cross-sectional image of the tested carbon fiber cable is output. When a defective cross-sectional image exists, the defect position is determined by the acquisition time of the detection voltage corresponding to the defective cross-sectional image.

[0015] Furthermore, the specific method of iteratively updating the initial sensitivity matrix in step S4 is:

[0016] S4-1. Iteratively update the initial sensitivity matrix. The formula for iterative update is:

[0017]

[0018] Among them, S0 is the initial sensitivity matrix, S1 is the matrix after the first iteration update, and G(S0) is the normalized pixel gray value vector data of the defect image corresponding to the sensitivity matrix S0;

[0019] The calculation formula of G(S0) is:

[0020]

[0021] Among them, S0 T is the transposed matrix of S0, U * is the set of detection voltages obtained by randomly setting the defects of the simulation model, is the initial voltage;

[0022] S4-2. Calculate the normalized pixel grayscale value vector data of the defect image corresponding to the iteratively updated sensitivity matrix, and the calculation formula is:

[0023]

[0024] S4-3. Calculate the loss function based on the normalized pixel gray value vector data of the defect image:

[0025] Loss||G * -G(S1)||

[0026] Among them, G * A set of normalized pixel grayscale value vector data of defect images obtained by randomly setting defects of the simulation model;

[0027] S4-4. Determine whether the loss function obtained in step S4-3 is less than 1×10 -3 Or iteratively update to a preset number of iterations. When the judgment result is yes, stop iterative updating and output the updated sensitivity matrix S1. When the judgment result is no, execute steps S4-5-S4-8;

[0028] S4-5. Iteratively update the sensitivity matrix. The formula for iterative update is:

[0029]

[0030] Among them, S i is the sensitivity matrix obtained after the i-th iteration update, S i-1 is the sensitivity matrix obtained after the i-1th iteration update, a is the learning rate, indicating the step size of the iteration, G(S i-1) is the sensitivity matrix S i-1 Normalized pixel gray value vector data of the corresponding defect image;

[0031] G(S i-1 ) is calculated as:

[0032]

[0033] Among them, S i T For S i The transposed matrix, U * is the set of detection voltages obtained by randomly setting the defects of the simulation model, is the initial voltage.

[0034] S4-6. Calculate the normalized pixel grayscale value vector data of the defect image corresponding to the iteratively updated sensitivity matrix, and the calculation formula is:

[0035]

[0036] S4-7. Calculate the loss function based on the normalized pixel gray value vector data of the defect image:

[0037] Loss||G * -G(S i )||

[0038] Among them, G * A set of normalized pixel grayscale value vector data of defect images obtained by randomly setting defects of the simulation model;

[0039] S4-8. Determine whether the loss function output in step S4-7 is less than 1×10 -3 Or iteratively update to a preset number of iterations. When the judgment result is no, repeat steps S4-5-S4-7 until the judgment result is yes, stop iterative updating, and output the sensitivity matrix of the last iterative update.

[0040] Furthermore, the specific method of constructing the defect location model is as follows:

[0041] S5-1. Improve the U-Net model;

[0042] S5-2. Annotate the defect image set obtained in step S3 to obtain an image set with annotated defects;

[0043] S5-3. According to the updated sensitivity matrix and the detection voltage set, the pixel value of each pixel in the defect image set obtained in step S3 is recalculated to reconstruct the defect image set;

[0044] S5-4. The images in the image set with marked defects and the images in the reconstructed defect image set are used as samples and input into the improved U-Net model for training to obtain a defect localization model for detecting intrinsic defects of carbon fiber material cables.

[0045] Furthermore, the specific way to improve the U-Net model is as follows: the double convolution of the U-Net deep learning network encoding part is replaced by a residual module. The structure of each residual module is as follows: the input of the residual module is divided into two paths, one path passes through the first convolution layer, the first batch normalization layer, the activation function, the second convolution layer and the second batch normalization layer, the other path passes through 1×1 convolution, and the outputs of the two paths are added and output after passing through the activation function.

[0046] Furthermore, the iteration cycle of defect localization model training is 100 epochs, the Batch Size is set to 16, the initial learning rate is set to 0.0001, and the loss function adopts the cross entropy loss function and Dice coefficient.

[0047] The present invention firstly uses the random gradient descent method to update the sensitivity matrix. The updated sensitivity matrix is ​​more suitable for imaging of specific object field information, reduces the influence of micro-vibration noise on imaging quality, and reconstructs a clearer defect image. Then, the U-Net network is used to construct a defect location model to improve the accuracy of the model in defect location, and the defect location is performed on the defect image of the tested carbon fiber cable, which weakens the influence of artifact noise and more intuitively displays the position and size information of the defect, which is conducive to the quantitative evaluation of the defects of the carbon fiber composite cable. The two stages of sensitivity optimization and defect image optimization are combined to realize the quantitative detection and visual characterization of the defects of the low-conductivity carbon fiber material cable.

[0048] The present invention is applicable to an electromagnetic tomography method for defects in low-conductivity carbon fiber composite cable, so that it can weaken micro-vibration and mechanical noise, reduce the influence of edge artifacts, and realize accurate identification of the position and shape of defects in carbon fiber composite cable. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the present invention.

[0050] Figure 2 It is a flow chart of randomly setting defects on the carbon fiber composite cable model.

[0051] Figure 3 It is the sensitivity matrix optimization flow chart.

[0052] Figure 4 It is the flowchart of defect image reconstruction in the second stage.

[0053] Figure 5 It is a schematic diagram of testing the carbon fiber cable under test.

[0054] Figure 6 It is a comparison chart of defect imaging at each stage. DETAILED DESCRIPTION

[0055] The present invention is further described below.

[0056] like Figure 1 As shown, the electromagnetic tomography flaw detection method of the carbon fiber cable provided by the present invention comprises the following steps:

[0057] S1. Construct an electromagnetic simulation model of the carbon fiber cable according to the specifications of the tested carbon fiber cable, apply current excitation to the three-dimensional electromagnetic simulation model of the carbon fiber cable to obtain the initial voltage, and establish the initial sensitivity matrix of the material field area of ​​the carbon fiber cable cross section through the field quantity extraction method.

[0058] The constructed simulation model is an ideal damage-free model, and the initial voltage obtained is the full-field voltage, and the specifications are the same as the parameters of the tested carbon fiber cable.

[0059] like Figure 2 As shown, the electromagnetic simulation model establishment and excitation detection process are realized by COMSOL Multiphysics software; the electromagnetic simulation model of the carbon fiber cable is mainly composed of a carbon fiber cable 1, an outer coil 2 and an inner coil 3. The carbon fiber cable 1 is composed of 7 carbon fiber rods, the length of a single carbon fiber rod is 100mm, the diameter is 8mm, the axial conductivity is set to 10060 (S / M), and the radial conductivity is 138 (S / M). The outer coil 2 is the excitation coil, and the inner coil 3 is the detection coil. The inner diameter of the inner coil and the outer coil are both 5mm, and the outer diameter is 10mm. The outer coil and the inner coil are both wound with copper wire. One inner coil corresponds to one outer coil, and the central axes of the corresponding inner coil and outer coil are on the same line.

[0060] The excitation current of the excitation coil is set to 1A, and the excitation frequency is set to 1MHz. When each excitation coil is excited, the 7 detection coils except the one below the excitation coil (that is, except the detection coil corresponding to the excitation coil) are used to collect detection voltage. During each detection process, the 8 excitation coils are excited in turn in a clockwise direction, so 56 voltage values ​​can be obtained in one detection process.

[0061] In the simulation model, a unit current is applied to each excitation coil and detection coil to extract the vector magnetic potential distribution of the field area of ​​the object being measured. The extracted independent vector magnetic potential distribution data is combined with the excitation coil and the detection coil in turn to calculate the conductivity sensitivity matrix. The formula is as follows:

[0062] S=-ω 2 AA A B

[0063] Where ω represents the operating frequency of the electromagnetic detection system, A is the outer coil, and A A is the vector magnetic potential distribution of the object field under the excitation condition A, B is the inner coil, A B is the vector magnetic potential distribution of the object field under B excitation.

[0064] S2. Randomly set the shape and position of the defect for the simulation model, apply excitation to the simulation model with the defect, and obtain a detection voltage; according to the defect shape and position set in the simulation model, draw a cross-sectional image of the simulation model where the defect is located to obtain a defect image; perform edge gradient filling on the defect image, and add Gaussian noise to the detection voltage.

[0065] In the simulation model, a defect is set at the cross section of the carbon fiber cable surrounded by the detection coil, and a difference is made between the rectangular structure of the preset specification and the geometric structure of the carbon fiber cable, and the internal boundary is preserved to form a defective carbon fiber composite cable structure model, and its axial conductivity is set to 10060S / m and radial conductivity to 138S / m. The preset specification of the rectangular structure used in the present invention is: 8mm long, 4mm wide, and 2mm high.

[0066] like Figure 2 As shown, the defect shape and position of the three-dimensional carbon fiber material cable simulation model are randomly set, the detection voltage value data of multiple coils excited in turn are collected, the detection voltage is obtained, and the cross-sectional defect image of the simulation model is drawn according to the set defect shape and position, and the corresponding relationship between the defect image and the detection voltage is established, so that the detection voltage and the defect image correspond to each other.

[0067] Randomly set defects are of the same type as the defects of the tested carbon fiber cables, such as broken strand defects.

[0068] Gradient filling of the edge of the defect image can effectively improve the smoothness of the defect imaging edge, and adding Gaussian noise to the detection voltage can increase the robustness of the defect location model.

[0069] S3. Repeat step S2 at least 600 times to obtain a defect image set of the electromagnetic simulation model of the carbon fiber cable and a set of detection voltages corresponding to various defects.

[0070] Defects are set multiple times to facilitate updating of the sensitivity matrix and provide samples for subsequent model training.

[0071] By rotating and moving the spatial position of the rectangular structure in step S2, a carbon fiber composite cable model with various defect types is formed after difference, and the detection voltage value data of multiple coils excited in turn are collected to obtain the detection voltage set corresponding to various defects.

[0072] S4. Iteratively update the initial sensitivity matrix according to the initial voltage, the preprocessed defect image and the detection voltage.

[0073] like Figure 3 As shown, in order to solve the problem that the sensitivity matrix obtained by the traditional field quantity extraction method is not suitable for low conductivity material detection, the present invention adopts the gradient descent idea to optimize the initial sensitivity matrix S0 and establishes a sensitivity matrix that is more suitable for low conductivity carbon fiber material cable defect imaging.

[0074] S4-1. Iteratively update the initial sensitivity matrix. The formula for iterative update is:

[0075]

[0076] Among them, S0 is the initial sensitivity matrix, S1 is the matrix after the first iteration update, and G(S0) is the normalized pixel gray value vector data of the defect image corresponding to the sensitivity matrix S0;

[0077] The calculation formula of G(S0) is:

[0078]

[0079] Among them, S0 T is the transposed matrix of S0, U * is the set of detection voltages obtained by randomly setting the defects of the simulation model, is the initial voltage;

[0080] S4-2. Calculate the normalized pixel grayscale value vector data of the defect image corresponding to the iteratively updated sensitivity matrix, and the calculation formula is:

[0081]

[0082] S4-3. Calculate the loss function based on the normalized pixel gray value vector data of the defect image:

[0083] Loss||G * -G(S1)|| (4)

[0084] Among them, G * A set of normalized pixel grayscale value vector data of defect images obtained by randomly setting defects of the simulation model;

[0085] S4-4. Determine whether the loss function obtained in step S4-3 is less than 1×10 -3Or iteratively update to a preset number of iterations. When the judgment result is yes, stop iterative updating and output the updated sensitivity matrix S1. When the judgment result is no, execute steps S4-5-S4-8;

[0086] S4-5. Iteratively update the sensitivity matrix. The formula for iterative update is:

[0087]

[0088] Among them, S i is the sensitivity matrix obtained after iterative update for the i-th time, S i-1 is the sensitivity matrix obtained after iterative update i-1 times, a is the learning rate, indicating the step size of the iteration, G(S i-1 ) is the sensitivity matrix S i-1 The normalized pixel gray value vector of the corresponding defect image.

[0089] G(S i-1 ) is calculated as:

[0090]

[0091] Among them, S i T For S i The transposed matrix, U * is a set of detection voltages obtained by randomly setting the defects of the simulation model. The detection voltage in each column is the detection voltage of each defect setting. is the initial voltage, The data in each column is the same, and the number of columns is the same as U * The same number of columns.

[0092] S4-6. Calculate the corresponding defect image according to the iteratively updated sensitivity matrix calculated by formula (5), and the calculation formula is:

[0093]

[0094] S4-7. Calculate the loss function based on the defect image:

[0095] Loss||G * -G(S i )|| (8)

[0096] Among them, G * It is a set of normalized pixel grayscale value vector data of defect images obtained by randomly setting defects of the simulation model, that is, a set of normalized pixel grayscale value vectors of all defect images preprocessed in step S2.

[0097] The loss function of formula (4) can also adopt the smoothed L1 loss function and L2 loss function. The formulas of the smoothed L1 loss function and L2 loss function are as follows:

[0098]

[0099] Among them, β is the switching threshold between L1 loss function and L2 loss function.

[0100] The smoothed L1 loss function and L2 loss function can be very robust to voltage signals containing noise interference, while taking into account the influence of small errors.

[0101] S4-8. Determine whether the loss function output in step S4-7 is less than 1×10 -3 Or iteratively update to a preset number of iterations. When the judgment result is no, repeat steps S4-5-S4-7 until the judgment result is yes, stop iterative updating, and output the sensitivity matrix of the last iterative update.

[0102] In each iteration, the parameter S is calculated i The gradient of , and updates the sensitivity matrix along the reverse direction of the gradient until the loss function is lower than 1×10 -3 Or the preset number of iterations stops the loop and outputs the optimal sensitivity matrix S i The damage function here can select other loss functions suitable for graph similarity evaluation, which can further improve the effect of sensitivity matrix optimization.

[0103] The defect image set is reconstructed according to the set of normalized pixel gray value vectors corresponding to the sensitivity matrix obtained by finally stopping the loop.

[0104] S5. Construct a defect location model for detecting defects in carbon fiber cable.

[0105] S5-1. Improve the U-Net model.

[0106] like Figure 4 As shown in the figure, the double convolution of the U-Net deep learning network is replaced by a residual module. The input of each residual module flows to two branches, one of which passes through the first convolution layer, the first batch normalization layer, the activation function, the second convolution layer, and the second batch normalization layer in sequence, and the other branch passes through the first convolution layer. The outputs of the two branches are added and passed through the activation function to obtain the output of the residual module.

[0107] The residual module can accelerate the training of the defect localization model and improve the stability of the defect localization model to reduce the risk of overfitting.

[0108] S5-2. Label the defect image set obtained in step S3 to obtain a set of images with labeled defects.

[0109] Each defect image in the preprocessed defect image set obtained in step S3 is processed, and the pixel value at the defect position is set to 255, and the pixel value at the non-defect position is set to 1, so as to obtain an image set with marked defects.

[0110] Among them, the image with marked defects, the detection voltage and the defect image correspond to each other.

[0111] S5-3. Reconstruct the defect image of step S3 using the initial voltage, the preprocessed detection voltage, and the iteratively updated sensitivity matrix.

[0112] The defect image is reconstructed by the LBP algorithm, and the set of normalized pixel gray value vector data of the defect image corresponding to the sensitivity matrix updated after the last iteration G(S i ) Each column of data is the pixel value corresponding to a defect image, and the defect image is reconstructed based on the pixel value.

[0113] S5-4. The images in the image set with marked defects and the images in the reconstructed defect image set are used as samples and input into the improved U-Net model for training to obtain a defect localization model for detecting intrinsic defects of carbon fiber material cables.

[0114] First, transfer learning is used to pre-train the weights of the defect localization model on the ImageNet large dataset and iterative training is performed. Then, the images with labeled defects and the reconstructed defect images are input into the defect localization model after the first training for training. The hyperparameters of model training are set as follows: the iteration cycle is 100 epochs, the batch size is set to 16, the initial learning rate is set to 0.0001, and the learning rate is adaptively adjusted according to the batch size and optimizer type. The loss function uses two loss functions: cross entropy loss function (Cross Entropy Loss) and Dice coefficient, and the Adam optimizer is used to optimize the learning parameters.

[0115] S6. Set up a detection probe, which has the structure of: a plurality of annular coils pointing to the center of the circle are arranged on the inner ring surface of the circular carrier, which are divided into detection coils and excitation coils. The number of the detection coils is the same as the number of the excitation coils, and they are evenly distributed in a double-layer ring around the object field, the inner coil is arranged as the detection coil, and the outer coil is arranged as the excitation coil; the detection probe moves from one end of the carbon fiber cable to be tested to the other end at a preset speed, completes the collection of the detection voltage every 300 milliseconds, and obtains a collection of the detection voltages of the carbon fiber cable to be tested.

[0116] like Figure 5As shown, the annular carrier in the present invention is a circuit board, and front-end circuits such as filtering and amplification are designed on the circuit board to reduce the influence of noise and weak signals, and control the excitation and collection of the coil. The excitation coil and the detection coil are both annular coils wound by copper wire. The distribution positions of the excitation coil and the detection coil are relatively, and the inner diameters of the detection coil and the excitation coil are both 5mm, and the outer diameters are both 10mm. The outer excitation coil is fixed on the circuit board, and an insulating material is used to pass through the annular detection coil and the excitation coil, and the central axis thereof is fixed on the same straight line. The intermediate object field area surrounded by the coil arrangement can be placed in the carbon fiber cable to be tested. The number of detection coils and excitation coils can be defined by oneself, and the number of detection coils and excitation coils depends on the imaging accuracy of the object field, the size of the object field and the coil arrangement. The number of detection coils and excitation coils of the present invention is 8.

[0117] The detection probe is placed on the carbon fiber cable to be tested, and the detection probe is perpendicular to the axial direction of the carbon fiber cable to be tested. The excitation current of the excitation coil is set to 1A, and the excitation frequency is set to 1MHz. The voltage of the detection coil is collected and recorded as the detection voltage. The detection probe moves from one end of the carbon fiber cable to the other end at a preset speed, and the detection voltage is collected every 300 milliseconds.

[0118] The preset speed of the present invention may be 10 mm / s.

[0119] S7. According to the set of detection voltages of the tested carbon fiber cable and the iteratively updated sensitivity matrix, the grayscale value of each pixel in the cross section of the tested carbon fiber cable is calculated, and a cross-sectional image of the tested carbon fiber cable is generated according to the grayscale value of each pixel. The cross-sectional image of the tested carbon fiber cable is input into the trained defect location model, and a clear cross-sectional image of the tested carbon fiber cable is output. When a defective cross-sectional image exists, the defect position is determined by the acquisition time of the detection voltage corresponding to the defective cross-sectional image.

[0120] Calculating defect imaging requires the detection voltage of a defect-free carbon fiber cable, which can be obtained by applying excitation to the model or by applying electrical excitation to the actual object (i.e., a defect-free carbon fiber cable). It can be the initial voltage of step S1 or the detection voltage of a defect-free carbon fiber cable with the same specifications as the carbon fiber cable being tested.

[0121] According to the detection voltage of the tested carbon fiber cable, the detection voltage of the defect-free carbon fiber cable and the iteratively updated sensitivity matrix, the grayscale value of each pixel of the cross section of the tested carbon fiber cable is calculated using formula (7), and each detection voltage collected in the set of detection voltages of the tested carbon fiber cable is calculated once. According to the grayscale value of each pixel of the cross section of the tested carbon fiber cable, a cross-sectional image of the tested carbon fiber cable is generated, and the cross-sectional image of the tested carbon fiber cable is input into the trained defect location model to obtain a clear cross-sectional image. When there is a defect, a clear intrinsic defect imaging can be obtained.

[0122] When a cross-sectional image with defects appears, the location of the defects of the tested carbon fiber cable can be calculated according to the acquisition time of the detection voltage corresponding to the cross-sectional image.

[0123] like Figure 6 As shown in the figure, after the first imaging and reconstruction, the defect position and shape are clearer and the outline is also clearer. Figure 6 The comparison results show that the defects of the traditional LBP algorithm imaging are submerged in the artifacts, which can only show the approximate defect location and it is difficult to reflect the precise size and location information of the defect. However, the present invention can achieve accurate visualization of the defects of carbon fiber cable.

Claims

1. A method for electromagnetic tomography flaw detection of carbon fiber cables, characterized in that: The steps include: S1. Construct an electromagnetic simulation model of a carbon fiber cable according to the specifications of the tested carbon fiber cable, apply current excitation to the three-dimensional electromagnetic simulation model of the carbon fiber cable to obtain an initial voltage, and establish an initial sensitivity matrix of the material field region of the cross-section of the carbon fiber cable by a field quantity extraction method; S2. randomly set the shape and position of the defect in the simulation model, apply excitation to the simulation model with the defect, and obtain a detection voltage; according to the shape and position of the defect set in the simulation model, draw a cross-sectional image of the simulation model where the defect is located to obtain a defect image; Perform edge gradient filling on the defect image and add Gaussian noise to the detection voltage; S3. Repeat step S2 at least 600 times to obtain a defect image set and a detection voltage set of the electromagnetic simulation model of the carbon fiber cable; S4. Iteratively update the initial sensitivity matrix according to the initial voltage, the defect image set and the detection voltage set; S5. Construct a defect location model for detecting intrinsic defects in carbon fiber cable; S6. Setting a detection probe, the structure of which is that a plurality of annular coils pointing to the center of the circle are set on the inner ring surface of the annular carrier, which are divided into detection coils and excitation coils. The number of detection coils is the same as the number of excitation coils, and they are evenly distributed in a double-layer ring around the object field. The inner coil is set as the detection coil, and the outer coil is set as the excitation coil; the detection probe moves from one end of the carbon fiber cable to be tested to the other end at a preset speed, and completes the collection of detection voltage every 300 milliseconds to obtain a collection of detection voltages of the carbon fiber cable to be tested; S7. According to the set of detection voltages of the tested carbon fiber cable and the iteratively updated sensitivity matrix, the grayscale value of each pixel in the cross section of the tested carbon fiber cable is calculated, and a cross-sectional image of the tested carbon fiber cable is generated according to the grayscale value of each pixel. The cross-sectional image of the tested carbon fiber cable is input into the trained defect location model, and a clear cross-sectional image of the tested carbon fiber cable is output. When a defective cross-sectional image exists, the defect position is determined by the acquisition time of the detection voltage corresponding to the defective cross-sectional image.

2. The electromagnetic tomography flaw detection method for carbon fiber cables according to claim 1, characterized in that: The specific method of iteratively updating the initial sensitivity matrix in step S4 is: S4-1. Iteratively update the initial sensitivity matrix. The formula for iterative update is: Among them, S0 is the initial sensitivity matrix, S1 is the matrix after the first iteration update, and G(S0) is the normalized pixel gray value vector data of the defect image corresponding to the sensitivity matrix S0; The calculation formula of G(S0) is: Among them, S0 T is the transposed matrix of S0, U * is the set of detection voltages obtained by randomly setting the defects of the simulation model, is the initial voltage; S4-2. Calculate the normalized pixel grayscale value vector data of the defect image corresponding to the iteratively updated sensitivity matrix, and the calculation formula is: S4-3. Calculate the loss function based on the normalized pixel gray value vector data of the defect image: Loss||G * -G(S1)|| Among them, G * A set of normalized pixel grayscale value vector data of defect images obtained by randomly setting defects of the simulation model; S4-4. Determine whether the loss function obtained in step S4-3 is less than 1×10 -3 Or iteratively update to a preset number of iterations. When the judgment result is yes, stop iterative updating and output the updated sensitivity matrix S1. When the judgment result is no, execute steps S4-5-S4-8; S4-5. Iteratively update the sensitivity matrix. The formula for iterative update is: Among them, S i is the sensitivity matrix obtained after the i-th iteration update, S i-1 is the sensitivity matrix obtained after the i-1th iteration update, a is the learning rate, indicating the step size of the iteration, G(S i-1 ) is the sensitivity matrix S i-1 Normalized pixel gray value vector data of the corresponding defect image; G(S i-1 ) is calculated as: Among them, S i T For S i The transposed matrix, U * is the set of detection voltages obtained by randomly setting the defects of the simulation model, is the initial voltage. S4-6. Calculate the normalized pixel grayscale value vector data of the defect image corresponding to the iteratively updated sensitivity matrix, and the calculation formula is: S4-7. Calculate the loss function based on the normalized pixel gray value vector data of the defect image: Loss||G * -G(S i )|| Among them, G * A set of normalized pixel grayscale value vector data of defect images obtained by randomly setting defects of the simulation model; S4-8. Determine whether the loss function output in step S4-7 is less than 1×10 -3 Or iteratively update to a preset number of iterations. When the judgment result is no, repeat steps S4-5-S4-7 until the judgment result is yes, stop iterative updating, and output the sensitivity matrix of the last iterative update.

3. The electromagnetic tomography flaw detection method for carbon fiber cables according to claim 1, characterized in that: The specific method of constructing the defect localization model is as follows: S5-1. Improve the U-Net model; S5-2. Annotate the defect image set obtained in step S3 to obtain an image set with annotated defects; S5-3. According to the updated sensitivity matrix and the detection voltage set, the pixel value of each pixel in the defect image set obtained in step S3 is recalculated to reconstruct the defect image set; S5-4. The images in the image set with marked defects and the images in the reconstructed defect image set are used as samples and input into the improved U-Net model for training to obtain a defect localization model for detecting intrinsic defects of carbon fiber material cables.

4. The electromagnetic tomography flaw detection method for carbon fiber cables according to claim 3, characterized in that: The specific way to improve the U-Net model is to replace the double convolution of the U-Net deep learning network encoding part with a residual module. The structure of each residual module is as follows: the input of the residual module is divided into two paths, one path passes through the first convolution layer, the first batch normalization layer, the activation function, the second convolution layer and the second batch normalization layer, the other path passes through 1×1 convolution, and the outputs of the two paths are added and output after passing through the activation function.

5. The electromagnetic tomography flaw detection method for carbon fiber cables according to claim 3, characterized in that: The iteration cycle of defect localization model training is 100 epochs, the batch size is set to 16, the initial learning rate is set to 0.0001, and the loss function uses the cross entropy loss function and the Dice coefficient.

Images