Composite material damage detection and positioning method based on neural network

Abstract

The invention provides a composite material damage detection and positioning method based on a neural network. The method comprises the following steps: acquiring a Lamb wave signal penetrating through a composite material; performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal; a damage detection Bayesian neural network model is constructed and trained, a trained damage detection Bayesian neural network model is obtained, and the trained damage detection Bayesian neural network model comprises a plurality of Bayesian linear layers connected in sequence; and inputting the optimized frequency domain Lamb wave signal into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material. According to the method, the Lamb wave signals are optimized, then the trained damage detection Bayesian neural network model is used for matching to obtain the damage positions of the correspondingly optimized frequency domain Lamb wave signals, and the damage positions of the composite material can be accurately measured without arranging a large number of sensors or depending on baseline signals.

Description

Technical Field

[0001] The present invention relates to the field of composite material damage detection, and in particular to a composite material damage detection and positioning method based on a neural network. Background Art

[0002] Carbon fiber reinforced polymer (CFRP) composites are widely used in aerospace and other fields due to their excellent mechanical properties. In practical applications, CFRP composite panels are prone to varying degrees of damage inside the composite material due to factors such as low-speed impact and fatigue environment. This type of damage is often not detectable by the naked eye, but it will affect the mechanical properties of the material in actual use. Therefore, it is necessary to establish a reliable structural health monitoring (SHM) system to ensure the integrity and safety of the composite structure.

[0003] In the prior art, for composite materials, especially CFRP composite materials, ultrasonic Lamb waves are generally used to perform large-scale detection of internal damage of composite materials. Lamb waves have a long propagation distance, and when passing through the internal damage position of the composite material, corresponding waveform changes will occur, which will be reflected at the receiving sensor. It is sensitive to a variety of hidden damages and helps to detect and locate internal damage of composite materials. However, Lamb waves have dispersion and multimodal characteristics, and are easily interfered by interlayer and boundary reflection echoes. In order to achieve the analysis and processing of Lamb wave waveforms, it is often necessary to predetermine the baseline signal of the composite material when it is not damaged, and a large number of sensors need to be set to form a sensor array to ensure a good damage measurement effect. However, in practical applications, it is impossible to collect the baseline signal of the composite material in advance, and it is difficult to set too many sensors on the installed composite material, resulting in the practicality of the relevant schemes for detecting internal damage of composite materials using Lamb waves. In addition, traditional detection methods rely on manual experience to judge and detect damage indexes, which increases the dependence on expert experience and reliable physical models.

[0004] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a composite material damage detection and positioning method based on a neural network, aiming to solve the problem that the composite material damage detection and positioning method in the prior art relies on pre-collected baseline signals and a large number of sensors, resulting in poor practicality.

[0006] The first aspect of the present invention provides a composite material damage detection and location method based on a neural network, comprising the steps of:

[0007] Acquire Lamb wave signals passing through composite materials;

[0008] Performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal;

[0009] Constructing a damage detection Bayesian neural network model, and training the damage detection Bayesian neural network model to obtain a trained damage detection Bayesian neural network model, wherein the trained damage detection Bayesian neural network model includes a plurality of Bayesian linear layers connected in sequence;

[0010] The optimized frequency-domain Lamb wave signal is input into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material.

[0011] In one implementation, the damage detection Bayesian neural network model is trained to obtain a trained damage detection Bayesian neural network model, specifically including:

[0012] Acquiring Lamb wave data and damage location data corresponding to the Lamb wave data as a training data set;

[0013] After preprocessing, the training data set is input into the damage detection Bayesian neural network model to extract damage information from the Lamb wave data, and match the damage information with the damage location data;

[0014] The uncertainty of the matching result is verified to obtain a damage information-damage position matching relationship that meets a preset credibility, and the training of the damage detection Bayesian neural network model is completed.

[0015] In one embodiment, the training data set is preprocessed, including: adding Gaussian noise data and Poisson noise data to the training data set to verify the robustness and reliability of the trained damage detection Bayesian neural network model to noise that may occur in an actual engineering environment.

[0016] In one embodiment, the optimized frequency-domain Lamb wave signal is input into the trained damage detection Bayesian neural network model to obtain the damage location of the composite material, specifically including:

[0017] Inputting the optimized frequency-domain Lamb wave signal into the trained damage detection Bayesian neural network model to extract damage information of the optimized frequency-domain Lamb wave signal;

[0018] According to the damage information-damage position matching relationship that meets the preset credibility, a nonlinear mapping relationship between the damage information and the damage position is constructed, thereby determining a nonlinear mapping between the Lamb wave and the damage position;

[0019] According to the nonlinear mapping between the obtained Lamb wave and the damage position, based on the Lamb wave signal to be detected, the detection and positioning of the unknown damage is realized;

[0020] In one embodiment, the damage information includes damage index, flight time, energy amplitude and similarity.

[0021] In one embodiment, the uncertainty verification includes:

[0022] Arbitrary uncertainty verification, wherein the arbitrary uncertainty verification extracts damage location data matched by the damage information and actual damage location data corresponding to the damage information, and determines the confidence of the damage information-damage location matching relationship through the variance between the damage location data and the actual damage location data; and

[0023] Cognitive uncertainty verification, the cognitive uncertainty verification randomly deactivates neurons in the damage detection Bayesian neural network model, selects several damage location data matched with the damage information, and determines the confidence of the damage information-damage location matching relationship through the mean of the sum of squares of the several damage location data and the square difference of the mean.

[0024] In one embodiment, obtaining a Lamb wave signal passing through a composite material specifically includes:

[0025] determining a region of the composite material to be inspected;

[0026] A pair of sensors are arranged at two ends of the area to be detected as an excitation sensor and a receiving sensor respectively;

[0027] The excitation sensor emits a Lamb wave signal, and the receiving sensor receives and acquires the Lamb wave signal passing through the area to be detected as the Lamb wave signal passing through the composite material.

[0028] In one implementation, performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal specifically includes:

[0029] Extracting an intrinsic mode function of the Lamb wave signal;

[0030] Screening out the components corresponding to the noise part in the intrinsic mode function, and mapping the remaining intrinsic mode function to the wavelet function space for decomposition;

[0031] Based on a preset threshold, the best reconstructed signal is obtained after several decompositions;

[0032] Reconstructing a reconstructed Lamb wave signal based on the optimal reconstruction signal;

[0033] The reconstructed Lamb wave signal is subjected to discrete Fourier transform to obtain the optimized frequency-domain Lamb wave signal.

[0034] The second aspect of the present invention discloses an intelligent terminal, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the composite material damage detection and positioning method based on a neural network as described in any of the above items are implemented.

[0035] A third aspect of the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the composite material damage detection and positioning method based on a neural network as described in any of the above items.

[0036] The present invention provides a composite material damage detection and positioning method based on a neural network, comprising the steps of: obtaining a Lamb wave signal passing through a composite material; performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal; constructing a damage detection Bayesian neural network model, training the damage detection Bayesian neural network model to obtain a trained damage detection Bayesian neural network model, wherein the trained damage detection Bayesian neural network model includes a plurality of Bayesian linear layers connected in sequence; and inputting the optimized frequency domain Lamb wave signal into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material. The present invention first optimizes the Lamb wave signal, and then uses the trained damage detection Bayesian neural network model to match and obtain the damage position of the corresponding optimized frequency domain Lamb wave signal, without setting a large number of sensors or relying on baseline signals, and can accurately determine the damage position of the composite material. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 It is a flowchart of the steps of the composite material damage detection and positioning method based on neural network according to the present invention;

[0039] Figure 2 It is a schematic diagram of the implementation process of the composite material damage detection and positioning method based on neural network according to the present invention;

[0040] Figure 3This is a schematic diagram of the architecture of the Bayesian neural network model for damage detection according to the present invention;

[0041] Figure 4 A schematic diagram for comparing different Lamb wave signals received by the composite material damage detection and positioning method based on a neural network according to the present invention;

[0042] Figure 5 It is a schematic diagram of the changes of the training loss and the verification loss of the damage detection Bayesian neural network model of the present invention with the number of training times in the process of damage position detection and positioning;

[0043] Figure 6 It is a schematic diagram comparing the detection and positioning results of the damage detection Bayesian neural network model of the present invention and CNN and LSTM for the damage position of the composite material;

[0044] Figure 7 Schematic diagram of the intelligent terminal of the present invention. DETAILED DESCRIPTION

[0045] The present invention provides a composite material damage detection and location method based on a neural network. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] Lamb waves have the characteristics of high penetration, long propagation distance, low attenuation, low cost and sensitivity to various defects in composite materials. They have broad application prospects in the field of damage detection of composite materials, especially CFRP composite materials. However, Lamb waves have dispersion characteristics when propagating inside composite materials. When encountering damaged positions, reflection, scattering, transmission and other phenomena will occur. When encountering interlayers or boundaries, reflection, refraction and other phenomena will also occur, resulting in the final collected wave signal being very complex, and it is difficult to directly apply the damage identification method based on time of flight (ToF) or other damage identification methods. Therefore, the Lamb wave composite material damage determination technology used in the prior art often needs to rely on the baseline signal, that is, the healthy signal when the composite material is intact as a benchmark, and it is necessary to set multiple sensors around the area to be detected to form a sensor array in order to determine the internal damage position of the composite material. However, this method is obviously not very practical and is not conducive to the further application of Lamb waves in the field of composite material damage determination. In view of this problem, the present invention discloses a composite material damage detection and positioning method based on a neural network. First, the acquired Lamb wave signal is optimized to reduce noise interference in the Lamb wave signal. Then, a trained damage detection Bayesian neural network model is used to extract damage information in the optimized frequency domain Lamb wave signal. The damage position of the composite material is determined based on the damage information matching, thereby completing the composite material damage measurement. There is no need to set up too many sensors and it does not rely on pre-collected baseline signals. The composite material damage detection and positioning method based on a neural network of the present invention can efficiently and accurately measure the internal damage position of the composite material, which is conducive to the further application of Lamb waves in the field of composite material damage measurement.

[0047] Specifically, Figure 1 As shown, the composite material damage detection and positioning method based on neural network of the present invention comprises the following steps:

[0048] S100. Acquire a Lamb wave signal passing through the composite material.

[0049] Taking advantage of the characteristics of Lamb wave, such as long propagation distance, low attenuation, low cost, and sensitivity to various defects, Lamb wave signals can be used to detect various types of damage (delamination, debonding, cracks, holes, etc.). The present invention first optimizes the Lamb wave signal to eliminate noise interference, obtains its frequency domain information based on discrete Fourier transform, extracts the damage information of the optimized frequency domain Lamb wave signal using a trained damage detection Bayesian neural network model, and finally outputs the damage position corresponding to the damage information, thereby realizing accurate detection and positioning of the damage position of the composite material under the condition of limited Lamb wave signal data, greatly improving the practicality of Lamb wave in the field of composite material damage determination.

[0050] Specifically, step S100 includes:

[0051] S110, determining a region of the composite material to be inspected;

[0052] S120, setting a pair of sensors at two ends of the area to be detected as an excitation sensor and a receiving sensor respectively;

[0053] S130, the excitation sensor emits a Lamb wave signal, and the receiving sensor receives and acquires the Lamb wave signal passing through the area to be detected as the Lamb wave signal passing through the composite material.

[0054] The composite material damage detection and positioning method based on neural network described in the present invention only uses a pair of sensors as an excitation sensor and a receiving sensor respectively. Therefore, when the excitation sensor transmits a Lamb wave signal, the receiving sensor will not only receive the Lamb wave signal passing through the area to be detected, but also receive the Lamb wave signal formed after the Lamb wave signal is reflected and refracted between the internal layers of the composite material, and after the reflection and refraction at the boundary of the composite material. If there is damage in the area to be detected, the Lamb wave signal received at the receiving sensor will further include a variety of scattered wave signals and reflected wave signals when passing through the damage position. Therefore, the composite material damage detection and positioning method based on neural network described in the present invention also needs to further optimize the acquired Lamb wave signal. Optionally, if Figure 2 As shown, the excitation sensor transmits the original input signal to the receiving sensor through the composite material, and then enters the next step after passing through the oscilloscope.

[0055] Furthermore, if Figure 1 As shown, the composite material damage detection and location method based on neural network of the present invention further includes, after step S100:

[0056] S200, performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal.

[0057] Specifically, the present invention combines EMD and wavelet transform to remove high-frequency noise in the signal, thereby obtaining a reconstructed Lamb wave signal, so as to reduce the interference of noise on the Lamb wave signal, and screen out the clutter caused by scattering, refraction or transmission when the Lamb wave signal propagates in the composite material; and then further optimizes the Lamb wave signal through discrete Fourier transform to obtain a frequency domain Lamb wave signal, so as to ensure more accurate measurement results in the subsequent damage measurement process of the composite material.

[0058] Specifically, step S200 includes:

[0059] S210, extracting the intrinsic mode function of the Lamb wave signal;

[0060] S220, filtering out the components corresponding to the noise part in the intrinsic mode function, and mapping the remaining intrinsic mode function to the wavelet function space for decomposition;

[0061] S230, performing several decompositions based on a preset threshold to obtain an optimal reconstructed signal;

[0062] S240, reconstructing a reconstructed Lamb wave signal based on the optimal reconstructed signal;

[0063] S250, performing discrete Fourier transform on the reconstructed Lamb wave signal to obtain the optimized frequency-domain Lamb wave signal.

[0064] Among them, the EMD method involves orthogonality and a single intrinsic component obtained after a recursive process. It decomposes the complex signal according to its time domain and identifies all intrinsic oscillation modes in the complex motion. In general, any signal is composed of different intrinsic mode functions (IMFs). The linear or nonlinear model of the extreme value and zero crossing of each point is the same. Any continuous crossing point has only one extreme value. Each mode should be independent of other modes. Complex signals can be decomposed into a finite number of IMF signals through EMD. Each component contains the characteristics of the original signal at different time scales. IMF must meet two conditions: first, the number of local extreme points and zero crossing points must be equal or at most one difference over the entire time range of the function. Second, the local IMF values ​​are symmetrical and have a mean of zero. The essence of the EMD method is to decompose the complex signal into a finite number of IMFs, and the IMF components are obtained by forcing a stable signal. For the Lamb wave signal X(t), after finding all the extreme points, the spline difference function e is used to calculate the local extreme value and zero crossing points. min (t) and e max (t) Calculate the average envelope function:

[0065] m(t)=(e max (t)+e min (t)) / 2.

[0066] The Lamb wave signal X(t) minus the average envelope function m(t) is the characteristic mode function IMF i (t), if the characteristic mode function cannot be successfully obtained, repeat the above steps to obtain:

[0067]

[0068] Where R(t) is the residual component, that is, the part that needs to be filtered out. In this way, the high-frequency noise in the Lamb wave signal can be effectively filtered out through EMD, thereby improving the accuracy of the signal and ensuring more accurate subsequent detection and positioning of composite material damage.

[0069] After determining the residual components and characteristic mode functions corresponding to the high-frequency noise in the original Lamb wave signal through EMD, the corresponding components are screened out, and then the wavelet transform maps the Lamb wave signal after EMD denoising to the wavelet function space, and decomposes the signal into two parts: large-scale approximation and detailed representation. The large-scale approximation part is further decomposed according to the definition to obtain a larger-scale approximation and finer detail resolution. In this way, by iterating multiple segmentations and selecting the corresponding requirements of the transform part, the denoising of the Lamb wave signal can be completed and the best reconstructed signal can be obtained. In this way, based on the best reconstructed signal, the reconstructed Lamb wave signal can be reconstructed. Optionally, the wavelet transform can reduce the autocorrelation of the signal, and then perform threshold analysis on the decomposed wavelet coefficients, adjust the wavelet coefficients within a specific threshold, and thus obtain the best reconstructed signal. Further, as Figure 2 As shown, after obtaining the reconstructed Lamb wave signal, a discrete Fourier transform (DFT) is performed to transform the time domain signal into a frequency domain signal, thereby reducing the amount of data in the signal while retaining the signal characteristics, so as to obtain an optimized frequency domain Lamb wave signal, and the optimized frequency domain Lamb wave signal is applied to subsequent neural network model training and final composite material damage location determination to ensure the training results and the final damage determination results.

[0070] Furthermore, if Figure 1 As shown, the composite material damage detection and location method based on neural network of the present invention further includes, after step S200:

[0071] S300, constructing a damage detection Bayesian neural network model, and training the damage detection Bayesian neural network model to obtain a trained damage detection Bayesian neural network model, wherein the trained damage detection Bayesian neural network model includes a plurality of Bayesian linear layers connected in sequence.

[0072] The present invention constructs and trains a Bayesian neural network model for damage detection, directly extracts damage information from Lamb wave signals, matches the damage information with the damage position, gives a confidence interval for the matching result through a probability model, and outputs the damage position that meets the preset credibility as the matched damage position, thereby achieving accurate detection and positioning of the damage position of the composite material.

[0073] The damage detection Bayesian neural network model of the present invention uses the weight parameter as a random variable and applies the Bayesian theorem to iteratively update the posterior distribution of the random variable. The probability of event A occurring when event B occurs is expressed as:

[0074]

[0075] In this way, through the likelihood probability P(B|A), based on our prior probability P(A) of the possibility of event A and the probability of event B P(B), we can get the posterior probability P(A|B).

[0076] In this process, we use maximum likelihood estimation (MLE) to determine the parameters in the likelihood function that maximizes the observed data:

[0077] θ * = argmax(P(D|θ));

[0078] The maximum a posteriori (MAP) estimate, which combines the prior probability and the likelihood probability, is used to determine:

[0079] θ * =argmax(P(θ|D))=argmax(P(D|θ)P(θ));

[0080] Where θ is the model parameter and D is the training data set.

[0081] The Bayesian neural network model for damage detection in the present invention solves all posterior probabilities P(θ|D), rather than the maximum posterior probability. Therefore, for the limited training data set D={(X 1 ,Y 1 ),(X 2 ,Y 2 ),…,(X m ,Y m )}, (m = 1, ..., n), where X is the training input data and Y is the predicted output data. The Bayesian neural network model for damage detection of the present invention can be expressed as:

[0082]

[0083] So after substituting into the total probability formula, we can get:

[0084]

[0085] Since the parameters are continuous values, the Bayesian neural network model for damage detection in the present invention can be further expressed as:

[0086]

[0087] However, in this case, no analytical solution can be obtained. Therefore, the damage detection Bayesian neural network model of the present invention uses variational inference (VI) to perform approximate calculations, that is, replacing the complete posterior distribution with a simpler distribution form, and then finding the parameters that make this simple distribution form closest to the posterior distribution, and solving this parameter by minimizing the difference between the two distributions. Specifically, assuming that the model is p(x, θ), where θ is the model parameter and x is the observed data, then variational inference finds a simpler distribution q φ (θ) makes it closest to the posterior distribution p(θ|x), that is, the optimization parameter φ is calculated by the following formula:

[0088]

[0089]

[0090] Where KL(q φ (θ)||p(θ|x)) is the KL divergence between the two distributions, and Represents for q φ Considering that P(x) is an omitable constant, the optimization problem is transformed into solving:

[0091]

[0092] In this way, an analytical solution can be obtained through optimization algorithms such as gradient descent method, thereby outputting the damage location of the composite material.

[0093] Specifically, Figure 3 As shown, the Bayesian neural network model for damage detection in the present invention is composed of As the input layer, the obtained optimized frequency domain Lamb wave signal is transmitted to a series of Bayesian linear layers. In this embodiment, a total of three Bayesian linear layers (Bayesian linear layer-1, Bayesian linear layer-2 and Bayesian linear layer-3) are set, and each Bayesian linear layer is connected by a Softplus function and a Dropout function, and finally the coordinate information of the corresponding composite material damage position is output. By setting a number of Bayesian linear layers connected in sequence, it can be ensured that the required damage information is extracted from the Lamb wave signal, and multiple uncertainty verifications can be performed at the same time, the uncertainty assessment effect is enhanced, and the stability of the overall model is increased. In this embodiment, through the closely connected Bayesian linear layers, it is not only guaranteed that the key damage information is obtained, but also that efficient learning can be carried out on the basis of ensuring the quantification of uncertainty, thereby ensuring the stability and reliability of the composite material damage position detection and positioning results.

[0094] Furthermore, the training process of the damage detection Bayesian neural network model includes:

[0095] H100, obtaining Lamb wave data and damage location data corresponding to the Lamb wave data as a training data set;

[0096] H200, after preprocessing the training data set, input the data set into the damage detection Bayesian neural network model, extract damage information from the Lamb wave data, and match the damage information with the damage location data;

[0097] H300 verifies the uncertainty of the matching result, obtains the damage information-damage location matching relationship that meets the preset credibility, and completes the training of the damage detection Bayesian neural network model.

[0098] Specifically, the Lamb wave data is obtained in the same manner as the Lamb wave signal in step S100, and a pair of sensors are used as an excitation sensor and a receiving sensor to obtain the Lamb wave signal passing through the area to be detected of the composite material, and then the corresponding Lamb wave data is obtained after the optimization process in step S200. Optionally, the optimized frequency domain Lamb wave signal is divided into a training set, a validation set and a test set according to a preset ratio, wherein the optimized frequency domain Lamb wave signal divided into the training set is the Lamb wave data, and the training data set of the damage detection Bayesian neural network model is formed in combination with the corresponding damage location data; and the optimized frequency domain Lamb wave signal divided into the validation set is used to perform performance verification and parameter adjustment on the damage detection Bayesian neural network model after training, and the optimized frequency domain Lamb wave signal divided into the test set is used to perform performance evaluation on the damage detection Bayesian neural network model after training, so as to ensure that a trained damage detection Bayesian neural network model that meets the requirements is finally obtained.

[0099] Optionally, before training the damage detection Bayesian neural network model, preprocessing the training data set includes adding Gaussian noise data and Poisson noise data to the training data set to verify the robustness and reliability of the trained damage detection Bayesian neural network model to noise that may appear in an actual engineering environment. The Gaussian noise data satisfies the Gaussian distribution, while the Poisson noise data satisfies the Poisson distribution, so as to simulate the noise data that may actually appear in reality and improve the anti-noise performance of the trained damage detection Bayesian neural network model.

[0100] Furthermore, the uncertainty verification includes arbitrary uncertainty verification and cognitive uncertainty verification, wherein the arbitrary uncertainty verification extracts the damage location data matched by the damage information and the actual damage location data corresponding to the damage information, and determines the confidence of the damage information-damage location matching relationship through the variance between the damage location data and the actual damage location data; while the cognitive uncertainty verification randomly deactivates neurons in the damage detection Bayesian neural network model, selects several damage location data matched by the damage information, and determines the confidence of the damage information-damage location matching relationship through the difference between the mean of the sum of squares of the several damage location data and the square of the mean.

[0101] Specifically, the arbitrary uncertainty reflects the variables and noise that naturally exist in the data, and reflects the deviation caused by sensors or environmental factors. Therefore, through the arbitrary uncertainty verification, the robustness of the damage detection Bayesian neural network model of the present invention and the damage determination effect in a real environment can be evaluated. Specifically, the mathematical form of the arbitrary uncertainty verification is:

[0102]

[0103] Where N is the number of samples, y i represents the real coordinates, f represents the Bayesian neural network model for damage detection according to the present invention, and the corresponding output is {f(x i ),log(σ 2 )}. In the loss function, f(x i ) is the predicted value of the model, log(σ 2 ) is the random uncertainty of the model on the data, that is, the logarithm of the data variance. The part before the plus sign in the formula corresponds to the residual in the regression model when matching the damage information with the damage location to capture the uncertainty associated with the model parameters; and the part after the plus sign It acts as a regularization term to prevent the uncertainty of the model output from being too high.

[0104] Specifically, epistemic uncertainty reflects the deviation caused by incomplete knowledge, that is, too little data, and can reflect the model's immunity to noise. Since the composite material damage detection and positioning method based on neural network described in the present invention only uses a pair of sensors to receive data, the evaluation of epistemic uncertainty is very important. In the Bayesian neural network model for damage detection described in the present invention, the uncertainty of the model is determined by observing how the initial model parameters are updated with the training data during the training process. In order to quantify epistemic uncertainty, we apply a dropout scheme, that is, in each training batch, during forward propagation, the activation values ​​of some neurons are stopped with a certain probability, so that the model generalizes more strongly and avoids dependence on certain local features. Specifically, the mathematical form of epistemic uncertainty verification is:

[0105]

[0106]

[0107] Among them, x * is the current input, f represents the Bayesian neural network model for damage detection according to the present invention, and the corresponding output is f(x * ,ω). Perform T forward propagations, each time applying the dropout scheme, and the output at the tth time is f t (x * ,ω t ). For all forward propagation steps, E[f(x * )] is the expected result of the prediction, that is, the average value of all outputs; Var[f(x * )] is the variance of the prediction, i.e., epistemic uncertainty.

[0108] Optionally, after arbitrary uncertainty verification and cognitive uncertainty verification, 90% is selected as the preset credibility, and only the damage location data and damage data whose uncertainty verification results fall within this range meet the requirements. The corresponding damage information-damage location matching relationship replaces the original parameters of the damage detection Bayesian neural network model described in the present invention to complete the training of the damage detection Bayesian neural network model.

[0109] Furthermore, if Figure 1 As shown, the composite material damage detection and location method based on neural network of the present invention further includes, after step S300:

[0110] S400, inputting the optimized frequency-domain Lamb wave signal into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material.

[0111] Specifically, after the optimized frequency domain Lamb wave signal is input into the trained Bayesian neural network model for damage detection, it passes through multiple Bayesian linear layers, and the corresponding damage information is extracted by the Bayesian linear layer, and a nonlinear mapping relationship between the damage information and the damage position is constructed according to the damage information-damage position matching relationship that meets the preset credibility, so as to determine the nonlinear mapping between the Lamb wave and the damage position, and then determine the damage position corresponding to the Lamb wave signal according to the nonlinear mapping, so as to realize the detection and positioning of the damage position in the area to be detected of the composite material. Specifically, the damage information includes damage index (DI), flight time (ToF) and energy amplitude (EnergyAmplitude), which are adaptively extracted from the Lamb wave based on the deep learning network. Specifically, after determining the nonlinear mapping between the Lamb wave and the damage position, the position coordinates of the damage position corresponding to the damage information on the composite material are determined by a regression algorithm, so as to locate the damage position, so as to realize the rapid and accurate detection and positioning of the internal damage of the composite material. Optionally, the present invention combines a Monte Carlo simulation algorithm to sample and calculate the average value and standard deviation of damage information to obtain the corresponding damage position on the composite material.

[0112] Based on the composite material damage detection and positioning method based on neural network, the present invention builds a composite material damage detection and positioning system based on neural network, including a pair of sensors, a Lamb wave generator (function generator) and an oscilloscope connected to the sensors, and a processing system connected to the oscilloscope, wherein the processing system includes:

[0113] A signal acquisition module, used for acquiring Lamb wave signals passing through the composite material;

[0114] A signal optimization module, used for performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal;

[0115] A model training module, used for constructing a damage detection Bayesian neural network model, training the damage detection Bayesian neural network model, and obtaining a trained damage detection Bayesian neural network model, wherein the trained damage detection Bayesian neural network model includes a plurality of Bayesian linear layers connected in sequence;

[0116] The data output module is used to input the optimized frequency domain Lamb wave signal into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material.

[0117] Furthermore, the present invention utilizes the neural network-based composite material damage detection and positioning system to perform performance testing on the neural network-based composite material damage detection and positioning method of the present invention, wherein the accuracy of composite material damage position detection and positioning is compared using a convolutional neural network (CNN) and a long short-term memory artificial neural network (LSTM) with the damage detection Bayesian neural network model of the present invention.

[0118] The experiment uses superimposed CFRP composite plates as the composite materials to be tested, wherein the CFRP composite plates themselves are not damaged, so metal weights are placed on the surface of the composite plates to simulate the effect of internal damage of the composite materials on the propagation of Lamb waves in the composite materials, so that the position of the metal weights can correspond to the damage position to be detected. The surface of the composite plate is divided into multiple squares, and the placement position of the metal weights is recorded and compared with the damage positions measured by the damage detection Bayesian neural network model, CNN and LSTM of the present invention, and the relative position relationship between the results is observed to determine the measurement accuracy of different models.

[0119] like Figure 4 The figure shows a schematic diagram comparing the Lamb wave passing through the undamaged area and the Lamb wave passing through the damaged area, where the light-colored lines are the Lamb wave waveforms passing through the undamaged area, and the dark-colored lines are the Lamb wave waveforms passing through the damaged area. It can be seen that after passing through the damaged area, due to the existence of scattering and reflection, not only the waveform of the direct Lamb wave changes, but also the subsequent boundary reflection wave and scattered wave will change, making the overall Lamb wave waveform more complex and containing more damage information related to the damage location.

[0120] like Figure 5 The figure shows a comparison diagram of the loss function change of the Bayesian neural network model for damage detection of the present invention during the training process and the verified loss function change. It can be seen from the figure that the loss function of the Bayesian neural network model for damage detection of the present invention tends to be stable after about 20 batches (epochs) of training, and the difference between the loss function change and the verified loss function change decreases rapidly, and remains stable in the subsequent training process. Therefore, it can be seen that the Bayesian neural network model for damage detection of the present invention has strong feature extraction capabilities, robustness against noise, learning capabilities, and model adaptability, and can quickly complete training, thereby realizing real-time detection and positioning of composite material damage positions.

[0121] like Figure 6The figure shows a schematic diagram comparing the accuracy of the damage detection Bayesian neural network model of the present invention, CNN and LSTM for damage position detection and positioning results. It can be seen from the figure that the damage positions determined by the damage detection Bayesian neural network model of the present invention are closer to the actual damage positions, and have the best visualization effect compared with CNN and LSTM, reflecting a better composite material damage position determination capability. Therefore, the composite material damage detection and positioning method based on neural network of the present invention utilizes the optimized frequency domain Lamb wave signal, combined with the damage detection Bayesian neural network model, to achieve rapid and accurate detection and positioning of the composite material damage position, without the need to collect baseline signals in advance or to arrange multiple sensors. Only a pair of sensors can be used to detect and position the damage position, which is conducive to the further application of Lamb waves in the field of composite material damage determination.

[0122] Based on the above-mentioned composite material damage detection and positioning method based on neural network, the present invention also provides an intelligent terminal, such as Figure 7 As shown, the intelligent terminal includes at least one processor, a display screen and a memory, and may also include a communication interface and a bus. The processor, the display screen, the memory and the communication interface can communicate with each other through the bus. The display screen is configured to display a preset user guide interface in the initial setting mode. The communication interface can transmit information. The processor can call the logic instructions in the memory to execute the method in the above embodiment.

[0123] In addition, the logic instructions in the memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0124] The memory, as a computer-readable storage medium, can be configured to store software programs, computer executable programs, such as program instructions or modules corresponding to the methods in the embodiments of the present invention. The processor executes functional applications and data processing by running the software programs, instructions or modules stored in the memory, that is, implementing the methods in the above embodiments.

[0125] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory. For example, a variety of media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, may also be a transient storage medium.

[0126] In addition, the specific process of loading and executing the multiple instruction processors in the above-mentioned storage medium and the terminal device has been described in detail in the above-mentioned method, and will not be described one by one here.

[0127] In summary, the present invention provides a composite material damage detection and positioning method based on a neural network, comprising the steps of: obtaining a Lamb wave signal passing through a composite material; performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal; constructing a damage detection Bayesian neural network model, training the damage detection Bayesian neural network model to obtain a trained damage detection Bayesian neural network model, wherein the trained damage detection Bayesian neural network model includes a plurality of Bayesian linear layers connected in sequence; and inputting the optimized frequency domain Lamb wave signal into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material. The present invention first optimizes the Lamb wave signal, and then uses the trained damage detection Bayesian neural network model to match the corresponding optimized frequency domain Lamb wave signal to obtain the damage position of the composite material, without setting a large number of sensors or relying on baseline signals, and can accurately detect and locate the damage position of the composite material.

[0128] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features therein may be replaced by equivalents. However, these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A composite material damage detection and location method based on neural network, characterized in that: Includes steps: Acquire Lamb wave signals passing through composite materials; Performing noise reduction and discrete Fourier transform on the Lamb wave signal to obtain an optimized frequency domain Lamb wave signal; Constructing a damage detection Bayesian neural network model, and training the damage detection Bayesian neural network model to obtain a trained damage detection Bayesian neural network model, wherein the trained damage detection Bayesian neural network model includes a plurality of Bayesian linear layers connected in sequence; The optimized frequency-domain Lamb wave signal is input into the trained damage detection Bayesian neural network model to obtain the damage position of the composite material.

2. The composite material damage detection and location method based on neural network according to claim 1 is characterized in that: The damage detection Bayesian neural network model is trained to obtain a trained damage detection Bayesian neural network model, specifically including: Acquiring Lamb wave data and damage location data corresponding to the Lamb wave data as a training data set; After preprocessing, the training data set is input into the damage detection Bayesian neural network model to extract damage information from the Lamb wave data, and match the damage information with the damage location data; The uncertainty of the matching result is verified to obtain a damage information-damage position matching relationship that meets a preset credibility, and the training of the damage detection Bayesian neural network model is completed.

3. The composite material damage detection and location method based on neural network according to claim 2 is characterized in that: The training data set is preprocessed, including: adding Gaussian noise data and Poisson noise data to the training data set to verify the robustness and reliability of the trained damage detection Bayesian neural network model to noise.

4. The composite material damage detection and location method based on neural network according to claim 2 is characterized in that: Inputting the optimized frequency-domain Lamb wave signal into the trained damage detection Bayesian neural network model to obtain the damage location of the composite material specifically includes: Inputting the optimized frequency-domain Lamb wave signal into the trained damage detection Bayesian neural network model to extract damage information of the optimized frequency-domain Lamb wave signal; According to the damage information-damage position matching relationship that meets the preset credibility, a nonlinear mapping relationship between the damage information and the damage position is constructed, thereby determining a nonlinear mapping between the Lamb wave and the damage position; According to the nonlinear mapping, the damage position corresponding to the Lamb wave signal is determined, so as to detect and locate the damage position of the composite material.

5. The composite material damage detection and location method based on neural network according to claim 4 is characterized in that: Extracting damage information of the optimized frequency-domain Lamb wave signal includes: adaptively extracting the damage information based on a deep learning network, wherein the damage information includes a damage index, a flight time, and an energy amplitude.

6. The composite material damage detection and location method based on neural network according to claim 2 is characterized in that: The uncertainty verification includes: Arbitrary uncertainty verification, wherein the arbitrary uncertainty verification extracts the damage location data matched by the damage information and the actual damage location data corresponding to the damage information, and determines the confidence of the damage information-damage location matching relationship through the variance between the damage location data and the actual damage location data; Cognitive uncertainty verification, the cognitive uncertainty verification randomly deactivates neurons in the damage detection Bayesian neural network model, selects several damage location data matched with the damage information, and determines the confidence of the damage information-damage location matching relationship by the difference between the mean of the sum of squares of the several damage location data and the square of the mean.

7. The composite material damage detection and location method based on neural network according to claim 1 is characterized in that: Acquire Lamb wave signals through composite materials, including: determining a region of the composite material to be inspected; A pair of sensors are arranged at two ends of the area to be detected as an excitation sensor and a receiving sensor respectively; The excitation sensor emits a Lamb wave signal, and the receiving sensor receives and acquires the Lamb wave signal passing through the area to be detected as the Lamb wave signal passing through the composite material.

8. The composite material damage detection and location method based on neural network according to claim 1, characterized in that: The Lamb wave signal is subjected to noise reduction and discrete Fourier transform to obtain an optimized frequency domain Lamb wave signal, specifically comprising: Extracting an intrinsic mode function of the Lamb wave signal; Screening out the components corresponding to the noise part in the intrinsic mode function, and mapping the remaining intrinsic mode function to the wavelet function space for decomposition; Based on a preset threshold, the best reconstructed signal is obtained after several decompositions; Reconstructing a reconstructed Lamb wave signal based on the optimal reconstruction signal; The reconstructed Lamb wave signal is subjected to discrete Fourier transform to obtain the optimized frequency-domain Lamb wave signal.

9. An intelligent terminal, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the composite material damage detection and positioning method based on neural network as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by the processor, the steps of the composite material damage detection and positioning method based on neural network as described in any one of claims 1 to 8 are implemented.

Images