Low-speed impact positioning treatment method based on carbon fiber composite material structure

Abstract

The invention discloses a low-speed impact positioning processing method based on a carbon fiber composite material structure. The method specifically comprises the following steps: preparing a piezoelectric sensor based on PVDF-TrFE; fast arrangement based on FastSAM is carried out; performing low-speed impact test on the composite material; the invention relates to a composite material impact positioning algorithm based on a dynamic network. According to the method, a triaxial sliding table signal acquisition system is established, the arrangement positions of sensors are determined by adopting FastSAM, a plurality of piezoelectric sensors are prepared on a composite material mechanism based on a polyvinylidene fluoride and trifluoroethylene copolymer PVDF-TrFE coating technology, a data set is constructed by carrying out a low-speed impact experiment on the composite material structure, and a attention mechanism network is constructed and trained. According to the invention, based on a restive network method, rapid impact positioning can be realized, and compared with other leading edge deep learning network architectures, the method has obvious improvement in positioning precision, and also can be used to realize signal processing in other ultrasonic guided wave fields.

Description

Technical Field

[0001] The present invention relates to a low-speed impact positioning technology for composite materials, and particularly to a low-speed impact positioning processing method based on a carbon fiber composite material structure. Background Art

[0002] Structural Health Monitoring (SHM) based on Lamb waves is a real-time, in-situ damage identification technology for engineering structures based on condition-based maintenance. Due to the advantages of Lamb waves such as long propagation distance, easy driving and receiving, and sensitivity to defects, it has the ability to solve the impact positioning problem on plate and shell structures. However, the SHM data acquisition based on Lamb waves relies on sparsely distributed sensors to capture the wave responses in a large scanning area, as well as the dispersion and multimodal characteristics of Lamb waves, which makes the signals complex and makes their analysis and interpretation a difficult task for engineers.

[0003] Traditional impact positioning algorithms are mainly based on physical models and signal processing techniques. For example, the triangulation method based on arrival time measures the time difference of the shock wave arriving at different sensors, and combines the known wave speed to determine the impact source position using the triangulation principle. The Time Difference of Arrival (TDOA) method calculates the time difference of the shock wave arriving at different sensors and draws a time difference distribution map, and determines the impact source position by analyzing the characteristics in the map. However, due to the anisotropic material properties of composite materials, it is necessary to first calculate the group velocity dispersion curve corresponding to the composite material structure. After obtaining accurate group velocity information in each direction, it is usually necessary to extract the single-frequency components of the impact signal data based on the wavelet transform method and improve the above TDOA algorithm to determine the impact position. The time reversal method determines the impact source position by time-reversing and superimposing the collected impact signals and using the coherence of the waves to enhance the signals from the direction of the impact source. The disadvantage of this method is that it usually requires the process of active excitation of the sensors, which is not conducive to the realization of fast impact positioning. The triangulation method based on arrival time measures the time difference of the shock wave arriving at different sensors, and combines the known wave speed to determine the impact source position using the triangulation principle. The characteristic of this type of method is that at least 6 sensors are required, which is not conducive to the layout and wiring during the actual implementation of impact positioning. The above impact positioning usually also has accuracy problems, and the error level is usually centimeter-level.

[0004] Deep learning generally includes constructing input-output sample sets, building neural network architectures, and training and testing. It can learn the hidden features in complex signals in Lamb waves, thereby achieving higher-precision impact source localization. Retentive network is a new deep learning network architecture based on the retention mechanism, which simultaneously satisfies the advantages of low-cost inference, powerful performance, and training parallelization in the field of natural language processing and other fields, and is conducive to achieving high-precision impact localization of composite materials. In addition, deep learning models have a variety of different network architectures, training methods, and learning strategies. Therefore, the applicability of the signals generated by composite material impact localization on different networks is a problem worthy of exploration. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention provides a method for low-speed impact localization processing based on a carbon fiber composite material structure, which is specifically as follows:

[0006] Step 1: Preparation of a piezoelectric sensor based on PVDF-TrFE;

[0007] Through screen printing technology, a single-component silver paste is brushed on the electrode area on the upper surface of the composite board to form a bridge with two ends. A plurality of bridges are arranged at a plurality of different positions dispersed on the upper surface of the composite board, one bridge at each position; the orientation of the plurality of bridges depends on the convenience of connecting the bottom and top electrodes.

[0008] For any one bridge, a rectangular strip connects the two ends, and one end becomes the bottom electrode of the piezoelectric sensor; after curing the silver paste that forms four bridges, acetone and N-N dimethylformamide are mixed into a solution in equal volume ratio, and then PVDF-TrFE is added to form a PVDF-TrFE coating; after magnetic stirring, the PVDF-TrFE coating is sprayed as evenly as possible on the bottom electrode; during the spraying process, the PVDF-TrFE coating only covers one end of the bridge, does not cover the rectangular strip and the other end, and the edge of the PVDF-TrFE coating reaches or exceeds the covered end; after annealing and corona polarization of the PVDF-TrFE coating, silver paste screen printing is used again above the PVDF-TrFE coating, corresponding to the top electrode of the piezoelectric sensor, and the projection of the top electrode on the horizontal plane is surrounded by the projection area of the PVDF-TrFE coating. After the top electrode is also cured at high temperature, the three-layer structure of the piezoelectric sensor is completed. The three-layer structure is, from top to bottom, the top electrode formed by the upper cured silver paste, the PVDF-TrFE coating, and the bottom electrode formed by the lower cured silver paste; the top electrode is used as the positive electrode and the bottom electrode is used as the negative electrode; since the current preparation and verification of flexible sensors are mainly for metal materials.

[0009] Step 2: Fast arrangement based on FastSAM;

[0010] An industrial camera is used to obtain an image of a carbon fiber composite material. The carbon fiber composite material where the sensor is to be arranged is placed in the center of the lens. The sensor position area in the composite material plate image is quickly obtained by the fast segmentation large model FastSAM. The fast segmentation large model FastSAM performs automatic and fast regional segmentation on the carbon fiber composite material to obtain a binary image corresponding to the segmented area. The Sobel edge detection in Python is used to extract the edge points of the carbon fiber composite plate from the binary image; two K-means clustering with different clustering numbers are performed on the edge points, and the sensor layout points are obtained by weighting. The weighting formula is shown in the following formula (1). In formula (1), the coordinate direction index m is x or y, and x and y represent the horizontal and vertical coordinate directions of the carbon fiber composite sample image in the Cartesian coordinate system, respectively. When m is x, after K-means clustering with n clustering points, the horizontal coordinate value k is output. n (m), the abscissa value k1(m) corresponding to the number of cluster points being 1 constitutes the abscissa of the cluster, and sp(m) represents the abscissa and ordinate values of the Cartesian coordinate system; this method can arbitrarily set the initial layout points of the sensor, and the number of sensor layout points corresponds to the number of k-means clusters;

[0011] sp(m)=0.7*k n (m)+0.3*k1(m) (1)

[0012] Step 3: Low-velocity impact test of composite materials

[0013] The specific process is as follows:

[0014] Step 1: The LabVIEW program on the PC host computer sets the motion position of the three-axis slide, and the closed-loop motor controls the motion of the three-axis slide in the xyz space rectangular coordinate system;

[0015] Step 2: At the moving position of the three-axis slide, a millimeter-scale ball freely falls onto the composite material plate, causing an impact. The vibration is transmitted to the piezoelectric sensor prepared in step 1 above and converted into a voltage waveform through the piezoelectric effect. When the voltage exceeds the threshold voltage, a signal is recorded.

[0016] Step 3: The mechanical vibration is converted into an electrical signal by the piezoelectric sensor prepared above and collected by a data acquisition board. The collected signal is saved and recorded in the PC host computer;

[0017] Step 4: Repeat the above impact process at different motion positions to construct an impact dataset;

[0018] Step 4: Composite material impact location algorithm based on retentive network;

[0019] The input of the Retentive network is a two-dimensional matrix composed of eight-channel one-dimensional Lamb wave signals corresponding to eight sensors, and a deep learning model is established to directly output the xy-axis coordinates; the Retentive network algorithm is as follows: the eight-channel one-dimensional Lamb wave signals are dimension-reduced by a linear layer and then pass through a layernorm layer;

[0020] X L = layerNorm(X) (2)

[0021] In Equation (2), X represents the input of the eight-channel one-dimensional signal, and layerNorm(.) represents the operation of the layernorm layer, that is, Equation (2) represents that X has passed through the operation of the layernorm layer to obtain the layer-normalized output X L ;

[0022] Q, K, and V respectively represent Query, Key, and Value. After passing through the layer normalization, the input undergoes three different operations. The following Equations (3-5) represent that after multiplying the input passing through the layernorm layer by the training weight matrix, a relative position embedding operation is further performed; in the equations, xpos represents the relative position embedding;

[0023] Q = xpos(X L W Q ) (3)

[0024] K = xpos(X L W K ) (4)

[0025] V = X L W K (5)

[0026] In Equations (3, 4), W Q , W K both represent trainable weights;

[0027] The calculation output of the relative position embedding xpos Enter is X L W, that is, using X L W represents the sum of X L W Q and X L W K ;

[0028] In the following formulas, vectors are represented in the form of [..,..,...]; for the calculation of the relative position embedding of Q in Equation (3), the calculation of scale is obtained according to Equation (6). For the calculation of the relative position embedding of K, the calculation of scale needs to take the reciprocal of the result of Equation (6); using X L Weven and X L W odd represent the odd-indexed columns and even-indexed columns of X L W respectively, and W even and W odd represent the corresponding weights of the odd and even-indexed columns respectively; take the opposite of X L W even and then recombine it with X L W odd to form X L W * , which is implemented through the concat function of the pytorch framework, and W * represents the weight after taking the opposite; M [1，0.01] represents a tensor of dimension (256, 2), and each row of the tensor is [1, 0.01]; represents the corresponding multiplication relationship; use ⊙ to represent dot multiplication, take the corresponding sine and cosine values and perform dot multiplication with the above scale to obtain the sine value sin of the relative position embedding M and the cosine value cos of the relative position embedding M ; obtain the operation result xpos of the xpos operation through Equation (9) out ;

[0029] [[ID=3⑧]]

[0030]

[0031] xpos out =X L W⊙cos M +X L W * ⊙sin M (9)

[0032] Introduce the original gamma matrix d matrix shown in Equation (11). The introduced d matrix needs to ensure the correspondence of the calculation dimensions in Equation (12), that is, the dimension of the d matrix meets the basic requirements of the matrix calculation in Equation (13); set a one-dimensional tensor γ, and set the length of the tensor to n_head. The maximum and minimum values of the elements in the tensor are a and b respectively; each element in the γ tensor is used as the base in the exponential operation, and use γ i to represent the i-th element in γ; the d matrix calculation is shown in the following Equation (11);

[0033]

[0034] y i =(QK T ☉D i )V (13)

[0035] yi = Retention(X) (14)

[0036] where n_sequence is the first intermediate parameter value, and the gamma matrix D i and the sub-output variable y i represent the results after the corresponding exponentiation operation and the results further calculated by Equation (13), respectively; the above operations from Equation (2) to Equation (13) are summarized by Equation (14), that is, Retention(X) represents the entire process from the initial input to the output of the sigmoid module;

[0037] The output y i is part of the final output of the multi-scale retention module; the (i + 1)-th element in γ is used to perform the above operation of Equation (14) with the same input X, and the results such as y1, y2... y i are merged in the third dimension (implemented by concat), and then the final output of the multi-scale retention module is obtained through the operation of Equation (15);

[0038] y = (sigmoid(XW G ) ⊙ GroupNorm(concat(y1, y2,..., y i )))W o (15)

[0039] In the formula, W G and W O are both weight terms, with dimensions [hidden_size, 2*hidden_size] and [2*hidden_size, hidden_size] respectively, and hidden_size represents the second intermediate parameter value; y and X represent the final output of the multi-scale retention module and the input after signal dimensionality reduction respectively, and y1, y2,..., y i represent the results of the operations of elements 1 to i in γ through Equation (13), and sigmoid(.), GroupNorm(.), concat(y1, y2,..., y i ) represent the sigmoid activation function, the group normalization function, and the tensor concatenation operation respectively;

[0040] After the operation of the above Equation (15), the operation result is further passed through a layernorm layer, a linear layer, and a GELU layer to obtain the output result of the retnet model, that is, the impact location horizontal and vertical coordinate values of the CFRP plate are obtained.

[0041] In step 1 of an embodiment of the present invention, the two end portions are two rectangular end portions with different sizes. The four sides of the two rectangular end portions are parallel to each other respectively, and a pair of opposite sides thereof are parallel to the long side of the rectangular strip. Only the large rectangular end portion is covered with a PVDF-TrFE coating, and then the top electrode is covered continuously.

[0042] In step 1 of a specific embodiment of the present invention, the silver paste is cured by drying in a vacuum dryer at 80 °C for 1 hour; the screen used for screen printing is made of 250-mesh polyester fabric; acetone and N-N dimethylformamide with a volume ratio of 50 / 50 vol% are mixed into a solution in an equal volume ratio, and then 70 / 30 mol% of PVDF-TrFE is added to form a PVDF-TrFE coating; the mass of PVDF-TrFE is 5% of the mass of the solution; after magnetic stirring for 6 h, the solution is sprayed on the bottom electrode; the PVDF-TrFE coating also needs to be annealed at 140 °C and corona polarized at 45 kv.

[0043] In another specific embodiment of the present invention, the shape of the top electrode is circular, and the top electrode needs to be cured at a high temperature of 80 °C.

[0044] In another embodiment of the present invention, the carbon fiber composite material is a CFRP plate, and the positive and negative electrodes are connected to enameled wires.

[0045] In Step 2 of step 3 of yet another specific embodiment of the present invention, the threshold voltage is set to 0.02 V.

[0046] In step 3 of still another specific embodiment of the present invention, the millimeter-sized ball has a diameter of 7 mm and a weight of 7 g.

[0047] In step 3 of still another specific embodiment of the present invention, 205 discrete impact points are uniformly arranged at a spacing of 10 mm in the training area of 180×180 mm 2 on the experimental composite plate, and the free fall of the millimeter-sized ball is repeated 3 - 15 times at each point.

[0048] In step 4 of another specific embodiment of the present invention, the eight-channel one-dimensional Lamb wave signal is dimensionally reduced by a linear layer, and the dimension changes from 6144 to 256.

[0049] In addition, in step 4 of a specific embodiment of the present invention, n_sequence is set to 256 and hidden_size is set to 8.

[0050] The characteristics of the Retentive network algorithm are suitable for low-speed impact localization on composite materials. A three-axis sliding table signal acquisition system is built, and FastSAM is used to determine the sensor layout positions. Four piezoelectric sensors are prepared on the composite material mechanism based on the copolymer (PVDF-TrFE) coating technology of polyvinylidene fluoride and trifluoroethylene. A dataset is constructed by conducting low-speed impact experiments on the composite material structure, and a retention mechanism network is constructed and trained. Based on the Retentive network method, the present invention can achieve rapid impact localization, with obvious improvement in localization accuracy compared to other advanced deep learning network architectures, and can also be used to process signals in other ultrasonic guided wave fields through this method.

[0051] To verify the superiority of the composite material impact localization method based on the Retentive network, two different sensor combinations, namely the dual-sensor and four-sensor cases, are used to test the impact on the carbon fiber composite laminate to verify the localization effect. Brief Description of the Drawings

[0052] Figure 1 Shows the schematic diagram of the piezoelectric sensor structure based on PVDF-TrFE;

[0053] Figure 2 Shows the algorithm flow chart of the sensor position layout based on FastSAM;

[0054] Figure 3 Shows the sensor layout position diagram of the composite laminate;

[0055] Figure 4 Shows the experimental schematic diagram of the construction of the composite material foreign object impact damage dataset;

[0056] Figure 5 Shows the impact localization model based on the Retentive network;

[0057] Figure 6 Shows the overall impact localization effect diagram of the four sensors;

[0058] Figure 7 Shows the overall impact localization error distribution of the four sensors. Detailed Description of the Preferred Embodiment

[0059] The present invention will be described in detail below with reference to the accompanying drawings.

[0060] The present invention proposes a method for low-speed impact positioning and processing based on a carbon fiber composite material structure. This method uses the FastSAM method to determine the layout positions of piezoelectric sensors (hereinafter simply referred to as "sensors") for carbon fiber laminates, and finally achieves high-precision composite material impact positioning through the retentive network algorithm. The specific implementation is as follows:

[0061] Step 1: Preparation of piezoelectric sensors based on PVDF-TrFE;

[0062] The composite material plate with a self-made three-layer structure piezoelectric sensor is as Figure 1 shown. First, through screen printing technology, as shown in Figure 2 , a single-component silver paste is brushed on the designated electrode area on the upper surface of the composite plate (see Figure 1 area A, and its specific position will be obtained through the result of step 2), forming a bridge with two rectangular ends of different sizes as shown in the lower right. Additionally, Figure 1 there are a total of four bridges numbered 0 - 3, located at four relatively scattered points on the upper surface of the composite plate. The four bridges seemingly face different directions. The reason is that since the four sensors (corresponding to the four bridges) are fabricated at one time, when the screen printing of one sensor (corresponding to one bridge) has just ended and it has not yet solidified, the screen printing process at the positions of the other three sensors cannot damage the uncured sensor. Therefore, the orientation of the screen printing stencil or the orientation of the sample needs to be changed. Thus, the positions and angles of the four bridges are different. There is no other special requirement, only to avoid damaging the uncured sensor during the preparation process. Figure 1 Connected between the two rectangular ends is a rectangular strip. The four sides of the two rectangular ends are respectively parallel to each other, and a pair of opposite sides is parallel to the long side of the rectangular strip (the above description of the shape is mainly to separate the bottom and top electrodes for convenient connection of the bottom electrode, and this practice can also facilitate the identification of the bottom and top electrodes. Other shapes commonly used in the art can also be used). The silver paste is cured by drying in a vacuum dryer at 80°C for 1 hour. The screen used for screen printing is made of 250-mesh polyester fabric. Acetone (50 / 50 vol%) and N-N dimethylformamide (DMF, Sigma-Aldrich) are mixed into a solution in equal volume ratio, and then PVDF-TrFE (70 / 30 mol%) is added to form a solution. The mass of PVDF-TrFE is 5% of the mass of this solution. After 6 hours of magnetic stirring, the solution is sprayed on the bottom electrode. The solution is sprayed on the bottom electrode as evenly as possible. During the spraying process, the solution only covers a small part of the silver paste, corresponding to the lower left corner part of the sensor structure in the B area of

[0063] the lower right corner sensor structure in Figure 1 , that is, Figure 1The part enclosed by the black hollow rectangular frame shown in region B, where the outermost rectangular frame represents Figure 1 Region A in Figure 1 , which is a schematic diagram of the structure of the sensor top electrode position in the thickness direction. This approach is used to prevent PVDF-TrFE from completely covering the bottom electrode, enabling the bottom electrode to conduct wiring tests. After annealing at 140 °C and corona polarization at 45 kV, PVDF-TrFE exhibits good piezoelectric properties. After that, silver paste is screen-printed again above the PVDF-TrFE coating part corresponding to the top electrode of the self-made sensor. The shape of the top electrode is circular, and the projection of the area covered by the solution surrounds the projection of the top electrode on the horizontal plane. The top electrode is above the bottom electrode. Generally, its projection on the horizontal plane will fall within the range of the bottom electrode, and the projections of the two on the horizontal plane cannot be offset, so they do not coincide at all. After the top electrode is also cured at 80 °C, the three-layer structure of the piezoelectric sensor is completed. The three-layer structure from top to bottom is the top electrode formed by the cured silver paste on the upper layer, the PVDF-TrFE coating, and the bottom electrode formed by the cured silver paste on the lower layer. After preparation, the upper electrode serves as the positive electrode and the lower electrode serves as the negative electrode. Subsequently, the positive and negative electrodes are connected to enameled wires to facilitate signal output to the computer terminal.

[0064] Step 2: Fast layout based on FastSAM;

[0065] To obtain the initial position layout of the self-made sensor for PVDF-TrFE, the idea of using computer vision to solve engineering problems was adopted. The specific content of the method is as follows: First, the Fast Segment Anything Model (FastSAM) (https: / / arxiv.org / abs / 2306.12156) was introduced. FastSAM is pre-trained with a large amount of image data and can achieve image segmentation in milliseconds. The method is to first obtain pictures of the CFRP plate through an industrial camera, place the CFRP plate where the sensor needs to be arranged in the center of the lens. The picture quickly obtains the CFRP position area in the composite material plate image through the FastSAM model. The model segments the plate area and can quickly obtain the corresponding binary image. The binary image combined with Sobel edge detection (Sobel in the cv2 library in Python) can more accurately extract the edge of the composite material plate. This method calculates the image gray gradient and extracts the edge information in the image (a similar method is mentioned in "Xiaokang Zhang, Runping Han. The Application of Mathematical Morphology and Sobel Operator in Infrared Image Edge Detection. International Industrial Informatics and Computer Engineering Conference (IIICEC 2015). January 2015."). It solves the problems of redundancy or omission existing in traditional edge detection methods. Perform K-means clustering with two different numbers of clusters on the edge points (KMeans in the sklearn.cluster library in Python), and obtain the sensor layout points through weighting. The weighting formula is shown in the following formula (1). In formula (1), the coordinate direction index m is x or y. x and y respectively represent the horizontal and vertical coordinate directions of the CFRP sample image (such as the sample image in Figure 1 ) in the Cartesian coordinate system. When m is x, after K-means clustering with the number of clustering points being n, the output abscissa value is k n (m), and the abscissa value k1(m) corresponding to the number of clustering points being 1 forms the abscissa of the cluster (on the basis of edge detection, the edge is composed of many points, and after K-means clustering, it can become several points, and the number of these clustering points can be actively set). sp(m) represents the horizontal and vertical coordinate values of the Cartesian coordinate system (when m is y, sp(m) is the ordinate). Taking the number of clusters being 4 as an example, the implementation steps of FastSAM-kmeans are as follows Figure 2 as shown Figure 3is the clustering result, and the 4 star-shaped points are used as the positions for constructing 4 different sensors. This method can arbitrarily set the initial layout points of the sensors, and the number of sensor layout points directly corresponds to and is equal to the number of k-means clusters. Figure 2 The final output results also show the sensor layout positions for the three cases of the number of layout points being 2, 4, and 6. Considering the impact location accuracy and the characteristics of wired sensors, 4 sensors are considered to be prepared. Figure 3 The "sensor electrode position" in it is the preparation position of the final self-made sensor on the forming plate.

[0066] sp(m) = 0.7*k n (m) + 0.3*k1(m) (1)

[0067] Step 3: Low-velocity impact experiment on the composite material;

[0068] To make full use of the information in the impact acoustic emission signal and construct a corresponding damage deep learning model for subsequent monitoring, an impact damage dataset is constructed by building an impact experiment platform, and the process is as Figure 4 shown.

[0069] The specific process is as follows:

[0070] Step1: The movement position of the three-axis slide table is set by the labview program in the PC host computer. By controlling the STM32 development board, the development board further controls the closed-loop motor, and finally the closed-loop motor controls the movement of the three-axis slide table in the xyz space rectangular coordinate system; this process is well-known to those skilled in the art and will not be repeated here;

[0071] Step2: At the movement position of the above three-axis slide table, a millimeter-sized ball freely falls and impacts on the composite material plate, and the vibration propagates to the piezoelectric sensor prepared in the above Step 1 and is converted into a voltage waveform through the piezoelectric effect. The threshold voltage is set to 0.02V, and the signal is recorded when the voltage exceeds the threshold voltage;

[0072] Step3: The mechanical vibration is converted into an electrical signal by the above-prepared piezoelectric sensor and collected through a data acquisition board card, and the collected signal is saved and recorded in the PC host computer.

[0073] Step4: Repeat the above impact process at different movement positions to construct an impact dataset. In a specific implementation of the present invention, the specific information of the millimeter-sized ball is a diameter of 7mm and a weight of about 7g.

[0074] In a specific embodiment of the present invention, 205 discrete impact points are evenly arranged at a spacing of 10mm in the training area of 180×180mm 2 on the experimental composite material plate, and the free fall of the millimeter-sized ball is repeated 3 - 15 times at each point.

[0075] Step 4: Composite Material Impact Location Algorithm Based on Retentive Network

[0076] The structure of the impact location algorithm is based on Retentive Network. The input of this network is a two-dimensional matrix composed of eight-channel one-dimensional Lamb wave signals corresponding to eight sensors (i.e., the one-dimensional signal dimensions are the same), and a deep learning model is established to directly output the xy-axis coordinates. The algorithm of Retentive Network (released in 2023, completed by the cooperation of Tsinghua University and Microsoft, Retentive Network A Successor to Transformer for Large Language Models, the link is: https: / / arxiv.org / abs / 2307.08621) is as follows: The eight-channel one-dimensional Lamb wave signals are reduced in dimension through a linear layer, and the dimension changes from 6144 to 256. Then, it passes through the layernorm layer. The LayerNorm layer is the layer normalization layer in the commonly used normalization layers. While keeping the input and output dimensions unchanged, the layernorm layer can reduce the overfitting problem and avoid the problem that the model cannot be trained caused by gradient explosion.

[0077] X L = laayerNorm(X) (2)

[0078] In formula (2), x represents the input of the eight-channel one-dimensional signal, and layerNorm(.) represents the operation of the layernorm layer. That is, formula (2) represents that X has undergone the operation of the layernorm layer to obtain the layer-normalized output X. L .

[0079] Similar to the operations of Q, K, and V in the attention mechanism, Q, K, and V respectively represent Query, Key, and Value. After passing through the layer normalization, the input undergoes three different operations. The following formulas (3 - 5) represent that after multiplying the input passing through the layerNorm layer by the training weight matrix, the relative position embedding operation is further performed. In the formula, xpos represents the relative position embedding, and this method comes from https: / / doi.org / 10.48550 / arXiv.2212.10554..

[0080] Q = xpos(X L W Q ) (3)

[0081] K = xpos(X L W K ) (4)

[0082] V = xpos(X L W K ) (5)

[0083] In equations (3, 4), W Q and W K both represent trainable weights.

[0084] Relative position embedding (xpos) is a technique used in the self-attention mechanism to encode the relative position information between data chunks in a sequence. It can better capture long-range dependencies and improve the performance of the model, and is suitable for establishing mapping relationships through each segment in the signal. The calculation input of xpos Enter is X L W, that is, using X L W to represent and X L W Q and X L W K .

[0085] Table 1: Details of the impact location model parameters

[0086]

[0087] The parameters and parameter values in Table 1 are all internal hyperparameters of the network used in this invention, without clear meaning, and are provided only for the sake of rigor.

[0088] In the following formulas, vectors are represented in the form of [..,..,...]. For the relative position embedding calculation of Q in equation (3), the calculation of scale is obtained according to equation (6). For the relative position embedding calculation of K, the calculation of scale needs to take the reciprocal of the result of equation (6). Using X L W even and X L W odd respectively represent the odd-index columns and even-index columns of X L W, and W even and W odd respectively represent the corresponding weights of the odd and even-index columns. Taking the dimensions (8, 256, 4) used in the research as an example, X L W even represents indices (:, :, 0) and (:, :, 2), and X L W odd represents indices (:, :, 1) and (:, :, 3). Taking X L W odd as an example, here the colon is used to represent taking all the data of X L M, but limited to the case where the third dimension index is 1 or 3. Taking XL W even After taking the opposite number, recombine with X L W odd Merge into X L W * ,W * represents the weight after taking the opposite number. Here, the merging refers to dimension merging, and this variable is required for the operation in Equation (9). In Equations (7, 8), M [1，0.01] represents a tensor with dimensions (256, 2). Each row of the tensor is [1, 0.01]. represents a corresponding multiplication relationship (the "corresponding multiplication relationship" is well-known to those skilled in the art). Taking Equation (7) as an example, after the operation of [0, 1, 2,..., i,..., 255] , each row of the tensor changes from [1, 0.01] to [i, 0.01 * i]. The first row of the tensor is [1, 0.01], and the last row is [255, 2.25]. Use ⊙ to represent dot multiplication, take the corresponding sine and cosine values and perform dot multiplication with the above scale to obtain the sine value sin of the relative position embedding M and the cosine value cos of the relative position embedding M . Finally, through Equation (9), the operation result xpos of the xpos operation is obtained out .

[0089]

[0090] xpos out = X L W ⊙ cos M + X L W * ☉ sin M (9)

[0091] For subsequent calculations, introduce the original gamma matrix d matrix shown in Equation (11). The introduced d matrix needs to ensure the correspondence of the calculation dimensions in Equation (12) (that is, the dimensions of the d matrix meet the basic requirements of the matrix calculation in Equation (13)). This calculation requires first setting a one-dimensional tensor γ, and the length of the tensor is set to n_head, represented by Equation (10). The maximum and minimum values of the elements in the tensor are a and b respectively. Each element in the γ tensor is used as the base in the exponential operation, and use γ i to represent the i-th element in γ. The d matrix calculation is as shown in Equation (11) below.

[0092]

[0093] y i = (QK T ☉ D i )V (13)

[0094] y i = Retemtion(X) (14)

[0095] Where n_sequence is an intermediate parameter value, set to a fixed value of 256 in the present invention, and the gamma matrix D i and the sub-output variable y i respectively represent the results after the corresponding power operation and the results further calculated by Equation (13); the above operations from Equation (2) to Equation (13) can be generally represented by Equation (14), that is, Retention(X) represents [[ID=II]] Figure 5 the entire process from the initial input to the output of the sigmoid module in

[0096] The output y after the above calculation i is Figure 5 a part of the final output of the multi-scale retention module in i Using the (i + 1)-th element in γ, perform the above operation of Equation (14) with the same input X, and combine the results such as y1, y2... y i (which is a three-dimensional tensor) in the third dimension, and then obtain the final output of the multi-scale retention module through the operation of Equation (15).

[0097] y = (sigmoid(XW G ) ⊙ GroupNorm(concat(y1, y2,..., y i )))W o (15)

[0098] In the formula, W G and W O are both weight terms, with dimensions [hidden_size, 2*hidden_size] and [2*hidden_size, hidden_size] respectively. hidden_size represents an intermediate parameter value, which is set to 8 in an embodiment of the present invention; y and X respectively represent Figure 5 the final output of the multi-scale retention module in i and the input after signal dimensionality reduction (see Equation (2)), and y1, y2,..., y i respectively represent the results of the operations of elements 1 to i in γ through Equation (13), sigmoid(.), GroupNorm(.), concat(y1, y2,..., y i) Represent the sigmoid activation function, group normalization function, and tensor concatenation operation respectively (the open-source deep learning framework for machine learning and deep learning is developed by Facebook, and the PyTorch version used is 1.16.0. There are call interfaces in both torch or torch.nn in the PyTorch library).

[0099] After the operation of the above formula (15), the operation result further passes through the layernorm layer, linear layer, and GELU layer to obtain the output result of the retnet model, and then the impact location horizontal and vertical coordinate values of the CFRP plate can be obtained, thus realizing the impact location on the composite material.

[0100] Retentive network has high application potential and is an alternative method to the underlying architecture model Transformer that supports large models. Based on the implementation of composite material impact location by retentive network, this invention introduces two advanced deep learning models, KAN and MAMBA, for comparison.

[0101] The above-mentioned advanced deep learning algorithms have shown varying degrees of advantages in the fields of NLP and CV, but there is still a blank in the research on composite material impact location. The performance comparison was carried out based on the advanced deep learning models MAMBA model, KAN, and the proposed retentive network model. Two composite material plates with different numbers and types of sensors were used to collect and construct impact location datasets in the same way, and the training parameters of the network architecture were controlled to be consistent. The descriptions of the parameters of the 3 models are shown in Table 1 below. The impact location effects based on the MAMBA model, KAN, and retnet model are as follows Figure 6 shown. It can be seen from this that retnet has the lowest average positioning error under various working conditions, and this comparative experiment also verifies the robustness of its high-precision impact location. The following are the main parameter descriptions during the actual program test, which are used to further supplement the parameters not mentioned above. Basically, they are non-critical parameters and are only elaborated here for completeness.

[0102] ·Retnet: At the end of the multi-scale retention module, apply the W in the hidden size setting (15) of the intermediate parameters of the network Figure 5 、W G 、W OWeight dimension. The size of the feed-forward network (FFN) corresponds to the output feature dimension of the input fully connected layer and the input feature dimension of the output fully connected layer in the RetNet feed-forward network module. In a specific embodiment of the present invention, to reduce the consumption during program operation, the RetNet network structure parameter layer (corresponding to "layers") is defaulted to 1, indicating the total number of cycles of RetNet. The tensor length n_heads in Equation (10) specifies how many groups the data is divided into for normalization. At the same time, it limits the step size of the γ vector, which helps with the calculation of xpos. The a and b in γ are equal to -ln32 and -ln512 respectively.

[0103] ·KAN: The parameter num_KAN corresponds to the KANConv2D layer and is the number of KANConv2D layers in the official open-source programming content. After testing with some data, the present invention uses 4 KANConv2D layers. spline_order is a complex parameter related to the dimension of the spline convolution part. At the same time, spline_order is mainly involved in the calculation of spline operations and multi-order spline bases, and the kernal_size, groups, padding, stride, and dilation of the base convolution part and the spline convolution part are kept consistent. More details can be clarified in the open-source code program link.

[0104] (KAN implementation reference code: https: / / github.com / IvanDrokin / torch-conv-kan).

[0105] ·MAMBA: d model is the dimension of the MAMBA model, and this parameter is used to ensure the consistency of the tensor dimensions of the input and output MAMBA models in most cases. d state is the SSM state expansion factor. d conv is the local convolution width, and expand is the block expansion factor. More details can be clarified in the open-source code program link.

[0106] (MAMBA implementation reference code: https: / / github.com / state-spaces / mamba )

[0107] Figure 6 Show the impact location effect Figure 7Show the average error distribution. Referring to the definition of the impact position prediction accuracy in the research paper "A hierarchical deep convolutional regression framework with sensor network fail-safe adaptation for acoustic-emission-based structural health monitoring", according to Equations (16, 17), the prediction accuracy Accuracy of impact localization is determined by normalizing the mean localization error (MLE) with respect to the size of the area under test. MLE is the average distance from the predicted position to the true coordinates of the impact source in N cases of impacts, and the specific calculation is as follows. The average localization accuracy is approximately 0.64%. Figure 6 The elliptical frames of the black curves in correspond to the localization results of different models for the same localization label. In all cases, our model is closer to the impact localization label represented by the black star; Figure 7 As can be seen, compared with other state-of-the-art deep learning models, the error distribution of our model is more concentrated around 0, indicating the superiority of the impact localization performance.

[0108]

[0109] Neural networks and various algorithms are involved in the above text. Wherever not described in detail, conventional technical means in the art are adopted.

Claims

1. A low-speed impact positioning and processing method based on a carbon fiber composite material structure, characterized in that, The details are as follows: Step 1: Preparation of a PVDF-TrFE-based piezoelectric sensor; By means of screen printing technology, a single-component silver paste is brushed on the electrode area on the upper surface of the composite board to form a bridge with two ends. A plurality of bridges are arranged at a plurality of different positions dispersed on the upper surface of the composite board, with one bridge at each position; The orientation of the plurality of bridges depends on the convenience of connecting the bottom and top electrodes; For any one bridge, a rectangular strip connects the two ends, and one of the ends becomes the bottom electrode of the piezoelectric sensor. After curing the silver paste forming four bridges, acetone and N,N-dimethylformamide are mixed into a solution in an equal volume ratio, and then PVDF-TrFE is added to form a PVDF-TrFE coating. After magnetic stirring, the PVDF-TrFE coating is sprayed on the bottom electrode as evenly as possible. During the spraying process, the PVDF-TrFE coating only covers one end of the bridge, not the rectangular strip and the other end, and the edge of the PVDF-TrFE coating reaches or exceeds the covered end. After annealing and corona polarization of the PVDF-TrFE coating, silver paste screen printing is used again above the PVDF-TrFE coating, corresponding to the top electrode of the piezoelectric sensor. The projection of the top electrode on the horizontal plane is surrounded by the projection area of the PVDF-TrFE coating. After the top electrode also undergoes high-temperature curing, the three-layer structure of the piezoelectric sensor is prepared. The three-layer structure from top to bottom is the top electrode formed by the cured silver paste in the upper layer, the PVDF-TrFE coating, and the bottom electrode formed by the cured silver paste in the lower layer; the top electrode serves as the positive electrode and the bottom electrode serves as the negative electrode. Since the preparation and verification of flexible sensors mainly target metal materials at present. Step 2: Fast arrangement based on FastSAM; Obtain pictures of carbon fiber composites through an industrial camera. Place the carbon fiber composites where sensors need to be arranged in the center of the camera lens. Use the FastSAM large model for fast segmentation to quickly obtain the sensor position area in the composite material plate image. The FastSAM large model for fast segmentation automatically and quickly segments the carbon fiber composites to obtain the corresponding binary image of the segmented area. Use the Sobel edge detection in Python to extract the edge points of the carbon fiber composite material plate from the binary image. Perform K-means clustering on the edge points with two different numbers of clusters, and obtain the sensor layout points through weighting. The weighting formula is shown in the following formula (1). In formula (1), the coordinate direction index m is x or y, where x and y respectively represent the horizontal and vertical coordinate directions of the carbon fiber composite sample image in the Cartesian coordinate system. When m is x, after K-means clustering with the number of clustering points being n, the abscissa value k(m) is output. The abscissa value k1(m) corresponding to the number of clustering points being 1 constitutes the abscissa of the cluster, and sp(m) represents the horizontal and vertical coordinate values in the Cartesian coordinate system. This method can arbitrarily set the initial layout points of the sensors, and the number of sensor layout points is equal to the number of K-means clusters. n (m), the abscissa value k1(m) corresponding to the number of clustering points being 1 constitutes the abscissa of the cluster, and sp(m) represents the horizontal and vertical coordinate values in the Cartesian coordinate system. This method can arbitrarily set the initial layout points of the sensors, and the number of sensor layout points is equal to the number of K-means clusters. sp(m) = 0.7 * k n (m) + 0.3 * k1(m) (1) Step 3: Low-velocity impact experiment on the composite material; The specific process is as follows: Step1: The motion position of the three-axis slide is set by the labview program in the PC host computer terminal, and the closed-loop motor controls the motion of the three-axis slide in the xyz space rectangular coordinate system; Step2: At the motion position of the three-axis slide, a millimeter-sized small ball freely falls and impacts on the composite material plate, causing vibration to propagate to the piezoelectric sensor prepared in Step 1 above, and is converted into a voltage waveform through the piezoelectric effect. When the voltage exceeds the threshold voltage, the signal is recorded; Step3: The mechanical vibration is converted into an electrical signal by the piezoelectric sensor prepared above and collected through a data acquisition board card, and the collected signal is saved and recorded in the PC host computer; Step4: Repeat the above impact process at different motion positions to construct an impact data set; Step 4: Composite material impact location algorithm based on the retentive network; The input of the Retentive network is a two-dimensional matrix composed of eight-channel one-dimensional Lamb wave signals corresponding to eight sensors, and a deep learning model is established to directly output the xy-axis coordinates. The Retentive network algorithm is as follows: The eight-channel one-dimensional Lamb wave signals are dimension-reduced through a linear layer and then pass through a layernorm layer; X L = layerNorm(X) (2) In Equation (2), X represents the input of an eight-channel one-dimensional signal, and layerNorm(.) represents the operation of the layernorm layer. That is, Equation (2) represents that X has undergone the operation of the layernorm layer to obtain the layer-normalized output X L ; Q, K, and V respectively represent Query, Key, and Value. After layer normalization, the input is subjected to three different operations. The following formula (3-5) represents that after the input passing through the layerNorm layer is multiplied by the training weight matrix, the relative position embedding operation is further performed; in the formula, xpos represents the relative position embedding. Q = xpos(X L W Q ) (3) K = xpos(X L W K ) (4) V = X L W K (5) In formulas (3, 4), W Q and W K both represent trainable weights; The calculation input of the relative position embedding xpos is X L W, that is, using X L W represents and X L W Q and X L W K ; Vectors are represented in the form of [..,..,...] in the following formula; for the calculation of the relative position embedding of Q in Equation (3), the scale is calculated according to Equation (6), and for the calculation of the relative position embedding of K, the scale is calculated by taking the reciprocal of the result of Equation (6); use X L W even and X L W odd represent the odd-indexed columns and even-indexed columns of X L W respectively, and W even and W odd represent the corresponding weights of the odd- and even-indexed columns respectively; Take X L W even After taking the opposite value, recombine with X L W odd Merge into X L W * , implemented through the concat function of the pytorch framework, where W * represents the weight after taking the opposite value; M [1，0.01] represents a tensor of dimension (256, 2), where each row of the tensor is [1, 0.01]; represents the corresponding multiplication relationship; use ⊙ to denote dot product, take the corresponding sine and cosine values and perform dot product with the above scale to obtain the sine value sin of the relative position embedding M and the cosine value cos of the relative position embedding M ; obtain the operation result xpos of the xpos operation through Equation (9) out ; xpos out = X L W⊙cos M + X L W * ⊙sin M (9) Introduce the original gamma matrix d matrix shown in formula (11). The introduced d matrix needs to ensure the correspondence of the calculation dimensions in formula (12), that is, the dimension of the d matrix meets the basic requirements of matrix calculation in formula (13); set a one-dimensional tensor γ, the length of the tensor is set to n_head, and the maximum and minimum values of the elements in the tensor are a and b respectively; each element in the γ tensor is used as the base in the exponential operation, and γ i represents the i-th element in γ; the calculation of the d matrix is shown in the following formula (11); y i = (QK T ⊙D i )V (13) y i = Retention(X) (14) where n_sequence is the first intermediate parameter value, and the gamma matrix D i and the sub-output variable y i respectively represent the results after the corresponding exponentiation operations and the results further calculated by Equation (13); the above operations from Equation (2) to Equation (13) are summarized by Equation (14), that is, Retention(X) represents the entire process from the initial input to the output of the sigmoid module; Output y i is part of the final output of the multi-scale retention module; take the (i + 1)-th element in γ, perform the operation of the above formula (14) with the same input X, and combine the results of y1, y2... y i etc. in the third dimension (implemented by concat), and then obtain the final output of the multi-scale retention module through the operation of formula (15); y = (sigmoid(XW G ) ⊙ Grouporm(concat(y1, y2,..., y i )))W O (15) Where, W G and W O are both weight terms, with dimensions [hidden_size, 2*hidden_size] and [2*hidden_size, hidden_size] respectively, and hidden_size represents the second intermediate parameter value; y and X represent the final output of the multi-scale retention module and the input after signal dimensionality reduction respectively, and y1, y2,..., y i represent the results of the operations of elements 1 to i in γ through Equation (13) respectively, and sigmoid(.), GroupNorm(.), and concat(y1, y2,..., y i ) represent the sigmoid activation function, the group normalization function, and the tensor concatenation operation respectively; After the operation of the above formula (15), the operation result further passes through the layernorm layer, the linear layer, and the GELU layer to obtain the output result of the retnet model, that is, the impact positioning horizontal and vertical coordinate values of the CFRP panel are obtained.

2. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, wherein In step 1, the two ends are two rectangular ends with different sizes. The four sides of the two rectangular ends are parallel to each other respectively, and a pair of opposite sides are parallel to the long side of the rectangular strip. Only the large rectangular end is covered with a PVDF-TrFE coating, and then the top electrode is covered.

3. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, wherein In step 1, the silver paste is cured by drying in a vacuum dryer at 80°C for 1 hour; the screen used for screen printing is made of 250-mesh polyester fabric; a solution is prepared by mixing acetone and N-N dimethylformamide in a volume ratio of 50 / 50 vol% and then adding PVDF-TrFE in a molar ratio of 70 / 30 mol% to form a PVDF-TrFE coating; the mass of PVDF-TrFE is 5% of the mass of the solution; after 6 hours of magnetic stirring, the solution is sprayed on the bottom electrode; the PVDF-TrFE coating also needs to be annealed at 140°C and corona polarized at 45 kV.

4. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, characterized in that The shape of the top electrode is circular, and the top electrode needs to be cured at a high temperature of 80°C.

5. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, characterized in that The carbon fiber composite material is a CFRP panel, and the positive and negative electrodes are connected to enameled wires.

6. The low-speed impact positioning processing method based on a carbon fiber composite material structure according to claim 1, wherein In Step 2 of step 3, the threshold voltage is set to 0.02V.

7. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, wherein, In step 3, the millimeter-sized ball has a diameter of 7 mm and a weight of 7 g.

8. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, characterized in that In Step 3, 205 discrete impact points are evenly distributed at an interval of 10 mm in the training area of 180×180 mm on the experimental composite plate, and free fall of millimeter-sized balls is repeated 3 - 15 times at each point. 2 ​ 9. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, characterized in that In step 4, the eight-channel one-dimensional Lamb wave signal is dimensionally reduced by the linear layer, and the dimension changes from 6144 to 256.

10. The low-speed impact positioning and processing method based on a carbon fiber composite material structure according to claim 1, characterized in that In step 4, n_sequence is set to 256, and hidden_size is set to 8.

Images