Impact load feature recognition method and system based on deep convolutional neural network, and medium
By combining a deep convolutional neural network with a PZT piezoelectric sensor and wavelet transform processing, the impact load characteristics of honeycomb sandwich panels are identified, solving the problem of inaccurate inversion in existing technologies and achieving efficient impact load identification and damage assessment.
Patent Information
- Application Number
- CN202510830069.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-10-17
AI Technical Summary
Existing technologies make it difficult to accurately invert the impact load characteristics of honeycomb sandwich panels. Especially in the aerospace field, the energy and material characteristics of the impact load are insufficiently estimated, affecting the accuracy of damage assessment.
A deep convolutional neural network is used to acquire signals through PZT piezoelectric sensors distributed at the four vertices of the honeycomb sandwich panel. After wavelet transform processing, the signals are input into the convolutional neural network to identify the energy characteristic parameters of the impact source, including diameter, material, and velocity.
It achieves accurate recognition of the impact load characteristics of honeycomb sandwich panels, maximizes the retention of data features, and the neural network structure is lightweight and easy to train and deploy, with an identification accuracy rate of up to 99.86%.
Smart Images
Figure CN120804566A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of honeycomb sandwich panel impact load identification. In particular, it relates to an impact load feature identification method and system based on deep convolutional neural network and a medium. BACKGROUND
[0002] Honeycomb sandwich panel is a lightweight material widely used in aerospace. Structures in the field of aerospace often face the risk of impact, such as debris impact in outer space. These potential threats pose a great danger to the structure, and special methods need to be developed to invert impact load features, mainly to estimate the energy and material characteristics of the impact load, and to play a role in damage assessment. How to develop a special method to invert the impact load features of honeycomb sandwich panel is a problem that needs to be solved. SUMMARY
[0003] The present application provides an impact load feature identification method and system based on deep convolutional neural network to accurately invert the impact load features of honeycomb sandwich panel.
[0004] To achieve the above-mentioned purpose, in a first aspect, the present application relates to an impact load feature identification method based on deep convolutional neural network, which is used to invert the energy features of impact sources on the surface of honeycomb sandwich panel structure, comprising:
[0005] When the impact source impacts the measured impact rectangular area, the signals obtained by the PZT piezoelectric sensors distributed at the four corners of the measured impact rectangular area are acquired, wherein the impact source occurs in the measured impact rectangular area;
[0006] The signals of the four PZT piezoelectric sensors are respectively processed by wavelet transform to obtain four gray scale images of wavelet transform correlation numbers, wherein the horizontal coordinate of the gray scale image of the wavelet transform correlation number corresponds to time, and the vertical coordinate corresponds to frequency, and the gray scale value of each pixel is the intensity of the signal at that frequency at that time;
[0007] The four gray scale images of wavelet transform correlation numbers obtained are input into a convolutional neural network for inverting energy feature parameters to obtain the predicted energy feature inversion in the wavelet transform value, wherein the energy feature parameters include the diameter, material and speed of the impact source.
[0008] To achieve the above-mentioned purpose, in a second aspect, the present application relates to an impact load feature identification system based on deep convolutional neural network, comprising: a signal acquisition module, which is used to acquire the signals obtained by the PZT piezoelectric sensors distributed at the four corners of the measured impact rectangular area when the impact source impacts the measured impact rectangular area, wherein the impact source occurs in the measured impact rectangular area;
[0009] a wavelet transform module for performing wavelet transform processing on the signals of the four PZT piezoelectric sensors, respectively, to obtain four grayscale images of wavelet transform correlation numbers, wherein the abscissa of the grayscale image of the wavelet transform correlation number corresponds to time, the ordinate corresponds to frequency, and the grayscale value of each pixel is the intensity of the frequency signal at that moment;
[0010] An inversion module is used to input the grayscale images of the four wavelet transform correlation numbers obtained respectively into the convolutional neural network for inverting energy characteristic parameters to obtain the energy characteristic inversion in the predicted wavelet transform value, where the energy characteristic parameters include the diameter, material and velocity of the impact source.
[0011] To achieve the above objectives, in a third aspect, the present invention also relates to a computer-readable storage medium, in which instructions are stored, and when the instructions are run, the above-mentioned impact load feature recognition method of a deep convolutional neural network is executed.
[0012] The present invention relates to a method, system, and medium for identifying impact load characteristics using a deep convolutional neural network. Compared with the prior art, the present invention has the following beneficial effects:
[0013] After wavelet transformation processing, both the time and frequency domain characteristics of the data are preserved, maximizing the characteristics of the original signal and facilitating subsequent neural network identification of energy signature parameters. Overall, the inversion convolutional neural network structure is lightweight, easy to train and deploy, and possesses sufficient expressive power. This invention can accurately invert the impact load characteristics of honeycomb sandwich panels. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 Schematic diagram of the flow of a method for identifying impact load characteristics using a deep convolutional neural network in Example 1.
[0015] Figure 2 This is the original signal before wavelet transformation of the impact load feature recognition method using a deep convolutional neural network in Example 1.
[0016] Figure 3 Schematic diagram of the two-dimensional correlation coefficient after wavelet transformation of an impact load feature recognition method using a deep convolutional neural network in Example 1.
[0017] Figure 4 Schematic diagram of the impact load inversion convolutional neural network structure of the impact load feature recognition method of a deep convolutional neural network in Example 1.
[0018] Figure 5 This is the loss function curve of the neural network model training of the experimental data set of Example 1 of the impact load feature recognition method using a deep convolutional neural network in Example 1.
[0019] Figure 6 The neural network model training loss function curve of the experimental data set of example 1 of the impact load feature recognition method of the deep convolutional neural network in embodiment one.
[0020] Figure 7 The predicted height scatter plot of example 1 of the impact load feature recognition method of the deep convolutional neural network in embodiment one.
[0021] Figure 8 The predicted diameter scatter plot of example 1 of the impact load feature recognition method of the deep convolutional neural network in embodiment one.
[0022] Figure 9 The predicted material scatter plot of example 1 of the impact load feature recognition method of the deep convolutional neural network in embodiment one.
[0023] Figure 10 The neural network model training loss function curve of the impact load feature recognition method of the deep convolutional neural network in embodiment one.
[0024] Figure 11 The structural schematic diagram of the impact load feature recognition system of the deep convolutional neural network in embodiment two. DETAILED DESCRIPTION
[0025] The application will be further described below in conjunction with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the application, and not to limit the application. In addition, it should be noted that, in order to facilitate the description, only the parts related to the application are shown in the drawings, not all the structures.
[0026] Embodiment one
[0027] The impact load feature recognition method of the deep convolutional neural network, please refer to Figures 1-5 The impact load feature recognition method of the deep convolutional neural network, please refer to
[0028] S101, when the impact source impacts the measured impact rectangular region, acquiring the signals acquired by the PZT piezoelectric sensor distributed at the four corners of the measured impact rectangular region, wherein the impact source occurs in the measured impact rectangular region. The rectangular four-corner sensor layout is adopted, the sensor arrangement is simple to paste, the signals with good symmetry can be collected, and when the monitoring area is expanded, the rectangular sensor network layout can be expanded.
[0029] Wherein, the original signal before wavelet transform is a time stress signal, such as Figure 2as shown.
[0030] S102 wavelet transform processing is carried out on signals of four PZT (Lead Zirconate Titanate Piezoelectric Sensor) piezoelectric sensors respectively, and four gray scale images of wavelet transform correlation numbers are obtained respectively, wherein the abscissa of the gray scale image of the wavelet transform correlation number corresponds to time, the ordinate corresponds to frequency, and the gray scale value of each pixel is the intensity of the frequency signal at the time.
[0031] The basic formula of wavelet transform is shown in the formula, wherein f(t) is the original signal, ψ(t) is the wavelet function, b is the translation parameter, i.e. the time dimension, and a is the scale parameter, i.e. the frequency dimension. A series of translation parameters and scale parameters are transformed to obtain different wavelet correlation coefficients, and the corresponding wavelet transform values can be represented as two-dimensional data with a and b as the axes. After wavelet transformation, the time domain characteristics and frequency domain characteristics of the data can be retained, the characteristics of the original signal can be maximized, and subsequent energy feature parameter recognition is facilitated.
[0032] S103 inputs the four gray scale images of wavelet transform correlation numbers obtained respectively into a convolutional neural network for energy feature parameter inversion to obtain predicted energy feature inversion in the wavelet transform values, and the energy feature parameters include the diameter, material and speed of the impact source.
[0033] Before S103, the training process of the convolutional neural network for energy feature parameter inversion includes:
[0034] S1031: generate a plurality of experimental energy feature parameters based on a preset diameter set, a preset material set and a preset speed set, perform an impact experiment based on the plurality of experimental energy feature parameters respectively, make the impact source impact the measured impact rectangular area to obtain experimental signals obtained by the four PZT piezoelectric sensors, and perform wavelet transform processing on the experimental signals to obtain gray scale images of the wavelet transform correlation numbers for training.
[0035] S1032: take the experimental energy feature parameters and the gray scale images of the wavelet transform correlation numbers corresponding to the experimental energy feature parameters as a training data set, and train the convolutional neural network for energy feature parameter inversion based on the training data set.
[0036] The training process of the convolutional neural network for inverting the energy characteristic parameter is specifically as follows: in order to develop an impact load inversion method for the experiment, the foregoing neural network inversion model is trained on 720 sets of experimental data. In terms of data set division, the overall data set is randomly divided, 70% of the data set is taken as a training data set, and 30% of the data set is taken as a test set. A mean square error function is used as a loss function. As shown in Figure 10 The best learning rate obtained by grid search is 0.001, and the best batch size is 4. The model is trained for 500 rounds in the optimal super parameter configuration in the experimental data set. The training loss curve is shown in FIG. add. The curve is stable and normal, and there is no overfitting phenomenon. The relative error is less than 5%.
[0037] A single sample input data includes four feature images, including a wavelet transform image of four surrounding sensor signals, as shown in Figure 4 The gray image of the wavelet transform correlation number of one of the sensor signals is shown in the embodiment. The convolutional neural network structure includes an input layer, a two-dimensional convolutional layer, a first activation function layer, a pooling layer, a flattening layer, a second activation function layer, a first full connection layer, a third activation function layer, the second full connection layer and the fourth activation function layer connected in sequence.
[0038] In the embodiment, the input layer inputs four-channel two-dimensional data (including a gray image of the wavelet transform correlation number of four sensors); the flattening layer flattens the multi-dimensional data output by the pooling layer into one-dimensional data and inputs the first full connection layer; the fourth activation function layer outputs three values representing the drop height, diameter and material of the impact source, wherein the material is a preset material code.
[0039] As shown in Figure 5 It is a convolutional neural network structure diagram for inverting the impact load characteristic parameter. The input layer is an input layer, and the input is four-channel two-dimensional data with a size of 100x100; the Conv is a two-dimensional convolutional layer, which is responsible for the main image recognition work; the Relu is an activation function layer, which can introduce a nonlinear mapping relationship into the neural network; the MaxPool is a pooling layer, which can extract main features and ignore secondary features to enhance the robustness of the model; the Gemm is a full connection layer, which is responsible for the complex final relationship mapping and decision-making part; the Flatten is a flattening layer, which flattens multi-dimensional data into one dimension to facilitate the connection of the full connection layer; the output is three values, which represent the drop height, diameter and material of the impact source in sequence, wherein the material does not have a continuous value, and can be coded for different materials to adapt to the model. Overall, the inversion model is lightweight, easy to train and deploy, and has sufficient expression ability.
[0040] In order to better illustrate the inversion accuracy of the scheme of the present application, the following example is given to test the effect of the impact load feature recognition method of the deep convolutional neural network, as shown in Figures 6-9 The method comprises the following steps:
[0041] In order to test the foregoing method, actual experimental verification is carried out. In the experiment, the honeycomb sandwich plate has a length and width of 200 mm and a thickness of 15 mm, and the material is an aluminum honeycomb structure. The center region of 120 mm x 120 mm of the honeycomb sandwich plate is taken as a monitoring region, the detection region is divided into 2 x 2 four rectangular sub-regions, each sub-region has the same size, which is 60 mm x 60 mm, and only the data of the four sensors around the impact point sub-region are taken for inversion. In the experiment, the drop ball free-fall impact method is used to apply the impact load, and the work condition parameters related to the energy of the drop ball include the material, diameter and drop height of the drop ball. The drop ball material is divided into tungsten steel and aluminum material, the drop ball diameter is divided into 6 mm, 8 mm and 10 mm, and the drop height is divided into 100 mm, 150 mm and 210 mm. Because the impact point has a certain randomness, the work condition is repeated 10 times on each sub-region, which reduces the error caused by accidental factors and also increases the data sample size. According to the above parameter combination, a total of 4 x 3 x 3 x 2 x 10 = 720 groups of experimental data are obtained after the experiment.
[0042] Firstly, the impact area positioning method is used to invert the impact area of the 720 groups of experiments, in the prediction of all the 720 groups of experimental conditions, 719 groups are correct and 1 group is wrong, the overall area recognition accuracy is as high as 99.86%, which shows that the impact positioning method is reliable and has high stability.
[0043] Next, the energy-related parameter inversion is carried out, at this time the energy-related parameters of the impact load identification are diameter, height and material, the diameter is divided into 6 mm, 8 mm and 10 mm, the height is divided into 100 mm, 150 mm and 210 mm, and the material is divided into 0 and 1, wherein 0 represents aluminum material and 1 represents tungsten steel material, the prediction result between 0-0.5 is aluminum material, and the prediction result between 0.5-1 is tungsten steel material, so the classification problem of the material can be converted into a regression problem. The foregoing neural network inversion model is trained on the 720 groups of experimental data. The grid search method is used to obtain the best learning rate of 0.001 and the best batch size of 4. The model is trained for 500 rounds in the experimental data set, and the training loss curve is as shown in Figure 7 The relative error is less than 5%.
[0044] Example Two
[0045] The application discloses an impact load characteristic identification system of a deep convolutional neural network, which is used for energy characteristic inversion of an impact source on a honeycomb sandwich plate structure surface and is realized by electronic device hardware with a central processing unit, and can be realized by a personal computer, a smart terminal, a local area network, a server and the like. Figure 11 The application comprises a signal acquisition module 61, a wavelet transform module 62 and an inversion module 63.
[0046] The signal acquisition module 61 is used for acquiring signals acquired by PZT piezoelectric sensors distributed at four corners of a measured impact rectangular region when an impact source impacts the measured impact rectangular region, wherein the impact source occurs in the measured impact rectangular region.
[0047] The wavelet transform module 62 is used for performing wavelet transform processing on signals of four PZT piezoelectric sensors respectively to obtain four gray scale images of wavelet transform correlation numbers respectively, wherein the horizontal coordinate of the gray scale image of the wavelet transform correlation number corresponds to time, the vertical coordinate corresponds to frequency, and the gray scale value of each pixel is the intensity of the frequency signal at the time.
[0048] The inversion module 63 is used for inputting the four gray scale images of wavelet transform correlation numbers obtained respectively into a convolutional neural network for energy characteristic parameter inversion to obtain predicted energy characteristic inversion in the wavelet transform value, wherein the energy characteristic parameters include the diameter, material and speed of the impact source.
[0049] The inversion module 63 further comprises a training data generation module 64 and a network training module 65.
[0050] The training data generation module 64 is used for generating a plurality of experimental energy characteristic parameters based on a preset diameter set, a preset material set and a preset speed set, performing an impact experiment based on the plurality of experimental energy characteristic parameters respectively, acquiring experimental signals acquired by four PZT piezoelectric sensors when an impact source impacts the measured impact rectangular region, and performing wavelet transform processing on the experimental signals to obtain gray scale images of wavelet transform correlation numbers for training.
[0051] The network training module 65 is used for taking the experimental energy characteristic parameters and the gray scale images of wavelet transform correlation numbers corresponding to the experimental energy characteristic parameters as a training data set, and training a convolutional neural network for energy characteristic parameter inversion based on the training data set.
[0052] In the embodiment, the convolutional neural network structure comprises an input layer, a two-dimensional convolutional layer, a first activation function layer, a pooling layer, a flattening layer, a second activation function layer, a first full connection layer, a third activation function layer, a second full connection layer and a fourth activation function layer which are sequentially connected.
[0053] The input layer input is two-dimensional data of 7 channels; the flattening layer flattens the multi-dimensional data output by the pooling layer into one-dimensional data and inputs the first full connection layer; the fourth activation function layer outputs three values representing the falling height, diameter and material of the impact source, wherein the material is a preset material code.
[0054] The impact load feature recognition system of the deep convolutional neural network of the embodiment has the same implementation process, method and effect as the impact load feature recognition method of the deep convolutional neural network described in embodiment one, and thus will not be described here.
[0055] Embodiment three
[0056] The present application relates to a kind of computer readable storage medium, storage medium has instruction, instruction executes the impact load feature recognition method of the deep convolutional neural network of embodiment one, its runtime implementation process, method and effect are same with the impact load feature recognition method of the deep convolutional neural network described in embodiment one, and thus will not be described here.
[0057] It should be noted that, in this paper, the term "including", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent in such process, method, article or device. Without more limitation, the element defined by the sentence "including a …" does not exclude the existence of other identical elements in the process, method, article or device including the element.
[0058] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.
Claims
1. A method for identifying impact load characteristics using a deep convolutional neural network, characterized in that: Used for inversion of energy characteristics of impact sources on the surface of honeycomb sandwich panel structures, including: When an impact source impacts the measured impact rectangular area, signals obtained by PZT piezoelectric sensors distributed at four corners of the measured impact rectangular area are obtained, wherein the impact source occurs in the measured impact rectangular area; The signals of the four PZT piezoelectric sensors are respectively subjected to wavelet transform processing to obtain four grayscale images of wavelet transform correlation numbers, wherein the horizontal axis of the grayscale image of the wavelet transform correlation number corresponds to time, the vertical axis corresponds to frequency, and the grayscale value of each pixel is the intensity of the frequency signal at that moment; The grayscale images of the four wavelet transform correlation numbers obtained are input into a convolutional neural network for inverting energy characteristic parameters to obtain the energy characteristic inversion in the predicted wavelet transform value, where the energy characteristic parameters include the diameter, material and velocity of the impact source.
2. The method for identifying impact load characteristics using a deep convolutional neural network according to claim 1, wherein: Before inputting the grayscale images obtained by respectively obtaining the four wavelet transform correlation numbers into the convolutional neural network for inverting the energy characteristic parameters, the method further includes: generating a plurality of experimental energy characteristic parameters based on a preset diameter set, a preset material set, and a preset velocity set, performing impact experiments based on the plurality of experimental energy characteristic parameters, causing an impact source to impact the measured impact rectangular area to obtain experimental signals obtained by four PZT piezoelectric sensors, and performing wavelet transform processing on the experimental signals to obtain grayscale images of the wavelet transform correlation numbers used for training; The experimental energy characteristic parameters and the grayscale images of the wavelet transform correlation numbers corresponding to the experimental energy characteristic parameters are used as training data sets, and the convolutional neural network for inverting the energy characteristic parameters is trained based on the training data sets.
3. The method for identifying impact load characteristics using a deep convolutional neural network according to claim 1, wherein: The convolutional neural network structure includes an input layer, a two-dimensional convolution layer, a first activation function layer, a pooling layer, a flattening layer, a second activation function layer, a first fully connected layer, a third activation function layer, the second fully connected layer and a fourth activation function layer connected in sequence.
4. The method for identifying impact load characteristics using a deep convolutional neural network according to claim 3, wherein: The input layer inputs 4-channel two-dimensional data; the flattening layer flattens the multi-dimensional data output by the pooling layer into one-dimensional data and inputs it into the first fully connected layer; the fourth activation function layer outputs 3 numerical values representing the falling height, diameter and material of the impact source, wherein the material is a preset material code.
5. A deep convolutional neural network impact load feature recognition system, characterized in that: Used for inversion of energy characteristics of impact sources on the surface of honeycomb sandwich panel structures, including: a signal acquisition module, configured to acquire signals acquired by PZT piezoelectric sensors distributed at four corners of the rectangular area under test when an impact source impacts the rectangular area under test, wherein the impact source occurs in the rectangular area under test; a wavelet transform module for performing wavelet transform processing on the signals of the four PZT piezoelectric sensors, respectively, to obtain four grayscale images of wavelet transform correlation numbers, wherein the abscissa of the grayscale image of the wavelet transform correlation number corresponds to time, the ordinate corresponds to frequency, and the grayscale value of each pixel is the intensity of the frequency signal at that moment; An inversion module is used to input the grayscale images of the four wavelet transform correlation numbers obtained respectively into the convolutional neural network for inverting energy characteristic parameters to obtain the energy characteristic inversion in the predicted wavelet transform value, where the energy characteristic parameters include the diameter, material and velocity of the impact source.
6. The impact load feature recognition system of a deep convolutional neural network according to claim 5, characterized in that: Before the inversion module, it also includes: a training data generation module, configured to generate a plurality of experimental energy characteristic parameters based on a preset diameter set, a preset material set, and a preset velocity set, conduct impact experiments based on the plurality of experimental energy characteristic parameters, cause an impact source to impact the measured impact rectangular area, obtain experimental signals obtained by the four PZT piezoelectric sensors, and perform wavelet transform processing on the experimental signals to obtain grayscale images of the wavelet transform correlation numbers used for training; The network training module is used to use the experimental energy characteristic parameters and the grayscale images of the wavelet transform correlation numbers corresponding to the experimental energy characteristic parameters as training data sets, and to train a convolutional neural network for inverting energy characteristic parameters based on the training data sets.
7. The impact load feature recognition system of a deep convolutional neural network according to claim 5, characterized in that: The convolutional neural network structure includes an input layer, a two-dimensional convolution layer, a first activation function layer, a pooling layer, a flattening layer, a second activation function layer, a first fully connected layer, a third activation function layer, the second fully connected layer and a fourth activation function layer connected in sequence.
8. The impact load feature recognition system of a deep convolutional neural network according to claim 7 is characterized in that: The input layer inputs 7-channel two-dimensional data; the flattening layer flattens the multi-dimensional data output by the pooling layer into one-dimensional data and inputs it into the first fully connected layer; the fourth activation function layer outputs 3 numerical values representing the falling height, diameter and material of the impact source, wherein the material is a preset material code.
9. A computer-readable storage medium, characterized in that: The storage medium stores instructions, which, when executed, execute the impact load feature recognition method of a deep convolutional neural network according to any one of claims 1 to 4.