Composite material plate structure impact position identification method based on impact stress wave deep learning
Through the deep learning model based on the TCN-CNN network, the characteristics of impact stress wave signals of composite material structures are extracted, and the mapping relationship between the impact response signals and positions is established, which solves the problems of low impact positioning accuracy and high resource consumption in the prior art. , realizes efficient and accurate impact position identification.
Patent Information
- Application Number
- CN202411226374.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-11-22
AI Technical Summary
The existing impact positioning methods for composite material structures have problems such as low accuracy, relying on prior knowledge, and consuming a lot of resources, making it difficult to accurately identify the impact source coordinates in composite material structures.
Using a deep learning model based on TCN-CNN network, the sequence characteristics and high-dimensional abstract characteristics of the impact stress wave signal are extracted through the time convolution and convolution module, and the mapping relationship between the impact response signal and the impact position is established in combination with the fully connected module. Deep learning technology does not need to rely on wave time and peak information to achieve the identification of impact locations.
It realizes efficient and accurate identification of the impact position of composite material structures, improves positioning accuracy, reduces dependence on prior knowledge, and reduces resource consumption.
Smart Images

Figure CN119004080A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of structural health monitoring, and in particular to a method for identifying the impact position of a composite material plate structure based on deep learning of impact stress waves. Background Art
[0002] Composite materials have the advantages of high specific strength, high specific modulus, strong designability and corrosion resistance. Therefore, they have been widely used in aerospace, rail transit and ship transportation in recent years. However, composite structures will inevitably suffer from external impacts during service, such as bird strikes, hail, runway gravel, etc., which will cause invisible tiny damage inside, resulting in a significant decrease in the bearing capacity of the structure, which seriously threatens the service safety of the structure. Therefore, it is crucial to develop online monitoring technology for composite structures that can locate external impact events in real time.
[0003] At present, the commonly used impact location methods at home and abroad mainly include time difference method, reference database method and machine learning method. The time difference method uses the geometric relationship between the wave arrival time difference, wave velocity and sensor position to achieve impact location, but due to the dispersion characteristics of the impact signal, external noise and boundary reflection and other interference, the time difference method is often limited in practical applications. The reference database method obtains reference signals from all sampling points on the structure to establish a database, and determines the impact position by comparing the similarity between the real impact signal and the reference signal. However, this method requires a lot of time and storage resources. The machine learning method needs to extract features based on prior knowledge and use a specific machine learning algorithm to achieve impact location. This method is highly dependent on prior knowledge and has poor generalization ability.
[0004] With the vigorous development of artificial intelligence technology, deep learning algorithms are increasingly used in the field of structural health monitoring. The impact location method based on the deep learning model has the ability to mine data independently of prior knowledge, and can obtain the deep mapping relationship between high-dimensional data and target parameters. The original impact signal is input into the designed deep network model, and the positioning result is directly output after processing and analysis. However, the existing impact location methods based on deep learning usually use simpler and shallower networks, which makes it difficult to obtain satisfactory positioning accuracy.
[0005] Therefore, in view of the shortcomings of existing technologies and the needs of engineering applications, there is an urgent need for a deep learning impact location method that can more accurately and efficiently identify the coordinates of the impact source. Summary of the invention
[0006] The purpose of the present application is to solve the problems in the prior art and to provide an improved method for identifying the impact position of a composite material plate structure.
[0007] To this end, some embodiments of the present application propose an impact position identification method for a composite plate structure based on deep learning of impact stress waves, which includes the steps of: establishing an impact response signal data set; establishing an impact position identification deep learning model based on a TCN-CNN network learning model, the TCN-CNN network learning model including a TCN module that uses the impact stress wave signal measured by each sensor as the impact response signal and outputs the sequence features of the impact stress wave signal; a CNN module that uses the sequence features of the impact stress wave signal as input and outputs high-dimensional abstract features of the impact stress wave signal; and a fully connected module that uses the high-dimensional abstract features of the impact stress wave signal as input and uses the impact source coordinates as output to establish a mapping relationship between the impact response signal and the impact position coordinates; wherein the time series features of the impact stress wave signal output by the TCN module are sent to the CNN module after convolution.
[0008] In some embodiments, the establishment of the impact response signal data set includes pasting a sensor on a composite material plate, performing impact experiments at different positions in the sensor envelope area, collecting the impact response signal in real time and recording the coordinate position, establishing a database and dividing the training set and the test set according to a preset ratio.
[0009] In some embodiments, a dilated convolution calculation method is used in the TCN module to increase the size of the convolution kernel by inserting a fixed number of spaces between elements of the standard convolution kernel.
[0010] In some embodiments, batch normalization (BN) and activation function processing are performed after each convolution calculation of the CNN module.
[0011] In some embodiments, the normalization process for each batch adds a little noise to prevent the model from overfitting.
[0012] In some embodiments, the activation function uses a ReLU activation function.
[0013] In some embodiments, the RMSProp optimization method is used to adjust the learning rate of each parameter by performing an exponentially weighted moving average on the square of the gradient, which includes first initializing the exponentially weighted moving average of the square of the gradient and the time step, then calculating the current gradient, and updating the exponentially weighted moving average of the square of the gradient and the learning rate, E[g 2 ]0=0,t=0;E[g 2 ] t =β[g 2 ] t-1 +(1-β)g t 2 ; Where E[g2 ] t is the exponentially weighted moving average of the squared gradient, g t is the current gradient, β is the decay rate, θ t is the parameter value at the current moment, η is the global learning rate, and ε is a small constant for numerical stability.
[0014] The beneficial effect of the proposed method of the present application is that the deep learning technology is used to identify the impact position without considering the arrival time, peak value and other information of the impact response signal, and the deep-level characteristics of the data can be mined and learned. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, 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 recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0016] Figure 1 Schematic diagram of impact damage experiment;
[0017] Figure 2a to Figure 2d This is a schematic diagram of the shock response signals collected by four sensors, from which we can see the signal form of the shock wave;
[0018] Figure 3a A deep learning network model diagram according to an embodiment of the present application;
[0019] Figure 3b A schematic diagram of TCN dilated convolution according to a method according to an embodiment of the present application;
[0020] Figure 3c A schematic diagram of a CNN two-dimensional convolution operation according to a method of an embodiment of the present application;
[0021] Figure 4a This is the loss function iteration graph. As can be seen from the figure, after multiple iterations of training, the loss function value gradually decreases, indicating that the model effect is good.
[0022] Figure 4b for Figure 4a A partial enlarged view of part A in the middle.
[0023] Figures 5a to 5d The schematic diagram between the real impact position and the model predicted position. Figures 5a to 5d It can be seen that the model proposed in this application predicts that the position distance is close. DETAILED DESCRIPTION
[0024] To make the technical solutions and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention:
[0025] The impact position identification method of composite plate structure based on impact stress wave deep learning specifically includes the following steps:
[0026] First, the piezoelectric sensor is pasted and a falling ball impact experiment is performed.
[0027] Piezoelectric sensors PZT, such as ceramic piezoelectric sensors, are glued to four locations of the composite material structure, such as Figure 1 As shown. The falling ball impact experiment is carried out at different positions. For example, the falling ball impact is carried out at the position of the impact source SoC, and the shock wave signal propagates to the piezoelectric sensor along the signal propagation path SP. The piezoelectric sensor collects the impact waveguide signal and records the coordinate position of the impact point. The signal examples collected by the four sensors PZT1 to PZT4 are shown as follows: Figure 2a to Figure 2d As shown.
[0028] Second, label the data and establish a database.
[0029] The collected impact response signals and the position coordinates of the falling ball are marked and a database is established.
[0030] Third, establish a deep learning network model.
[0031] The deep learning network model established is the TCN-CNN (time convolution-convolution) network learning model. The network model structure is as follows: Figure 3a As shown. The collected impulse response signal is a complex sequence signal containing time information. The dimension of each data line is recorded as C×H×W (1×4×5000), where C, H, and W represent the data channel, height, and width, respectively. The input data is processed by the time convolution network module (TCN module) 10 on each sensor signal to learn the time series characteristics of the impulse response signal. Then the signal time series characteristics of the four sensors are spliced and the high-dimensional abstract features of the signal are learned through convolution calculation of the CNN module 20. As shown Figure 3b As shown, the calculation method of hollow convolution can be used in the temporal convolution network module 10, which increases the size of the convolution kernel by inserting a fixed number of spaces between the elements of the standard convolution kernel without increasing the number of parameters or the amount of calculation. This method can increase the receptive field of the convolution layer, that is, the size of the area covered by a single convolution operation in the input feature map, allowing the network to capture a wider range of contextual information without significantly increasing the computational burden. The calculation process is shown in formula (1).
[0032]
[0033] Where i is the position index in the output sequence, k is the size of the convolution kernel, w is the convolution kernel, x[i] is the input sequence, f[i] is the output sequence, l is the offset position index of the convolution kernel, indicating the lth position of the convolution kernel. d is the dilation factor, which defines the step size of the convolution kernel jumping on the input sequence.
[0034] like Figure 3c The figure shows a schematic diagram of the CNN two-dimensional convolution operation. Figure 3c As shown in the figure, in this CNN network, the convolution kernel of each convolution layer is 3x3 and the stride is 2.
[0035] The calculation process of CNN two-dimensional convolution is shown in formula (2): Among them, k is the size of the convolution kernel, k=3, f(i,j) is the concatenated data after calculation by the TCN module, g(ki,kj) is the convolution kernel, and b(i,j) is the bias.
[0036] After each convolution calculation of the CNN module 20, batch normalization (BN) and ReLU activation function processing are performed to improve the training efficiency and performance of the neural network. Batch normalization normalizes each small batch of data so that the distribution of each layer input remains relatively stable, reduces the dependence on parameter initialization during training, and reduces the sensitivity of the model to the initial weights. The normalization process of each batch adds a little noise, which has a regularization effect to a certain extent and helps prevent the model from overfitting. The ReLU activation function can increase the nonlinearity of the model, help alleviate the gradient vanishing problem in deep networks, and enable deeper networks to be successfully trained. The calculation process of batch normalization is shown in formula (3):
[0037]
[0038] Among them, x i represents the i-th sample, m is the batch size, is the mean, is the variance, ε is a small constant used to prevent division by zero, and γ and β are learnable parameters for scaling and offset, respectively.
[0039] Finally, the data is transformed into a one-dimensional signal feature through the full connection operation of the full connection module 30, and finally a mapping relationship is established with the impact source SoC coordinate (x, y). MSE Loss (also known as square error loss) is a loss function used to measure the difference between two vectors. The formula of MSE Loss is shown in the following formula (4):
[0040]
[0041] in, They represent the true value and corresponding predicted value of the i-th sample respectively, and N is the number of samples.
[0042] MSE Loss calculates the mean of the squared errors between model predictions and labels, where the goal is to minimize the difference between predicted values and actual values, and is often used in regression task analysis.
[0043] In the process of deep learning back propagation, the optimizer is an optimization algorithm for solving the loss function, guiding the various parameters of the loss function to update the appropriate values in the right direction, so that the parameter values of the updated loss function are constantly approaching the global minimum. The RMSProp optimization method is adopted in the present invention, which is an adaptive learning rate optimization algorithm, which aims to solve the problem of learning rate adjustment in the gradient descent process. By performing an exponentially weighted moving average on the square of the gradient, the learning rate of each parameter is adjusted to ensure that the parameter update is stable and converges quickly. It can effectively prevent the oscillation phenomenon in the gradient descent process by performing an exponentially weighted moving average on the square of the gradient. The optimization algorithm first initializes the exponentially weighted moving average and time step of the square of the gradient, and then calculates the current gradient, and updates the exponentially weighted moving average and learning rate of the square of the gradient, as shown in formula (5).
[0044] E[g 2 ]0=0,t=0
[0045] E[g 2 ] t =βE[g 2 ] t-1 +(1-β)g t 2
[0046]
[0047] Where E[g 2 ] t is the exponentially weighted moving average of the squared gradient, g t is the current gradient, β is the decay rate, θ t is the parameter value at the current moment, η is the global learning rate, and ε is a small constant used for numerical stability. The recommended default value is 1e-8.
[0048] Experimental example:
[0049] To verify the effectiveness of the method of the present invention, Figure 1As shown in Figure 2, an impact test was conducted on a carbon fiber composite reinforced wall panel with an overall size of 500mm×500mm. Four piezoelectric sensors (PZT) were arranged at the four corners 70mm away from the boundary, and the sensor spacing was 360mm×360mm. The impact test was conducted at different positions within the sensor envelope area, and the impact response signals were collected in real time, as shown in Figure 2, which shows the signals measured by the four sensors. A total of 1745 impact response signals were collected.
[0050] The algorithm of the present invention was implemented through PyTorch induction experiment and two NVIDIA GeForce RTX3090s. Adam was used to optimize the model parameters, the weight decay was 1e-5, and the initial learning rate was 1e-4. The cosine annealing algorithm was used to gradually reduce the learning rate. The ratio of the training set and the test set was 8:2, the batch size was set to 32, and a total of 100 epochs were iterated. The loss curve is shown in the figure. Figure 4a , Figure 4b As shown in Figure 2, the impact position identification results are as follows: Figures 5a to 5d As shown in the figure. After multiple iterations of training, R 2 The value of 0.9815 indicates that the model has a good fitting effect.
[0051] The results show that the deep learning algorithm can be used to learn the relationship between the impact response signal and the impact position without analyzing the signal characteristics. The deep learning method for composite material impact position identification has broad application prospects.
[0052] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
[0053] The data set sent to the deep learning network has a total of 1745 data items, each of which has a dimension of 4 times 5000, where 4 refers to the data measured by four sensors and each sensor has 5000 sampling points.
[0054] The four TCN modules of the deep learning network structure learn the data from four sensors respectively. The input of the TCN module is the shock stress wave signal measured in the experiment, and the output is the shock stress wave signal feature. The advantages of the TCN module are that it can effectively capture the dependency relationship over a long period of time, has stable gradients, supports parallel computing, flexibly controls the receptive field, reduces overfitting through parameter sharing, and is more stable during back propagation.
[0055] The input of the CNN module is the shock stress wave signal feature, and the output is the shock wave signal feature. After the full connection processing of the full connection module, the output obtains the shock source coordinates. After the TCN module processing, the calculation of the CNN module can learn the information on the four sensors and perform feature fusion on the four shock wave response signals.
[0056] It should be understood that the method of the present application may also be provided with the detection of impact stress wave signals by more sensors, such as 5 or 6 sensors, or fewer sensors, such as 3 sensors.
[0057] The number of convolutional layers in the CNN module can be set to one or more layers according to the learning effect.
[0058] It should be understood that the detailed description of the technical solutions of the present invention by means of the preferred embodiments is illustrative rather than restrictive. A person skilled in the art may modify the technical solutions described in each example or replace some of the technical features by equivalents based on reading the specification of the present invention, but these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each example of the present invention.
Claims
1. Impact position identification method of composite plate structure based on deep learning of impact stress wave, Features Includes steps: Step S1: Establishing an impulse response signal data set; Step S2: Establishing a deep learning model for impact position recognition based on a TCN-CNN network learning model, wherein the TCN-CNN network learning model includes a TCN module that uses the impact stress wave signal measured by each sensor as the impact response signal and outputs the sequence features of the impact stress wave signal; a CNN module that uses the sequence features of the impact stress wave signal as input and outputs the high-dimensional abstract features of the impact stress wave signal; and a fully connected module that uses the high-dimensional abstract features of the impact stress wave signal as input and outputs the impact source coordinates to establish a mapping relationship between the impact response signal and the impact position coordinates; Among them, the time series features of the impact stress wave signal output by the TCN module are sent to the CNN module after convolution.
2. The method for identifying impact positions of composite material plate structures based on deep learning of impact stress waves according to claim 1, It is characterized in that The method of establishing a shock response signal data set includes pasting a sensor on a composite material plate, performing shock experiments at different positions in the sensor envelope area, collecting shock response signals in real time and recording coordinate positions, establishing a database and dividing the training set and the test set according to a preset ratio.
3. The method for identifying impact positions of composite material plate structures based on deep learning of impact stress waves according to claim 1, It is characterized in that In the TCN module, a dilated convolution calculation method is used to increase the size of the convolution kernel by inserting a fixed number of spaces between the elements of the standard convolution kernel.
4. The method for identifying impact positions of composite material plate structures based on deep learning of impact stress waves according to claim 1, It is characterized in that Batch normalization (BN) and activation function processing are performed after each convolution calculation of the CNN module.
5. The method for identifying impact positions of composite material plate structures based on deep learning of impact stress waves according to claim 1, It is characterized in that The normalization process for each batch adds a little noise to prevent the model from overfitting.
6. The method for identifying impact positions of composite material plate structures based on deep learning of impact stress waves according to claim 1, It is characterized in that The activation function adopts the ReLU activation function.
7. The method for identifying impact positions of composite material plate structures based on deep learning of impact stress waves according to claim 1, It is characterized in that The RMSProp optimization method is used to adjust the learning rate of each parameter by performing an exponentially weighted moving average on the square of the gradient. It includes first initializing the exponentially weighted moving average of the square of the gradient and the time step, then calculating the current gradient, and updating the exponentially weighted moving average of the square of the gradient and the learning rate. Where E[g 2 ] t is the exponentially weighted moving average of the squared gradient, g t is the current gradient, β is the decay rate, θ t is the parameter value at the current moment, η is the global learning rate, and ε is a small constant for numerical stability.
Citation Information
Patent Citations
Impact localization method based on phase-sensitive optical reflection and deep learning of convolutional neural network
CN108645498A
Concrete structure internal defect imaging system and method based on sound vibration signal deep learning
CN113219063A
Composite material structure impact positioning method based on CNN-LSTM-Attention
CN118069979A
Cited By
Small-scale sensing array composite material impact positioning method based on deep learning
CN119846559A
Method for impact localization of small scale sensor array composites based on deep learning
CN119846559B
Recognition method for impact position of net-shaped structure based on bionic principle
CN120850833A
