Low-speed impact positioning system and method for composite laminated structure

Abstract

The invention discloses a low-speed impact positioning system and method for a composite laminated structure, and relates to the technical field of impact positioning, and the system comprises a signal collection module which is used for obtaining an impact response signal on a composite laminated structure test piece; the signal processing module is used for executing discrete wavelet transform and short-time Fourier transform on the collected impact response signal; the image enhancement module is used for carrying out image enhancement processing on the short-time Fourier time-frequency spectrogram; the model construction module is used for constructing a composite laminated structure impact positioning model; and the impact positioning module is used for performing impact positioning on the image data through a composite laminated structure impact positioning model, determining an impact position and performing positioning error calculation and result analysis on a prediction result. According to the technical scheme, the impact position of the composite laminated structure under the low-speed impact condition can be accurately recognized, the impact positioning precision and efficiency can be improved, and high application value is achieved.

Description

Technical Field

[0001] The invention relates to the technical field of impact positioning, and in particular to a low-speed impact positioning system and method for a composite laminated structure. Background Art

[0002] Compared with traditional materials such as aluminum and steel, composite laminates have been widely used in aerospace, wind turbines, shipbuilding and other fields due to their light weight, high strength and good fatigue resistance. However, in actual engineering projects, such materials are extremely susceptible to low-speed impacts caused by gravel, bird strikes, and tool drops during manufacturing and maintenance, which can cause invisible internal damage and pose a hidden danger to the overall balance of the structure. In this context, impact location technology, as a key link in the structural health monitoring process, is crucial to ensuring the stability of the structure.

[0003] Shock location is a passive structural health monitoring technology that directly uses the response signal generated by the structure itself for structural health monitoring. At present, most of the existing shock location methods have defects in one aspect. For example, the shock location method based on time difference has been widely studied and applied in practical engineering. It uses the geometric relationship between the wave arrival time difference, wave velocity and sensor position to achieve shock location. However, due to the interference caused by the dispersion characteristics of the shock signal, external noise and boundary reflection, this method is difficult to obtain accurate wave arrival time, which is easily limited in practical applications; time reversal focusing method, this method does not need to obtain the wave arrival time and waveguide structure in advance, has good focusing ability and noise resistance, and is suitable for complex material structures, but this method requires expensive full-wave field measurement equipment to obtain the transfer function or requires a large number of sensors to synthesize the focusing signal, which limits the wide application of this method to a certain extent; the shock location method based on the database obtains the reference signal from the sampling point on the structure to establish a database, and determines the location of the shock by comparing the similarity between the real shock signal and the reference signal. It can adapt to more complex positioning environments. However, the positioning accuracy of this method is too dependent on the data in the database and has great limitations.

[0004] Therefore, on this basis, the present application proposes a low-speed impact positioning system and method for a composite laminated structure. Summary of the invention

[0005] Technical Purpose In order to solve the above problems, the purpose of the present invention is to provide a low-speed impact positioning system and method for a composite laminate structure, which can not only be applied to various complex material structures, but also can realize the precise positioning of the impact load. The system and method have a wide range of application scenarios, can realize the accurate identification of the impact position of the composite laminate structure under low-speed impact conditions, and can improve the impact positioning accuracy and efficiency, and has low application cost.

[0006] Technical Solution In order to achieve the above-mentioned purpose, the present invention provides a low-speed impact positioning system and method for a composite laminated structure, which combines discrete wavelet transform and short-time Fourier transform signal processing methods to process the impact response signal, and converts the one-dimensional time domain signal into a two-dimensional short-time Fourier time spectrum diagram on the basis of extracting the high-frequency component of the signal to display the intensity distribution of the signal at different times and frequencies. The DenseNet network and CBAM attention mechanism are applied to the positioning of the impact load, which solves the problem that the weak impact signal is easily submerged by strong noise interference on a one-dimensional time scale when the sensor nodes are sparse.

[0007] In a first aspect, the present invention provides a composite laminate structure low-speed impact positioning system, comprising: A signal acquisition module, used to obtain an impact response signal on a composite laminate structure specimen; A signal processing module, used for performing discrete wavelet transform and short-time Fourier transform on the collected impulse response signal; An image enhancement module is used to perform image enhancement processing on the short-time Fourier spectrum diagram; A model building module, used to build an impact positioning model for a composite laminate structure; The impact positioning module is used to perform impact positioning on image data through the composite laminate structure impact positioning model, determine the impact position, and calculate the positioning error and analyze the results of the prediction results.

[0008] Furthermore, the signal acquisition module obtains the impact response signal on the composite laminate structure specimen through the strain sensor.

[0009] Furthermore, the number of the strain sensors is preferably 4, which are respectively deployed at four corner points of a selected square monitoring area on the specimen, and their axes are parallel to the diagonals of the squares at the corner points.

[0010] Furthermore, a two-dimensional rectangular coordinate system is established on the square monitoring area of ​​the composite laminated structure with the lower left vertex of the monitoring area as the coordinate origin, and the monitoring area is divided into grids of moderate size and uniform distribution to locate the impact position.

[0011] Furthermore, the grid points are repeatedly struck multiple times by an impact hammer so that the strain sensor can collect impact signals.

[0012] Furthermore, the signal processing module removes the low-frequency components in the impulse response signal by performing discrete wavelet decomposition on the collected impulse response signal, performs inverse discrete wavelet transform on the high-frequency components in the impulse response signal to obtain a one-dimensional time domain reconstructed signal, and performs short-time Fourier transform on the one-dimensional time domain reconstructed signal to obtain a two-dimensional short-time Fourier time spectrum. This enhances the visualization effect of the impulse signal and is conducive to the deep learning network to extract features from the impulse response signal.

[0013] Furthermore, image enhancement processing is performed on the two-dimensional short-time Fourier spectrum map to expand the image data volume of the two-dimensional short-time Fourier spectrum map, thereby obtaining an image sample data set.

[0014] Furthermore, the image sample data set includes multiple groups of image data and corresponding label information, and is divided into a training set and a validation set according to a certain ratio.

[0015] Furthermore, a portion of image data is selected as a test data set, and no image enhancement processing is performed on this portion of image data.

[0016] Furthermore, the maximum and minimum values ​​of the label information corresponding to the image data in the training set and the test set are normalized to improve the convergence and generalization capabilities of the model. The normalization formula is:

[0017] In the formula, is the original coordinate value; is the normalized coordinate value; is the maximum value of the original data; is the minimum value of the original data.

[0018] Furthermore, the DenseNet-CBAM model is implemented based on the DenseNet-121 network and the CBAM attention mechanism, and is trained by the AdamW optimizer. The learning rate is preset to 0.001, the weight decay coefficient is preset to 1e-4, the learning rate is adjusted by the cosine annealing scheduling algorithm, the number of iterations of model training is preset to 50, the image training batch is preset to 10, and the Huber function is selected as the loss function. The loss function formula is:

[0019] In the formula, is a true value; is the predicted value; To adjust the threshold parameter of the loss function expression, it is preferably 3.0.

[0020] Through the dense connection structure of DenseNet, feature reuse and information flow are increased, ensuring that each layer can directly receive feedback from the error signal and self-adjust accordingly to optimize the final positioning result, effectively weakening the interference of noise in actual acquisition, which is conducive to improving the accuracy of impact positioning. By optimizing the model's selection of feature channels and enhancing the expression ability of key areas of the image, the impact positioning performance of the regression model and the accuracy of the impact prediction results are improved.

[0021] Furthermore, the training process of the DenseNet-CBAM model includes: (1) Convert the size of the color image in the image sample dataset to , and enter a The convolutional layer with a step size of 2 extracts image features and then inputs a A max pooling layer with a stride of 2 is used to reduce the size of the feature map; (2) The CBAM module is deployed at the post-maximum pooling layer of the DenseNet-121 model and at the pre-maximum position of the first DenseBlock module. The CBAM module receives the output of the maximum pooling layer and includes a channel attention module and a spatial attention module. (3) First, enter the channel attention module, enter the global maximum pooling layer and the global average pooling layer, and obtain two one-dimensional vectors. The two one-dimensional vectors are respectively sent to a shared multi-layer perceptron for processing. The multi-layer perceptron consists of two convolutional layers, where the first layer is used for dimensionality reduction and the second layer is used for dimensionality increase to restore the original number of channels. Between the two layers, there are ReLU activation functions and Batch Normalization layers. The two vectors processed by the multi-layer perceptron are added together, and the channel attention weight map is obtained through the sigmoid activation function. The channel attention weight map and the input feature map are multiplied element by element to generate the input features required by the subsequent modules. (4) Next, enter the spatial attention module and perform channel-based global maximum pooling and global average pooling operations to obtain two two-dimensional vectors. The two two-dimensional vectors are concatenated in the channel dimension and then enter a The convolution layer with a step size of 1 performs convolution operation, obtains the spatial channel attention weight map through the sigmoid activation function, and multiplies the spatial channel attention weight map with the input feature map of the module to obtain the data processed by the CBAM module; (5) The data processed by the CBAM module enters the first Dense Block module. The first DenseNet module contains six Dense Layers, each of which consists of two Batch Normalization layers, two ReLU activation layers, and one A convolutional layer and a The convolutional layers are arranged in the following order: ; (6) The output data of the first Dense Block module is input to the first Transition Layer, which contains a Batch Normalization layer, a ReLU activation layer, and a The convolutional layer and a The average pooling layer is arranged in the following order: ; The Transition Layer is used to compress the image. Assuming that the number of channels of the feature map obtained by connecting the Dense Block to the Transition Layer is m, θm features can be generated after the Transition Layer, where θ is the compression coefficient, and the range is ,when When , the number of features does not change after the Transition Layer, that is, there is no compression state. When , the features will be compressed; (7) The DenseNet-CBAM model includes four Dense Block modules and three Transition Layer layers. Each Dense Block contains 6, 12, 24 and 16 Dense Layer structures respectively. The Dense Blocks are connected through the Transition Layer layer. The data output by the last Dense Block enters a A global average pooling layer and a two-dimensional fully connected layer finally obtain the two-dimensional coordinates of the impact position.

[0022] Furthermore, the impact location model of the composite laminated structure is used to perform impact location on the image data to obtain the predicted coordinate value of the impact position, and the error calculation formula of the predicted coordinate value is:

[0023] In the formula, is the positioning error value; is the abscissa of the actual impact position; is the abscissa of the predicted impact position; is the ordinate of the actual impact position; is the ordinate of the predicted impact position.

[0024] Furthermore, the system also includes a reinforcement learning module, which selects the next execution action based on the current sensor layout and the location of the hitting point. The execution action includes adjusting the sensor layout and selecting the hitting location. After the execution action is completed, the environmental feedback is calculated based on the accuracy of the impact positioning, and then the optimal sensor layout and hitting point location are searched based on the maximized cumulative feedback.

[0025] Furthermore, the reinforcement learning module also includes a historical impact response matrix, including response data of all past impact points.

[0026] Furthermore, the environmental feedback defines feedback obtained after performing an action, for example, if the distance between the actual impact position and the predicted position is less than a certain threshold, positive feedback is given, and if it exceeds the threshold, negative feedback is given.

[0027] Furthermore, the reinforcement learning module accumulates environmental feedback through a policy gradient method, and the policy gradient formula is as follows:

[0028] in, For the time step Strategy parameters of For the time step The updated strategy parameters of is the time step The learning rate; is the gradient of the performance function with respect to the policy parameters; Represents the weight attenuation coefficient, which is used for regularization to prevent overfitting; Represents the L2 norm of the policy parameters.

[0029] It improves the accuracy of impact positioning by continuously optimizing the sensor layout and impact point location, which can reduce the workload of sample collection and training. The algorithm is relatively simple, reducing the complexity of the algorithm.

[0030] Furthermore, the system also includes a linear regression module for preliminarily estimating the location of the impact when the amount of data is small or the computing resources are limited. This module enables the system to quickly build a model and make a preliminary estimate when the amount of data is small or the computing resources are limited, and find the optimal regression coefficient and intercept by minimizing the error between the predicted value and the true value, thereby improving the estimation accuracy of the impact location.

[0031] In a second aspect, the present invention further provides a method for low-speed impact positioning of a composite laminate structure, the method being based on the system described in the first aspect, comprising: The method comprises: Acquiring impact response signals on composite laminate structure specimens through sensors; Performing discrete wavelet transform and short-time Fourier transform on the collected impulse response signal; Perform image enhancement processing on the short-time Fourier spectrum; Constructing impact location model of composite laminated structure based on DenseNet-CBAM model; The impact location model of the composite laminated structure is used to locate the impact of the image data, determine the impact position, and calculate the positioning error and analyze the results of the prediction results.

[0032] In the third aspect, the present invention also provides a computer device, comprising a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the computer device performs at least one step of the aforementioned method for low-speed impact positioning of a composite laminated structure.

[0033] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, at least one step of the aforementioned method for low-speed impact positioning of a composite laminate structure is implemented.

[0034] The present invention combines the signal processing methods of discrete wavelet transform and short-time Fourier transform to process the impact response signal. On the basis of extracting the high-frequency components of the signal, the one-dimensional time domain signal is converted into a two-dimensional short-time Fourier time spectrum to show the intensity distribution of the signal at different times and frequencies. The DenseNet network and CBAM attention mechanism are applied to the positioning of the impact load, solving the problem that the weak impact signal is easily submerged by strong noise interference on a one-dimensional time scale when the sensor nodes are sparse. The system and method have a wide range of application scenarios, can realize the accurate identification of the impact position of the composite laminated structure under low-speed impact conditions, and can improve the impact positioning accuracy and efficiency, with low application cost.

[0035] Beneficial Effects By implementing the composite laminate structure low-speed impact positioning system and method provided by the present invention, the following technical effects are achieved: (1) The present invention combines the signal processing methods of discrete wavelet transform and short-time Fourier transform to process the impulse response signal. On the basis of extracting the high-frequency components of the signal, the one-dimensional time domain signal is converted into a two-dimensional short-time Fourier time spectrum to display the intensity distribution of the signal at different times and frequencies. This enhances the visualization effect of the impulse signal and facilitates the deep learning network to extract features from the impulse response signal.

[0036] (2) The DenseNet network and CBAM attention mechanism are applied to the location of impact loads, which solves the problem that weak impact signals are easily submerged by strong noise interference on a one-dimensional time scale when the sensor nodes are sparse. Through the dense connection structure of DenseNet, feature reuse and information flow are increased, ensuring that each layer can directly receive feedback from the error signal and self-adjust to optimize the final positioning result, effectively weakening the interference of noise in actual acquisition, which is conducive to improving the accuracy of impact positioning. By optimizing the model's selection of feature channels and enhancing the expression ability of key areas of the image, the impact positioning performance of the regression model and the accuracy of the impact prediction results are improved.

[0037] (3) Accumulate environmental feedback through the policy gradient method, and search for the optimal sensor layout and impact point location based on maximizing the accumulated feedback. By continuously optimizing the sensor layout and impact point location, the accuracy of impact positioning can be improved, the workload of sample collection and training can be reduced, the algorithm is relatively simple, and the complexity of the algorithm is reduced.

[0038] (4) The linear regression algorithm is used to preliminarily estimate the location of the impact when the amount of data is small or the computing resources are limited. This allows the system to quickly build a model and make a preliminary estimate when the amount of data is small or the computing resources are limited. The optimal regression coefficient and intercept are found by minimizing the error between the predicted value and the true value, thereby improving the estimation accuracy of the impact location. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to make the above-mentioned composite laminated structure low-speed impact positioning system and method of the present invention more obvious and easy to understand, the drawings required for use in the specific implementation methods of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 A schematic diagram showing a process flow of a low-speed impact positioning method for a composite laminate structure; Figure 2 A schematic diagram showing a signal acquisition system for a low-speed impact positioning method for a composite laminate structure based on a DenseNet-CBAM model; Figure 3 A schematic diagram showing signal acquisition of a low-speed impact positioning method for a composite laminate structure based on the DenseNet-CBAM model; Figure 4 A schematic diagram showing the monitoring area of ​​the composite laminate structure and the relative positions of the strain sensors; Figure 5 A schematic diagram showing the structure of the DenseNet-CBAM model; Figure 6 Box plots showing the positioning results analysis of the three methods. DETAILED DESCRIPTION

[0041] Embodiment 1: A composite laminate structure low-speed impact positioning system and method are provided, and a flow chart of the composite laminate structure low-speed impact positioning method is shown in FIG. Figure 1 As shown, the composite laminate structure low-speed impact positioning system includes: a signal acquisition module, a signal processing module, an image enhancement module, a model building module and an impact positioning module, specifically the composite laminate structure low-speed impact positioning system and method of the steps described below.

[0042] The signal acquisition module is used to obtain the impact response signal on the composite laminate structure specimen.

[0043] like Figure 2 , Figure 3 and Figure 4 As shown, in The composite laminate specimens were selected The square monitoring area is symmetrically pasted with four strain sensors at the four corners of the monitoring area. The axial direction of the strain sensor is parallel to the diagonal of the square at the corner, and the angle between the sides of the monitoring area is 45°. A two-dimensional rectangular coordinate system is established on the monitoring area of ​​the composite laminate structure, and the lower left corner vertex of the monitoring area is defined as the origin of the coordinate system. , define the x-axis as a horizontal straight line starting from the origin, and the y-axis as a straight line starting from the origin and perpendicular to the x-axis. Divide the monitoring area into moderately sized and evenly distributed grids, generating a total of 81 grid points, use an impact hammer to repeatedly knock the 81 grid points 3 times, and then randomly select 5 grids and 15 non-grid points for knocking once, set the sampling frequency of the signal acquisition device to 10kHz, and collect the impact signal collected by the strain sensor.

[0044] The signal processing module is used to perform discrete wavelet transform and short-time Fourier transform on the collected impulse response signal.

[0045] In order to reduce the interference of low-frequency components in the collected impulse response signal, the Db6 wavelet function is used to perform 6-layer discrete wavelet decomposition on the impulse response signal. After removing the low-frequency components, the remaining high-frequency components are subjected to inverse discrete wavelet transform to obtain a one-dimensional time domain reconstructed signal. In order to intuitively show the characteristics of the impulse response signal, the one-dimensional time domain reconstructed signal is subjected to short-time Fourier transform, and a Hamming window function is used. The window length is set to 32, and the overlap length is set to 24, and finally a two-dimensional short-time Fourier frequency spectrum is obtained.

[0046] The image enhancement module is used to perform image enhancement processing on the short-time Fourier spectrum.

[0047] The 81 grid points were repeatedly tapped three times, and the obtained two-dimensional short-time Fourier spectrum was enhanced to expand the image sample data volume. Gaussian, Poisson, multiplicative and salt and pepper noises were added to the image respectively to expand the image sample data set by five times to obtain 1215 images. The 1215 images were then flipped, histogram equalized and contrast enhanced to expand the image sample data set by four times again to obtain a total of 4860 images, which finally constituted the image sample data set. The image sample data set also included label information corresponding to the image data, and the data set was divided into a training set and a validation set according to a certain ratio.

[0048] The maximum and minimum values ​​of the label information corresponding to the image data in the training set and the test set are normalized to improve the convergence and generalization capabilities of the model. The normalization formula is:

[0049] In the formula, is the original coordinate value; is the normalized coordinate value; is the maximum value of the original data; is the minimum value of the original data.

[0050] The model building module is used to build a composite laminate structure impact positioning model. The DenseNet-CBAM model structure is as follows: Figure 5 shown.

[0051] The DenseNet-CBAM model is implemented based on the DenseNet-121 network and the CBAM attention mechanism, and is trained using the AdamW optimizer. The learning rate is preset to 0.001, the weight decay coefficient is preset to 1e-4, and the learning rate is adjusted using the cosine annealing scheduling algorithm. The number of iterations of model training is preset to 50, the image training batch is preset to 10, and the Huber function is selected as the loss function. The loss function formula is:

[0052] In the formula, is a true value; is the predicted value; To adjust the threshold parameter of the loss function expression, it is preferably 3.0.

[0053] The training process of the DenseNet-CBAM model includes: (1) Convert the size of the color image in the image sample dataset to , and enter a The convolutional layer with a step size of 2 extracts image features and then inputs a A max pooling layer with a stride of 2 is used to reduce the size of the feature map; (2) The CBAM module is deployed at the post-maximum pooling layer of the DenseNet-121 model and at the pre-maximum position of the first DenseBlock module. The CBAM module receives the output of the maximum pooling layer and includes a channel attention module and a spatial attention module. (3) First, enter the channel attention module, enter the global maximum pooling layer and the global average pooling layer, and obtain two one-dimensional vectors. The two one-dimensional vectors are respectively sent to a shared multi-layer perceptron for processing. The multi-layer perceptron consists of two convolutional layers, where the first layer is used for dimensionality reduction and the second layer is used for dimensionality increase to restore the original number of channels. Between the two layers, there are ReLU activation functions and Batch Normalization layers. The two vectors processed by the multi-layer perceptron are added together, and the channel attention weight map is obtained through the sigmoid activation function. The channel attention weight map and the input feature map are multiplied element by element to generate the input features required by the subsequent modules. (4) Next, enter the spatial attention module and perform channel-based global maximum pooling and global average pooling operations to obtain two two-dimensional vectors. The two two-dimensional vectors are concatenated in the channel dimension and then enter a The convolution layer with a step size of 1 performs convolution operation, obtains the spatial channel attention weight map through the sigmoid activation function, and multiplies the spatial channel attention weight map with the input feature map of the module to obtain the data processed by the CBAM module; (5) The data processed by the CBAM module enters the first Dense Block module. The first DenseNet module contains six Dense Layers, each of which consists of two Batch Normalization layers, two ReLU activation layers, and one A convolutional layer and a The convolutional layers are arranged in the following order: ; (6) The output data of the first Dense Block module is input to the first Transition Layer, which contains a Batch Normalization layer, a ReLU activation layer, and a The convolutional layer and a The average pooling layer is arranged in the following order: ; The Transition Layer is used to compress the image. Assuming that the number of channels of the feature map obtained by connecting the Dense Block to the Transition Layer is m, θm features can be generated after the Transition Layer, where θ is the compression coefficient, and the range is ,when When , the number of features does not change after the Transition Layer, that is, there is no compression state. When , the features will be compressed; (7) The DenseNet-CBAM model includes four Dense Block modules and three Transition Layer layers. Each Dense Block contains 6, 12, 24 and 16 Dense Layer structures respectively. The Dense Blocks are connected through the Transition Layer layer. The data output by the last Dense Block enters a A global average pooling layer and a two-dimensional fully connected layer finally obtain the two-dimensional coordinates of the impact position.

[0054] The impact location module is used to locate the impact of image data through the composite laminate structure impact location model, determine the impact position, and calculate the positioning error and analyze the results of the prediction results.

[0055] The impact location model of the composite laminated structure is used to perform impact location on the image data to obtain the predicted coordinate value of the impact position. The error calculation formula of the predicted coordinate value is:

[0056] In the formula, is the positioning error value; is the abscissa of the actual impact position; is the abscissa of the predicted impact position; is the ordinate of the actual impact position; The model provided by the present application has lower error and smaller model parameter volume, and the impact positioning performance of the regression model and the accuracy of the impact prediction results can be guaranteed.

[0057] Figure 6 The box plot analysis of the positioning results of the DenseNet121-CBAM composite laminated structure impact positioning model described in this application and other models is shown. Figure 6 It can be seen that the model of the present application has higher positioning accuracy and lower error value compared with other models, and can be used for low-speed impact positioning of composite laminated structures.

[0058] Embodiment 2: Based on the foregoing embodiments, the system also includes a reinforcement learning module, which selects the next execution action according to the current sensor layout and the location of the striking point. The execution action includes adjusting the sensor layout and selecting the striking location. After the execution action is completed, the environmental feedback is calculated according to the accuracy of the impact positioning, and then the optimal sensor layout and striking point location are searched according to the maximized cumulative feedback.

[0059] The reinforcement learning module also includes a historical impact response matrix, including response data of all past impact points.

[0060] The environmental feedback defines the feedback obtained after performing the action, for example, if the distance between the actual impact position and the predicted position is less than a certain threshold, positive feedback is given, and if it exceeds the threshold, negative feedback is given.

[0061] The reinforcement learning module accumulates environmental feedback through the policy gradient method. The policy gradient formula is as follows:

[0062] in, For the time step Strategy parameters of For the time step The updated strategy parameters of is the time step The learning rate; is the gradient of the performance function with respect to the policy parameters; Represents the weight attenuation coefficient, which is used for regularization to prevent overfitting; The L2 norm of the strategy parameter is represented. By continuously optimizing the sensor layout and the impact point position, the accuracy of impact positioning can be improved, the workload of sample collection and training can be reduced, the algorithm is relatively simple, and the complexity of the algorithm is reduced. Verification shows that when obtaining an average error similar to that of the above embodiment, only about 80% of the training set data is required to complete it, indicating that adding a reinforcement learning module helps to reduce the workload of model training while maintaining high accuracy of impact positioning, effectively expand the applicable working conditions of the model, and reduce the time and economic costs of training.

[0063] Embodiment 3: On the basis of the above-mentioned embodiment, the system adds a linear regression module for preliminarily estimating the impact position when the data volume is not large or the computing resources are limited.

[0064] First, the impact response signals are collected from multiple strain sensors on the composite laminate structure, and the actual impact location coordinates are marked for each data point.

[0065] The sensor data most relevant to the impact location is selected as the feature, and the features are standardized or normalized to be on the same scale.

[0066] The strain response value of each sensor is taken as a dimension of the feature vector.

[0067] The data is divided into a training set and a test set, and the data in the training set is used to fit the model. The calculation formula of the least squares method is:

[0068] in, is the updated bias term; is the initial value of the bias term, which indicates the default prediction of the model when there is no feature information; is the adjustment coefficient of the bias term, which is used to control the dynamic adjustment range of the bias term; is the number of samples in the dataset; For the The actual impact position of the data points; is the average of the target values; is the updated weight vector; For the The feature vector of the data point; is the mean value of the feature vector.

[0069] The mean square error indicator is used to ensure that the prediction error of the model is within an acceptable range.

[0070] Assume there is a dataset: Dataset 1: Data point 1, sensor 1 strain x1 is 2, sensor 2 strain x2 is 3, impact position y is 5; Dataset 2: data point 2, sensor 1 strain x1 is 4, sensor 2 strain x2 is 6, impact position y is 10; Dataset 3: Data point 3, sensor 1 strain x1 is 5, sensor 2 strain x2 is 7, and impact position y is 12.

[0071] Assumptions , ;

[0072]

[0073]

[0074] The effect of the linear regression module is shown in Table 1.

[0075] Table 1. Summary of linear regression module effects

[0076] The prediction errors were low in all experiments, indicating that the linear regression model had high prediction accuracy and was able to learn from the sensor strain data and predict the impact location well.

[0077] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable non-transient storage media containing computer-usable program code.

[0078] The present invention can provide computer program instructions to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the system.

[0079] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions of the system.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions of the system.

Claims

1. A low-speed impact positioning system for a composite laminate structure, characterized in that: include: A signal acquisition module, used to obtain an impact response signal on a composite laminate structure specimen; A signal processing module, used for performing discrete wavelet transform and short-time Fourier transform on the collected impulse response signal; An image enhancement module is used to perform image enhancement processing on the short-time Fourier spectrum diagram; A model building module, used to build an impact positioning model for a composite laminate structure; The impact positioning module is used to perform impact positioning on image data through the composite laminate structure impact positioning model, determine the impact position, and calculate the positioning error and analyze the results of the prediction results.

2. The system according to claim 1, characterized in that: The signal acquisition module acquires the impact response signal on the composite laminate structure specimen through a strain sensor, and the strain sensor is deployed at the corner point of the selected monitoring area on the specimen.

3. The system according to claim 1, characterized in that: The signal processing module removes low-frequency components in the impulse response signal by performing discrete wavelet decomposition on the collected impulse response signal, performs inverse discrete wavelet transform on the high-frequency components in the impulse response signal to obtain a time domain reconstructed signal, and performs short-time Fourier transform on the time domain reconstructed signal to obtain a short-time Fourier time-frequency spectrum.

4. The system according to claim 1, characterized in that: The composite laminate structure impact location model is constructed based on the DenseNet-CBAM model; The DenseNet-CBAM model is implemented based on the DenseNet-121 network and the CBAM attention mechanism, trained by the AdamW optimizer, adjusted the learning rate by the cosine annealing scheduling algorithm, and selected the Huber function as the loss function. The loss function formula is: In the formula, is a true value; is the predicted value; It is the threshold parameter for adjusting the loss function expression.

5. The system according to claim 4, characterized in that: The impact location model of the composite laminated structure is used to perform impact location on the image data to obtain the predicted coordinate value of the impact position. The error calculation formula of the predicted coordinate value is: In the formula, is the positioning error value; is the abscissa of the actual impact position; is the abscissa of the predicted impact position; is the ordinate of the actual impact position; is the ordinate of the predicted impact position.

6. The system according to claim 1, characterized in that: The system also includes a reinforcement learning module, which selects the next execution action based on the current sensor layout and the location of the hitting point. The execution action includes adjusting the sensor layout and selecting the hitting location. After the execution action is completed, the environmental feedback is calculated based on the accuracy of the impact positioning, and then the optimal sensor layout and hitting point location are searched based on the maximized cumulative feedback.

7. The system according to claim 1, characterized in that: The system also includes a linear regression module for preliminarily estimating the location where the impact occurs when the amount of data is small or the computing resources are limited.

8. A method for low-speed impact positioning of a composite laminate structure, characterized in that: The method is implemented based on the system according to any one of claims 1 to 7: The method comprises: Acquiring impact response signals on composite laminate structure specimens through sensors; Performing discrete wavelet transform and short-time Fourier transform on the collected impulse response signal; Perform image enhancement processing on the short-time Fourier spectrum; Constructing impact location model of composite laminated structure based on DenseNet-CBAM model; The impact location model of the composite laminated structure is used to locate the impact of the image data, determine the impact position, and calculate the positioning error and analyze the results of the prediction results.

9. A computer device comprising a processor and a memory, wherein the processor is connected to the memory, and the memory is used to store a computer program, wherein: The processor is configured to execute the computer program stored in the memory, so that the computer device performs at least one step of the method according to claim 8.

10. A computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, characterized in that: The computer program implements at least one step of the method of claim 8 when executed.

Images