A method for dynamic modeling of composite material structures in electric aircraft based on attention meta-learning
By using a Transformer model based on attention meta-learning, the problems of insufficient accuracy and generalization ability of traditional dynamic modeling methods under small sample data conditions are solved, realizing efficient dynamic modeling of composite material structures of electric aircraft and improving the adaptability and prediction accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENYANG AEROSPACE UNIVERSITY
- Filing Date
- 2025-07-18
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional dynamic experiments are complex and expensive, and traditional dynamic modeling methods lack accuracy and generalization ability under small sample data conditions, making it difficult to meet the dynamic modeling requirements of composite material structures for electric aircraft.
We employ an attention-based meta-learning Transformer model, dynamically adjusting attention weights through meta-training and testing phases to quickly adapt to new tasks, achieving multi-task learning and knowledge transfer, and constructing a dynamic model of composite material structures for small-sample electric aircraft.
It improves the accuracy and generalization ability of dynamic modeling under small sample data conditions, reduces computation time, and enhances the model's adaptability to new tasks and prediction accuracy.
Smart Images
Figure CN121072017B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of dynamics modeling technology, specifically to a method for dynamics modeling of composite material structures of electric aircraft based on attention meta-learning. Background Technology
[0002] With green aviation becoming a focal point, electric aircraft offer a viable technological solution for achieving green aviation due to their low carbon emissions. Composite material structures possess numerous advantages, including high specific strength, specific stiffness, lightweight properties, fatigue resistance, and corrosion resistance. Therefore, the RX series general aviation electric aircraft utilizes various composite material composite structures, such as all-composite fuselages, composite wing boxes, and composite assembly structures, which account for over 80% of the aircraft's composition. Due to the lightweighting trend in electric aircraft, the RX series requires these complex composite material composite structures to possess superior dynamic performance to reduce the probability of failure. However, traditional dynamic experiments on complex composite material composite structures are extremely complex and challenging, requiring specialized molds and fixtures for forming specimens, time-consuming and expensive vibration testing, and difficulty in controlling the experimental environment temperature—all of which limit their development. Therefore, accurate dynamic modeling methods are of great importance.
[0003] Dynamics modeling can be broadly categorized into two types: model-based methods and data-driven methods. Model-based methods rely heavily on extensive expertise, and the models typically exhibit weak generalization ability. With the development of artificial intelligence, deep learning and machine learning, due to their powerful feature extraction and generalization capabilities, have been widely applied to dynamics modeling. However, both rely on large amounts of data and exhibit lower accuracy and generalization ability under conditions of small sample sizes. Conversely, meta-learning algorithms have attracted significant interest due to their superior generalization ability, and they are particularly effective in handling tasks with small sample sizes, complex systems, and nonlinear characteristics. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method for dynamic modeling of composite material structures of small-sample electric aircraft based on attention meta-learning. Specifically, it introduces an attention meta-learning Transformer model, which can more effectively and accurately model the dynamics of composite material structures of small-sample electric aircraft. This solves the problem that existing model-based methods rely heavily on a large amount of professional knowledge, and the models usually exhibit weak generalization ability.
[0005] To achieve the above objectives, this invention provides the following technical solution: a method for dynamic modeling of composite material structures in small-sample electric aircraft based on attention meta-learning, comprising the following steps:
[0006] Step 1: The dataset used comes from vibration experiments conducted on composite laminated wing box structures, composite laminated air beam structures, and composite conical-cylindrical composite shells under different temperature conditions. Given the limited data, various composite structures and ambient temperatures were divided into independent task samples. Each task includes vibration test data for the composite laminated wing box structure, composite laminated air beam structure, and composite conical-cylindrical composite shell at their respective temperatures; the data was then divided into a meta-training set and a meta-test set.
[0007] Step 2: Dataset partitioning. Specifically, vibration test data of CLWBs structures at 20℃, 50℃, 100℃ and 125℃ are designated as test tasks; vibration test data of CLWBs structures under the remaining temperature conditions and CLABs and CCCCs under all conditions are designated as training tasks; each task is divided into support set and query set in a 4:1 ratio.
[0008] Step 3: Import the support set data and labels from the meta-training set into the model for training. Train the AMLT model until convergence, using the Adam optimization algorithm and the MSE loss function. For each task, quickly update the model parameters using their respective support set data to achieve optimal parameters for the current task environment; the update formula is: Secondly, the AMLT model uses a meta-learner to identify shared features in multi-temperature and multi-structure tasks, and dynamically adjusts the attention weights of new tasks based on the similarity between the new task and the existing dataset; specifically, tasks that show high similarity to the target task are given higher weights, while those with low similarity are given lower weights.
[0009] The query set data and labels from the meta-training set are imported into the model for testing. The meta-learner uses query sets from these different tasks to evaluate the performance of the base learner, and updates the initial model parameters again by calculating the error and performing backpropagation; the update formula is: Repeat this process until training for all batch tasks is complete, and until convergence is achieved on all structure and temperature tasks.
[0010] Step 4: In the meta-testing phase, the AMLT model is trained using support set data partitioned from the test set, and the model parameters are updated to enhance its adaptation to new tasks. Subsequently, the query set partitioned from the test set is input into the model to evaluate the model's performance on new tasks, resulting in the dynamic model of CLWBs.
[0011] Step 5: Construct the following networks respectively: Model Agnostic Meta-Learning Transformer Network, Transformer Network, Deep Convolutional Neural Network, Long Short-Term Memory Neural Network, Convolutional Neural Network-Long Short-Term Memory Neural Network, and Bidirectional Long Short-Term Memory Network. Input the same training set into the model for training, and input the test set and labels into the trained model to obtain the amplitude-frequency response curves of CLWBs under these models respectively.
[0012] Step 6: Compare the dynamic modeling results under the seven models. The root mean square error (RMSE) is used to evaluate the prediction results. The formulas for calculating MSE and RMSE are as follows: In the formula, For predicted values, This is the actual value. n To test the sample length, the accuracy of the model is evaluated by comparing MSE and RMSE.
[0013] Preferably, step 1 normalizes the original vibration data. This method can transform the data into the range [0, 1]. The normalization operation for the original vibration data is expressed as follows: in x Indicates the initial data. x max and x min express x The maximum and minimum values, y This represents a normalized vector.
[0014] Preferably, the similarity in step 3 is determined by the Pearson similarity coefficient between different structural and temperature characteristics, and the formula for calculating the similarity coefficient is: in r Represents the similarity coefficient under one-dimensional features; n express n Dimensional features.
[0015] This invention provides a few-sample dynamics modeling method for composite material structures of electric aircraft based on attention meta-learning, which has the following advantages: It innovatively integrates meta-learning with dynamics modeling of composite material structures of electric aircraft. The AMLT model can leverage the rapid task adaptability and multi-task learning capabilities of MAML to quickly transfer and apply knowledge to new tasks, resulting in higher adaptability and generalization ability when facing new tasks. Since the AMLT model is trained on the dataset of the current task each time, it allows the network model to adjust its parameters according to the specific training task. Simultaneously, the AMLT model dynamically adjusts the influence weights for the target task during meta-training through an attention mechanism. Tasks exhibiting higher feature consistency with the target task receive higher weights, thus preserving prominent cross-task features. This mitigates the significant prediction errors caused by compromise models when predicting dynamic characteristics. By enabling the model to learn "meta-knowledge" across various tasks, optimal model initialization parameters for multiple tasks can be obtained. This helps to quickly iterate and converge the model with minimal data during actual prediction, thereby reducing the training data requirements of traditional neural networks. Furthermore, the model can quickly adapt to new tasks through several gradient descent steps, thus exhibiting optimal performance on these tasks. Compared to traditional deep learning models, the AMLT model has a strong ability to adapt to new tasks, faster computation time, and higher prediction accuracy in the dynamic modeling of composite material composite structures in electric aircraft. Attached Figure Description
[0016] Figure 1 This is a flowchart of the dynamic modeling method for composite material structures of electric aircraft based on meta-learning in this invention.
[0017] Figure 2 This is a flowchart of the AMLT model training process.
[0018] Figure 3 This is a schematic diagram of the training rate with and without meta-learning.
[0019] Figure 4 This is a comparison chart based on different deep learning dynamics modeling methods. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Please see Figure 1-4This invention provides a technical solution: a method for dynamic modeling of composite material structures of electric aircraft based on attention meta-learning, comprising the following steps:
[0022] Step 1: The dataset used comes from vibration experiments conducted on composite laminated wing box structures, composite laminated air beam structures, and composite conical-cylindrical composite shells under different temperature conditions. Given the limited data, various composite structures and ambient temperatures were divided into independent task samples. Each task includes vibration test data for the composite laminated wing box structure, composite laminated air beam structure, and composite conical-cylindrical composite shell at their respective temperatures; the data was then divided into a meta-training set and a meta-test set.
[0023] Step 2: Dataset partitioning. Specifically, vibration test data of CLWBs structures at 20℃, 50℃, 100℃ and 125℃ are designated as test tasks; vibration test data of CLWBs structures under the remaining temperature conditions and CLABs and CCCCs under all conditions are designated as training tasks; each task is divided into support set and query set in a 4:1 ratio.
[0024] Step 3: Import the support set data and labels from the meta-training set into the model for training. Train the AMLT model until convergence, using the Adam optimization algorithm and the MSE loss function. For each task, quickly update the model parameters using their respective support set data to achieve optimal parameters for the current task environment; the update formula is: Secondly, the AMLT model uses a meta-learner to identify shared features in multi-temperature and multi-structure tasks, and dynamically adjusts the attention weights of new tasks based on the similarity between the new task and the existing dataset; specifically, tasks that show high similarity to the target task are given higher weights, while those with low similarity are given lower weights.
[0025] The query set data and labels from the meta-training set are imported into the model for testing. The meta-learner uses query sets from these different tasks to evaluate the performance of the base learner, and updates the initial model parameters again by calculating the error and performing backpropagation; the update formula is: Repeat this process until training for all batch tasks is complete, and until convergence is achieved on all structure and temperature tasks.
[0026] Step 4: In the meta-testing phase, the AMLT model is trained using support set data partitioned from the test set, and the model parameters are updated to enhance its adaptation to new tasks. Subsequently, the query set partitioned from the test set is input into the model to evaluate the model's performance on new tasks, resulting in the dynamic model of CLWBs.
[0027] Step 5: Construct the following networks respectively: Model Agnostic Meta-Learning Transformer Network, Transformer Network, Deep Convolutional Neural Network, Long Short-Term Memory Neural Network, Convolutional Neural Network-Long Short-Term Memory Neural Network, and Bidirectional Long Short-Term Memory Network. Input the same training set into the model for training, and input the test set and labels into the trained model to obtain the amplitude-frequency response curves of CLWBs under these models respectively.
[0028] Step 6: Compare the dynamic modeling results under the seven models. The root mean square error (RMSE) is used to evaluate the prediction results. The formulas for calculating MSE and RMSE are as follows: In the formula, For predicted values, This is the actual value. n To test the sample length, the accuracy of the model is evaluated by comparing MSE and RMSE.
[0029] Furthermore, step 1 normalizes the original vibration data. This method transforms the data into the range [0, 1]. The normalization operation for the original vibration data is expressed as follows: in x Indicates the initial data. x max and x min express x The maximum and minimum values, y This represents a normalized vector.
[0030] Furthermore, the similarity in step 3 is determined by the Pearson similarity coefficient between different structural and temperature characteristics. The formula for calculating the similarity coefficient is as follows: in r Represents the similarity coefficient under one-dimensional features; n express n Dimensional features.
[0031] Those skilled in the art will be able to operate the technology in this case sequentially, and the specific operating sequence should refer to the following working principle. The detailed connection means are well-known in the art. The following mainly introduces the working principle and process. Example
[0032] Step 1: The dataset used comes from vibration experiments conducted on composite laminated wing box structures, composite laminated air beam structures, and composite conical-cylindrical composite shells under different temperature conditions. Given the limited data, various composite structures and ambient temperatures were divided into independent task samples. Each task includes vibration test data for the composite laminated wing box structure, composite laminated air beam structure, and composite conical-cylindrical composite shell at their respective temperatures. The data was divided into a meta-training set and a meta-test set. The original vibration data was normalized, transforming the data to the range [0, 1]. The normalization operation for the original vibration data is expressed as: in x Indicates the initial data. x max and x min express x The maximum and minimum values, y This represents a normalized vector.
[0033] Step 2: Dataset Partitioning. Specifically, vibration test data of CLWBs structures at 20℃, 50℃, 100℃, and 125℃ are designated as test tasks. Vibration test data of CLWBs structures under residual temperature conditions, as well as vibration test data of CLABs and CCCCs under all conditions, are designated as training tasks. Each task is divided into a support set and a query set in a 4:1 ratio.
[0034] Step 3: Import the support set data and labels from the meta-training set into the model for training. Train the AMLT model until convergence, using the Adam optimization algorithm and the MSE loss function. For each task, quickly update the model parameters using their respective support set data to achieve optimal parameters for the current task environment. The update formula is: Secondly, the AMLT model employs a meta-learner to identify shared features in multi-temperature and multi-structure tasks, dynamically adjusting the attention weights of new tasks based on their similarity to existing datasets. Specifically, tasks exhibiting high similarity to the target task are assigned higher weights, while those with low similarity are assigned lower weights. Similarity is determined by the Pearson similarity coefficient between different structural and temperature features, calculated using the following formula: in r Represents the similarity coefficient under one-dimensional features; n express n Dimensional features.
[0035] The query set data and labels from the meta-training set are imported into the model for testing. The meta-learner uses query sets from these different tasks to evaluate the performance of the base learner, and updates the initial model parameters again by calculating the error and performing backpropagation. The update formula is: Repeat this process until training for all batch tasks is complete, and until convergence is achieved across all structure and temperature tasks.
[0036] Then, in the meta-testing phase, the AMLT model is trained using support set data partitioned from the test set, and the model parameters are updated to enhance its adaptation to new tasks. Subsequently, the query set partitioned from the test set is input into the model to evaluate the model's performance on new tasks, resulting in the dynamic model of CLWBs.
[0037] Then, we constructed a model-agnostic meta-learning Transformer network, Transformer network, deep convolutional neural network, long short-term memory neural network, convolutional neural network-long short-term memory neural network, and bidirectional long short-term memory network, respectively. We input the same training set into the model for training, and input the test set and labels into the trained model to obtain the amplitude-frequency response curves of CLWBs under these models.
[0038] Finally, the dynamic modeling results under the seven models were compared. The root mean square error (RMSE) was used to evaluate the prediction results. The formulas for calculating MSE and RMSE are as follows: In the formula, For predicted values, This is the actual value. n To test the sample length, the model's accuracy is evaluated by comparing MSE and RMSE.
[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for dynamic modeling of composite material structures in small-sample electric aircraft based on attention meta-learning, characterized in that, Includes the following steps: Step 1: The dataset used comes from vibration experiments conducted on composite laminated wing box structures, composite laminated air beam structures, and composite conical-cylindrical composite shells under different temperature conditions. Due to limited data, various composite structures and ambient temperatures are divided into independent task samples. Each task includes vibration test data of composite laminated wing box structures, composite laminated air beam structures, and composite conical-cylindrical composite shells at their respective temperatures. The data is divided into a meta-training set and a meta-test set. Step 2: Dataset partitioning. Specifically, vibration test data of CLWBs structures at 20℃, 50℃, 100℃ and 125℃ are designated as test tasks; vibration test data of CLWBs structures under the remaining temperature conditions and CLABs and CCCCs under all conditions are designated as training tasks; each task is divided into support set and query set in a 4:1 ratio. Step 3: Import the support set data and labels from the meta-training set into the model for training. Train the AMLT model until convergence, using the Adam optimization algorithm and the MSE loss function. For each task, quickly update the model parameters using their respective support set data to achieve optimal parameters for the current task environment; the update formula is: Secondly, the AMLT model uses a meta-learner to identify shared features in multi-temperature and multi-structure tasks, and dynamically adjusts the attention weights of the new task based on the similarity between the new task and the existing dataset. Specifically, tasks that show high similarity to the target task are given higher weights, while those with low similarity are given lower weights. The query set data and labels in the meta-training set are imported into the model for testing. The meta-learner uses query sets from these different tasks to evaluate the performance of the base learner and updates the initial model parameters again by calculating the error and performing backpropagation. Its update formula is: Repeat this process until training for all batch tasks is complete, and until convergence is achieved on all structure and temperature tasks. Step 4: In the meta-testing phase, the AMLT model is trained using support set data partitioned from the test set, and the model parameters are updated to enhance its adaptation to new tasks. Subsequently, the query set partitioned from the test set is input into the model to evaluate the model's performance on the new task, thus obtaining the dynamic model of CLWBs. Step 5: Construct the following networks respectively: Model Agnostic Meta-Learning Transformer Network, Transformer Network, Deep Convolutional Neural Network, Long Short-Term Memory Neural Network, Convolutional Neural Network-Long Short-Term Memory Neural Network, and Bidirectional Long Short-Term Memory Network. Input the same training set into the model for training, and input the test set and labels into the trained model to obtain the amplitude-frequency response curves of CLWBs under these models respectively. Step 6: Compare the dynamic modeling results under the seven models. The root mean square error (RMSE) is used to evaluate the prediction results. The formulas for calculating MSE and RMSE are as follows: In the formula, For predicted values, This is the actual value. n To test the sample length, the accuracy of the model is evaluated by comparing MSE and RMSE.
2. The method for modeling the dynamics of composite material structures in small-sample electric aircraft based on attention meta-learning according to claim 1, characterized in that, Step 1 normalizes the original vibration data. This method transforms the data into the range [0, 1]. The normalization operation for the original vibration data is expressed as follows: Where x represents the initial data, xmax and xmin represent the maximum and minimum values of x, and k represents the normalized vector.
3. The method for modeling the dynamics of composite material structures in small-sample electric aircraft based on attention meta-learning according to claim 1, characterized in that, The similarity in step 3 is determined by the Pearson similarity coefficient between different structural and temperature characteristics. The formula for calculating the similarity coefficient is as follows: Where r represents the similarity coefficient under one-dimensional features; m represents m-dimensional features.