Submerged unmanned ship resistance forecasting method based on combined neural network

Abstract

The invention provides a submerged unmanned ship resistance forecasting method based on a combined neural network. The submerged unmanned ship resistance forecasting method comprises the following steps: acquiring a multi-source data set of ship model experiment and simulation calculation; performing data enhancement on the multi-source data set by using a WGAN-GP model to obtain an enhanced data set; respectively inputting the enhanced data set into two deep neural networks for training to obtain resistance predicted values of the two deep neural networks; and constructing a fusion network, and carrying out dynamic weighting on the resistance prediction values of the two deep neural networks to obtain a final resistance prediction value after fusion. According to the method, the multi-source data is enhanced through the WGAN-GP model, the two deep neural networks are used for respectively learning the resistance mapping relation in the experimental data and the simulation data, and the prediction values of the two neural networks are dynamically weighted through the fusion network, so that high-precision resistance prediction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-mode vehicle resistance prediction, in particular, especially relates to a submarine floating unmanned ship resistance prediction method based on a combined neural network. BACKGROUND

[0002] The submarine floating unmanned ship is a new type of multi-mode unmanned high-performance ship with water surface navigation, semi-submersible concealment and submarine navigation functions, and has the functions of "water surface-underwater" cross-domain navigation. It can navigate at high speed on the water surface, and can avoid waves, underwater obstacles or implement concealment by cross-medium diving. It has outstanding advantages in applications such as three-dimensional detection, underwater concealment, communication relay and obstacle avoidance. The proxy model technology is widely used in ship design, which can replace the traditional experiment and calculation method to a certain extent to obtain the required data at low cost. However, in the field of new ships such as submarine floating unmanned ships, the limited number of real ships restricts the smooth implementation of building a proxy model through a large number of samples. At the same time, the multi-source heterogeneous and scarce sample data lead to uncertainty in the sample data and the proxy model. Under the traditional proxy model technology framework, the multi-source data needs to be corrected in accuracy and clustered in weight level, and the data set scale needs to be uniformly represented and the reliability needs to be evaluated.

[0003] The resistance prediction proxy model based on deep learning is an important branch of the improvement research of traditional proxy models. Through neural network technology, the nonlinear relationship between various physical quantities of the vehicle in a complex flow field environment is fitted, and the prediction accuracy and generalization performance of the model are improved. In the research of using neural networks to build ship resistance prediction proxy models, the scarcity of high-precision data and experimental samples is the main problem affecting the prediction accuracy and generalization ability of the model. The common method is to perform data enhancement, transfer learning or build variable reliability models on experimental samples.

[0004] Data enhancement is a technology that modifies existing data slightly or generates synthetic samples to expand the training data set and improve the generalization ability of the model. Common methods include geometric transformation (rotation, flipping, scaling, etc.), color space adjustment, noise injection, and GAN generated synthetic data. The overall process is: analyze the characteristics of the original data, select the enhancement strategy, generate new samples, and merge into an expanded training set. As a representative of generative enhancement, GAN can learn the data distribution and generate high-quality synthetic samples, breaking through the limitations of traditional methods, and is especially suitable for data-scarce or high-dimensional scenarios The existing technology still has significant defects in data acquisition, model construction and multi-source data fusion, and it is difficult to achieve high-precision and high-efficiency resistance prediction, which restricts the intelligent development of submarine floating unmanned ships. Therefore, there is an urgent need for a technical solution that can effectively integrate experimental and simulation data and improve the prediction performance through a deep learning model. SUMMARY

[0005] To address the aforementioned technical problems, this invention provides a method for predicting the resistance of submersible unmanned surface vessels (USVs) based on combined neural networks. This invention enhances multi-source data using the WGAN-GP model, and utilizes two deep neural networks to learn the resistance mapping relationship between experimental and simulated data, respectively. Then, a fusion network dynamically weights the predicted values ​​from the two neural networks, thereby achieving high-precision resistance prediction.

[0006] The technical means employed in this invention are as follows: A method for predicting the drag of a submerged unmanned surface vessel based on a combined neural network includes: Obtain multi-source datasets from ship model experiments and simulation calculations; The WGAN-GP model was used to augment the multi-source dataset to obtain an augmented dataset. The augmented dataset is input into two deep neural networks (DNNI and DNNII) for training to obtain the resistance prediction values ​​of the two deep neural networks. A fusion network is constructed, and the resistance prediction values ​​of the two deep neural networks are dynamically weighted to obtain the final resistance prediction value after fusion.

[0007] Furthermore, the step of obtaining the multi-source dataset for ship model experiments and simulation calculations includes: Design a submersible unmanned vessel and build a physical model of the vessel. A ship model experiment was conducted in the towing tank to test the hydrostatic resistance at multiple speed points under different diving depth conditions. Numerical simulations were performed using STAR-CCM software. The Reynolds-averaged (RANS) method was used, and the SST k-ω turbulence model was employed. The mesh was refined at the bow and stern and in the free surface region of the ship to obtain simulation data.

[0008] Furthermore, the step of using the WGAN-GP model to augment the multi-source dataset to obtain the augmented dataset includes: The objective function of the WGAN-GP model is defined as follows:

[0009] in, For discriminator, For generator, For the discriminator to input The output value, Generate sample distribution for the generator Expectations The weights of the gradient penalty term are... For sampling distribution Expectations the output value of the discriminator for the input , the L2 norm of the gradient of the discriminator at point for the input , the real data distribution, the generated distribution; The percentage quantile coverage ratio, feature entropy, and nearest neighbor distance ratio are selected as the standards for measuring the quality of the generated data of the WGAN-GP. The enhanced data with a difference between the percentage quantile coverage ratio value and 1 not exceeding 0.03, a feature entropy not less than 3, and a nearest neighbor distance ratio value not less than 0.8 are selected as the effective augmented data set.

[0010] Further, the calculation formula of the percentage quantile coverage ratio value is as follows:

[0011] wherein, denotes the percentage quantile coverage ratio value, denotes the number of variables, denotes the quantile of the generated data on the i-th variable, denotes the quantile of the original data on the i-th variable, denotes the quantile of the generated data on the i-th variable, denotes the quantile of the original data on the i-th variable.

[0012] Further, the calculation formula of the feature entropy is as follows:

[0013] wherein, denotes the feature entropy value, denotes the number of discrete intervals divided, denotes the interval index, denotes the probability that the value of the i-th feature in the generated data falls in the j-th interval.

[0014] ​​​​​​​​​​Further, the calculation formula of the nearest neighbor distance ratio (Nearest Neighbor Distance Ratio) is as follows:

[0015] wherein, denotes the nearest neighbor distance ratio, denotes the number of data points in the generated data set, denotes the number of data points in the original data set, denotes the generated data set, denotes the original data set, denotes a sample point in the generated data set, denotes a sample point in the original data set; Further, the two deep neural networks are DNNI deep neural network and DNNII deep neural network, both of which adopt a hidden layer structure of 256, 128, and 64, and learn the mapping relationship of the unmanned ship depth, speed, and resistance in their respective data based on the expanded training sample set and the simulation data set.

[0016] Further, the DNNI deep neural network and the DNNII deep neural network both adopt MSE as the loss function, the optimizer is Adam (Adaptive Moment Estimation), and the ReLU activation function is adopted, and the expression is: .

[0017] Further, the fusion network learns the mapping relationship of the resistance prediction values of the two deep neural networks to the real resistance based on the real resistance data, and the specific steps for optimizing the hyperparameters of the hidden layer structure and the activation function are as follows: Perform grid search optimization for each hyperparameter combination; Train for 100 epochs using the Adam optimizer and the MSE loss function; Calculate the , MAE, and RMSE performance indicators on the test set; Record the hyperparameter combination and the corresponding performance indicators.

[0018] Compared with the prior art, the present application has the following advantages: 1. The present application uses deep neural networks to learn and model multi-source data, which can more accurately capture the complex nonlinear relationship between the resistance of the submerged floating unmanned ship and the depth, speed, and other parameters, thereby significantly improving the accuracy of resistance prediction and providing a more reliable basis for the design and optimization of the unmanned ship.

[0019] 2、The application can effectively solve the problem of multi-source data scarcity by using the WGAN-GP model to perform data enhancement on experimental data and simulation data, fully utilize limited experimental data and simulation data, expand the size of the training data set, and improve the generalization ability of the model, so that a high-quality prediction model can be constructed even in the case of small sample.

[0020] 3、The application can automatically adjust the weight according to the characteristics of different data sources and the reliability of the prediction value, realize the optimal fusion of multi-source data, effectively avoid the uncertainty caused by the difference in the reliability of multi-source data, and further improve the accuracy and stability of the prediction result.

[0021] 4、The application reduces the dependence on a large number of experimental samples, reduces the experimental cost and time investment, and at the same time, through the combination of efficient numerical simulation and deep learning, the data can be quickly acquired and processed, and the process of unmanned ship resistance prediction is accelerated, and the research and development efficiency is improved In summary, the application of the application is not only suitable for submarine floating unmanned ships, but also can be popularized to resistance prediction of other new multi-hull state vehicles, and provides technical support for intelligent development in related fields, and promotes the wide application and in-depth development of unmanned ship technology in stereoscopic detection, underwater concealment, communication relay, obstacle avoidance and other application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0023] Figure 1 The method flowchart of the application.

[0024] Figure 2 The scatter plot of the WGAN-GP after inspection, respectively expanding 10 times, 40 times and 160 times data of the experimental training sample set.

[0025] Figure 3 The fusion network hyperparameter optimization result comparison chart.

[0026] Figure 4 The combined neural network workflow chart.

[0027] Figure 5 The test set performance result chart of the combined neural network under different training data amounts.

[0028] Figure 6This is a weight distribution diagram of two deep neural networks (DNNI and DNNII) under different training data volumes. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0031] like Figure 1 As shown, this invention provides a method for predicting the resistance of submersible unmanned vessels based on combined neural networks, including: S1. Obtain multi-source datasets for ship model experiments and simulation calculations; S2. Use the WGAN-GP model to augment the multi-source dataset to obtain the augmented dataset; S3. Input the augmented dataset into two deep neural networks (DNNI, DNNII) for training to obtain the resistance prediction values ​​of the two deep neural networks; S4. Construct a fusion network and dynamically weight the resistance prediction values ​​of the two deep neural networks to obtain the final resistance prediction value after fusion.

[0032] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11. Design a submersible unmanned vessel and build a real vessel model. S12. Conduct ship model experiments in the towing tank to test the hydrostatic resistance at multiple speed points under different diving depth conditions. S13. Numerical simulation was performed using STAR-CCM software. Based on the Reynolds-averaged (RANS) method, the SST k-ω turbulence model was adopted, and the mesh was refined in the bow and stern and free surface regions of the ship. Simulation data were obtained by dividing the mesh computational domain under different working conditions and performing numerical calculations.

[0033] In this embodiment, the ship model experimental data and numerical simulation experiments use a submersible unmanned surface vessel (USV) as the target ship type, considering only the resistance characteristics of the main hull and ignoring the influence of appendages on resistance, to obtain experimental and simulation data. The specific layout of the ship model towing experiment is as follows: the dimensions of the pool are 160.0m × 7.0m × 3.7m (length × width × water depth). Using a submersible USV as the target ship type, ship model experiments and numerical simulation experiments are carried out, setting six different diving depth conditions from 0# to 5#, which correspond to diving depths of 0m, 0.054m, 0.32m, 0.48m, 0.64m, and 0.96m, respectively. The hydrostatic resistance was tested at multiple velocity points from 0.4 m / s to 1.7 m / s under various operating conditions, namely 0.4 m / s, 0.6 m / s, 0.8 m / s, 1.0 m / s, 1.1 m / s, 1.2 m / s, 1.3 m / s, 1.4 m / s, 1.5 m / s, 1.6 m / s and 1.7 m / s. The corresponding Froude numbers (Fr) ranged from 0.1010 to 0.4291, and the Reynolds numbers (Re) ranged from 0.5284 × 10⁶ to 2.2459 × 10⁶.

[0034] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. Define the objective function of the WGAN-GP model as follows:

[0035] in, For discriminator, For generator, For the discriminator to input The output value, Generate sample distribution for the generator Expectations The weights of the gradient penalty term are... For sampling distribution Expectations For the discriminator to input The output value, For the discriminator at point Input The L2 norm of the gradient, For the true data distribution, To generate a distribution; S22. The percentage quantile distance (Quantile Coverage Ratio), feature entropy, and nearest neighbor distance ratio are selected as standards for measuring the quality of data generated by WGAN-GP; where: The formula for calculating the percentage quantile distance value is as follows:

[0036] in, This represents the percentage quantile distance value. Indicates the number of variables. Indicates that the generated data is in the first... On each variable Quantiles Indicates the original data at the 1st... On each variable Quantiles Indicates that the generated data is in the first... On each variable Quantiles Indicates the original data at the 1st... On each variable Quantiles.

[0037] The formula for calculating the feature entropy is as follows:

[0038] in, Represents the feature entropy value. This indicates the number of discrete intervals that have been divided. Indicates a range index. This indicates that in the generated data, the first... The value of the first feature falls within the first... The probability of each interval; The formula for calculating the Nearest Neighbor Distance Ratio is as follows:

[0039] in, This represents the nearest neighbor distance ratio. This indicates the number of data points in the generated dataset. This indicates the number of data points in the original dataset. This indicates the generation of a dataset. Represents the original dataset. This represents the sample points in the generated dataset. This represents the sample points in the original dataset; S23. Select augmented data with a percentage quantile distance value that differs from 1 by no more than 0.03, a feature entropy value that is no less than 3, and a nearest neighbor distance ratio that is no less than 0.8 as effective augmented datasets. For example... Figure 2As shown, these are the results of WGAN-GP expanding the training sample set of experimental data by 10 times, 40 times, and 160 times respectively, while meeting the requirements.

[0040] In a specific implementation, as a preferred embodiment of the present invention, the two deep neural networks in step S3 are the DNNI deep neural network and the DNNII deep neural network, both of which adopt a hidden layer structure of 256, 128, and 64, respectively, and learn the mapping relationship from the unmanned vessel's diving depth, speed to resistance in their respective data based on the expanded training sample set and the simulation dataset.

[0041] In a specific implementation, as a preferred embodiment of the present invention, both the DNNI deep neural network and the DNNII deep neural network use MSE as the loss function, Adam (Adaptive Moment Estimation) as the optimizer, and ReLU activation function as the expression: .

[0042] In this embodiment, Adam is a gradient descent-based optimization algorithm that adaptively adjusts the learning rate by calculating the first moment estimate (mean) and second moment estimate (uncentered variance) of the gradient and combining them with an exponentially weighted average.

[0043] In a specific implementation, as a preferred embodiment of the present invention, the fusion network learns the mapping relationship between the resistance prediction values ​​of the two deep neural networks and the actual resistance based on real resistance data. The specific steps for optimizing the hyperparameters of its hidden layer structure and activation function are as follows: Perform grid search optimization for each hyperparameter combination; Train for 100 epochs using the Adam optimizer and the MSE loss function; Calculate on the test set MAE and RMSE performance indicators; Record the hyperparameter combinations and their corresponding performance indicators.

[0044] In this embodiment, the hyperparameter optimization results are shown below. Figure 3 Based on this, a fusion network is constructed. Combining the two DNNs from step 3, the prediction workflow of the entire ensemble neural network model is shown in the diagram. Figure 4 After the model was built, it was trained using experimental and simulation data augmented by 10, 40, and 160 times, respectively, to test the performance of the combined neural network on the test set. The performance and weight distribution of the dynamic weight fusion model under different data augmentation amounts are shown below. Figure 5 , Figure 6 As shown.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network, characterized in that, include: Obtain multi-source datasets from ship model experiments and simulation calculations; The WGAN-GP model was used to augment the multi-source dataset to obtain an augmented dataset. The augmented dataset is input into two deep neural networks for training to obtain the resistance prediction values ​​of the two deep neural networks. A fusion network is constructed, and the resistance prediction values ​​of the two deep neural networks are dynamically weighted to obtain the final resistance prediction value after fusion.

2. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 1, characterized in that, The steps for obtaining the multi-source dataset for ship model experiments and simulation calculations include: Design a submersible unmanned vessel and build a physical model of the vessel. A ship model experiment was conducted in the towing tank to test the hydrostatic resistance at multiple speed points under different diving depth conditions. Numerical simulations were performed using STAR-CCM software. Based on the Reynolds-averaged method, the SST k-ω turbulence model was adopted, and the mesh was refined in the bow, stern, and free surface regions of the ship to obtain simulation data.

3. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 1, characterized in that, The steps of using the WGAN-GP model to augment the multi-source dataset to obtain the augmented dataset include: The objective function of the WGAN-GP model is defined as follows: in, For discriminator, For generator, For the discriminator to input The output value, Generate sample distribution for the generator Expectations The weights of the gradient penalty term are... For sampling distribution Expectations For the discriminator to input The output value, For the discriminator at point Input The L2 norm of the gradient, For the true data distribution, To generate a distribution; Percentile distance, feature entropy value, and nearest neighbor distance ratio are selected as standards for measuring the quality of data generated by WGAN-GP; Augmented data with a percentile quantile distance value that is no more than 0.03 from 1, a feature entropy value that is no less than 3, and a nearest neighbor distance ratio that is no less than 0.8 are selected as effective augmented datasets.

4. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 3, characterized in that, The formula for calculating the percentage quantile distance value is as follows: in, This represents the percentage quantile distance value. Indicates the number of variables. Indicates that the generated data is in the first... On each variable Quantiles Indicates the original data at the 1st... On each variable Quantiles Indicates that the generated data is in the first... On each variable Quantiles Indicates the original data at the 1st... On each variable Quantiles.

5. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 3, characterized in that, The formula for calculating the feature entropy is as follows: in, Represents the feature entropy value. This indicates the number of discrete intervals that have been divided. Indicates a range index. This indicates that in the generated data, the first... The value of the first feature falls within the first... The probability of each interval.

6. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 3, characterized in that, The formula for calculating the Nearest Neighbor Distance Ratio is as follows: in, This represents the nearest neighbor distance ratio. This indicates the number of data points in the generated dataset. This indicates the number of data points in the original dataset. This indicates the generation of a dataset. Represents the original dataset. This represents the sample points in the generated dataset. This represents the sample points in the original dataset.

7. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 1, characterized in that, The two deep neural networks are the DNNI deep neural network and the DNNII deep neural network, both of which adopt a hidden layer structure of 256, 128, and 64. They learn the mapping relationship from the unmanned vessel's diving depth, speed and resistance in their respective data based on the expanded training sample set and the simulation dataset.

8. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 7, characterized in that, Both the DNNI deep neural network and the DNNII deep neural network use MSE as the loss function, Adam as the optimizer, and ReLU activation function, with the following expression: 。 9. The method for predicting the resistance of a submersible unmanned surface vessel based on a combined neural network according to claim 1, characterized in that, The fusion network learns the mapping relationship between the resistance predictions of two deep neural networks and the actual resistance based on real resistance data. The specific steps for optimizing the hyperparameters of its hidden layer structure and activation function are as follows: Perform grid search optimization for each hyperparameter combination; Train for 100 epochs using the Adam optimizer and the MSE loss function; Calculate on the test set MAE and RMSE performance indicators; Record the hyperparameter combinations and their corresponding performance metrics.

Images