A ship resistance prediction method based on image learning

Abstract

The present invention discloses a ship resistance prediction method based on image learning, belonging to the field of ship equipment technology. The method comprises: taking a ship type as a sample, wherein a sample includes a ship type image and a label; using the ship type image as input to express the ship type characteristics, wherein the ship type image content is a side view of the ship type, and the color or grayscale of the pixel points represents the width information of the corresponding hull surface value point; the label in the sample is the resistance data corresponding to the corresponding ship type; using a deep learning framework to build a convolutional neural network model; inputting the ship type image and corresponding label of the ship type sample into the convolutional neural network model for training, and obtaining the parameters of the convolutional neural network model. When predicting the resistance of a new ship, the ship type image of the new ship is input into the convolutional neural network model to obtain the resistance prediction result. The present invention uses images as input to express ship type characteristics, effectively overcoming the shortcomings of previous approximate model methods, such as limited versatility and scalability.

Description

Technical Field

[0001] The present invention belongs to the technical field of ship equipment, and more specifically, relates to a ship resistance prediction method based on image learning. Background Art

[0002] The overall design of modern ships places higher demands on the efficiency of ship navigation performance prediction. During the ship demonstration and design process, multiple rounds of ship design and optimization are required according to the requirements of the overall plan. High-precision simulation calculations are time-consuming and difficult to meet the needs of rapid optimization in practical applications. Constructing an approximate model is a way to improve optimization efficiency. The approximate model constructs an approximate function based on the input and output of sample data to achieve approximation of the complex design space. At present, the commonly used approximate technologies mainly include the response surface method, variable precision model, Kriging model, radial basis function model, etc., and have been widely studied in the field of ship optimization.

[0003] Establishing an approximate model requires a sample space. This sample space must be generated using a specific sample point generation method based on a given range of design variables. Design variables often vary for different optimization cases. Consequently, the approximate model is deeply tied to the design variables, and an approximate model constructed for one case cannot be accumulated and applied to subsequent optimization problems. Applying existing sample data to a new optimization problem requires redefining and extracting parameters, which is a significant workload. Therefore, a non-parametric database construction method—one that is independent of design variables—is required to achieve universality and scalability.

[0004] The non-parametric database construction method directly uses the geometric information of the hull surface, namely the form value points on the hull surface. A hull surface requires thousands of form value points to fully define it. Conventional approximate models are unable to handle such high-dimensional data, requiring the introduction of convolutional neural network technology from deep learning.

[0005] Regarding the input data structure, previous methods used an expanded structure, arranging all the model points sequentially to form an N×3 matrix format, where N is the number of model points. This expanded structure disrupts the proximity of the model points, hindering deep learning algorithms. Furthermore, different ship types have different numbers of model points, so the method only works for similar ship types with the same model point topology, limiting its versatility and scalability. Summary of the Invention

[0006] In response to the above-mentioned defects or improvement needs of the existing technology, the present invention proposes a ship resistance prediction method based on image learning, which uses images as input to express ship type characteristics, effectively overcoming the shortcomings of previous approximate model methods such as poor versatility and scalability.

[0007] To achieve the above objectives, the present invention provides a ship resistance prediction method based on image learning, comprising:

[0008] Take a ship type as a sample, where a sample includes a ship type image and a label;

[0009] The ship image is used as input to express the ship features. The ship image content is the side view of the ship, and the color or grayscale of the pixel represents the width information of the corresponding hull surface value point;

[0010] The labels in the sample are the resistance data corresponding to the corresponding ship type;

[0011] A deep learning framework is used to build a convolutional neural network model, which includes an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer;

[0012] The ship type images and corresponding labels of the ship type samples are input into the convolutional neural network model for training to obtain the parameters of the convolutional neural network model. When predicting the resistance of a new ship, the ship type images of the new ship are input into the convolutional neural network model to obtain the resistance prediction result.

[0013] In some optional embodiments, the ship-shaped image is based on a three-dimensional ship-shaped surface and is generated by general three-dimensional modeling software, corresponding numerical simulation software or a self-written program.

[0014] In some optional implementation schemes, the generation of the ship image should be carried out under a unified standard and should meet the following three conditions: the ship image size, display range and viewport direction are defined to ensure that the ship image size, display range and viewport obtained from different ship files are consistent; the RGB color range of the ship image display is determined to ensure that the coordinate information corresponding to the color of the ship images obtained from different ship files is consistent; the resolution and format of the output ship image are determined to ensure that the format of the ship images obtained from different ship IGES is consistent.

[0015] In some optional implementation schemes, the resistance data is selected as the whole ship resistance or related resistance coefficient as needed, and the resistance coefficient is obtained by corresponding numerical calculation methods, such as potential flow method, CFD method, etc.

[0016] In some optional embodiments, for the input layer, the boat-shaped image is taken as input, and the boat-shaped image is first converted into a three-dimensional matrix of RGB pixel brightness, each pixel brightness value on each color channel is a feature, and each feature is normalized and standardized.

[0017] In some optional implementation schemes, for the convolution layer, the feature map output by the input layer is first subjected to a convolution operation to obtain a new feature map, and then the convolution layer feature map is output through an activation function, wherein the convolution operation of the convolution layer is implemented by a convolution kernel, and the parameters of the convolution kernel include size and number of channels, and the parameters of the convolution operation include step size and padding.

[0018] In some optional implementation schemes, for the pooling layer, on the one hand, the size of the feature map obtained by the convolution layer is reduced to simplify the computational complexity, and on the other hand, the feature map obtained by the convolution layer is compressed and the main features are extracted.

[0019] In some optional implementation schemes, for the fully connected layer, the fully connected layer is composed of multiple neurons that are fully connected, and the features extracted layer by layer by CNN are output to the fully connected layer, so that the entire CNN is globally trained using a gradient descent algorithm.

[0020] In some optional implementation schemes, the output layer is used for predicting the output of samples, and can be used for both regression analysis and classification tasks depending on the output model.

[0021] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0022] This paper proposes a ship resistance prediction method based on image learning. This method, independent of the design methods and parameters of the ship's lines, can predict the ship's resistance coefficient with high accuracy based solely on an image of the ship's curved surface. The method is highly scalable and can be used to predict the hydrodynamic forces of various types of ships by constructing a large database of imaged ship models. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 1 is a flow chart of a method for predicting ship resistance based on image learning provided by an embodiment of the present invention;

[0024] Figure 2 This is an example of a ship surface visualization of different ship schemes provided by an embodiment of the present invention. The black area in the figure is the null value area, and the darker the color, the larger the Y coordinate value.

[0025] Figure 3 This is a schematic diagram of the structure of a convolutional neural network provided by an embodiment of the present invention. The convolutional neural network comprises an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer.

[0026] Figure 4The embodiment of the present invention provides a method for obtaining a series of ship-shaped surfaces by using a radial basis function method for ship-shaped transformation, wherein one control point is selected at the bow and stern to control the change of the ship-shaped surface.

[0027] Figure 5 This is a schematic diagram of a convolutional neural network structure and parameters provided by an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of how MSE changes with the number of training times, provided by an embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram of a comparison between a predicted value and an actual value provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0031] Example 1

[0032] like Figure 1 As shown, a flow chart of a ship resistance prediction method based on image learning provided by an embodiment of the present invention includes:

[0033] S1: Take a ship type as a sample, where a sample includes a ship type image and a label;

[0034] S2: Use a ship image as input to express ship features. The ship image is a side view of the ship, and the color or grayscale of the pixel represents the width information of the hull surface value point at that location.

[0035] In the embodiment of the present invention, the ship image can be generated based on the three-dimensional ship surface through general three-dimensional modeling software, corresponding numerical simulation software, or a self-written program. The generation of the ship image should be performed under a unified standard and should ensure that:

[0036] a. Define the ship image size, display range and viewport direction to ensure that the ship image size, display range and viewport obtained from different ship files are consistent;

[0037] b. Determine the RGB color range of the ship image display to ensure that the coordinate information corresponding to the color of the ship image obtained from different ship files is consistent;

[0038] c. Determine the resolution and format of the output ship image to ensure that the formats of the ship images obtained by IGES of different ship types are consistent;

[0039] In addition, considering that the curvature of the ship surface along the length direction is small, scaling can be performed in the length direction to compress the size of the ship image.

[0040] Figure 2 Examples of ship surface images for different ship type schemes are given.

[0041] S3: The label in the sample is the resistance data corresponding to the ship type;

[0042] Among them, the resistance data can be selected as the whole ship resistance or related resistance coefficient as needed. The resistance coefficient can be obtained by corresponding numerical calculation methods, such as potential flow method, CFD method, etc.

[0043] S4: Establish a convolutional neural network. Use deep learning frameworks such as Pytorch to build a convolutional neural network model. The convolutional neural network model consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Intermediate layers such as the convolutional layer and the pooling layer are also called hidden layers.

[0044] a. Input layer

[0045] Taking a color boat image as an example, a color boat image is first converted into a three-dimensional matrix of RGB pixel brightness. Each pixel brightness value in each color channel is a feature. To make the pixel brightness values more evenly distributed and reduce the difficulty of model learning, the data needs to be normalized and standardized.

[0046] b. Convolutional layer

[0047] The convolutional layer is a core component of a CNN. The feature map output by the previous layer is first convolved to create a new feature map, which is then output through an activation function to create the convolutional layer feature map. The convolution operation in the convolutional layer is implemented using a convolution kernel, which acts as a filter. Different convolution kernels correspond to different feature extraction capabilities. The main parameters of the convolution kernel include size and number of channels, while the main parameters of the convolution operation include stride and padding.

[0048] c. Pooling layer

[0049] The pooling layer has two main functions: first, it reduces the size of the feature map, thereby simplifying computational complexity; second, it compresses features and extracts key features. Pooling functions typically replace the neighborhood of a certain location in the spatial representation of the convolutional layer's output features with the overall statistics of that neighborhood. For example, a max pooling operation extracts the maximum value within a square region as the feature representation of that square region.

[0050] d. Fully connected layer

[0051] The fully connected layer is composed of multiple neurons that are fully connected. The features extracted layer by layer by CNN are output to the fully connected layer, which allows the entire CNN to be trained globally using algorithms such as gradient descent.

[0052] e. Output layer

[0053] The output layer is used for sample prediction output. Depending on the output model, it can be used for regression analysis or classification tasks.

[0054] The structure of convolutional neural network is as follows Figure 3 shown.

[0055] S5: Input the ship images and labels of a large number of ship samples into the convolutional neural network model for training to obtain the parameters of the convolutional neural network model;

[0056] S6: When predicting the resistance of a new ship, the ship type image of the new ship is input into the convolutional neural network model to obtain the resistance prediction result.

[0057] Example 2

[0058] The method of the present invention is described in detail below by taking an example.

[0059] (1) Prepare samples;

[0060] In the embodiment of the present invention, a sample is generated using a ship form transformation method based on a master model. The sample is generated using the standard model Series 60 ship model as the master model. The main scale elements of the Series 60 ship model are shown in Table 1 below.

[0061] Table 1 Main dimension elements of Series 60 ship model

[0062]

[0063] The radial basis function method is used to transform the ship shape to obtain a series of ship shape surfaces. A total of three variables are selected to control the change of the ship shape. First, a control point is selected at the bow and stern, and the Y coordinates of these two points are selected as variables, such as Figure 4 As shown in Figure 2, a scaling factor α is given to the Y coordinates of all spatial point clouds on the hull to control the change in the Y coordinates of all points on the hull surface. The details of the three variables are shown in Table 2 below.

[0064] Table 2 Ranges of control variables

[0065]

[0066] A uniform experimental design method was used to sample the aforementioned variables, resulting in a total of 9,983 samples. The resulting ship surface was then converted into a bitmap. The label associated with each image is the wave-making resistance coefficient of the corresponding ship, calculated using the three-dimensional potential flow method.

[0067] (2) Establish a convolutional neural network;

[0068] In the embodiment of the present invention, the established convolutional neural network includes one input layer, three convolutional layers, three pooling layers, one fully connected layer and one output layer. The specific parameters of each layer are as follows: Figure 5 shown.

[0069] a. Input layer

[0070] The input layer is an S60-based ship image. A higher resolution results in a more accurate prediction model, but this also increases both modeling and prediction time. Therefore, a moderate resolution is recommended. The input image is a 150×150 pixel color image, or a 150×150 two-dimensional matrix, with values between 0 and 255.

[0071] b. First convolutional layer

[0072] Based on the resolution of the input image, a convolution kernel size of 3×3 is selected, the number of output channels of the convolution kernel is 64, the stride is 2, and no padding is used. After the convolution response and the nonlinear activation function ReLu operation, 64 feature maps with a resolution of 74×74 are obtained.

[0073] c. First max pooling layer

[0074] A larger pooling region and a longer stride mean a greater reduction in feature map resolution and faster computation, but this also results in greater information loss and the risk of underfitting. Therefore, appropriate pooling parameters are crucial. A pooling region size of 2×2 and a stride of 2 ensure that adjacent pooling regions do not overlap. After pooling, 64 feature maps with a resolution of 37×37 are obtained.

[0075] d. Second convolutional layer

[0076] The convolution kernel size is selected as 3×3, the number of output channels of the convolution kernel is 128, the stride is 2, and no padding is used. After the convolution response and the nonlinear activation function ReLu operation, 128 feature maps with a resolution of 18×18 are obtained.

[0077] e. Second maximum pooling layer

[0078] The pooling area size is 2×2, and the stride is 2, that is, adjacent pooling areas do not overlap. After pooling, 128 feature maps with a resolution of 9×9 are obtained.

[0079] f. The third convolutional layer

[0080] The convolution kernel size is selected as 3×3, the number of output channels of the convolution kernel is 256, the stride is 2, and no padding is used. After the convolution response and the nonlinear activation function ReLu operation, 256 feature maps with a resolution of 4×4 are obtained.

[0081] g. The third maximum pooling layer

[0082] The pooling area size is 2×2, and the stride is 2, that is, adjacent pooling areas do not overlap. After pooling, 256 feature maps with a resolution of 2×2 are obtained.

[0083] h. Fully connected layer

[0084] The number of neurons in the fully connected layer depends on the resolution and number of feature maps in the previous layer. A fully connected layer is set with 1024 neurons.

[0085] i. Output layer

[0086] The output channel for regression problems is 1.

[0087] (3) Model training;

[0088] The order of the 9,983 samples was shuffled, and 60% was randomly selected as the training set. The normalized wave-making drag coefficient was used as the label value, and the CNN model was trained 200 times. A random 20% of the samples outside the training set were used as the validation set to verify convergence. The Adam optimizer was selected, and the batch size of the image input was set to 512, meaning that 512 ship image samples were input per batch for training. The learning rate was set to 1e-3, and the convolution kernel parameters were randomly initialized.

[0089] The curve of mean square error MSE changing with the number of training times during training is as follows Figure 6 As shown in the figure, the training MSE decreases very rapidly and has a trend of continuing to decrease, indicating that the degree of model fitting has reached a very high level.

[0090] (4) Model testing.

[0091] The remaining 20% of samples are used as a test set and input into the trained CNN model to obtain the wave resistance coefficient. The prediction time is much less than 1S. The comparison between the predicted value and the true value is shown in the figure below. Figure 7 As shown in Figure 2, the average error of the entire test set is about 3%. It can be seen that the prediction accuracy of the model is high.

[0092] Through the above-mentioned technical solution, the present invention avoids the definition and selection of feature parameters through image-based data input. This is independent of the type value point selection rules and topological structure, and can easily achieve database accumulation and expansion. At the same time, convolutional neural networks can achieve regression prediction for high-dimensional problems with fewer parameters, thus achieving a transition from limited feature parameter input to complete geometric data.

[0093] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0094] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A ship resistance prediction method based on image learning, characterized in that: include: Take a ship type as a sample, where a sample includes a ship type image and a label; The ship image is used as input to express the ship features. The ship image content is the side view of the ship, and the color or grayscale of the pixel represents the width information of the corresponding hull surface value point; The labels in the sample are the resistance data corresponding to the corresponding ship type; A deep learning framework is used to build a convolutional neural network model, which includes an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer; The ship type images and corresponding labels of the ship type samples are input into the convolutional neural network model for training to obtain the parameters of the convolutional neural network model. When predicting the resistance of a new ship, the ship type images of the new ship are input into the convolutional neural network model to obtain the resistance prediction result.

2. The method according to claim 1, characterized in that The ship image is based on the three-dimensional ship surface and is generated by general three-dimensional modeling software, corresponding numerical simulation software or a self-written program.

3. The method according to claim 2, characterized in that The generation of the ship image should be carried out under a unified standard and should meet the following three conditions: the ship image size, display range and viewport direction should be defined to ensure that the ship image size, display range and viewport obtained from different ship files are consistent; the RGB color range of the ship image display should be determined to ensure that the coordinate information corresponding to the color of the ship image obtained from different ship files is consistent; the resolution and format of the output ship image should be determined to ensure that the format of the ship image obtained from different ship IGES files is consistent.

4. The method according to claim 3, characterized in that The resistance data may be selected from the whole ship resistance or related resistance coefficient according to needs, and the resistance coefficient is obtained by a corresponding numerical calculation method.

5. The method according to any one of claims 1 to 4, characterized in that For the input layer, the boat image is taken as input and first converted into a three-dimensional matrix of RGB pixel brightness. Each pixel brightness value on each color channel is a feature, and each feature is normalized and standardized.

6. The method according to claim 5, characterized in that For the convolution layer, the feature map output by the input layer is first subjected to a convolution operation to obtain a new feature map, and then the convolution layer feature map is output through the activation function, wherein the convolution operation of the convolution layer is implemented by the convolution kernel, and the parameters of the convolution kernel include size and number of channels, and the parameters of the convolution operation include step size and padding.

7. The method according to claim 6, characterized in that For the pooling layer, on the one hand, the size of the feature map obtained by the convolution layer is reduced to simplify the calculation complexity, and on the other hand, the feature map obtained by the convolution layer is compressed and the main features are extracted.

8. The method according to claim 7, characterized in that For the fully connected layer, the fully connected layer is composed of multiple neurons that are fully connected. The features extracted layer by layer by CNN are output to the fully connected layer, so that the entire CNN is globally trained using a gradient descent algorithm.

9. The method according to claim 8, characterized in that The output layer is used for sample prediction output. Depending on the output model, it can be used for both regression analysis and classification tasks.

Images