Main body structure design model training method and main body structure generation method

Abstract

The invention relates to a training method for providing an underwater vehicle main body structure design model and an underwater vehicle main body structure generation method. The training method of the underwater vehicle main body structure design model comprises the following steps: firstly, generating a noise adding main body structure based on a randomly initialized ordinal number stamp k, a sample main body structure and randomly initialized Gaussian noise generation data; secondly, determining predicted Gaussian noise data based on the ordinal number stamp k, the performance parameters of the sample main body structure and the noise adding main body structure by adopting a main body structure design model; and finally, determining a training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and training the main structure design model based on the training loss function. Under the condition that only few sample main body structures exist, enough training sample data used for main body structure design model training can be generated, and model training under a small amount of sample main body structure data is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a subject structure design model training method and a subject structure generation method. BACKGROUND

[0002] Underwater vehicles such as submarines, underwater detectors, etc. need to dive to the target depth for navigation or operation, and the subject structure thereof needs to withstand the external water pressure at the target depth. The subject structure includes a pressure hull and a rib, wherein the rib enhances the shell stiffness of the pressure hull by supporting the pressure hull, so as to avoid deformation or collapse of the pressure hull under water pressure. In addition, in order to ensure the underwater maneuverability and endurance of the underwater vehicle, the weight of the subject structure needs to be strictly limited.

[0003] In order to ensure that the subject structure of the underwater vehicle meets the pressure index required by the target working condition and minimizes the weight, the conventional design method is to first perform structural design, and then perform numerical simulation based on the design mechanism to verify whether the structural design is reasonable. The previous design method has a long design cycle, low optimization efficiency, and may be limited by the designer's thinking and unable to achieve structural innovation design. With the development of artificial intelligence technology, automatic generation methods based on deep learning have been applied to the field of mechanical structure design. The automatic generation method based on deep learning needs to rely on a large amount of labeled historical data or labeled simulation data. The subject structure of the underwater vehicle is special in structure and has a small amount of product types, which causes the sample data for supervised learning to be scarce, making it difficult to realize automatic design by using conventional deep learning automatic methods, and more intuitively, it is difficult to realize the construction of the subject structure design model of the underwater vehicle. SUMMARY

[0004] The embodiments of the present application provide a subject structure design model training method of an underwater vehicle and a subject structure generation method of an underwater vehicle.

[0005] In a first aspect, the embodiments of the present application provide a subject structure design model training method of an underwater vehicle, which comprises: first generating a noisy subject structure based on a randomly initialized ordinal stamp k, a sample subject structure and randomly initialized generated Gaussian noise data; then determining predicted Gaussian noise data based on the ordinal stamp k, the performance parameters of the sample subject structure and the noisy subject structure by using a subject structure design model; and finally determining a training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and training the subject structure design model based on the training loss function.

[0006] For the same sample main body structure, the number of Gaussian noise data that can be randomly initialized is large, and the number of ordinal stamps k that can be randomly initialized is large, so that multiple sets of training data for implementing the main body structure design model training can be obtained in the case of one sample main body structure. Correspondingly, in the case of only a small number of sample main body structure, enough training sample data for main body structure design model training can be generated. That is, the training method of the main body structure design model provided in the embodiment of the application can realize model training with a small amount of sample main body structure data.

[0007] In some possible implementations, the foregoing training method further includes: determining a denoised main body structure by using the noisy main body structure and the predicted Gaussian noise data; performing stress analysis on each structural unit in the denoised main body structure to determine the predicted stress of each structural unit under the environmental condition corresponding to the performance parameter; simultaneously determining the real stress of each structural unit in the sample main body structure under the environmental condition corresponding to the performance parameter; and then calculating the stress prediction loss function based on the predicted stress and the real stress of each structural unit. Correspondingly, the steps of training the model are: first, determining the noise prediction loss function based on the generated Gaussian noise data and the predicted Gaussian noise data; and then determining the training loss function based on the stress prediction loss function and the noise prediction loss function.

[0008] In the model training process, determining the training loss function based on the stress prediction loss function can constrain the main body structure design model to meet reasonable performance parameters, so that the predicted main body structure output by the main body structure design model meets the corresponding stress constraint.

[0009] In some possible implementations, the stress analysis on each structural unit in the denoised main body structure to determine the predicted stress of each structural unit under the environmental condition corresponding to the performance parameter includes: determining the topological relationship of each structural unit, and constructing a topological graph based on the topological relationship; and analyzing the topological graph by using a pre-trained structural stress analysis model to determine the predicted stress of each structural unit; the structural stress analysis model is a graph neural network model.

[0010] In some possible implementations, the foregoing training method further includes: performing process compliance analysis on the denoised main body structure to determine a process constraint loss function; and correspondingly, before training the main body structure design model, first, determine the training loss function based on the stress prediction loss function, the noise prediction loss function, and the process constraint loss function.

[0011] In some possible implementations, before generating the noisy main body structure based on the randomly initialized ordinal stamp k, the sample main body structure and the randomly initialized generated Gaussian noise data, the method further includes: generating combined design parameters based on various design parameters, and generating a corresponding sample main body structure based on each set of generated design parameters; determining a performance parameter of each sample main body structure, the performance parameter including a maximum pressure value, a true stress of each structure unit at the maximum pressure value, and a weight of the sample main body structure.

[0012] By generating the sample main body structure through combination of various design parameters, a large number of samples can be provided for model training, and the quality of the model obtained through training can be improved accordingly.

[0013] In some possible implementations, the main body structure design model includes a conditional encoder and a denoising diffusion probability network.

[0014] The main body structure design model is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the performance parameter of the sample main body structure and the noisy main body structure.

[0015] The performance parameter of the sample main body structure is processed by using the conditional encoder to obtain a diffusion condition vector.

[0016] The denoising diffusion probability network is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the diffusion condition vector and the noisy main body structure.

[0017] The main body structure design model is trained based on a training loss function, including: the conditional encoder and the denoising diffusion probability network are jointly trained based on the loss function.

[0018] In a second aspect, an embodiment of the present application provides a main body structure generation method of an underwater vehicle, including:

[0019] Determining a required performance parameter that needs to be met by a main body structure of an underwater vehicle to be designed, and initializing an ordinal stamp k=K and initial Gaussian noise data.

[0020] Cyclically performing the following steps until k=-1:

[0021] The main body structure design model is used to determine the noise data to be denoised in the kth step based on the ordinal stamp k, the required performance parameter and the denoised main body structure in the kth step, and; when k=K, the denoised main body structure is the initial Gaussian noise data.

[0022] The noise data to be denoised in the kth step is subtracted from the denoised main body structure in the kth step to obtain the denoised main body structure in the k-1th step.

[0023] Updating k=k-1.

[0024] The de-noised main body structure determined in the last step is taken as the generated main body structure.

[0025] In some possible implementation manners, the generated main body structure is analyzed to determine a performance parameter of the generated main body structure.

[0026] In a third aspect, an embodiment of the present application provides a construction device of a main body structure design model of an underwater vehicle, including:

[0027] The structure noise adding unit is configured to generate a noise-added main body structure based on the randomly initialized ordinal stamp k, the sample main body structure, and the randomly initialized generated Gaussian noise data.

[0028] The noise prediction unit is configured to determine predicted Gaussian noise data based on the ordinal stamp k, the performance parameter of the sample main body structure, and the noise-added main body structure by using the main body structure design model.

[0029] The model training unit is configured to determine a training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and train the main body structure design model based on the training loss function.

[0030] In a fourth aspect, an embodiment of the present application provides a main body structure generation device of an underwater vehicle, including an initial parameter determination unit, a de-noising unit, and a main body structure determination unit.

[0031] The initial parameter determination unit is configured to determine a required performance parameter of a main body structure of an underwater vehicle to be designed, and initialize the ordinal stamp k=K and initial Gaussian noise data.

[0032] The de-noising unit is configured to, in a case that k>0, make k=k-1; determine noise data to be de-noised based on the ordinal stamp k, the required performance parameter, and a de-noised main body structure by using the main body structure design model; and subtract the noise data to be de-noised from the initial Gaussian noise data to obtain an updated de-noised main body structure until k=-1; wherein in a case that k=K, the de-noised main body structure is the initial Gaussian noise data.

[0033] The main body structure determination unit is configured to take the de-noised main body structure determined in the last step as a generated main body structure.

[0034] The beneficial effects of the above second aspect to fourth aspect can be referred to the related description in the first aspect, and are not described herein again for brevity. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a structural schematic diagram of the main body structure design model to be trained in an embodiment of the present application;

[0036] Figure 2 is a flowchart of a training method of the main body structure design model provided in an embodiment of the present application;

[0037] Figure 3 is a flow chart of a method for generating a main structure of an underwater vehicle provided by an embodiment of the present application;

[0038] Figure 4 shows all flow charts of the embodiments of the present application;

[0039] Figure 5 is a structural schematic diagram of a construction device of a design model of a main structure of an underwater vehicle provided by an embodiment of the present application;

[0040] Figure 6 is a flow chart of a method for generating a main structure of an underwater vehicle provided by an embodiment of the present application; DETAILED DESCRIPTION

[0041] The term "and / or" mentioned in the present application is a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that there are three cases of A alone, A and B together, and B alone. The symbol " / " in the present application means that the associated objects are or, for example, A / B means A or B.

[0042] The terms "first" and "second" and the like in the description and claims of the present application are used to distinguish different objects, and are not used to describe a specific order of the objects. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, which is only a distinguishing way adopted in the description of the same attribute objects in the embodiments of the present application. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, so that the process, method, system, product or equipment containing a series of units does not have to be limited to those units, but can include other units not clearly listed or inherent to these processes, methods, systems, products or equipment.

[0043] In the embodiments of the present application, the words such as "exemplary" or "for example" are used to mean as an example, illustration or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. On the contrary, the use of "exemplary" or "for example" and the like is intended to present the relevant concept in a specific way.

[0044] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more, for example, a plurality of processing units means two or more processing units, and the like; a plurality of elements means two or more elements, and the like.

[0045] In order to solve the problem that it is difficult to realize the construction of the main body structure design model of the underwater vehicle by using the conventional supervised learning method due to the lack of sample data, an embodiment of the present application provides a new training method of the main body structure design model of the underwater vehicle (hereinafter referred to as the main body structure design model training method) using the denoising diffusion principle. The main body structure design model training method generates a noisy main body structure by actively adding noise to the sample main body structure, and then realizes the main body structure design model training method based on the denoising of the noisy main body structure.

[0046] Before analyzing the main body structure design model training method provided by the embodiment of the present application, the main body structure design model provided by the embodiment of the present application is first introduced briefly. Figure 1 is the structural schematic diagram of the main body structure design model to be trained by the embodiment of the present application. As shown in Figure 1 , the main body structure design model 100 includes a conditional encoder 101 and a denoising diffusion probability network 102.

[0047] The conditional encoder 101 is an encoder for converting performance parameters and the like into a diffusion condition vector, specifically an encoder for converting target pressure value, main body structure weight upper limit, material type and other key process information into a diffusion condition vector. In a specific implementation, the conditional encoder can adopt a network with a cross-attention mechanism or a self-attention mechanism, so as to be able to fully mine the mutual influence relationship between different performance parameters, and then realize more diversified layout exploration.

[0048] The denoising diffusion probability network 102 is a backbone network for predicting and determining the predicted Gaussian noise data directly based on the diffusion condition vector and the noisy main body structure. Considering that the object to be processed in the embodiment of the present application is the main body structure, the denoising diffusion probability network can adopt a network such as 3D-UNet.

[0049] Corresponding to the adoption of the 3D-UNet network, the aforementioned noisy main body structure and Gaussian noise data should be 3D voxel grid data. The 3D-Unet is a network with an encoder-decoder structure. The encoder part gradually extracts high-level features of the noisy main body structure through a series of convolution and pooling operations, while reducing the data dimension. The decoder part gradually restores the high-level features to complete 3D voxel grid through deconvolution and upsampling operations, so as to accurately capture the spatial structure features of the rib and determine the predicted Gaussian noise data. It can be imagined that through the random noise based on the diffusion model and the 3D-UNet processing, combined with the cross-attention layer condition injection, diversified layout exploration is realized and local optimum is avoided.

[0050] The training method of the main body structure design model provided in the embodiment of the present application is analyzed as follows. The main body structure is trained as a black box model first, and after the core training method is introduced, the main body structure design model is treated as a white box model, and the training process of the internal module is analyzed. Figure 2 is a flowchart of the training method of the main body structure design model provided in the embodiment of the present application. As shown in Figure 2 , the training method of the main body structure design model provided in the embodiment of the present application includes S110-S130.

[0051] S110: generating a noisy main body structure based on a randomly initialized ordinal stamp k, a sample main body structure and randomly initialized generated Gaussian noise data.

[0052] In the embodiment of the present application, the randomly determined ordinal stamp k is , the ordinal stamp of the maximum noise adding step determined in advance.

[0053] The sample main body structure is sample data for training the main body structure design model. The sample main body structure can be an actual main body structure of an underwater vehicle that has been designed, or a simulation main body structure designed by using simulation design software. In the embodiment of the present application, the sample main body structure is represented by .

[0054] The generated Gaussian noise data complies with standard Gaussian distribution noise data.

[0055] Based on the randomly initialized ordinal stamp k, the sample main body structure and the generated Gaussian noise data , the noisy main body structure is generated specifically as , wherein , is the signal-to-noise ratio coefficient used in the first step.

[0056] S120: determining predicted Gaussian noise data based on the ordinal stamp k, the performance parameters of the sample main body structure and the noisy main body structure by using the main body structure design model.

[0057] The performance parameters of the sample main body structure are, for example, the maximum pressure resistance value, the actual stress of each structure unit under the maximum pressure resistance value, the weight of the sample main body structure, etc., wherein the necessary data is the maximum pressure resistance value.

[0058] In the case of treating the main body structure design model as a black box model, the predicted noise sequence is determined as based on the ordinal stamp k, the performance parameters of the sample main body structure and the noisy main body structure.

[0059] S130: Determine a training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and train the subject structure design model based on the training loss function.

[0060] In specific implementations, the training loss function can be a mean square error loss function In other embodiments, the training loss function can also be an absolute error loss function, a logarithmic loss function, a Huber loss function, etc., and the embodiments of the present application are not limited thereto.

[0061] As previously analyzed, in actual applications, there can be a large number of randomly initialized generated Gaussian noise data, and there can also be a large number of randomly initialized ordinal stamps k, so that a plurality of sets of training data for implementing training of the subject structure design model can be obtained in the case of only one sample subject structure. Correspondingly, in the case of only a plurality of sample subject structures, a sufficient number of training sample data for training the subject structure design model can also be generated. That is, the training method of the subject structure design model provided by the embodiments of the present application can realize model training with a small amount of sample subject structure data.

[0062] In some embodiments, the training method of the subject structure design model can include S140-S170 in addition to the foregoing S110-S130.

[0063] S140: Determine a denoised subject structure using the noisy subject structure and the predicted Gaussian noise data.

[0064] Determining a denoised subject structure using the noisy subject structure and the predicted Gaussian noise data can be performed by characterization.

[0065] S150: Perform stress analysis on each structural unit in the denoised subject structure to determine the predicted stress of each structural unit under the performance parameter corresponding environment condition.

[0066] In some embodiments, finite element analysis software can be used to perform stress analysis on the denoised subject structure. The structural model corresponding to the denoised subject structure can be input into the finite element analysis software, and the environment condition can be set to determine the predicted stress of each structural unit under the corresponding environment by the finite element analysis software.

[0067] In some other embodiments, a structural stress analysis model can be employed to analyze the stress of the denoised subject structure. The structural stress analysis model is a prediction model trained by a sample subject structure and actual stresses of each structural unit therein. In practical applications, the structural stress analysis model can be a Graph Neural Network (GNN). Correspondingly, the stress analysis of the denoised subject structure by the structural stress model specifically comprises: first, determining the topological relationship of each structural unit, and constructing a topological graph based on the topological relationship; the edges of the topological graph can represent the force transmission relationship between different structural units; then, applying external forces under environmental conditions on the pressure hull in the topological graph to obtain the initial stresses received by each structural unit; finally, analyzing the topological graph by the structural stress analysis model to determine the predicted stresses of each structural unit.

[0068] In practical applications, the structural stress analysis model can abstract the pressure hull and the rib structure of the subject structure into a topological graph structure, wherein the nodes in the topological graph structure represent the pressure hull units or the rib units, and the edges represent the connection relationship between the pressure hull units and the rib units. Through training of large sample data, the structural stress analysis model can quickly and accurately predict the stress values of each node unit according to the input topological graph.

[0069] S160: Determine the real stresses of each structural unit in the sample subject structure under the environmental conditions corresponding to the performance parameters.

[0070] Determining the real stresses of each structural unit in the sample subject structure under the environmental conditions corresponding to the performance parameters can be achieved by the foregoing method using finite element analysis software, can be achieved by using a structural stress analysis model, or can be directly determined from experimental data corresponding to the sample subject structure.

[0071] S170: Calculate a stress prediction loss function based on the predicted stresses and the real stresses of each structural unit.

[0072] Calculating the stress prediction loss function based on the predicted stresses and the real stresses of each structural unit can be a mean square error loss function, an absolute error loss function, a logarithmic loss function, etc., and the embodiments of the present application are not limited thereto.

[0073] In the case of performing the foregoing S140-S170, the foregoing S130 specifically comprises S131-S132 as follows.

[0074] S131: Determine a noise prediction loss function based on the generated Gaussian noise data and the predicted Gaussian noise data. The noise prediction loss function can employ the indicated above.

[0075] S132: determining a training loss function based on the stress prediction loss function and the noise prediction loss function.

[0076] Specifically, the training loss function is determined based on the stress prediction loss function and the noise prediction loss function, which can be wherein is the noise prediction loss function, is the stress prediction loss function, and is a preset weight, is the training loss function.

[0077] Based on the stress prediction loss function, the training loss function is determined, which can constrain the subject structure design model to meet reasonable performance parameters, and thus the predicted subject structure output by the subject structure design model can meet the corresponding stress constraints with high probability.

[0078] In actual applications, the generated subject structure should meet certain process rules, otherwise it is difficult to manufacture even if the corresponding subject structure is designed. For example, the thickness of the pressure-resistant shell in the subject structure should be smooth to avoid sudden changes in thickness that cause stress concentration, and the spacing between adjacent ribs in the subject structure should be greater than 50 mm to ensure that the welding tool can operate smoothly.

[0079] To enable the subject structure design model to learn process rules, in some embodiments, the training method of the subject structure design model further includes S180-S190 as follows.

[0080] S180: performing process compliance analysis on the denoised subject structure to determine a process constraint loss function.

[0081] In the embodiments of the present application, the process rules can be converted into a differentiable loss function, which represents the degree to which the designed structure meets the process rules, and the loss function will rapidly increase with the increase of the degree of non-compliance with the process rules, and this loss function is a convex loss function. It can be imagined that the process constraint loss function can punish the design scheme that does not meet the process requirements more.

[0082] In the case of performing the foregoing S180, the foregoing S170 can specifically be: determining the training loss function based on the stress prediction loss function, the noise prediction loss function, and the process constraint loss function.

[0083] If the process constraint loss function adopts indicates, the training loss function can be determined based on the stress prediction loss function, the noise prediction loss function, and the process constraint loss function, which can be indicates.

[0084] It can be conceived that, in the case of simultaneously considering the stress prediction loss function, the noise prediction loss function and the process loss function, the possibility of the subject structure prediction model generated by training to meet the physical performance requirements and the actual manufacturing process from the source is improved.

[0085] In the case of determining the training loss function by comprehensively considering the stress prediction loss function, the noise prediction loss function and the process loss function, and training the subject structure construction model by using the training loss function, the learning rate can be set to 0.0005. In order to improve the data quality and the model training effect, the sample subject structures can be divided into a training set, a validation set and a test set according to a ratio of 7:1:2 in the training process. The training set is used for learning and updating the model parameters, the validation set is used for monitoring the performance of the model in the training process and preventing overfitting, and the test set is used for finally evaluating the generalization ability of the model.

[0086] In the model training process, the validation set is used to monitor the performance of the model. In a specific application, when the training loss function reaches a minimum value in the validation set, and the pressure resistance value index of the generated subject structure reaches a standard rate of more than 95% and the process rule meets a rate of more than 98%, it is determined that the model has reached a good training effect, and the model training can be ended.

[0087] As previously analyzed, in some embodiments, the sample subject structure is a simulation subject structure designed by using a simulation design software. Specifically, the process of designing the simulation subject structure by using the simulation design software includes S210-S220.

[0088] S210: Obtain a combined design parameter based on various design parameter combinations.

[0089] The design parameters can include parameters of the pressure-resistant shell and parameters of the ribs. The parameters of the pressure-resistant shell include the thickness of the pressure-resistant shell and the material of the pressure-resistant shell; the parameters of the ribs include the rib thickness, the rib width, the number of ribs, the rib spacing, the rib cross-section shape parameter, the rib material, etc. In a specific implementation, the rib thickness can be set to 5-20mm, the rib width can be set to 20-100mm, the number of ribs can be set to 8-30, the spacing between adjacent ribs can be set to 50-300mm, and the rib cross-section shape parameter can be an I-shaped section, etc. After obtaining the values of various design parameters in the numerical range, the combined design parameters can be obtained by combining the aforementioned values.

[0090] S220: Generate a corresponding sample subject structure based on each group of generated design parameters.

[0091] After obtaining the combined design parameters, the corresponding sample body structure can be automatically generated by a finite element analysis software (such as Abaqus software). It is conceivable that by using the combined design parameters and combining the finite element analysis software automatic generation, the sample body structure conforming to the actual physical law can be obtained more accurately.

[0092] In a specific implementation, after obtaining the sample body structure, the finite element analysis software can also be used to perform finite element analysis on the sample body structure to determine the pressure resistance value of the sample body structure, the stress state of each internal structural unit under a specific pressure resistance value, the overall weight and other performance index parameters, and then provide corresponding data for training the body structure design model. It should be noted here that when performing finite element analysis, the elastic modulus, Poisson's ratio and other mechanical parameters of the actual material used need to be determined to ensure that the sample body structure simulates the actual application of the body structure.

[0093] It should also be noted that when using the finite element analysis software to perform finite element analysis, a reasonable water pressure range (from normal pressure to extreme pressure in deep sea, 0-1000 MPa) should be set according to the target application scenario. When performing finite element analysis, the boundary conditions need to be clearly defined, for example, the ends of the pressure-resistant shell body need to be fixed as constraints to simulate the installation state in actual work.

[0094] In actual application, in order to obtain a sufficient number of sample body structures, a special parameterized script such as Python-Abaqus interface script can be written, and the aforementioned script can be used to select and generate automatic combined design parameters within various design parameter ranges, and trigger the finite element analysis software to automatically generate sample body structures based on the combined design parameters and perform finite element analysis to obtain corresponding performance parameters.

[0095] It is conceivable that there is a close physical correlation between the geometric parameters and the performance parameters in the data set. For example, according to the principles of mechanics, increasing the rib thickness will significantly improve the pressure resistance of the structure, but at the same time, it will inevitably increase the overall weight; adjusting the rib spacing will affect the stress distribution uniformity of the structure. This strong correlation conforms to the basic laws of structural mechanics, providing a solid physical basis for subsequent model learning and prediction.

[0096] In the case of obtaining combined design parameters using the foregoing method and generating multiple sample body structures based on the combined design parameters, the multiple sample body structure data covers the conventional design range and various extreme working conditions, and the corresponding performance parameters also conform to various submersion depth conditions, which can make the data distribution reflect various possible application environment scenarios, and thus the body structure design model obtained based on the sample body structure has wide applicability.

[0097] It should be noted that in order to facilitate model training, the aforementioned sample body structure, performance parameters are quantified and normalized before use, eliminating the influence of different parameters due to the difference in dimension and value range, so that the data is more suitable for model training and learning. For example, the pressure resistance value and the weight are normalized to the interval [0, 1]. In addition, for non-numeric parameters such as material quality, they are converted into numerical form through coding. For example, titanium alloy is coded as 1, and aluminum alloy is coded as 2.

[0098] In practical applications, the sample body structure is generated based on high-precision finite element analysis software and is subjected to finite element analysis. A small number of sample body structures are also subjected to physical test verification, and the error between the performance parameter indicators obtained by the finite element analysis software and the actual physical experiment results is controlled to be less than 5%. The reliability and usability of the sample body structure are also verified, and the sample body structure can replace the real sample structure data for model training.

[0099] In addition, it should be noted that in practical applications, the finite element analysis software used in addition to Abaqus can be ANSYS, COMSOL Multiphysics and LS-DYNA.

[0100] The foregoing describes the training process of the body structure design model as a black box model. The following analyzes the training process of the body structure design model as a white box model. The body structure design model is analyzed as a white box model. In the case of taking the body structure design model as the aforementioned white box model (i.e., including the conditional encoder and the denoising diffusion probability network), the training process is as follows: when performing S120, first, the conditional encoder is used to process the performance parameters of the sample body structure to obtain a diffusion condition vector; then, the denoising diffusion probability network is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the diffusion condition vector and the noisy body structure. Correspondingly, S130 trains the body structure design model based on the training loss function, specifically: the conditional encoder and the denoising diffusion probability network are collaboratively trained based on the training loss function.

[0101] After the training of the body structure design model is implemented, the body structure design model can be used for underwater vehicle body structure prediction. Figure 3 is a flowchart of a method for generating an underwater vehicle body structure provided by an embodiment of the present application. As shown in Figure 3 the method for generating an underwater vehicle body structure includes S210-S240.

[0102] S210: Determine the required performance parameters that the to-be-designed underwater vehicle body structure needs to meet, and initialize the ordinal stamp k=K and the initial Gaussian noise data.

[0103] The performance parameters to be satisfied by the main body structure of the underwater vehicle to be designed are determined in advance, including the pressure resistance value of the pressure hull and the overall weight.

[0104] The timestamp k=K of the current denoising step is initialized, which is such that the main body structure design model starts to denoise the initial Gaussian noise data from the pre-determined Kth denoising step. As analyzed above, is the ordinal stamp of the maximum denoising step.

[0105] The initial Gaussian noise data is random noise data conforming to .

[0106] S220: Based on the ordinal stamp k, the performance parameters and the denoised main body structure, the main body structure design model is used to determine the noise data to be denoised.

[0107] In specific implementation, S220 includes the following steps (1)-(3).

[0108] (1) Random Gaussian noise is generated for each denoised noise data of the kth step: in the case of k>0 , in the case of k=0 .

[0109] (2) The main body structure design model is used to process the kth denoised noise data , the performance parameters and the ordinal stamp k+1 to obtain the denoised noise data .

[0110] It should be noted that when k=K, the denoised main body structure is the initial Gaussian noise data.

[0111] S230: The noise data to be denoised is subtracted from the initial Gaussian noise data to obtain the updated denoised main body structure.

[0112] The denoised noise data is removed from the kth denoised main body structure to obtain the k-1th updated denoised main body structure . .

[0113] S240: Update k=k-1, and determine whether k is -1; if not, execute S220; if yes, execute S250.

[0114] S250: The denoised main body structure determined in the last step is taken as the generated main body structure.

[0115] Because the main body structure design model in the embodiment of the present application is trained by the method in the foregoing, the main body structure generated by using the main body structure design model can also meet the corresponding conditions.

[0116] After obtaining the generated main body structure by the foregoing method, the performance parameters of the generated main body structure need to be determined to determine whether to adopt the generated main body structure. In specific implementation, the generated main body structure can be analyzed by a finite element analysis model to obtain parameters such as pressure resistance value, stress value of each structural unit, and weight.

[0117] In some other embodiments, the generated main body structure can also be processed by a pre-trained forward prediction model to determine the performance parameters of the generated main body structure. The task of the forward prediction model is to predict the corresponding weight and pressure resistance value according to the input main body structure. The forward prediction model can be a model using a convolutional neural network as the core architecture. It includes an input layer, a convolutional layer, and a fully connected layer.

[0118] The input layer is used to receive the quantized vector of parameters such as rib thickness, width, number, spacing distribution, rib material, etc. after standardization processing, for example, a 128-dimensional vector. The vector inputs the key information of the main body structure into the network in a compact form. In specific implementation, the input layer

[0119] The convolutional layer extracts local features between parameters by sliding a convolutional kernel of a preset size on the input vector. For example, it can capture the synergistic effect between rib thickness and spacing, and the potential relationship between width and number, etc. The design of the convolutional layer can effectively reduce the data dimension while retaining the key feature information.

[0120] The fully connected layer integrates the local features extracted by the convolutional layer to form a global understanding of the input geometric parameters. Through the calculation of the fully connected layer, a two-dimensional vector is finally output, and the two values of the two-dimensional vector correspond to the predicted weight value and pressure resistance value, respectively.

[0121] In specific application, a weight loss function and a pressure resistance value loss function can be used to form a joint loss function for training the parameters of the forward prediction model , wherein , are the predicted weight and pressure resistance value of the model, respectively, and are the true weight and true pressure resistance value, respectively, and The preset weight is used. In the implementation, when the parameters of the forward prediction model are updated by the back propagation algorithm, the Adam optimizer can be used to adjust the weight, and the learning rate is set to 0.001 to control the step of weight update, and ensure that the model can gradually converge to the optimal solution in the training process. In the specific training process, the performance of the model is monitored by using the validation set. By calculating the loss function on the validation set and observing its trend. When the loss function no longer decreases significantly on the validation set, or fluctuates, it is considered that the model may have reached a good training state, at which point the training is stopped. Generally, after about 300 iterations of training, the model can achieve good prediction accuracy, ensuring that the prediction error is less than 3%. In this way, the optimal number of iterations of the model is determined, so that the model can fully learn the data features on the training set while also having good generalization ability on the validation set and test set.

[0122] In some embodiments, other types of model architectures can also be used to build the forward prediction model. For example, a graph neural network can be used to establish the geometry of the main structure by constructing nodes (pressure hull and rib nodes) and edges (connections between pressure hull and ribs, ribs and ribs), and then capturing the topological association between components to achieve pressure value prediction and weight prediction.

[0123] In some other embodiments, a Transformer architecture can also be used to process the long-distance spatial dependence relationship of the pressure hull and rib network in the main structure using a self-attention mechanism to achieve pressure value prediction and weight prediction. In some embodiments, a hybrid forward prediction model can also be constructed using, for example, a graph neural network and a convolutional neural network to achieve pressure value prediction and weight prediction.

[0124] After obtaining the predicted pressure value as described above, it can then be determined whether the main structure generated above meets the design requirements based on the predicted pressure value. In addition, in specific applications, the structure stress analysis model described above can also be used to analyze the obtained main structure to determine whether the stress of each structural unit exceeds the corresponding structural size and material limit. In addition, it can also be determined from the structure data in the main structure model whether the structure meets the processing technology. If it does not meet the corresponding processing technology, the corresponding part of the foregoing main structure can be forcibly corrected according to the processing technology to obtain a corrected main structure. Then, the pressure value prediction and stress analysis of each structural unit are performed on the corrected main structure.

[0125] Figure 4 All flowcharts of the embodiments of the present application are shown, which show all steps of generating a reasonable main structure. Figure 4As shown, according to the foregoing method, first is sample data collection, specifically, a large amount of data is generated by using a parameterized script, and the corresponding sample body structure is automatically generated by using the large amount of data, and finite element analysis is performed on each sample body structure by using a finite element analysis software to obtain corresponding performance parameters. Subsequently, model training is performed, that is, the training of the body structure design model is realized by using the construction method of the body structure design model mentioned in the foregoing, and the training of the forward prediction model is realized; then is the use of the foregoing body structure design model and the forward prediction model, specifically, the performance parameters required are input into the body structure design model, the body structure design model is used to automatically generate the generated body structure, the foregoing body structure is forcibly modified according to the production process parameters, and finally the structure stress analysis model and the forward prediction model mentioned in the foregoing are used to analyze the generated body structure to determine whether the generated body structure meets the performance requirements, and further determine whether the automatic body structure generation is successful.

[0126] In practical applications, it is feasible to obtain original data according to the foregoing method. The finite element analysis software Abaqus is widely used in the industry, has mature technical documents and user community support. Although it requires certain programming and finite element knowledge to write Python-Abaqus interface scripts for parameterized simulation, Python, as a widely used programming language, has a relatively simple and understandable syntax, and the official Abaqus documentation provides detailed API interface descriptions and example code frameworks. Technical personnel can refer to the examples in the document about parameterized modeling, batch task submission and result export, combined with the geometric parameter range of the pressure hull rib (such as rib thickness 5-20mm, spacing 50-300mm, etc.), to write scripts to realize automatic data generation. Specifically, by defining geometric parameter variables, setting water pressure conditions (0-100MPa) and boundary conditions, the script can be used to iteratively generate models with different parameter combinations and run simulations, and finally export a dataset containing body structure geometric parameters, weight, pressure resistance and stress distribution. Even for technical personnel who are not familiar with script writing, they can also refer to publicly available Abaqus secondary development cases (quickly master the relevant methods to ensure that 100,000 sets of training data covering extreme working conditions can be generated as required.

[0127] In practical applications, in addition to the Abaqus software can be used to generate sample body structure, other finite element analysis software can also be used to generate sample body structure data. For example, ANSYS, COMSOL, Multiphysics or LS-DYNA can be used. Among them: ANSYS has the characteristics of supporting nonlinear analysis of complex structure, suitable for pressure shell simulation scenarios containing material plasticity, geometric buckling, etc.; COMSOL Multiphysics: good at multi-physical field coupling analysis, can consider the performance data generation of the rib under the combined load of water pressure, temperature, etc. simultaneously; LS-DYNA has the advantage of simulation precision in dynamic load (such as underwater impact) scenarios, which can supplement the data of extreme dynamic working conditions. Of course, other finite element analysis software can also be used to generate sample body structure. Such as ANSYS, ABAQUS or COMSOL, a three-dimensional model can be established, considering material nonlinearity, geometric nonlinearity, initial manufacturing defects, complex loading conditions and other factors, shell buckling analysis, yield analysis and even instability simulation. This method supports accurate simulation of special-shaped structures (such as shell with cabin section, opening, reinforcing rib). In specific applications, classical models such as Timoshenko shell stability theory and Von Mises yield criterion can also be used to calculate the pressure resistance value of the main structure with regular geometry and idealized load, especially the ratio of wall thickness to diameter of the long cylindrical shell and spherical shell main structure to realize the theoretical calculation of the pressure resistance value. In addition, the main structure determined by physical experiment data can also be supplemented as a sample main structure to ensure that the main structure design model has good generalization ability.

[0128] The model architecture involved in the present application can be reused mature deep learning framework and public network structure, and those skilled in the art can realize it through existing tools. The CNN network adopted by the forward prediction model can be directly based on the basic modules of PyTorch or TensorFlow to complete the construction of the convolutional layer and the fully connected layer, and the network parameters (such as the convolution kernel size 3x3 and the fully connected layer dimension 256) are the industry regular settings, without special innovation. The 3D-UNet backbone network in the subject structure prediction model can reuse the open source project (such as the 3D-UNet implementation in the MONAI library), and only needs to adjust the input layer dimension according to the rib bone voxel grid resolution (128x128x128); the conditional encoder can be realized by referring to the multi-head attention mechanism in the Transformer architecture, and the diffusion condition vector is associated and mapped with the 3D feature map; the real-time finite element proxy model based on GNN can be constructed by using the graph convolution layer (GCNConv) in the PyTorch Geometric library, and the stress prediction can be realized by defining the node features (rib element geometric parameters) and edge features (connection relationship). In addition, the process rule loss function is a differentiable loss function (such as distance constraint loss and thickness gradient loss), which can be realized by basic mathematical operations (such as absolute value and squared difference), and the formula is clear and easy to convert into code, so that the technical personnel do not need to rely on special algorithms to complete the model building.

[0129] The parameter setting and optimization method involved in the foregoing model training process are standard processes in the field of deep learning, which are operable. The training environment can be built based on a single server equipped with an NVIDIA GPU (such as RTX 3090 or A100), and the PyTorch / TensorFlow framework supports GPU acceleration maturely without additional special hardware configuration. In terms of training parameters, the learning rate (0.001 for the forward model and 0.0005 for the inverse model), the number of iterations (300-500 rounds), and the optimizer (Adam) are all conventional choices, and the technical personnel can fine-tune them according to the performance of the validation set (such as loss value change). During the data flow process, from the original data loading, standardization processing to input model calculation loss, all can be realized automatically through Python data processing library (such as Pandas, NumPy) and deep learning framework data loading tool without manual intervention. Even if overfitting problem may occur during the training process, the technical personnel can solve it through common regularization methods (such as Dropout layer and early stopping strategy), to ensure that the generalization performance of the model on the test set meets the standard (such as pressure prediction error <5% and process rule satisfaction rate >98%).

[0130] The practical application of intelligent design requires no complex operations; technicians can generate design schemes through simple interface calls. Input of process requirement parameters (target pressure resistance, weight limit, etc.) can be achieved through a visual interface (such as a simple tool developed based on PyQt). The parameter transformation logic (standardization, feature mapping) of the conditional encoder can be directly embedded into the input processing module via Python function coding. The closed-loop process of initial layout generation by the diffusion generator, stress analysis by the GNN surrogate model, and verification and correction by the process rule engine can be automated by writing serial scripts. Technicians only need to trigger the start command, and the system can complete iterative corrections and output results within 15 minutes. The final CAD model export can be achieved by calling the API of 3D modeling software (such as FreeCAD) to convert the voxel mesh into .step format. The included dimensions (thickness, spacing, etc.) can be directly extracted from the geometric parameters output by the model, ensuring that the generated design scheme can be directly used for manufacturing or simulation verification. The entire application process requires no professional AI knowledge; ordinary designers can master the operation methods after simple training.

[0131] In addition to the aforementioned methods, this application also provides a construction device 500 for a design model of the main structure of an underwater vehicle. Figure 5 This is a schematic diagram of the construction device 500 for the design model of the main structure of an underwater vehicle provided in this application embodiment. Figure 5 As shown, the main structure design model construction device 500 includes a structure noise addition unit 501, a noise prediction unit 502, and a model training unit 503.

[0132] The structure noise-adding unit 501 is used to generate a noise-adding main structure based on the randomly initialized ordinal stamp k, the sample main structure, and the randomly initialized generated Gaussian noise data.

[0133] The noise prediction unit 502 is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the performance parameters of the sample main structure, and the noisy main structure by adopting the main structure design model.

[0134] The model training unit 503 is used to determine the training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and to train the main structure design model based on the training loss function.

[0135] In some embodiments, the construction apparatus 500 of the main body structure design model further comprises a denoising main body structure determination unit and a stress determination unit. The denoising main body structure determination unit is configured to determine a denoising main body structure by using the noisy main body structure and the predicted Gaussian noise data. The stress determination unit is configured to perform stress analysis on each structural unit in the denoising main body structure, determine the predicted stress of each structural unit under the environmental condition corresponding to the performance parameter, and determine the real stress of each structural unit in the sample main body structure under the environmental condition corresponding to the performance parameter. Correspondingly, the model training unit 503 calculates a stress prediction loss function based on the predicted stress and the real stress of each structural unit, and determines a noise prediction loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and then determines a training loss function based on the stress prediction loss function and the noise prediction loss function.

[0136] In some embodiments, the stress determination unit determines the topological relationship of each structural unit, and constructs a topological graph based on the topological relationship. Then, a pre-trained structural stress analysis model is used to analyze the topological graph to determine the predicted stress of each structural unit. The structural stress analysis model is a graph neural network model.

[0137] In some embodiments, the model training unit 503 is further configured to perform process compliance analysis on the denoising main body structure to determine a process constraint loss function. Correspondingly, the training loss function is then determined based on the stress prediction loss function, the noise prediction loss function, and the process constraint loss function.

[0138] In some embodiments, the construction apparatus of the main body structure design model further comprises a sample generation unit. The sample generation unit first combines various design parameters to obtain combined design parameters, and generates a corresponding sample main body structure based on each group of generated design parameters. Then, the performance parameters of each sample main body structure are determined, including the maximum pressure value, the real stress of each structural unit under the maximum pressure value, and the weight of the sample main body structure.

[0139] In some embodiments, the main body structure design model comprises a conditional encoder and a denoising diffusion probability network.

[0140] The noise prediction unit 502 first processes the performance parameters of the sample main body structure by using the conditional encoder to obtain a diffusion condition vector. Then, the denoising diffusion probability network is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the diffusion condition vector, and the noisy main body structure. Correspondingly, the model training unit 503 jointly trains the conditional encoder and the denoising diffusion probability network based on the loss function.

[0141] The embodiments of the present application also provide an underwater vehicle main body structure generation method. Figure 6 is a flow chart of an underwater vehicle main body structure generation apparatus 600 method provided by the embodiments of the present application. As shown inFigure 6 As shown, the underwater vehicle body structure generation apparatus 600 comprises an initial parameter determination unit 601 and a de-noising unit 602.

[0142] The initial parameter determination unit 601 is configured to determine the required performance parameters of the underwater vehicle body structure to be designed, and initialize the ordinal stamp k=K and initial Gaussian noise data.

[0143] The de-noising unit 602 is configured to, in the case of k>0, make k=k-1; adopt the body structure design model, determine the noise data to be de-noised based on the ordinal stamp k, the required performance parameters and the de-noised body structure; and subtract the noise data to be de-noised from the initial Gaussian noise data to obtain an updated de-noised body structure until k=-1; wherein when k=K, the de-noised body structure is the initial Gaussian noise data.

[0144] The body structure determination unit is configured to determine the de-noised body structure of the last step as the generated body structure.

[0145] In some embodiments, the underwater vehicle body structure generation apparatus 600 further comprises a performance parameter determination unit 603. The performance parameter determination unit 603 is configured to analyze the generated body structure and determine the performance parameters of the generated body structure.

[0146] The embodiment of the present application provides a computing device, which can implement the training method of the underwater vehicle body structure design model as mentioned above and the underwater vehicle body structure generation method as mentioned above.

[0147] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, when the computer program is executed by a processor, the method mentioned above is implemented.

[0148] The embodiment of the present application provides a chip, which comprises at least one processor and an interface, the at least one processor determines program instructions or data through the interface; the at least one processor is configured to execute the program instructions to implement the method mentioned above.

[0149] The embodiment of the present application provides a computer program or a computer program product, which comprises instructions, when the instructions are executed, the computer executes the method mentioned above.

[0150] Those skilled in the art should further understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in a general manner. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0151] The above detailed description of the application serves to further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above is only a specific implementation of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A training method for a design model of the main structure of an underwater vehicle, characterized in that, include: A noisy main structure is generated based on randomly initialized ordinal stamp k, sample main structure, and randomly initialized Gaussian noise data; Using the aforementioned main structure design model, based on the ordinal stamp k, the performance parameters of the sample main structure, and the noisy main structure, the predicted Gaussian noise data is determined; The training loss function is determined based on the generated Gaussian noise data and the predicted Gaussian noise data, and the main structure design model is trained based on the training loss function.

2. The training method according to claim 1, characterized in that, The method further includes: The denoising main structure is determined using the noisy main structure and the predicted Gaussian noise data; Stress analysis is performed on each structural unit in the noise reduction main structure to determine the predicted stress of each structural unit under the environmental conditions corresponding to the performance parameters. Determine the actual stress of each structural unit in the main structure of the sample under the environmental conditions corresponding to the performance parameters; The stress prediction loss function is calculated based on the predicted stress and the actual stress of each structural unit. The step of determining the training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data includes: determining a noise prediction loss function based on the generated Gaussian noise data and the predicted Gaussian noise data; The training loss function is determined based on the stress prediction loss function and the noise prediction loss function.

3. The training method according to claim 2, characterized in that, The stress analysis of each structural unit in the noise reduction main structure to determine the predicted stress of each structural unit under the environmental conditions corresponding to the performance parameters includes: Determine the topological relationships of each of the structural units, and construct a topological graph based on the topological relationships; The topology is analyzed using a pre-trained structural stress analysis model to determine the predicted stress of each structural unit; the structural stress analysis model is a graph neural network model.

4. The training method according to claim 2 or 3, characterized in that, Also includes: The process compliance degree of the denoising main structure is analyzed to determine the process constraint loss function; The step of determining the training loss function based on the stress prediction loss function and the noise prediction loss function includes: The training loss function is determined based on the stress prediction loss function, the noise prediction loss function, and the process constraint loss function.

5. The training method according to any one of claims 1-3, characterized in that, Before generating the noisy main structure based on the randomly initialized ordinal stamp k, the sample main structure, and the randomly initialized generated Gaussian noise data, the method further includes: Based on the combination of various design parameters, combined design parameters are obtained, and corresponding sample main structures are generated based on the generated design parameters of each group. The performance parameters of each of the sample main structures are determined, including the maximum pressure resistance value, the actual stress of each structural unit under the maximum pressure resistance value, and the weight of the sample main structure.

6. The training method according to any one of claims 1-3, characterized in that, The main structure design model includes a conditional encoder and a denoising diffusion probability network; The main structure design model is adopted, and the predicted Gaussian noise data is determined based on the ordinal stamp k, the performance parameters of the sample main structure, and the noisy main structure. The performance parameters of the main structure of the sample are processed using the conditional encoder to obtain the diffusion condition vector; The denoising diffusion probability network is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the diffusion condition vector, and the denoising main structure. The step of training the main structure design model based on the training loss function includes: jointly training the conditional encoder and the denoising diffusion probability network based on the loss function.

7. A method for generating the main structure of an underwater vehicle, characterized in that, include: Determine the required performance parameters that the main structure of the underwater vehicle to be designed must meet, and initialize the ordinal stamp k = K and the initial Gaussian noise data; Repeat the following steps until k = -1: Using the main structure design model, based on the ordinal stamp k, the required performance parameters, and the denoising main structure of step k, the noise data to be denoised in step k is determined, and; wherein when k = K, the denoising main structure is the initial Gaussian noise data; Subtract the noise data to be denoised in step k from the denoised main structure in step k to obtain the denoised main structure in step k-1. Update k = k-1; The denoising structure determined in the final step is used as the generated structure.

8. The method according to claim 7, characterized in that, Also includes: The generated main structure is analyzed to determine its performance parameters.

9. A device for constructing a design model of the main structure of an underwater vehicle, characterized in that, include: The structure noise-adding unit is used to generate a noisy main structure based on the randomly initialized ordinal stamp k, the sample main structure, and the randomly initialized generated Gaussian noise data; The noise prediction unit is used to determine the predicted Gaussian noise data based on the ordinal stamp k, the performance parameters of the sample main structure, and the noisy main structure, using the main structure design model. The model training unit is used to determine the training loss function based on the generated Gaussian noise data and the predicted Gaussian noise data, and to train the main structure design model based on the training loss function.

10. A device for generating the main structure of an underwater vehicle, characterized in that, include: The initial parameter determination unit is used to determine the required performance parameters of the main structure of the underwater vehicle to be designed, and to initialize the ordinal stamp k = K and the initial Gaussian noise data; A denoising unit is used to ensure that k = k-1 when k > 0; using the main structure design model, based on the ordinal stamp k, the required performance parameters, and the denoising main structure, to determine the noise data to be denoised; and to subtract the noise data to be denoised from the initial Gaussian noise data to obtain an updated denoising main structure until k = -1. When k = K, the denoising main structure is the initial Gaussian noise data; The main structure determination unit is used to take the denoised main structure determined in the last step as the generated main structure.

Images