CN-122020173-A - Training method, evaluating method, device and equipment for box structure implosion evaluating model

CN122020173ACN 122020173 ACN122020173 ACN 122020173ACN-122020173-A

Abstract

The invention relates to the field of ship cabin implosion assessment, in particular to a training method, an assessment device and equipment for a box-type structure implosion assessment model. The method comprises the steps of obtaining dimensionless feature vectors and failure mode labels corresponding to original working condition parameters of ship cabin implosion, training a topology-aware classification network according to the dimensionless feature vectors and the failure mode labels, constructing and preheating a double-flow regression network, wherein one flow learns a global displacement reference based on data fitting, the other flow learns a local correction field based on physical equation constraint, applying boundary constraint to fusion results of the global displacement reference and the local correction field, inverting adjustable parameters in an optimized physical equation according to the constrained fusion results, and carrying out combined fine adjustment on the classification network, the double-flow regression network and the optimized adjustable parameters to obtain a final evaluation model. The improvement of the accuracy, efficiency and reliability of the ship cabin implosion assessment is realized.

Inventors

Lu Zhengcao
TAN XIAOJUN
SUN CHUANJIE
LU YONGGANG
SHI XIAOHAI
YU CHUNXIANG

Assignees

中国工程物理研究院总体工程研究所

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. The training method for the box structure implosion assessment model is characterized by comprising the following steps of: Acquiring dimensionless feature vectors, failure mode labels and deflection labels corresponding to original working condition parameters of ship cabin internal explosion, wherein the original working condition parameters comprise total explosion energy in a ship cabin, geometric dimensions and yield strength of a dominant structure, and the failure mode labels are used for representing failure conditions of the dominant structure; training a topology-aware classification network according to the dimensionless feature vector and the failure mode label; Constructing and preheating a double-flow regression network, wherein one flow learns a global displacement reference based on data fitting, and the other flow learns a local correction field based on physical equation constraint; Applying boundary constraint to the fusion result of the global displacement reference and the local correction field, and inverting and optimizing adjustable parameters in the physical equation according to the constrained fusion result; And carrying out joint fine adjustment on the classification network, the double-flow regression network and the optimized adjustable parameters according to the deflection label to obtain a final evaluation model.
2. The method of claim 1, wherein the obtaining a dimensionless feature vector corresponding to an original operating condition parameter of a ship cabin implosion comprises: According to the total explosion energy in the original working condition parameters, the plate thickness, the yield strength and the characteristic height of the dominant structure, calculating the dimensionless damage number : , Wherein, the For the total explosion energy, the total explosion energy is, In order to achieve a yield strength, the material, For characterizing the height dimension of the cabin or box structure in a direction perpendicular to the analyzed support structure, Is the plate thickness; obtaining a geometric stiffness factor according to the ratio of the characteristic length of the dominant structure to the plate thickness; Taking the characteristic length as a characteristic scale factor; And forming the dimensionless feature vector by the dimensionless damage number, the geometric rigidity factor and the feature scale factor.
3. The method of claim 2, wherein training the topology aware classification network based on the dimensionless feature vectors and the failure mode labels comprises: Inputting the dimensionless feature vector into a preset topology-aware classification subnet configured to identify a macroscopic failure mode of structural failure; calculating cross entropy loss according to the failure mode probability vector output by the topology perception classification sub-network and the failure mode label; Optimizing parameters of the topology aware classification sub-network by back propagation to minimize the cross entropy loss, resulting in a classification network.
4. A method according to claim 3, wherein said constructing and preheating a dual-flow regression network comprises: after the dimensionless feature vector and the failure mode probability vector are spliced, a preset data private sub-network is input, and the data private sub-network is configured to learn a displacement magnitude standard determined by a statistical rule; Calculating the mean square error loss of full-field displacement field data corresponding to the global displacement reference standard value and the original working condition parameter output by the data special sub-network, wherein the full-field displacement field data is used for representing the displacement condition of the dominant structure; Optimizing parameters of the data private sub-network through back propagation to minimize the mean square error loss, and obtaining a preheated data expert module; combining the dimensionless feature vector with a space coordinate grid defined on a structural calculation domain, and inputting the combination into a physical expert sub-network; calculating a first physical equation residual error generated after the local correction field output by the physical special sub-network is substituted into a preset rigid-plastic large deformation control equation; and obtaining a preheated physical expert module by optimizing parameters of the physical private sub-network to minimize the first physical equation residual error.
5. The method of claim 4, wherein the data-specific subnetwork comprises a plurality of fully-connected layers connected in sequence, the last connected layer being an output layer with a linear activation function for mapping network-learned features to scalar values representing global displacement references; the physical private subnetwork comprises a plurality of sequentially connected residual blocks, each residual block comprising two fully connected layers and a jump connection for adding an input of each residual block to an output of a second fully connected layer.
6. The method of claim 4, wherein after applying a boundary constraint to the fusion result of the global displacement reference and the local correction field, inverting and optimizing the adjustable parameters in the physical equation according to the constrained fusion result comprises: fusing the global displacement reference field output by the data expert module with the local correction field output by the physical expert module to form a preliminary prediction displacement field; applying geometric hard boundary constraint to the preliminary predicted displacement field to obtain a constrained predicted displacement field; Setting an equivalent energy transfer coefficient in the physical equation as a trainable parameter; substituting the constraint prediction displacement field into the preset rigid-plastic large deformation control equation under the condition of fixing all network weights to calculate a second physical equation residual error; And updating the equivalent energy transfer coefficient through a gradient descent algorithm to minimize the second physical equation residual error, so as to obtain the optimal physical parameters.
7. The method according to claim 1, wherein the performing joint fine tuning on the classification network, the dual-flow regression network, and the optimized adjustable parameters according to the deflection label to obtain a final evaluation model includes: Thawing all parameters of the classification network, the double-flow regression network and the optimized adjustable parameters; Constructing a composite loss function, wherein the composite loss function is obtained by weighted summation of data mean square error loss and physical equation residual error loss; and taking the composite loss function as a total target, and performing end-to-end optimization on the whole network according to the deflection label to obtain a final evaluation model.
8. The method for evaluating the explosion damage of the ship cabin is characterized by comprising the following steps of: Acquiring current working condition parameters of a ship cabin to be evaluated; carrying out dimension analysis on the current working condition parameters to construct a dimensionless feature vector comprising dimensionless damage numbers, geometric stiffness factors and feature scale factors; inputting the dimensionless feature vector into a pre-trained fusion evaluation model to obtain a final full-field displacement field prediction result, wherein the fusion evaluation model is obtained according to the method of any one of claims 1-7.
9. The utility model provides a box structure implosion evaluation model trainer which characterized in that includes: The system comprises an acquisition module, a failure mode label and a deflection label, wherein the acquisition module is used for acquiring dimensionless characteristic vectors, failure mode labels and deflection labels corresponding to original working condition parameters of the internal explosion of a ship cabin, the original working condition parameters comprise total explosion energy in the ship cabin, the geometric dimension and yield strength of a dominant structure, and the failure mode labels are used for representing failure conditions of the dominant structure; the classification training module is used for training a topology-aware classification network according to the dimensionless feature vector and the failure mode label; the preheating module is used for constructing and preheating a double-flow regression network, wherein one flow learns a global displacement reference based on data fitting, and the other flow learns a local correction field based on physical equation constraint; The inversion module is used for inverting and optimizing the adjustable parameters in the physical equation according to the constrained fusion result after applying boundary constraint to the fusion result of the global displacement reference and the local correction field; and the fine tuning module is used for carrying out joint fine tuning on the classification network, the double-flow regression network and the optimized adjustable parameters according to the deflection label to obtain a final evaluation model.
10. An electronic device comprising at least one processor, at least one memory, and computer program instructions stored in the memory, which when executed by the processor, implement the method of any of claims 1-8.

Description

Training method, evaluating method, device and equipment for box structure implosion evaluating model Technical Field The invention relates to the field of ship cabin implosion assessment, in particular to a training method, an assessment device and equipment for a box-type structure implosion assessment model. Background The explosion of the ship cabin is a typical destructive form in modern sea warfare. Unlike free field explosions, the cabin implosion environment is extremely complex involving multiple reflections, convergence of shock waves and accumulation of quasi-static air pressure (QSP). The structure may exhibit complex topological phase transitions from continuous "high deflection plastic deformation" to discrete "edge tearing/shear failure" under strong pulse loading. The existing assessment method has the following remarkable defects that the actual combat requirement is difficult to meet: the conventional pure data driven AI model is essentially based on data fitting of specific dimension analysis, relies on specific experimental data, cannot process complex geometric configuration and multi-working condition coupling, and cannot give full-field deformation details. Disclosure of Invention The invention aims to provide a training method, an evaluating device and equipment for an implosion evaluating model of a box structure, which solve the problems in the prior art. The invention is realized by the following technical scheme: in a first aspect, an embodiment of the present invention provides a training method for an implosion assessment model of a box structure, including: Acquiring dimensionless feature vectors and failure mode labels corresponding to original working condition parameters of ship cabin internal explosion, wherein the original working condition parameters comprise total explosion energy in a ship cabin, the geometric dimension and yield strength of a dominant structure, and the failure mode labels are used for representing failure conditions of the dominant structure; training a topology-aware classification network according to the dimensionless feature vector and the failure mode label; Constructing and preheating a double-flow regression network, wherein one flow learns a global displacement reference based on data fitting, and the other flow learns a local correction field based on physical equation constraint; Applying boundary constraint to the fusion result of the global displacement reference and the local correction field, and inverting and optimizing adjustable parameters in the physical equation according to the constrained fusion result; and carrying out joint fine adjustment on the classification network, the double-flow regression network and the optimized adjustable parameters to obtain a final evaluation model. Preferably, the obtaining the dimensionless feature vector corresponding to the original working condition parameter of the ship cabin implosion includes: According to the total explosion energy in the original working condition parameters, the plate thickness, the yield strength and the characteristic height of the dominant structure, calculating the dimensionless damage number : , Wherein, the For the total explosion energy, the total explosion energy is,In order to achieve a yield strength, the material,For characterizing the height dimension of the cabin or box structure in a direction perpendicular to the analyzed support structure,Is the plate thickness; obtaining a geometric stiffness factor according to the ratio of the characteristic length of the dominant structure to the plate thickness; Taking the characteristic length as a characteristic scale factor; And forming the dimensionless feature vector by the dimensionless damage number, the geometric rigidity factor and the feature scale factor. Preferably, the training topology aware classification network according to the dimensionless feature vector and the failure mode label includes: Inputting the dimensionless feature vector into a preset topology-aware classification subnet configured to identify a macroscopic failure mode of structural failure; calculating cross entropy loss according to the failure mode probability vector output by the topology perception classification sub-network and the failure mode label; Optimizing parameters of the topology aware classification sub-network by back propagation to minimize the cross entropy loss, resulting in a classification network. Preferably, the constructing and preheating a dual-flow regression network includes: after the dimensionless feature vector and the failure mode probability vector are spliced, a preset data private sub-network is input, and the data private sub-network is configured to learn a displacement magnitude standard determined by a statistical rule; Calculating the mean square error loss of full-field displacement field data corresponding to the global displacement reference standard value and the original working condition pa