Box-type structure implosion evaluation model training method, box-type structure implosion evaluation model evaluation method, device and equipment

Abstract

The invention relates to the field of ship cabin internal explosion assessment, in particular to a box type structure implosion assessment model training method, assessment method, device and equipment. Obtaining a dimensionless feature vector and a failure mode label corresponding to the original working condition parameters of the explosion in the ship cabin; training a topology-aware classification network according to the dimensionless feature vectors and the failure mode labels; a double-flow regression network is constructed and preheated, 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; after boundary constraint is applied to the fusion result of the global displacement reference and the local correction field, the adjustable parameters in the physical equation are inversed and optimized according to the constrained fusion result; and performing joint fine tuning on the classification network, the double-flow regression network and the optimized adjustable parameters to obtain a final evaluation model. The precision, efficiency and reliability of ship cabin indoor explosion evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of implosion assessment in ship compartments, specifically to a training method, assessment method, device, and equipment for a box-type structure implosion assessment model. Background Technology

[0002] Shipboard compartment explosions are a typical form of damage in modern naval warfare. Unlike free-field explosions, compartment explosions occur in extremely complex environments, involving multiple reflections and convergences of shock waves, as well as the accumulation of quasi-static pressure (QSP). Under strong pulse loads, structures exhibit a complex topological phase transition, ranging from continuous large-deflection plastic deformation to discrete edge tearing / shear failure.

[0003] Existing assessment methods have the following significant drawbacks, making them difficult to meet practical needs:

[0004] Traditional pure data-driven AI models: Traditional methods are essentially data fitting based on specific dimensional analysis, rely on specific experimental data, cannot handle complex geometric configurations and multi-condition coupling, and cannot provide details of deformation across the entire field. Summary of the Invention

[0005] The purpose of this invention is to provide a training method, evaluation method, device and equipment for evaluating the implosion assessment model of a box-shaped structure, which solves the problems in the prior art.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for training a box-shaped structure implosion assessment model, comprising:

[0008] Obtain the dimensionless feature vector and failure mode label corresponding to the original operating condition parameters of the explosion inside the ship's compartment; the original operating condition parameters include the total explosion energy inside the compartment, the geometric dimensions of the dominant structure, and the yield strength, and the failure mode label is used to characterize the failure status of the dominant structure;

[0009] A topology-aware classification network is trained based on the dimensionless feature vector and the failure mode label.

[0010] A dual-stream regression network is constructed and preheated, in which one stream learns a global displacement benchmark based on data fitting, and the other stream learns a local correction field based on physical equation constraints.

[0011] After applying boundary constraints to the fusion result of the global displacement reference and the local correction field, the adjustable parameters in the physical equation are inverted and optimized based on the constrained fusion result.

[0012] The classification network, the two-stream regression network, and the optimized adjustable parameters are jointly fine-tuned to obtain the final evaluation model.

[0013] Preferably, obtaining the dimensionless feature vector corresponding to the original operating parameters of the ship's internal explosion includes:

[0014] Calculate the dimensionless damage number based on the total explosive energy, plate thickness of the dominant structure, yield strength, and characteristic height from the original working condition parameters. :

[0015] ,

[0016] in, For total burst energy, For yield strength, The characteristic height is used to characterize the height dimension of a compartment or box-shaped structure in the direction perpendicular to the supporting structure being analyzed. For plate thickness;

[0017] The geometric stiffness factor is obtained based on the ratio of the characteristic length of the dominant structure to the plate thickness.

[0018] The feature length is used as the feature scale factor;

[0019] The dimensionless damage number, geometric stiffness factor, and characteristic scale factor are used to construct a dimensionless eigenvector.

[0020] Preferably, training the topology-aware classification network based on the dimensionless feature vector and the failure mode label includes:

[0021] The dimensionless feature vector is input into a preset topology-aware classification subnet, which is configured to identify macroscopic failure modes of structural damage.

[0022] The cross-entropy loss is calculated based on the failure mode probability vector output by the topology-aware classification subnet and the failure mode label.

[0023] The classification network is obtained by optimizing the parameters of the topology-aware classification subnet through backpropagation to minimize the cross-entropy loss.

[0024] Preferably, the construction and preheating of the dual-stream regression network includes:

[0025] The dimensionless feature vector is concatenated with the failure mode probability vector and then input into a preset data expert subnet. The data expert subnet is configured to learn a displacement magnitude benchmark determined by statistical laws.

[0026] The mean square error loss is calculated by comparing the global displacement reference scalar value output by the data expert subnet with the full-field displacement data corresponding to the original working condition parameters. The full-field displacement data is used to characterize the displacement of the dominant structure.

[0027] The parameters of the data expert subnet are optimized by backpropagation to minimize the mean square error loss, thus obtaining a preheated data expert module.

[0028] The dimensionless feature vector is combined with the spatial coordinate grid defined on the structural computation domain and input into the physics expert subnetwork;

[0029] The first physical equation residual is generated after substituting the local correction field output by the physical expert subnetwork into the preset rigid-plastic large deformation control equation.

[0030] By optimizing the parameters of the physics expert subnetwork to minimize the residual of the first physics equation, a preheated physics expert module is obtained.

[0031] Preferably, the data expert subnet includes multiple fully connected layers connected in sequence, with the last connected layer being an output layer with a linear activation function, used to map the features learned by the network to a scalar value representing a global displacement reference.

[0032] The physical expert subnet includes multiple sequentially connected residual blocks. Each residual block includes two fully connected layers and a skip connection, which is used to add the input of each residual block to the output of the second fully connected layer.

[0033] Preferably, after applying boundary constraints to the fusion result of the global displacement reference and the local correction field, the adjustable parameters in the physical equation are inverted and optimized based on the constrained fusion result, including:

[0034] The global displacement reference field output by the data expert module and the local correction field output by the physics expert module are fused to form a preliminary predicted displacement field.

[0035] Apply geometric hard boundary constraints to the preliminary predicted displacement field to obtain the constrained predicted displacement field;

[0036] Set the equivalent energy transfer coefficient in the physical equation as a trainable parameter;

[0037] Under the condition of fixing all network weights, the constrained predicted displacement field is substituted into the preset rigid-plastic large deformation control equation to calculate the residual of the second physical equation;

[0038] The equivalent energy transfer coefficients are updated using the gradient descent algorithm to minimize the residuals of the second physical equation, thereby obtaining the optimal physical parameters.

[0039] Preferably, the step of jointly fine-tuning the classification network, the two-stream regression network, and the optimized adjustable parameters based on the deflection label to obtain the final evaluation model includes:

[0040] Unfreeze all parameters of the classification network, the two-stream regression network, and the optimized adjustable parameters;

[0041] A composite loss function is constructed, which is obtained by weighted summation of data mean square error loss and physical equation residual loss.

[0042] Using the composite loss function as the overall objective, the entire network is optimized end-to-end based on the deflection label to obtain the final evaluation model.

[0043] Secondly, embodiments of the present invention provide a method for assessing explosion damage inside ship compartments, including:

[0044] Obtain the current operating parameters of the ship's compartments to be evaluated;

[0045] Dimensional analysis is performed on the current working condition parameters to construct a dimensionless feature vector containing dimensionless damage number, geometric stiffness factor and characteristic scale factor;

[0046] The dimensionless feature vector is input into the pre-trained fusion evaluation model to obtain the final full-field displacement field prediction result. The fusion evaluation model is obtained according to the method of the first aspect.

[0047] Thirdly, embodiments of the present invention provide a training device for a box-shaped structure implosion evaluation model, comprising:

[0048] The acquisition module is used to acquire dimensionless feature vectors, failure mode labels, and deflection labels corresponding to the original operating condition parameters of the explosion inside the ship's compartment. The original operating condition parameters include the total explosion energy inside the compartment, the geometric dimensions of the dominant structure, and the yield strength. The failure mode labels are used to characterize the failure status of the dominant structure.

[0049] The classification training module is used to train a topology-aware classification network based on the dimensionless feature vector and the failure mode label.

[0050] The preheating module is used to build and preheat the dual-stream regression network, in which one stream learns the global displacement benchmark based on data fitting, and the other stream learns the local correction field based on physical equation constraints.

[0051] The inversion module is used to apply boundary constraints to the fusion result of the global displacement reference and the local correction field, and then invert and optimize the adjustable parameters in the physical equation based on the constrained fusion result.

[0052] The fine-tuning module is used to jointly fine-tune the classification network, the two-stream regression network, and the optimized adjustable parameters based on the deflection labels to obtain the final evaluation model.

[0053] Fourthly, embodiments of the present invention provide an electronic device, including: 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 the first aspect described above.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] This solution improves the accuracy, efficiency, and reliability of shipboard compartment blast assessment by constructing a fusion evaluation model combining topological classification and dual-stream regression, deeply integrating data-driven and physics-driven approaches. The model uses dimensionless features and failure labels to train a classification network to identify macroscopic damage patterns. Simultaneously, it learns global displacement benchmarks through data experts and local correction fields through physics experts, and introduces geometric hard constraints and physical equation residual optimization. Finally, a unified evaluation framework is obtained through joint fine-tuning. Its technical advantages are: it inherits the ability of neural networks to mine complex patterns from massive amounts of data, while ensuring physical consistency of the solutions by embedding rigid-plastic control equations and boundary constraints; it improves computational efficiency by several orders of magnitude while maintaining high-fidelity prediction accuracy, achieving a leap from hourly simulation to millisecond-level evaluation; and it possesses strong generalization ability, capable of reasonably extrapolating conditions outside the training distribution, providing an efficient and reliable analytical tool for ship blast-resistant structure design, safety assessment, and protection optimization. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0057] Figure 1 Flowchart of the training method for the box-type structure implosion evaluation model provided by the present invention Figure 1 ;

[0058] Figure 2 Flowchart of the training method for the box-type structure implosion evaluation model provided by the present invention Figure 2 ;

[0059] Figure 3 A schematic diagram illustrating the classification performance provided by this invention;

[0060] Figure 4 Scatter plot of deflection predictions for the model provided by this invention on the training set (a) and the validation set (b);

[0061] Figure 5This is a schematic diagram comparing the residuals of the calculated deflection at y=0 with those of the experiment in the embodiments provided by the present invention;

[0062] Figure 6 A schematic diagram of the training device for the box-shaped implosion evaluation model provided by the present invention;

[0063] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0066] It should be noted that all actions involving the acquisition of signals, information, or data in this invention are carried out in compliance with the relevant data protection laws and regulations of the locality and with authorization from the owner of the relevant device.

[0067] Example 1

[0068] Please see Figure 1 This invention provides a method for training a box-type structure implosion assessment model, comprising:

[0069] S1. Obtain the dimensionless feature vector, failure mode label, and deflection label corresponding to the original operating condition parameters of the ship's compartment explosion; the original operating condition parameters include the total explosion energy in the compartment, the geometric dimensions of the dominant structure, and the yield strength, and the failure mode label is used to characterize the failure status of the dominant structure.

[0070] This step prepares structured data for model training. The original operating condition parameters refer to a quantified set of physical conditions and environmental properties that trigger an internal explosion, including but not limited to the total energy released by the explosive, the planar dimensions and thickness of the impacted bulkhead, and the mechanical strength indices of the constituent materials. These parameters directly describe the initial state of the load and structure in numerical form. Failure mode labels are category markers for the final failure mode of the structure under explosive loads, such as overall large deformation, edge tearing, or shear failure, providing a supervised target for subsequent classification learning. Deflection labels are the set of spatial vectors of material particles in a ship's box-shaped structure (such as bulkheads, decks, and rib frames) under internal explosive loads, from their initial position to their deformed position. These can be obtained from high-fidelity simulations (such as explicit dynamic finite element analysis) or limit experiments. Dimensionless eigenvectors are a set of dimensionless numerical combinations obtained by mathematically transforming the original operating condition parameters. Their construction aims to eliminate differences in magnitude and units between different physical quantities, projecting the high-dimensional heterogeneous original parameters into a low-dimensional space with inherent physical consistency, thereby improving the numerical stability and cross-scale generalization potential of the model learning.

[0071] This step first transforms the discrete set of operating parameters into numerical samples in a unified format based on the principle of physical similarity. Then, a dimensionless method is used to extract the core feature combinations that dominate the structural response. This process systematically organizes the multidimensional, coupled physical variables in the engineering problem into standardized inputs and corresponding output labels that can be directly processed by machine learning algorithms, laying a reliable data foundation for building a data-driven evaluation model. This preprocessing method reduces the complexity of subsequent network learning tasks and minimizes ill-conditioned numerical calculations caused by differences in physical dimensions.

[0072] In some embodiments, obtaining the dimensionless feature vector corresponding to the original operating condition parameters of the ship's internal explosion includes:

[0073] Calculate the dimensionless damage number based on the total explosive energy, plate thickness of the dominant structure, yield strength, and characteristic height from the original working condition parameters. :

[0074] ,

[0075] in, For total burst energy, For yield strength, The characteristic height is used to characterize the height dimension of a compartment or box-shaped structure in the direction perpendicular to the supporting structure being analyzed. For plate thickness;

[0076] The geometric stiffness factor is obtained based on the ratio of the characteristic length of the dominant structure to the plate thickness.

[0077] The feature length is used as the feature scale factor;

[0078] The dimensionless damage number, geometric stiffness factor, and characteristic scale factor are used to construct a dimensionless eigenvector.

[0079] Specifically, total explosive energy represents the total energy input from the outside and is the driving force behind the destruction.

[0080] Yield strength refers to a material's inherent ability to resist deformation.

[0081] Plate thickness refers to the geometric dimension of a structure in the thickness direction, which directly affects its cross-sectional bending and tensile strength.

[0082] Characteristic height refers to the dimension of the compartment in the direction perpendicular to the plate plane. Its introduction is a key innovation because the quasi-static pressure effect formed by the multiple reflections of the implosion shock wave in the enclosed space has an intensity and range directly related to the compartment volume. Together they constitute a dimension related to the volume of energy absorption.

[0083] It is the ratio of the external explosive impact energy to the energy dissipation potential of the structure's internal plastic deformation. A small value indicates a robust structure or a large compartment (energy is dispersed), meaning the explosive energy is insufficient to cause severe plastic deformation, and only minor deformation may occur. A high value indicates a large explosive yield or a relatively fragile structure, where the external energy input far exceeds the structure's dissipation capacity, potentially leading to large deformation, tearing, or even shear failure. Therefore, It is the decisive order parameter for predicting which damage mode the structure will eventually enter.

[0084] The geometric stiffness factor L / H, physically equivalent to the slenderness ratio of a structure, directly determines the dominant load-bearing mechanism and deformation pattern of a plate structure during the large deformation plastic stage. When this ratio is large, the structural behavior tends to be dominated by membrane tension, resulting in smooth deformation; when the ratio is small, the bending effect is more significant, and local plastic hinges are easily formed. This factor-guided model distinguishes the differences in deformation modes of different geometric configurations under the same load.

[0085] The feature scale factor L plays a crucial role in explicitly introducing physical scale information to compensate for prediction biases caused by the difference between scaled-down models and full-size prototypes. Due to nonlinear factors such as material strain rate sensitivity, the strict geometric similarity law often fails under explosive impact. By directly using the planar feature length L as an input feature, the model can automatically learn and correct for scale effects during training, thereby achieving reliable generalization predictions from small laboratory models to large-scale real-world ships.

[0086] S2. Train a topology-aware classification network based on the dimensionless feature vector and the failure mode label;

[0087] Topology-aware classification networks are specially designed neural network models that function to identify and distinguish different macroscopic failure modes in structural responses. This network treats failure mode discrimination as an independent classification task, aiming to learn the decision boundaries between different failure modes from input features.

[0088] This step uses the dimensionless feature vector output from step S1 as input and the corresponding failure mode label as the training target, adjusting the weight parameters within the network through an optimization algorithm. During training, the network learns to extract abstract representations strongly correlated with failure modes from complex feature combinations, ultimately outputting a probability distribution to quantify the likelihood that the input condition belongs to various preset failure modes. By prioritizing failure mode discrimination, the model logically decouples the discrete category identification problem from the subsequent continuous physical field prediction problem. This design enables the model to make accurate mode judgments in advance when facing extreme conditions that may lead to discontinuous topological changes in the structure, providing correct physical scene guidance for subsequent regression predictions and avoiding potential conflicts and performance degradation that may occur when a single model learns different types of mapping relationships.

[0089] In some implementations, S2, training a topology-aware classification network based on the dimensionless feature vector and the failure mode label, includes:

[0090] S21. Input the dimensionless feature vector into a preset topology-aware classification subnet, which is configured to identify macroscopic failure modes of structural damage.

[0091] Specifically, the topology-aware classification subnet is a dedicated artificial neural network module designed to distinguish and classify different macroscopic failure modes exhibited by structures under extreme loads. This network receives normalized operating condition characteristics as input, extracts abstract patterns related to failure modes through multi-layer nonlinear transformations, and ultimately outputs a discrete probability distribution to quantify the likelihood that the input operating condition belongs to various preset failure modes.

[0092] This step takes the dimensionless feature vector generated in the previous step and feeds it as a standardized input data stream to the input layer of the topology-aware classification subnet. This subnet, based on its pre-defined weight matrix and activation function, performs layer-by-layer transformation and information extraction on the input features. The activation function of the final layer of the subnet normalizes the network computation results into a probability vector, where each element's value ranges from zero to one, and the sum of all elements is one. This represents the network's confidence assessment of the various failure modes that the current operating condition may trigger.

[0093] S22. Calculate the cross-entropy loss based on the failure mode probability vector output by the topology-aware classification subnet and the failure mode label;

[0094] Specifically, cross-entropy loss is an objective function widely used in machine learning classification tasks to measure the difference between the probability distribution predicted by the model and the true label distribution. Its mathematical properties dictate that the loss value reaches its theoretical minimum when the predicted probabilities perfectly match the true labels; the greater the deviation, the greater the loss value. This loss function provides a clear and differentiable gradient direction for optimizing model parameters.

[0095] This step compares the failure mode probability vector output by the topology-aware classification subnet with the true failure mode labels annotated in the training data. The labels are typically represented using one-hot encoding, where the position corresponding to the true class is a 1, and the remaining positions are zeros. The calculation involves taking the logarithm of the probability vector, multiplying it by the label vector, and summing the results to obtain a scalar loss value. This scalar value quantifies the overall degree of discrepancy between the prediction results and the actual situation under the current network parameters. This calculation is part of the differentiable operation, ensuring that the gradient can propagate back along the computation graph.

[0096] S23. Optimize the parameters of the topology-aware classification subnet through backpropagation to minimize the cross-entropy loss, thereby obtaining the classification network.

[0097] Specifically, backpropagation is the core algorithm for training multi-layer neural networks. It uses a chain rule to calculate the contribution of each parameter in the network to the total loss, i.e., the gradient, layer by layer, starting from the loss value of the output layer. Parameters refer to the adjustable weights and biases in the neural network, which determine the mapping relationship from input to output. Optimization refers to the systematic process of iteratively adjusting these parameters to make the loss function value decrease.

[0098] This step uses the calculated cross-entropy loss value to initiate the backpropagation process. The algorithm automatically calculates the partial derivative of the loss function with respect to each trainable parameter in the classification subnet. Subsequently, based on the update rules of gradient descent or its variants, these gradient information are used to iteratively adjust the network parameters. Each adjustment moves the network a small step in the parameter space along the direction of loss descent. After multiple rounds of such forward prediction, loss calculation, backpropagation, and parameter update cycles on the training dataset, the network's internal representation gradually evolves, and its output probability distribution tends to align with the statistical characteristics of the true labels. When the training process converges, i.e., the loss value stabilizes at a low level and the classification accuracy reaches the preset requirement, the parameter state of the topology-aware classification subnet is fixed, thus obtaining a classification network with stable failure mode discrimination capabilities. This network forms the logical basis for subsequent hierarchical evaluation processes.

[0099] Furthermore, the classification network can be a topology-aware classification stream:

[0100] Input: Logarithmic feature space .

[0101] Function: As a "gating" module, it outputs a failure mode probability vector. The subsequent regression network is activated only when the condition is determined to be a continuous deformation mode (Mode I). This logically avoids the risk of unfitness when forcibly fitting a continuous model to a tearing condition (discontinuous field).

[0102] S3. Construct and preheat a dual-stream regression network, in which one stream learns the global displacement benchmark based on data fitting, and the other stream learns the local correction field based on physical equation constraints.

[0103] The dual-stream regression network is a parallel neural network architecture containing two functionally independent sub-network branches, called the data expert stream and the physical expert stream, respectively. The global displacement baseline is a scalar or low-dimensional vector output by the data expert stream, representing an estimate of the overall magnitude or average level of structural deformation under specific conditions, primarily derived from learning the statistical regularities of existing observational data. The local correction field is a spatial correlation function output by the physical expert stream, describing the local detailed changes and adjustments of the displacement field at different spatial locations of the structure, above the global baseline.

[0104] This step involves a phased warm-up training of the dual-stream regression network. First, the data expert stream learns under supervision, aiming to predict a baseline value that best fits the average level of existing full-field displacement data based on the input conditions, thereby quickly anchoring the magnitude range of deformation. Subsequently, the physics expert stream warms up in an unsupervised manner. Its training objective is not to directly fit the data, but rather to make its output modified field function satisfy a pre-defined fundamental physical governing equation describing the structural mechanical behavior as closely as possible. Specifically, the spatial derivative of the network's output field is calculated using automatic differentiation techniques, and the residual is calculated by substituting it into the governing equation. Minimizing this residual guides the network to learn a field function shape that conforms to physical laws. This decoupled warm-up strategy allows the two expert networks to initially grasp their respective core tasks: one focuses on the statistical trends presented by the data, and the other focuses on the differential relationships governed by physical laws, laying the foundation for subsequent collaborative work.

[0105] In some implementations, S3, constructing and preheating the dual-stream regression network, includes:

[0106] S31. After concatenating the dimensionless feature vector with the failure mode probability vector, the concatenation is input into a preset data expert subnet, which is configured to learn a displacement magnitude benchmark determined by statistical laws.

[0107] This step feeds a combined feature vector representing the physical nature of the working condition and the expected failure mode into a data expert subnet specifically designed for learning a global displacement magnitude benchmark. This subnet performs a nonlinear transformation on the input based on its internal mapping relationships, ultimately outputting a single scalar value. This scalar value aims to statistically predict the overall level of the structure's displacement response under the corresponding working condition, providing a stable magnitude reference for subsequent detailed reconstruction of the physical field. The establishment of this benchmark does not rely on spatial coordinate information but focuses on uncovering the correlation between input features and the final deformation scale.

[0108] S32. Calculate the mean square error loss between the global displacement reference scalar value output by the data expert subnet and the full-field displacement data corresponding to the original working condition parameters. The full-field displacement data is used to characterize the displacement of the dominant structure.

[0109] This step establishes a numerical link between the scalar values ​​output by the data expert subnet and the global statistics of actual physical observations. By calculating the square of the difference between the two, a scalar loss value is obtained, quantifying the degree of prediction deviation. This calculation provides a numerical carrier for the core supervision signal, and its output directly reflects the accuracy of displacement magnitude estimation under the current network parameters. This loss value provides a clear and differentiable optimization objective and gradient direction for the subsequent parameter optimization process.

[0110] S33. Optimize the parameters of the data expert subnet through backpropagation to minimize the mean square error loss, thereby obtaining a preheated data expert module;

[0111] This step drives the systematic adjustment of the data expert subnet parameters to reduce the error between its predicted scalar and the true statistical benchmark. Based on the gradient information calculated by the loss function, the adjustable parameters within the network are iteratively updated in the direction of error reduction. After multiple rounds of training iterations, the data expert subnet gradually converges to a stable state, and its output can reliably reflect the typical magnitude range of displacement response under different working conditions and failure modes. The parameterized model obtained at this point is the preheated data expert module, which has the ability to provide reasonable prior displacement magnitudes for unknown working conditions.

[0112] S34. Combine the dimensionless feature vector with the spatial coordinate grid defined on the structural computation domain and input it into the physics expert subnetwork;

[0113] This step constructs a combined input for the physics expert sub-network, incorporating both global state information and local location information. By concatenating the feature vector describing the global operating conditions with the coordinates of discrete points within the spatial computational domain, the network is provided with the fundamental independent variables and boundary conditions necessary for generating the spatial field function. This input structure explicitly tells the network that its task is to learn a continuous function defined in the spatial domain, given global physical constraints, the shape of which is determined by both the coordinates and the operating conditions.

[0114] S35. Calculate the physical equation residuals generated after substituting the local correction field output by the physical expert subnetwork into the preset rigid-plastic large deformation control equation.

[0115] The core of this step lies in evaluating the consistency between the output field of the physics expert sub-network and the fundamental physical laws. Using automatic differentiation techniques, the spatial partial derivatives required to satisfy the physical equations are directly calculated from the network output field. Subsequently, the output field values, their derivatives, and operating parameters are substituted into the preset governing equations, and the differences between the left and right sides of the equations are calculated at global sampling points. These differences are normalized to obtain a scalar residual characterizing the overall deviation of the network's predicted field from the physical constraints. This residual serves as an unsupervised training signal, independent of any real displacement data, and generated solely from the physical laws themselves.

[0116] S36. By optimizing the parameters of the physical expert sub-network to minimize the residuals of the physical equations, a preheated physical expert module is obtained.

[0117] This step uses the residuals of the physical equations as the optimization objective to perform unsupervised parameter training on the physics expert sub-network. The gradient of the residuals with respect to the network parameters is calculated through backpropagation, and the parameters are iteratively adjusted accordingly, so that the network's output field function and its differential properties gradually approximate the preset governing equations. This process internalizes the differential constraints and conservation laws inherent in the equations within the network. When the training converges and the residuals stabilize at a low level, it indicates that the network has learned to generate a spatial field function that is mathematically consistent with the physical model. The network obtained at this point is the preheated physics expert module, possessing the ability to construct complex spatial deformation forms based on physical principles.

[0118] In some embodiments, the data expert subnet includes multiple fully connected layers connected in sequence, with the last connected layer being an output layer with a linear activation function, used to map the features learned by the network to a scalar value representing a global displacement reference.

[0119] The physical expert subnet includes multiple sequentially connected residual blocks. Each residual block includes two fully connected layers and a skip connection, which is used to add the input of each residual block to the output of the second fully connected layer.

[0120] Specifically, multiple sequentially connected fully connected layers constitute a standard multilayer perceptron. The output of each layer serves as the input to the next layer, allowing information to flow unidirectionally. The final layer employs a linear activation function, meaning its output is a weighted sum of the input features plus a bias, without undergoing a nonlinear transformation.

[0121] In this architecture, the input combined feature vector (working condition and pattern probabilities) first passes through several fully connected layers with nonlinear activation functions. These layers are responsible for nonlinear transformation and high-order abstraction of the features, extracting deep patterns related to the global displacement magnitude from the original input. Since the learning target is a scalar independent of spatial location, the network does not require a complex structure to capture spatial correlations. The final linear output layer directly maps the high-dimensional feature representation learned by the preceding nonlinear layers to a one-dimensional numerical space, which is interpreted as the global displacement baseline. Linear activation ensures that the output value can cover any range of real numbers to adapt to displacement predictions of different magnitudes. This relatively simple and direct feedforward structure is well-suited for learning a deterministic mapping from finite high-dimensional features to a single continuous value, enabling efficient and stable regression of the overall displacement level. Its structural complexity matches the task difficulty, avoiding unnecessary parameter redundancy and overfitting risks.

[0122] Each residual block is a basic computational unit containing short-circuit connections. Within a residual block, the input first passes through a first fully connected layer and a nonlinear activation layer, then through a second fully connected layer. Finally, the output of this second fully connected layer is directly added to the original input of the block, and after activation, the output is released. Multiple such blocks are sequentially connected to form a deep network.

[0123] In this architecture, the input (a combination of coordinates and conditions) flows through the first residual block. Skip connections allow the input information of the block to be losslessly passed to the output, where it is fused with the information after two layers of nonlinear transformations. This design mitigates the gradient vanishing problem that can occur in deep networks, enabling the network to be built and trained more effectively. Multiple sequentially connected residual blocks can perform deep and complex nonlinear transformations on the input features. For physics expert networks, the task is to learn a highly complex spatial field function that must satisfy specific partial differential equations. This requires the network not only to fit complex nonlinear forms but also to accurately compute the higher-order derivatives of its output with respect to the input coordinates. Deep networks provide powerful function approximation capabilities to capture the complex forms of the field, while the residual structure ensures effective gradient backpropagation at such depths, making it possible to compute high-precision, stable spatial derivatives through automatic differentiation. Therefore, this deep residual architecture is particularly suitable for physics-driven learning tasks that require accurate modeling of complex spatial differential relationships, enabling the network to generate mathematically smooth modified fields that satisfy deep physical constraints.

[0124] S4. After applying boundary constraints to the fusion result of the global displacement reference and the local correction field, the adjustable parameters in the physical equation are inverted and optimized based on the constrained fusion result.

[0125] This step is crucial after the warm-up of the dual-stream regression network. Its purpose is to calibrate the uncertainty parameters in the physical model using the initial displacement field predicted by the network, with fixed network weights. Specifically, firstly, the global displacement benchmark learned by data experts is fused with the local correction field learned by physics experts, and geometric hard constraints consistent with the actual deformation mode of the structure are applied to form a more physically reasonable trial function. Subsequently, this constrained displacement field is substituted into the rigid-plastic governing equations describing the large deformation behavior of the structure. By optimizing key parameters in the equations (such as the equivalent energy transfer coefficient), the residuals of the equations are minimized, thereby retrieving the physical parameter values ​​that best explain the current prediction results. This process enables the calibration of the first-principles model with data-driven prediction results, providing a more physically consistent starting point for subsequent joint fine-tuning.

[0126] In some implementations, S4, after applying boundary constraints to the fusion result of the global displacement reference and the local correction field, optimizes the adjustable parameters in the physical equations based on the constrained fusion result, including:

[0127] S41. The global displacement reference field output by the data expert module and the local correction field output by the physics expert module are fused to form a preliminary predicted displacement field.

[0128] This step performs the initial fusion of the outputs from the dual-stream regression network. The data expert module outputs a scalar field uniformly distributed in space, i.e., the global displacement reference field, which provides a benchmark for the magnitude of deformation. The physics expert module outputs a field that varies with spatial coordinates, i.e., the local correction field, which provides local details and physical form of deformation. The two are then weighted and fused by adding them point-by-point or by using Softmax activation to obtain the weights of the two expert networks, resulting in a complete spatial displacement field prediction that integrates global statistical regularities and local physical constraints—the preliminary predicted displacement field. This field is the direct object of subsequent physical consistency assessment and parameter optimization; it simultaneously carries the magnitude information learned from the data and the differential structure information learned from the physical equations.

[0129] Furthermore, the physical-data dual-stream regression employs a residual fusion mechanism and includes two parallel expert modules:

[0130] Data Expert: Receives only operating parameters and classification probabilities, and learns the mapping. . Used to anchor a global displacement benchmark determined by statistical laws.

[0131] Physics Expert: Receives spatial coordinates (X, Y) and operating parameters, and learns the local differential corrections that satisfy PDE constraints. .

[0132] Final output: .

[0133] S42. Apply geometric hard boundary constraints to the preliminary predicted displacement field to obtain the constrained predicted displacement field;

[0134] To ensure that the predicted displacement field strictly conforms to the physical facts of the structural support boundaries (such as fixed or simply supported) and to suppress boundary noise commonly found in numerical solutions, it is necessary to impose geometric hard constraints on the initial predicted displacement field. Specifically, this is achieved by constructing a distance function mask or trial function related to the normalized coordinates of the structure, forcing the displacement field to satisfy preset displacement and rotation conditions at the boundaries. For example, at a fixed support edge, the displacement and normal gradients are forced to zero. After this processing, the initial predicted displacement field is corrected into a geometrically completely compatible constrained predicted displacement field, whose shape more closely resembles actual physical deformation, especially reflecting the characteristics of plastic hinge formation during large deformation stages.

[0135] Specifically, to address boundary numerical noise and conform to the physical fact of plastic hinge formation during large deformation, this invention constructs a trial function that satisfies geometric compatibility. For normalized coordinates... For the plate frame, construct a distance function mask:

[0136] ;

[0137] The network's final physical field output Defined as:

[0138] ;

[0139] Technical effect: This structure ensures that at the boundary ( The displacement is always zero (W=0), but the normal rotation angle is allowed to evolve freely. This is consistent with the physical mechanism of plastic slippage at the edge of the structure under implosion high pressure, and avoids the non-physical high-frequency oscillations (spectral bias) caused by forced zero rotation angle (fixed support).

[0140] S43. Set the equivalent energy transfer coefficient in the physical equation as a trainable parameter;

[0141] This step specializes in a key coefficient embedded in the physical governing equations. The equivalent energy transfer coefficient is a scalar parameter with a clear physical meaning, macroscopically characterizing the efficiency of converting explosive chemical energy into structural plastic deformation work, integrating complex fluid-structure interaction and energy dissipation mechanisms. In traditional methods, this coefficient is often estimated using empirical formulas, which introduces uncertainty. In this step, the coefficient is transformed from a fixed constant into a trainable variable in the neural network computation graph. This means that in subsequent optimization processes, its value, like the weights of the neural network, can be adjusted via gradient descent based on feedback from the loss function, thus freeing its value from prior assumptions and allowing it to be determined by both observational data and physical laws.

[0142] S44. Under the condition of fixing all network weights, substitute the constrained predicted displacement field into the preset rigid-plastic large deformation control equation to calculate the residual of the second physical equation.

[0143] In this optimization sub-step, the weight parameters of all neural networks (including the classification network, data expert, and physics expert) are frozen and no longer updated. At this point, the constrained predicted displacement field obtained in the previous step, satisfying the boundary conditions, along with the current equivalent energy transfer coefficient to be optimized, is substituted into the preset rigid-plastic large deformation governing equation. The error generated when this displacement field is used as a solution to the equation is calculated; this is the residual of the second physical equation. This residual measures the degree of deviation between the displacement field predicted by the network and the first physical principles under the current equivalent energy transfer coefficient.

[0144] The governing equations for rigid-plastic large deformation can be ,in, It is a dimensionless Laplace operator (characterizing plastic film force).

[0145] It is the equivalent energy transfer coefficient.

[0146] The normalized equivalent quasi-static load term can be obtained through... get.

[0147] Parameter inversion mechanism: This is set as the only trainable physical parameter in the network. During training, an automatic differentiation mechanism is used to automatically derive the optimal value by minimizing the physical residual and data error. This value macroscopically characterizes the effective proportion of explosive chemical energy converted into structural plastic work, achieving adaptive compensation for complex fluid-structure interaction effects.

[0148] To give the neural network physical interpretability, this invention embeds a governing equation based on the Rigid-Plastic Large Deflection Theory into the loss function.

[0149] Physical assumptions: For the scenario of a large deformation due to an explosion inside a ship's compartment, it is assumed that the maximum deflection w is much greater than the plate thickness H (i.e., At this point, the bending stiffness of the structure is negligible, and the load-bearing mechanism is dominated by in-plane membrane force.

[0150] Dimensional equilibrium equations: The geometric configuration at the end of the structure satisfies the following plastic membrane force equilibrium equations:

[0151]

[0152] in, The force per unit width of the fully plastic film, For the Laplace operator, This is an equivalent quasi-static load.

[0153] Dimensionless transformation: To eliminate the difference in physical magnitudes, the above equations are mapped to a dimensionless characteristic space. Dimensionless variables are defined. The derived dimensionless evolution equation is as follows:

[0154]

[0155] in:

[0156] The first two items Characterizes the dimensionless internal plastic film force (resistance);

[0157] The normalized external load term is derived from the dimensionless impulse factor. Decide;

[0158] (Key parameter): Defined as the equivalent energy transfer coefficient. In this invention, It is no longer a fixed empirical constant, but is set as a trainable parameter in the neural network. It represents the effective proportion of explosive chemical energy converted into structural plastic work.

[0159] S45. Update the equivalent energy transfer coefficient using the gradient descent algorithm to minimize the residual of the second physical equation and obtain the optimal physical parameters.

[0160] This step focuses on optimizing the equivalent energy transfer coefficient specifically, using the physical equation residuals as the sole optimization objective. Since all neural network weights are fixed, the gradient of the loss function with respect to this coefficient can be clearly obtained through backpropagation of the computational graph. This gradient indicates the direction and magnitude in which the equivalent energy transfer coefficient should be adjusted to reduce the physical residuals (i.e., improve the compatibility between the predicted field and the theoretical equations). Using a gradient descent algorithm, the value of this coefficient is iteratively updated along the reverse direction of the gradient. After multiple iterations, the coefficient converges to a value that brings the second physical equation residuals to a local minimum or a stable low value. The coefficient value obtained at this point represents the optimal physical parameter derived through the joint driving force of data and physics. Mathematically, this parameter value represents an equivalent representation value that, for the currently trained network model and its predictions, achieves the best compromise between the simplified theoretical physical model and complex actual observations, thus realizing adaptive data calibration for unknown and complex physical processes.

[0161] S5. Based on the deflection labels, the classification network, the dual-stream regression network, and the optimized adjustable parameters are jointly fine-tuned to obtain the final evaluation model.

[0162] Joint fine-tuning is the final training phase conducted after the initial training and parameter inversion of each component of the network, releasing the frozen state of all parameters, with the goal of overall optimization. The final evaluation model refers to a complete neural network system that, after the complete training process, integrates topology classification, data regression, and physical constraint capabilities, and can be used for end-to-end prediction of new operating conditions.

[0163] This step integrates the classification network, two-stream regression network, and optimal adjustable parameters obtained in the previous steps into a complete computational graph. Based on this, a composite loss function is defined, which simultaneously incorporates the mean squared error loss from data fitting and the residual loss from the physical equations. Subsequently, using all the training data prepared in step S1 as input, and this composite loss as the optimization objective, end-to-end gradient descent optimization is performed on all parameters of the entire ensemble model based on the model's final prediction output and deflection labels. During this stage, the classification network, data expert stream, physical expert stream, and physical parameters can all be finely adjusted based on the total loss feedback, prompting each component to work collaboratively while maintaining its core capabilities, achieving a balance between overall prediction accuracy and physical consistency. Through this comprehensive optimization process, the model ultimately forms a unified and efficient inference system capable of coherently completing failure mode identification and high-fidelity displacement field prediction after receiving dimensionless features of new operating conditions.

[0164] In some implementations, S5 involves jointly fine-tuning the classification network, the two-stream regression network, and the optimized adjustable parameters to obtain the final evaluation model, including:

[0165] S51. Unfreeze all parameters of the classification network, the two-stream regression network, and the optimized adjustable parameters;

[0166] This step removes the update restrictions on all learnable variables in the model. In previous stages, the weights of the classification network, data expert module, and physics expert module have completed initial training and stabilized, while the equivalent energy transfer coefficients have been independently inverted. This step resets these partially optimized parameter states to an updatable state. This means that in subsequent optimization loops, gradient information can be applied to every adjustable part of the entire network, including the weights and biases of each network layer, as well as the physical parameters as special nodes. This operation places the previously obtained local optima or suboptimal solutions obtained in stages and tasks within a unified global optimization framework, providing a prerequisite for the final coordination and fine-tuning of the model's components.

[0167] S52. Construct a composite loss function, which is obtained by weighted summation of data mean square error loss and physical equation residual loss;

[0168] This step defines a multi-objective optimization function to comprehensively guide the model's final behavior. The composite loss function is a weighted linear combination of multiple independent loss terms. Specifically, the mean squared error loss measures the numerical closeness between the model's predicted deflection and the deflection label; the physical equation residual loss measures the physical consistency of the predicted displacement field in satisfying the fundamental mechanical governing equations. Each loss term is assigned a pre-defined non-negative weight coefficient, which determines the relative importance or priority of each task in the final optimization objective. Through weighted summation, the three training objectives—classification accuracy, data fit, and physical consistency—that might otherwise compete or conflict, are unified and quantified into a single, differentiable scalar total loss. This total loss provides a unique and explicit mathematical objective for subsequent global optimization.

[0169] Furthermore, the training objective of this invention is to minimize the composite loss function. It consists of two parts: data fitting error and physical equation residuals.

[0170] The formula is expressed as:

[0171]

[0172] The specific definitions are as follows:

[0173] , These are coefficients.

[0174] Data mean square error loss ( ):

[0175] The magnitude of the physical field is used to anchor it. The mean square error (MSE) is used to calculate the deviation between the predicted value and the deflection label (experimental / fine simulation data):

[0176]

[0177] in, The model predicts the first Predict the deflection of the target. For the first The deflection labels corresponding to the original working parameters The number of high-fidelity samples (e.g., 82 groups).

[0178] Physical equation residual loss ( ):

[0179] Differential manifold structures used to constrain physical fields. The residuals of the governing equations are calculated at collocational points throughout the domain:

[0180]

[0181] in, The number of sampling points used to calculate the physical residual. The first output of the model Predicted deflection of the target at each sampling point location For the first A normalized equivalent quasi-static load term. In this calculation... It participates in gradient backpropagation as a variable. The network will search for an optimal one. This minimizes the physical residual.

[0182] Weighting: To balance the dominant role of sparse data with the regularization effect of physical equations, weighting coefficients are set. .

[0183] S53. Using the composite loss function as the overall objective, perform end-to-end optimization on the entire network to obtain the final evaluation model.

[0184] This step uses a composite loss function as a global guide to perform final joint training on the ensemble network. Preprocessed training data (condition features, spatial coordinates) is input into the ensemble network with all parameters unfrozen, and a complete forward propagation is performed. This yields the failure mode probabilities, global baseline field, local correction field, and fused displacement field sequentially, while simultaneously calculating the losses for classification, data fitting, and physics. These are then summarized into a total loss. Using the backpropagation algorithm, the gradient of the total loss with respect to each trainable parameter in the network (including network weights and physical parameters) is calculated. These gradients comprehensively reflect the combined impact of each parameter adjustment on pattern classification accuracy, data fitting precision, and compliance with physical laws.

[0185] An optimization algorithm is employed to synchronously iteratively update all parameters based on this comprehensive gradient information. During this process, the classification network may fine-tune its decision boundary to better coordinate with regression prediction; the data expert module may fine-tune its magnitude output to better fit the data under physical constraints; the physics expert module may fine-tune its field function shape to simultaneously adapt to data distribution and better physical parameters; and the equivalent energy transfer coefficient may also be fine-tuned around its inversion value to seek a better balance point within the collaborative framework. After multiple iterations, training stops when the total loss converges to a stable low value and the losses of each component reach an acceptable balance. The resulting complete network system, where all component parameters have been collaboratively optimized for overall performance, is the final fusion evaluation model. This model possesses end-to-end inference capabilities, capable of coherently, stably, and balancedly outputting high-confidence failure mode discrimination and high-fidelity physical field prediction after receiving the original operating condition parameters.

[0186] Furthermore, the model hyperparameters are configured as follows:

[0187]

[0188] For example, such as Figure 2 The diagram shown is a flowchart of the training process, which mainly includes the following steps:

[0189] 1. Input and Gating Networks

[0190] Input includes:

[0191] Topology Classifier

[0192] Condition Information (such as boundary conditions, physical parameters, etc.)

[0193] physical variables ;

[0194] Coordinate information ;

[0195] 2. Shuangliu Expert Network

[0196] Stream 1: Data Expert

[0197] It is a fully connected network (FC) that uses the tanh activation function to learn global, data-driven patterns as a global baseline, suitable for regions without strong physical constraints.

[0198] Stream 2: Physics Expert

[0199] Composed of multiple ResBlocks, it learns local, physics-driven residual corrections and is suitable for regions with well-defined physical laws (such as near boundaries and singularities).

[0200] 3. Integration and Constraints

[0201] The output of the gated network is activated by Softmax, resulting in the weights of the two expert networks.

[0202] Fusion output = Weight 1 × Data expert output + Weight 2 × Physics expert output.

[0203] Boundary functions are used to apply physical boundary conditions.

[0204] Loss calculations combine: data fitting loss (such as MSE), physical constraint loss (such as PDE residuals), and boundary condition loss.

[0205] Training effect Figure 3-5 As shown, where Figure 3 A schematic diagram illustrating the classification performance of an embodiment of the model; Figure 4 Scatter plots showing the deflection predictions of the model on the training set (a) and the validation set (b); Figure 5 This is a schematic diagram comparing the residuals of the calculated deflection at y=0 with those of the experiment in the example.

[0206] Example 2

[0207] This invention provides a method for assessing internal explosion damage in ship compartments, including:

[0208] Obtain the current operating parameters of the ship's compartments to be evaluated;

[0209] The current operating condition parameters are a quantitative set that must fully encompass the key factors that dominate the structural dynamic response. This includes at least: the total chemical energy of the explosive charge that triggers the explosion (total blast energy), the planar characteristic dimensions and thickness (geometric dimensions) of the impacted compartment walls, the yield strength of the constituent materials, and the height of the compartment perpendicular to the walls. These parameters collectively and uniquely define a specific implosion physics problem and serve as the raw data foundation for all subsequent calculations and inferences. These parameters may be obtained from design drawings, sensor readings, or mission scenarios; their accuracy and completeness directly affect the reliability of the final evaluation results.

[0210] Dimensional analysis is performed on the current working condition parameters to construct a dimensionless feature vector containing dimensionless damage number, geometric stiffness factor and characteristic scale factor;

[0211] This step transforms the original working condition parameters, which have different physical units and dimensions, into a standardized, dimensionless mathematical characteristic representation. This process is based on dimensional analysis principles and specific physical similarity criteria. First, dimensionless damage numbers are calculated based on the total burst energy, plate thickness, material yield strength, and compartment height. These numbers characterize the competition between the external load input and the dissipation potential within the structure. Next, the geometric stiffness factor is calculated based on the ratio of the panel's planar characteristic dimension to its thickness. This factor determines the structure's stiffness characteristics and dominant deformation mode during large deformation stages. Finally, the planar characteristic dimension of the panel itself is introduced as a characteristic scale factor. Combining these three calculated dimensionless numbers in a fixed order constitutes a three-dimensional dimensionless eigenvector. This vector eliminates the differences in physical dimensions and numerical scales of the original parameters, mapping the specific engineering problem to a low-dimensional characteristic space with inherent physical consistency. This process is a prerequisite for ensuring that the pre-trained model can correctly identify and handle the working condition.

[0212] The dimensionless feature vector is input into the pre-trained fusion evaluation model to obtain the final full-field displacement field prediction result. The fusion evaluation model is obtained according to the method of Example 1.

[0213] The pre-trained fusion evaluation model is a complex computational system with fixed parameters that has undergone complete multi-stage course training. This system integrates multiple mechanisms, including topology-aware classification, data-driven regression, and physics-driven regression. After inputting the dimensionless feature vectors constructed in the previous step, the model first implicitly performs rapid failure mode discrimination within its internal logic, and then activates its dual-stream regression component. The data expert module provides a global displacement benchmark, and the physics expert module provides a local correction field that satisfies physical constraints. These two are automatically fused within the model and subjected to hard boundary constraints.

[0214] The model performs a series of pre-defined, efficient forward calculations, ultimately outputting a complete full-field displacement field defined over the structural computational domain. This prediction result is a rapid and comprehensive mathematical deduction of the current working condition based on statistical regularities learned from historical data and differential constraints learned from physical laws. This displacement field prediction result is the direct output of the assessment and can be used to visualize deformation patterns, extract maximum deflection, assess stress distribution, or serve as a quantitative basis for further damage control decisions. The entire prediction process is extremely fast, achieving a second-level mapping from specific working condition parameters to high-fidelity physical field prediction results.

[0215] Example 3

[0216] Please see Figure 6 This invention provides a training device for a box-shaped structure implosion assessment model, comprising:

[0217] The acquisition module 601 is used to acquire the dimensionless feature vector sum, failure mode label and deflection label corresponding to the original working condition parameters of the explosion inside the ship's compartment; the original working condition parameters include the total explosion energy inside the compartment, the geometric dimensions of the dominant structure and the yield strength, and the failure mode label is used to characterize the failure status of the dominant structure.

[0218] The classification training module 602 is used to train a topology-aware classification network based on the dimensionless feature vector and the failure mode label.

[0219] The preheating module 603 is used to construct and preheat the dual-stream regression network, in which one stream learns a global displacement benchmark based on data fitting, and the other stream learns a local correction field based on physical equation constraints.

[0220] The inversion module 604 is used to apply boundary constraints to the fusion result of the global displacement reference and the local correction field, and then invert and optimize the adjustable parameters in the physical equation based on the constrained fusion result.

[0221] The fine-tuning module 605 is used to jointly fine-tune the classification network, the two-stream regression network, and the optimized adjustable parameters based on the deflection label to obtain the final evaluation model.

[0222] It should be noted that each module and unit in the box-type structure implosion assessment model training device in this embodiment corresponds one-to-one with each step in the box-type structure implosion assessment model training method in the aforementioned embodiment. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned box-type structure implosion assessment model training method, and will not be repeated here.

[0223] Example 4

[0224] Please see Figure 7 This embodiment provides an electronic device, including at least one processor 701 and a memory 702. Optionally, the device further includes a communication component 703. The processor 701, memory 702, and communication component 703 are connected via a bus 704.

[0225] In a specific implementation, at least one processor 701 executes computer execution instructions stored in memory 702, causing at least one processor 701 to perform the above-described method.

[0226] The specific implementation process of processor 701 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0227] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0228] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0229] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0230] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0231] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0232] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0233] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0234] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0235] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0236] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0237] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0238] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0239] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for an assessment model of an implosion in a box-shaped structure, characterized in that, include: Obtain the dimensionless feature vector, failure mode label, and deflection label corresponding to the original operating conditions parameters of the ship's internal explosion. The original operating condition parameters include the total explosive energy inside the hull, the geometric dimensions of the dominant structure, and the yield strength. The failure mode label is used to characterize the failure status of the dominant structure. A topology-aware classification network is trained based on the dimensionless feature vector and the failure mode label. A dual-stream regression network is constructed and preheated, in which one stream learns a global displacement benchmark based on data fitting, and the other stream learns a local correction field based on physical equation constraints. After applying boundary constraints to the fusion result of the global displacement reference and the local correction field, the adjustable parameters in the physical equation are inverted and optimized based on the constrained fusion result. The classification network, the two-stream regression network, and the optimized adjustable parameters are jointly fine-tuned based on the deflection labels to obtain the final evaluation model.

2. The method according to claim 1, characterized in that, The acquisition of the dimensionless feature vector corresponding to the original operating condition parameters of the ship's internal explosion includes: Calculate the dimensionless damage number based on the total explosive energy, plate thickness of the dominant structure, yield strength, and characteristic height from the original working condition parameters. : ， in, For total burst energy, For yield strength, The characteristic height is used to characterize the height dimension of a compartment or box-shaped structure in the direction perpendicular to the supporting structure being analyzed. For plate thickness; The geometric stiffness factor is obtained based on the ratio of the characteristic length of the dominant structure to the plate thickness. The feature length is used as the feature scale factor; The dimensionless damage number, geometric stiffness factor, and characteristic scale factor are used to construct a dimensionless eigenvector.

3. The method according to claim 2, characterized in that, The step of training a topology-aware classification network based on the dimensionless feature vector and the failure mode label includes: The dimensionless feature vector is input into a preset topology-aware classification subnet, which is configured to identify macroscopic failure modes of structural damage. The cross-entropy loss is calculated based on the failure mode probability vector output by the topology-aware classification subnet and the failure mode label. The classification network is obtained by optimizing the parameters of the topology-aware classification subnet through backpropagation to minimize the cross-entropy loss.

4. The method according to claim 3, characterized in that, The construction and preheating of the dual-stream regression network includes: The dimensionless feature vector is concatenated with the failure mode probability vector and then input into a preset data expert subnet. The data expert subnet is configured to learn a displacement magnitude benchmark determined by statistical laws. The mean square error loss is calculated by comparing the global displacement reference scalar value output by the data expert subnet with the full-field displacement data corresponding to the original working condition parameters. The full-field displacement data is used to characterize the displacement of the dominant structure. The parameters of the data expert subnet are optimized by backpropagation to minimize the mean square error loss, thus obtaining a preheated data expert module. The dimensionless feature vector is combined with the spatial coordinate grid defined on the structural computation domain and input into the physics expert subnetwork; The first physical equation residual is generated after substituting the local correction field output by the physical expert subnetwork into the preset rigid-plastic large deformation control equation. By optimizing the parameters of the physics expert subnetwork to minimize the residual of the first physics equation, a preheated physics expert module is obtained.

5. The method according to claim 4, characterized in that, The data expert subnet includes multiple fully connected layers connected in sequence, with the last connected layer being an output layer with a linear activation function, used to map the features learned by the network to scalar values ​​representing global displacement references; The physical expert subnet includes multiple sequentially connected residual blocks. Each residual block includes two fully connected layers and a skip connection, which is used to add the input of each residual block to the output of the second fully connected layer.

6. The method according to claim 4, characterized in that, After applying boundary constraints to the fusion result of the global displacement reference and the local correction field, the adjustable parameters in the physical equation are inverted and optimized based on the constrained fusion result, including: The global displacement reference field output by the data expert module and the local correction field output by the physics expert module are fused to form a preliminary predicted displacement field. Apply geometric hard boundary constraints to the preliminary predicted displacement field to obtain the constrained predicted displacement field; Set the equivalent energy transfer coefficient in the physical equation as a trainable parameter; Under the condition of fixing all network weights, the constrained predicted displacement field is substituted into the preset rigid-plastic large deformation control equation to calculate the residual of the second physical equation; The equivalent energy transfer coefficients are updated using the gradient descent algorithm to minimize the residuals of the second physical equation, thereby obtaining the optimal physical parameters.

7. The method according to claim 1, characterized in that, The step of jointly fine-tuning the classification network, the two-stream regression network, and the optimized adjustable parameters based on the deflection label to obtain the final evaluation model includes: Unfreeze all parameters of the classification network, the two-stream regression network, and the optimized adjustable parameters; A composite loss function is constructed, which is obtained by weighted summation of data mean square error loss and physical equation residual loss. Using the composite loss function as the overall objective, the entire network is optimized end-to-end based on the deflection label to obtain the final evaluation model.

8. A method for assessing internal explosion damage in ship compartments, characterized in that, include: Obtain the current operating parameters of the ship's compartments to be evaluated; Dimensional analysis is performed on the current working condition parameters to construct a dimensionless feature vector containing dimensionless damage number, geometric stiffness factor and characteristic scale factor; The dimensionless feature vector is input into the pre-trained fusion evaluation model to obtain the final full-field displacement field prediction result, wherein the fusion evaluation model is obtained according to any one of the methods described in claims 1-7.

9. A box-shaped implosion assessment model training device, characterized in that, include: The acquisition module is used to acquire the dimensionless feature vector, failure mode label, and deflection label corresponding to the original operating condition parameters of the ship's internal explosion. The original operating condition parameters include the total explosive energy inside the hull, the geometric dimensions of the dominant structure, and the yield strength. The failure mode label is used to characterize the failure status of the dominant structure. The classification training module is used to train a topology-aware classification network based on the dimensionless feature vector and the failure mode label. The preheating module is used to build and preheat the dual-stream regression network, in which one stream learns the global displacement benchmark based on data fitting, and the other stream learns the local correction field based on physical equation constraints. The inversion module is used to apply boundary constraints to the fusion result of the global displacement reference and the local correction field, and then invert and optimize the adjustable parameters in the physical equation based on the constrained fusion result. The fine-tuning module is used to jointly fine-tune the classification network, the two-stream regression network, and the optimized adjustable parameters based on the deflection labels to obtain the final evaluation model.

10. An electronic device, characterized in that, include: 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 as described in any one of claims 1-8.

Images