A rapid damage assessment method for ship structures under fire based on numerical simulation and MLP

Abstract

The present invention provides a method for rapidly assessing damage to ship structures under fire based on numerical simulation and MLP. The method includes: constructing a thermomechanical coupling model based on PyroSim and ANSYS software, comprehensively considering the dual effects of heat conduction and structural mechanics to perform a thermomechanical coupling numerical simulation of the ship under fire; employing a multi-layer perceptron model to perform deep learning training on the simulated data to construct a deep perceptron model to assess structural damage under ship fire; and designing GPU acceleration logic to optimize the computational efficiency of key steps through GPU acceleration, thereby achieving the goal of rapidly assessing structural damage under ship fire. The technical solution of the present invention not only significantly reduces computational costs but also provides efficient and accurate damage assessment results, providing new technical support and solutions for the research of ship structural safety under fire scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of ship structure damage assessment, and in particular to a method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP. Background Art

[0002] With the increasing number and complexity of shipping routes, structural damage assessment for ships in extreme scenarios such as fires has become a crucial research topic for ensuring safe ship operation, improving shipping efficiency, and reducing economic losses. Currently, there are three main types of techniques for assessing ship structural damage in fire scenarios: those based on full-scale ship testing, those based on scaled-down model testing, and those based on thermomechanical coupling analysis using numerical simulations. The full-scale ship testing approach suffers from high experimental costs and large errors due to the uneven distribution of detection data in the combustion environment. While the scaled-down model approach offers lower costs compared to full-scale ship testing, the combustion process at scale differs significantly from the actual large-scale combustion process physically, and this approach also cannot guarantee the accuracy of the prediction results.

[0003] In recent years, with the significant improvement of computer computing power and the continuous improvement of numerical solution methods, numerical simulation and rapid solution analysis have provided more new ideas for the assessment of ship structure damage under fire. However, the current assessment methods for ship structure damage still have the following problems: (1) At present, the deformation and damage that ships are prone to in a fire environment are very complex in their physical processes, mainly involving the dynamic changes of structural mechanical parameters caused by heat conduction and temperature changes during the fire. However, existing simulation methods are difficult to effectively handle these complex thermal coupling processes, making it difficult to accurately simulate ship structures under fire conditions. (2) Due to the large size and complex structure of ships, the thermal coupling process under fire has highly nonlinear characteristics. Relying solely on numerical simulation often requires a large amount of computing resources, which is difficult to meet the real-time requirements in practical applications.

[0004] At present, the numerical simulation methods used for damage assessment of ship structures under fire mainly rely on finite element analysis and fire dynamics simulation. These methods can more accurately simulate the evolution of temperature, stress and deformation fields during the fire process, providing an important basis for evaluating the mechanical properties of ship structures in high-temperature environments. However, since ship structures usually have large-scale and complex characteristics, and the thermal-mechanical coupling effects under fire involve the interaction of multiple physical fields, relying solely on numerical simulation for research often requires a large amount of computing resources, which is difficult to meet the real-time and high-efficiency requirements in practical applications. Especially when responding to sudden fire incidents, the computational efficiency of traditional methods is insufficient, and they cannot quickly provide effective damage assessment results, which in turn affects the timeliness of emergency decision-making. Summary of the Invention

[0005] To address the aforementioned technical issues, a method for rapid assessment of ship structure damage under fire, based on numerical simulation and MLP, is provided. This method overcomes the shortcomings of existing methods based on full-scale ship structure damage testing and numerical simulation testing of scaled-down models under fire. It also addresses the problem that existing numerical simulation-based methods fail to consider the effects of complex hull heat conduction and that serial calculations are too slow to meet practical engineering needs.

[0006] The technical means adopted in the present invention are as follows:

[0007] A rapid damage assessment method for ship structures under fire based on numerical simulation and MLP, including:

[0008] S1. Based on PyroSim and ANSYS software, a thermal-mechanical coupling model was constructed, which comprehensively considered the dual effects of heat conduction and structural mechanics, and numerically simulated the thermal-mechanical coupling of the ship under fire.

[0009] S2. Use a multi-layer perceptron model to perform deep learning training on simulated data and build a deep perceptron model to assess structural damage under the action of ship fire.

[0010] Furthermore, step S1 specifically includes:

[0011] S11. Constructing ship structure model:

[0012] According to the structural drawings and design requirements of the target ship, the geometric model of the ship cabin is created in PyroSim software. The constructed geometric model is meshed, and an appropriate mesh size is selected to balance calculation accuracy and efficiency. The minimum mesh unit should be larger than the thinnest wall thickness of the hull.

[0013] S12. Set thermal combustion parameters:

[0014] Determine the fire form, set the location, size and combustion characteristics of the fire source according to the fire form, set the physical properties of the ship's structural materials, the steel type of the ship's bulkheads, the shell properties of different parts, set the boundary conditions of the fire simulation, the initial temperature of the simulated environment, the atmospheric pressure and ventilation conditions;

[0015] S13, simulated fire values:

[0016] Select large eddy simulation and set the calculation time step and calculation accuracy to solve the physical model;

[0017] S14. Result output and analysis:

[0018] After the run is completed, the fire simulation result data is output, the simulation results are quantitatively analyzed, and all heat flux q data of all simulated positions within the set time are extracted and saved as a csv file;

[0019] S15. Initialization of ANSYS thermal-mechanical coupling simulation:

[0020] Start the Workbench tool in ANSYS software and establish the transient thermal analysis module and the transient structural analysis module. Use SolidWorks software to draw a ship model with the same size, coordinate system, and scale as PyroSim. Save the model in SAT file format and import it into the transient thermal analysis module. Set the relevant boundary conditions according to the fire environment simulated by PyroSim.

[0021] S16. Solve the thermal load:

[0022] Import the temperature field data pre-processed in step S14 into the transient thermal analysis module as the node temperature distribution input corresponding to the time step;

[0023] At the same time, the heat flux data is applied to the surface of the fire source area to simulate the actual effects of heat radiation and heat convection on the ship structure under fire conditions;

[0024] Configure the time step and time sequence of heat load application;

[0025] Run the transient thermal analysis module to solve the distribution and change law of the temperature field in the ship structure and obtain the temperature T distribution results during the entire fire process;

[0026] Ensure numerical convergence during the analysis process and verify through visualization tools whether the calculated temperature field results are consistent with the fire simulation expectations;

[0027] S17. Thermal coupling data transfer:

[0028] The results of transient thermal analysis, i.e. the calculated temperature distribution data of the entire ship, are transferred to the transient structural analysis module as thermal boundary condition input to simulate the impact of temperature changes on the thermal stress and thermal deformation of the ship structure.

[0029] S18. Solve the structural mechanics load:

[0030] Run the transient structural analysis module to calculate the thermal stress and thermal strain distribution of the ship structure under the thermal load conditions solved in step S16; extract the stress σ and strain ε of key nodes or components, and analyze whether there is a risk of material yield or fracture.

[0031] Furthermore, in step S15, all model edge lines are set to a fixed temperature to adapt to the integrity of the ship under real conditions and ensure that the simulation scenario conforms to the actual operating conditions of the ship.

[0032] Furthermore, step S2 specifically includes:

[0033] S21. Standardization of fire and mechanical data:

[0034] The temperature T and heat flux q in the fire scene, as well as the mechanical parameters stress σ and strain ε corresponding to the damage of the ship structure, are standardized and converted into a standard normal distribution with a mean of zero and a standard deviation of one to eliminate dimensional differences.

[0035] S22. Dataset division:

[0036] Divide the data set after the normalization process in step S21 into a training set, a validation set, and a test set according to the time series;

[0037] S23. Training model design and initialization:

[0038] A prediction model based on a multi-layer perceptron was designed. The network structure consists of an input layer, a hidden layer, and an output layer. The input layer has two nodes, corresponding to the input variables fire parameter temperature T and heat flux q, respectively. The hidden layer consists of two layers, the first layer has 64 neurons, and the second layer has 32 neurons. Both layers use ReLU as the activation function. The output layer has two nodes, corresponding to the predicted stress σ and strain ε, respectively.

[0039] The output of layer l is calculated as follows:

[0040] z (l) =W (l) a (l) +b (l)

[0041] a (l) =f(z (l) )

[0042] Among them, W (l) is the weight matrix, b (l) is the bias vector, a (l) =f(z (l) ) is the result of the activation function;

[0043] Calculate the final output as follows:

[0044]

[0045] Where L represents the index of the last layer;

[0046] S24. Loss function definition:

[0047] Define the mean square error as the loss function, the formula is as follows:

[0048]

[0049] Among them, N is the number of samples, that is, the number of training set samples, is the predicted value, Y i The loss function is used to measure the difference between the model prediction and the actual value.

[0050] S25. Optimizer selection and model training:

[0051] Set the initial learning rate to 0.001, select the Adam optimizer, and perform parameter optimization using the following formula:

[0052]

[0053] The training set is input into the prediction model designed in step S23 for forward propagation. After the prediction value is calculated, the error is evaluated using the loss function defined in step S24. Backward propagation is then performed to calculate the gradient, and the optimizer is used to update the weights and biases as follows:

[0054]

[0055] S26. Validation set evaluation:

[0056] After each round of training, the validation set divided in step S22 is input into the model to evaluate the model performance and record the validation error. If the validation error continues to increase, it is necessary to return to step S23 to adjust the model hyperparameters such as the number of hidden layers, the number of neurons, and the learning rate to ensure the optimal model configuration.

[0057] S27, test set performance test:

[0058] Input the test set in step S22 into the model to test its generalization ability on unseen data; verify the reliability of the final model evaluation function by calculating the prediction accuracy and error index. If the reliability verification fails, repeat step S22 until the reliability is verified;

[0059] S28. Model prediction application:

[0060] The temperature T and heat flux q of the target fire scene are input into the trained model in step S23, and the corresponding stress σ and strain ε are directly output.

[0061] Furthermore, step S21 specifically includes:

[0062] S211, the input of each sample is [T,q], for the input data X∈R 2N , N is the number of samples, and the standardization process is performed as follows:

[0063]

[0064] Among them, μ X is the mean of the input data, σ X is the standard deviation of the input data;

[0065] S212. The output data Y, namely stress σ and strain ε, are normalized using the following formula:

[0066]

[0067] Among them, μ Y is the mean of the output data, σ Y is the standard deviation of the output data.

[0068] Furthermore, step S22 specifically includes:

[0069] S221, assuming there are N samples in total, each sample corresponds to a time point; the data is from t1 to t N Permutation, {t1,t2,......,t N} is a time series, of which the training set accounts for 80%, the validation set and the test set each account for 10%, assuming:

[0070]

[0071] Among them, N train N is the number of training set samples, retaining the largest integer; val N is the number of samples in the validation set, retaining the largest integer; test is the number of test set samples, including the remaining data;

[0072] S222. Divide the data into three parts directly in chronological order, including:

[0073] Training set: starting from the first data t to the Nth train Data until;

[0074] Validation set: from the Nth train +1 data Start, to Nth train +N val Data until;

[0075] Test set: from the Nth train +N val +1 data Start from the beginning and end at the Nth data.

[0076] Furthermore, step S3 specifically includes:

[0077] S31. Design the kernel function PK1 for standardization of fire and mechanical data:

[0078] Assign a GPU thread to each data sample to calculate the mean and standard deviation of temperature T, heat flux q, stress σ, and strain ε, and perform normalization to convert them into a standard normal distribution;

[0079] S32. Design the time series data segmentation kernel function PK2:

[0080] The normalized dataset is divided into training, validation, and test sets by time series using a kernel function. Each data sample corresponds to a thread, which quickly completes data partitioning and allocates it to different memory areas.

[0081] S33. Design the forward propagation calculation kernel function PK3:

[0082] During the model training phase, a GPU thread is assigned to each hidden layer neuron. The forward propagation of temperature T and heat flux q is completed through parallel computing, including the weighted summation from the input layer to the hidden layer and the ReLU activation function calculation, ensuring that large-scale training data is efficiently input into the model.

[0083] S34. Design back propagation and gradient calculation kernel function PK4:

[0084] Used to calculate the gradient of the back-propagation phase, each weight parameter is assigned a GPU thread; by parallelizing the calculation of the gradient of the loss function with respect to weights and biases, the efficiency of back-propagation is significantly improved;

[0085] S35. Design optimizer updates kernel function PK5:

[0086] Parallelize the update logic of the Adam optimizer; assign a thread to each weight and bias parameter, and update the model parameters based on the calculated gradient; complete the optimizer's learning rate adjustment and momentum calculation through GPU acceleration;

[0087] S36. Design the verification set performance evaluation kernel function PK6:

[0088] After each round of training, a thread is assigned to each validation set sample to quickly complete the validation set error calculation; at the same time, the validation error trend is recorded for real-time adjustment of model hyperparameter configuration;

[0089] S37. Design the test set performance test kernel function PK7:

[0090] Compare the predictions of the test set with the true values ​​and calculate the error index; assign a thread to each test sample for parallel processing to evaluate the model's generalization ability on unseen data;

[0091] S38. Design model prediction application kernel function PK8:

[0092] The temperature T and heat flux q of the target fire scene are input into the trained model, and each input sample is assigned a thread to parallelly calculate the output stress σ and strain ε.

[0093] Furthermore, in step S3, the operation logic and sequence of each designed kernel function include:

[0094] S301, running the fire and mechanical data standardization processing kernel function PK1, the time series data segmentation kernel function PK2, and the forward propagation calculation kernel function PK3 to complete data standardization processing, data set division, and model initialization respectively;

[0095] S302, enter the time loop, run the back propagation and gradient calculation kernel function PK4 according to the input data and time step of the current fire scene to calculate the feature map of the input data and update the model state;

[0096] S303. Based on the model state at the current time step, the optimizer updates the kernel function PK5, performs forward propagation to calculate the predicted values ​​of the structural parameters in the fire scenario, and simultaneously runs the validation set performance evaluation kernel function PK6 to evaluate the prediction error.

[0097] S304: Based on the results of the optimizer update kernel function PK5 and the validation set performance evaluation kernel function PK6, the test set performance test kernel function PK7 and the model prediction application kernel function PK8 are sequentially run to complete the loss function calculation and parameter optimization respectively;

[0098] S305: Based on the results of the test set performance test kernel function PK7 and the model prediction application kernel function PK8, the performance of the model on the validation set is evaluated. At the same time, the damage data of the current time step is copied from the video memory to the internal memory and saved to the local computer for result analysis.

[0099] S306. After step S305, verify whether it is within the predetermined error range. If so, the calculation is terminated; if not, repeat steps S303-S305 until the preset time step or training termination condition is reached.

[0100] Compared with the prior art, the present invention has the following advantages:

[0101] This invention provides a rapid assessment method for ship structure damage under fire based on numerical simulation and MLP. By combining PyroSim and ANSYS software, a systematic thermal-mechanical coupling model is established, comprehensively considering the dual effects of heat conduction and structural mechanics. This model accurately simulates the heat conduction path and mechanical damage to ship structures under fire, ensuring the accuracy and reliability of fire simulations, effectively addressing the shortcomings of existing technologies in coupled structural damage analysis under fire.

[0102] 2. This invention provides a method for rapidly assessing damage to ship structures under fire based on numerical simulation and MLP. This method employs a multi-layer perceptron (MLP) model to perform deep learning training on simulated data, constructing a model capable of rapidly assessing damage to ship structures under fire. This trained MLP model boasts efficient computing power and high prediction accuracy. GPU acceleration further enhances computational efficiency, enabling the model to assess fire damage in real time in practical applications.

[0103] 3. The present invention provides a method for rapid damage assessment of ship structures under fire based on numerical simulation and MLP, which can not only significantly reduce the computational cost, but also provide efficient and accurate damage assessment results, providing new technical support and solutions for the research on ship structure safety under fire scenarios.

[0104] Based on the above reasons, the present invention can be widely promoted in fields such as ship structure damage assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0106] Figure 1 The present invention is a flow chart of the thermal-mechanical coupling numerical simulation of a ship under fire based on PyroSim and ANSYS.

[0107] Figure 2 This is the model training flow chart of the present invention.

[0108] Figure 3 This is the framework diagram of the deep perception machine kernel function of the present invention.

[0109] Figure 4 This is the operation logic diagram of each kernel function of the present invention. DETAILED DESCRIPTION

[0110] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0111] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0112] This method, which builds upon the existing approach of assessing ship structure fire damage solely through PyroSim fire simulation and ANSYS thermal coupling, innovatively combines a multi-layer perceptron with OpenCL-based accelerated computation technology to fully account for the impact of combustion on the ship's overall structure. This method effectively addresses the difficulty of traditional thermal coupling computation speeds in meeting engineering requirements. By integrating relevant detection data with model computation results, it enables rapid assessment of ship structure damage under fire.

[0113] The principle of numerical fire simulation on ships using PyroSim, based on its built-in large eddy simulation (LES) model and fire dynamics equations, aims to accurately predict key physical parameters during a fire, such as gas temperature and total combustion heat flux, under set conditions, and thus reproduce the actual fire results. By simulating physical phenomena such as thermal radiation, convective heat transfer, and combustion reactions during a fire, PyroSim can effectively reproduce the fire's propagation path and impact range within the ship's structure. Its core working mechanism includes the following steps: First, the initial computational domain for the fire numerical simulation is established by modeling and meshing the ship's cabin geometry; second, the fire model is driven by input parameters such as combustion source characteristics, ventilation conditions, and material properties; finally, the fire dynamics equations are iteratively solved to output the spatiotemporal distribution of key physical quantities during the fire, such as temperature fields, smoke concentrations, and heat fluxes. This simulation method provides reliable technical support for studying the thermodynamic effects of fire on ship structures, with high engineering applicability and analytical accuracy. (The thermal-mechanical coupling between ANSYS and PyroSim in this invention focuses only on the data generated by the model numerical simulation, not the software's principles.)

[0114] The present invention provides a method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP, comprising:

[0115] S1. Based on PyroSim and ANSYS software, a thermal-mechanical coupling model was constructed, which comprehensively considered the dual effects of heat conduction and structural mechanics, and numerically simulated the thermal-mechanical coupling of the ship under fire.

[0116] S2. Use a multi-layer perceptron (MLP) model to perform deep learning training on simulated data and build a deep perceptron model to assess structural damage under ship fire.

[0117] S3. Design GPU acceleration logic to optimize the computational efficiency of key steps through GPU acceleration to achieve the goal of rapid assessment of structural damage under ship fire.

[0118] When specifically implemented, as a preferred embodiment of the present invention, Figure 1 As shown, step S1 specifically includes:

[0119] S11. Constructing ship structure model:

[0120] Based on the target ship's structural drawings and design requirements, the ship's cabin geometry was created in PyroSim. Key components, including cabin walls, doors, windows, vents, and other key structural components, must be accurately reproduced. The constructed geometry was then meshed, with an appropriate mesh size chosen to balance accuracy and efficiency. The minimum mesh element must be larger than the thinnest wall thickness of the hull.

[0121] S12. Set thermal combustion parameters:

[0122] Determine the fire type and, based on the fire type, set the source location, size, and combustion characteristics, such as fuel type, maximum heat flux, and mass loss rate. Common fires during navigation are caused by leaks of fuels such as gasoline and natural gas. Also, set the physical properties of the ship's structural materials, the steel type of the bulkhead, and the hull properties of different locations, such as thermal conductivity, specific heat capacity, density, and flash point. Set the boundary conditions for the fire simulation, the initial temperature of the simulated environment, atmospheric pressure, and ventilation conditions, such as natural or mechanical ventilation.

[0123] S13, simulated fire values:

[0124] Select Large Eddy Simulation (LES) and set the calculation time step and calculation accuracy to solve the physical model to ensure the numerical stability of the simulation and the accuracy of the results;

[0125] S14. Result output and analysis:

[0126] After the run is completed, the fire simulation result data is output, including the temperature field, heat flux distribution and fire propagation path, etc., to conduct quantitative analysis of the simulation results. The tool fds2ascii provided in the FDS program is used to extract all the heat flux q data of all simulated locations within the set time and save it as a csv file for subsequent processing;

[0127] S15. Initialization of ANSYS thermal-mechanical coupling simulation:

[0128] Launch the Workbench tool in ANSYS software and establish the Transient Thermal Analysis module (Transient Thermal) and the Transient Structural Analysis module (Transient Structural). Use SolidWorks software to draw a ship model with the same size, coordinate system, and scale as PyroSim. Save the model as a SAT file and import it into the Transient Thermal Analysis module. Set relevant boundary conditions based on the fire environment simulated by PyroSim.

[0129] S16. Solve the thermal load:

[0130] Import the temperature field data pre-processed in step S14 into the transient thermal analysis module as the node temperature distribution input corresponding to the time step;

[0131] At the same time, the heat flux data is applied to the surface of the fire source area to simulate the actual effects of heat radiation and heat convection on the ship structure under fire conditions;

[0132] Configure the time step and the time series of heat load application to accurately restore the dynamic heat transfer process during the fire;

[0133] Run the transient thermal analysis module to solve the distribution and change law of the temperature field in the ship structure and obtain the temperature T distribution results during the entire fire process;

[0134] Ensure numerical convergence during the analysis process and verify through visualization tools whether the calculated temperature field results are consistent with the fire simulation expectations;

[0135] S17. Thermal coupling data transfer:

[0136] The results of transient thermal analysis, i.e. the calculated temperature distribution data of the entire ship, are transferred to the transient structural analysis module as thermal boundary condition input to simulate the impact of temperature changes on the thermal stress and thermal deformation of the ship structure.

[0137] S18. Solve the structural mechanics load:

[0138] Run the transient structural analysis module to calculate the thermal stress and thermal strain distribution of the ship structure under the thermal load conditions solved in step S16; extract the stress σ and strain ε of key nodes or components, and analyze whether there is a risk of material yield or fracture.

[0139] In specific implementation, as a preferred embodiment of the present invention, in step S15, all model edge lines are set to a fixed temperature to adapt to the integrity of the ship under real conditions and ensure that the simulation scene conforms to the actual operating conditions of the ship.

[0140] When specifically implemented, as a preferred embodiment of the present invention, Figure 2 As shown, step S2 specifically includes:

[0141] S21. Standardization of fire and mechanical data:

[0142] The temperature T and heat flux q in the fire scene, as well as the mechanical parameters stress σ and strain ε corresponding to the damage of the ship structure, are standardized and converted into a standard normal distribution with a mean of zero and a standard deviation of one to eliminate dimensional differences.

[0143] S22. Dataset division:

[0144] Divide the data set after the normalization process in step S21 into a training set, a validation set, and a test set according to the time series;

[0145] S23. Training model design and initialization:

[0146] A prediction model based on a multi-layer perceptron (MLP) was designed. The network structure consists of an input layer, a hidden layer, and an output layer. The input layer has two nodes, corresponding to the input variables fire parameter temperature T and heat flux q, respectively. The hidden layer consists of two layers, the first layer has 64 neurons, and the second layer has 32 neurons. Both layers use ReLU as the activation function. The output layer has two nodes, corresponding to the predicted stress σ and strain ε, respectively.

[0147] The output of layer l is calculated as follows:

[0148] z (l) =W (l) a (l) +b (l)

[0149] a (l) =f(z (l) )

[0150] Among them, W (l) is the weight matrix, b (l) is the bias vector, a (l) =f(z (l) ) is the result of the activation function;

[0151] Calculate the final output as follows:

[0152]

[0153] Where L represents the index of the last layer;

[0154] S24. Loss function definition:

[0155] Define the mean square error (MSE) as the loss function, the formula is as follows:

[0156]

[0157] Among them, N is the number of samples, that is, the number of training set samples, is the predicted value, Y i The loss function is used to measure the difference between the model prediction and the actual value.

[0158] S25. Optimizer selection and model training:

[0159] Set the initial learning rate to 0.001, select the Adam optimizer, and perform parameter optimization using the following formula:

[0160]

[0161] The training set is input into the prediction model designed in step S23 for forward propagation. After the prediction value is calculated, the error is evaluated using the loss function defined in step S24. Backward propagation is then performed to calculate the gradient, and the optimizer is used to update the weights and biases as follows:

[0162]

[0163] S26. Validation set evaluation:

[0164] After each round of training, the validation set divided in step S22 is input into the model to evaluate the model performance and record the validation error. If the validation error continues to increase, it is necessary to return to step S23 to adjust the model hyperparameters such as the number of hidden layers, the number of neurons, and the learning rate to ensure the optimal model configuration.

[0165] S27, test set performance test:

[0166] Input the test set in step S22 into the model to test its generalization ability on unseen data; verify the reliability of the final model evaluation function by calculating the prediction accuracy and error index. If the reliability verification fails, repeat step S22 until the reliability is verified;

[0167] S28. Model prediction application:

[0168] The temperature T and heat flux q of the target fire scene are input into the trained model in step S23, and the corresponding stress σ and strain ε are directly output.

[0169] In specific implementation, as a preferred embodiment of the present invention, step S21 specifically includes:

[0170] S211, the input of each sample is [T,q], for the input data X∈R 2N , N is the number of samples, and the standardization process is performed as follows:

[0171]

[0172] Among them, μ X is the mean of the input data, σ X is the standard deviation of the input data;

[0173] S212. The output data Y, namely stress σ and strain ε, are normalized using the following formula:

[0174]

[0175] Among them, μ Y is the mean of the output data, σ YThe purpose of this step is to make the model learn the output more efficiently and avoid deviations during training.

[0176] In specific implementation, as a preferred embodiment of the present invention, step S22 specifically includes:

[0177] S221, assuming there are N samples in total, each sample corresponds to a time point; the data is from t1 to t N Permutation, {t1,t2,......,t N} is a time series, of which the training set accounts for 80%, the validation set and the test set each account for 10%, assuming:

[0178]

[0179] Among them, N train N is the number of training set samples, retaining the largest integer; val N is the number of samples in the validation set, retaining the largest integer; test is the number of test set samples, including the remaining data;

[0180] S222. Divide the data into three parts directly in chronological order, including:

[0181] Training set: starting from the first data t to the Nth train Data until;

[0182] Validation set: from the Nth train +1 data Start, to Nth train +N val Data until;

[0183] Test set: from the Nth train +N val +1 data Start from the beginning and end at the Nth data.

[0184] When specifically implemented, as a preferred embodiment of the present invention, Figure 3 As shown, step S3 specifically includes:

[0185] S31. Design the kernel function PK1 for standardization of fire and mechanical data:

[0186] A GPU thread is assigned to each data sample to calculate the mean and standard deviation of temperature T, heat flux q, stress σ, and strain ε, and then normalize them to convert them into a standard normal distribution. Parallel computing can significantly improve the efficiency of standardization for large-scale sample sets.

[0187] S32. Design the time series data segmentation kernel function PK2:

[0188] The normalized dataset is divided into training, validation, and test sets by time series using a kernel function. Each data sample corresponds to a thread, which quickly completes data partitioning and allocates it to different memory areas.

[0189] S33. Design the forward propagation calculation kernel function PK3:

[0190] During the model training phase, a GPU thread is assigned to each hidden layer neuron. The forward propagation of temperature T and heat flux q is completed through parallel computing, including the weighted summation from the input layer to the hidden layer and the ReLU activation function calculation, ensuring that large-scale training data is efficiently input into the model.

[0191] S34. Design back propagation and gradient calculation kernel function PK4:

[0192] Used to calculate the gradient of the back-propagation phase, each weight parameter is assigned a GPU thread; by parallelizing the calculation of the gradient of the loss function with respect to weights and biases, the efficiency of back-propagation is significantly improved;

[0193] S35. Design optimizer updates kernel function PK5:

[0194] Parallelize the update logic of the Adam optimizer; assign a thread to each weight and bias parameter, and update the model parameters based on the calculated gradient; complete the optimizer's learning rate adjustment and momentum calculation through GPU acceleration;

[0195] S36. Design the verification set performance evaluation kernel function PK6:

[0196] After each round of training, a thread is assigned to each validation set sample to quickly complete the validation set error calculation; at the same time, the validation error trend is recorded for real-time adjustment of model hyperparameter configuration;

[0197] S37. Design the test set performance test kernel function PK7:

[0198] Compare the predictions of the test set with the true values ​​and calculate the error metric (such as MSE). Each test sample is assigned a thread for parallel processing to evaluate the model's generalization ability on unseen data.

[0199] S38. Design model prediction application kernel function PK8:

[0200] The temperature T and heat flux q of the target fire scene are input into the trained model, and each input sample is assigned a thread to parallelly calculate the output stress σ and strain ε.

[0201] Through the design and implementation of the above kernel function, the computational efficiency of the fire structure damage assessment process can be significantly improved, meeting the needs of rapid assessment, and providing an efficient solution for the safety assessment of complex ship structures.

[0202] When specifically implemented, as a preferred embodiment of the present invention, Figure 4 As shown, in step S3, the operation logic and sequence of each designed kernel function include:

[0203] S301, running the fire and mechanical data standardization processing kernel function PK1, the time series data segmentation kernel function PK2, and the forward propagation calculation kernel function PK3 to complete data standardization processing, data set division, and model initialization respectively;

[0204] S302, enter the time loop, run the back propagation and gradient calculation kernel function PK4 according to the input data and time step of the current fire scene to calculate the feature map of the input data and update the model state;

[0205] S303. Based on the model state at the current time step, the optimizer updates the kernel function PK5, performs forward propagation to calculate the predicted values ​​of the structural parameters in the fire scenario, and simultaneously runs the validation set performance evaluation kernel function PK6 to evaluate the prediction error.

[0206] S304: Based on the results of the optimizer update kernel function PK5 and the validation set performance evaluation kernel function PK6, the test set performance test kernel function PK7 and the model prediction application kernel function PK8 are sequentially run to complete the loss function calculation and parameter optimization respectively;

[0207] S305: Based on the results of the test set performance test kernel function PK7 and the model prediction application kernel function PK8, the performance of the model on the validation set is evaluated. At the same time, the damage data of the current time step is copied from the video memory to the internal memory and saved to the local computer for result analysis.

[0208] S306. After step S305, verify whether it is within the predetermined error range. If so, the calculation is terminated; if not, repeat steps S303-S305 until the preset time step or training termination condition is reached.

[0209] Through the above technical route, using deep learning models and GPU accelerated design, we can efficiently complete the rapid assessment of ship structure damage under the action of fire, and provide support for subsequent structural safety analysis.

[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP, characterized by: include: S1. Based on PyroSim and ANSYS software, a thermal-mechanical coupling model was constructed, which comprehensively considered the dual effects of heat conduction and structural mechanics, and numerically simulated the thermal-mechanical coupling of the ship under fire. S2. Use a multi-layer perceptron model to conduct deep learning training on simulated data and build a deep perceptron model to assess structural damage under ship fire. S21. Standardization of fire and mechanical data: Temperature in fire scene and heat flux , and the mechanical parameter stress corresponding to the damage of the ship structure and strain Perform standardization processing to convert it into a standard normal distribution with a mean of zero and a standard deviation of one to eliminate dimensional differences; S22. Dataset division: Divide the data set after the normalization process in step S21 into a training set, a validation set, and a test set according to the time series; S23. Training model design and initialization: The prediction model based on the multi-layer perceptron is designed. The network structure includes input layer, hidden layer and output layer. The input layer has two nodes, corresponding to the input variables fire parameter temperature and heat flux The hidden layer consists of two layers. The first layer has 64 neurons and the second layer has 32 neurons. Both layers use ReLU as the activation function. The output layer has 2 nodes, corresponding to the predicted stress and strain ; Calculate the The output of the layer is as follows: in, is the weight matrix, is the bias vector, is the result of the activation function; Calculate the final output as follows: in, Indicates the index of the last layer; S24. Loss function definition: Define the mean square error as the loss function, the formula is as follows: in, is the number of samples, i.e. the number of samples in the training set, is the predicted value, is the actual value, and the loss function is used to measure the gap between the model prediction and the actual value; S25. Optimizer selection and model training: Set the initial learning rate to 0.001, select the Adam optimizer, and perform parameter optimization using the following formula: The training set is input into the prediction model designed in step S23 for forward propagation. After the prediction value is calculated, the error is evaluated using the loss function defined in step S24. Backward propagation is then performed to calculate the gradient, and the optimizer is used to update the weights and biases as follows: S26. Validation set evaluation: After each round of training, the validation set divided in step S22 is input into the model to evaluate the performance of the model and record the validation error. If the validation error continues to increase, it is necessary to return to step S23 to adjust the model hyperparameters such as the number of hidden layers, the number of neurons, and the learning rate. S27, test set performance test: Input the test set in step S22 into the model to test its generalization ability on unseen data; verify the reliability of the final model evaluation function by calculating the prediction accuracy and error index. If the reliability verification fails, repeat step S22 until the reliability is verified; S28. Model prediction application: The temperature of the target fire scene and heat flux Input into the model trained in step S23 and directly output the corresponding stress and strain ; S3. Design GPU acceleration logic to optimize the computational efficiency of key steps through GPU acceleration to achieve the goal of rapid assessment of structural damage under ship fire.

2. A method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP according to claim 1, characterized in that: Step S1 specifically includes: S11. Build ship structure model: According to the structural drawings and design requirements of the target ship, the geometric model of the ship cabin is created in PyroSim software. The constructed geometric model is meshed, and an appropriate mesh size is selected to balance calculation accuracy and efficiency. The minimum mesh unit should be larger than the thinnest wall thickness of the hull. S12. Set thermal combustion parameters: Determine the fire form, set the location, size and combustion characteristics of the fire source according to the fire form, set the physical properties of the ship's structural materials, the steel type of the ship's bulkheads, the shell properties of different parts, set the boundary conditions of the fire simulation, the initial temperature of the simulated environment, the atmospheric pressure and ventilation conditions; S13, simulated fire values: Select large eddy simulation and set the calculation time step and calculation accuracy to solve the physical model; S14. Result output and analysis: After the operation is completed, the fire simulation result data is output, the simulation results are quantitatively analyzed, and the total heat flux of all simulated positions within the set time is extracted. Data and save it as a csv file; S15. Initialization of ANSYS thermal-mechanical coupling simulation: Start the Workbench tool in ANSYS software and establish the transient thermal analysis module and the transient structural analysis module. Use SolidWorks software to draw a ship model with the same size, coordinate system, and scale as PyroSim. Save the model in SAT file format and import it into the transient thermal analysis module. Set the relevant boundary conditions according to the fire environment simulated by PyroSim. S16. Solve the thermal load: Import the temperature field data pre-processed in step S14 into the transient thermal analysis module as the node temperature distribution input corresponding to the time step; At the same time, the heat flux data is applied to the surface of the fire source area to simulate the actual effects of heat radiation and heat convection on the ship structure under fire conditions; Configure the time step and time sequence of heat load application; Run the transient thermal analysis module to solve the distribution and change law of the temperature field in the ship structure and obtain the temperature during the entire fire process Distribution results; Ensure numerical convergence during the analysis process and verify through visualization tools whether the calculated temperature field results are consistent with the fire simulation expectations; S17. Thermal coupling data transfer: The results of transient thermal analysis, i.e. the calculated temperature distribution data of the entire ship, are transferred to the transient structural analysis module as thermal boundary condition input to simulate the impact of temperature changes on the thermal stress and thermal deformation of the ship structure. S18. Solve the structural mechanics load: Run the transient structural analysis module to calculate the thermal stress and thermal strain distribution of the ship structure under the thermal load conditions solved in step S16; extract the stress of key nodes or components and strain , and analyze whether there is a risk of material yielding or fracture.

3. The method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP according to claim 2 is characterized in that: In step S15, all model edge lines are set to a fixed temperature to adapt to the integrity of the ship under real conditions and ensure that the simulation scenario conforms to the actual operating conditions of the ship.

4. The method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP according to claim 1 is characterized in that: Step S21 specifically includes: S211, the input of each sample is , for the input data , is the number of samples, and the normalization process is as follows: in, is the mean of the input data, is the standard deviation of the input data; S212: Output data , that is, stress and strain , perform standardization, the formula is as follows: in, is the mean of the output data, is the standard deviation of the output data.

5. The method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP according to claim 1 is characterized in that: Step S22 specifically includes: S221, assuming there are samples, each sample corresponds to a time point; data is sorted by time from arrive arrangement, is a time series, of which the training set accounts for 80%, the validation set and the test set each account for 10%. Let: in, is the number of training set samples, retaining the largest integer; The number of samples in the validation set is the maximum integer. is the number of test set samples, including the remaining data; S222. Divide the data into three parts directly in chronological order, including: Training set: from the first data Start to Data until; Validation set: From Data Start to Data until; Test set: From Data Start to Until the data.

6. The method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP according to claim 1 is characterized in that: Step S3 specifically includes: S31. Design the fire and mechanical data standardization kernel function PK1: Assign a GPU thread to each data sample to calculate the temperature , heat flux ,stress and strain The mean and standard deviation of are normalized and converted into standard normal distribution; S32. Design the time series data segmentation kernel function PK2: The normalized dataset is divided into training, validation, and test sets by time series using a kernel function. Each data sample corresponds to a thread, which quickly completes data partitioning and allocates it to different memory areas. S33. Design the forward propagation calculation kernel function PK3: During the model training phase, a GPU thread is assigned to each hidden layer neuron; temperature is calculated in parallel. and heat flux The forward propagation, including the weighted summation from the input layer to the hidden layer and the ReLU activation function calculation, ensures that large-scale training data can be efficiently input into the model; S34. Design back propagation and gradient calculation kernel function PK4: Used to calculate the gradient of the back-propagation phase, each weight parameter is assigned a GPU thread; by parallelizing the calculation of the gradient of the loss function with respect to weights and biases, the efficiency of back-propagation is significantly improved; S35. Design optimizer updates kernel function PK5: Parallelize the update logic of the Adam optimizer; assign a thread to each weight and bias parameter, and update the model parameters based on the calculated gradient; complete the optimizer's learning rate adjustment and momentum calculation through GPU acceleration; S36. Design the verification set performance evaluation kernel function PK6: After each round of training, a thread is assigned to each validation set sample to quickly complete the validation set error calculation; at the same time, the validation error trend is recorded for real-time adjustment of model hyperparameter configuration; S37. Design the test set performance test kernel function PK7: Compare the predictions of the test set with the true values ​​and calculate the error index; assign a thread to each test sample for parallel processing to evaluate the model's generalization ability on unseen data; S38. Design model prediction application kernel function PK8: The temperature of the target fire scene and heat flux Input the trained model, and assign one thread to each input sample to parallelly calculate the output stress and strain .

7. A method for rapid assessment of ship structure damage under fire based on numerical simulation and MLP according to claim 6, characterized in that: In step S3, the operation logic and sequence of each designed kernel function include: S301, running the fire and mechanical data standardization processing kernel function PK1, the time series data segmentation kernel function PK2, and the forward propagation calculation kernel function PK3 to complete data standardization processing, data set division, and model initialization respectively; S302, enter the time loop, run the back propagation and gradient calculation kernel function PK4 according to the input data and time step of the current fire scene to calculate the feature map of the input data and update the model state; S303. Based on the model state at the current time step, the optimizer updates the kernel function PK5, performs forward propagation to calculate the predicted values ​​of the structural parameters in the fire scenario, and simultaneously runs the validation set performance evaluation kernel function PK6 to evaluate the prediction error. S304: Based on the results of the optimizer update kernel function PK5 and the validation set performance evaluation kernel function PK6, the test set performance test kernel function PK7 and the model prediction application kernel function PK8 are sequentially run to complete the loss function calculation and parameter optimization respectively; S305: Based on the results of the test set performance test kernel function PK7 and the model prediction application kernel function PK8, the performance of the model on the validation set is evaluated. At the same time, the damage data of the current time step is copied from the video memory to the internal memory and saved to the local computer for result analysis. S306. After step S305, verify whether it is within the predetermined error range. If so, terminate the calculation; if not, repeat steps S303-S305 until the preset time step or training termination condition is reached.

Images