Deepwater explosion container operation and maintenance scheme generation method based on generative adversarial network

Abstract

This invention discloses a method for generating operation and maintenance schemes for deep-water explosive containers based on generative adversarial networks (CGANs), relating to the field of deep-water explosive container safety operation and maintenance technology. Based on the maximum strain prediction value output by the optimal GRU strain prediction model, the fuzziness of material performance parameters is first processed using a fuzzy interval reliability model to obtain a reliability interval. Then, using this reliability interval as input, a random interval hybrid reliability model is combined to fuse the randomness of the predicted strain, calculating a reliability index η used to quantify the container's safety status. This invention generates virtual data that conforms to the real distribution and has physical rationality through conditional generative adversarial networks (CGANs), effectively compensating for the lack of historical data under extreme conditions such as high explosive dosage and high water pressure, enhancing the generalization ability of data-driven models, and significantly improving strain prediction accuracy in scarce sample intervals.

Description

Technical Field

[0001] This invention relates to the field of deep-sea explosive container safety operation and maintenance technology, and in particular to a method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks. Background Technology

[0002] As the core equipment for underwater explosion tests, deep-water explosion containers play a crucial role in simulating deep-water explosion impacts, evaluating weapon damage effectiveness, and verifying the blast resistance of underwater structures. Their operating environment is characterized by high hydrostatic pressure, strong impact loads, and multi-parameter coupling. During long-term use, the container wall will experience strain growth due to fatigue accumulation. If timely and targeted maintenance measures are not taken, structural failure or even safety accidents may occur. Therefore, accurate monitoring of its operating status and the scientific development of maintenance plans are of great significance.

[0003] Currently, the operation and maintenance of deep-water explosion containers mainly rely on traditional empirical methods or statistical analysis based on limited experimental data, which has the following significant limitations: 1. Data scarcity limits prediction accuracy: Experiments under extreme conditions such as high charge and high water pressure are costly and risky, resulting in a severe lack of historical data for such conditions. Conventional data-driven models show significantly increased prediction errors in scarce sample intervals, making it difficult to support reliable maintenance decisions. 2. Disconnect between physical constraints and data modeling: Existing data augmentation methods mostly rely on pure mathematical interpolation or statistical fitting, failing to fully incorporate the positive correlation between strain and charge / water pressure, as well as physical laws such as fatigue cumulative effects, resulting in virtual... The simulated data may not reflect the actual engineering situation, leading to a decrease in the model's generalization ability; third, the reliability analysis method is singular: traditional reliability assessments mostly use deterministic or purely stochastic models, which make it difficult to simultaneously quantify the fuzziness of material performance parameters (such as the range uncertainty of elastic modulus) and the randomness of load parameters (such as strain fluctuations), resulting in biases in the determination of safety status and a lack of accurate basis for maintenance level classification; fourth, the maintenance plan lacks dynamism: existing strategies are mostly based on fixed thresholds and do not dynamically adjust in conjunction with the real-time operating parameters of the container and the predicted reliability indicators, making it difficult to adapt to the maintenance needs under different working conditions, which may lead to over-maintenance or under-maintenance problems.

[0004] In recent years, the application of Generative Adversarial Networks (GANs) in data augmentation has provided a new approach to solving the problem of sample scarcity. Through adversarial learning between generators and discriminators, GANs can generate virtual data that conforms to the real distribution. Deep learning models (such as GRUs) have shown advantages in dynamic parameter prediction due to their ability to capture temporal dependencies. The fusion of multi-source reliability analysis methods has made it possible to quantify the uncertainty of complex parameters. However, how to organically combine these technologies to construct an integrated solution that balances physical rationality, prediction accuracy, and maintenance dynamism remains a key problem that urgently needs to be solved in the field of deep-water explosive container operation and maintenance. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks. The technical solution adopted is as follows: A method for generating operation and maintenance solutions for deep-sea explosive containers based on generative adversarial networks includes the following steps: Step 1: Collect historical operational data of deep-water explosion containers, and sequentially perform outlier removal, missing value supplementation, and standardization processing. Combined with correlation analysis, clarify the physical constraint relationship between the dosage, hydrostatic pressure, number of tests, and maximum strain, and construct a structured dataset. Step 2: Construct a conditional generative adversarial network (CGAN) with embedded physical constraint modules. CGAN includes a generator and a discriminator. The generator generates virtual strain data that conforms to physical laws by taking random noise and working condition parameters as input. The discriminator judges the authenticity of the input real or virtual strain data and the corresponding working condition parameters. Step 3: Train CGAN using the WGAN algorithm, and merge virtual data that meets the quality requirements with real data to form an expanded training dataset; Step 4: Construct a gated recurrent neural network strain prediction model GRU based on the expanded training dataset, optimize the model parameters through cross-validation, evaluate the model performance using mean absolute percentage error and coefficient of determination, and determine the optimal GRU strain prediction model. Step 5: Based on the maximum strain prediction value output by the optimal GRU strain prediction model, the fuzziness of the material performance parameters is first processed by the fuzzy interval reliability model to obtain the reliability interval. Then, the reliability interval is used as input and combined with the random interval hybrid reliability model to fuse the randomness of the predicted strain, and the reliability index η used to quantify the safety status of the container is calculated. Step 6: Based on the reliability index η, the operating status of the deep-water explosion container is divided into three maintenance levels: early warning level, attention level, and emergency level. Correspondingly, routine monitoring strategies, fatigue mitigation strategies, and emergency repair strategies are generated to form a dynamic maintenance plan.

[0006] Optionally, the historical operational data of the deep-water explosion vessel in step 1 includes historical data on explosive charge, hydrostatic pressure, number of tests, and corresponding maximum strain.

[0007] Optionally, in step 1, outliers in the historical data are removed using box plots, and missing strain data are supplemented using cubic spline interpolation, where the interpolation process follows the physical laws of fatigue accumulation; the processed data are standardized, and the physical constraint relationship between drug dosage, water pressure, number of tests and maximum strain is clarified through correlation analysis and visualization methods, thereby constructing a structured dataset.

[0008] Optionally, in step 2, the scarce working condition samples are given a higher weight. The scarce working condition is a working condition with high dosage and high water pressure. Specifically, the working condition with high dosage and high water pressure is a working condition with a dosage greater than 80% of the maximum dosage in historical data and a static water pressure greater than 80% of the maximum static water pressure in historical data. The discriminator assigns a weight to the scarce working condition samples that is twice the weight of the regular working condition samples.

[0009] Optionally, in step 2, the physical constraint module constrains physical laws by adding a penalty term to the loss function of the generator.

[0010] Optionally, in step 4, the GRU model parameters are optimized through cross-validation, and the mean absolute error (MAPE) and coefficient of determination are used. As a performance evaluation metric, cross-validation is k-fold cross-validation, with k ranging from 5 to 10. The GRU model parameters to be optimized include the number of hidden layer nodes, learning rate, and number of iterations. The optimization objective is to minimize the MAPE of the test set.

[0011] Optionally, the performance evaluation metrics in step 4 require that the mean absolute percentage error (MAPE) of the test set be less than 8% and R0 be less than 8%. 2 Greater than 0.9.

[0012] Optionally, the material property parameters in step 5 include the elastic modulus E and the yield strength. In the fuzzy interval reliability model, triangular fuzzy functions or trapezoidal fuzzy functions are used to describe the uncertainty of material performance parameters.

[0013] Optionally, the reliability index thresholds η corresponding to the three maintenance levels in step 6 are as follows: η≥3.0 is the early warning level, corresponding to the conventional monitoring strategy, which records strain data every 10 tests; 1.5≤η<3.0 is the attention level, corresponding to the fatigue mitigation strategy, which is achieved by reducing the dosage or hydrostatic pressure and shortening the monitoring cycle to every 5 tests; η<1.5 is the emergency level, corresponding to the emergency repair strategy, which involves immediately shutting down the machine for repair and replacing fatigue-damaged components.

[0014] Optionally, the formula for calculating the reliability index η is: ; in The mean and standard deviation of the yield strength. These are statistical parameters of the load stress.

[0015] In summary, the present invention includes at least one of the following beneficial technical effects: by generating virtual data that conforms to the real distribution and has physical rationality through conditional generative adversarial networks (CGAN), it effectively makes up for the lack of historical data under extreme working conditions such as high drug dosage and high water pressure, enhances the generalization ability of data-driven models, and significantly improves the strain prediction accuracy in scarce sample intervals. By embedding a physical constraint module into the generator loss function of CGAN, the generated data is forced to satisfy the positive correlation between strain and drug dosage and water pressure, as well as the fatigue accumulation effect through a penalty term. This avoids the engineering violation problems that may be caused by traditional pure mathematical interpolation or statistical augmentation methods, and improves the credibility and usability of the generated data. By combining the fuzzy interval reliability model and the stochastic-interval hybrid reliability model, the fuzziness of material performance parameters (such as elastic modulus and yield strength) and the randomness of load parameters are handled respectively. The safety status of the container is comprehensively quantified, overcoming the limitations of the traditional single reliability model in modeling complex uncertainties and improving the accuracy and comprehensiveness of safety threshold determination. Based on real-time predicted strain and reliability index η, the maintenance levels are dynamically divided into three levels: early warning, attention, and emergency, and corresponding strategies (routine monitoring, parameter adjustment, and shutdown maintenance) are generated. This realizes the transformation from fixed threshold maintenance to state-driven dynamic maintenance, which avoids the cost waste caused by over-maintenance and prevents the operational risks caused by insufficient maintenance. By collaboratively designing the entire process from data preprocessing, virtual enhancement, predictive modeling, reliability analysis to maintenance generation, an integrated intelligent operation and maintenance solution is formed, which significantly improves the operational safety and maintenance targeting of deep-water explosion containers under complex high-pressure and impact load environments, and has important engineering application value. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks. Figure 2 This is a schematic diagram of data collection and preprocessing in step 1; Figure 3 This is a schematic diagram of virtual data generation based on CGAN in step 2; Figure 4 This is a schematic diagram of the construction and optimization of the GRU strain prediction model based on fused data in step 3; Figure 5 This is a schematic diagram of the reliability analysis of the deep-water explosion container based on the dual model in step 4; Figure 6 This is a schematic diagram of the deep-water explosion container operation and maintenance plan generated in step 5; Figure 7 This is a schematic diagram showing the arrangement of the strain and acceleration sensors; Figure 8These are the strain and acceleration waveforms obtained from explosion experiments with and without pressurization of explosives of different masses: Figures a(1), b(1), c(1) and d(1) are the strain and acceleration waveforms of explosion experiments with 5g TNT without pressurization, 10g TNT without pressurization, 5g TNT + 2MPa pressurization and 10g TNT + 2MPa pressurization, respectively; Figures a(2), b(2), c(2) and d(2) are the local magnified views of the corresponding working conditions. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings.

[0018] This invention discloses a method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks.

[0019] Reference Figures 1-8 Example 1, a method for generating an operation and maintenance scheme for deep-sea explosive containers based on generative adversarial networks, includes the following steps: Step 1: Collect historical operational data of deep-water explosion containers, and sequentially perform outlier removal, missing value supplementation, and standardization processing. Combined with correlation analysis, clarify the physical constraint relationship between the dosage, hydrostatic pressure, number of tests, and maximum strain, and construct a structured dataset. Step 2: Construct a conditional generative adversarial network (CGAN) with embedded physical constraint modules. CGAN includes a generator and a discriminator. The generator generates virtual strain data that conforms to physical laws by taking random noise and working condition parameters as input. The discriminator judges the authenticity of the input real or virtual strain data and the corresponding working condition parameters. Step 3: Train CGAN using the WGAN algorithm, and merge virtual data that meets the quality requirements with real data to form an expanded training dataset; Step 4: Construct a gated recurrent neural network strain prediction model GRU based on the expanded training dataset, optimize the model parameters through cross-validation, evaluate the model performance using mean absolute percentage error and coefficient of determination, and determine the optimal GRU strain prediction model. Step 5: Based on the maximum strain prediction value output by the optimal GRU strain prediction model, the fuzziness of the material performance parameters is first processed by the fuzzy interval reliability model to obtain the reliability interval. Then, the reliability interval is used as input and combined with the random interval hybrid reliability model to fuse the randomness of the predicted strain, and the reliability index η used to quantify the safety status of the container is calculated. Step 6: Based on the reliability index η, the operating status of the deep-water explosion container is divided into three maintenance levels: early warning level, attention level, and emergency level. Correspondingly, routine monitoring strategies, fatigue mitigation strategies, and emergency repair strategies are generated to form a dynamic maintenance plan.

[0020] Example 2: The historical operating data of the deep-water explosion container in step 1 includes historical data on explosive charge, hydrostatic pressure, number of tests, and corresponding maximum strain.

[0021] In Example 3, in step 1, outliers in the historical data are removed using box plots, and missing strain data are supplemented using cubic spline interpolation, where the interpolation process follows the physical laws of fatigue accumulation. The processed data is standardized, and the physical constraint relationship between the dosage, water pressure, number of tests and maximum strain is clarified through correlation analysis and visualization methods, thereby constructing a structured dataset.

[0022] In Example 4, in step 2, the scarce working condition samples are given higher weights. The scarce working condition is a working condition with high dosage and high water pressure. Specifically, the working condition with high dosage and high water pressure is a working condition with a dosage greater than 80% of the maximum dosage in historical data and a static water pressure greater than 80% of the maximum static water pressure in historical data. The discriminator assigns twice the weight of the conventional working condition samples to the scarce working condition samples.

[0023] In Example 5, in step 2, the physical constraint module constrains physical laws by adding a penalty term to the loss function of the generator.

[0024] Example 6: In step 4, the GRU model parameters are optimized through cross-validation, and the mean absolute error (MAPE) and coefficient of determination are used. As a performance evaluation metric, cross-validation is k-fold cross-validation, with k ranging from 5 to 10. The GRU model parameters to be optimized include the number of hidden layer nodes, learning rate, and number of iterations. The optimization objective is to minimize the MAPE of the test set.

[0025] Example 7, in step 4, the model performance evaluation criteria require the mean absolute percentage error (MAPE) of the test set to be less than 8% and R... 2 Greater than 0.9.

[0026] Example 8, the material performance parameters in step 5 include elastic modulus E and yield strength. In the fuzzy interval reliability model, triangular fuzzy functions or trapezoidal fuzzy functions are used to describe the uncertainty of material performance parameters.

[0027] In Example 9, the reliability index thresholds η corresponding to the three maintenance levels in step 6 are as follows: η≥3.0 is the early warning level, corresponding to the conventional monitoring strategy, which records strain data every 10 tests; 1.5≤η<3.0 is the attention level, corresponding to the fatigue mitigation strategy, which is achieved by reducing the dosage or hydrostatic pressure and shortening the monitoring cycle to every 5 tests; η<1.5 is the emergency level, corresponding to the emergency repair strategy, which involves immediately stopping the machine for repair and replacing fatigue-damaged components.

[0028] Example 10, the formula for calculating the reliability index η is: ; in The mean and standard deviation of the yield strength. These are statistical parameters of the load stress.

[0029] Step 1: Cleaning and preprocessing of operational data from deep-water explosion containers; ① Data Collection and Integration: Collect historical experimental data, including independent variables (drug dosage) hydrostatic pressure Number of trials ) and dependent variable (maximum stress on container wall) This integrates multi-source records (such as sensor monitoring data and test reports) to form a structured dataset. ,in This represents the total sample size.

[0030] ② Outlier removal: Outliers are identified using box plots, and the quartiles of the strain data are calculated. (Lower quartile) (Upper quartile) and interquartile range It will exceed the range The samples are identified as outliers and removed to avoid data deviations caused by sensor malfunctions or operational errors.

[0031] ③ Missing value imputation: For a small number of missing strain data Cubic spline interpolation is used to supplement the interval. For missing values ​​within the range, construct an interpolation function; , Interpolation conditions are met: , The second derivative is continuous. This ensures that the interpolation results conform to the smooth growth pattern of fatigue accumulation.

[0032] ④ Data Standardization and Correlation Analysis: z-score standardization is used to eliminate the influence of dimensions. The formula is as follows: ,in For raw data , The mean, The standard deviation is given. The correlation strength between the independent variable and strain is calculated using the Pearson correlation coefficient: Combine visualization to clarify physical constraints and Positive correlation ( ,and The positive correlation (fatigue cumulative effect) provides a constraint basis for subsequent model construction.

[0033] Step 2: Virtual data generation based on Conditional Generative Adversarial Network (CGAN); ①CGAN model structure design: Generator : with random noise and operating parameters As input, virtual strain is generated through a 3-layer fully connected neural network. An embedded physical constraint module is used to add a penalty term to the loss function:

[0034] in The penalty coefficient (ranging from [1,5]) ensures that the generated data meets the following conditions. , , The physical laws.

[0035] Discriminator Input the actual strain or virtual strain and corresponding operating parameters Output the probability of data authenticity For scarce operating conditions (such as...) and ,in , The sample with the highest value in historical data is assigned a weight of ω=2 (the regular sample has ω=1) to improve the quality of generating scarce intervals.

[0036] ② Model Training: Stable training is achieved using the WGAN algorithm, with alternating optimization of the generator and discriminator. Discriminator loss (including weights): And satisfy the Lipschitz constraint through gradient clipping; Generator loss: ; The training process is iterated 1000 times. Every 50 iterations, the Wasserstein distance between the generated data and the real data is calculated, and the process stops after convergence.

[0037] ③ Data Fusion: Assess the quality of virtual data from two aspects: statistical consistency (mean and variance deviation <5%) and physical rationality (satisfying the constraints in step 1). Merge qualified data with real data to form an expanded dataset. .

[0038] Step 3: Construction and optimization of the GRU strain prediction model based on fused data; ① Model Structure: A gated recurrent neural network (GRU) is constructed to capture the time dependency between the number of trials and strain. The input layer is... The hidden layer contains 64 neurons, and the output layer is the predicted strain. The core formulas include: Update Gate Control the proportion of historical information retained; Reset door Control the proportion of historical information updates; Candidate hidden state ; Hidden state ; in For the sigmoid function, For element-wise product, These are the parameters to be optimized.

[0039] ② Training and optimization: The model was divided into training and testing sets in an 8:2 ratio and trained using the Adam optimizer. Parameters were optimized using 5-fold cross-validation, with the goal of minimizing the mean absolute percentage error (MAPE) on the testing set to determine the optimal parameter combination.

[0040] ③Performance evaluation: The accuracy of the model is quantified using indicators. ; ; in To determine the sample size of the test set, the following is required: and .

[0041] Step 4: Reliability analysis of deep-water explosion container based on dual model; Fuzzy interval reliability model: Material property parameters (elastic modulus) Yield strength The variable is considered as a fuzzy variable, and its uncertainty is described using a triangular fuzzy function: ; in As the center value, The fuzzy radius is 10% of the mean.

[0042] Failure criterion: The stress corresponding to the maximum strain Calculate the reliability interval The impact of fuzziness in response parameters on the safety status.

[0043] Stochastic-interval hybrid reliability model: Material parameters are statistically transformed into interval variables (e.g.) ), and by combining the stochasticity of the strain predicted by GRU, the reliability index is calculated: ; in The mean and standard deviation of the yield strength. These are statistical parameters of the load stress.

[0044] Step 5: Generation of operation and maintenance plan for deep-water explosion container; Based on reliability indicators Three maintenance levels are defined: Warning level ( The container is in a safe state, and a conventional monitoring strategy is used. Strain data is recorded every 10 tests, and no adjustment of experimental parameters is required. Attention level ( ( ) In a sub-safe state, fatigue accumulation can be slowed down and the monitoring cycle shortened by reducing the dosage or water pressure; Emergency level ( If the machine is nearing failure, immediately stop it for inspection and repair, and replace any fatigue-damaged parts.

[0045] The following specific embodiments illustrate the implementation principle of the present invention: Taking a certain type of deep-water explosion container as the research object, Figure 1 This paper demonstrates the basic process of a method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks, including the following steps: Step 1: Data collection and preprocessing of deep-water explosion vessel operation: Figure 2 This demonstrates the basic workflow of data collection and preprocessing, specifically including the following steps: Data collection and integration: Historical experimental data were collected from N=783 groups, including the independent variable (drug dosage). hydrostatic pressure Number of trials ) and dependent variable (maximum strain of container wall) Example data is shown in Table 1, which integrates multi-source records (such as sensor monitoring data and test reports) to form a structured dataset. ,in This represents the total sample size.

[0046] Table 1. Examples of Experimental Data Samples

[0047] Outlier removal: Outliers were identified using box plots, and the quartiles of the strain data were calculated. (Lower quartile) (Upper quartile) and interquartile range It will exceed the range The samples are identified as outliers and removed to avoid data deviations caused by sensor malfunctions or operational errors.

[0048] In one implementation, box plot analysis was performed on the strain values ​​of all 783 sets of data, and the lower quartile Q1 was calculated to be 68.2, the upper quartile Q3 to be 132.5, and the interquartile range (IQR) to be 64.3. Samples exceeding the range [68.2-1.5×64.3, 132.5+1.5×64.3], i.e., [-28.25, 228.95], were identified as outliers (a total of 21 sets were removed).

[0049] Missing value supplementation: For a small amount of missing strain data Cubic spline interpolation is used to supplement the interval. For missing values ​​within the range, construct an interpolation function; ; Interpolation conditions are met: , The second derivative is continuous. This ensures that the interpolation results conform to the smooth growth pattern of fatigue accumulation.

[0050] Data standardization and correlation analysis: z-score standardization is used to eliminate the influence of dimensions. The formula is as follows: ,in For raw data , The mean, The standard deviation is given. The correlation strength between the independent variable and strain is calculated using the Pearson correlation coefficient: Combine visualization to clarify physical constraints and Positive correlation ( ,and The positive correlation (fatigue cumulative effect) provides a constraint basis for subsequent model construction.

[0051] In one implementation, the 762 valid data sets after removing outliers are Z-score standardized, and the mean and standard deviation of each parameter (drug dosage q, hydrostatic pressure p, number of tests n, maximum strain ε) are calculated. For example, for the drug dosage q, μ q =2.15, σ q =1.08; Calculate the Pearson correlation coefficient to verify the physical constraints: r q,ε =0.76 (strong positive correlation) r p,ε =0.69 (strong positive correlation) r n,ε=0.82 (strong positive correlation, confirming the cumulative effect of fatigue).

[0052] Step 2: Virtual data generation based on CGAN; Figure 3 The basic process of generating virtual data based on CGAN is demonstrated, specifically including the following steps: CGAN model structure design: Generator : with random noise and operating parameters As input, virtual strain is generated through a 3-layer fully connected neural network. An embedded physical constraint module is used to add a penalty term to the loss function: ; in The penalty coefficient (ranging from [1,5]) ensures that the generated data meets the following conditions. , , The physical laws.

[0053] In one implementation, the generator is a 3-layer fully connected network (4 nodes in the input layer, 128 nodes in the hidden layer, and 1 node in the output layer), with penalty coefficients of λ1=2, λ2=2, and λ3=3. Discriminator Input the actual strain or virtual strain and corresponding operating parameters Output the probability of data authenticity For scarce operating conditions (such as...) and ,in , The sample with the highest value in historical data is assigned a weight of ω=2 (the regular sample has ω=1) to improve the quality of generating scarce intervals.

[0054] In one implementation, the discriminator assigns high weight to scarce operating condition samples and determines q based on historical data. max =10kg, p max =30MPa, the scarce working condition is defined as q>8kg and p>24MPa, there are 41 groups of such samples, and the weight ω=2.

[0055] Model training: Stable training is achieved using the WGAN algorithm, with alternating optimization of the generator and discriminator: Discriminator loss (including weights): And satisfy the Lipschitz constraint through gradient clipping; Generator loss: ; The training process is iterated 1000 times. Every 50 iterations, the Wasserstein distance between the generated data and the real data is calculated, and the process stops after convergence.

[0056] In one implementation, the WGAN algorithm is used for training 1000 times. The training process is stable, and the Wasserstein distance steadily decreases with the number of iterations, converging after the 850th iteration (distance < 0.1).

[0057] Data fusion: The quality of virtual data is assessed from two aspects: statistical consistency (mean and variance deviation <5%) and physical plausibility (satisfying the constraints in step 1). Qualified data is then merged with real data to form an expanded dataset. .

[0058] In one implementation, 2000 sets of high-quality virtual data are generated. After evaluation, the mean strain deviation and variance deviation between the virtual data and the real data are <3%, and they meet physical constraints. Finally, the dataset D is fused. fusion It contains 762 + 2000 = 2762 groups of samples.

[0059] Step 3: Construction and optimization of the GRU strain prediction model based on fused data; Figure 4 This demonstrates the basic process of building and optimizing a GRU strain prediction model based on fused data, specifically including the following steps: Model Structure: A gated recurrent neural network (GRU) is constructed to capture the time dependency between the number of trials and strain. The input layer is... The hidden layer contains 64 neurons, and the output layer is the predicted strain. The core formulas include: Update Gate Control the proportion of historical information retained; Reset door Control the proportion of historical information updates; Candidate hidden state ; Hidden state ; in For the sigmoid function, For element-wise product, These are the parameters to be optimized.

[0060] Training and optimization: The model was divided into training and testing sets in an 8:2 ratio and trained using the Adam optimizer. Parameters were optimized using 5-fold cross-validation, with the goal of minimizing the mean absolute percentage error (MAPE) on the testing set to determine the optimal parameter combination.

[0061] In one implementation, the fused dataset is divided into a training set (2209 groups) and a test set (553 groups) in an 8:2 ratio. The Adam optimizer is used, and the optimal hyperparameters are determined by 5-fold cross-validation: learning rate lr=0.001 and number of iterations epochs=300.

[0062] Performance evaluation: Model accuracy is quantified using indicators. ; ; in To determine the sample size of the test set, the following is required: and .

[0063] In one implementation, the model performance is evaluated on a test set (M=553), and the following calculations are obtained: =6.8%; =0.923.

[0064] Step 4: Reliability analysis of deep-water explosion container based on dual model; Figure 5 The basic workflow for reliability analysis of deep-water explosion containers based on a dual-model approach is demonstrated, specifically including the following steps: ① Construct a fuzzy interval reliability model to handle the fuzziness of material performance parameters: Material property parameters (elastic modulus) Yield strength The variable is considered as a fuzzy variable, and its uncertainty is described using a triangular fuzzy function: ; in As the center value, The fuzzy radius is 10% of the mean.

[0065] The failure criterion is defined as the stress corresponding to the maximum strain. The reliability interval is obtained by calculating the transition range from "completely reliable" to "completely failed" using fuzzy mathematics. This is to reflect the impact of material parameter ambiguity on safety status.

[0066] ② Construct a stochastic-interval hybrid reliability model to handle the randomness of load parameters and calculate reliability indices: Material parameters are statistically transformed into interval variables (e.g.) That is, using the reliability interval obtained in the previous step. As input, the reliability index is calculated by incorporating the stochasticity of the strain predicted by GRU. The calculation formula is as follows: ; in The mean and standard deviation of the yield strength. For the statistical parameters of load stress (where, and These are the mean and standard deviation of the strain values ​​predicted by the GRU model in step 3, respectively.

[0067] Security threshold quantification: The reliability index η, as a unified indicator for quantifying the safety status of a container, will be used for maintenance level classification and strategy generation in step 5.

[0068] In one implementation, the material parameters are treated as triangular fuzzy variables, with the elastic modulus E being the center value a. E =206 GPa, fuzzy radius b E =20.6 GPa; Yield strength σ s : Central value a s =895MPa, blur radius b s =89.5MPa (10%); for a certain predicted strain ε pred =168.5µε, calculate its stress σ=E⋅ε pred Through the fuzzy failure criterion σ≥σ s The reliability interval is obtained as [ R min, R [max]=[0.91,0.98]; Convert the fuzzy material parameters into intervals: E∈[185.4,226.6]GPa, σs∈[805.5,984.5]MPa; Combine the strain distribution predicted by GRU (for a specific working condition, the predicted strain mean με=170.2, standard deviation σε=14.5), calculate the load stress distribution parameters: μσload=E⋅με, σσload=E⋅σε; Substitute into the reliability index formula, calculate: η=2.4.

[0069] Step 5: Generation of operation and maintenance plan for deep-water explosion container; Figure 6 This demonstrates the basic process for generating operation and maintenance plans for deep-water explosion containers, specifically including the following steps: Based on reliability indicators The maintenance levels are divided into three levels (Table 2 shows examples of three typical operating conditions): Warning level ( The container is in a safe state, and a conventional monitoring strategy is used. Strain data is recorded every 10 tests, and no adjustment of experimental parameters is required. Attention level ( ( ) In a sub-safe state, fatigue accumulation can be slowed down and the monitoring cycle shortened by reducing the dosage or water pressure; Emergency level ( If the machine is nearing failure, immediately stop it for inspection and repair, and replace any fatigue-damaged parts.

[0070] Table 2 Examples of maintenance levels and strategies under three typical operating conditions

[0071] To fully verify the effectiveness and engineering applicability of the method described in this invention, a system test was conducted based on all experimental data obtained from the deep-water explosion vessel acceptance test. The test covered four typical operating conditions, making full use of the charts and tables in the experimental data to ensure the comprehensiveness and reliability of the verification.

[0072] Test conditions and data sources: Tests are based on sensor placement schemes (e.g.) Figure 7 As shown in the figure, a complete dataset was used, comprising six strain gauges (1#-6#) and one acceleration gauge. The test included four typical operating conditions: ①5g TNT explosion experiment without pressure: its strain and acceleration waveforms are as follows Figure 8 As shown in -a(1) and 8-a(2); ② 10g TNT explosion experiment without pressure: its strain and acceleration waveforms are as follows Figure 8 As shown in -b(1) and 8-b(2); ③ 5g TNT + 2MPa pressurized explosion experiment: its strain and acceleration waveforms are as follows Figure 8 -c(1) and 8-c(2) are shown; ④ 10g TNT + 2MPa pressure explosion experiment: its strain and acceleration waveforms are as follows Figure 8 As shown in -b(1) and 8-b(2).

[0073] Detailed experimental data for each working condition are shown in Table 3.

[0074] Table 3. Interpretation results of experimental data for explosive containers with different TNT equivalents

[0075] Note: "*" indicates that the accelerometer was damaged in the first two experiments.

[0076] Test steps and result analysis: Data preprocessing and feature extraction: Four sets of complete experimental data were extracted from Table 1. The maximum strain at each measuring point was below the plastic limit (1500 με), indicating that the explosion container was in a safe working condition. Combined with... Figure 8 Waveform feature analysis revealed that: Unpressurized operating conditions ( Figure 8 -a、 Figure 8 -b) The strain waveform exhibits typical characteristics of an explosion shock response; Pressurized operating conditions ( Figure 8 -c、 Figure 8 -d) The increase in strain waveform amplitude verifies the rule in conclusion b that "strain generally increases with the increase of equivalent". The data from measuring points 5 and 6 show a certain degree of randomness (strain at point 5 in Table 1: 407→263→305→277), which is related to the special nature of their locations.

[0077] CGAN virtual data generation and verification: Using the four complete experimental data sets in Table 1 as the training set, 400 sets of virtual data were successfully generated. The generated data was compared with the measured data: The statistical deviation between the virtual data and the measured data in Table 1 is less than 5%; The generated data strictly follows the physical constraints of ∂ε / ∂q≥0 and ∂ε / ∂p≥0; The randomness characteristics of measuring points #5 and #6 were successfully simulated.

[0078] Performance validation of the GRU strain prediction model: The GRU model is trained using the expanded dataset to make predictions for various operating conditions: ①5gTNT without pressure: predicted maximum strain 502με, measured value 600με; ② 10g TNT without pressure: predicted maximum strain 588με, measured value 614με; ③5gTNT+2MPa working condition: predicted maximum strain 482με, measured value 527με; ④10gTNT+2MPa working condition: predicted maximum strain 602με, measured value 614με.

[0079] All predicted values ​​were well below the plastic limit (1500 με), further verifying that the container was in a safe working condition.

[0080] Reliability analysis and safety assessment: Based on the prediction results, reliability analysis showed that the reliability index η was greater than 1.5 for all operating conditions. ①5g TNT operating condition: η=3.2 (early warning level); ②10g TNT operating condition: η=2.8 (level of concern); ③ Pressurized condition: η=2.1-2.4 (Level of concern).

[0081] The results show that the container meets the design specifications under all test parameters, indicating that the design is reasonable and the manufacturing is reliable.

[0082] Maintenance strategy generation and optimization: Based on the reliability analysis results, a tiered maintenance plan is generated: ①η≥3.0: Warning level, routine monitoring (corresponding to 5g TNT operating conditions); ②1.5≤η<3.0: Level of concern, adjust test parameters (corresponding to 10gTNT and pressurized conditions); ③ All solutions ensure that the container is in a safe working condition.

[0083] Test conclusion: The method of this invention has verified the following important characteristics through system testing: (1) Safety verification: All predicted strain values ​​(502-602με) are far below the plastic limit (1500με), proving that the container is always in a safe working state; (2) Physical laws are followed: The generated data strictly follows the positive correlation between strain and drug dosage and water pressure; (3) Special working conditions: The randomness of measuring points 5# and 6# was successfully handled, demonstrating good engineering adaptability; (4) Design rationality verification: The reliability indicators are all within the safe range, which proves that the container design is reasonable and the processing is reliable; (5) Engineering practicality: The prediction model based on multi-source data fusion showed good performance under various working conditions, with MAPE=7.8% and R 2 =0.908.

[0084] This test case, through comprehensive utilization of the charts and tables from four sets of experiments, fully verifies the effectiveness, reliability, and engineering applicability of the method of this invention in the operation and maintenance of deep-water explosive containers. The test results are highly consistent with the experimental conclusions, demonstrating that the method can effectively support the safe operation and maintenance decisions of explosive containers, and providing sufficient data support and a verification basis for the practical application of the patented technology.

[0085] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for generating a deep water explosion containment vessel operation and maintenance scheme based on a generative adversarial network, characterized in that: Includes the following steps: Step 1: Collect historical operational data of deep-water explosion containers, and sequentially perform outlier removal, missing value supplementation, and standardization processing. Combined with correlation analysis, clarify the physical constraint relationship between the dosage, hydrostatic pressure, number of tests, and maximum strain, and construct a structured dataset. Step 2: Construct a conditional generative adversarial network (CGAN) with embedded physical constraint modules. CGAN includes a generator and a discriminator. The generator generates virtual strain data that conforms to physical laws by taking random noise and working condition parameters as input. The discriminator judges the authenticity of the input real or virtual strain data and the corresponding working condition parameters. Step 3: Train CGAN using the WGAN algorithm, and merge virtual data that meets the quality requirements with real data to form an expanded training dataset; Step 4: Construct a gated recurrent neural network strain prediction model GRU based on the expanded training dataset, optimize the model parameters through cross-validation, evaluate the model performance using mean absolute percentage error and coefficient of determination, and determine the optimal GRU strain prediction model. Step 5: Based on the maximum strain prediction value output by the optimal GRU strain prediction model, the fuzziness of the material performance parameters is first processed by the fuzzy interval reliability model to obtain the reliability interval. Then, the reliability interval is used as input and combined with the random interval hybrid reliability model to fuse the randomness of the predicted strain, and the reliability index η used to quantify the safety status of the container is calculated. Step 6: Based on the reliability index η, the operating status of the deep-water explosion container is divided into three maintenance levels: early warning level, attention level, and emergency level. Correspondingly, routine monitoring strategies, fatigue mitigation strategies, and emergency repair strategies are generated to form a dynamic maintenance plan.

2. The deep water explosion vessel operation scheme generation method based on a generative adversarial network according to claim 1, characterized in that: The historical operational data of the deep-water explosion vessel in step 1 includes historical data on explosive charge, hydrostatic pressure, number of tests, and corresponding maximum strain.

3. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 2, characterized in that: In step 1, outliers in the historical data are removed using box plots, and missing strain data are supplemented using cubic spline interpolation, where the interpolation process follows the physical laws of fatigue accumulation. The processed data is standardized, and the physical constraints between the dosage, water pressure, number of tests, and maximum strain are clarified through correlation analysis and visualization methods, thereby constructing a structured dataset.

4. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 3, characterized in that: In step 2, the scarce working condition samples are given higher weights. The scarce working condition is the working condition with high dosage and high water pressure. Specifically, the working condition with high dosage and high water pressure is the working condition with a dosage greater than 80% of the maximum dosage in historical data and a static water pressure greater than 80% of the maximum static water pressure in historical data. The discriminator assigns twice the weight of the scarce working condition samples as the conventional working condition samples.

5. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 4, characterized in that: In step 2, the physical constraint module constrains physical laws by adding a penalty term to the loss function of the generator.

6. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 5, characterized in that: In step 4, the GRU model parameters are optimized through cross-validation, and the mean absolute error (MAPE) and coefficient of determination are used. As a performance evaluation metric, cross-validation is k-fold cross-validation, with k ranging from 5 to 10. The GRU model parameters to be optimized include the number of hidden layer nodes, learning rate, and number of iterations. The optimization objective is to minimize the MAPE of the test set.

7. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 6, characterized in that: The performance evaluation of the model in Step 4 requires the mean absolute percentage error MAPE of the test set to be less than 8% and R 2 greater than 0.

9.

8. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 7, characterized in that: The material property parameters in step 5 include the elastic modulus E and the yield strength. In the fuzzy interval reliability model, triangular fuzzy functions or trapezoidal fuzzy functions are used to describe the uncertainty of material performance parameters.

9. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 8, characterized in that: The reliability index thresholds η corresponding to the three maintenance levels in step 6 are as follows: η≥3.0 is the early warning level, which corresponds to the conventional monitoring strategy, and the strain data is recorded every 10 tests; 1.5≤η<3.0 is the attention level, which corresponds to the fatigue mitigation strategy, which is achieved by reducing the dosage or hydrostatic pressure and shortening the monitoring cycle to every 5 tests. If η < 1.5, it is an emergency level, corresponding to an emergency maintenance strategy, which involves immediately shutting down the machine for maintenance and replacing fatigue-damaged parts.

10. The method for generating operation and maintenance schemes for deep-sea explosive containers based on generative adversarial networks according to claim 9, characterized in that: The formula for calculating the reliability index η is: ； in The mean and standard deviation of the yield strength. These are statistical parameters of the load stress.

Images