Liquid hydrogen storage tank structure system two-process reliability evaluation method based on hybrid model

Abstract

The invention provides a liquid hydrogen storage tank structure system two-process reliability evaluation method based on a hybrid model, and relates to the technical field of reliability engineering and evaluation.The method comprises the steps that a thermal-mechanical coupling analysis model is used for analyzing the specific structure and the operation working condition of a large liquid hydrogen storage tank, and a simulation result is obtained; identifying key factors influencing the reliability of the storage tank by utilizing a simulation result in combination with expert knowledge, historical data and failure physical analysis to obtain key failure risk factors; constructing the key failure risk factors into a polymorphic fault tree structure, and mapping the polymorphic fault tree structure to obtain a dynamic Bayesian network model; and performing double-process quantitative reliability evaluation on the simulation result by using a dynamic Bayesian network model and a mixed agent model to obtain a hierarchical reliability evaluation result. According to the method, the problem that the reliability of the large liquid hydrogen storage tank is difficult to comprehensively and accurately assess by the existing single system reliability method or structure reliability method is solved.

Description

Technical Field

[0001] This specification relates to the fields of reliability engineering and assessment technology, and in particular to a dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model. Background Technology

[0002] Large liquid hydrogen storage tanks are critical equipment for hydrogen energy storage and transportation. They operate under harsh conditions such as extremely low temperatures (approximately 20K) and complex external environments (e.g., variable loads), and involve complex thermal-structural coupling effects. Failure in any part of the tank's insulation structure (e.g., vacuum jackets, multi-layer insulation materials), pressure-bearing structure (inner and outer liner, welds), or supporting structure can lead to catastrophic accidents. Therefore, conducting accurate and comprehensive reliability assessments of large liquid hydrogen storage tanks is a core element in ensuring their design rationality and operational safety.

[0003] Existing reliability assessment methods can be broadly categorized into system and structural methods. System reliability methods (such as FMEA, FTA, and BN) excel at describing logical relationships but struggle to directly integrate with the complex physical fields (such as stress and temperature) coupled with failure mechanisms calculated by finite element models (FEM). In particular, while dynamic Bayesian networks (DBNs) can handle time-varying conditions, deep integration with physical models remains a challenge. Structural reliability methods (such as FORM, MCS, and surrogate models) focus on the physical level. Monte Carlo simulation (MCS), while highly accurate, incurs enormous computational costs when combined with FEM, hindering engineering applications; approximate analytical methods (FORM / SORM) have limited accuracy for highly nonlinear problems. Surrogate models (such as Kriging) significantly improve computational efficiency, but traditional applications are often limited to single components or limit state equations, making it difficult to extend to dynamic system-level assessment frameworks involving multiple components, multiple levels, and multiple failure mode interactions.

[0004] The reliability issues of large liquid hydrogen storage tanks are characterized by multi-physics coupling, multi-level structure, dynamic evolution, and interaction of multiple failure modes. Multi-physics coupling involves strong coupling effects between cryogenic heat transfer, fluid flow, and structural stress and strain. The multi-level structure extends from basic materials and welds (component level) to vacuum interlayers and supporting structures (partial level), and then to the tank body and safety systems (system level), where failures can occur at different levels and influence each other. Dynamic evolution means that material properties may degrade over time, insulation performance may decline, loads may change, and failure modes (such as crack propagation) are time-varying. The coexistence and interaction of multiple failure modes may involve simultaneous occurrences of insufficient structural strength, instability, fatigue fracture, insulation failure, leakage, and other modes, which may mutually promote each other (e.g., stress concentration caused by temperature changes accelerates crack propagation).

[0005] Existing single system reliability or structural reliability methods are insufficient to comprehensively and accurately assess the reliability of large liquid hydrogen storage tanks. Qualitative / semi-quantitative methods cannot quantify risks; traditional quantitative methods struggle to incorporate complex physical mechanisms; structural reliability methods focus on the component level, making it difficult to grasp the overall system function and the interaction of multiple components; while surrogate models can improve efficiency, they need to be effectively integrated into a system-level dynamic evaluation framework. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model provided by this invention solves the problem that existing single system reliability methods or structural reliability methods are insufficient to comprehensively and accurately assess the reliability of large liquid hydrogen storage tanks.

[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model, comprising: S1: Using a thermo-coupling analysis model, the specific structure and operating conditions of a large liquid hydrogen storage tank are analyzed, and simulation results are obtained. S2: Using simulation results, combined with expert knowledge, historical data and failure physics analysis, identify the key factors affecting the reliability of the storage tank and obtain the key failure risk factors; S3: Construct key failure risk factors into a polymorphic fault tree structure, and map the polymorphic fault tree structure to obtain a dynamic Bayesian network model. S4: Using a dynamic Bayesian network model and a hybrid surrogate model, a two-stage quantitative reliability assessment of the simulation results is performed to obtain hierarchical reliability assessment results, thus completing the two-stage reliability assessment of the liquid hydrogen storage tank structure system; the hybrid surrogate model is obtained through training.

[0008] Further, S1 includes: The specific structure and operating conditions of a large liquid hydrogen storage tank are analyzed to obtain the corresponding mesh-divided three-dimensional geometric model. By using a thermo-coupling analysis model, the three-dimensional geometric model of the mesh is processed by setting boundary conditions, and multi-physics coupling is performed to obtain simulation results.

[0009] Further, S2 includes: By using simulation results, combined with expert knowledge, historical data, and failure physics analysis, the possible failure modes of large liquid hydrogen storage tanks are identified, and the failure mode identification results are obtained. Based on the failure mode identification results, risk factors are traced and identified to obtain the key factors affecting the reliability of the storage tank. The key factors affecting the reliability of storage tanks are quantified and screened to identify key failure risk factors.

[0010] Furthermore, the key factors affecting the reliability of the storage tank are quantified and screened to obtain the key failure risk factors, which include: The key factors affecting the reliability of storage tanks are quantified. For risk factors with clear physical properties and direct measurability, parameterization is performed. For process-related, state-related, or risk factors that are difficult to describe directly with a single numerical value, qualitative classification is performed to obtain the quantification results of key factors. Based on the quantitative results of key factors, the risk priority number of key factors is calculated using the expert evaluation method to obtain the first group of candidate key risk factors. Based on the quantification results of key factors, the standardized local sensitivity index was calculated using the preliminary judgment method of sensitivity analysis to obtain the second group of candidate key risk factors. By combining the first group of candidate key risk factors and the second group of candidate key risk factors, a review list is constructed to obtain the key failure risk factors.

[0011] Further, S3 includes: Based on key failure risk factors, by defining the top event and system decomposition, identifying event states, selecting logic gates and constructing a tree structure, a polymorphic fault tree structure is obtained. An initial mapping is performed from the polymorphic fault tree structure to the dynamic Bayesian network model to obtain the conditional probability table structure. Based on the conditional probability table structure, the conditional probability of each node in the initial time slice dynamic Bayesian network is quantified, time dependence is introduced, and key failure risk factors are taken as nodes of the dynamic Bayesian network to obtain the dynamic Bayesian network model.

[0012] Furthermore, the hybrid agent model includes: The global trend regression module is used to map and linearly combine the input vector to be predicted to obtain the global trend prediction result. The local bias correction module is used to calculate the residuals between the input vector to be predicted and the training dataset to obtain the local bias correction value for the global trend prediction. The output layer is used to comprehensively analyze the global trend prediction results and the local deviation correction values ​​of the global trend prediction to obtain the prediction values ​​of the hybrid surrogate model.

[0013] Furthermore, the expression for the global trend prediction result is: ; The expression for the local deviation correction value of the global trend prediction is: ; The expression for the predicted value of the hybrid proxy model is: ; ; ; in, This indicates the overall trend prediction result; represents the input point to be predicted; j represents the index of the basis function, from 1 to p; p represents the total number of basis functions in the global trend model; This represents the regression coefficient corresponding to the j-th basis function; This represents the j-th basis function acting on the input point x; This represents a row vector consisting of all basis functions; This represents a p-dimensional column vector consisting of all regression coefficients; This represents the local deviation correction value for the global trend prediction, calculated by the Gaussian process model; This represents the covariance between the new point x and all training points. This represents the inverse of the covariance matrix K between training sample points; This represents the output vector of the training sample; Let p represent the n×p matrix formed by the training input samples under the basis functions. This represents the best estimate of the regression coefficients. This represents the final predicted value of the hybrid proxy model for input point x. This represents the transpose of the design matrix F.

[0014] Furthermore, the training process of the hybrid agent model includes: Key parameters affecting the reliability of storage tanks are sampled, and each training input point is simulated using a thermo-mechanical coupling analysis model to obtain the physical response field. A failure criterion function is established based on the physical failure mechanism, and the event probability is calculated on the physical response field to obtain a training sample set containing multiple sample pairs. Using a training sample set containing multiple sample pairs, a numerical optimization algorithm is used to find the parameter combination that maximizes the probability of the training data appearing. The coefficients of the global trend function and the hyperparameters of the Gaussian process covariance function of the hybrid agent model are optimized to obtain the trained hybrid agent model.

[0015] Furthermore, the expression for the loss function of the hybrid agent model is: ; ; ; in, Let represent the negative log-likelihood function, which is the objective function for model training and optimization; Represents the hyperparameters of the kernel function; This represents the process standard deviation of a Gaussian process. This represents the complete training dataset; Indicates the number of training samples; Represents the determinant operation of a matrix; Indicates by hyperparameters θ The correlation matrix of the decision; Indicates all n A column vector consisting of the true output values ​​of each training sample; Let n×p represent the design matrix formed by all training input samples under the basis functions; This represents the best estimate of the regression coefficients. Let represent the condensed negative log-likelihood function, which is a function only of the hyperparameter θ; This represents the maximum likelihood estimate of the process variance.

[0016] Further, S4 includes: The simulation results are analyzed using a dynamic Bayesian network model to obtain the physical parameters at the current time step; By using a hybrid proxy model, the physical parameters of the current time step are analyzed to obtain the conditional probabilities of each basic event at the current time step. By using the conditional probabilities of each basic event at the current time step, the state of the corresponding basic event node in the dynamic Bayesian network model is updated to obtain the updated dynamic Bayesian network model. By using the updated dynamic Bayesian network model, forward reasoning is performed on the simulation results to obtain hierarchical reliability assessment results, thus completing the dual-process reliability assessment of the liquid hydrogen storage tank structure system.

[0017] The beneficial effects of this invention are as follows: This invention provides a dual-process reliability assessment method for liquid hydrogen storage tank structure systems based on a hybrid model. Combining the advantages of MFTA, DBN and hybrid Kriging, it achieves a comprehensive, accurate and efficient quantitative assessment of the structure-system coupling, multi-level and dynamic reliability of large liquid hydrogen storage tanks. The assessment results can be directly used to guide the design optimization, operation monitoring, maintenance strategy formulation and remaining life prediction of the storage tank, thereby improving the intrinsic safety level and economy of large liquid hydrogen storage tanks. (1) By integrating system reliability (MFTA, DBN) and structural reliability (proxy model linking physical model) methods and considering multi-physics coupling, a more comprehensive and accurate assessment of the complex failure behavior of large liquid hydrogen storage tanks is achieved. (2) The innovative introduction of DBN enables the model to quantitatively describe time-varying processes such as damage accumulation and performance degradation, overcoming the limitations of static models and realizing dynamic assessment of the reliability of the storage tank throughout its entire life cycle. (3) The application of MFTA allows for modeling of multiple intermediate failure states of components, which is more in line with the gradual development process of failure in engineering practice, significantly improving the expressive power of the model and the precision of the assessment. (4) The self-constructed hybrid Kriging proxy model significantly reduces the number of calls to time-consuming physical simulations, greatly improving computational efficiency while ensuring accuracy, making it possible to conduct quantitative assessments of the entire lifecycle of large and complex systems. (5) The clear hierarchical decomposition (component-part-system) and structure-system dual-process assessment make the assessment logic clear, enabling a systematic consideration of failure transmission and impact between different levels, and realizing an integrated assessment from physical failure to system functional failure. (6) It not only provides quantitative reliability indicators, but also identifies key risk factors and their impact patterns through DBN analysis capabilities, providing strong scientific support for the optimized design, risk warning, and maintenance decisions of storage tanks. Attached Figure Description

[0018] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart of a dual-process reliability assessment method for a liquid hydrogen storage tank structure system based on a hybrid model, as shown in some embodiments of this specification. Figure 2 This is an exemplary framework diagram of a dual-process reliability assessment method for a liquid hydrogen storage tank structure system based on a hybrid model, as shown in some embodiments of this specification. Detailed Implementation

[0019] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0020] Example Figure 1 This is an exemplary flowchart of a dual-process reliability assessment method for a liquid hydrogen storage tank structure system based on a hybrid model, as shown in some embodiments of this specification. Figure 1 and Figure 2 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0021] S1: Using a thermo-coupling analysis model, the specific structure and operating conditions of a large liquid hydrogen storage tank are analyzed, and simulation results are obtained.

[0022] In some embodiments, the processor can, for the specific structure and operating conditions of a large liquid hydrogen storage tank, utilize numerical simulation techniques such as finite element analysis (FEA) to establish a thermo-mechanical coupling analysis model that describes the internal fluid flow, heat transfer between the tank wall and insulation structure, and the structural stress-strain response caused by temperature changes. This thermo-mechanical coupling analysis model forms the basis for subsequent risk identification and parameter mapping.

[0023] The specific structure and operating conditions of large liquid hydrogen storage tanks are determined based on the design drawings. A precise three-dimensional geometric model is established, including inner and outer tanks, insulation layers (such as vacuum jackets and fillers), supporting structures, and pipe interfaces. High-quality mesh generation is performed on the model, especially in stress concentration areas, near welds, and areas with large temperature gradients, to ensure calculation accuracy.

[0024] The simulation results extract key physical field data, such as temperature distribution, stress / strain distribution, heat flux density, displacement, and fluid velocity / pressure in critical components like the tank wall, welds, and support structures, especially how these physical quantities change over time. These results will serve as the basis for subsequent steps.

[0025] In some embodiments, the processor can analyze the specific structure and operating conditions of a large liquid hydrogen storage tank to obtain a corresponding meshed three-dimensional geometric model; using a thermo-coupling analysis model, the processor processes the meshed three-dimensional geometric model by setting boundary conditions, performs multi-physics coupling, and obtains simulation results.

[0026] In some embodiments, the processor can input the material properties of each component required for the model, paying particular attention to the mechanical properties (elastic modulus, yield strength, tensile strength, fatigue characteristics, etc.) and thermophysical properties (thermal conductivity, specific heat capacity, coefficient of thermal expansion, etc.) of the material at liquid hydrogen temperature (approximately 20K), taking into account the temperature dependence of the material properties.

[0027] Boundary conditions include thermal boundaries: ambient temperature, solar radiation, internal liquid hydrogen temperature, vacuum level of the vacuum interlayer (affecting residual gas heat conduction and convection), and the radiative heat transfer coefficient of the multilayer insulation material. Force boundaries include: internal pressure, liquid gravity, external wind load, seismic load, pipe connection force, and supporting structural constraints. Fluid boundaries include: inlet and outlet velocities and pressures.

[0028] In some embodiments, the multiphysics coupling process includes: setting up thermal-structural, thermal-fluid, or thermal-fluid-structure interaction (TFSI) coupling analysis in finite element analysis software; selecting an appropriate solver and time step; and performing transient or steady-state simulations to simulate the multiphysics response of the storage tank under preset operating conditions (such as steady-state storage, dynamic filling / discharging, emergency depressurization, etc.).

[0029] S2: Using simulation results, combined with expert knowledge, historical data, and failure physics analysis, identify the key factors affecting the reliability of the storage tank and obtain the key failure risk factors.

[0030] In some embodiments, the processor can identify key factors affecting the reliability of the storage tank based on the simulation results of the thermo-coupling analysis model, combined with expert knowledge, historical data and failure physics analysis, such as: material defects, weld quality, vacuum degree changes, loosening of support structure, external loads (wind load, earthquake), fluctuations in operating parameters (pressure, liquid level), and degradation of material low-temperature performance.

[0031] The key failure risk factors are a list of key failure risk factors that are controllable in number and highly representative. This list will serve as the direct basis and focus of attention for defining basic events in the subsequent construction of the Multimorphic Fault Tree (MFTA) model, and for defining root nodes and conditional probability tables in the construction of the Dynamic Bayesian Network (DBN) model.

[0032] In some embodiments, the processor can utilize simulation results, combined with expert knowledge, historical data, and failure physics analysis, to identify possible failure modes of large liquid hydrogen storage tanks and obtain failure mode identification results; based on the failure mode identification results, risk factor tracing and identification are performed to obtain key factors affecting the reliability of the storage tank; and the key factors affecting the reliability of the storage tank are quantified and screened to obtain key failure risk factors.

[0033] Failure Mode and Effects Identification (FMEA) results combine the simulation results from the first phase, the physical mechanisms of liquid hydrogen storage tank failure (such as cryogenic brittleness, fatigue accumulation, material degradation, vacuum failure mechanisms, etc.), expert experience, and historical accident data to systematically identify possible failure modes for large liquid hydrogen storage tanks. Examples include structural instability (buckling), material yielding or fracture, weld fatigue cracking, vacuum interlayer failure leading to a sharp decline in insulation performance, support structure failure, and seal leakage.

[0034] The key factors affecting the reliability of storage tanks are tracing the root causes or influencing factors (i.e., risk factors) for each failure mode. These factors can be categorized as follows: Inherent factors: inherent material defects, insufficient design margins, manufacturing process deviations (such as welding quality). Operational factors: overpressure, temperature shocks caused by rapid charging and discharging, number of cyclic loads, improper maintenance. Environmental factors: extreme external temperatures, corrosive environments, accidental impacts. Degradation factors: material aging, vacuum level decay over time, cumulative fatigue damage.

[0035] In some embodiments, the processor can quantify the key factors affecting the reliability of the storage tank. For risk factors with clear physical properties and direct measurability, parameterization is performed. For process-related, state-related, or risk factors that are difficult to describe directly with a single numerical value, qualitative classification is performed to obtain the key factor quantification results. Based on the quantitative results of key factors, the risk priority number of key factors is calculated using the expert evaluation method to obtain the first group of candidate key risk factors. Based on the quantification results of key factors, the standardized local sensitivity index was calculated using the preliminary judgment method of sensitivity analysis to obtain the second group of candidate key risk factors. By combining the first group of candidate key risk factors and the second group of candidate key risk factors, a review list is constructed to obtain the key failure risk factors.

[0036] In some embodiments, the processor can quantify or qualitatively classify each factor in the identified set of risk factors. For risk factors with clearly defined and directly measurable physical properties, parameterization is performed. This process includes determining the baseline value of the value, the statistical distribution type (e.g., normal distribution, Weibull distribution) and its distribution parameters (e.g., mean, standard deviation), or a functional expression of its change over time. Quantifiable factors include, but are not limited to: the initial defect size of the material, the geometric discontinuity parameters of the weld, the amplitude and frequency of external cyclic loads (e.g., wind load, temperature cycling), the pressure fluctuation range of the internal medium, and the annual vacuum decay rate of the vacuum interlayer. For process-related, state-related, or risk factors that are difficult to describe directly with a single numerical value, qualitative classification is performed. This process includes defining clearly defined, discrete, and ordered state levels, and assigning a clear physical or engineering meaning to each level. For example, qualitative classification factors might define the manufacturing process quality of the weld as "excellent," "medium," and "poor"; and the quality of maintenance as "standard," "average," and "inadequate."

[0037] In some embodiments, the processor may employ a hybrid screening strategy that combines qualitative expert evaluation with quantitative sensitivity analysis. Expert evaluation method: Using the Failure Mode and Effects Analysis (FMEA) framework, experts from multiple fields (covering materials, design, manufacturing, operation and maintenance, etc.) evaluate each risk factor. Preliminary ranking is achieved by calculating the Risk Priority Number (RPN) for each risk factor. The formula for calculating the RPN is: ; Wherein, S (Severity) represents the severity level of the consequences of the failure mode caused by the risk factor on the tank system; O (Occurrence) represents the probability or frequency of the occurrence of the risk factor throughout the entire life cycle of the tank; and D (Detection) represents the ease with which the risk factor or its early signs can be detected by existing monitoring and detection methods before a failure occurs or its consequences become severe (Note: the higher the detection difficulty, the larger the D value). Based on the preset RPN threshold or by selecting several factors ranked first in the RPN calculation results, the first group of candidate critical risk factors is formed.

[0038] In some embodiments, the processor can perform a preliminary sensitivity analysis on parameterized risk factors that have a simplified physical model, analytical expression, or empirical formula relating to a key performance indicator, to quantitatively assess the impact of that factor on the performance indicator. A simplified performance function Y = f(X1, X2, ..., X...) is established. n), where Y is an indicator characterizing a key performance of the storage tank (e.g., stress concentration factor at a specific location, fatigue crack propagation life, heat flux density of the insulation layer, etc.), X i These are the input risk factor parameters. Calculate X for each input factor. i The standardized local sensitivity index S of the output Y i , no m The calculation formula is as follows: ; in, Y / X i Let the performance function Y be the input parameter X at the design baseline. i The partial derivative of X represents the instantaneous rate of change of the output with respect to the input; i,0 For input parameter X i The design baseline or average value; Y0 is the output performance index Y calculated when all input parameters are taken as baseline values. The larger the absolute value of the sensitivity index, the more sensitive the tank performance is to fluctuations in this risk factor parameter. Factors are ranked according to their absolute values, and the top-ranked factors are selected to form the second group of candidate key risk factors.

[0039] In some embodiments, the processor can combine a first group of candidate critical risk factors and a second group of candidate critical risk factors to form a final list of critical failure risk factors. The screening principle is to identify risk factors that are highly important in both the expert evaluation method (high RPN value) and the preliminary sensitivity analysis method (high sensitivity index) as critical risk factors. For factors that show high importance only in one method (e.g., factors whose sensitivity is not significant due to model simplification but which experts consider to have extremely high risk, or factors with low RPN but extremely high sensitivity), a second review is conducted by an expert group, taking into account engineering practice experience, to decide whether to include them in the list.

[0040] S3: Construct a polymorphic fault tree structure for key failure risk factors, and map the polymorphic fault tree structure to obtain a dynamic Bayesian network model.

[0041] In some embodiments, the processor can use system-level functional failures of the storage tank (such as pressure failure, insulation failure, structural instability, etc.) as the top event, and decompose system failures into component-level and element-level failure events in a top-down manner. Considering that components have multiple failure states (such as minor leakage, severe leakage, complete fracture), the MFTA method is adopted, using polymorphic logic gates (such as AND gates, OR gates, k / n gates) to describe the logical relationships between failure events at different levels and of different components, thus constructing an MFTA model that can reflect the complex failure logic of the storage tank. Based on the topology of the MFTA, it is mapped to the initial time-slice structure of the DBN. Events in the MFTA correspond to nodes in the DBN, and the logical relationships are transformed into a conditional probability table (CPT) between nodes. A time dimension is introduced to establish the transition probabilities of nodes between adjacent time slices to describe the changes in system state (such as damage accumulation, performance degradation, failure mode evolution) over time. The weight information of risk factors can be used to correct the CPT or influence the transition probabilities. Thus, a DBN model that can capture the dynamic evolution characteristics of storage tank reliability is constructed.

[0042] In some embodiments, the processor can obtain a polymorphic fault tree structure by defining top events and system decomposition based on key failure risk factors, identifying event states, selecting logic gates and constructing a tree structure; An initial mapping is performed from the polymorphic fault tree structure to the dynamic Bayesian network model to obtain the conditional probability table structure. Based on the conditional probability table structure, the conditional probability of each node in the initial time slice dynamic Bayesian network is quantified, time dependence is introduced, and key failure risk factors are taken as nodes of the dynamic Bayesian network to obtain the dynamic Bayesian network model.

[0043] In some embodiments, the construction process of the polymorphic fault tree structure includes: selecting a system-level functional failure of the storage tank as the top event of the MFTA, such as "overall tank failure" or more specifically, "major leak in the storage tank". Following the hierarchical structure of system → subsystem → component → element, the top event is decomposed level by level down to the basic event (typically corresponding to component-level failure or external event). For events with multiple failure degrees or modes (especially components and elements), their discrete multi-states are defined. For example, the state of "vacuum interlayer insulation performance" can be defined as (excellent, slight decrease, significant decrease, complete failure); the state of "weld condition" can be defined as (intact, microcrack present, crack propagation to critical size, through fracture). MFTA logic gates (such as AND gates, OR gates, PAND gates, k / n gates, etc.) are used to connect events at different levels, accurately describing the causal relationships and logical combinations between them. Special attention is paid to using logic gates capable of handling multi-state inputs. For example, the "moderate failure" state of a component is composed of the "slight failure" states of multiple components below it, combined through a k / n gate. Ultimately, this is used to construct a complete MFTA graph or logical expression.

[0044] In some embodiments, the processor can map the constructed MFTA structure to the network structure of the DBN at the initial time slice (t=0). Event nodes in the MFTA correspond to random variable nodes in the DBN, and the state space of the nodes is consistent with the states defined in the MFTA. The logical gate relationships of the MFTA are transformed into directed arcs between DBN nodes, and the structure of the conditional probability table (CPT) is determined accordingly.

[0045] In some embodiments, the processor can quantify the conditional probability P(child node state | parent node state combination) of each node in the initial time slice DBN based on physical models, expert knowledge, statistical data, or the basic event probabilities of MFTA (if known). Nodes in the DBN whose states change over time are identified (e.g., nodes representing damage accumulation, performance degradation, crack propagation). A transition probability model P(node ​​state_t | node state_{t-1}, other relevant parent node states_{t or t-1}) between time slices is defined for these nodes. This model should reflect the underlying physical processes (e.g., determining the transition probability of crack states based on fatigue cumulative damage theory). Identified key risk factors are used as nodes in the DBN (evidence nodes or parameter nodes affecting CPT / transition probabilities). For example, "weld quality" (state: good / medium / poor) can affect the prior probability of "initial crack state of weld" and the transition probability of crack propagation.

[0046] S4: Using a dynamic Bayesian network model and a hybrid surrogate model, a two-stage quantitative reliability assessment of the simulation results is performed to obtain hierarchical reliability assessment results, thus completing the two-stage reliability assessment of the liquid hydrogen storage tank structure system; the hybrid surrogate model is obtained through training.

[0047] In some embodiments, the processor can construct a hybrid surrogate model (hybrid Kriging surrogate model) that efficiently connects a thermo-coupled physical model with a probabilistic reliability model. The input parameters (such as material properties, loads, boundary conditions) or output responses (such as stress or temperature at a specific location) of the thermo-coupled model are used as input, and the occurrence probabilities of basic events in the MFTA or the conditional probabilities of corresponding nodes in the DBN are used as output. The hybrid Kriging model is trained by obtaining training samples through running a small number of thermo-coupled model simulations (compared to the number required for direct MCS). The hybrid surrogate model (e.g., combining Gaussian processes and radial basis functions) aims to balance local accuracy and global trend fitting capability, improving the surrogate efficiency and accuracy for complex physical responses.

[0048] In some embodiments, the hybrid agent model includes: The global trend regression module is used to map and linearly combine the input vector to be predicted to obtain the global trend prediction result. The local bias correction module is used to calculate the residuals between the input vector to be predicted and the training dataset to obtain the local bias correction value for the global trend prediction. The output layer is used to comprehensively analyze the global trend prediction results and the local deviation correction values ​​of the global trend prediction to obtain the prediction values ​​of the hybrid surrogate model.

[0049] In some embodiments, the input vector x∈Rd of the hybrid proxy model consists of a set of key parameters that can significantly affect the reliability of the storage tank, where d is the dimension of the input parameters. This vector contains one or a combination of the following two types of parameters: ① External input parameters that directly affect the physical response: the magnitude of external loads (such as wind pressure, seismic acceleration), ambient temperature, internal medium pressure, material property parameters (such as elastic modulus, yield strength), geometric defect size, etc. ② Key physical response parameters calculated by the thermo-mechanical coupling model of the first stage: the maximum stress amplitude of a specific high-stress area on the storage tank structure, the highest operating temperature, the strain rate of key welds, the heat flux density of the vacuum interlayer, etc.

[0050] In some embodiments, the output scalar y of the hybrid proxy model is a probability value representing the likelihood of a specific failure event occurring. ① The probability of a basic event occurring in the MFTA model constructed in the third stage. ② The conditional probability of a leaf node (root node) given evidence in the DBN model constructed in the fourth stage.

[0051] In some embodiments, the processor can input vectors x Through a set of predefined basis functions f ( x )=[ f 1( x ), f 2( x ),..., f p ( x )] T Mapped to a p 3D eigenvectors. For example, for a second-order polynomial trend, The feature vector is then compared with a previously trained feature vector. p Dimensional weight (regression coefficient) vector β Perform linear combinations.

[0052] In some embodiments, the processor can process the residuals of the training data (the difference between the true value and the global trend prediction). (Where F is an n×p basis function matrix) is modeled as a zero-mean Gaussian process. Based on the conditional distribution theory of Gaussian processes, the residual information of the training data is used to predict the residual at the new point x.

[0053] In some embodiments, the expression for the global trend prediction result is: ; In some embodiments, the expression for the local deviation correction value of the global trend forecast is: ; In some embodiments, the expression for the predicted value of the hybrid surrogate model is: ; ; ; in, This indicates the overall trend prediction result; represents the input point to be predicted; j represents the index of the basis function, from 1 to p; p represents the total number of basis functions in the global trend model; This represents the regression coefficient corresponding to the j-th basis function; This represents the j-th basis function acting on the input point x; This represents a row vector consisting of all basis functions; This represents a p-dimensional column vector consisting of all regression coefficients; This represents the local deviation correction value for the global trend prediction, calculated by the Gaussian process model; This represents the covariance between the new point x and all training points. This represents the inverse of the covariance matrix K between training sample points; This represents the output vector of the training sample; Let p represent the n×p matrix formed by the training input samples under the basis functions. This represents the best estimate of the regression coefficients. This represents the final predicted value of the hybrid proxy model for input point x. This represents the transpose of the design matrix F.

[0054] In some embodiments, the training process of the hybrid agent model includes: Key parameters affecting the reliability of storage tanks are sampled, and each training input point is simulated using a thermo-mechanical coupling analysis model to obtain the physical response field. A failure criterion function is established based on the physical failure mechanism, and the event probability is calculated on the physical response field to obtain a training sample set containing multiple sample pairs. Using a training sample set containing multiple sample pairs, a numerical optimization algorithm is used to find the parameter combination that maximizes the probability of the training data appearing. The coefficients of the global trend function and the hyperparameters of the Gaussian process covariance function of the hybrid agent model are optimized to obtain the trained hybrid agent model.

[0055] In some embodiments, the processor can generate a set of representative training sample pairs. Sampling is performed within the d-dimensional input parameter space using a Design of Experiments (DOE) method, and a set of training points {x1, x2, …, x} with good space-filling properties is generated using the Latin hypercube sampling (LHS) method. n} where n is the total number of training samples. For each training input point xi: ① Use it as input and run the high-precision thermo-mechanical coupling physical model (such as a finite element model) established in the first stage to perform a complete simulation calculation, obtaining the detailed physical response field corresponding to the input point. ② Extract the key physical response quantities related to the predetermined failure mode from the response field, denoted as vector zi. ③ Apply a predefined failure criterion function or probability transformation function g(zi) to calculate the output event probability y corresponding to the physical response. i Ultimately, this results in a training dataset containing n sample pairs. .

[0056] The failure criterion function is established based on physical failure mechanisms: ① If the failure mode is material yielding, the probability of stress exceeding the limit can be calculated based on the statistical distribution of stress value and material yield strength. ② If the failure mode is fatigue crack propagation, the probability of crack propagation to the critical size can be calculated by integrating Paris's law.

[0057] In some embodiments, the processor may choose a hybrid Kriging model as the surrogate model. The advantage of this model lies in its structure, which combines a global deterministic trend function with a local stochastic process (Gaussian process), allowing for excellent fitting of complex global nonlinear trends and local detail fluctuations. i. Model Form Selection: ① Global Trend Function: Based on prior knowledge of the problem, a global multinomial function or a radial basis function (RBF) network can be selected. ② Local Stochastic Process: A zero-mean Gaussian process is used, whose characteristics are determined by the covariance function (or kernel function). A Gaussian kernel function is preferred due to its good smoothness.

[0058] In some embodiments, the goal of training the hybrid Kriging model is to determine the unknown parameters of the model using the training dataset D, including the coefficients of the global trend function and the hyperparameters (such as variance and length scale) of the Gaussian process covariance function. Maximum likelihood estimation (MLE) is preferably employed, using a numerical optimization algorithm to find the parameter combination that maximizes the probability of occurrence in the training data.

[0059] In some embodiments, K-fold cross-validation is used to evaluate the generalization ability and accuracy of the trained model. The training set is randomly divided into K subsets, and K-1 subsets are used for training in turn, while the remaining subset is used for testing. This process is repeated K times, and the average error metric (such as root mean square error RMSE) is used as the evaluation of the final accuracy of the model.

[0060] In some embodiments, the expression for the loss function of the hybrid agent model is: ; ; ; in, Let represent the negative log-likelihood function, which is the objective function for model training and optimization; Represents the hyperparameters of the kernel function; This represents the process standard deviation of a Gaussian process. This represents the complete training dataset; Indicates the number of training samples; Represents the determinant operation of a matrix; Indicates by hyperparameters θ The correlation matrix of the decision; Indicates all n A column vector consisting of the true output values ​​of each training sample; Let n×p represent the design matrix formed by all training input samples under the basis functions; This represents the best estimate of the regression coefficients. Let represent the condensed negative log-likelihood function, which is a function only of the hyperparameter θ; This represents the maximum likelihood estimate of the process variance.

[0061] In some embodiments, the training process employs numerical optimization algorithms such as gradient descent, conjugate gradient, or genetic algorithms to find the negative log-likelihood function. The hyperparameter θ* that reaches its minimum value. Once the optimal hyperparameter is determined, all other parameters of the model... This was also determined; among them, ( K ij = σ 2 R ij ).

[0062] In some embodiments, the processor can identify three main levels of evaluation work: Component level: focusing on the failure probability of the material itself (e.g., yielding, fracture) and welds (e.g., fatigue, defect propagation). Part level: focusing on the functional failure probability of components composed of parts, such as the thermal insulation performance of vacuum jackets and the load-bearing capacity of supporting structures. System level: focusing on the functional failure probability of the overall system composed of parts, such as the sealing performance, pressure resistance, and stability of storage tanks. Ensuring information flow: When designing the evaluation process, it is necessary to ensure that the failure probability information at lower levels can be used as input to influence the state evaluation at higher levels (reflected in the DBN model).

[0063] The hierarchical reliability assessment results spatially decompose the tank reliability assessment into three interrelated levels: component level (such as the fatigue failure probability of welds and the thermal conductivity of insulation materials), part level (such as the vacuum retention capacity of the vacuum jacket and the load-bearing capacity of the supporting structure), and system level (such as the overall sealing performance, pressure bearing capacity, and insulation efficiency of the tank). This ensures that the assessment covers all critical levels and their interactions.

[0064] In some embodiments, the processor can utilize a pre-constructed DBN model, combined with the real-time or near-real-time mapping from physical parameters to probabilistic parameters provided by a hybrid Kriging proxy model, to perform quantitative reliability assessments. The assessment process is characterized by a "dual-flow" approach: Structural flow: assessing the probability of physical failure of components and parts (e.g., stress exceeding limits, fatigue life exhaustion), these results can serve as inputs to the system flow. System flow: assessing the probability of failure of system-level functions, comprehensively considering the functional states and logical relationships of each component, as well as the impact of structural failure on system functionality. Using a forward inference algorithm (e.g., a forward algorithm) of the DBN, the probability of occurrence of the system's top event at different time points is calculated, yielding the reliability curve of the storage tank over time.

[0065] In some embodiments, the processor can use a dynamic Bayesian network model to analyze the simulation results and obtain the physical parameters of the current time step; By using a hybrid proxy model, the physical parameters of the current time step are analyzed to obtain the conditional probabilities of each basic event at the current time step. By using the conditional probabilities of each basic event at the current time step, the state of the corresponding basic event node in the dynamic Bayesian network model is updated to obtain the updated dynamic Bayesian network model. By using the updated dynamic Bayesian network model, forward reasoning is performed on the simulation results to obtain hierarchical reliability assessment results, thus completing the dual-process reliability assessment of the liquid hydrogen storage tank structure system.

[0066] In some embodiments, the processor can perform sensitivity analysis or importance analysis based on the DBN evaluation results, quantify the impact of each key failure risk factor on the overall or specific failure mode reliability of the tank, identify the main risk sources and weak links, and reveal the spatiotemporal evolution characteristics of the impact of risk factors.

[0067] In some embodiments, the processor can set the DBN to a state at an initial time t=0. For each time step t within the evaluation time range: obtain the physical parameters of the current time step: from a synchronously running dynamic thermo-coupling simulation, or based on a preset running profile; invoke the surrogate model: input the obtained physical parameters into the hybrid Kriging model trained in the fifth stage to quickly predict the conditional probabilities of each basic event at the current time step; update the DBN evidence: use the probabilities predicted by the surrogate model as evidence to update the state of the corresponding basic event node in the DBN; perform DBN inference: run the forward inference algorithm of the DBN (such as the forward algorithm, the connection tree algorithm, or an approximate inference algorithm suitable for large networks such as particle filtering) to calculate the state probability distribution of all nodes (including the top event node) in the DBN at the current time step t; reflect dual-process coupling: ensure that in DBN inference, the failure probability of structural level nodes (such as the probability of support fracture) can correctly affect the probability of system level functional nodes (such as overall stability) through the connection relationship of the DBN. Record the probability of the top event (system failure) at each time step to form a curve of the system failure probability changing over time, which is the dynamic reliability / unreliability assessment result of the storage tank. Simultaneously, the changes in the failure probability of each subsystem and component over time can also be obtained.

[0068] In some embodiments, the processor can systematically change the node parameters (such as prior probabilities and CPT entries) representing key risk factors in the DBN, observe the degree of change in the failure probability of the top event, and thus determine which risk factors have the greatest impact on system reliability. The processor calculates the probability contribution of each basic or intermediate event in the DBN to the occurrence of the top event (such as Birnbaum importance, criticality, etc.) to identify "weak links" in the system. The DBN model is used to simulate reliability performance under different operating strategies, maintenance measures, or unexpected event scenarios to support decision-making. Based on the above analysis results, the processor clarifies the key risk sources, main failure paths, and the trend of risk evolution over time, and proposes targeted design improvements, risk control, or operation and maintenance recommendations.

[0069] In some embodiments of this specification, a dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model is provided. Combining the advantages of MFTA, DBN, and hybrid Kriging, a comprehensive, accurate, and efficient quantitative assessment of the structure-system coupling, multi-level, and dynamic reliability of large liquid hydrogen storage tanks is achieved. The assessment results can be directly used to guide the design optimization, operation monitoring, maintenance strategy formulation, and remaining life prediction of the storage tank, thereby improving the intrinsic safety level and economy of large liquid hydrogen storage tanks. (1) By integrating system reliability (MFTA, DBN) and structural reliability (proxy model linking physical model) methods and considering multi-physics coupling, a more comprehensive and accurate assessment of the complex failure behavior of large liquid hydrogen storage tanks is achieved. (2) The innovative introduction of DBN enables the model to quantitatively describe time-varying processes such as damage accumulation and performance degradation, overcoming the limitations of static models and achieving dynamic assessment of the reliability of the storage tank throughout its entire life cycle. (3) The application of MFTA allows for modeling of multiple intermediate failure states of components, which is more in line with the gradual development process of failure in engineering practice, significantly improving the expressive power of the model and the precision of the assessment. (4) The self-constructed hybrid Kriging proxy model significantly reduces the number of calls to time-consuming physical simulations, greatly improving computational efficiency while ensuring accuracy, making it possible to conduct quantitative assessments of the entire lifecycle of large and complex systems. (5) The clear hierarchical decomposition (component-part-system) and structure-system dual-process assessment make the assessment logic clear, enabling a systematic consideration of failure transmission and impact between different levels, and realizing an integrated assessment from physical failure to system functional failure. (6) It not only provides quantitative reliability indicators, but also identifies key risk factors and their impact patterns through DBN analysis capabilities, providing strong scientific support for the optimized design, risk warning, and maintenance decisions of storage tanks.

Claims

1. A dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model, characterized in that, include: S1: Using a thermo-coupling analysis model, the specific structure and operating conditions of a large liquid hydrogen storage tank are analyzed, and simulation results are obtained. S2: Using simulation results, combined with expert knowledge, historical data and failure physics analysis, identify the key factors affecting the reliability of the storage tank and obtain the key failure risk factors; S3: Construct key failure risk factors into a polymorphic fault tree structure, and map the polymorphic fault tree structure to obtain a dynamic Bayesian network model. S4: Using a dynamic Bayesian network model and a hybrid surrogate model, a two-stage quantitative reliability assessment of the simulation results is performed to obtain hierarchical reliability assessment results, thus completing the two-stage reliability assessment of the liquid hydrogen storage tank structure system; the hybrid surrogate model is obtained through training.

2. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, S1 includes: The specific structure and operating conditions of a large liquid hydrogen storage tank are analyzed to obtain the corresponding mesh-divided three-dimensional geometric model. By using a thermo-coupling analysis model, the three-dimensional geometric model of the mesh is processed by setting boundary conditions, and multi-physics coupling is performed to obtain simulation results.

3. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, S2 includes: By using simulation results, combined with expert knowledge, historical data, and failure physics analysis, the possible failure modes of large liquid hydrogen storage tanks are identified, and the failure mode identification results are obtained. Based on the failure mode identification results, risk factors are traced and identified to obtain the key factors affecting the reliability of the storage tank. The key factors affecting the reliability of storage tanks are quantified and screened to identify key failure risk factors.

4. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 3, characterized in that, The key factors affecting the reliability of storage tanks are quantified and screened to identify the key failure risk factors, which include: The key factors affecting the reliability of storage tanks are quantified. For risk factors with clear physical properties and direct measurability, parameterization is performed. For process-related, state-related, or risk factors that are difficult to describe directly with a single numerical value, qualitative classification is performed to obtain the quantification results of key factors. Based on the quantitative results of key factors, the risk priority number of key factors is calculated using the expert evaluation method to obtain the first group of candidate key risk factors. Based on the quantification results of key factors, the standardized local sensitivity index was calculated using the preliminary judgment method of sensitivity analysis to obtain the second group of candidate key risk factors. By combining the first group of candidate key risk factors and the second group of candidate key risk factors, a review list is constructed to obtain the key failure risk factors.

5. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, S3 includes: Based on key failure risk factors, by defining the top event and system decomposition, identifying event states, selecting logic gates and constructing a tree structure, a polymorphic fault tree structure is obtained. An initial mapping is performed from the polymorphic fault tree structure to the dynamic Bayesian network model to obtain the conditional probability table structure. Based on the conditional probability table structure, the conditional probability of each node in the initial time slice dynamic Bayesian network is quantified, time dependence is introduced, and key failure risk factors are taken as nodes of the dynamic Bayesian network to obtain the dynamic Bayesian network model.

6. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, The hybrid agent model includes: The global trend regression module is used to map and linearly combine the input vector to be predicted to obtain the global trend prediction result. The local bias correction module is used to calculate the residuals between the input vector to be predicted and the training dataset to obtain the local bias correction value for the global trend prediction. The output layer is used to comprehensively analyze the global trend prediction results and the local deviation correction values ​​of the global trend prediction to obtain the prediction values ​​of the hybrid surrogate model.

7. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 6, characterized in that, The expression for the global trend prediction result is: ； The expression for the local deviation correction value of the global trend prediction is: ； The expression for the predicted value of the hybrid proxy model is: ； ； ； in, This indicates the overall trend prediction result; This represents the input point to be predicted; j The index represents the basis function, from 1 to p; p represents the total number of basis functions in the global trend model; Indicates the first j The regression coefficients corresponding to each basis function; Indicates action on the input point x The first j One basis function; This represents a row vector consisting of all basis functions; It represents the combination of all regression coefficients. p 3D column vector; This represents the local deviation correction value for the global trend prediction, calculated by the Gaussian process model; Indicates a new point x Covariance with all training points Represents the covariance matrix between training sample points K The inverse matrix; This represents the output vector of the training sample; Let p represent the n×p matrix formed by the training input samples under the basis functions. This represents the best estimate of the regression coefficients. This indicates that the hybrid proxy model applies to the input point. x The final predicted value, Design matrix F The transpose of .

8. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, The training process of the hybrid agent model includes: Key parameters affecting the reliability of storage tanks are sampled, and each training input point is simulated using a thermo-mechanical coupling analysis model to obtain the physical response field. A failure criterion function is established based on the physical failure mechanism, and the event probability is calculated on the physical response field to obtain a training sample set containing multiple sample pairs. Using a training sample set containing multiple sample pairs, a numerical optimization algorithm is used to find the parameter combination that maximizes the probability of the training data. The coefficients of the global trend function and the hyperparameters of the Gaussian process covariance function of the hybrid agent model are optimized to obtain the trained hybrid agent model.

9. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, The expression for the loss function of the hybrid agent model is: ； ； ； in, Let represent the negative log-likelihood function, which is the objective function for model training and optimization; Represents the hyperparameters of the kernel function; This represents the process standard deviation of a Gaussian process. This represents the complete training dataset; Indicates the number of training samples; Represents the determinant operation of a matrix; Indicates by hyperparameters θ The correlation matrix of the decision; Indicates all n A column vector consisting of the true output values ​​of each training sample; Let n×p represent the design matrix formed by all training input samples under the basis functions; This represents the best estimate of the regression coefficients. Let represent the condensed negative log-likelihood function, which is a function only of the hyperparameter θ; This represents the maximum likelihood estimate of the process variance.

10. The dual-process reliability assessment method for liquid hydrogen storage tank structural systems based on a hybrid model according to claim 1, characterized in that, S4 includes: The simulation results are analyzed using a dynamic Bayesian network model to obtain the physical parameters at the current time step; By using a hybrid proxy model, the physical parameters of the current time step are analyzed to obtain the conditional probabilities of each basic event at the current time step. By using the conditional probabilities of each basic event at the current time step, the state of the corresponding basic event node in the dynamic Bayesian network model is updated to obtain the updated dynamic Bayesian network model. By using the updated dynamic Bayesian network model, forward reasoning is performed on the simulation results to obtain hierarchical reliability assessment results, thus completing the dual-process reliability assessment of the liquid hydrogen storage tank structure system.

Images