Liquid hydrogen filling parameter identification method fusing fuzzy classification and generative adversarial network

Abstract

The application discloses a liquid hydrogen filling parameter identification method fusing fuzzy classification and a generative adversarial network, and relates to the technical field of parameter identification of a space filling system. Firstly, a low-temperature liquid hydrogen filling mechanism coupling model is constructed to determine a set of to-be-identified parameters and input and output variables. A piecewise linear membership function is constructed based on a differential pressure of a conveying pipeline and a filling flow rate, and a filling process is divided into three working conditions, i.e., a large-flow filling, a deceleration filling and automatic supplementing, through fuzzy classification. The liquid hydrogen filling mechanism coupling model is taken as a core calculation model of a generative adversarial network generator, a reward signal is constructed in combination with discriminator output, a policy gradient and Monte Carlo sampling, and the to-be-identified parameters are corrected through the alternation iteration training of the generator and the discriminator. After the convergence condition is met, the model parameters after identification are output. The application realizes data-mechanism fusion of liquid hydrogen filling parameter identification, and can improve the calculation accuracy of a model on state variables such as filling pressure and flow rate and the adaptability of the model under different filling working conditions.

Description

Technical Field

[0001] This invention relates to the field of parameter identification technology for aerospace refueling systems, and in particular to a method for identifying liquid hydrogen refueling parameters by fusing fuzzy classification and generative adversarial networks. Background Technology

[0002] Cryogenic liquid hydrogen refueling at space launch sites is a crucial step in launch vehicle missions, and the operational stability and control precision of this process directly impact the reliability and safety of the launch mission. To conduct simulation analysis, state prediction, and operational monitoring of the refueling process, it is typically necessary to establish a liquid hydrogen refueling model that reflects the dynamic characteristics of the actual refueling process. The liquid hydrogen refueling system involves multiple coupled processes, including pressurization of the ground storage tank's gas pillow, vaporization and reflux in the vaporizer, flow in the cryogenic delivery pipeline, changes in the opening of regulating valves, and gas pillow venting from the onboard storage tank. It is characterized by nonlinearity, multivariable coupling, significant changes in operational phases, and inconsistent sampling periods across multiple devices.

[0003] In liquid hydrogen refueling models, parameters such as the flow characteristic parameters of the vaporization branch valves, the vaporization flow correction coefficient, the friction coefficient of the delivery pipeline, the local resistance coefficient of the regulating valve, the discharge coefficient of the onboard tank, and the comprehensive resistance correction term are typically difficult to obtain accurately from field sensors. These parameters affect the model's calculations of state variables such as refueling pressure, flow rate, liquid level, and pipeline pressure difference; therefore, they need to be identified and corrected using actual sampled data. Accurately identifying these key parameters is a crucial foundation for improving the simulation accuracy and state prediction capabilities of liquid hydrogen refueling models.

[0004] Model parameter identification is a method to obtain unknown model parameters by minimizing the residual between the model output and the actual measured values, given the model structure. Essentially, model parameter identification is an optimization problem, seeking optimal values ​​for unknown parameters based on actual measurement data to ensure that the calculated values ​​of the model are as consistent as possible with the actual values. Currently, optimization methods for nonlinear model structure parameter identification include genetic algorithms and particle swarm optimization algorithms. However, for liquid hydrogen refueling models with complex nonlinearity and strong multivariate coupling, the identification performance is easily affected by the number of samples, sample coverage, and data quality. Complex identification optimization problems are prone to getting trapped in local optima, leading to low convergence accuracy and difficulty in convergence, resulting in inaccurate parameter identification for liquid hydrogen refueling models. Furthermore, for the liquid hydrogen refueling process, the frequency of refueling tasks is relatively limited, and the data distribution varies across different refueling stages, causing single data fitting methods to suffer from unstable parameter estimation or insufficient model adaptability under multiple operating conditions. In recent years, Generative Adversarial Networks (GANs), with their ability to learn sample distributions and generate simulated data, have provided a new approach for parameter identification in complex systems. However, the liquid hydrogen refueling process has different operating stages, such as high-flow refueling, deceleration refueling, and automatic replenishment. The changes in pressure difference and flow rate are different in different stages. If the operating conditions are not distinguished, the stability of parameter identification may be affected.

[0005] In summary, existing methods for identifying liquid hydrogen refueling parameters still have certain shortcomings in terms of maintaining physical constraints, adapting to multiple refueling stages, processing discrete time series data, and the stability of parameter identification under conditions of insufficient sample coverage. These shortcomings make it difficult to fully meet the needs of model correction, simulation analysis, and state prediction in the liquid hydrogen refueling process.

[0006] Therefore, it is necessary to propose a method for identifying liquid hydrogen refueling parameters that integrates fuzzy classification and generative adversarial networks. Summary of the Invention

[0007] To address the aforementioned issues, this application proposes a liquid hydrogen refueling parameter identification method that integrates fuzzy classification and generative adversarial networks. This method aims to solve the problem that existing liquid hydrogen refueling parameter identification technologies cannot simultaneously achieve physical interpretability, multi-condition adaptability, discrete temporal adaptability, and identification accuracy for scarce samples, thus failing to meet the requirements for high-fidelity modeling and safe operation control in the aerospace cryogenic liquid hydrogen refueling process.

[0008] On the one hand, this application proposes a method for identifying liquid hydrogen refueling parameters that integrates fuzzy classification and generative adversarial networks, including the following steps: S1. Construct a coupled model of cryogenic liquid hydrogen refueling mechanism and determine the model input variables, output variables and parameter set to be identified; The coupled model of the cryogenic liquid hydrogen refueling mechanism integrates the physical laws of the entire process, including the ground storage tank, vaporization branch, delivery pipeline, on-board tank, and venting circuit. It includes the liquid level change equation, the gas state equation of the storage tank cushion, the pipeline pressure difference coupling equation, the refueling flow rate equation, and the tank venting flow rate equation. The set of parameters to be identified includes vaporization flow correction coefficient, pipeline friction coefficient, regulating valve local resistance coefficient, and pipeline flow coefficient. The input variable is a historical time-series vector composed of the injection system pressure, flow rate, temperature, and valve opening; the output variable is a future time-series vector of the injection system pressure and flow rate. S2. Collect time-series observation data from sensors during the liquid hydrogen refueling process and construct a complete refueling data sequence; S3. Construct a fuzzy classification working condition identification model. Based on the two types of state variables, namely the pressure difference of the delivery pipeline and the injection flow rate, establish linear membership functions corresponding to the high flow injection, deceleration injection and automatic replenishment conditions respectively. Calculate the comprehensive membership degree of the high-flow-rate refueling condition, the deceleration refueling condition, and the automatic replenishment condition, and divide the refueling process into three operating conditions based on the principle of maximum membership degree: high-flow-rate refueling, deceleration refueling, and automatic replenishment. S4. Initialize the GAN network structure: The generator uses the liquid hydrogen refueling mechanism coupling model described in S1 as the core calculation model. After inputting time series data and parameters to be identified, it outputs a simulated refueling state sequence. The discriminator uses a temporal neural network structure to receive the real betting state sequence or the simulated betting state sequence output by the generator, and outputs the probability that the input sequence belongs to the real betting sample data. S5. Based on the annotated data sequence obtained in S2, the generator is first pre-trained, and then the discriminator is pre-trained by combining real samples and simulated samples output by the generator. S6. Alternately iterate the adversarial training of the generator and discriminator until the preset convergence condition is met, and finally output the target parameters to be identified after convergence within the generator. In the generator training phase, the policy gradient algorithm and Monte Carlo sampling are combined to construct a reward signal, and the parameters to be identified inside the coupled model are updated in reverse according to the reward signal. After each round of generator parameter updates, a new simulated sample is generated using the current generator, and the discriminator is retrained by combining it with the real sampled sample. The convergence condition includes at least one of the following: The discriminator output probability tends to stabilize, the generator loss function changes by less than a preset threshold, the iteration update amplitude of the parameters to be identified converges, and the time sequence error between the simulated state sequence and the actual sampling is less than a preset error threshold.

[0009] Furthermore, the expression for the coupled model of the cryogenic liquid hydrogen refueling mechanism in S1 is as follows: ; in, This refers to the liquid level in the surface storage tank. Add more flow to the main source. For the vaporization branch flow rate, For ground storage tanks at liquid level The corresponding liquid surface area; The liquid level in the rocket's onboard tank; The cross-sectional area of ​​the rocket's storage tank; This refers to the pressure of the air cushion in the ground storage tank. This refers to the volume of the air cushion in the ground storage tank. The compressibility factor is... The gas constant is... The temperature of the air cushion in the ground storage tank. This refers to the molar mass of hydrogen gas. The density of liquid hydrogen, The pressure of the air cushion in the rocket's storage tank; This refers to the volume of the air cushion in the rocket's storage tank. The compressibility coefficient of hydrogen gas inside the rocket's onboard storage tank; The temperature of the air cushion in the rocket's storage tank. The density of hydrogen gas inside the rocket's storage tank's gas pillow; The volumetric flow rate of the air cushion vented from the rocket's onboard storage tank. t For the time variable during the refueling process; The expression for the coupling relationship between sub-models in the coupled model of cryogenic liquid hydrogen refueling mechanism is as follows: ; ; ; ; ; ; ; ; in, This represents the total volume of the surface storage tanks; For ground storage tanks at liquid level The corresponding liquid hydrogen volume; This refers to the total volume of the arrow's onboard storage tanks; This refers to the inlet pressure of the delivery pipeline; It is the acceleration due to gravity; This refers to the outlet pressure of the delivery pipeline; The height difference between the bottom of the rocket's onboard storage tank and the bottom of the ground storage tank; Differential pressure can be used for the delivery pipeline; This refers to the cross-sectional area of ​​the pipeline. The diameter of the conveying pipeline; This refers to the length of the transport pipeline; This is the friction coefficient along the friction path; For the first The local resistance coefficient of a valve or local component; The sum of the local resistance coefficients of each valve and local component, Q is the main injection flow rate, and the state quantity of the air cushion in the ground storage tank is defined as... ， The air cushion state quantity of the rocket's onboard storage tank is defined as follows: .

[0010] Furthermore, in S3, regarding the pressure difference in the delivery pipeline... For the main refueling flow rate Q, establish piecewise linear membership functions corresponding to the high-flow refueling condition, the deceleration refueling condition, and the automatic replenishment condition, respectively.

[0011] For any state variable x, where ,set up Thresholds are assigned to the operating conditions corresponding to this state variable, where to This is the transition range from automatic refueling to deceleration refueling. to The main range for deceleration and refueling operations. to This is the transition range from deceleration refueling conditions to high-flow refueling conditions.

[0012] The automatic processing condition membership function is: ; The membership function for the deceleration and refueling condition is: ; The membership function for high-flow-rate refueling is: ; in, , and Representing state variables respectively x Membership degree of automatic replenishment condition, deceleration refueling condition, and high flow rate refueling condition.

[0013] For the pressure difference of the delivery pipeline ,available , and For the main betting flow Q, we can obtain , and .

[0014] Each threshold can be determined based on the liquid hydrogen refueling procedure settings, equipment operating parameters, test data, or expert experience.

[0015] The formula for calculating the membership degree of the three types of refueling conditions is as follows: ; ; ; in, To increase the overall membership degree of the operating conditions for high flow rates, To determine the overall membership of the deceleration and refueling operation, To automatically replenish the overall membership degree of the processing conditions; , , These represent the degree of membership of the pressure state to the three types of working conditions, respectively. , , These represent the membership degree of the flow state to the three types of operating conditions, respectively.

[0016] The operating condition determination rule is: when When the current refueling process is in automatic refueling mode, it is determined that the current refueling process is in automatic refueling mode; when When the current refueling process is determined to be in a deceleration refueling condition; when At that time, it is determined that the current refueling process is in a high-flow refueling condition.

[0017] Furthermore, after S3 divides the work conditions into three categories, during the training of the generative adversarial network, corresponding data intervals, initial parameter values, or model correction terms can be configured for different work conditions to reduce the impact of data differences at different infeeding stages on parameter identification results.

[0018] Furthermore, the generator has no independent pure neural network mapping layer, and uses the differential equations of the liquid hydrogen refueling mechanism as the core calculation model. All parameters to be optimized are interpretable physical parameters of cryogenic fluid engineering, and the generation simulation time series fully meets the thermodynamic and fluid dynamic constraints of liquid hydrogen.

[0019] Furthermore, the discriminator is constructed using any one of CNN, RNN, LSTM or Transformer temporal neural networks to extract dynamic features of pressure, flow, and temperature time series and to determine the source of the samples.

[0020] Furthermore, in S5, the generator pre-training adopts maximum likelihood estimation or minimum residual loss to enable the mechanistic model output time series to initially fit the real injection data, eliminating the oscillation non-convergence problem in adversarial training. The discriminator pre-training process involves labeling real sequences as positive samples and generating simulated sequences as negative samples. Optimization is achieved by minimizing the discriminator's cross-entropy loss, which is expressed as follows: ; Where, in the formula, This indicates that the data comes from a real data sample distribution. This indicates that the data comes from the generator's simulated sample distribution. Given an input variable (X) and an output sequence (Y), the discriminator determines the probability that the sequence belongs to the true sample data. To calculate the expected value of the distribution of simulated samples generated by the generator, To calculate the mathematical expectation of the distribution of the real sample data, For discriminator model parameters, These are the generator model parameters.

[0021] Furthermore, the expression for the objective function during generator training in S6 is: ; in, This represents the target sequence generated by the generator. The action-value function is the function of the discriminator in the state. At that time, according to the strategy Take action The resulting cumulative rewards; generated by the generator model Generate analog output sequence The generator generates results at each time t based on the results from the previous time. Generate current output S represents the operating condition label output by S3. The expression for the Monte Carlo action-value function is: ; Where C is the number of Monte Carlo samplings, b(˙) is the variance suppression baseline constant, D(˙) is the true probability output by the discriminator, n is the sequence number of the nth Monte Carlo sampling, and N is the number of Monte Carlo samplings. The generator is given a sequence of input variables, where t is the current generation time. The expression for parameter gradient update is: ; Where T is the total timing length. For the generator objective function with respect to parameters gradient, For the generator in the first t Output generated at time 1 Seeking expectations, To generate the current output for the generator given the generated sequence and input variables. The probability of.

[0022] Furthermore, Monte Carlo sampling generates N simulated betting sequences for the same historical time series, and takes the average of the discrimination probabilities of multiple sequences as the cumulative reward, thereby reducing the variance of the single-step time series reward and improving the stability of parameter identification under critical and abnormally scarce conditions.

[0023] On the other hand, this application proposes a liquid hydrogen refueling parameter identification device that integrates fuzzy classification and generative adversarial networks, comprising: Data acquisition module, storage module, working condition fuzzy classification module, mechanism modeling module, GAN training module, parameter output module; The data acquisition module is used to collect time-series data on pressure, flow rate, temperature, and valve opening of the liquid hydrogen refueling pipeline and transmit them to the storage module; The storage module stores the coupling model of the cryogenic liquid hydrogen refueling mechanism, the membership function, the GAN network weights, and historical observation data. The fuzzy classification module is used to execute the fuzzy classification logic of the refueling conditions. Based on the piecewise linear membership function of the pressure difference of the delivery pipeline and the refueling flow rate, it calculates the comprehensive membership degree of the high flow refueling condition, the deceleration refueling condition and the automatic replenishment condition respectively, and outputs the corresponding refueling condition label according to the principle of maximum membership degree. The mechanism modeling module constructs the coupled model of the differential equation system and defines the parameter set to be identified and the input-output time series vectors. The GAN training module has a built-in generator and discriminator. The generator uses the liquid hydrogen refueling mechanism coupling model as the core calculation model, and constructs a reward signal by combining the discriminator output, policy gradient and Monte Carlo sampling. The parameters to be identified are corrected by alternating iterative training of the generator and discriminator. When the convergence condition is met, the identified model parameters are output. The parameter output module outputs the updated parameters to be identified in the converged generator and uses them as the identification results of key parameters of the liquid hydrogen refueling model for liquid hydrogen refueling simulation and online monitoring. Each module is interconnected via industrial control bus or Ethernet communication, and the processor uniformly schedules and executes the complete identification process.

[0024] In summary, the liquid hydrogen refueling parameter identification method of the present invention, which integrates fuzzy classification and generative adversarial networks, has the following advantages compared with traditional technologies: 1. Improved the rationality of parameter identification in the refueling model. Using the refueling process mechanism model as the core of the generator, valve characteristic parameters, pipeline resistance parameters, vaporization correction parameters, etc. are used as parameters to be identified. Through adversarial training between real sampled data and generator simulated data, the model parameters are iteratively corrected, making the model output results closer to the actual refueling process data, and improving the accuracy of calculation and prediction of state variables such as pressure and flow rate. 2. The model's adaptability to different operating conditions has been enhanced. By introducing fuzzy classification, the operating conditions during the refueling process are determined based on key state variables such as differential pressure and flow rate. Appropriate data intervals, initial parameter values, or model correction terms are selected according to different operating conditions. This reduces the impact of operating condition changes on parameter identification results, making the model more adaptable to both normal and fluctuating operating conditions, and providing a reference for simulation analysis and operational monitoring of the refueling process. 3. Provides corrected model parameters for refueling process prediction and simulation analysis. Through alternating training of the generator and discriminator, the parameters to be identified in the generator are finally output as corrected parameters for the refueling model. Substituting these parameters into the refueling process mechanism model, they can be used for the calculation, prediction, and simulation analysis of state variables such as pressure and flow rate, providing a reference for subsequent refueling process operation monitoring.

[0025] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] Figure 1 A schematic diagram of a generative adversarial network model; Figure 2 A schematic diagram illustrating the implementation steps of a generative adversarial algorithm for parameter-oriented identification. Detailed Implementation

[0027] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0028] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0029] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0030] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0031] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0032] This application achieves the identification and correction of key parameters of the model through an adversarial learning process of "generation-discrimination-feedback", which utilizes the ideas of generative adversarial game and parameter adaptive optimization.

[0033] First, a game-theoretic framework is constructed between the generator and the discriminator. The generator takes the injection process mechanism model as its core and generates a simulated output sequence based on time series data such as flow rate, pressure, temperature, and valve opening collected by sensors. The discriminator is used to distinguish between the real sampled data and the simulated data output by the generator, so as to constrain the consistency between the generator output and the actual injection process.

[0034] Meanwhile, a parameter identification module is embedded in the generator, which, based on the dynamic comparison between real and simulated data during adversarial training, corrects key parameters in the model such as valve characteristic parameters, pipeline resistance parameters, and vaporization correction parameters.

[0035] In addition, by leveraging the adversarial training characteristics of GANs, the dynamic changes in the data during the injection process can be further utilized to improve the adaptability of model parameter estimation to different operating conditions and alleviate the problem of parameters being difficult to fit accurately in complex injection scenarios.

[0036] Finally, by leveraging the iterative optimization capabilities of GANs, the model output quantities such as pressure or flow rate output by the generator are gradually made closer to the real sampled data, thereby realizing the identification and correction of the added model parameters.

[0037] Example 1 This embodiment applies to a cryogenic liquid hydrogen refueling system at a space launch site. The system includes a ground-based liquid hydrogen storage tank, a vaporization reflux branch, a cryogenic delivery pipeline, a flow control valve, an onboard liquid hydrogen storage tank, and a tank venting circuit. It is used to accurately identify key resistance and flow correction parameters in the refueling mechanism model, providing support for digital twin simulation and operational status monitoring of the refueling process. Figure 2 As shown, the specific implementation steps are as follows: Step 1: Construct a coupled model of the cryogenic liquid hydrogen refueling mechanism and determine the parameters to be identified and the input and output variables.

[0038] By integrating the state equations for the ground-based storage tank's air cushion, the onboard storage tank's air cushion, the liquid level change equation, the pipeline pressure difference equation, and the delivery pipeline flow rate equation, a dynamic coupled model of the cryogenic liquid hydrogen refueling process can be constructed. This model uses the liquid level in the ground-based storage tank, the liquid level in the onboard storage tank, and the state variables of the ground-based storage tank's air cushion and the onboard storage tank's air cushion as the main state variables, and incorporates the pipeline pressure difference equation. and adding flow Achieve coupling between the sub-models.

[0039] The system of differential equations is as follows: ; in This refers to the liquid level in the surface storage tank. Add more flow to the main source. For the vaporization branch flow rate, For ground storage tanks at liquid level The corresponding liquid surface area; The liquid level in the rocket's onboard tank; The cross-sectional area of ​​the rocket's storage tank; This refers to the pressure of the air cushion in the ground storage tank. This refers to the volume of the air cushion in the ground storage tank. The compressibility factor is... The gas constant is... The temperature of the air cushion in the ground storage tank. This represents the molar mass of hydrogen gas. The density of liquid hydrogen, The pressure of the air cushion in the rocket's storage tank; This refers to the volume of the air cushion in the rocket's storage tank. The compressibility coefficient of hydrogen gas inside the rocket's onboard storage tank; The temperature of the air cushion in the rocket's storage tank. The density of hydrogen gas inside the rocket's storage tank's gas pillow; The volumetric flow rate of the air cushion vented from the rocket's storage tank.

[0040] The coupling relationship between the sub-models is as follows: ; ; ; ; ; ; ; ; in, This represents the total volume of the surface storage tanks; For ground storage tanks at liquid level The corresponding liquid hydrogen volume; This refers to the total volume of the arrow's onboard storage tanks; This refers to the inlet pressure of the delivery pipeline; It is the acceleration due to gravity; This refers to the outlet pressure of the delivery pipeline; The height difference between the bottom of the rocket's onboard storage tank and the bottom of the ground storage tank; Differential pressure can be used for the delivery pipeline; This refers to the cross-sectional area of ​​the pipeline. The diameter of the conveying pipeline; This refers to the length of the transport pipeline; This is the friction coefficient along the friction path; For the first The local resistance coefficient of a valve or local component; The sum of the local resistance coefficients of each valve and local component, Q is the main injection flow rate, and the state quantity of the air cushion in the ground storage tank is defined as... ， The air cushion state quantity of the rocket's onboard storage tank is defined as follows: .

[0041] Based on the mechanism model of liquid hydrogen refueling, parameters that are difficult to measure directly by sensors but affect the accuracy of the model output are selected as parameters to be identified. These parameters include flow coefficient, pipeline specific resistance coefficient, regulating valve resistance coefficient, and vaporization process correction coefficient, denoted as... .in, These represent the key parameters that need to be identified and corrected in the injection model.

[0042] The input variables for parameter identification include pipeline pressure at historical sampling times. , increase flow ,temperature Valve opening and sampling time , can be represented as: ; in, Indicates from Time to The pressure sequence at any given time, Indicates from Time to The injection flow sequence at any given time. Indicates from Time to Temperature sequence at time, Indicates from Time to The sequence of valve openings at different times. Indicates the current sampling time.

[0043] The output variable of parameter identification is the state variable that the model needs to fit, which can be pressure, flow rate, or temperature at future sampling times, and can be expressed as: ; Step 2: Collect observation data of the injection process and construct a training dataset.

[0044] The system collects observational data (time-series data from pressure sensors, flow sensors, temperature sensors, and valve position sensors) generated during the refueling process, covering the entire operation phase from pressurization and precooling to low-flow refueling, high-flow refueling, and replenishment / release. The raw data undergoes denoising, outlier removal, and normalization preprocessing, and is then divided into training samples according to fixed time windows to form the data sequence to be identified. ,in , This represents the observation vector of the system at the t-th sampling time. This represents the actual output value of the injection process at time t.

[0045] Step 3: Implement adaptive classification of refueling conditions based on fuzzy classification.

[0046] Based on the key state variables of the refueling system, fuzzy membership functions were designed using expert experience. During the refueling process, the main variables affecting the refueling conditions include pipeline pressure. , increase flow Pipeline temperature and valve opening All of the above variables can be collected by sensors or control systems and used to determine the current operating status of the refueling process.

[0047] For the pressure difference of the delivery pipeline For the injection flow rate Q, establish piecewise linear membership functions corresponding to high-flow injection, deceleration injection, and automatic replenishment conditions, respectively.

[0048] For any state variable x, where The threshold for classifying operating conditions is determined based on the refueling procedure settings, equipment operating parameters, test data, or expert experience. And satisfy .

[0049] Among them, a smaller value corresponds to automatic replenishment mode, a value in the middle range corresponds to deceleration refueling mode, and a larger value corresponds to high flow refueling mode.

[0050] The automatic processing condition membership function is: ; The membership function for the deceleration and refueling condition is: ; The membership function for high-flow-rate refueling is: ; in, , and Representing state variables respectively xMembership degree of automatic replenishment condition, deceleration refueling condition, and high flow rate refueling condition.

[0051] For the pressure difference of the delivery pipeline ,available , and For the main betting flow Q, we can obtain , and .

[0052] Each threshold can be determined based on the liquid hydrogen refueling procedure settings, equipment operating parameters, test data, or expert experience.

[0053] After calculating the membership degrees of the pipeline pressure difference and the injection flow rate to the three types of operating conditions separately, the minimum operator is used to calculate the comprehensive membership degree of the three injection conditions: ; ; ; in, To increase the overall membership degree of the operating conditions for high flow rates, To determine the overall membership of the deceleration and refueling operation, To automatically replenish the overall membership degree of the processing conditions; , , These represent the degree of membership of the pressure state to the three types of working conditions, respectively. , , These represent the membership degree of the flow state to the three types of operating conditions, respectively.

[0054] The current refueling condition is determined based on the principle of maximum membership: when At that time, it is determined that the current refueling process is in automatic refueling mode; when At that time, it is determined that the current refueling process is in a deceleration refueling condition; when At that time, it is determined that the current refueling process is in a high-flow refueling condition.

[0055] Based on the above fuzzy classification results, the corresponding data range, initial parameter value, or model correction term can be selected according to different operating conditions, so that the generator can be trained in combination with the corresponding mechanism model parameters under different operating conditions, reducing the problem of identification bias in the process of refueling under multiple operating conditions by a single model.

[0056] Step 4: Initialize and generate the adversarial network framework, such as... Figure 1 As shown.

[0057] Initialize the generator model Discriminator Model .

[0058] Generator Model Used to generate the output sequence of a simulation system based on random variables, its internal parameters The parameters to be identified in the corresponding system model. Generator model. Based on the mechanism model of the injection process, it is used to determine the input variables. and parameters to be identified Generate a simulated output sequence. The simulated sequence output by the generator can be represented as: ; in, This represents the state variables such as simulated pressure, simulated flow rate, or simulated temperature output by the generator.

[0059] Discriminator Model This is used to determine whether the input sequence comes from real sampled data or simulated data output by the generator. The discriminator's input includes the real output sequence. and the simulated sequence output by the generator Its output is the probability that the sequence belongs to the real sampled data. Through adversarial training between the generator and the discriminator, the parameters to be identified in the generator are continuously corrected. This allows the generator's output to gradually approximate the actual injection process data.

[0060] Step 5: Pre-training of the generator and discriminator.

[0061] Step 5.1: Based on the acquired data sequence X, in the initialization phase before the discriminator participates in training, the maximum likelihood estimation method or the residual loss minimization method is used to train the generator model. Pre-train the generator model. Based on the mechanism model of the filling process, a simulated output sequence is generated according to the pressure, flow rate, temperature, valve opening and parameters to be identified at historical sampling times.

[0062] The goal of pre-training is to make the simulated pressure, flow, or temperature sequences output by the generator statistically close to the real sampled data, so that the generator initially has the ability to fit the distribution of real filling process data, and provides stable initial conditions for subsequent adversarial training between the generator and the discriminator.

[0063] Step 5.2: Construct the feedback information required for generator training.

[0064] Based on pre-trained generator model Input historical sampling data and the current parameters to be identified, and generate a simulated output sequence. The simulated output sequence from the generator and the actual sampling sequence are then input together into the discriminator. The discriminator outputs the probability that the simulated sequence belongs to the real sampled data. This discriminant probability is used as the reward signal for generator training and is used to update the generator parameters in the future.

[0065] Step 5.3: Based on the given generator, pre-train the discriminator model.

[0066] Step 5.3.1: Given a generator model Under these conditions, training samples for the discriminator are constructed using real sampled data and simulated data output by the generator. Real sampled data includes output sequences of pressure, flow rate, and temperature collected by sensors; simulated data consists of simulated output sequences of pressure, flow rate, or temperature generated by the generator based on historical input variables and the current parameter to be identified. Real sampled sequences are labeled as real samples, and simulated sequences output by the generator are labeled as generated samples, used to train the discriminator model. .

[0067] Step 5.3.2, Discriminator Model This is used to determine whether the input sequence comes from real sampled data or simulated data output by the generator. As one implementation method, a discriminator model can be built using a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), or other neural network structures suitable for time series discrimination. If a CNN structure is used, local variation features in time series data such as pressure, flow rate, and temperature can be extracted through convolutional operations, thereby reducing model training complexity and improving sequence discrimination capabilities.

[0068] Step 5.3.3: During training, when the input sequence is real sampled data, the discriminator... The goal of the discriminator is to make the output probability D(x) approach 1; when the input sequence is the simulated data G(z) output by the generator G, the goal of the discriminator D is to correctly determine the data source and make D(G(z)) approach 0. Here, X represents the input variable sequence, Y represents the true output sequence, and Ŷ represents the simulated output sequence of the generator. The discriminator D is optimized by minimizing the discriminator loss function to obtain the optimal solution. ; In the formula, Y∈ This indicates that the data comes from a real data sample distribution. This indicates that the data comes from the generator's simulated sample distribution. Given an input variable (X) and an output sequence (Y), the discriminator determines the probability that the sequence belongs to the true sample data. To calculate the expected value of the distribution of simulated samples generated by the generator, To calculate the mathematical expectation of the distribution of the real sample data, For discriminator model parameters, These are the generator model parameters.

[0069] Step 5.4: Enter the adversarial training phase of the generator.

[0070] Based on the parameter identification of the refueling model, the generator G generates a simulated output sequence according to the input variables and the current parameter to be identified θ. The simulated output sequence includes refueling process state variables such as pressure and flow rate. The discriminator D is used to determine whether the input sequence comes from real sampled data or from simulated data generated by the generator. By combining GAN technology with the policy gradient method, the discriminator output is fed back to the generator as a reward signal, thereby updating the parameter to be identified in the generator. The steps in this stage are as follows: Step 5.4.1, the objective function of the generator can be expressed as: ; In the formula, This represents the target sequence generated by the generator. The action-value function is the function of the discriminator in the state. At that time, according to the strategy Take action The resulting cumulative rewards. (Generator model) Generate analog output sequence The generator generates results at each time t based on the results from the previous time. Generate current output .

[0071] Step 5.4.2: Analyze the generator's output at time t. Calculate the corresponding reward function: (1) The probability that the discriminator considers the generated sequence to belong to the true sample distribution is used as the reward value: ; In the formula, This represents the baseline value used to reduce the variance of reward values; it is usually set as a constant.

[0072] (2) The remaining unknown sequences are sampled using a Monte Carlo search strategy based on policy G to estimate the action value function of the intermediate state. The discriminator obtains N reward values ​​based on the sampled N sequences. Then, the Nth Monte Carlo search strategy can be expressed as: ; In the formula, This is the current state. Let G be the sequence length sampled based on policy G. The length of the sequence samples for the Monte Carlo search.

[0073] (3) If the final reward value of the intermediate state is the average of N reward values, then the action value function can be represented in two cases as follows: ; Where C is the number of Monte Carlo samplings, b(˙) is the variance suppression baseline constant, D(˙) is the true probability output by the discriminator, n is the sequence number of the nth Monte Carlo sampling, and N is the number of Monte Carlo samplings. The generator is given a sequence of input variables, where t is the current generation time.

[0074] Step 5.4.3: Based on the calculated reward function, the discriminator is dynamically updated to improve the generator's performance. The discriminator can also be further refined based on the generator's data sequence. After updating the discriminator, the generator is updated alternately. The generator's parameters are calculated by solving for its gradient according to its objective function. ; Where T is the total timing length. For the generator objective function with respect to parameters gradient, For the generator in the first t Output generated at time 1 Seeking expectations, To generate the current output for the generator given the generated sequence and input variables. The probability of.

[0075] Through the above steps, the generator continuously corrects the parameters to be identified based on the reward signal fed back by the discriminator, so that the simulated pressure, simulated flow and other sequences output by the generator gradually approach the real sampled data, thereby realizing the identification and correction of key parameters of the injection model.

[0076] Step 5.5: Enter the discriminator retraining process.

[0077] After the generator parameters are updated, the current generator model is used. The simulated injection output sequence is regenerated and combined with the real sampling data to form the discriminator training samples. The discriminator model is then retrained. By retraining the discriminator model, the discriminator can more accurately distinguish between the real sampled sequence and the simulated sequence output by the current generator, thereby providing a more accurate reward signal for the next round of generator parameter updates.

[0078] Step 6: Train the generator and discriminator alternately and output the recognition results.

[0079] Repeat steps 5.4 and 5.5 to train the generator model and discriminator model alternately. In each training round, the generator generates a simulated output sequence based on the input data and the current parameters to be identified. The discriminator judges the difference between the simulated output sequence and the real sampled sequence and uses the judgment result as a feedback signal to correct the parameters to be identified in the generator. .

[0080] Training stops when preset convergence conditions are met. These convergence conditions include, but are not limited to: the discriminator output stabilizing, the generator loss function changing less than a preset threshold, and the parameters to be identified... The update magnitude tends to stabilize, or the error between the generator output sequence and the real sampled sequence is less than a preset threshold.

[0081] After the convergence condition is met, output the generator model. Parameters in This serves as the final identification result for the key parameters of the injection model.

[0082] Example 2 This embodiment provides a generative adversarial algorithm step for identifying parameters of an injection model.

[0083] The training of generative adversarial networks (GANs) involves training both the generator model and the discriminator model, requiring alternating updates of both during training. Combining this with a cryogenic liquid hydrogen refueling mechanism model, the steps of a GAN algorithm for identifying refueling model parameters are as follows: Step (1) Collect observation data of the cryogenic liquid hydrogen refueling process to obtain the data sequence to be identified. ; Step (2) Determine the parameters to be identified based on the injection mechanism model. And determine the input and output variables; Step (3) Initialize the generator model Discriminator Model ; Step (4) Based on data sequence For generator models Pre-training was performed to enable it to initially generate simulated injection data; Step (5) Based on the given generator model Generating simulated data and combining it with real sampled data to improve the discriminator model. Perform pre-training; Step (6) involves training the generator model through the following three steps: 1. Generator Model Based on input variables and current parameters Generate a simulated output sequence; 2. Discriminator Model The simulated output sequence is discriminated, and the discrimination result is used as a reward signal; 3. Based on the reward signal, update the parameters to be identified in the generator model using the policy gradient method. ; Step (7) Retrain the discriminator model: Train the discriminator model together with the simulated data generated by the current generator model and the real sampled data. And update the discriminator parameters. ; Step (8) Repeat steps (6) and (7) to iteratively train the generator model and discriminator model alternately until the convergence condition is met, and output the parameters in the generator model. This serves as the final identification result for the key parameters of the injection model.

[0084] Example 3 This embodiment proposes a liquid hydrogen refueling parameter identification device that integrates fuzzy classification and generative adversarial networks, including: Data acquisition module, storage module, working condition fuzzy classification module, mechanism modeling module, GAN training module, parameter output module; The data acquisition module is used to collect time-series data on pressure, flow rate, temperature, and valve opening of the liquid hydrogen refueling pipeline and transmit them to the storage module; The storage module stores the coupling model of the cryogenic liquid hydrogen refueling mechanism, the membership function, the GAN network weights, and historical observation data. The fuzzy classification module is used to execute the fuzzy classification logic of the refueling conditions. Based on the piecewise linear membership function of the pressure difference of the delivery pipeline and the refueling flow rate, it calculates the comprehensive membership degree of the high flow refueling condition, the deceleration refueling condition and the automatic replenishment condition respectively, and outputs the corresponding refueling condition label according to the principle of maximum membership degree. The mechanism modeling module constructs the coupled model of the differential equation system and defines the parameter set to be identified and the input-output time series vectors. The GAN training module has a built-in generator and discriminator. The generator uses the liquid hydrogen refueling mechanism coupling model as the core calculation model, and constructs a reward signal by combining the discriminator output, policy gradient and Monte Carlo sampling. The parameters to be identified are corrected by alternating iterative training of the generator and discriminator. When the convergence condition is met, the identified model parameters are output. The parameter output module outputs the updated parameters to be identified in the converged generator and uses them as the identification results of key parameters of the liquid hydrogen refueling model for liquid hydrogen refueling simulation and online monitoring. Each module is interconnected via industrial control bus or Ethernet communication, and the processor uniformly schedules and executes the complete identification process.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for identifying liquid hydrogen refueling parameters by integrating fuzzy classification and generative adversarial networks, characterized in that, Includes the following steps: S1. Construct a coupled model of cryogenic liquid hydrogen refueling mechanism and determine the model input variables, output variables and parameter set to be identified; The coupled model of the cryogenic liquid hydrogen refueling mechanism integrates the physical laws of the entire process, including the ground storage tank, vaporization branch, delivery pipeline, on-board tank, and venting circuit. It includes the liquid level change equation, the gas state equation of the storage tank cushion, the pipeline pressure difference coupling equation, the refueling flow rate equation, and the tank venting flow rate equation. The set of parameters to be identified includes vaporization flow correction coefficient, pipeline friction coefficient, regulating valve local resistance coefficient, and pipeline flow coefficient. The input variable is a historical time-series vector composed of the injection system pressure, flow rate, temperature, and valve opening; the output variable is a future time-series vector of the injection system pressure and flow rate. S2. Collect time-series observation data from sensors during the liquid hydrogen refueling process and construct a complete refueling data sequence; S3. Construct a fuzzy classification working condition identification model. Based on two types of state variables, namely the pressure difference of the delivery pipeline and the injection flow rate, establish linear membership functions corresponding to the high flow injection, deceleration injection and automatic replenishment conditions respectively. The membership degree of the three working conditions is calculated based on the membership degree of the state variables, and the working condition to which the current refueling process belongs is determined according to the principle of the maximum membership degree of the comprehensive working condition. S4. Initialize the GAN network structure: The generator uses the liquid hydrogen refueling mechanism coupling model described in S1 as the calculation model, inputs time series data and parameters to be identified, and outputs a simulated refueling state sequence. The discriminator uses a temporal neural network structure to receive the real betting state sequence or the simulated betting state sequence output by the generator, and outputs the probability that the input sequence belongs to the real betting sample data. S5. Based on the annotated data sequence obtained in S2, the generator is first pre-trained, and then the discriminator is pre-trained by combining real samples and simulated samples output by the generator. S6. Alternately iterate the adversarial training of the generator and discriminator until the preset convergence condition is met, and finally output the target parameters to be identified after convergence within the generator. In the generator training phase, the policy gradient algorithm and Monte Carlo sampling are combined to construct a reward signal, and the parameters to be identified inside the coupled model are updated in reverse according to the reward signal. After each round of generator parameter updates, a new simulated sample is generated using the current generator, and the discriminator is retrained by combining it with the real sampled sample. The convergence condition includes at least one of the following: The discriminator output probability tends to stabilize, the generator loss function changes by less than a preset threshold, the iteration update amplitude of the parameters to be identified converges, and the time sequence error between the simulated state sequence and the actual sampling is less than a preset error threshold.

2. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 1, characterized in that, The expression for the coupled model of the cryogenic liquid hydrogen refueling mechanism in S1 is: ； in, This refers to the liquid level in the surface storage tank. Add more flow to the main source. For the vaporization branch flow rate, The liquid level in the rocket's onboard tank; The cross-sectional area of ​​the rocket's storage tank; This refers to the pressure of the air cushion in the ground storage tank. This refers to the volume of the air cushion in the ground storage tank. The compressibility factor is... The gas constant is The temperature of the air cushion in the ground storage tank. This refers to the molar mass of hydrogen gas. The density of liquid hydrogen, The pressure of the air cushion in the rocket's storage tank; This refers to the volume of the air cushion in the rocket's storage tank. The compressibility coefficient of hydrogen gas inside the rocket's onboard storage tank; The temperature of the air cushion in the rocket's storage tank. The density of hydrogen gas inside the rocket's storage tank's gas pillow; The volumetric flow rate of the air cushion vented from the rocket's onboard storage tank. t For the time variable during the refueling process; The expression for the coupling relationship between sub-models in the coupled model of cryogenic liquid hydrogen refueling mechanism is as follows: ； ； ； ； ； ； ； ； in, This represents the total volume of the surface storage tanks; For ground storage tanks at liquid level The corresponding liquid hydrogen volume; This refers to the total volume of the arrow's onboard storage tanks; This refers to the inlet pressure of the delivery pipeline; It is the acceleration due to gravity; This refers to the outlet pressure of the delivery pipeline; The height difference between the bottom of the rocket's onboard storage tank and the bottom of the ground storage tank; Differential pressure can be used for the delivery pipeline; This refers to the cross-sectional area of ​​the pipeline. The diameter of the conveying pipeline; This refers to the length of the transport pipeline; This is the friction coefficient; For the first The local resistance coefficient of a valve or local component; The sum of the local resistance coefficients of each valve and local component, Q is the main injection flow rate, and the state quantity of the air cushion in the ground storage tank is defined as follows: ， The air cushion state quantity of the rocket's onboard storage tank is defined as follows: .

3. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 2, characterized in that, In S3, pressure difference and flow rate adopt a unified form of piecewise linear membership function; For the pressure difference of the delivery pipeline For the main injection flow rate Q, establish piecewise linear membership functions corresponding to the high-flow injection, deceleration injection, and automatic replenishment conditions, respectively; For state variables x ,in Establish membership functions for the high-flow-rate refueling condition, the deceleration refueling condition, and the automatic replenishment condition, respectively, denoted as . , and ; in, Indicates the pressure difference in the delivery pipeline. Q This indicates that the flow rate has been increased. L This indicates a high-flow-rate refueling condition. D This indicates a deceleration and refueling operation. A Indicates automatic processing status; The formula for calculating the membership degree of the three types of working conditions is as follows: ； ； ； in, To increase the overall membership of the operating conditions for high-flow rates, To determine the overall membership of the deceleration and refueling operation, To automatically replenish the overall membership degree of the processing conditions; , , These represent the degree of membership of the pressure state to the three types of working conditions, respectively. , , These represent the membership degree of the flow state to the three types of operating conditions, respectively. The operating condition determination rules are as follows: when At that time, it is determined that the current refueling process is under high-flow refueling conditions; when At that time, it is determined that the current refueling process is in a deceleration refueling condition; when At that time, it is determined that the current refueling process is in automatic refueling mode.

4. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 3, characterized in that, After S3 divides the work conditions into three categories, during the GAN training process, different initial values ​​of parameters, training data selection rules and model correction terms are configured for different work conditions to suppress parameter identification drift caused by work condition switching.

5. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 4, characterized in that, The generator has no independent pure neural network mapping layer. It uses the differential equations of the liquid hydrogen refueling mechanism as the calculation model. All parameters to be optimized are interpretable physical parameters of cryogenic fluid engineering. The generation simulation time series fully satisfies the thermodynamic and fluid dynamic constraints of liquid hydrogen.

6. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 5, characterized in that, The discriminator is constructed using any one of CNN, RNN, LSTM or Transformer time-series neural networks to extract dynamic features of pressure and flow time series and determine the source of samples.

7. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 6, characterized in that, In S5, the generator pre-training adopts maximum likelihood estimation or minimum residual loss to enable the mechanistic model output time series to initially fit the real injection data and eliminate the oscillation non-convergence problem in adversarial training. The discriminator pre-training process involves labeling real sequences as positive samples and generating simulated sequences as negative samples. Optimization is achieved by minimizing the discriminator's cross-entropy loss, which is expressed as follows: ； Where, Y∈ This indicates that the data comes from a real data sample distribution. This indicates that the data comes from the generator's simulated sample distribution. Given an input variable (X) and an output sequence (Y), the discriminator determines the probability that the sequence belongs to the true sample data. To calculate the expected value of the distribution of simulated samples generated by the generator, To calculate the expected value of the distribution of the real sample data, For discriminator model parameters, These are the generator model parameters.

8. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 7, characterized in that, The expression for the objective function during generator training in S6 is: ； in, This represents the target sequence generated by the generator. The action value function is the function of the discriminator in the state. At that time, according to the strategy Take action The resulting accumulated rewards; By generator model Generate analog output sequence The generator generates results at each time t based on the results from the previous time. Generate current output S represents the operating condition label output by S3. The expression for the Monte Carlo action-value function is: ； Where C is the number of Monte Carlo samplings, b(˙) is the variance suppression baseline constant, D(˙) is the true probability output by the discriminator, n is the sequence number of the nth Monte Carlo sampling, and N is the number of Monte Carlo samplings. The generator is given a sequence of input variables, where t is the current generation time. The expression for parameter gradient update is: ； Where T is the total timing length. For the generator objective function with respect to parameters gradient, For the generator in the first t Output generated at time 1 Seeking expectations, To generate the current output for the generator given the generated sequence and input variables. The probability of.

9. The liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks according to claim 8, characterized in that, Monte Carlo sampling generates N simulated betting sequences for the same historical time series. The average of the discrimination probabilities of multiple sequences is taken as the cumulative reward, which reduces the variance of the single-step time series reward and improves the stability of parameter identification under critical and abnormally scarce conditions.

10. A liquid hydrogen refueling parameter identification device integrating fuzzy classification and generative adversarial networks is used to implement the liquid hydrogen refueling parameter identification method integrating fuzzy classification and generative adversarial networks as described in any one of claims 1-9, characterized in that, include: Data acquisition module, storage module, working condition fuzzy classification module, mechanism modeling module, GAN training module, parameter output module; The data acquisition module is used to collect time-series data on pressure, flow rate, temperature, and valve opening of the liquid hydrogen refueling pipeline and transmit them to the storage module; The storage module stores the coupling model of the cryogenic liquid hydrogen refueling mechanism, the membership function, the GAN network weights, and historical observation data. The fuzzy classification module is used to execute the fuzzy classification logic of the refueling conditions. Based on the piecewise linear membership function of the pressure difference of the delivery pipeline and the refueling flow rate, it calculates the comprehensive membership degree of the high flow refueling condition, the deceleration refueling condition and the automatic replenishment condition respectively, and outputs the corresponding refueling condition label according to the principle of maximum membership degree. The mechanism modeling module constructs the coupled model of the differential equation system and defines the parameter set to be identified and the input-output time series vectors. The GAN training module has a built-in generator and discriminator. The generator uses the liquid hydrogen refueling mechanism coupling model as the calculation model and is trained iteratively using the policy gradient and Monte Carlo sampling alternately until convergence. The parameter output module outputs the updated parameters to be identified in the converged generator and uses them as the identification results of key parameters of the liquid hydrogen refueling model for liquid hydrogen refueling simulation and online monitoring. Each module is interconnected via industrial control bus or Ethernet communication, and the processor uniformly schedules and executes the complete identification process.

Images