Fluidized system digital twinborn construction and state evaluation method

Abstract

The invention relates to the technical field of electronic digital data processing, and discloses a fluidization system digital twinning construction and state evaluation method, which comprises the following steps: carrying out CFD (Computational Fluid Dynamics) solution on various gas-solid fluidization working conditions, and establishing a numerical simulation database; according to the characteristics of the gas-solid fluidization system, determining a core reward function and physical constraints, training a reinforcement learning agent, and establishing a fluidization system digital twinborn model base; constructing a fluidized system environment according to the twin targets, and calculating multi-physical field distribution and discrete particle state evolution at subsequent moments; and using a multidisciplinary fusion degradation model to analyze the stress condition of the material at each position in a selected time period, calculating the fatigue condition and progressive loss rate of the material, performing system state evaluation, and outputting a result. The problem that in the prior art, simulation precision and efficiency of a fluidization system are difficult to consider at the same time is solved, and the purposes of efficient and accurate simulation, high operation reliability, comprehensive evaluation and analysis, convenient maintenance and high safety are achieved.

Description

Technical Field

[0001] The present invention relates to the field of physical technologies, and particularly to electronic digital data processing. Background Art

[0002] Fluidization systems are widely used in multiple engineering industries such as energy, chemical engineering, mining, metallurgy, food, pharmaceuticals, and biology, such as fluidized bed boilers, pneumatic conveying, and particle spray coating in food and pharmaceuticals. Fluidization technology can effectively improve the utilization rate of raw materials; however, due to the extremely complex fluidization system, which involves complex interactions between multiple physical fields, and the difficulty in measuring the multiphase flow parameters themselves, and the high opacity of the system, there is still great room for improvement in the design, operation, and maintenance of such systems. Simulation can provide more refined guidance for the operation and adjustment of fluidization systems, but the existing CFD (Computational Fluid Dynamics) simulation methods are slow, lack key models, and have poor accuracy. Therefore, when using the results of existing CFD technologies as the data source for digital twins, it often leads to distorted prediction results, so it cannot be used as a direct tool for establishing digital twins. At the same time, fluidization systems are prone to material degradation due to reasons such as thermal stress and particle erosion, and it is necessary to evaluate their reliability in a timely manner; and this problem involves multiple disciplines, and it is difficult to achieve only relying on CFD technology.

[0003] For example, the Chinese patent with publication number CN112546971B discloses a dense phase fluidized reaction control device and method, and provides the following technical scheme: the present invention discloses a dense phase fluidized reaction control device and method, including a dense phase fluidized reaction gas supply system and a dense phase fluidized reaction control system, the dense phase fluidized reaction gas supply system includes a raw gas I unit, a raw gas II unit, an inert gas unit and a dense phase fluidized reaction gas distribution unit; the dense phase fluidized reaction control system includes a dense phase bed fluidization monitoring unit, a fluidization monitoring data calculation and analysis unit, an expert judgment unit, a target controller unit and other instrument signal units; the present invention also provides a computer autonomous learning dense phase fluidized reaction control method, by analyzing the flow characteristics of the reaction bed, controlling the flow rate and reaction activity of the raw gas entering the dense phase area of ​​the reactor, and then controlling the dense phase fluidized reaction process. The present invention has a high degree of automation, efficient and convenient data analysis, and has the characteristics of machine learning, and has broad application prospects in the field of dense phase fluidized reaction control. However, the above-mentioned dense phase fluidized reaction control device and method has a large internal cavity volume of the fluidized system, and the difficulty of multiphase flow detection and its accuracy cannot be solved by multi-distribution monitoring; as for the prediction calculation and control process, it completely relies on machine learning analysis of measurement data, and cannot achieve a complete grasp of the distribution of system state parameters at the digital twin level. Therefore, it is difficult to achieve a complete description of the state of the fluidized system through existing analysis and adjustment methods for scenarios where the operating conditions such as load changes and particle property deviations of the fluidized system change, which makes it difficult to achieve ideal operating results. Summary of the invention

[0004] The present invention solves the problems in the prior art that it is difficult to strike a balance between simulation accuracy and efficiency of the fluidization system, the lack of physical mechanism in the machine learning model resulting in poor extrapolation ability, and the failure of material degradation assessment to fully cover the multi-field coupling effect through reinforcement learning coupled CFD. A method for constructing and state assessing a digital twin of a fluidization system is proposed, which achieves the goals of efficient and accurate simulation, high operational reliability, comprehensive evaluation and analysis, convenient maintenance and high safety, thereby providing a high-precision and fast-response technical solution for the operation diagnosis of the fluidization system.

[0005] To achieve the above object, the present invention adopts the following technical solution: A method for constructing and evaluating a digital twin of a fluidized system comprises the following steps: S1: Perform CFD solutions for various gas-solid fluidization conditions and establish a numerical simulation database; S2: According to the characteristics of the gas-solid fluidization system, determine the core reward function and physical constraints, train the reinforcement learning agent, and establish the digital twin model base of the fluidization system; S3: Construct a fluidization system environment based on the twin target, and calculate the multi-physical field distribution and the evolution of the discrete particle state at subsequent moments; S4: Use a multi-disciplinary fusion degradation model to analyze the stress conditions of the materials at each position within the selected time period, calculate the material fatigue and the progressive loss rate, evaluate the system state, and output the results.

[0006] The present invention applies a reinforcement learning paradigm based on CFD to efficiently calculate the multi-physical field distribution and complex particle behavior in the fluidization system environment; considering the multi-physical field effects in the fluidization reactor, a multi-disciplinary fusion material degradation evaluation method is constructed; the multi-physical field and discrete particle evolution information obtained statistically is combined with the degradation evaluation method to accurately evaluate the system state reasonably, and risk warnings and maintenance suggestions are put forward.

[0007] Preferably, the step S2 includes the following steps: S2.1: Based on the characteristics of the fluidization system, select a reinforcement learning algorithm, define the discrete particles as embodied agents, and the flow field and multi-physical fields as a dynamic interaction environment; S2.2: Construct a particle state space and an action space; S2.3: Design a core reward function based on the energy-minimized multi-scale model, with the normalized suspension transportation energy consumption as the optimization target and adding physical constraint conditions; S2.4: Construct a priority experience replay pool based on the numerical simulation database and perform dynamic sampling. When the model meets the conditions, complete the construction of the digital twin base.

[0008] By visualizing the discrete particles as agents, the deep integration of physical mechanisms and reinforcement learning is realized. The physical constraint reward function designed in combination with the energy-minimized multi-scale model not only ensures the physical rationality of the calculation process but also realizes the accurate mapping of the complex behavior of the fluidization system.

[0009] Preferably, the physical constraint conditions include particle packing constraints, grid crossing scale constraints, and particle event constraints. The particle packing constraint specifically means that the particle volume fraction does not exceed the close-packed limit, and the distance between any two particles is not less than the sum of their radii; the grid crossing scale constraint specifically means that the particle displacement within a single time step is less than the grid size where the particle is located; the particle event constraint specifically means that each particle undergoes at most one state transition within a single time step.

[0010] Adopting a triple physical constraint mechanism, the particle packing constraint effectively avoids non-physical particle penetration phenomena, the grid crossing constraint ensures the numerical stability of space-time discretization, and the particle event constraint accurately controls the state transition frequency. The synergistic effect of the three significantly improves the calculation accuracy of the digital twin and solves the technical bottlenecks of strong grid dependence and high calculation divergence risk in traditional CFD.

[0011] Preferably, in step S3, it includes initializing the fluidized system to be simulated, obtaining the discrete particle state at the current moment, and starting time advancement; according to the multi-physical field environment, using an agent to obtain particle actions; updating the discrete particle state at the next moment, interacting with the multi-physical field environment, and updating the field distribution; continuing time advancement until the end point of the simulation.

[0012] A closed-loop iterative space-time advancement mechanism is constructed. Through the two-way coupling of agent decision-making and multi-physical field update, the dynamic interactive simulation of the fluid-structure coupling process is realized, significantly improving the calculation accuracy.

[0013] Preferably, in step S2.3, the energy-minimized multi-scale model takes the normalized suspension transportation energy consumption as the optimization target, where the suspension transportation energy consumption is jointly determined by the density difference between particles and gas, the difference in volume fractions of dense-phase and dilute-phase particles, the proportion of the dense-phase volume, and the gas velocity term.

[0014] Preferably, step S4 includes the following steps: S4.1: Based on the discrete particle evolution data, analyze the particle impact events near the area of interest including the inner wall and internal components of the fluidized system, and calculate the stress, wear probability, and wear amount. S4.2: Calculate the thermal stress, and determine whether there is stress concentration or environmental factor influence. If so, enter step S4.3; if not, enter step S4.4. S4.3: Calculate the material strength reduction coefficient, output the total stress based on the material strength reduction coefficient, and enter step S4.5. S4.4: Output the total stress based on the stress. S4.5: Evaluate the state of the fluidized system.

[0015] The intelligent discrimination of complex working conditions is realized through a step-by-step decision tree structure. By introducing the dynamic calculation of the material strength reduction coefficient, the multiple effects of structural defects and environmental corrosion are effectively integrated, solving the technical problem of stress evaluation distortion in traditional single-field analysis.

[0016] Preferably, step S4.1 includes the following steps: S4.11: Calculate the particle event stress, and determine whether there is a non-contact field effect. If so, enter step S4.12; if not, enter step S4.13. S4.12: Calculate the non-contact field force, and use the sum of the particle event stress and the non-contact field force as the result to replace the particle event stress for subsequent calculations. S4.13: Calculate the material wear amount.

[0017] Preferably, the evaluation of the fluidization system state specifically includes determining the corresponding maximum number of cycles according to the stress amplitude at each position, determining the initial thickness of the material structure corresponding to each position, normalizing the calculated stress cycle times and progressive wear amounts to obtain a set of state evaluation values for the fluidization system. When the stress cycle times in a local area approach the limit value given based on the material, or its wear rate is higher than the given limit value, a risk reminder is sent to the operation or management personnel.

[0018] Through the dual determination of the stress cycle ratio and the wear rate threshold, a multi-dimensional early warning of material failure risk is achieved. The use of a dynamic safety factor mechanism to replace the fixed threshold significantly improves the adaptability of the state evaluation to working conditions, and can effectively avoid false alarms and missed alarms especially in the variable load operation scenario.

[0019] Preferably, the material wear amount is included in the progressive loss rate statistics and is accurately calculated using the finite element method according to the stress received, or calculated using the fragmentation kernel function and the sub-particle distribution function model; the material strength reduction coefficient includes magnifying the stress effect in the stress concentration area according to the reduction rate, and for environmental factors, based on the multi-physical field distribution evolution data, marking the oxidation-reduction range of each area, calculating the slagging and corrosion rates, obtaining the change curve of the material strength reduction coefficient over time, and magnifying the stress effect in this area according to the changing reduction rate.

[0020] Preferably, the state space includes particle position, particle velocity, gas velocity, and attribute parameters, and the action space includes displacement change, velocity adjustment, and attribute change.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0022] 1. Through the integration of reinforcement learning and CFD, the present invention optimizes the problems of large computational resource consumption and low parallel efficiency in traditional numerical simulations. Based on the embodied intelligence modeling at the particle scale, a multi-physical field update mechanism is established under physical constraint conditions, and the sample utilization rate is improved by combining the experience replay technique. Compared with the traditional CFD method, the present invention significantly shortens the simulation time while maintaining the analysis accuracy, making it possible to realize the real-time monitoring of industrial-scale fluidization systems and effectively balancing the engineering contradiction between accuracy and efficiency.

[0023] 2. By transforming the energy minimization principle of the fluidization system into a reinforcement learning reward function, the present invention constructs a mechanism and data dual-driven modeling framework. Based on the state and action spaces defined by particle kinematics and integrating the environmental interaction rules of the EMMS mesoscale theory, the machine learning process has a clear hydrodynamic basis. This training mechanism under physical constraints improves the generalization ability of the model under variable working conditions and provides a more reliable decision-making basis for the operation and maintenance of industrial systems.

[0024] 3. The present invention integrates multidisciplinary models such as fluid-structure interaction, thermodynamics, and chemical kinetics to construct a structural integrity assessment framework covering multiple factors such as mechanical stress, thermal stress, and environmental impact. Through the superposition analysis of multiple physical fields and the application of dynamic reduction factors, the comprehensiveness of degradation prediction is significantly improved, and it can effectively identify complex damage modes that are easily overlooked by traditional single models, providing more accurate spatio-temporal early warnings for preventive maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the overall flowchart of a method for constructing a digital twin and state assessment of a fluidized system according to the present invention.

[0026] Figure 2 It is the calculation flowchart of a multidisciplinary integrated degradation model for a method for constructing a digital twin and state assessment of a fluidized system according to the present invention.

[0027] Figure 3 It is the graph of the total coal feeding amount change under variable load operation for a method for constructing a digital twin and state assessment of a fluidized system according to the present invention.

[0028] Figure 4 It is the comparison graph of the operating value and the twin value of the bed material temperature at the rear wall under variable load for a method for constructing a digital twin and state assessment of a fluidized system according to the present invention.

[0029] Figure 5 It is the graph of the inner wall state assessment result of the fluidized system in operation mode A for a method for constructing a digital twin and state assessment of a fluidized system according to the present invention.

[0030] Figure 6 It is the graph of the inner wall state assessment result of the fluidized system in operation mode B for a method for constructing a digital twin and state assessment of a fluidized system according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] To make the objectives, technical solutions, and advantages of the present disclosure clearer, the following will further describe the embodiments of the present disclosure in detail with reference to the drawings. The proportions of the components are not drawn according to the actual proportions, and the proportions and dimensions shown in the drawings should not limit the essential technical solutions of the present invention. These embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described.

[0032] See Figure 1-6 As shown, a method for constructing a digital twin and state assessment of a fluidized system includes the following steps: S1: Perform CFD solutions for multiple gas-solid fluidization conditions to establish a numerical simulation database; S2: Determine the core reward function and physical constraints according to the characteristics of the gas-solid fluidization system, train the reinforcement learning agent, and establish the digital twin model base of the fluidization system; S3: Construct the fluidization system environment according to the twin target, and calculate the distribution of multiple physical fields and the evolution of discrete particle states at subsequent moments; S4: Use the multidisciplinary fusion degradation model to analyze the stress conditions of materials at each position within the selected time period, calculate the material fatigue and progressive loss rate, evaluate the system state, and output the results.

[0033] Such as Figure 1 And Figure 2 In an embodiment shown, Figure 1 This is the overall flowchart of a method for constructing and state evaluating a digital twin of a fluidization system according to the present invention, Figure 2 This is the calculation flowchart of the multidisciplinary fusion degradation model of a method for constructing and state evaluating a digital twin of a fluidization system according to the present invention. The present invention relates to a method for constructing and state evaluating a digital twin of a fluidization system, including the following steps: Step 1: Construction of the numerical simulation database Perform numerical solutions for various gas-solid fluidization conditions based on CFD technology, establish a numerical simulation database including the distribution of multiple physical fields and particle motion characteristics, and obtain high-precision data through fine grid division.

[0034] Step 2: Construction and training of the reinforcement learning model Definition of the algorithm framework: Select reinforcement learning algorithms including but not limited to DQN, DDPG, or SAC. Define discrete particles as embodied agents, and the flow field and multiple physical fields as the dynamic interaction environment; Definition of the state and action spaces: The state space includes particle position, velocity, gas velocity in the grid where the particle is located, and particle attributes; The action space includes displacement change, velocity adjustment, and attribute change; Design the reward function, based on the energy minimum multi-scale (EMMS) model, with the normalized suspension transportation energy consumption as the optimization goal. And impose physical constraints, including: Particle packing constraint: The particle volume fraction in the grid ≤ the close-packed limit, and the distance between any two particles ≥ the sum of the radii; Grid crossing constraint: The particle displacement within a single time step < the grid size; Particle event constraint: At most one state transition within a single time step.

[0035] Construct a prioritized experience replay pool based on the numerical simulation database, and complete the construction of the digital twin base when the model prediction error is less than the set value.

[0036] Step 3: Real-time simulation and multi-field coupling calculation The system is initialized to load the initial state parameters of the target system; the particle actions are obtained through the agent policy function; the particle state is updated and the NS equations are solved to update the gas-phase field distribution; the iteration is advanced until the end of the simulation cycle, and the multi-physical field distribution and the particle motion trajectory are synchronously recorded.

[0037] Step 4: Multidisciplinary integrated material degradation assessment According to the presence or absence of non-contact field effects, calculate the stress and the material wear amount, and calculate the thermal stress, and correct according to stress concentration and environmental factors: calculate the material strength reduction factor for the stress concentration area and the environmental correction area.

[0038] State assessment decision, match the material S-N curve to obtain the maximum number of cycles, calculate the cumulative damage; normalize the wear amount, and give corresponding alarms.

[0039] In this embodiment, specifically as follows: Use a fine grid and an accurate model, invest sufficient computing power to perform CFD numerical simulation calculations for multiple groups of fluidization conditions, and establish a database. Regarding the model: It is sufficient to obtain accurate numerical simulation results of the fluidization process for constructing the database and can be applied to the present invention. In the present invention, three embodiments are exemplified: Embodiment 1: Use the fully resolved simulation (FRS), also known as the true direct numerical simulation (True DNS) method, divide the grid scale below the particle scale, and perform multi-phase direct numerical simulation at the sub-particle scale.

[0040] Embodiment 2: Use the DNS (Direct Numerical Simulation) method, divide the grid scale to the turbulent Kolmogorov dissipation scale, and at the same time regard the discrete particles as point sources or pseudo-fluids for solution.

[0041] Embodiment 3: Using the large eddy simulation (Large Eddy Simulation) method or solving the Reynolds-averaged Navier-Stokes equations (Reynolds-Averaged Navier-Stokes Equations) on a coarser grid is also acceptable, which does not affect the applicability of the method, but the data accuracy will decrease accordingly.

[0042] Apply the reinforcement learning paradigm in CFD, establish the digital twin base, and use a series of physical mechanisms such as the principle of minimum energy as the training objective and constraints, which can effectively improve the model accuracy, accelerate the convergence rate, and enhance the interpretability. The detailed description of the adaptability of reinforcement learning (RL) and the fluidization system is as follows: The fluid flow field (distribution) is regarded as the environment; discrete particles are regarded as embodied agents, and the value ranges of various attribute variables (particle size, velocity, temperature, composition, etc.) of the discrete particles are the state space. The interaction between the air flow and the discrete particles corresponds to the interaction between the environment and the agent. Among them, the air flow dragging the particles (the action of the environment on the agent) determines the actions of the particles, and the particles blocking the air flow (the action of the agent on the environment) can be described as the result of performing actions in the current state. According to the actions, the state parameters of the agent change within the state space range (old state + action = new state, and both the old and new states should fall within the state space). Therefore, in the solution process, physical constraints must be imposed on the actions of the agent.

[0043] The present invention has no special requirements for RL algorithms, and common algorithms such as Q-Learning, DQN (Deep Q-Network), SAC (Soft Actor-Critic), DDPG (Deep Deterministic Policy Gradient), and Dyna-Q are all applicable.

[0044] It should be noted that there are multiple classification methods for RL. Considering that the fluidization system description problem to be solved is a real and mechanistic industrial process, it can also be discussed under the classification of model-based and model-free reinforcement learning. The technical route of the present invention is applicable to both types of reinforcement learning methods.

[0045] Discrete particles are regarded as agents, and the calculation of particle motion is regarded as a Markov Decision Process (MDP). During the process of training the agent, the elements used are as follows: (1) Objective function: The principle of minimum energy, that is, the energy consumption for the suspended transportation of particles per unit mass tends to be minimized, Nst -> min.

[0046] In this embodiment, the calculation method of the suspended transportation energy used refers to the energy-minimization multi-scale model (EMMS).

[0047] (2) State space S, which includes the spatial position where the particle is located, the particle velocity, the gas phase velocity at the position where the particle is located, and other custom variables.

[0048] More variables Ψ can be added to the state space according to the actual problem, such as discrete particle information such as diameter and density in a polydisperse particle system, the particle component composition, temperature, turbulent kinetic energy, etc. in a chemical reaction system.

[0049] (3) Action space A, which includes the particle movement distance, the particle velocity change value, and the change amplitude of other custom variables, is jointly determined by the agent's current state St and the state space S.

[0050] Specifically: reference can be made to "physical constraints".

[0051] (4) Reward function rt, whose expression is a function of the suspended transportation energy from the starting moment of reward accumulation to the current moment and the cumulative reward decay factor.

[0052] Regarding the reward function and normalization, in RL, the reward function is the most core, but its given method can also be freely designed (taking AlphaGo as an example, the design method of the reward function does not affect the realization of the purpose of "playing Go", only affecting the game situation). Therefore, the highlight of the present invention is to use the "principle of minimum energy" to design the reward function. As for what kind of normalization method to adopt and what value to take for the reward decay at each step, they are relatively free. In this embodiment, the decay factor used is 0.80 - 0.99.

[0053] (5) Physical constraints: Particle packing constraint: for any grid, the volume fraction of particles therein is not greater than its packing limit: ε s ≤ ε s,cp , where ε is the particle volume fraction, and the subscript s means the solid phase solid, and cp represents the close - packing limit of particles (generally taken as about 0.63 for single - size spherical particles).

[0054] For any two particles, the distance between the coordinates of their centroid positions is not less than the sum of the radii of the two particles. Assuming that all particles are spherical and do not undergo extrusion deformation. Where D ij is the distance between the centers of the spheres of the two particles numbered i , j , and d is the particle diameter.

[0055] Handling method when the constraints are not met: It can be processed as if the particles collide. It is considered that the particles collide and generate displacement (state change). Hard - sphere, soft - sphere and other models can be used for calculation to update the particle state. Because distance detection is included in such models, it can ensure that the new state obtained after calculation meets the constraint conditions.

[0056] The above collisions can also be handled in a simplified manner. For example, referring to the processing method of the multiphase particle-in-cell method (MP-PIC), instead of directly calculating particle collisions, a body force is directly applied to the particles along the direction of the particle stress gradient to cause displacement (the result is also a change in state, which can also be called position offset or state offset), so that the state of the displaced particles (agents) satisfies the packing constraint. The method of applying the body force is not uniquely determined and can refer to the prior art.

[0057] Mesh crossing scale constraint: within any time step, the movement distance of any particle is less than the size of the current grid it is in. This constraint is a direct constraint on the time slot. If it is not satisfied, the time slot can only be reduced until the movement distance of the particle within a single time slot meets the constraint.

[0058] Particle event constraint: within the time scale of one action and interaction, each particle can have at most one state change, and each state change is not the superposition of the effects of two state changes.

[0059] Explanation of the particle event constraint: This constraint is similar to the mesh crossing scale constraint and is intended to limit the time slot used to divide the continuous time flow in the RL paradigm. The mesh crossing scale constraint is a limitation on the single movement distance of the particle (agent), and the particle event constraint mainly limits the number of particle events (such as collisions), generally limited to 1 time. Within each time slot, the behavior of the agent and the state of the environment are sampled and updated, which can be understood as "taking a picture". It is close to the concept of the time step in simulation calculations.

[0060] Construct a fluidized system environment for the twin target, and calculate the distribution of multiple physical fields and the evolution of discrete particle states during time advancement. The steps are as follows: Determine the state space S, action space A, and reward function R of the solid-phase particles (agents); Put part of the particle information (s, a, r, s') in the database into the experience replay pool; The present invention has no special requirements for the sampling method, and the prior art can be referred to.

[0061] Initialize the fluidized system to be simulated and obtain the discrete particle state s at the current moment t , start time advancement; according to the multi-physical field environment, use the agent to obtain the particle action a t ; update the discrete particle state s at the next moment t+1 , interact with the multi-physical field environment, update the field distribution; continue time advancement until the end point of the simulation.

[0062] By statistically calculating the stress on the materials in the local area during the operation process, the material loss rate x is progressively calculated: When applying this degradation model for analysis, the fatigue characteristics of the corresponding materials should be determined first. Specifically: The structural materials of the fluidized system are generally steel materials. According to the evolution of the multi-physical field distribution in the fluidized system and the calculation results of particle impact events, the alternating stress on the materials is much lower than the yield limit of the materials, which is identified as high-cycle fatigue. For the types of materials used in the fluidized system (such as: 20# steel, heat-resistant steel, stainless steel, carbon steel, alloys, etc.), the fatigue characteristics of the materials can be determined according to their S-N curves.

[0063] When using the S-N curve, the corresponding fatigue life (N, number of cycles) can be found according to the stress amplitude (S).

[0064] The stress F on the material at any position can be expressed as the superposition of the following several actions. Specifically: ① Stress part of particle events (F1) Based on the discrete particle evolution data, analyze the particle impact events near the areas of concern such as the inner wall and internal components of the fluidized system, and calculate the stress, wear probability and wear amount.

[0065] In the case of no non-contact field force: The impact stress F1 is calculated according to the energy conservation and momentum balance of dynamics, and analyzed and calculated according to elastic / inelastic collisions; the wear probability P x and the wear amount x are included in the progressive loss rate statistics. The finite element method can be used for accurate calculation according to the stress, or it can be simplified to a crushing process and calculated using the crushing kernel function and the sub-particle distribution function model.

[0066] At the event level, it mainly includes the following data: occurrence time, particle size, impact angle, relative velocity before and after the impact event; At the statistical level, it mainly includes the following data: particle material, hardness, sphericity, collision frequency, average momentum loss.

[0067] In the case of having non-contact field force: Use F1 + F2 to calculate the wear probability P x and the wear amount x.

[0068] ② Non-contact field action part (F2) If the electrostatic effects between particles and the wall, and between particles in the fluidization system are significant, or there are non-contact field effects such as magnetic fields and electric fields, then based on the field force data (or field strength and particle charge data), before incorporating the stress obtained in ① into the fatigue characteristics calculation of the S-N curve, the increase or decrease value of the true collision stress caused by the non-contact field force F2 is superimposed, and then the material strain and fatigue characteristics are calculated.

[0069] ③ Thermal stress part (F3) Based on the multi-physical field distribution evolution data, the thermal stress F3 suffered by the materials at various places in the fluidization system is calculated.

[0070] It is mainly calculated according to the temperature distribution. The local structure can be simplified for direct calculation. For example, the one-dimensional fully constrained plane stress can be written as: thermal stress = elastic modulus × coefficient of linear expansion × temperature change. A more rigorous approach is to perform a finite element simulation of elasticity mechanics on the real structure based on the known temperature distribution.

[0071] In addition, in special cases, the material strength reduction coefficient a should also be defined: ① Stress concentration and tissue defect part (a1) According to the design and construction drawings of the fluidization system, locate the stress concentration areas, mainly including pipe connection turning points, wall areas with protrusions or notches, and internal components.

[0072] Define the material strength reduction coefficient a1 for this area, and amplify the stress effect on this area according to the reduction rate.

[0073] The reduction coefficient value can be manually defined from the perspective of safe use; or the finite element simulation method can be used to calculate the true stress conditions at the defect under several working conditions based on the real structure with defects, and the reduction coefficient value or function that is more in line with the actual situation can be obtained by fitting.

[0074] ② Environmental factor part (a2) Based on the multi-physical field distribution evolution data, mark the oxidation-reduction atmosphere in each area, calculate the slagging and corrosion rates, obtain the change curve of the material strength reduction coefficient a2 over time, and amplify the stress effect on this area according to the changing reduction rate. If there are other coatings such as metal spraying and high-temperature coatings, this part is calculated again when the real-time cumulative wear amount exceeds the coating thickness. The slagging problem can be simplified using the "collision + probability adhesion + accumulation" model. For the corrosion problem, the corrosion rate can be estimated according to the gas component concentration distribution based on empirical formulas.

[0075] The present invention does not exclude any slagging corrosion model or calculation method, because the multi-physical field distribution evolution data has been obtained in the previous steps, which is sufficient to meet the calculation requirements of different precision models.

[0076] Based on the calculated material fatigue condition and progressive loss rate, evaluate the state of the fluidization system and issue instructions or suggestions.

[0077] Evaluation steps: I. Material fatigue: Determine the corresponding maximum number of cycles N according to the stress amplitude at each position max ; Material wear: Determine the initial thickness c of the material structure corresponding to each position ini .

[0078] For the case where the stress amplitude span is large, N can be estimated by interpolation, fitting, etc. max .

[0079] II. At any concerned moment t, normalize the calculated number of stress cycles N t and progressive wear amount c t to obtain the state evaluation value set of the fluidization system at time t. The evaluation value set should consist of at least fatigue evaluation and wear evaluation; There can be various normalization methods in this embodiment, which can refer to the prior art.

[0080] III. Customization: Decision interval (segmented) or safety factor (one-size-fits-all) for determining when to issue what instructions or strategies. That is, when the number of stress cycles in a local area approaches the limit value given based on the material S-N curve, or its wear rate is higher than the given limit value, a risk reminder is issued to the operation or management personnel.

[0081] In this embodiment, the safety factor mode is used: If it is considered that there is a risk of local material structure fracture when N t / N max > 0.8, then define 0.8 (or a smaller value) as the stress fatigue safety factor y. When N t / N max ≥ y, a structural fatigue maintenance reminder is issued.

[0082] In actual operation, different safety factors can be defined according to the structure and material characteristics of each position, as well as the requirements of operation safety margin.

[0083] Taking the wear evaluation as an example again, the decision interval mode is used: If it is considered that there may be a furnace shutdown accident when the water wall position is at c t / c ini > 0.3, then the following evaluation configuration can be carried out: Decision interval: c t / c ini = 0~0.1, Strategy: Only dynamically update the visualization effect and do not mark it as a risk item; Decision interval: c t / c ini = 0.1 - 0.2, Strategy: Mark it as a risk item, estimate the remaining reliable operation time (using methods such as polynomial extrapolation and prediction simulation), send it to management / operation personnel, and arrange a maintenance plan; Decision interval: c t / c ini = 0.2 - 0.3, Strategy: Key maintenance reminder; Decision interval: c t / c ini > 0.3, Strategy: Major operation risk warning.

[0084] In actual operation, different decision intervals can also be defined for materials at different positions. For example, the alarm threshold can be increased for materials at non-critical positions (the importance of other walls must be lower than that of the water circulation tube bundle) to avoid over-maintenance and waste of maintenance efforts.

[0085] The following is the application situation of this embodiment: 1. Fluidization system operation analysis - laboratory scale Taking the riser of a circulating fluidized bed as an example, the particle mass flow rate at the outlet directly reflects the transport capacity of the fluidizing air (gas) for solid particles and can prove whether the simulation analysis results are accurate.

[0086] Apply the present invention to a riser with a diameter of 0.09 m and a height of 10.2 m at the laboratory scale, where the solid particle diameter is 54 um and the density is 930 kg / m 3 . Select a total of 6 working conditions with fluidization air velocities Ug of 1.52 m / s and 2.10 m / s and material amounts of 15 kg, 25 kg, and 35 kg respectively. The calculation results are shown in Table 1.

[0087] It can be seen that the accuracy of the analysis results of the existing CFD technology is low (in six working conditions, at most, the solid particle mass flow rate is overestimated by 973%); while the accuracy of the present invention is significantly improved.

[0088] Table 1

[0089] Table 2

[0090] Table 1 and Table 2 are for U gComparison of calculation effects at different values. Source of experimental data exp.: Li J H, Tung Y, Kwauk M. Axial voidage profiles of fast fluidized beds in different operating regions. Proceedings of the Second International Conference on Circulating Fluidized Beds, Compiégne, France, 193 - 203, 1988 (Li Jinghai, et al. Axial voidage profiles of fast fluidized beds in different operating conditions. The Second International Conference on Circulating Fluidized Beds, Compiégne, France, 193 - 203, 1988).

[0091] 2. Analysis of the operation of the fluidization system - industrial scale For a 130t / h cogeneration fluidized bed boiler, the change in the total coal feed under variable load operation on the demand side is as follows Figure 3 (The data sampling time interval is 1 min). The bed material temperature at the rear wall can reflect the fluidization and combustion organization in the furnace. The comparison between the measured operating values and the calculated values of the twin is as follows Figure 4 .

[0092] 3. Analysis of the material deterioration of the fluidization system Using different operating modes (A - hot state, B - cold state), the visualized results after quantifying the evaluation of the inner wall state of the fluidization system are as follows Figure 5 、 6 shown. The darker the color, the more serious the deterioration. By comparison, particle impact is still the main cause of system material deterioration, but the influence of thermal stress, non - contact force field and environmental factors on deterioration makes the material deterioration results calculated solely based on particle collision significantly distorted. Under the calculation system proposed in the present invention, the evaluation of the material deterioration of the fluidization system under the coupling action of multiple fields can be effectively realized; and the present invention applies the reinforcement learning paradigm in the original CFD architecture, with high calculation efficiency, meeting the dual requirements of the digital twin for response efficiency and calculation accuracy, and can effectively realize the "transparency" of the industrial fluidization system, providing a reliable guarantee for its stable operation and efficient maintenance.

[0093] The present invention is not limited to the above - mentioned embodiments. No matter what changes are made in its shape or material composition, as long as the structural design provided by the present invention is adopted, it is a deformation of the present invention and should be considered within the protection scope of the present invention.

Claims

1. A method for constructing a digital twin and evaluating the state of a fluidization system, characterized in that It includes the following steps: S1: Conduct CFD solutions for various gas-solid fluidization working conditions to establish a numerical simulation database; S2: Determine the core reward function and physical constraints according to the characteristics of the gas-solid fluidization system, train the reinforcement learning agent, and establish the base of the digital twin model of the fluidization system; S3: Construct the environment of the fluidization system according to the twin target, calculate the distribution of multiple physical fields and the evolution of discrete particle states at subsequent moments; S4: Use the multidisciplinary degradation model to analyze the stress conditions of the materials at each position within the selected time period, calculate the material fatigue and progressive loss rate, evaluate the system state, and output the results.

2. The digital twin construction and state evaluation method of a fluidization system according to claim 1, characterized in that The step S2 includes the following steps: S2.1: Based on the characteristics of the fluidization system, select the reinforcement learning algorithm, define the discrete particles as embodied agents, and the flow field and multiple physical fields as the dynamic interaction environment; S2.2: Construct the particle state space and action space; S2.3: Design the core reward function based on the energy-minimized multi-scale model, with the normalized suspension transportation energy consumption as the optimization target and add physical constraint conditions; S2.4: Construct a priority experience replay pool based on the numerical simulation database and perform dynamic sampling. When the model meets the conditions, complete the construction of the digital twin base.

3. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 2, characterized in that, The physical constraint conditions include particle packing constraints, grid crossing scale constraints, and particle event constraints. The particle packing constraint is specifically that the particle volume fraction does not exceed the close-packed limit, and the distance between any two particles is not less than the sum of their radii; the grid crossing scale constraint is specifically that the particle displacement within a single time step is less than the grid size where it is located; the particle event constraint is specifically that each particle undergoes at most one state transition within a single time step.

4. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 2 or 3, characterized in that In the step S3, it includes initializing the fluidization system to be simulated, obtaining the discrete particle state at the current moment, and starting time advancement; according to the multi-physical field environment, using the agent to obtain the particle actions; Update the discrete particle state at the next moment, interact with the multi-physical field environment, and update the field distribution; continue time advancement until the end point of the simulation.

5. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 2, characterized in that, In the step S2.3, the energy-minimized multi-scale model takes the normalized suspension transportation energy consumption as the optimization target, where the suspension transportation energy consumption is jointly determined by the density difference between the particles and the gas, the difference in the volume fractions of the dense and dilute phase particles, the proportion of the dense phase volume, and the gas velocity term.

6. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 2 or 3, characterized in that The step S4 includes the following steps: S4.1: Based on the discrete particle evolution data, analyze the particle impact events near the concerned areas including the inner wall and internal components of the fluidization system, and calculate the stress, wear probability, and wear amount; S4.2: Calculate the thermal stress, and judge whether there is stress concentration or environmental factor influence. If so, enter step S4.3; if not, enter step S4.4; S4.3: Calculate the material strength reduction coefficient, output the total stress based on the material strength reduction coefficient, and enter step S4.5; S4.4: Output the total stress based on the stress; S4.5: Evaluate the state of the fluidization system.

7. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 6, characterized in that, In the step S4.1, it includes the following steps: S4.11: Calculate the particle event stress, and judge whether there is a non-contact field effect. If so, enter step S4.12; if not, enter step S4.13; S4.12: Calculate the non-contact field force, and use the sum of the particle event stress and the non-contact field force as the result to replace the particle event stress for subsequent calculations; S4.13: Calculate the material wear amount.

8. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 6, characterized in that, The evaluation of the fluidized bed system state specifically includes: determining the corresponding maximum cycle number according to the stress amplitude at each position, determining the initial thickness of the material structure corresponding to each position, normalizing the calculated stress cycle number and progressive wear amount to obtain a set of state evaluation values for the fluidized bed system. When the stress cycle number in a local area approaches the limit value given based on the material, or its wear rate is higher than the given limit value, a risk reminder is given to the operation or management personnel.

9. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 7, characterized in that, The material wear amount is included in the progressive loss rate statistics and is accurately calculated using the finite element method according to the applied stress, or calculated using the fragmentation kernel function and sub-particle distribution function model; the material strength reduction coefficient includes magnifying the stress effect in the stress concentration area according to the reduction rate. For environmental factors, based on the multi-physical field distribution evolution data, mark the oxidation-reduction range of each area, calculate the slagging and corrosion rates, obtain the change curve of the material strength reduction coefficient over time, and magnify the stress effect in this area according to the changing reduction rate.

10. A method for constructing a digital twin and evaluating the state of a fluidization system according to claim 2, characterized in that, The state space includes particle position, particle velocity, gas velocity, and attribute parameters, and the action space includes displacement change, velocity adjustment, and attribute change.

Images