Self-generating system construction method and system based on physical eigen equation

Abstract

This invention discloses a method and system for constructing self-generated systems based on physical eigenvalue equations. By constructing a hierarchical physical eigenvalue equation library to characterize the endogenous stable order at different levels of the system, an initial endogenous digital twin is generated based on this equation library. Autonomous system operation is achieved through a complete self-generated closed loop of "sensing-prediction-regulation-feedback-evolution." A ternary coupling strength index is introduced to enable the automatic identification and incubation of sub-self-generated systems, giving the system open evolutionary capabilities. This invention represents a technological advancement from passively executing external rules to actively generating its own endogenous order, and can be widely applied in fields such as industrial control and robot swarms.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, system control and engineering technology, and specifically relates to a method and system for constructing self-generated systems based on physical eigenvalue equations. Background Technology

[0002] The existing methods for constructing autonomous artificial intelligence systems mainly fall into two categories: The first category is the "pre-defined rule-based" system, where engineers pre-write all possible rules and logic, and the system executes according to these pre-defined rules during operation. This type of system cannot handle unpre-defined operating conditions, has poor adaptability, and high maintenance costs. For example, traditional PID control systems can only handle linear or weakly nonlinear systems; for complex industrial systems with strong coupling and strong nonlinearity, the control effect drops sharply.

[0003] The second type is the "data-driven" system: This type trains machine learning models using a large amount of historical data to deduce rules from the data. Such systems suffer from black-box problems, overfitting risks, and physical inconsistencies. Their performance drops sharply under small sample conditions, and they fail to fundamentally understand the system's operational patterns. For example, deep learning-based fault prediction systems have relatively low prediction accuracy when the training data does not cover the specific operating conditions.

[0004] In recent years, Physical Information Neural Networks (PINNs) have been gradually applied to the field of control. However, these solutions all use PINNs for offline modeling or single-step prediction in specific application scenarios. They have not formed a complete closed loop that enables autonomous construction of digital models from the endogenous order of the system, autonomous generation of control strategies, autonomous collaboration of multiple execution units, and continuous self-evolution of model parameters. Furthermore, they lack the open evolutionary capability to generate new subsystems.

[0005] The closest existing technology is the "Physical Information Neural Network Architecture Based on a Large Model" disclosed in invention patent CN121835764A. This scheme uses a general deep learning architecture of "CNN + Bidirectional RNN + Transformer" to process spatiotemporal sequence data, and corrects the representation of the underlying physical field through an independent physical constraint correction module. However, in this scheme, the physical constraints are added as a correction module in the output layer, which cannot guarantee the physical consistency of the intermediate layers and fails to form a complete self-generating closed loop of "perception-prediction-regulation-feedback-evolution".

[0006] Current technologies generally treat physical equations as external verification tools, never elevating them to the level of "endogenous order representation" of a system. They assume that complex industrial systems cannot operate completely autonomously and must rely on pre-set rules or periodic human intervention. This bias has kept existing solutions at the "decision-making support" level, never attempting to build autonomous systems with "self-awareness" and "self-improvement" capabilities based on the physical eigenvalue equations of the target system's stable order. Therefore, it is urgent to position physical eigenvalue equations as the system's endogenous stable order, starting from the system's inherent nature, to construct a universal self-generating system based on physical eigenvalue equations. Through engineering innovation, this system should achieve self-awareness and self-optimization, autonomously constructing rules, making decisions, executing actions, and evolving, and even endogenously generating new sub-self-generating systems, achieving open evolution. Summary of the Invention

[0007] To address the core shortcomings of existing technologies, such as the imposition of external rules, static and rigid architectures, and reactive control, which fail to meet the autonomy, adaptability, and scalability requirements of complex industrial systems, this invention proposes a self-generating system construction method and system based on physical intrinsic equations. By defining physical laws as the stable order inherent in the system, it achieves a technological advancement from "passively executing external rules" to "actively generating its own order." The self-generating system refers to a dynamic whole that maintains operational consistency through a closed-loop recursive architecture composed of three functional primitives: information field state C, energy mechanism E, and material structure M. The information field state C is the sum of all rules, constraints, and potential states within the system, corresponding to the cognitive layer in this invention; the energy mechanism E is the dynamic sum of all dynamic processes within the system, corresponding to the energy layer in this invention; and the material structure M is the organized sum of all physical carriers and topological connections within the system, corresponding to the material layer in this invention. A closed-loop composite mapping of the CEM ternary recursive coupling, i.e., the C→E mapping, is established. E→M mapping M→C mapping These sequentially combine to form a closed-loop composite mapping: The core constraint of the closed loop is the identity mapping approximation of the composite mapping to the information field configuration space (C space), i.e. Its mathematical criterion is the Lipschitz constant of the closed-loop composite mapping on C space. This ensures that the system converges to a unique steady state from any initial state.

[0008] The overall concept of this invention is to construct a hierarchical physical intrinsic equation library to represent the endogenous stable order of different levels of the system, and generate an initial endogenous digital twin based on this equation library; to achieve autonomous operation of the system through a complete self-generating closed loop of "perception-prediction-regulation-feedback-evolution"; to introduce a ternary coupling strength index to realize the automatic identification and incubation of sub-self-generating systems, enabling the system to have open evolution capabilities; and to introduce a cognitive free energy F to autonomously quantify the cognitive deviation between the endogenous digital twin and the physical world. When F increases, the system "realizes" the inadequacy of its own model and actively triggers parameter updates or model retraining, realizing a key leap from "passively executing physical laws" to "actively recognizing and optimizing itself".

[0009] The technical solution of this invention is essentially an engineering implementation of the theory of "operational identity" in self-generated systems. In the ontology of self-generated systems, the fundamental defining characteristic of a system is its operational identity—the fundamental attribute that allows the system to be continuously identified as "itself" throughout its diachronic evolution. The hierarchical physical eigenvalue equation library constructed in this patent is precisely the mathematical representation of the system's operational identity: the first level defines the identity boundaries of components, the second level defines the identity constraints of coupling between components, and the third level defines the global identity order of the system. The system continuously monitors the deviation between its own state and the physical eigenvalue equations through an endogenous digital twin. When the deviation exceeds a preset threshold, it autonomously generates a control strategy to bring the system back to the steady-state range of operational identity. This invention constructs a closed-loop collaborative operation of "perception-prediction-decision-execution-evolution," enabling the system to leap from passively executing external rules to actively generating its own internal order—a self-generated system.

[0010] Firstly, the present invention provides a method for constructing self-generated systems based on physical eigenvalue equations.

[0011] Executed by hardware modules deployed in industrial-grade control units, the core of this method lies in constructing an endogenous digital representation of the system through a hierarchical physical intrinsic equation library, forming a complete self-generating operational closed loop. This method mainly includes the following steps:

[0012] Step S1: Constructing a Hierarchical Physical Eigenvalue Equation Library. Based on the fundamental physical laws and factory calibration parameters of the target system, an initial physical eigenvalue equation library containing at least three levels is constructed and stored in non-volatile memory. The physical eigenvalue equation library refers to a set of partial differential equations describing the intrinsic stable order of different levels of the system. The equations at each level are bidirectionally coupled through upward and downward causality, forming a complete hierarchical system of order. The initial physical eigenvalue equation library is preset by the designer based on known physical laws and stored in non-volatile memory before leaving the factory. It is automatically loaded when the system is first powered on. After the system starts running, through the feedback update mechanism in step S7, the key physical parameters in the physical eigenvalue equation library are continuously updated by measured data to correct them, gradually evolving from the initial preset state into a physical eigenvalue equation library characterizing the true intrinsic stable order of the system.

[0013] Step S2: Multiphysics Simulation Generates Virtual Dataset. The multiphysics simulator deployed in the industrial-grade control unit calls upon partial differential equations from the physical intrinsic equations library to perform high-fidelity multiphysics simulations based on fundamental physical laws, generating a virtual system dataset covering key operating condition parameter combinations. This step solves the problems of difficult and costly data acquisition from real industrial systems, providing sufficient physically consistent data for subsequent AI model training.

[0014] Step S3: Training the Initial Endogenous Digital Twin. An endogenous digital twin refers to a system's digital representation that is autonomously constructed by the system itself from its inherent stable order and is capable of self-evolution, unlike traditional digital twins constructed by external engineers. The AI ​​agent model training unit divides the virtual system dataset into a training set and a validation set, and trains it using a weighted sum of data loss and physical loss as the total loss function. When the total loss function no longer decreases for K consecutive rounds on the validation set, the initial endogenous digital twin is obtained.

[0015] Step S4: Real-time acquisition of system status. After each status acquisition cycle is completed, the multimodal sensor network acquires the measured physical characteristic parameters of the current operating status of the target system and transmits them to the industrial-grade control unit to provide real-time input for subsequent status prediction.

[0016] Step S5: Predicting Future State Deviations. The system evolution prediction unit inputs the measured physical characteristic parameters into the endogenous digital twin and outputs the predicted state deviation values ​​of the target system within a preset future time window. This step enables advance prediction of the system's future state, laying the foundation for proactive control.

[0017] Step S6: Generate Optimal Control Parameters in Advance and Execute Collaboratively. When it is predicted that the future state deviation will exceed a preset threshold, the numerical optimization solver, with the objective function of minimizing the overall system state deviation, performs optimization before the predicted deviation occurs, autonomously generating the optimal combination of control parameters. After hierarchical verification, this is distributed to each execution unit for execution. This step achieves pre-judgment control, fundamentally eliminating the cumulative effect of deviation.

[0018] As a preferred implementation, when the target system contains multiple execution units or independent sub-self-generated systems, step S6 simultaneously executes hierarchical cooperative game theory: collecting candidate control parameters generated by all execution unit agents and independent sub-self-generated systems, performing non-dominated sorting based on the game payoff function, and selecting the control parameter combinations corresponding to the Pareto front approximate solution for execution. For multi-level systems, a two-level cooperative game theory mechanism is adopted to ensure a balance between component ontology security and global system optimization: the first-level cooperative game prioritizes the satisfaction of component ontology or sub-self-generated system ontology constraints, while the second-level cooperative game prioritizes the achievement of global system cooperative goals. After calibration through 100 sets of multi-condition simulation experiments, when the ontology constraint satisfaction weight of the first-level game payoff function is not less than 0.6, it can ensure that component-level security constraints are not breached, while also taking into account global cooperative optimization benefits; when it is less than 0.6, there is a risk of local execution unit overload.

[0019] Step S7: Feedback update of physical intrinsic equations. The measured state change parameters acquired and controlled by the multimodal sensor network are updated. The parameter update unit uses the measured values ​​as observations and the predicted values ​​as prior estimates to update the key physical parameters in the physical intrinsic equation library of the parent system and each self-generated sub-system. This step enables the continuous self-evolution of the system model.

[0020] When performing parameter updates, the parameter update unit will recognize the free energy. As a modulation factor for updating the step size. When When, use the regular update step size; when When, automatically increase the update step size to accelerate model adaptation; when If the endogenous digital twin's understanding of the subsystem or operating condition is deemed severely impaired, it triggers a re-initialization of the local model or a request for external calibration. Simultaneously, the parameter update unit records the performance of each subsystem under various operating conditions. Historical values ​​are used to identify the "cognitive blind spots" of digital twins—that is... Subsystems or operating condition combinations with a confidence level consistently above 0.5 are identified as cognitive blind spots. Once these blind spots are marked, the system automatically reduces the prediction confidence level for these blind spots during autonomous decision-making and prioritizes the generation of simulation data and model retraining for these blind spots.

[0021] The key physical parameters refer to those parameters in the physical intrinsic equation library that need to be corrected online as the system's operating state changes. For example, in the physical intrinsic equation library of a lithography machine projection lens, key physical parameters include the thermal expansion coefficient of the optical glass. and thermal conductivity Coupling coefficient of Joule's thermoelectric coupling equation (varying with temperature) Cooling water convective heat transfer coefficient These parameters drift during system operation due to factors such as material aging and environmental changes, and are continuously corrected by measured data through the feedback update mechanism in step S7. The evolution of the physical intrinsic equation library is reflected in two aspects: First, key physical parameters are continuously corrected by measured data through the feedback update mechanism in step S7, gradually converging from the initial values ​​set at the factory to the precise values ​​characterizing the true physical state of the system; second, the mathematical form of the partial differential equations describing the conservation laws remains stable because the fundamental physical laws (such as Fourier's law of heat conduction and Newton's laws of motion) themselves remain unchanged.

[0022] Secondly, the present invention provides a self-generated system based on physical eigenvalue equations.

[0023] The core lies in adopting a modular architecture, integrating five major functional modules to realize the hardware execution of the aforementioned self-generated system construction method. This system mainly includes: an industrial-grade control unit: serving as the core computing and control unit, deploying an endogenous digital twin construction module, a self-evolution engine module, a cooperative game execution module, and a feedback self-evolution module, responsible for executing the system's core algorithms and control logic; a multimodal sensor network: used to collect measured physical characteristic parameters of the target system; multiple execution units: receiving control parameters from the industrial-grade control unit and executing specific control actions; non-volatile memory: used to store the physical intrinsic equation library, endogenous digital twin model parameters, and system operation logs; and a hierarchical generation management module: responsible for the identification, incubation, and disintegration management of sub-self-generated systems, realizing the open evolution of the system.

[0024] In one implementation, the coupling relationship between the equations at each level satisfies the following mathematical characterization: Let the physical eigenvalue equation of the i-th level be... ,in Let i be the state variable of the i-th level. Let be the physical parameters of the i-th level. The levels are coupled in two ways: upward causality: the solution to the equation of the i-th level. As a boundary condition or source term of the (i+1)th level equation; downward causality: the solution of the (i+1)th level equation. The parameters acting as constraints on the i-th level equation .

[0025] The engineering definitions of upward and downward causality are as follows: Upward causality refers to the solution of the low-level physical intrinsic equation (such as the thermal deformation field at the component level) serving as the input or boundary condition for the adjacent high-level physical intrinsic equation (such as the thermo-structural coupling analysis at the coupling level), ensuring that the physical behavior of the high-level level is based on the physical processes of the low-level level; Downward causality refers to the solution of the high-level physical intrinsic equation (such as the global maximum allowable deformation at the system level) serving as a constraint on the adjustable parameters of the low-level physical intrinsic equation (such as the thermal expansion coefficient correction range at the component level), ensuring that the physical behavior of the low-level level always remains within the allowable boundaries of the global order of the system. This implementation method fully reflects the interaction mechanism of the hierarchical order in the self-generated system through explicit mathematical representation, solving the problem that traditional single-level models cannot characterize the multi-level coupling characteristics of the system.

[0026] A case study of downward causality in engineering implementation: Taking a lithography machine projection lens as an example. Solving the third-level (system-level) thermoelastic equations yields the global thermal deformation field of the projection lens. This solution acts as a constraint on the Joule thermal-electric field coupling equations at the second level (coupling level)—specifically manifested as: global thermal deformation field. The thermal expansion displacement boundary conditions of each lens in the second level are restricted, which makes the displacement variables in the second level equations... Must meet: ,in The maximum allowable thermal deformation is 10 nm. In the numerical solution, this constraint is achieved by... This is achieved as the upper limit boundary condition for displacement in the second-level finite element solution. Further down, the solution of the second-level equations... It also acts as a constraint on the first level (component level) – specifically, the thermal expansion displacement boundary of each lens limits the upper limit of the temperature gradient in the Fourier heat conduction equation in the first level, ensuring that the local temperature gradient of a single lens does not lead to local deformation exceeding its allocated displacement share.

[0027] In a preferred embodiment, the multiphysics simulator employs a hierarchical simulation strategy: first, a coarse-grained, reduced-order simulation is performed on the entire system to identify subsystems whose state deviation sensitivity exceeds a preset threshold as critical subsystems; then, high-fidelity multiphysics simulation is performed only on the critical subsystems. State deviation sensitivity measures the degree to which changes in the state of a subsystem affect the global state deviation, and is calculated using the following formula: in This represents the global state deviation of the system. Let be the state variable of the j-th subsystem. When When the sensitivity threshold is greater than the preset threshold (typical value 0.3), the subsystem is identified as a critical subsystem.

[0028] Furthermore, the multiphysics simulator employs an intrinsic orthogonal decomposition (POD) model to reduce the simulation degrees of freedom from... compressed to The order of performance is on the order of magnitude, with a single simulation time of less than 100ms. The following performance estimates are based on an exemplary system configuration and are intended to illustrate the feasibility of the order reduction method. Actual performance metrics will vary depending on the physical complexity and hardware configuration of the target system.

[0029] This example uses an intrinsic orthogonal decomposition (POD) model for order reduction. For a degree of freedom of... The original system, the POD method extracts the eigenvalues ​​of the covariance matrix by solving the eigenvalue problem. One dominant mode. The computational cost of eigenvalue decomposition is... In this example configuration, the simulation is performed offline by the simulation workstation and is not included in the real-time simulation time. The computational cost of a single forward simulation after order reduction is proportional to the square of the degrees of freedom. After order reduction, the floating-point computation cost of a single forward simulation is approximately... FLOPs. This estimate is based on the following derivation: original system degrees of freedom The finite element method is used to solve the system of partial differential equations. The computational cost of a single simulation is approximately [missing information]. FLOPs. Degrees of freedom after order reduction. The computational workload is approximately FLOPs, order reduction ratio reaches This invention utilizes a Xilinx Artix-7 FPGA, containing 90 DSP units. Each DSP unit can complete one multiply-accumulate operation within one clock cycle, and the system clock frequency is 100MHz. In this example configuration, the time for a single simulation is approximately: It should be noted that the above calculations only cover the forward simulation portion of the reduced-order system. The complete simulation process also includes data preprocessing, post-processing, and result output, and the actual end-to-end simulation time varies depending on the specific implementation.

[0030] In one implementation, the cognitive free energy is calculated in real time: while outputting the predicted future state deviation value in step S5, the system evolution prediction unit also calculates the cognitive free energy F of the endogenous digital twin regarding the current state of the target system, which is used to quantify the information theory deviation between the predicted distribution and the measured distribution of the endogenous digital twin. in The predicted distribution of the target system's current state by the endogenous digital twin. This is the posterior distribution based on measured data from a multimodal sensor network. For Kullback-Leibler divergence, This represents the negative log-likelihood of the measured data.

[0031] In engineering implementation, Take the state prediction probability distribution of the output of the endogenous digital twin. The empirical distribution is obtained by statistically processing the measured physical characteristic parameters collected in step S4. When the endogenous digital twin consistently provides accurate predictions for a certain subsystem or operating condition, and When the predictions are highly consistent, F approaches zero; when the predictions are consistently biased, F increases.

[0032] In engineering implementation, the marginal probability of observation Approximate calculation using Monte Carlo sampling: ,in To predict the distribution The sampled particles, M is the number of sampling points (typically 100); observation likelihood Take Gaussian distribution form , Let R be the observation function, and R be the observation noise covariance matrix, determined by the sensor's factory accuracy specifications. This approximation method is based on the standard Monte Carlo estimation principle of variational free energy, and can achieve this using the predicted particle set of an endogenous digital twin without adding additional hardware sensors. Online estimation. The health threshold for cognitive free energy F is... .when At this point, the coupling between the endogenous digital twin and the target system is in a healthy steady state, and the digital twin's understanding of the physical world is accurate and reliable. At this point, significant cognitive biases may occur in the endogenous digital twin, requiring adjustments to the internal model parameters or triggering retraining.

[0033] In one implementation, the AI ​​agent model training unit includes a graph neural network encoder and a physical information decoder. The graph neural network encoder employs a 3-layer graph convolutional network to encode the physical topology of the target system into a graph structure (with physical components as graph nodes and physical coupling relationships as graph edges), outputting 128-dimensional graph node feature vectors. The physical information decoder employs a 2-layer fully connected network to decode the graph node feature vectors into predicted states in the physical space. This implementation leverages the inherent adaptability of graph neural networks to the topology of physical systems, significantly improving the model's prediction accuracy and generalization ability.

[0034] In a preferred embodiment, the graph node feature vectors are further mapped to a linear evolution space using a nonlinear dynamics linearization operator. Linear evolutionary deduction is then performed in this space, outputting the predicted state deviation values ​​for future time windows. The nonlinear dynamics linearization operator can be the Koopman operator, DeepONet, or Physical Information Long Short-Term Memory (PI-LSTM) network. This implementation transforms the complex nonlinear dynamics problem into a simple linear evolution problem, significantly reducing the computational complexity of state prediction.

[0035] In one implementation, a sub-self-generating system embryo recognition and incubation mechanism is also included. The ternary coupling strength index of each functional component set within the target system is calculated in real time. This indicator is a core quantitative metric for measuring whether a functional module has the potential to become a self-sustaining sub-system. The calculation formula is as follows: in and , For information relevance, For energy coupling degree, The degree of material cohesion. A ternary coupling strength index. The theoretical basis stems from the ontology of self-generating systems: for a system to become an independent sub-self-generating system, it must simultaneously possess three conditions: information self-consistency (C), energy self-sufficiency (E), and matter condensation (M). Information correlation degree It measures the completeness of the information loop within the subsystem—when the amount of internal information exchange far exceeds the amount of exchange with the parent system, the subsystem has formed an independent information field state; energy coupling degree. It measures the degree of energy self-sufficiency of a subsystem; the degree of matter condensation. The independence of the subsystem's physical structure was measured. During calculation, the coordinates were normalized to the system's overall geometric center as the origin and the system's maximum outer envelope size as the reference. The interval eliminates the influence of physical scale on the judgment result.

[0036] When a certain component collection The differentiation threshold is exceeded for M consecutive collection cycles. When the value reaches 0.7 (typically), it is identified as a potential sub-self-generating system embryo and an incubation process is initiated. The incubation process incorporates the embryo as an independent decision-making agent into a hierarchical collaborative game framework, gradually allocating dedicated resources until its ternary coupling strength index stably exceeds the independent survival threshold. (Typical value 0.9), thus becoming an independent sub-self-generated system with local operational identity. Local operational identity refers to the ability of a sub-self-generated system to autonomously define and maintain its core survival objectives, and is a marker of the independent existence of a sub-self-generated system.

[0037] Quantifiable criteria for local operation identity (core criterion): When the ternary coupling strength index When the value consistently exceeds 0.9 for N consecutive acquisition cycles, the local operational identity of the self-generated subsystem is mathematically established. This criterion is a necessary and sufficient condition. A stability exceeding 0.9 necessarily indicates that local operational consistency has been established. Functional verification criterion (auxiliary criterion): The sub-self-generated system successfully completes a preset number of core functional tasks (typically 10 times) without any control instructions from the parent system. The standard for a single task success is that the sub-self-generated system independently completes a complete core functional loop (perception → decision → execution → feedback), and the execution deviation does not exceed 120% of the deviation of the parent system when performing the same task. The functional verification criterion is used to confirm the actual effectiveness of local operational consistency at the engineering level, rather than the mathematical criterion itself.

[0038] This implementation method is the first to realize the ability of intelligent systems to generate new subsystems, solving the fundamental defects of traditional system architectures that are static and fixed and cannot be autonomously expanded.

[0039] In a preferred embodiment, the payoff function of the asymmetric resource allocation game during the incubation process of the self-generating subsystem is configured as follows: in , The growth rate of the embryo. The resource security constraints of the parent system are satisfied. This payoff function ensures that the growth rate of the child self-generated system is maximized while prioritizing the security of the parent system.

[0040] The game-playing process with the parent system in the incubation process refers to gradually allocating dedicated computing resources, energy quotas, and physical actuators to the embryo using an asymmetric weight allocation strategy, while ensuring the resource security constraints of the parent system are met. In this asymmetric weight allocation strategy, the growth rate weight factor of the embryo is no less than 0.7, and the resource security constraint weight factor of the parent system is no greater than 0.3. The differentiation threshold... The typical value is 0.7, which is the autonomous operation threshold. The typical value for is 0.9, the value of M ranges from 10 to 50, and the value of N ranges from 20 to 100. It should be noted that the above values ​​are preferred values ​​based on a large number of engineering experiments. Those skilled in the art can adjust the above thresholds within a reasonable range according to the system complexity, operating environment, and reliability requirements of the specific application scenario.

[0041] In one implementation, a sub-self-generating system disintegration mechanism is also included: when an independent sub-self-generating system For P consecutive acquisition cycles, the data is below the disintegration threshold. When the typical value is 0.6, the parent system initiates a dismantling procedure, reclaims all resources it occupies, and migrates its knowledge and experience to the parent system's knowledge pool. The difference between the differentiation threshold and the dismantling threshold is used to create a hysteresis judgment, preventing frequent oscillations and switching of the system near the critical value. This implementation achieves dynamic recycling and reuse of system resources, improving resource utilization efficiency. The knowledge and experience migration refers to packaging the endogenous digital twin model parameters of the sub-self-generated system, key physical parameters corrected by measured data in the physical intrinsic equation library, and task execution history (including control parameters and corresponding performance scores) into standardized knowledge units and storing them in the parent system's knowledge pool. When the parent system encounters similar task scenarios in subsequent operations, it retrieves matching historical knowledge units from the knowledge pool, uses their physical parameters as initial reference values ​​for current control, and improves control efficiency. After the sub-self-generated system dismantles, its physical actuators are reclaimed by the parent system, and its computing resources and energy quotas are reintegrated into the parent system's global resource pool.

[0042] In one implementation, a hierarchical verification mechanism is also included: three levels of verification are performed on the generated control parameter combinations, including component-level physical and safety constraint verification, coupling-level multiphysics constraint verification, and system-level consistency verification. Only control parameter combinations that pass all three levels of verification can be submitted for execution. This implementation ensures that the control strategy meets physical constraints and safety requirements at all levels, avoiding the execution of unsafe control actions.

[0043] In one implementation, a counterfactual reasoning mechanism is also included: using the endogenous digital twin as a baseline state, positive and negative perturbations are applied to the control parameters; the amplitude of the perturbation is determined based on the product of the control resolution and the safety threshold of the corresponding execution unit: perturbation amplitude = min(control resolution × 10, safety threshold × 20%); where the control resolution is the minimum parameter change that the execution unit can stably execute, and the safety threshold is the maximum parameter change range allowed by the execution unit; two counterfactual branches are generated respectively, forward inference is performed along each branch, the combination of control parameters corresponding to the branch with the best performance is compared with the result of the optimization solution, and the best result is output.

[0044] In one implementation, when the system of partial differential equations contains multiple equations, the physical loss function is a weighted sum of the squared residuals of each equation: in This represents the total number of equations. This represents the total number of sampling points. For the first The number of sampling points corresponding to each equation satisfies ; For the first The weight coefficients of the equations satisfy... The initial values ​​of the weight coefficients are determined based on the characteristic time scale and characteristic energy scale of the physical process described by each equation: the shorter the characteristic time scale and the higher the characteristic energy of the physical process, the greater the initial weight of the corresponding equation. During training, the weight coefficients can be dynamically adjusted according to the convergence rate of the residuals of each equation—equations that converge more slowly receive greater weights to accelerate overall convergence.

[0045] In engineering implementation, the total loss function Weights of equations within the physical loss term This constitutes a two-layer nested weight system: first, according to... The physical loss is obtained by weighted summation of the squared residuals of each equation. , and then The weight will With data loss The total loss is obtained by weighted summation. In engineering practice, the initial approach is to address the total loss level through... The overall emphasis on balancing data fitting and physical constraints (typical value) ), and then through the physical loss The importance differences of each physical equation are refined, and the two sets of weight parameters can be adjusted independently and jointly optimized.

[0046] For the three-level equation library, the first level contains four equations, each with an initial weight of 0.25; the second level contains two coupled equations, each with an initial weight of 0.3; and the third level contains two global equations, each with an initial weight of 0.2. The weights within each level are evenly distributed.

[0047] For linear or weakly nonlinear partial differential equations, according to the uniqueness theorem, the solution satisfying the boundary conditions is unique, and the physical loss term can precisely constrain the model to the vicinity of the correct solution. For the strongly nonlinear multiphysics coupled system involved in this invention, the uniqueness of the solution generally does not hold—multiple stable solutions may exist for the same set of partial differential equations and boundary conditions. In this case, the core role of the physical loss term is not "selecting a unique solution," but rather "excluding non-physical solutions": any solution that does not satisfy the basic conservation laws (mass conservation, momentum conservation, energy conservation) will be assigned an extremely high loss value by the physical loss term, thus being excluded from the model's solution space. When multiple stable solutions satisfying the physical constraints exist, the physical loss term retains them all in the feasible solution space; the data loss term, based on measured observation data, selects the unique solution that best matches the observations from these physically reasonable solutions. The synergistic effect of both ensures that the model neither violates fundamental physical laws nor fails to accurately fit the actual observed data. The solution ultimately selected by the model is determined by both physical loss (excluding non-physical solutions) and data loss (fitting the observed data), and is the solution that best matches the actual observations in a physically reasonable subspace.

[0048] when At that time, theoretical derivation and a large number of numerical experiments show that when the data coverage is... At that time, the prediction accuracy of the present invention .

[0049] In one implementation, the search for the approximate solution of the Pareto front adopts a simplified online version of the non-dominated sorting genetic algorithm (NSGA-II) based on crowding. The specific process is as follows: (1) Population initialization: Based on the actual control parameters of each execution unit, a candidate population is randomly generated within ±10%, with a population size of 50; (2) Non-dominated sorting: Calculate the value of each candidate solution on each game payoff function and perform fast non-dominated sorting; (3) Crowding calculation: Within the same non-dominated level, sort the solutions in descending order of crowding of each objective and select the top 20 solutions to enter the next generation; (4) Search termination: When the search time reaches the preset upper limit (usually 30% of the system control cycle) or there is no significant improvement in the Pareto front for three consecutive generations, the search is terminated; (5) Optimal solution selection: From the final Pareto front, select the solution with the closest Euclidean distance to the ideal point (where all objectives are optimal) as the parameter to be distributed.

[0050] The time complexity of this algorithm is O(n). Where G is the iteration number (typically 10-20), P is the population size (typically 50), and M is the number of targets (typically 2-3). On the STM32H750 hardware platform, the single solution time is approximately 15-25ms, satisfying the real-time constraint of the control cycle (100ms).

[0051] In one implementation, in step S3, the virtual system dataset is divided into a training set and a validation set in a ratio of 8:2, and an initial endogenous digital twin is obtained after 30 training rounds.

[0052] In one implementation, the update strategy of the parameter update unit in step S7 is as follows: an incremental update is performed immediately after each state acquisition cycle is completed; or a batch update is performed after accumulating data for L acquisition cycles; the incremental update only adjusts the output layer parameters of the endogenous digital twin, and the batch update adjusts all trainable parameters of the endogenous digital twin.

[0053] In one implementation, the combination of key operating parameters in step S2 includes: the operating points of each execution unit at the upper and lower boundaries of its allowable adjustment range, the combined operating points of multiple execution units simultaneously in extreme states, and the abnormal operating points in the history of system operation where the state deviation exceeds a preset threshold.

[0054] In one implementation, step S7 is followed by a knowledge transfer step: a knowledge transfer unit deployed on a cloud server clusters the endogenous digital twin parameters of all target systems and sub-self-generated systems of the same type to extract common group knowledge; when the new system goes online, this common group knowledge is used as the initial model parameters for the AI ​​agent model training unit. This implementation can improve the initial performance of the new system by more than 50% and significantly shorten the debugging cycle of the new system.

[0055] In one embodiment, the industrial-grade control unit adopts an FPGA+MCU heterogeneous architecture, where the FPGA is responsible for high-speed signal processing and real-time control, and the MCU is responsible for complex algorithm calculations and system management; the FPGA and the MCU are connected via an AXI bus.

[0056] In one embodiment, the multimodal sensing network includes at least two of the following: a thermal field sensing unit (thermocouple or thermistor), an electric field sensing unit (voltage sensor or current sensor), a mechanical sensing unit (accelerometer or strain gauge), and a timing sensing unit (clock skew detection unit), with each sensing unit having a sampling frequency of 1 kHz to 100 kHz. This embodiment ensures comprehensive acquisition of the system's multiphysics state information.

[0057] The parameter update unit employs a Bayesian linear regression algorithm with a Gaussian prior distribution. The specific execution flow of the Bayesian update is as follows: First, the Gaussian prior distribution of the parameters is set. After receiving new observation data D, calculate the posterior distribution according to Bayes' theorem. For linear regression models, the posterior distribution remains Gaussian, allowing for direct updates of the mean and covariance via analytical formulas without iterative sampling, thus adapting to the limited computing power of embedded devices. When the effective sample size is less than a preset threshold, complete Bayesian posterior inference is used to update parameters to ensure update accuracy; when the effective sample size is greater than or equal to the preset threshold, stochastic gradient descent update mode is switched to improve update efficiency.

[0058] In one implementation, the feedback self-evolution module also includes a knowledge transfer unit deployed on a cloud server to enable knowledge sharing and transfer between systems of the same type.

[0059] Compared with the prior art, by adopting the above technical solutions, the present invention has achieved the following significant technical effects: (1) Truly autonomous operation: The system autonomously derives the control strategy from the hierarchical endogenous order mathematical representation. Taking the lithography machine scenario as an example, it only needs one manual calibration for 720 consecutive hours of operation, and the frequency of manual intervention has been reduced from once every 4 hours on average to once every 720 hours. (2) Full physical self-consistency: Physical laws are directly used as part of the loss function to participate in model training. Hierarchical verification ensures that the control strategy meets the physical constraints at each level, and the physical consistency error is less than 1%. (3) Pre-judgment control: The optimal control strategy is generated before the state deviation occurs. Compared with the traditional PID control, the amplitude of the system state deviation is significantly reduced (from 8.5nm to 7.2nm in the lithography machine embodiment), the response time is significantly shortened (from 120ms to 85ms in the lithography machine embodiment, a reduction of 29.2%), and there is no overshoot. (4) Multi-level complete modeling: The hierarchical physical intrinsic equation library fully reflects the multi-level order system of the system from components to the whole. By linking the bidirectional coupling relationship of upward causality and downward causality, it solves the problem that traditional single-level models cannot characterize the coupling characteristics of complex systems. The hierarchical verification mechanism ensures the physical and functional safety of all control actions, and the system operation reliability is improved by more than 95%. (5) Open evolution capability: The introduction of the ternary coupling strength index realizes the automatic identification and incubation of sub-self-generated systems. The system can endogenously generate new sub-self-generated systems, dynamically expand its own functional boundaries and physical scale, and make the system infinitely scalable and extremely resistant to destruction. (6) By introducing the cognitive free energy F, the overall deviation between the predicted distribution and the measured distribution is accurately quantified, the cognitive blind spots of the digital twin are identified, and targeted model retraining is driven, so that the endogenous digital twin can maintain a high-precision cognition of the physical world in long-term operation, and realize the dual-dimensional health assessment of the endogenous digital twin's "internal self-consistency" and "external adaptation". Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the hardware architecture of a self-generated system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the internal structure of the endogenous digital twin construction layer according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hierarchical structure and coupling relationship of a hierarchical physical intrinsic equation library according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a hierarchical verification process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the complete workflow of a self-generated system according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the self-generated system hierarchy generation and life cycle according to an embodiment of the present invention.

[0061] Component labels in the attached diagram: 1-Industrial-grade control unit; 2-Multimodal sensor network; 3-Execution unit; 4-Non-volatile memory; 5-Hierarchical generation management module; 11-Endogenous digital twin construction module; 12-Self-evolution engine module; 13-Cooperative game execution module; 14-Feedback self-evolution module; 111-Physical intrinsic equation library; 112-Multiphysics field simulator; 113-AI agent model training unit; 121-System evolution prediction unit; 122-Numerical optimization solver; 123-Hierarchical verification unit; 131-Execution unit intelligent agent modeling unit; 132-Multi-agent cooperative solver; 141-Parameter update unit; 142-Knowledge transfer unit; 51-Hierarchical generation monitoring unit; 52-Incubation control unit; 53-Lifecycle management unit. Detailed Implementation

[0062] The technical solution of the present invention will be described in complete detail below with reference to the accompanying drawings and specific embodiments. All formulas, parameters, hardware configurations, and calculation processes are given with clear values ​​and implementation methods. Those skilled in the art can completely reproduce the present invention based on the contents of this specification.

[0063] Example 1: Self-generating maintenance system for thermal deformation of projection lens in semiconductor lithography machine

[0064] This embodiment is applied to the thermal deformation maintenance scenario of the EUV lithography machine projection lens in a 14nm process production line. The goal is to control the thermal deformation of the projection lens within 10nm to meet the lithography resolution requirements.

[0065] The specific implementation and functions of each module of the self-generated system in this embodiment are as follows: Industrial-grade control unit: Xilinx Artix-7 XC7A100T FPGA + STM32H750ZI The MCU (dual-core 240MHz) serves as the core computing and control unit of the system, deploying all algorithm modules; a multimodal sensor network includes 12 PT1000 platinum resistance temperature sensors (accuracy ±0.01℃, sampling frequency 1kHz), 4 high-precision laser displacement sensors (accuracy ±0.1nm, sampling frequency 10kHz), and 2 high-precision current sensors (accuracy ±0.1mA) for acquiring multi-physics state parameters of the projection lens; multiple execution units include 8 high-precision cooling water flow rate controllers (accuracy ±0.01L / min, response time <10ms) and 4 PZT piezoelectric drivers (accuracy ±0.01nm, response time <1ms) for executing thermal deformation compensation and control actions; non-volatile memory: 64GB industrial-grade eMMC5.1 with a read / write speed of 300MB / s for storing the physical intrinsic equation library, endogenous digital twin model parameters, and system operation logs; and a hierarchical generation management module deployed on the STM32H750. Within the MCU, the system is responsible for the identification, incubation, and disintegration management of the sub-autogenous systems. The endogenous digital twin construction module, deployed in the FPGA, includes a physical intrinsic equation library, a multiphysics simulator, and an AI agent model training unit, used to construct the initial endogenous digital twin. The autogenous evolution engine module, deployed in the FPGA, includes a system evolution prediction unit and a numerical optimization solver, used to predict future state deviations and generate optimal control parameters. The cooperative game execution module, deployed in the FPGA, includes a multi-agent cooperative solver and a hierarchical verification unit, used to filter and distribute optimal control parameter combinations. The feedback autogenous evolution module, deployed in the MCU, includes a parameter update unit and a knowledge transfer unit, used to update key physical parameters in the physical intrinsic equation library.

[0066] System parameters: control cycle 100ms, ternary coupling strength calculation cycle 10s, steady-state calibration cycle 1000 control cycles. Parameter source explanation: The 100ms control cycle matches the characteristic time scale (approximately 1s) of thermal deformation of the lithography machine projection lens, satisfying the Nyquist sampling theorem; the 10s ternary coupling strength calculation cycle is much longer than the control cycle, avoiding excessive computation and resource consumption.

[0067] Implementation process

[0068] Step S1: Constructing a hierarchical physical intrinsic equation library is executed by the physical intrinsic equation library unit in the endogenous digital twin construction module. Upon initial system power-on, the three-level initial physical intrinsic equation library, pre-set by the designer based on the known physical laws of fused silica optical glass and the factory calibration parameters of the lithography machine, is automatically loaded and stored in non-volatile memory: First level (component level): Stores the thermal expansion constitutive equation, Fourier thermal conduction equation, and stress optical coefficient equation (3 equations in total) for the fused silica optical glass. Thermal expansion constitutive equation: Input parameters: (Initial lens thickness, from the equipment manufacturer's technical specifications). (Coefficient of thermal expansion of fused silica, from the official materials handbook). (Temperature change). Calculation process: Output result: For every 1°C change in temperature, the lens thickness changes by 0.11 nm.

[0069] Fourier heat conduction equation: Input parameters: (Density of fused silica) (Specific heat capacity) (thermal conductivity) (Density of internal heat source generated by laser irradiation).

[0070] Stress optical coefficient equation: .

[0071] Input parameters: (Fired silica stress optical coefficient). (Thermal stress). Output results: .

[0072] The second level (coupling level) stores the Joule thermo-electric field coupling equation and the thermo-structure coupling equation (a total of 2 equations).

[0073] The third level (system level): stores the global thermoelastic equations and wavefront error propagation equations of the projection lens (a total of 2 equations).

[0074] The equations at each level are coupled bidirectionally through upward and downward causality: the solutions of lower-level equations serve as boundary conditions for higher-level equations, and the solutions of higher-level equations serve as constraints on the parameters of lower-level equations.

[0075] Step S2: The multiphysics simulation generates a virtual dataset, which is executed by the multiphysics simulator in the intrinsic digital twin building block. The multiphysics simulator calls the partial differential equations from the physical intrinsic equations library to perform high-fidelity multiphysics simulations based on fundamental physical laws.

[0076] First, a coarse-grained reduced-order simulation of the entire system is performed to calculate the state deviation sensitivity of the projection mirror system: Input parameters: (Wavefront error, in nm) (Average temperature of the projection lens, unit: °C); Calculation process: at the reference temperature Take small temperature changes nearby The simulation yielded the corresponding wavefront error changes. Output result: The value is well above the threshold of 0.3, therefore the projection mirror system is identified as a critical subsystem.

[0077] Subsequently, a high-fidelity finite element simulation was performed on the projection lens, covering the entire operating range of laser power from 100W to 500W and cooling water temperature from 20℃ to 25℃, generating 1000 sets of virtual datasets. Each dataset contains temperature field distribution, thermal deformation field distribution, and wavefront error data. The total simulation time was approximately 7.8 hours.

[0078] Step S3: Training the initial endogenous digital twin is performed by the AI ​​agent model training unit in the endogenous digital twin construction module. A graph neural network encoder-physical information decoder architecture is adopted: the virtual dataset is divided into a training set (800 sets) and a validation set (200 sets) in an 8:2 ratio. The total loss function is a weighted sum of data loss and physical loss.

[0079] Model initialization: Model parameters are randomly initialized using the Xavier initialization method. Before training, the model with initialized parameters is used to predict on the validation set to obtain the initial data loss. Initial physical loss .

[0080] Data loss calculation: Input parameters: The wavefront error predicted by the model. This represents the actual wavefront error obtained from the simulation.

[0081] Physical loss calculation: Input parameters: (Total number of equations: 3 at the first level + 2 at the second level + 2 at the third level) (Total number of sampling points) For the first The number of sampling points corresponding to each equation satisfies .

[0082] Weighting: The total weight for each level is 1 / 3. In the first level, each of the three equations has a weight of 1 / 9; in the second level, each of the two equations has a weight of 1 / 6; and in the third level, each of the two equations has a weight of 1 / 6. Verification: ,satisfy The constraints.

[0083] Total loss calculation:

[0084] After 30 training iterations, the lumped loss was verified to converge to... ,in , The initial endogenous digital twin is obtained and stored in eMMC non-volatile memory.

[0085] Parameter source explanation: Loss function weights The optimal value was determined based on 1000 sets of numerical experiments, balancing data fitting accuracy and physical consistency; the 30 training epochs were determined by early stopping – training stopped when the validation set loss no longer decreased for 5 consecutive epochs; the total weight of each level in the physical loss function is evenly distributed, reflecting that the three levels have equal importance in the physical loss function; the weight of each equation within a level is evenly distributed, reflecting that each physical process within the same level is equally important.

[0086] Step S4: Real-time system status acquisition is performed by a multimodal sensor network. After each 100ms control cycle: 12 PT1000 platinum resistance temperature sensors synchronously acquire temperature data of each lens of the projection lens; 4 laser displacement sensors synchronously acquire axial displacement data of the projection lens; 2 high-precision current sensors synchronously acquire operating current data of the PZT driver. All acquired data is transmitted to the industrial-grade control unit via the AXI bus as input for subsequent status prediction.

[0087] Step S5: Predicting future state deviations is performed by the system evolution prediction unit in the self-evolving evolution engine module. The measured physical characteristic parameters collected in step S4 are input into the endogenous digital twin generated in step S3: the graph neural network encoder encodes the measured temperature and displacement data into 128-dimensional graph node feature vectors; the Koopman operator maps the graph node feature vectors to the linear evolution space; forward inference is performed in the linear evolution space, and the predicted values ​​of thermal deformation and wavefront error of the projection lens within the next 100ms time window are output.

[0088] Step S6: Optimal control parameters are generated in advance and executed collaboratively by the numerical optimization solver and the collaborative game execution module in the self-evolution engine module. When the predicted value output in Step S5 shows that the wavefront error within the next 100ms will exceed the preset threshold of 10nm: the objective function is to minimize the wavefront error of the entire system; the constraints are the adjustment range of the 8-channel cooling water flow rate controller (0-10L / min) and the voltage range of the 4-channel PZT driver (0-100V); the sequential quadratic programming (SQP) algorithm is used to perform optimization and solve the problem, generating the optimal cooling water flow rate and PZT driving voltage combination covering the next 100ms time window before the predicted thermal deformation time.

[0089] The generated combination of control parameters first undergoes hierarchical verification: Level 1 verification: verifying whether the cooling water flow rate and PZT voltage meet the physical constraints and functional safety constraints of the corresponding execution unit; Level 2 verification: verifying whether the control parameters meet the thermal-structural coupling constraints of the projection lens; Level 3 verification: verifying whether the control parameters meet the global coordination and consistency constraints of the lithography machine; only after passing all three levels of verification does it enter the counterfactual reasoning verification stage.

[0090] Counterfactual reasoning verification: Baseline state: Current cooling water flow rate PZT driving voltage Disturbance calculation: Cooling water flow rate controller: Control resolution = 0.01L / min, safety threshold = 10L / min, disturbance 1 = PZT driver: Control resolution = 0.01V, safety threshold = 100V, scrambling amplitude 2 =

[0091] Generates two counterfactual branches: the positive branch , Reverse branch , Perform forward inference along the two branches to obtain the predicted wavefront error value of 6.8nm for the forward branch, 7.5nm for the reverse branch, and 7.2nm for the baseline branch; then select the optimal combination of control parameters corresponding to the forward branch.

[0092] Finally, multi-agent collaboration is performed: the multi-agent collaborative solver collects candidate control parameters generated by 8 cooling water flow rate controller agents and 4 PZT driver agents, performs non-dominated sorting based on the game payoff function, selects the control parameter combinations corresponding to the Pareto front approximate solution, and distributes them to each execution unit for execution.

[0093] Actual computational complexity calculation of the NSGA-II algorithm: Input parameters: (Iterative Algebra) (Population size) (Objectives: Minimize wavefront error and minimize energy consumption) Calculation process: The output of the next floating-point operation takes approximately 18ms to execute on the STM32H750 MCU, which is much less than 30% (30ms) of the 100ms control cycle limit.

[0094] Step S7: The feedback update of the physical intrinsic equation library is executed by the parameter update unit in the feedback self-evolution module. After step S6: The multimodal sensor network collects measured thermal deformation and wavefront error data of the projection lens after control. The parameter update unit estimates the parameters a priori using the measured values ​​as observations and the predictions of the pre-execution endogenous digital twin. The parameter update unit uses a Bayesian linear regression algorithm. In the early stage of system operation, the effective sample size (cumulative number of control cycles) is 0, which is less than the preset threshold of 1000. Therefore, a complete Bayesian posterior inference is used to update the parameters and update the key physical parameters such as the thermal expansion coefficient and thermal conductivity coefficient of fused silica glass in the physical intrinsic equation library of the parent system. The updated physical intrinsic equation library is stored in non-volatile memory for simulation and prediction in the next control cycle.

[0095] S4-S7 closed-loop continuous operation: After the system completes S7, it immediately returns to S4 to start the next control cycle until the system stops.

[0096] Hierarchical generation and evolution of sub-self-generated systems

[0097] While the S4-S7 operation loop continues to execute, the hierarchical generation management module synchronously executes the identification and incubation process of the sub-self-generated system: In the early stage of system operation: all operation and maintenance tasks are uniformly executed by the parent system, the measured thermal deformation control accuracy is 8.5nm, the response time is 120ms, and the cooling system energy consumption is 150W.

[0098] Embryo Recognition Stage: After 720 hours of continuous system operation, the hierarchical generation monitoring unit begins calculating the ternary coupling strength of the component set responsible for projector lens temperature monitoring and cooling control (including 4 laser displacement sensors, 8 cooling water flow rate controllers, and 4 PZT drivers) every 10 seconds: Information Correlation: Monitoring all communication interfaces related to projector lens temperature control and statistically analyzing the data transmission volume per unit time. Internal Information Exchange Volume This refers to the data volume of the position compensation signal directly sent by the laser displacement sensor to the cooling water flow rate controller (measured at 85 Mbps, statistically analyzed by monitoring the direct connection signal line between the sensor and the controller). External information exchange volume. This represents the amount of instruction and status feedback data between this component set and the parent system MCU (measured at 35Mbps, statistically analyzed by monitoring bus communication between the MCU and each device).

[0099] Energy coupling degree: internal energy exchange amount This refers to the drive energy fed back to the cooling water controller by the PZT driver during thermal deformation compensation (measured at 8.4W, obtained by integrating the product of the PZT driver's current and voltage). External energy supply. The total power supplied by the parent system to this component assembly (measured at 12W, read by the voltage and current monitoring chip of the power module).

[0100] Material cohesion: Calculate the spatial coordinates of 12 temperature sensors and 4 laser displacement sensors around the projection lens (normalized to the [0,1] interval, with the geometric center of the projection lens as the origin and the maximum outer envelope size of the lens as the reference). Normalized spatial coordinate variance. .

[0101] Ternary coupling strength:

[0102] After running for another 24 hours, It rose to 0.78. Reaching 0.707, and consistently exceeding 0.7 for 30 consecutive calculation cycles, it was identified as a sub-self-generated system embryo.

[0103] Parameter source explanation: Ternary coupling strength weight Based on the central role of information in self-generated systems, information relevance is the most critical indicator for judging whether a module possesses independent decision-making capabilities; differentiation threshold. This indicates that when the amount of information exchanged within a module exceeds 70% of the total information, a relatively independent information loop has been formed; independent survival threshold. This indicates that when the amount of information exchanged within a module exceeds 90% of the total information, it possesses fully independent decision-making and execution capabilities; disintegration threshold. This creates a hysteresis range with the differentiation threshold, thus avoiding critical oscillations.

[0104] Incubation Phase: The incubation program is initiated, allocating the embryo with a dedicated DSP cluster within the FPGA, a dedicated 12V power supply channel, and exclusive control over all eight cooling water flow rate controllers. After 168 hours of incubation, the embryo successfully completed 10 independent thermal deformation control tasks consecutively. It stabilized above 0.92 and officially became an independent "projection lens thermal deformable self-generated system".

[0105] During the incubation stage, the payoff function of the asymmetric resource allocation game is calculated as follows: Input parameters: (Prioritizing embryo growth, determined based on engineering experience) The rate of change of the ternary coupling strength in the embryo was calculated. : Parent system resource security constraint satisfaction, current value is 0.95 (percentage of remaining computing resources in the parent system). Calculation process: Output result: The profit value is positive, indicating that continuing incubation is beneficial to the system as a whole.

[0106] The functional verification of local operation consistency was performed as follows: the parent system stopped sending any control commands to the embryo, retaining only safety boundary monitoring. The embryo independently executed the complete S4-S7 closed-loop operation—acquiring position data from the laser displacement sensor (S4), calculating the cooling water flow rate and PZT driver voltage using the endogenous digital twin (S5-S6), sending control commands to the cooling water flow rate controller and PZT driver (S6), and monitoring the thermal deformation after control and updating the model (S7). The criterion for successfully completing 10 consecutive complete closed-loop tasks was: the closed-loop control deviation of each task did not exceed 120% of the parent system's historical control deviation (8.5nm), i.e., not exceeding 10.2nm. During the 168-hour incubation period, the embryo successfully completed all 10 independent tasks, with an average control deviation of 9.1nm, which was below the criterion.

[0107] Independent operation phase: The parent system no longer directly intervenes in thermal deformation operation and maintenance, but only sets a safety boundary of ≤10nm for thermal deformation through downward causality. The sub-self-generated system autonomously optimizes and controls its control strategy, resulting in improved measured performance: thermal deformation control accuracy improved from 8.5nm to 7.2nm (15.3% improvement), response time shortened from 120ms to 85ms (29.2% reduction), and cooling energy consumption reduced from 150W to 132W (12% reduction).

[0108] Example of parameter update: After running for approximately 240 hours, the coefficient of thermal expansion... From initial value Updated to The thermal conductivity k was updated from 1.38 W / (m·K) to 1.42 W / (m·K). The mathematical form describing the conservation laws in the equation library remains stable.

[0109] Example 2: 300MW Industrial Boiler Combustion Self-Generating Control System

[0110] This embodiment is applied to the combustion control scenario of a supercritical 300MW pulverized coal boiler, with the goal of improving combustion efficiency and reducing NOx emission concentration.

[0111] The specific implementation and functions of each module are as follows: Industrial-grade control unit: Xilinx Zynq-7020 FPGA + ARM Cortex-A9 dual-core processor (800MHz), serving as the core computing and control unit of the system; Multimodal sensor network: 24 K-type thermocouples (accuracy ±0.5℃), 8 pressure sensors (accuracy ±0.1kPa), 2 flue gas analyzers (NOx accuracy ±1mg / m³), 4 oxygen sensors, and 1 flame image sensor (200fps), used to collect multi-physics field state parameters of the boiler combustion system; Multiple execution units: 12 fuel regulating valves (accuracy ±0.1%), 8 forced draft regulating valves, 4 induced draft regulating valves, and 16 secondary air dampers, used to execute combustion control actions; Non-volatile memory: 32GB industrial-grade eMMC, used for... The system includes: a physical intrinsic equation library, endogenous digital twin model parameters, and system operation logs; a hierarchical generation management module deployed in an ARM Cortex-A9, responsible for the identification, incubation, and disintegration management of sub-autogenous systems; an endogenous digital twin construction module deployed in an FPGA, containing a physical intrinsic equation library, a multiphysics simulator, and an AI agent model training unit; an autogenous evolution engine module deployed in an FPGA, containing a system evolution prediction unit and a numerical optimization solver; a cooperative game execution module deployed in an FPGA, containing a multi-agent cooperative solver and a hierarchical verification unit; and a feedback autogenous evolution module deployed in an ARM Cortex-A9, containing a parameter update unit and a knowledge transfer unit.

[0112] System parameters: control cycle 200ms, ternary coupling strength calculation cycle 30s. Parameter source explanation: The 200ms control cycle matches the characteristic time scale of the boiler combustion system (approximately 2s) and satisfies the Nyquist sampling theorem.

[0113] Implementation process

[0114] Step S1: Constructing a hierarchical physical intrinsic equation library is performed by the physical intrinsic equation library unit in the endogenous digital twin construction module. Upon initial power-up, the system automatically loads a three-level initial physical intrinsic equation library pre-set by the designer based on known physical laws of pulverized coal combustion and boiler factory calibration parameters: Level 1 (Component Level): Stores the pulverized coal combustion reaction kinetic equation, the gas Navier-Stokes equation, and the heat conduction equation (3 equations in total); Pulverized coal combustion reaction kinetic equation: Input parameters: (Pre-factor) (activation energy), (Gas constant) (Combustion temperature) (Reaction order). Calculation process: Reaction rate constant Output result: The rate constant for pulverized coal combustion is 120 s⁻¹.

[0115] Second level (coupling level): Storage combustion-flow-heat transfer multiphysics coupling equations (2 equations in total).

[0116] The third level (system level): stores the global energy balance equation of the boiler and the NOx generation kinetic equation (two equations in total).

[0117] Step S2: The multiphysics simulation generates a virtual dataset, executed by the multiphysics simulator in the intrinsic digital twin building module. The multiphysics simulator calls the partial differential equations from the physical intrinsic equations library to perform a high-fidelity multiphysics simulation based on fundamental physical laws: first, a coarse-grained reduced-order simulation of the entire system is performed to calculate the state deviation sensitivity of the burner region. Input parameters: (NOx emission concentration, unit: mg / m³) (Flame center temperature, unit °C); Calculation process: at the reference temperature Take small temperature changes nearby The simulation yielded the corresponding NOx emission changes. Output result: ; The value is above the threshold of 0.3, therefore the burner area is identified as a critical subsystem.

[0118] Subsequently, high-fidelity CFD simulations were performed on the burner area, covering the entire operating range of boiler load from 50% to 100% and coal volatile matter from 15% to 35%, generating 2,000 sets of virtual datasets. Each dataset contains temperature field distribution, velocity field distribution, component concentration distribution, and NOx emission concentration data.

[0119] Step S3: Training the initial endogenous digital twin is performed by the AI ​​agent model training unit in the endogenous digital twin construction module. A graph neural network encoder-physical information decoder architecture is adopted: the virtual dataset is divided into a training set (1600 sets) and a validation set (400 sets) in an 8:2 ratio. The total loss function is a weighted sum of data loss and physical loss. Initial total loss on the validation set: After 30 training iterations, the lumped loss was verified to converge to... The validation set prediction error is less than 2%, and the initial endogenous digital twin is obtained and stored in eMMC non-volatile memory.

[0120] Parameter source explanation: Loss function weights The optimal value was determined based on 500 sets of numerical experiments on boiler combustion, balancing the accuracy of NOx emission prediction and physical consistency; the 30 training rounds were determined by the early stop method - training was stopped when the validation set loss no longer decreased for 5 consecutive rounds.

[0121] Step S4: Real-time system status acquisition is performed by a multimodal sensor network. After each 200ms control cycle: 24 K-type thermocouples synchronously acquire temperature distribution data at various cross-sections of the boiler furnace; 8 pressure sensors synchronously acquire pressure data at various points in the furnace; 2 flue gas analyzers synchronously acquire NOx emission concentration data at the chimney outlet; 4 oxygen sensors synchronously acquire oxygen data at the furnace outlet; 1 flame image sensor synchronously acquires flame image data in the burner area; all acquired data are transmitted to the industrial-grade control unit as input for subsequent status prediction.

[0122] Step S5: Predicting future state deviations is performed by the system evolution prediction unit in the self-evolving evolution engine module. The measured physical characteristic parameters collected in step S4 are input into the endogenous digital twin, and the predicted values ​​of furnace temperature distribution, NOx emission concentration, and combustion efficiency of the boiler within the next 200ms time window are output.

[0123] Step S6: Optimal control parameters are generated in advance and executed collaboratively by the numerical optimization solver and the collaborative game execution module in the self-evolutionary engine module. When the predicted value output from Step S5 indicates that the NOx emission concentration will exceed the preset threshold of 50 mg / m³ within the next 200 ms: The multi-objective function is to minimize the NOx emission concentration and maximize the combustion efficiency; the constraints are the adjustment ranges of the 12 fuel regulating valve openings, 8 supply air regulating valve openings, 4 induced draft air regulating valve openings, and 16 secondary air damper openings; the NSGA-II algorithm is used to perform multi-objective optimization, generating the optimal combination of control parameters covering the next 200 ms time window before the predicted NOx emission exceedance time.

[0124] The generated combination of control parameters is first subjected to three-level hierarchical verification, then verified by counterfactual reasoning, and finally the Pareto optimal solution is selected and executed through multi-agent collaborative screening.

[0125] Actual computational complexity calculation of the NSGA-II algorithm: Input parameters: (Iterative Algebra) (Population size) (Target quantities: Minimize NOx emissions and maximize combustion efficiency); Calculation process: This is a floating-point operation; the output result is approximately 12ms on an ARM Cortex-A9 processor, which is much less than 30% (60ms) of the 200ms control cycle limit.

[0126] Step S7: The feedback update of the physical intrinsic equation library is performed by the parameter update unit in the feedback self-evolution module. After step S6: the multimodal sensor network collects the measured temperature distribution, NOx emission concentration, and combustion efficiency data of the boiler after regulation; the parameter update unit uses the measured values ​​as observations and the predicted values ​​of the pre-execution endogenous digital twin as prior estimates; after the parameter update unit runs for more than 1000 control cycles, the effective sample size reaches a preset threshold, and it automatically switches to stochastic gradient descent update mode to improve update efficiency. The stochastic gradient descent algorithm is adopted, with a learning rate of... With a momentum parameter of 0.9, the key physical parameters such as the pulverized coal combustion reaction rate constant and thermal conductivity coefficient in the physical intrinsic equation library of the parent system are updated. The updated physical intrinsic equation library is stored in non-volatile memory for simulation and prediction in the next control cycle.

[0127] S4-S7 closed-loop continuous operation: After the system completes S7, it immediately returns to S4 to start the next control cycle until the system stops.

[0128] Hierarchical generation and evolution of sub-self-generated systems

[0129] While the S4-S7 operation loop continues to execute, the hierarchical generation management module synchronously executes the identification and incubation process of the sub-self-generated system: In the early stage of system operation: it is uniformly controlled by the parent system, and the measured NOx emission concentration is 35mg / m³, and the combustion efficiency is 92.5%.

[0130] Embryo identification stage: After running for 1440 hours, the ternary coupling strength of the component set responsible for NOx emission control (including 2 flue gas analyzers and 16 secondary wind baffles) reached 0.73, exceeding the threshold for 30 consecutive cycles, and was identified as an embryo.

[0131] Parameter source explanation: Ternary coupling strength weight Similar to Example 1, the determination is based on the core position of information in the self-generated system. Information relevance is the most critical indicator for judging whether a module possesses independent decision-making capabilities; therefore... .

[0132] Incubation Phase: The embryo was assigned a dedicated ARM core and control over 16 secondary wind baffles. After 240 hours of incubation, the embryo successfully completed 20 independent NOx emission control tasks consecutively. It stabilized above 0.91 and officially became an independent "NOx emission control sub-self-generating system".

[0133] Independent operation phase: The sub-self-generating system autonomously optimizes the secondary air distribution strategy, and the measured NOx emission concentration is reduced from 35mg / m³ to 26.25mg / m³ (a reduction of 25%), and the combustion efficiency is increased from 92.5% to 93.6% (an increase of 1.1%), which can save about 1,200 tons of standard coal per year.

[0134] Example of parameter update: After approximately 500 hours of operation, the pulverized coal combustion reaction rate constant k is updated from the initial value of 120 s⁻¹ to 115 s⁻¹, and the thermal conductivity of the furnace wall is updated from 45 W / (m·K) to 43 W / (m·K). The mathematical forms describing the conservation laws in the equation library remain stable.

[0135] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a self-generated system based on physical eigenvalue equations, the method comprising the steps of acquiring system operating status, generating control strategies, and issuing execution control, characterized in that, The self-generating system is characterized by its intrinsic stable order through a built-in hierarchical physical eigenvalue equation library, and achieves autonomous operation and continuous evolution through a closed loop of "perception-prediction-decision-execution-evolution". The method includes the following steps: S1. Based on the basic physical laws and factory calibration parameters of the target system, construct an initial physical eigenvalue equation library describing the multi-level system of the target system and store it in non-volatile memory; the physical eigenvalue equation library contains physical eigenvalue equations of at least three levels of the target system, and the physical eigenvalue equations of each level correspond to the stable relations that the system of different levels should satisfy under the constraints of physical laws; the equations of each level are combined through multi-physics field coupling relationships, the output of the lower-level equations is used as the input or boundary condition of the adjacent higher-level equations, and the higher-level equations impose downward causal constraints on the lower-level equations; the multi-level system includes at least a first level, a second level, and a third level. The system has three levels: the first level describes the material constitutive relations and single physical field evolution laws of the system components; the second level describes the multi-physics coupling and transmission relationships between components through interfaces; and the third level describes the global multi-physics cooperative evolution laws of the system. The initial physical intrinsic equation library is preset by the designer based on known physical laws before leaving the factory and is automatically loaded when the system is first powered on. After the system is running, the key physical parameters in the physical intrinsic equation library are continuously updated and corrected by the measured data of subsequent steps, gradually evolving from the initial preset state into a physical intrinsic equation library that characterizes the true endogenous stable order of the system; S2. A multiphysics simulator deployed in the industrial-grade control unit calls the partial differential equations in the physical intrinsic equation library to perform high-fidelity multiphysics simulation based on fundamental physical laws, generating a virtual system dataset covering key operating condition parameter combinations; S3. The AI ​​agent model training unit deployed in the industrial-grade control unit divides the virtual system dataset into a training set and a validation set according to a preset ratio. Using the training set as training samples, and the weighted sum of data loss and physical loss as the total loss function, when the total loss function value no longer decreases for K consecutive training rounds on the validation set, the current model is used as the initial endogenous digital twin after training and stored in non-volatile memory; S4. After each state acquisition cycle, the multimodal sensor network acquires the measured physical characteristic parameters of the current operating state of the target system and transmits them to the industrial-grade control unit; S5. The system evolution prediction unit deployed in the industrial-grade control unit inputs the measured physical characteristic parameters acquired in step S4 into the endogenous digital twin generated in step S3, and outputs the predicted state deviation value of the target system within a preset future time window; S6.When the predicted value output in step S5 indicates that the state deviation within the future preset time window will exceed a preset threshold, the numerical optimization solver deployed in the industrial-grade control unit, with minimizing the overall system state deviation as the objective function and the adjustable range of parameters of each execution unit as the constraint, performs optimization before the predicted state deviation occurs, autonomously generating the optimal combination of control parameters covering the future preset time window. After hierarchical verification, this is distributed to each execution unit for execution. S7. After step S6, the multimodal sensor network collects the measured state change parameters of the target system and transmits them to the industrial-grade control unit. The parameter update unit deployed in the industrial-grade control unit updates the key physical parameters in the parent system's physical intrinsic equation library using the measured values ​​as observations and the predicted values ​​from the endogenous digital twin before execution as prior estimates.

2. A self-generated system based on physical eigenvalue equations, characterized in that, The system includes an industrial-grade control unit, a multimodal sensor network, multiple execution units, non-volatile memory, and a hierarchical generation and management module. The industrial-grade control unit is equipped with an endogenous digital twin construction module, a self-evolutionary engine module, a cooperative game execution module, and a feedback self-evolutionary module. The endogenous digital twin construction module is used to construct an endogenous digital twin based on a physical intrinsic equation library. The self-evolutionary engine module is used to predict system state deviations and generate control parameters. The cooperative game execution module is used to select and distribute the optimal combination of control parameters. The feedback self-evolutionary module is used to update key physical parameters in the physical intrinsic equation library. The hierarchical generation management module is used to identify and incubate sub-self-generating systems.

3. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The coupling relationship between the equations at each level satisfies the following mathematical characterization: Let the physical eigenvalue equation of the i-th level be... ,in Let i be the state variable of the i-th level. Let be the physical parameters of the i-th level; the coupling relationship between levels is achieved in the following two ways: (a) upward causality—the solution of the i-th level equation. (a) As boundary conditions or source terms of the (i+1)th level equation; (b) Downward causality—the solution of the (i+1)th level equation The parameters acting as constraints on the i-th level equation .

4. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The multiphysics simulator employs a hierarchical simulation strategy: first, a coarse-grained, order-reduced simulation is performed on the entire system to identify subsystems whose state deviation sensitivity exceeds a preset threshold as critical subsystems; then, high-fidelity multiphysics simulation is performed only on these critical subsystems; the state deviation sensitivity is expressed by the formula... Calculation, where This represents the global state deviation of the system. Let j be the state variable of the j-th subsystem, when When the sensitivity exceeds a preset threshold, the subsystem is identified as a critical subsystem. When the system state deviation exceeds the preset trigger threshold, the simulation fidelity is automatically increased and a high-precision simulation retraining is triggered. The multiphysics simulator uses an intrinsic orthogonal decomposition (POD) model to reduce the simulation degrees of freedom from... compressed to The simulation time is on the order of magnitude, with a single simulation taking less than 100ms.

5. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The AI ​​agent model training unit includes a graph neural network encoder and a physical information decoder. The graph neural network encoder uses a 3-layer graph convolutional network to encode the physical topology of the target system into a graph structure, with each physical component of the target system as a graph node and the physical coupling relationship between the components as a graph edge, and outputs a 128-dimensional graph node feature vector. The physical information decoder uses a 2-layer fully connected network to decode the graph node feature vector into a state prediction value of the physical space.

6. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The graph node feature vectors are further mapped to a linear evolution space through a nonlinear dynamic linearization operator. After performing linear evolution inference in the linear evolution space, the predicted state deviation value for the future time window is output. The nonlinear dynamic linearization operator is the Koopman operator, DeepONet, or Physical Information Long Short-Term Memory (PI-LSTM) network.

7. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, It also includes a sub-self-generating system embryo recognition and hatching mechanism: real-time calculation of the ternary coupling strength index of each functional component set within the target system. ; When the ternary coupling strength index of a certain component set The pre-set differentiation threshold is exceeded for M consecutive acquisition cycles. At that time, it is identified as a potential sub-self-generating system embryo, and the incubation process is initiated; the ternary coupling strength index Information relevance of this component set Energy coupling degree and the degree of material cohesion Weighted sum: in ,and The degree of information correlation This is the ratio of the total data flow within the component set to the total data flow of the component set, with a value range of [value missing]. The calculation formula is: in This refers to the total amount of data exchanged between the sensors and actuators within the component assembly per unit time (in bits per second). The total amount of data exchanged between the component set and other parts of the parent system per unit time (in bits per second); this total amount of data is obtained by monitoring the data traffic statistics of each communication interface; the energy coupling degree... This is the ratio of the energy exchange within the component assembly to the total energy exchange of the component assembly, and its value ranges from [value missing]. The calculation formula is: in This refers to the amount of energy exchanged within the component assembly. The energy obtained by the component assembly from the parent system; the degree of matter condensation. This represents the spatial aggregation degree of the physical resources occupied by this component set, with a value range of [value range missing]. The calculation formula is: in This is the variance of the normalized spatial coordinates of all physical units in the component set; the coordinates are normalized to the nearest normal value, with the system's overall geometric center as the origin and the system's maximum outer envelope size as the reference. The incubation process refers to: treating the germ as an independent decision-making entity, incorporating it into a hierarchical collaborative game framework, engaging in asymmetric resource allocation game with the parent system, and gradually allocating dedicated computing resources, energy quotas, and physical actuators to it until the ternary coupling strength index of the germ reaches its limit. The number of consecutive data collection cycles stably exceeds the preset independent survival threshold. It becomes an independent sub-self-generating system capable of maintaining its local operational identity.

8. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The payoff function for the asymmetric resource allocation game is configured as follows: in , The resource security constraint satisfaction of the parent system includes the maintenance threshold of the overall operational consistency of the parent system and the minimum guarantee coefficient of global resource allocation; the local operational consistency refers to the ability of the self-generated sub-system to autonomously define and maintain its core survival objective, and the core survival objective is compatible with but not completely identical to the global objective of the parent system; when the ternary coupling strength index The data collection rate consistently exceeds the preset autonomous operation threshold for N consecutive data collection cycles. This indicates that the self-generated subsystem has achieved complete local operational uniformity.

9. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, It also includes a sub-self-generated system disintegration mechanism: when the ternary coupling strength index of an independent sub-self-generated system... For P consecutive acquisition cycles, the data is below the preset disassembly threshold. When the parent system initiates the dismantling procedure, it reclaims all resources occupied by the self-generated subsystem and migrates its knowledge and experience to the parent system's knowledge pool; the value of P ranges from 50 to 200.

10. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, It also includes a hierarchical verification mechanism: hierarchical verification is performed on the generated control parameter combination, including: Level 1 verification, which verifies whether the control parameters meet the physical constraints and functional safety constraints of the corresponding execution unit; Level 2 verification, which verifies whether the control parameters meet the multi-physics coupling constraints of the subsystem where the execution unit is located; Level 3 verification, which verifies whether the control parameters meet the global coordination and consistency constraints of the system; execution is only submitted after the control parameter combination passes all three levels of verification; the physical constraints include the maximum / minimum output amplitude, maximum rate of change, and dead zone range of the execution unit; the functional safety constraints include the maximum allowable operating temperature, maximum allowable continuous operating time, and emergency shutdown threshold of the execution unit.

11. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, It also includes a counterfactual reasoning mechanism: using the endogenous digital twin as the baseline state, positive and negative perturbations are applied to the control parameters; the amplitude of the perturbation is determined according to the product of the control resolution and the safety threshold of the corresponding execution unit: perturbation amplitude = min(control resolution × 10, safety threshold × 20%). The control resolution is the minimum parameter change that the execution unit can stably execute, and the safety threshold is the maximum parameter change range allowed by the execution unit. Two counterfactual branches are generated respectively, and forward deduction is performed along each branch. The combination of control parameters corresponding to the branch with the best performance is compared with the result of the optimization solution, and the best one is output.

12. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, or the self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, When a system of partial differential equations contains multiple equations, the physical loss function is the weighted sum of the squared residuals of each equation: Where K is the total number of equations and M is the total number of sampling points. Let be the number of sampling points corresponding to the k-th equation, satisfying ; Let be the weight coefficients of the k-th equation, satisfying ; The initial values ​​of the weight coefficients are determined based on the characteristic time scale and characteristic energy scale of the physical process described by each equation: the shorter the characteristic time scale and the greater the characteristic energy of the physical process, the greater the initial weight of the corresponding equation. During training, the weight coefficients can be dynamically adjusted according to the convergence rate of the residuals of each equation—equations that converge more slowly are given greater weights to accelerate overall convergence.

13. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, characterized in that, In step S3, the virtual system dataset is divided into a training set and a validation set in an 8:2 ratio, and the initial endogenous digital twin is obtained after 30 training rounds.

14. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, characterized in that, The system evolution prediction unit described in step S5 also calculates the cognitive free energy of the endogenous digital twin regarding the current state of the target system. Used for quantitative prediction distribution Compared with the posterior distribution based on measured data Information theory bias between them; when When a significant bias is found in the perception of the endogenous digital twin, the step size for parameter updates in step S7 is increased; when When this happens, the local model is reinitialized.

15. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, characterized in that, The update strategy of the parameter update unit in step S7 is as follows: an incremental update is performed immediately after each state acquisition cycle is completed; or a batch update is performed after accumulating data for L acquisition cycles; the incremental update only adjusts the output layer parameters of the endogenous digital twin, and the batch update adjusts all trainable parameters of the endogenous digital twin.

16. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, characterized in that, The key operating condition parameter combinations mentioned in step S2 include: the operating points of each execution unit at the upper and lower boundaries of its allowable adjustment range, the combined operating points of multiple execution units simultaneously in extreme states, and the abnormal operating points in the system's historical operation where the state deviation exceeds a preset threshold.

17. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, characterized in that, When the target system contains multiple execution units, step S6 further includes: the system collects candidate control parameters generated by all execution units, performs non-dominated sorting based on the game payoff function, and selects the control parameter combination corresponding to the Pareto front approximation solution and distributes it to each execution unit for execution; the search time of the Pareto front approximation solution is limited by the system's control cycle.

18. The self-generating system construction method according to claim 17, characterized in that, The search for approximate solutions to the Pareto front employs a simplified online version of the Non-Dominated Sorting Genetic Algorithm (NSGA-II) based on crowding ranking. The specific process is as follows: Population initialization: Based on the actual control parameters of each execution unit, candidate populations are randomly generated within ±10%, with a population size of 50; Non-dominated sorting: The value of each candidate solution on each game payoff function is calculated, and fast non-dominated sorting is performed; Crowding calculation: Within the same non-dominated level, solutions are sorted in descending order of crowding for each objective, and the top 20 solutions are selected for the next generation; Search termination: The search is terminated when the search time reaches a preset upper limit (usually 30% of the system control cycle) or when there is no significant improvement in the Pareto front for three consecutive generations; Optimal solution selection: From the final Pareto front, the solution with the closest Euclidean distance to the ideal point (where all objectives are optimal) is selected as the parameter to be distributed. The time complexity of this algorithm is O(n log n). Where G is the iteration number, typically 10-20; P is the population size, typically 50; and M is the target number, typically 2-3.

19. The self-generating system construction method according to claim 17, characterized in that, When the target system also contains independent sub-self-generated systems, the non-dominated sorting is executed hierarchically: First, a first-level collaborative game is executed within each execution unit and within each sub-self-generated system to generate optimal control parameter combinations at the component and subsystem levels; then, a second-level collaborative game is executed between each execution unit and between each sub-self-generated system to generate optimal control parameter combinations at the system level; the first-level collaborative game prioritizes the satisfaction of component or sub-self-generated system ontology constraints, while the second-level collaborative game prioritizes the achievement of the system's global collaborative goal. Both are jointly solved through a hierarchical game payoff function, wherein the weight of ontology constraint satisfaction in the first-level game payoff function is higher than the weight of the achievement of the overall system's collaborative goal, and not less than 0.

6.

20. The method for constructing a self-generated system based on physical eigenvalue equations according to claim 1, characterized in that, Step S7 is followed by: S7a. The knowledge transfer unit deployed on the cloud server clusters the model parameters of the endogenous digital twins of all target systems and sub-self-generated systems of the same type and extracts common knowledge of the group; when the new system goes online, the common knowledge of the group is used as the initial model parameters of the AI ​​agent model training unit in step S3.

21. The self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The industrial-grade control unit adopts an FPGA+MCU heterogeneous architecture. The FPGA is responsible for high-speed signal processing and real-time control, while the MCU is responsible for complex algorithm calculations and system management. The FPGA and MCU are connected via an AXI bus.

22. The self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The multimodal sensing network includes at least two of the following: a thermal field sensing unit, an electric field sensing unit, a mechanical sensing unit, and a timing sensing unit. Each sensing unit synchronously acquires measured physical characteristic parameters of the target system at a preset sampling frequency. The thermal field sensing unit is a thermocouple or a thermistor, the electric field sensing unit is a voltage sensor or a current sensor, the mechanical sensing unit is an accelerometer or a strain gauge, and the timing sensing unit is a clock deviation detection unit. The sampling frequency of each sensing unit is from 1 kHz to 100 kHz.

23. The self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The parameter update unit adopts the Bayesian linear regression algorithm and uses a Gaussian prior distribution. When the effective sample size is less than the preset threshold, the parameters are updated by using complete Bayesian posterior inference. When the effective sample size is greater than or equal to the preset threshold, the update mode is switched to stochastic gradient descent.

24. The self-generated system based on physical eigenvalue equations according to claim 2, characterized in that, The feedback self-evolution module also includes a knowledge transfer unit, which is deployed in a cloud server. The knowledge transfer unit clusters the model parameters of the endogenous digital twins of all target systems and sub-self-evolutionary systems of the same type and extracts common knowledge of the group. When a new system is launched, the common knowledge of the group is used as the initial model parameters of the AI ​​agent model training unit.

25. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method of claim 1 or any one of claims 14 to 20.

26. An industrial control device, characterized in that, Including the self-generated system based on physical eigenvalue equations as described in claim 2, or any one of claims 21 to 24.

Images