Multi-working-condition parameter optimization method and device for aero-engine system

Abstract

The invention relates to a multi-working-condition parameter optimization method and device for an aero-engine system. The method comprises the steps that a target optimization task is obtained, and a plurality of components and a plurality of working states of a target aero-engine system are determined; constructing an optimization model with a neural network structure based on a plurality of components and a plurality of working states of the target aero-engine system; in each model iteration process, training the initial optimization model based on a preset first optimization strategy and a preset second optimization strategy until the trained optimization model meets a preset condition, and obtaining a trained optimization model; and inputting the target adjustment direction into the optimization model to obtain an optimization result. According to the method, the problem of unified optimization of heterogeneous components is solved by adopting an alternate optimization strategy, the limitation of a traditional single modeling mode is broken through, different types of model components are coordinated and processed under a unified calculation framework, and a brand-new technical path is provided for multi-working-condition optimization of a complex engineering system.

Description

Technical Field

[0001] This disclosure relates to the field of engine optimization technology, and in particular to a method and apparatus for multi-condition parameter optimization of aero-engine systems. Background Technology

[0002] In aero-engine systems, how to use artificial intelligence (AI) architecture to handle complex input-output relationships at the component level, prediction, and system-level coupling optimization problems? Specifically, this involves using AI to learn nonlinear input-output mapping relationships in key components that cannot be accurately described by traditional formulas, establishing parameter optimization algorithms that can handle the complex coupling effects of preceding components on subsequent components, and achieving global parameter optimization under various operating conditions. This ensures that an optimal parameter combination that can adapt to multiple operating conditions is found, thereby solving the technical bottlenecks faced by traditional engine modeling methods in handling nonlinear coupling relationships between components and multi-condition parameter optimization. Summary of the Invention

[0003] In view of this, this disclosure proposes a method and apparatus for multi-condition parameter optimization of aero-engine systems.

[0004] According to one aspect of this disclosure, a method for multi-condition parameter optimization of an aero-engine system is provided. The method includes: acquiring a target optimization task for a target aero-engine system; determining multiple components and multiple operating states of the target aero-engine system; constructing an optimization model with a neural network structure based on the multiple components and multiple operating states of the target aero-engine system, the optimization model including a condition neural network corresponding to each operating state, each condition neural network including multiple component layers, each component layer corresponding to different components; training an initial optimization model based on a preset first optimization strategy and a second optimization strategy during each model iteration until the trained optimization model meets preset conditions, thereby obtaining a trained optimization model; determining a target adjustment direction for the target aero-engine system according to the target optimization task, and inputting the target adjustment direction into the optimization model to obtain an optimization result, the optimization result including the parameter adjustment amount of the components required to meet the target adjustment direction.

[0005] By reconstructing the engine system into a neural network and employing an alternating optimization strategy to solve the unified optimization problem of heterogeneous components, this approach overcomes the limitations of traditional single-modeling paradigms. It coordinates the processing of different types of model components within a unified computational framework, providing a novel technical path for multi-condition optimization of complex engineering systems. This multi-condition parameter optimization method addresses the challenges of parameter coupling and nonlinear optimization under multiple operating conditions in aero-engines through hierarchical modeling and hybrid optimization strategies for differentiable / non-differentiable components. Considering the differences in characteristics between differentiable and non-differentiable components in the engine system, the component layers of the operating condition neural network adopt differentiated designs. For example, differentiable components (such as compressors and turbines) use traditional neural network structures, constructing continuous mapping relationships through linear transformations and differentiable activation functions (such as ReLU), maintaining overall differentiability. Non-differentiable components (such as ignition systems and bleed valves) can be embedded with a hybrid structure of decision trees and neural networks (such as differentiable adaptive neural trees ANT), transforming discrete threshold logic into continuous probability outputs through a "soft decision path," thus solving the problem of insufficient modeling capability for discrete events by traditional neural networks. This hierarchical approach enables the optimization model to accurately capture the step characteristics of non-differentiable components while maintaining global differentiability, reducing parameter prediction errors under multiple operating conditions compared to a single neural network model. Furthermore, a closed-loop training mechanism is formed by the first optimization strategy (data-driven) and the second optimization strategy (physical constraints). This dual-strategy collaborative training helps improve training efficiency and convergence stability. The data-driven strategy uses historical operating data (such as fuel flow and exhaust temperature) to train the neural network, implementing gradient descent through the Adam optimizer. The physical constraint strategy introduces mechanistic model regularization terms (such as the GRI-Mech combustion mechanism) into the non-differentiable component layer to avoid data overfitting.

[0006] In one possible implementation, the method further includes: classifying multiple components of the target aero-engine system into differentiable components and non-differentiable components according to a preset classification strategy, wherein the first optimization strategy corresponds to the differentiable components and the second optimization strategy corresponds to the non-differentiable components.

[0007] In this way, by distinguishing between differentiable and non-differentiable components through a pre-defined classification strategy and matching them with a dedicated optimization strategy, a refined optimization framework of "feature adaptation" is constructed. This optimization paradigm of "feature classification-strategy adaptation" solves the inherent contradiction of the traditional method of "using a continuous model to describe discrete behavior".

[0008] In one possible implementation, the classification strategy includes at least one of the following: classifying components whose state parameters change continuously with the input as differentiable components and components whose state parameters change abruptly as non-differentiable components; classifying components whose behavior is affected by a continuous physical field as differentiable components and components whose behavior is affected by discrete events as non-differentiable components; classifying components whose geometric features exhibit continuous curvature changes as differentiable components and components whose geometric structures have discontinuous boundaries as non-differentiable components; classifying components whose output response exhibits a linear or smooth nonlinear relationship as differentiable components and components whose output response exhibits a threshold effect as non-differentiable components.

[0009] In this way, by accurately classifying differentiable and non-differentiable components using the above four classification strategies, a scientific "digital anatomy" standard is provided for the hybrid modeling of engine systems. This not only improves modeling accuracy and significantly reduces component characteristic fitting errors, but also optimizes computational efficiency, increasing the iteration speed of the optimization model by several times. Furthermore, the classification strategy enables the control algorithm to achieve "characteristic adaptation," which can enhance control robustness and improve stability under extreme operating conditions.

[0010] In one possible implementation, training an initial optimization model according to the first optimization strategy includes: fixing the network parameters of the component layer of all non-differentiable components, and performing gradient optimization on the differentiable components using a multi-working-state backpropagation algorithm.

[0011] In this way, by fixing the parameters of the non-differentiable component layer and optimizing the differentiable components through a multi-working-state backpropagation algorithm, a dynamic balance between physical constraints and data-driven approaches is achieved in hybrid modeling. This "fixed-optimized" layered training strategy perfectly balances model flexibility and physical reliability. During the differentiable component optimization phase, fixing the network parameters of all non-differentiable components and using a multi-working-state backpropagation algorithm to efficiently optimize the gradients of the differentiable physical model and other differentiable components fully leverages the advantages of gradient information in parameter space navigation.

[0012] In one possible implementation, training an initial optimization model according to the second optimization strategy includes: fixing the network parameters of the component layers of all differentiable components, and performing gradient-free optimization on the non-differentiable components using a Bayesian optimization algorithm based on Gaussian processes.

[0013] Thus, by fixing the network parameters of differentiable components and employing the Gaussian process-based Bayesian optimization algorithm, this method achieves gradient-free, precise optimization of non-differentiable components, significantly improving the optimization efficiency and robustness of complex engine systems. Furthermore, by fixing the parameters of the differentiable components, Bayesian optimization concentrates computational resources on the high-influence parameters of the non-differentiable components, shortening the optimization cycle. This hierarchical optimization strategy redefines the intelligent optimization paradigm for complex systems. Differentiable components provide stable physical constraint boundaries, while Bayesian optimization efficiently explores the black-box characteristics of non-differentiable components within these boundaries. In the non-differentiable component optimization stage, fixing the network parameters of the differentiable components and employing the Gaussian process-based Bayesian optimization method enables intelligent gradient-free optimization of the non-differentiable components, effectively balancing exploration and utilization by constructing a surrogate model.

[0014] In one possible implementation, the method further includes: calculating a loss function value based on a dynamic balance coefficient during each model iteration; and obtaining a trained optimized model when the loss function value is less than a preset threshold.

[0015] In this way, an adaptive parameter adjustment mechanism is further introduced. The optimization process is automated by using dynamic balancing coefficients. The adaptive parameter adjustment mechanism can automatically adjust the weight relationship between the output of key components and the overall component output according to the optimization process. In the early stage of optimization, more attention is paid to the rapid improvement of overall performance, and in the later stage of optimization, the fine-tuning of key components is gradually strengthened to ensure that the entire optimization process has both good convergence stability and can achieve a high-quality optimization solution.

[0016] According to another aspect of this disclosure, a multi-condition parameter optimization device for an aero-engine system is provided. The device includes: a task acquisition module, used to acquire a target optimization task for a target aero-engine system and determine multiple components and multiple operating states of the target aero-engine system; a model construction module, used to construct an optimization model with a neural network structure based on the multiple components and multiple operating states of the target aero-engine system, the optimization model including a condition neural network corresponding to each operating state, each condition neural network including multiple component layers, each component layer corresponding to different components; a model training module, used to train an initial optimization model based on a preset first optimization strategy and a second optimization strategy during each model iteration until the trained optimization model meets preset conditions, thereby obtaining a trained optimization model; and a parameter optimization module, used to determine a target adjustment direction for the target aero-engine system according to the target optimization task, and input the target adjustment direction into the optimization model to obtain an optimization result, the optimization result including the parameter adjustment amount of the components required to meet the target adjustment direction.

[0017] In this way, by reconstructing the engine system into a neural network and employing an alternating optimization strategy to solve the unified optimization problem of heterogeneous components, the limitations of the traditional single modeling paradigm are broken through. Different types of model components are coordinated and processed within a unified computational framework, providing a novel technical path for multi-condition optimization of complex engineering systems. Through hierarchical modeling and hybrid optimization strategies for differentiable / non-differentiable components, the challenges of parameter coupling and nonlinear optimization under multiple operating conditions in aero-engines are solved. Considering the differences in characteristics between differentiable and non-differentiable components in the engine system, the component layers of the operating condition neural network adopt differentiated designs. For example, differentiable components (such as compressors and turbines) use traditional neural network structures, constructing continuous mapping relationships through linear transformations and differentiable activation functions (such as ReLU), maintaining overall differentiability. For non-differentiable components (such as ignition systems and bleed valves), a hybrid structure of decision trees and neural networks (such as the differentiable adaptive neural tree ANT) can be embedded. Through a "soft decision path," discrete threshold logic is transformed into continuous probability outputs, solving the problem of insufficient modeling capability of traditional neural networks for discrete events. This hierarchical approach enables the optimization model to accurately capture the step characteristics of non-differentiable components while maintaining global differentiability, reducing parameter prediction errors under multiple operating conditions compared to a single neural network model. Furthermore, a closed-loop training mechanism is formed by the first optimization strategy (data-driven) and the second optimization strategy (physical constraints). This dual-strategy collaborative training helps improve training efficiency and convergence stability. The data-driven strategy uses historical operating data (such as fuel flow and exhaust temperature) to train the neural network, implementing gradient descent through the Adam optimizer. The physical constraint strategy introduces mechanistic model regularization terms (such as the GRI-Mech combustion mechanism) into the non-differentiable component layer to avoid data overfitting.

[0018] In one possible implementation, the device further includes a classification module for: classifying multiple components of the target aero-engine system into differentiable components and non-differentiable components according to a preset classification strategy, wherein the first optimization strategy corresponds to the differentiable components and the second optimization strategy corresponds to the non-differentiable components.

[0019] In this way, by distinguishing between differentiable and non-differentiable components through a pre-defined classification strategy and matching them with a dedicated optimization strategy, a refined optimization framework of "feature adaptation" is constructed. This optimization paradigm of "feature classification-strategy adaptation" solves the inherent contradiction of the traditional method of "using a continuous model to describe discrete behavior".

[0020] In one possible implementation, the classification strategy includes at least one of the following: classifying components whose state parameters change continuously with the input as differentiable components and components whose state parameters change abruptly as non-differentiable components; classifying components whose behavior is affected by a continuous physical field as differentiable components and components whose behavior is affected by discrete events as non-differentiable components; classifying components whose geometric features exhibit continuous curvature changes as differentiable components and components whose geometric structures have discontinuous boundaries as non-differentiable components; classifying components whose output response exhibits a linear or smooth nonlinear relationship as differentiable components and components whose output response exhibits a threshold effect as non-differentiable components.

[0021] In this way, by accurately classifying differentiable and non-differentiable components using the above four classification strategies, a scientific "digital anatomy" standard is provided for the hybrid modeling of engine systems. This not only improves modeling accuracy and significantly reduces component characteristic fitting errors, but also optimizes computational efficiency, increasing the iteration speed of the optimization model by several times. Furthermore, the classification strategy enables the control algorithm to achieve "characteristic adaptation," which can enhance control robustness and improve stability under extreme operating conditions.

[0022] In one possible implementation, training an initial optimization model according to the first optimization strategy includes: fixing the network parameters of the component layer of all non-differentiable components, and performing gradient optimization on the differentiable components using a multi-working-state backpropagation algorithm.

[0023] In this way, by fixing the parameters of the non-differentiable component layer and optimizing the differentiable components through a multi-working-state backpropagation algorithm, a dynamic balance between physical constraints and data-driven approaches is achieved in hybrid modeling. This "fixed-optimized" layered training strategy perfectly balances model flexibility and physical reliability. During the differentiable component optimization phase, fixing the network parameters of all non-differentiable components and using a multi-working-state backpropagation algorithm to efficiently optimize the gradients of the differentiable physical model and other differentiable components fully leverages the advantages of gradient information in parameter space navigation.

[0024] In one possible implementation, training an initial optimization model according to the second optimization strategy includes: fixing the network parameters of the component layers of all differentiable components, and performing gradient-free optimization on the non-differentiable components using a Bayesian optimization algorithm based on Gaussian processes.

[0025] In this way, by fixing the network parameters of differentiable components and employing the Gaussian process-based Bayesian optimization algorithm, gradient-free precise optimization of non-differentiable components is achieved, significantly improving the optimization efficiency and robustness of complex engine systems. Furthermore, by fixing the parameters of the differentiable components, Bayesian optimization concentrates computational resources on the high-influence parameters of the non-differentiable components, shortening the optimization cycle. This hierarchical optimization strategy redefines the intelligent optimization paradigm for complex systems. Differentiable components provide stable physical constraint boundaries, while Bayesian optimization efficiently explores the black-box characteristics of non-differentiable components within these boundaries. In the non-differentiable component optimization stage, fixing the network parameters of the differentiable components and employing the Gaussian process-based Bayesian optimization method enables intelligent gradient-free optimization of the non-differentiable components, effectively balancing the relationship between exploration and utilization by constructing a surrogate model.

[0026] In one possible implementation, the device further includes an iterative judgment module, used to: calculate the loss function value based on the dynamic balance coefficient during each model iteration; and obtain the trained optimized model if the loss function value is less than a preset threshold.

[0027] In this way, an adaptive parameter adjustment mechanism is further introduced. The optimization process is automated by using dynamic balancing coefficients. The adaptive parameter adjustment mechanism can automatically adjust the weight relationship between the output of key components and the overall component output according to the optimization process. In the early stage of optimization, more attention is paid to the rapid improvement of overall performance, and in the later stage of optimization, the fine-tuning of key components is gradually strengthened to ensure that the entire optimization process has both good convergence stability and can achieve a high-quality optimization solution.

[0028] According to another aspect of this disclosure, a multi-condition parameter optimization apparatus for an aero-engine system is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.

[0029] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0030] According to another aspect of this disclosure, a computer program product is provided, including a computer program or a non-volatile computer-readable storage medium carrying the computer program, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0031] This disclosure identifies multiple components and operating states of a target aero-engine system by acquiring a target optimization task for that system. Based on these components and operating states, an optimization model with a neural network structure is constructed. This model includes a condition neural network corresponding to each operating state, with each layer containing multiple component layers, each corresponding to a different component. During each model iteration, the initial optimization model is trained using a preset first and second optimization strategy until the trained model meets preset conditions, resulting in a trained optimization model. Finally, a target adjustment direction for the target aero-engine system is determined according to the target optimization task. The target adjustment direction is input into the optimization model to obtain the optimization result, which includes the parameter adjustment amount of the components required to meet the target adjustment direction. This allows the engine system to be reconstructed into a neural network form, employing an alternating optimization strategy to solve the unified optimization problem of heterogeneous components. This breaks through the limitations of traditional single modeling paradigms, coordinating the processing of different types of model components within a unified computational framework, providing a new technical path for multi-condition optimization of complex engineering systems. Through hierarchical modeling and hybrid optimization strategies for differentiable / non-differentiable components, the problem of parameter coupling and nonlinear optimization under multiple operating conditions of aero-engines is solved. Considering the differences in characteristics between differentiable and non-differentiable components in the engine system, the component layer of the operating condition neural network adopts a differentiated design. This hierarchical processing enables the optimization model to accurately capture the step characteristics of non-differentiable components while maintaining global differentiability, reducing parameter prediction errors under multiple operating conditions compared to a single neural network model. Furthermore, the closed-loop training mechanism formed by the first and second optimization strategies helps improve training efficiency and convergence stability.

[0032] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0033] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0034] Figure 1 A flowchart illustrating the parameter optimization method provided in an embodiment of this disclosure is shown.

[0035] Figure 2 A schematic diagram of a neural network provided in an embodiment of this disclosure is shown.

[0036] Figure 3 A block diagram of the parameter optimization apparatus provided in an embodiment of this disclosure is shown. Detailed Implementation

[0037] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0038] As used herein, the terms “comprising,” “including,” “having,” or variations thereof are open-ended and include one or more of the stated features, integrals, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integrals, elements, steps, components, functions, or groups thereof.

[0039] When an element is referred to as “connected,” “coupled,” “responding,” or a variation thereof relative to another element, it may be directly connected, coupled, or responding to another element, or there may be an intermediate element present.

[0040] Although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0041] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0042] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0043] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0044] To address the aforementioned technical problems, this disclosure provides a method for optimizing multiple operating state parameters of an aero-engine system. The method acquires a target optimization task for the target aero-engine system and determines multiple components and operating states of the target aero-engine system. Based on the multiple components and operating states of the target aero-engine system, an optimization model with a neural network structure is constructed. The optimization model includes a condition neural network corresponding to each operating state, and each condition neural network includes multiple component layers, with each component layer corresponding to a different component. During each model iteration, the initial optimization model is trained based on a preset first optimization strategy and a second optimization strategy until the trained optimization model meets preset conditions, resulting in a trained optimization model. The method then determines the target optimization task for the aero-engine system. The target adjustment direction of the target aero-engine system is input into the optimization model to obtain the optimization result. The optimization result includes the parameter adjustment amount of the components required to meet the target adjustment direction. By reconstructing the engine system into a neural network and using an alternating optimization strategy to solve the unified optimization problem of heterogeneous components, the limitations of the traditional single modeling paradigm are broken. Different types of model components are coordinated and processed under a unified computing framework, providing a new technical path for multi-condition optimization of complex engineering systems. Furthermore, through hierarchical modeling and hybrid optimization strategies for differentiable / non-differentiable components, the problems of parameter coupling and nonlinear optimization under multiple operating conditions of aero-engines are solved. Considering the differences in characteristics between differentiable and non-differentiable components in the engine system, the component layer of the operating condition neural network adopts a differentiated design. This hierarchical processing enables the optimization model to accurately capture the step characteristics of non-differentiable components while maintaining global differentiability, reducing parameter prediction errors under multiple operating conditions compared to a single neural network model. Moreover, a closed-loop training mechanism is formed by the first optimization strategy and the second optimization strategy. This dual-strategy collaborative training helps to improve training efficiency and convergence stability.

[0045] Now combined Figure 1 and Figure 2 This disclosure provides an illustrative description of a method for optimizing multiple operating state parameters of an aero-engine system, based on embodiments of the present disclosure. This method can be executed on a server or terminal. Figure 1 As shown, this method may include the following steps S101 to S104.

[0046] Step S101: Obtain the target optimization task for the target aero-engine system and determine multiple components and multiple operating states of the target aero-engine system.

[0047] The target aero-engine system is an engine system of current interest. The engine system may include multiple components such as the crankshaft and connecting rod mechanism, valve train, fuel supply mechanism, cooling mechanism, lubrication mechanism, and starting mechanism.

[0048] Step S102: Construct an optimization model with a neural network structure based on multiple components and multiple operating states of the target aero-engine system. The optimization model includes a working condition neural network corresponding to each operating state. Each working condition neural network includes multiple component layers, and each component layer corresponds to a different component.

[0049] The specific neural network used in the optimization model can be flexibly selected according to actual needs, and this disclosure does not limit this aspect.

[0050] Specifically, the engine system can be reconstructed as a multi-layer neural network structure, where each engine component corresponds to a network layer, such as... Figure 2 As shown, the output of the previous component layer and the parameters of the current component layer together determine the output of the current component layer. Simultaneously, each working condition or state is constructed as an independent neural network, and expressed through a unified mathematical formula. Establish a complete optimization model, in which, This indicates the initial position and status information of each component, with components numbered 1, 2, ... , Indicates component layer parameters, , Indicates the first The component layer parameters of each component. Under this architecture, the functions of each layer... It can be flexibly configured as a mechanistic model or an AI model according to actual needs, forming a hybrid modeling architecture. For the non-differentiable component layer, a Gaussian model is used as a surrogate model, and an input-output mapping is constructed based on historical data. The output formula of each component layer can be expressed as follows: .in, Indicates the first The output of each component layer Indicates the first Functions of each component layer Indicates the first The output of each component layer Indicates the first The network parameters of each component layer are used. This design not only fully preserves the physical connections and energy transfer characteristics between engine components, but also lays a solid mathematical foundation for the unified expression and collaborative optimization of heterogeneous models. Each operating condition neural network achieves accurate modeling of physical laws under specific operating conditions through parameter coupling at the component layer. The operating condition neural network shares feature extraction layers through transfer learning, reducing the need for new operating condition data.

[0051] Step S103: In each model iteration, the initial optimization model is trained based on the preset first optimization strategy and the second optimization strategy until the trained optimization model meets the preset conditions, and the trained optimization model is obtained.

[0052] The preset conditions can be that the loss function value meets a threshold, or that the number of model iterations meets a threshold.

[0053] Step S104: Determine the target adjustment direction for the target aero-engine system according to the target optimization task, and input the target adjustment direction into the optimization model to obtain the optimization result. The optimization result includes the parameter adjustment amount of the components required to meet the target adjustment direction.

[0054] The target adjustment direction indicates a general direction for engine adjustment. For example, maximum thrust or minimum fuel consumption.

[0055] Thus, this multi-condition parameter optimization method solves the challenges of parameter coupling and nonlinear optimization under multiple operating conditions in aero-engines by employing a hierarchical modeling and hybrid optimization strategy for differentiable / non-differentiable components. Addressing the differences in characteristics between differentiable and non-differentiable components in the engine system, the component layers of the operating condition neural network adopt differentiated designs. For example, differentiable components (such as compressors and turbines) utilize traditional neural network structures, constructing continuous mapping relationships through linear transformations and differentiable activation functions (such as ReLU), maintaining overall differentiability. Non-differentiable components (such as ignition systems and bleed valves) can be embedded with a hybrid structure of decision trees and neural networks (such as the differentiable adaptive neural tree ANT), transforming discrete threshold logic into continuous probability outputs through a "soft decision path," thus overcoming the insufficient modeling capability of traditional neural networks for discrete events. This hierarchical approach enables the optimization model to accurately capture the step characteristics of non-differentiable components while maintaining global differentiability, reducing parameter prediction errors under multiple operating conditions compared to a single neural network model. Furthermore, a closed-loop training mechanism is formed by the first optimization strategy (data-driven) and the second optimization strategy (physical constraint). This dual-strategy collaborative training helps to improve training efficiency and convergence stability. The data-driven strategy uses historical operating data (such as fuel flow and exhaust temperature) to train the neural network and implements gradient descent through the Adam optimizer. The physical constraint strategy introduces a mechanism model regularization term (such as the GRI-Mech combustion mechanism) to the non-differentiable component layer to avoid data overfitting.

[0056] The method may further include: classifying multiple components of the target aero-engine system into differentiable components and non-differentiable components according to a preset classification strategy, wherein a first optimization strategy corresponds to the differentiable components and a second optimization strategy corresponds to the non-differentiable components. In this way, by distinguishing between differentiable and non-differentiable components through a preset classification strategy and matching them with specific optimization strategies, a refined optimization framework of "feature adaptation" is constructed.

[0057] For example, for differentiable components such as compressor blade angle of attack and turbine disk speed, the first optimization strategy employs a neural network optimization algorithm based on gradient descent. Taking a high-pressure compressor as an example, a CNN-LSTM hybrid network containing 100,000 sets of operating condition data is trained to accurately fit the continuous function relationship of the flow-pressure ratio (R²=0.992). Under typical operating conditions of Mach 0.8 and cruising altitude of 11 km, this strategy can find the optimal stator blade angle within 20 iterations, improving compressor efficiency by 1.8%, which translates to an annual fuel saving of 280 tons per engine.

[0058] For example, for non-differentiable components such as fuel injector atomization and ignition systems, the second optimization strategy integrates expert rules and reinforcement learning. For instance, the combustion chamber ignition system adopts a hybrid strategy of "threshold triggering + Q-learning". When the fuel pressure is lower than the critical value (1.2MPa), a predefined rule is invoked to ensure that the ignition energy is ≥25mJ; after exceeding the threshold, the spark plug gap is dynamically adjusted through deep reinforcement learning (adjustment accuracy ±0.1mm).

[0059] This "feature classification-strategy adaptation" optimization paradigm resolves the inherent contradiction of traditional methods that "use continuous models to describe discrete behavior".

[0060] The classification strategy may include: classifying components whose state parameters change continuously with the input as differentiable components, such as the surge boundary of a high-pressure compressor. As the inlet flow rate gradually decreases from the design point, the compressor efficiency decreases smoothly through a continuous function (such as the Moore-Greitzer model), and its characteristic curve is differentiable under most operating conditions; and classifying components whose state parameters change abruptly as non-differentiable components. For example, the spray characteristics of a fuel nozzle show that when the fuel pressure is below the critical value (usually 0.8-1.2 MPa), the fuel is injected in the form of a liquid column. After exceeding the critical pressure, it suddenly atomizes into micron-sized droplets. This "liquid column-atomization" conversion process cannot be described by a single continuous function.

[0061] The classification strategy may include: classifying components whose behavior is affected by a continuous physical field as differentiable components. Differentiable components can be components described by partial differential equations (PDEs), such as the heat conduction process of turbine blades. The gas temperature forms a continuous temperature field inside the blade through convection-conduction coupling, and its distribution satisfies Fourier's law. The temperature derivatives with respect to time and space exist. Under steady-state conditions, the temperature field gradient changes continuously, supporting thermal stress prediction based on neural networks. Components whose behavior is affected by discrete events are classified as non-differentiable components. Non-differentiable components exhibit switching characteristics. A typical example is the combustion chamber ignition system. When the spark plug discharges, plasma is generated instantaneously (lasting about 0.5 ms), causing the local mixture temperature to jump from the ambient temperature (300 K) to the ignition temperature (1800 K). This energy release process is a typical discrete event.

[0062] The classification strategy may include: classifying components with continuous curvature changes in geometric features as differentiable components, such as the airfoil surface of a fan blade, where the chord length, camber, and thickness change continuously from the blade root to the blade tip according to aerodynamic requirements, which can be accurately described by NURBS curves, and the influence of its geometric parameters (such as the location of maximum thickness) on aerodynamic performance can be quantified by derivatives; and classifying components with discontinuous boundaries in their geometric structure as non-differentiable components, such as the tenon groove of a turbine disk, where the mating surface between the tenon teeth and the blade tenon has a stepped structure, the contact stress is concentrated at the corners, and the stress distribution function is not differentiable at the geometric abrupt change point.

[0063] The classification strategy may include: classifying components whose output response exhibits a linear or smooth nonlinear relationship as differentiable components, such as fuel regulators, where the fuel flow rate increases continuously with the control signal as the throttle lever angle increases from 20° to 80°, and its transfer function is differentiable under normal operating conditions; and classifying components whose output response exhibits a threshold effect as non-differentiable components, where the response of non-differentiable components exhibits a threshold effect, such as vent valves, where the valve is completely closed when the compressor outlet pressure is lower than the set value (e.g., 1.2 times the design pressure), and after exceeding the threshold, the valve opening increases linearly with the pressure, and its characteristic curve forms a "bend" at the threshold point.

[0064] Thus, this method accurately classifies differentiable and non-differentiable components through the above four classification strategies, providing a scientific "digital anatomy" standard for hybrid modeling of engine systems. This not only improves modeling accuracy and significantly reduces component characteristic fitting errors, but also optimizes computational efficiency, increasing the iteration speed of the optimization model by several times. Furthermore, the classification strategy enables the control algorithm to achieve "characteristic adaptation," enhancing control robustness and improving stability under extreme operating conditions.

[0065] To effectively address the technical challenge of coexisting differentiable and non-differentiable components in hybrid modeling, this method designs a unique alternating optimization strategy. This strategy decomposes the optimization process into two coordinated stages based on the differentiability of the components. In the differentiable component optimization stage, the network parameters of all non-differentiable components are fixed, and a multi-condition backpropagation algorithm is used to efficiently optimize the gradients of differentiable components, including the differentiable physical model, fully leveraging the advantages of gradient information in parameter space navigation. In the non-differentiable component optimization stage, the network parameters of differentiable components are fixed, and a Gaussian process-based Bayesian optimization method is used to intelligently optimize the non-differentiable components without gradients, effectively balancing exploration and utilization by constructing a surrogate model.

[0066] Specifically, training an initial optimization model based on the first optimization strategy may include: fixing the network parameters of the component layers of all non-differentiable components, and using a multi-working-state backpropagation algorithm to perform gradient optimization on the differentiable components.

[0067] For example, once the network parameters of non-differentiable components (such as ignition systems and bleed valves) are fixed, the backpropagation process only needs to calculate the gradients of differentiable components (such as compressors and turbines), reducing the parameter optimization dimension from over 5000 to around 1200. This significantly improves gradient calculation efficiency and shortens the convergence cycle of complex system optimization. By freezing the network parameters of non-differentiable components, ensuring that their physical behavior strictly follows expert rules or mechanistic models, the failure rate of non-differentiable components can be reduced. For example, the spray characteristic parameters of fuel injectors (atomization angle, droplet diameter) are locked to experimental measurements, preventing the neural network from producing non-physical optimization results due to data noise.

[0068] In this way, by fixing the parameters of the non-differentiable component layer and optimizing the differentiable components through the backpropagation algorithm of multiple working states, a dynamic balance between physical constraints and data-driven approaches is achieved in hybrid modeling. This "fixed-optimized" layered training strategy perfectly balances model flexibility and physical reliability.

[0069] Specifically, training the initial optimization model according to the second optimization strategy may include: fixing the network parameters of the component layer of all differentiable components, and using a Bayesian optimization algorithm based on Gaussian processes to perform gradient-free optimization on the non-differentiable components.

[0070] For example, for non-differentiable components such as fuel injector atomization switching and bleed valve threshold response, Gaussian process models replace traditional gradient calculations with probability distribution modeling, perfectly adapting to discontinuous characteristics. Taking combustion chamber ignition system optimization as an example, the Bayesian optimization algorithm only requires 28 samples (compared to 120 samples by traditional trial and error methods) to find the optimal combination of ignition energy parameters, increasing the high-altitude ignition success rate from 82% to 99.3%. Thus, gradient-free optimization overcomes the non-differentiable barrier, improving the accuracy of parameter adjustment under extreme operating conditions.

[0071] Thus, by fixing the network parameters of differentiable components and employing the Gaussian process Bayesian optimization algorithm, this method achieves gradient-free, precise optimization of non-differentiable components, significantly improving the optimization efficiency and robustness of complex engine systems. Furthermore, by fixing the parameters of the differentiable components, Bayesian optimization concentrates computational resources on the high-influence parameters of the non-differentiable components, shortening the optimization cycle. This hierarchical optimization strategy redefines the intelligent optimization paradigm for complex systems; differentiable components provide stable physical constraint boundaries, while Bayesian optimization efficiently explores the black-box characteristics of non-differentiable components within these boundaries.

[0072] The two optimization phases are organically coordinated through an interactive iterative mechanism. In each iteration, the system alternately optimizes differentiable and non-differentiable components. The optimization progress is judged by monitoring the convergence characteristics of the multi-condition loss function (see below). The optimization process terminates when the loss change between consecutive iterations is less than a preset threshold or the maximum number of iterations is reached. This alternating optimization mechanism not only makes full use of the optimization characteristics of different types of components, but also ensures that the entire hybrid system can converge to the true global optimum through global coordination.

[0073] In the model optimization process, this method further introduces an adaptive parameter adjustment mechanism, which dynamically balances the coefficients. To achieve automated control of the optimization process, Indicates the number of iterations The dynamic equilibrium coefficient under the following conditions Indicates the initial coefficients. This indicates the preset parameters. The adaptive parameter adjustment mechanism can automatically adjust the weight relationship between the output of key components and the overall component output according to the optimization process. In the early stage of optimization, it focuses more on the rapid improvement of overall performance, and in the later stage of optimization, it gradually strengthens the fine-tuning of key components, ensuring that the entire optimization process has both good convergence stability and can achieve a high-quality optimization solution.

[0074] Accordingly, to address the complexity of multi-condition optimization, embodiments of this disclosure construct a unified loss function that balances key performance characteristics with overall performance. In each working state, the maximum value term for that working state is determined based on the maximum value output of each component layer in the corresponding working condition neural network. Furthermore, the output values ​​of all component layers in the corresponding working condition neural network are determined, and the weighted sum term under the corresponding working condition is determined based on all output values ​​and the preset dynamic balance coefficient. The first loss for each working state is determined based on the maximum value and weighted sum term, and the sum of the first losses for all working states is used as the loss function value for training the optimization model. In this way, by constructing a multi-condition loss function by fusing the maximum value term and the weighted sum term, the optimization model training achieves a dual-objective optimization of "extreme constraint-global balance". The maximum value of the output of each component layer is used as a constraint term to ensure that the training process strictly follows the physical limit boundary. The dynamic balance coefficient realizes differentiated weighting of the contribution of different component layers, avoiding the "local optimum trap". The composite loss function of the maximum value term and the weighted sum term makes the gradient direction clearer, and the composite loss function reduces the overall convergence time and significantly reduces the consumption of computational resources.

[0075] The loss function constructed in this embodiment employs a dual constraint mechanism. A maximum value term ensures that key components maintain excellent performance under various operating conditions, avoiding local performance bottlenecks. Simultaneously, a weighted summation term guarantees the performance balance of the entire system, preventing over-optimization of certain components at the expense of others. Finally, the average output value across multiple operating conditions can be used as the final loss function output. This design achieves a fundamental shift from single-condition local optimization to multi-condition global optimization.

[0076] The method may further include: calculating the loss function value based on the dynamic balance coefficient during each model iteration; and obtaining the trained optimized model when the loss function value is less than a preset threshold.

[0077] In this way, by constructing a multi-condition loss function by fusing the maximum value term and the weighted sum term, the optimization model training achieves a dual-objective optimization of "extreme constraint-global balance". The maximum value of the output of each component layer is used as a constraint term to ensure that the training process strictly follows the physical limit boundary. The dynamic balance coefficient realizes differentiated weighting of the contribution of different component layers, avoiding the "local optimum trap". The composite loss function of the maximum value term and the weighted sum term makes the gradient direction clearer, and the composite loss function reduces the overall convergence time and significantly reduces the consumption of computational resources.

[0078] For example, the complete pseudocode flow of this method can be found in Table 1.

[0079] Table 1. Complete pseudocode of this method

[0080] The parameter optimization method provided in this disclosure can intelligently optimize key parameters of an engine under various operating conditions, better overcoming the problem of traditional methods requiring individual parameter adjustments for different operating conditions. It provides an optimal parameter solution for the engine that can adapt to multiple operating conditions, and has significant engineering practical value. This technology uses neural network modeling, combined with backpropagation and Bayesian optimization methods to form a combined intelligent agent, achieving global parameter optimization across operating conditions. It is highly adaptable and has a wide range of applications.

[0081] This method reconstructs the traditional engine system into a multi-layer neural network structure, combining neural network optimization and Bayesian optimization to achieve deep integration of the mechanistic model and the AI ​​model. Each engine component corresponds to a network layer, maintaining physical connections while laying the foundation for unified optimization of heterogeneous models. Regarding the design of the alternating optimization strategy, to address the challenge of coexisting differentiable and non-differentiable components, the optimization process is decomposed into two coordinated stages: a multi-condition backpropagation algorithm is used for differentiable components, while a Gaussian-based Bayesian optimization is used for non-differentiable components. Interactive iteration ensures convergence to the global optimum. This represents a significant breakthrough in multi-condition global optimization. Furthermore, by constructing a unified loss function and employing a dual constraint mechanism, an optimal parameter combination adaptable to various operating conditions can be found, overcoming the technical bottleneck of traditional methods requiring separate parameter tuning for different operating conditions. Simultaneously, computational efficiency is significantly improved, with initial training reduced to within 30 minutes and incremental training to within 2 minutes. An adaptive parameter adjustment mechanism is introduced to automate the optimization process.

[0082] This disclosure also provides a multi-condition parameter optimization device for an aero-engine system. The device includes: a task acquisition module, used to acquire a target optimization task for a target aero-engine system and determine multiple components and multiple operating states of the target aero-engine system; a model construction module, used to construct an optimization model with a neural network structure based on the multiple components and multiple operating states of the target aero-engine system, the optimization model including a condition neural network corresponding to each operating state, each condition neural network including multiple component layers, each component layer corresponding to different components; a model training module, used to train an initial optimization model based on a preset first optimization strategy and a second optimization strategy during each model iteration until the trained optimization model meets preset conditions, thereby obtaining a trained optimization model; and a parameter optimization module, used to determine a target adjustment direction for the target aero-engine system according to the target optimization task, and input the target adjustment direction into the optimization model to obtain an optimization result, the optimization result including the parameter adjustment amount of the components required to meet the target adjustment direction.

[0083] In this way, by reconstructing the engine system into a neural network and employing an alternating optimization strategy to solve the unified optimization problem of heterogeneous components, the limitations of the traditional single modeling paradigm are broken through. Different types of model components are coordinated and processed within a unified computational framework, providing a novel technical path for multi-condition optimization of complex engineering systems. Through hierarchical modeling and hybrid optimization strategies for differentiable / non-differentiable components, the challenges of parameter coupling and nonlinear optimization under multiple operating conditions in aero-engines are solved. Considering the differences in characteristics between differentiable and non-differentiable components in the engine system, the component layers of the operating condition neural network adopt differentiated designs. For example, differentiable components (such as compressors and turbines) use traditional neural network structures, constructing continuous mapping relationships through linear transformations and differentiable activation functions (such as ReLU), maintaining overall differentiability. For non-differentiable components (such as ignition systems and bleed valves), a hybrid structure of decision trees and neural networks (such as the differentiable adaptive neural tree ANT) can be embedded. Through a "soft decision path," discrete threshold logic is transformed into continuous probability outputs, solving the problem of insufficient modeling capability of traditional neural networks for discrete events. This hierarchical approach enables the optimization model to accurately capture the step characteristics of non-differentiable components while maintaining global differentiability, reducing parameter prediction errors under multiple operating conditions compared to a single neural network model. Furthermore, a closed-loop training mechanism is formed by the first optimization strategy (data-driven) and the second optimization strategy (physical constraints). This dual-strategy collaborative training helps improve training efficiency and convergence stability. The data-driven strategy uses historical operating data (such as fuel flow and exhaust temperature) to train the neural network, implementing gradient descent through the Adam optimizer. The physical constraint strategy introduces mechanistic model regularization terms (such as the GRI-Mech combustion mechanism) into the non-differentiable component layer to avoid data overfitting.

[0084] In one possible implementation, the device further includes a classification module for: classifying multiple components of the target aero-engine system into differentiable components and non-differentiable components according to a preset classification strategy, wherein the first optimization strategy corresponds to the differentiable components and the second optimization strategy corresponds to the non-differentiable components.

[0085] In this way, a refined optimization framework of "feature adaptation" is constructed by distinguishing differentiable and non-differentiable components through a pre-defined classification strategy and matching them with specific optimization strategies. For example, for differentiable components such as compressor blade angle of attack and turbine disk speed, the first optimization strategy adopts a neural network optimization algorithm based on gradient descent. Taking a high-pressure compressor as an example, a CNN-LSTM hybrid network containing 100,000 sets of operating condition data is trained to accurately fit the continuous function relationship of flow-pressure ratio (R²=0.992). Under typical operating conditions of Mach number 0.8 and cruising altitude of 11km, this strategy can find the optimal stator blade angle within 20 iterations, improving compressor efficiency by 1.8%, which translates to an annual fuel saving of 280 tons per engine. For non-differentiable components such as fuel nozzle atomization and ignition system, the second optimization strategy integrates expert rules and reinforcement learning. For example, the combustion chamber ignition system employs a hybrid strategy of "threshold triggering + Q-learning." When the fuel pressure is below a critical value (1.2 MPa), predefined rules are invoked to ensure that the ignition energy is ≥25 mJ. After exceeding the threshold, the spark plug gap is dynamically adjusted through deep reinforcement learning (adjustment accuracy ±0.1 mm). This optimization paradigm of "characteristic classification - strategy adaptation" resolves the inherent contradiction of traditional methods that "use continuous models to describe discrete behavior."

[0086] In one possible implementation, the classification strategy includes at least one of the following: classifying components whose state parameters change continuously with the input as differentiable components and components whose state parameters change abruptly as non-differentiable components; classifying components whose behavior is affected by a continuous physical field as differentiable components and components whose behavior is affected by discrete events as non-differentiable components; classifying components whose geometric features exhibit continuous curvature changes as differentiable components and components whose geometric structures have discontinuous boundaries as non-differentiable components; classifying components whose output response exhibits a linear or smooth nonlinear relationship as differentiable components and components whose output response exhibits a threshold effect as non-differentiable components.

[0087] In this way, by accurately classifying differentiable and non-differentiable components using the above four classification strategies, a scientific "digital anatomy" standard is provided for the hybrid modeling of engine systems. This not only improves modeling accuracy and significantly reduces component characteristic fitting errors, but also optimizes computational efficiency, increasing the iteration speed of the optimization model by several times. Furthermore, the classification strategy enables the control algorithm to achieve "characteristic adaptation," which can enhance control robustness and improve stability under extreme operating conditions.

[0088] In one possible implementation, training an initial optimization model according to the first optimization strategy includes: fixing the network parameters of the component layer of all non-differentiable components, and performing gradient optimization on the differentiable components using a multi-working-state backpropagation algorithm.

[0089] In this way, by fixing the parameters of the non-differentiable component layer and optimizing the differentiable components through a multi-working-state backpropagation algorithm, a dynamic balance between physical constraints and data-driven approaches is achieved in hybrid modeling. This "fixed-optimized" layered training strategy perfectly balances model flexibility and physical reliability. During the differentiable component optimization phase, fixing the network parameters of all non-differentiable components and using a multi-working-state backpropagation algorithm to efficiently optimize the gradients of the differentiable physical model and other differentiable components fully leverages the advantages of gradient information in parameter space navigation.

[0090] In one possible implementation, training an initial optimization model according to the second optimization strategy includes: fixing the network parameters of the component layers of all differentiable components, and performing gradient-free optimization on the non-differentiable components using a Bayesian optimization algorithm based on Gaussian processes.

[0091] In this way, by fixing the network parameters of differentiable components and employing the Gaussian process-based Bayesian optimization algorithm, gradient-free precise optimization of non-differentiable components is achieved, significantly improving the optimization efficiency and robustness of complex engine systems. Furthermore, by fixing the parameters of the differentiable components, Bayesian optimization concentrates computational resources on the high-influence parameters of the non-differentiable components, shortening the optimization cycle. This hierarchical optimization strategy redefines the intelligent optimization paradigm for complex systems. Differentiable components provide stable physical constraint boundaries, while Bayesian optimization efficiently explores the black-box characteristics of non-differentiable components within these boundaries. In the non-differentiable component optimization stage, fixing the network parameters of the differentiable components and employing the Gaussian process-based Bayesian optimization method enables intelligent gradient-free optimization of the non-differentiable components, effectively balancing the relationship between exploration and utilization by constructing a surrogate model.

[0092] In one possible implementation, the device further includes an iterative judgment module, used to: calculate the loss function value based on the dynamic balance coefficient during each model iteration; and obtain the trained optimized model if the loss function value is less than a preset threshold.

[0093] In this way, an adaptive parameter adjustment mechanism is further introduced. The optimization process is automated by using dynamic balancing coefficients. The adaptive parameter adjustment mechanism can automatically adjust the weight relationship between the output of key components and the overall component output according to the optimization process. In the early stage of optimization, more attention is paid to the rapid improvement of overall performance, and in the later stage of optimization, the fine-tuning of key components is gradually strengthened to ensure that the entire optimization process has both good convergence stability and can achieve a high-quality optimization solution.

[0094] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0095] This disclosure also provides a multi-condition parameter optimization device for an aero-engine system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0096] This disclosure also provides a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0097] This disclosure also provides a computer program product, including a computer program or a non-volatile computer-readable storage medium carrying the computer program, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0098] Figure 3 A block diagram of a parameter optimization apparatus provided in an embodiment of this disclosure is shown. For example, apparatus 1900 may be provided as a server or terminal device. (Refer to...) Figure 3 The apparatus 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0099] Device 1900 may also include a power supply component 1926 configured to perform power management of device 1900, a wired or wireless network interface 1950 configured to connect device 1900 to a network, and an input / output interface 1958 (I / O interface). Device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM macOS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0100] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of the device 1900 to perform the above-described method.

[0101] Computer-readable storage media can be tangible devices capable of holding and storing programs / instructions used by instruction execution devices. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0102] The computer program (or computer-readable program instructions) described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage medium in the respective computing / processing device.

[0103] The computer program (or computer program instructions) used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions to implement various aspects of this disclosure.

[0104] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0105] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0106] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0107] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0108] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for optimizing multi-condition parameters of an aero-engine system, characterized in that, The method includes: Obtain the target optimization task for the target aero-engine system, and determine multiple components and multiple operating states of the target aero-engine system; An optimization model with a neural network structure is constructed based on multiple components and multiple operating states of the target aero-engine system. The optimization model includes a working condition neural network corresponding to each operating state. Each working condition neural network includes multiple component layers, and each component layer corresponds to a different component. In each model iteration, the initial optimization model is trained based on the preset first optimization strategy and the second optimization strategy until the trained optimization model meets the preset conditions, and the trained optimization model is obtained. The target adjustment direction for the target aero-engine system is determined based on the target optimization task, and the target adjustment direction is input into the optimization model to obtain the optimization result. The optimization result includes the parameter adjustment amount of the components required to meet the target adjustment direction.

2. The method according to claim 1, characterized in that, The method further includes: According to a preset classification strategy, multiple components of the target aero-engine system are divided into differentiable components and non-differentiable components. The first optimization strategy corresponds to the differentiable components, and the second optimization strategy corresponds to the non-differentiable components.

3. The method according to claim 2, characterized in that, The classification strategy includes at least one of the following: Components whose state parameters change continuously with the input are classified as differentiable components, and components whose state parameters change abruptly are classified as non-differentiable components. Components whose behavior is affected by continuous physical fields are classified as differentiable components, and components whose behavior is affected by discrete events are classified as non-differentiable components. Components with continuous curvature changes in geometric features are classified as differentiable components, while components with discontinuous boundaries in their geometric structure are classified as non-differentiable components. Components whose output response exhibits a linear or smooth nonlinear relationship are classified as differentiable components, while components whose output response exhibits a threshold effect are classified as non-differentiable components.

4. The method according to claim 2, characterized in that, Training an initial optimization model based on the first optimization strategy includes: By fixing the network parameters of the component layer for all non-differentiable components, gradient optimization of the differentiable components is performed using a multi-working-state backpropagation algorithm.

5. The method according to claim 2, characterized in that, Training an initial optimization model based on the second optimization strategy includes: By fixing the network parameters of the component layer of all differentiable components, gradient-free optimization is performed on the non-differentiable components using a Bayesian optimization algorithm based on Gaussian processes.

6. The method according to claim 1, characterized in that, The method further includes: During each model iteration, the loss function value is calculated based on the dynamic equilibrium coefficient; When the loss function value is less than a preset threshold, the trained optimized model is obtained.

7. A multi-condition parameter optimization device for an aero-engine system, characterized in that, The device includes: The task acquisition module is used to acquire the target optimization task for the target aero-engine system and determine multiple components and multiple operating states of the target aero-engine system. The model building module is used to build an optimization model with a neural network structure based on multiple components and multiple operating states of the target aero-engine system. The optimization model includes a working condition neural network corresponding to each operating state. Each working condition neural network includes multiple component layers, and each component layer corresponds to a different component. The model training module is used to train the initial optimization model based on the preset first optimization strategy and the second optimization strategy in each model iteration until the trained optimization model meets the preset conditions and the trained optimization model is obtained. The parameter optimization module is used to determine the target adjustment direction for the target aero-engine system according to the target optimization task, and input the target adjustment direction into the optimization model to obtain the optimization result. The optimization result includes the parameter adjustment amount of the components required to meet the target adjustment direction.

8. A multi-condition parameter optimization device for an aero-engine system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, or a non-volatile computer-readable storage medium carrying a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images