A method for extracting failure feature based on mechanism model reinforced learning

Abstract

This invention belongs to the field of aero-engine fault diagnosis technology, specifically relating to a fault feature extraction method based on a mechanism model using reinforcement learning. The specific technical solution involves: constructing a high-precision aero-engine blade-level performance simulation model; using this model as a reinforcement learning agent; building a reinforcement learning environment; training the reinforcement learning model in a specific aero-engine fault environment; and when the reinforcement learning model converges, extracting aero-engine fault features and constructing an aero-engine fault feature library. This model can simulate faults in different working blades, supporting the simulation and feature enhancement of faults in different levels of working blades within the same component, achieving the goal of refined aero-engine blade-level diagnosis. Simultaneously, it utilizes reinforcement learning to adaptively learn fault features to address the problems of limited fault samples and the inability to obtain unknown fault features.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine fault diagnosis technology, specifically relating to a fault feature extraction method based on mechanism model reinforcement learning. Background Technology

[0002] Aero engines and gas turbines operate in complex and harsh environments. The high temperature and pressure conditions of rotating components cause their performance to degrade at a faster rate than other components, making them more prone to failure, engine damage, and serious safety accidents. Therefore, monitoring, tracking, and evaluating the health status of engines, especially the turbomachinery performance, can accurately and quickly predict and locate potential faults, thereby improving engine safety and reducing maintenance costs.

[0003] Engines, due to wear and tear, design flaws, and external damage, can develop various failure modes, such as blade chipping or breakage in the airflow components. These different failure modes alter the characteristics of corresponding engine components, causing engine operating parameters to deviate from their healthy baseline. The deviation of engine component characteristic parameters (such as compressor flow rate and efficiency) from the healthy baseline is called a fault characteristic, which is the dependent variable in the transmission of fault effects. These fault characteristics are often unmeasurable parameters. To obtain fault characteristics, it is necessary to establish the correlation between measurable engine operating parameters (such as temperature, pressure, rotor speed, and fuel flow rate at various engine sections) and fault characteristics. A common approach is to construct a deviation function using the engine's main performance and process parameters as the optimization objective, and the component degradation factor as the optimized parameter, thereby correcting the engine component characteristics and obtaining the fault characteristics.

[0004] Engine faults typically have few features, and some are unknown, making it difficult for traditional methods to meet the need for an enhanced fault feature database. Currently, engine fault feature extraction usually involves building engine component-level models and using whole-engine test data under real fault conditions to correct component characteristics, obtaining engine component characteristics that deviate from a healthier state, and thus obtaining the component's fault features.

[0005] Engine component-level modeling is based on component-level engine modeling theory. It decomposes the engine structure into major components, from the engine inlet to the exhaust nozzle. Based on aerodynamic and thermodynamic principles, it establishes equations for gas flow and thermodynamic processes one by one. Then, based on physical equilibrium relationships such as engine flow balance and pressure balance, it obtains a system of simultaneous equations that can simulate the engine's common operating state. Finally, it solves this nonlinear system of equations to obtain the parameters of various relevant sections of the engine. The dataset describing the performance parameters of engine components, such as flow rate, efficiency, pressure ratio, and speed, within the full envelope is called the component characteristic map. During the calculation process, the state of the components needs to be marked on the component characteristic map to obtain the corresponding component performance parameters.

[0006] For rotating components like compressors and turbines, a fault might only occur in a single stage of blades, leading to a degradation of the entire component's performance parameters. Establishing only component-level fault characteristics is insufficient for the refined diagnosis of engine faults. Furthermore, in actual aero-engine operation, the relationship between faults and symptoms is often not a simple one-to-one correspondence; one fault may correspond to multiple symptoms, and conversely, one symptom may be caused by multiple faults. Moreover, due to the limited availability of fault samples, actual aero-engine operating data alone is insufficient to encompass all fault characteristics, and there are still unknown and difficult-to-classify types of aero-engine faults. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a fault feature extraction method based on a mechanism model using reinforcement learning. This technique can be used to establish a high-precision engine blade-level performance simulation model, achieving the goal of refined engine blade-level diagnosis. At the same time, it utilizes reinforcement learning to adaptively learn fault features to solve the problems of insufficient fault samples and the inability to obtain unknown fault features.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a fault feature extraction method based on a mechanism model, which involves constructing a high-precision aero-engine blade-level performance simulation model, using this model as a reinforcement learning agent, building a reinforcement learning environment, training the reinforcement learning model in a specific aero-engine fault environment, and extracting aero-engine fault features when the reinforcement learning model converges, thereby constructing an aero-engine fault feature library; the reinforcement learning environment includes a state space and an action space.

[0009] Preferred method: The high-precision aero-engine blade-level performance simulation model construction method is as follows: replace the characteristic diagram of the rotating component of the aero-engine with a single-stage working blade diagram, correct the single-stage working blade characteristic diagram using component test data, construct a high-precision aero-engine blade-level performance simulation model, calculate component performance parameters, and correct the calculated component performance parameters using whole-engine test data.

[0010] Preferred method: The method for constructing the high-precision aero-engine blade-level performance simulation model is as follows:

[0011] S1. Obtain the working characteristic diagram of a single-stage blade: Based on the engine blade test, obtain the performance parameters of the blade under different operating conditions, and use the interpolation algorithm based on the test discrete points to obtain the continuous performance parameters of the blade under different operating conditions, and form the working characteristic diagram of a single-stage blade.

[0012] S2. Correcting the single-stage working blade characteristic diagram using component test data: Based on the principle that the blades of each stage in the rotating component follow the same rotation speed and flow balance, the inlet and outlet aerodynamic parameters of each stage blade in the component are obtained from the single-stage blade working characteristic diagram. The single-stage performance parameters are calculated using the inlet and outlet aerodynamic parameters of each stage blade, and then the single-stage working blade characteristic diagram is corrected using component test data.

[0013] S3. Construct a high-precision simulation model of aero-engine blade-level performance: Determine the calculation components according to the type of aero-engine;

[0014] S4. Calculation Initialization: Obtain the corrected single-stage working blade characteristic curve and the general characteristic curve of other components; for the engine, give the throttle position, flight Mach number, flight altitude and atmospheric environment; select the aero-engine regulation law;

[0015] S5, correction factors for structural components and blade characteristics, and initialization;

[0016] S6. Perform thermodynamic calculations from the undisturbed section far in front of the aero-engine to the aero-engine exit section;

[0017] S7. Based on the continuous flow and power balance within the aero-engine, construct a common working equation;

[0018] S8. Solve the common working equations to obtain the performance parameters of the component cross section and the blade cross section;

[0019] S9. Based on the performance parameters of the component cross section and blade cross section in step S8, and the test data obtained from the faulty aero-engine, construct an optimization objective function;

[0020] S10. Solve the optimization objective function to obtain the component and blade characteristic correction factors;

[0021] S11. Determine whether the optimization objective function has converged. If it has not converged, repeat steps S6-S10. If it has converged, obtain the component and blade characteristic correction factor parameters and construct the fault characteristics.

[0022] Preferably, in step S5, the formula for calculating the blade-level correction factor is as follows:

[0023]

[0024] C πI C represents the pressure ratio / pressure drop ratio correction factor for the first-stage working blades. ηI π represents the efficiency correction factor for the first-stage working blade. act This represents the boost ratio / drop ratio parameter obtained from experimental data, π. ref This indicates the pressure ratio / pressure drop ratio parameter in the uncorrected working blade characteristic diagram, η.act η represents the efficiency parameter obtained from experimental data. ref This indicates the efficiency parameters in the uncorrected working blade characteristic diagram;

[0025] And / or, the formula for calculating the correction factor for intake duct components is as follows:

[0026]

[0027] σ act σ represents the total pressure recovery coefficient obtained from experimental data. ref This represents the uncorrected total pressure recovery factor;

[0028] And / or, the formula for calculating the correction factor for combustion chamber components is as follows:

[0029]

[0030] η act η represents the combustion efficiency obtained from experimental data. ref σ represents the uncorrected combustion efficiency. act σ represents the total pressure recovery coefficient obtained from experimental data. ref This represents the uncorrected total pressure recovery coefficient.

[0031] Preferably, in step S9, the objective function is optimized as follows:

[0032] e I =(X ref / X act ) I -1.0 (I = 1, 2, ..., n)

[0033]

[0034] F = F(C) π1 C πn C η1 C ηn ,I,R η η σ )

[0035] X act X is the selected measurement parameter obtained from the faulty engine test. ref The parameters are calculated using the uncorrected blade stage and component characteristic curves; F is the constructed deviation optimization function, a I As a weighting factor;

[0036] The final optimization objective function can be expressed as a function of the blade stage and component correction factors.

[0037] The preferred reinforcement learning process is as follows:

[0038] T1. Building an intelligent agent model and fault environment for an aero-engine: The high-precision aero-engine blade-level performance simulation model is used as the intelligent agent; for the engine, the fault environment includes throttle position, flight Mach number, flight altitude, atmospheric conditions, and regulation rules;

[0039] T2. Construct a reinforcement learning interactive environment in a fault environment, including the state parameter space and action space of the reinforcement learning algorithm;

[0040] T3. Based on prior knowledge of aero-engines, construct an evaluation function for aero-engine intelligent agents;

[0041] T4. The aero-engine intelligent agent is trained by interacting with the fault environment to obtain fault characteristics that meet given fault conditions.

[0042] T5. Determine whether the reinforcement learning algorithm has converged. If it has converged, write the fault feature vector and the corresponding performance parameter threshold obtained in step T4 into the database to construct the aero-engine fault feature library. If it has not converged, adjust the performance parameter threshold under the aero-engine fault conditions and proceed to step T4.

[0043] Preferably, in step T2, the state parameter space consists of performance parameters calculated by the aero-engine intelligent agent model, including thrust and fuel consumption rate.

[0044] And / or, the action space is a parameter vector composed of correction factors for aero-engine blades and components:

[0045] [I, F] π1 F η1 C π1 C η1 , ..., R η R σ T π1 T η1 ,...]

[0046] I represents the correction factor for the total pressure recovery coefficient of the intake components, F π1 F η1 C represents the boost ratio of the first-stage fan blades and the efficiency correction factor, respectively. π1 C η1 R represents the pressure ratio and efficiency correction factor of the first stage blades of the high-pressure compressor, respectively. η R σ T represents the combustion efficiency of the combustion chamber, the total pressure recovery coefficient correction factor, and T, respectively. π1 T η1These represent the pressure ratio and efficiency correction factor of the first-stage turbine blades, respectively; the value range of each correction factor is [0,1], where 0 indicates complete damage and 1 indicates a healthy state.

[0047] Preferably, step T4 includes:

[0048] T41. Initialize the aircraft engine fault environment;

[0049] T42. Initialize the state parameters, which are the performance parameters calculated by the high-precision aero-engine blade-level performance simulation model under healthy conditions, and input them into the neural network;

[0050] T43. Neural network output action: Output vectors corresponding to blade and component-level correction factors;

[0051] T44. Actions executed by the aero-engine intelligent agent model: The inputs to the aero-engine intelligent agent model include aero-engine fault environment parameters and neural network action parameters;

[0052] T45, State Update: The aero-engine intelligent agent model calculates new state parameters and obtains a reward value based on the evaluation function;

[0053] T46, Experience Pool;

[0054] T47. Neural network parameter update: Randomly obtain this set of data from the experience pool and update the neural network parameters according to the data update strategy;

[0055] T48. Save action parameters: When the algorithm converges, save the action parameters, i.e., the fault feature vector.

[0056] Correspondingly: an electronic device, comprising:

[0057] One or more processors;

[0058] Storage device for storing one or more programs;

[0059] When the one or more programs are executed by the one or more processors, the one or more processors implement the fault feature reinforcement learning extraction method based on the mechanism model.

[0060] Correspondingly: a computer-readable medium storing a computer program that, when executed by a processor, implements the fault feature extraction method based on a mechanism model through reinforcement learning.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] By constructing a high-precision aero-engine blade-level performance simulation model, this model is used as a reinforcement learning agent. The model is trained in a specific aero-engine fault environment. If the trained model converges, aero-engine fault features are extracted, and an aero-engine fault feature library is constructed. This high-precision aero-engine blade-level performance simulation model can simulate faults in different working blades, supporting the simulation and feature enhancement of faults in different levels of working blades within the same component, and supporting refined engine fault diagnosis. Through the agent's autonomous decision-making, the high-precision aero-engine blade-level performance simulation model is embedded into the fault adaptive optimization source domain. Engine performance parameter degradation thresholds are set, and fault feature extraction is completed through reinforcement learning adaptive learning. The fault features and their corresponding fault manifestations are then incorporated into the fault feature library. Attached Figure Description

[0063] Figure 1 This is a flowchart of the fault feature extraction method based on mechanism model according to the present invention.

[0064] Figure 2 This is a flowchart illustrating the construction of the high-precision engine blade-level performance simulation model for this invention.

[0065] Figure 3 This is a flowchart of the reinforcement learning-based aero-engine fault feature extraction method of the present invention;

[0066] Figure 4 This is a schematic diagram illustrating the construction of the high-precision engine blade-level performance simulation model of the present invention;

[0067] Figure 5 This is a diagram illustrating the framework of the reinforcement learning principle of this invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0069] like Figure 1 As shown, this invention discloses a fault feature extraction method based on a mechanism model using reinforcement learning. Its core idea is to: construct a high-precision aero-engine blade-level performance simulation model; use the established blade-level performance simulation model as a reinforcement learning agent; establish a reinforcement state space and action space to complete the reinforcement learning environment construction; train the reinforcement learning model in a specific aero-engine fault environment; if the trained reinforcement learning model converges, extract aero-engine fault features and construct an aero-engine fault feature library.

[0070] The fault feature extraction method of the present invention mainly includes two parts: one is the construction method of a high-precision aero-engine blade-level performance simulation model, and the other is the aero-engine fault feature extraction method based on reinforcement learning of the high-precision aero-engine blade-level performance simulation model.

[0071] The core idea of ​​the high-precision aero-engine blade-level performance simulation model construction method is as follows: Replace the traditional characteristic diagrams of rotating components in an aero-engine with single-stage working blade diagrams. Correct the single-stage working blade characteristic diagrams using component test data to reflect the performance of a single-stage blade under the combined working environment of the components. Construct a blade-level performance simulation model, calculate component performance parameters based on multiple corrected single-stage working blade characteristic diagrams, and then correct the calculated component performance parameters using whole-engine test data to reflect the current performance state of the engine components. The rotating components include the compressor and turbine components.

[0072] Because multiple single-stage blade characteristic maps are used, the performance states of different stages of blades within the same component can be obtained. Specifically, the simplex method is recommended for correcting the characteristic maps of aero-engine components, but the use of other solution methods still falls within the scope of this patent.

[0073] Furthermore, such as Figure 2 As shown, the method for constructing the blade-level performance simulation model is as follows:

[0074] S1. Obtaining the single-stage blade operating characteristic diagram: The performance parameters of the blade under different operating conditions are obtained. Based on the discrete test points, an interpolation algorithm is used to obtain the continuous performance parameters of the blade under different operating conditions, forming a single-stage blade operating characteristic diagram. According to aero-engine blade tests, different operating conditions include different speeds and different flow rates. Blade types include, but are not limited to, fan blades, low-pressure compressor blades, high-pressure compressor blades, high-pressure turbine blades, and low-pressure turbine blades.

[0075] S2. Correcting the single-stage working blade characteristic diagram using component test data: Based on the principle that the blades of each stage in the rotating component follow the same rotation speed and flow balance, the inlet and outlet aerodynamic parameters and single-stage performance parameters of each stage blade in the component are obtained using the single-stage blade working characteristic diagram. The single-stage performance parameters are then calculated using the inlet and outlet aerodynamic parameters of each stage blade obtained from the component test data, and the single-stage working blade characteristic diagram is corrected accordingly.

[0076] S3. Construct a blade-level performance simulation model: Determine the components to be calculated based on the type of aero-engine, such as the inlet, fan / compressor, combustion chamber, turbine, and nozzle, depending on the direction of medium flow. Different types of engines have different structures.

[0077] S4. Calculation Initialization: Obtain the corrected single-stage working blade characteristic curve and the general characteristic curve of other components; for the engine, give the throttle position, flight Mach number, flight altitude and atmospheric environment; select the aero-engine adjustment law.

[0078] S5, the correction factor for structural components and blade characteristics, and initialization.

[0079] S6. Perform thermodynamic calculations from the undisturbed section at the far front of the aero-engine to the exit section. Specifically, for rotating components, the thermodynamic calculations begin with the first-stage blade. Based on the inlet parameters, trial performance parameters, and the characteristic curve of the first-stage working blade, the exit thermodynamic parameters of that stage blade are obtained. The calculation method for subsequent stages is similar, until the exit section thermodynamic parameters of the last stage blade of the rotating component are calculated; these are the exit section thermodynamic parameters of that component. For other components, the exit section thermodynamic parameters are calculated from the inlet section thermodynamic parameters based on trial parameters and the general component characteristic diagram.

[0080] S7. Based on the continuous flow and power balance within the aero-engine, construct common working equations. The number of common working equations is the same as the number of trial parameters to satisfy the equation closure condition. The trial parameters are the unknowns encountered during the thermodynamic calculations in step S6.

[0081] S8. Solve the common working equations to obtain the performance parameters of the component cross section and the blade cross section;

[0082] S9. Based on the performance parameters of the component cross section and blade cross section calculated in step S8, and the test data obtained from the faulty aero-engine, construct an optimization objective function.

[0083] S10. Solve the objective function to obtain the component and blade characteristic correction factors. The optimization methods include, but are not limited to, the simplex method.

[0084] S11. Determine whether the optimization objective function has converged. If it has not converged, repeat steps S6-S10. If it has converged, obtain the component and blade characteristic correction factor parameters. The vector composed of these parameters is the fault characteristic.

[0085] Furthermore, in step S5, the formula for calculating the blade-level correction factor is as follows:

[0086]

[0087] C πI C represents the pressure ratio / pressure drop ratio correction factor for the first-stage working blades. ηI π represents the efficiency correction factor for the first-stage working blade. act This represents the boost ratio / drop ratio parameter obtained from experimental data, π. refThis indicates the pressure ratio / pressure drop ratio parameter in the uncorrected working blade characteristic diagram, η. act η represents the efficiency parameter obtained from experimental data. ref This indicates the efficiency parameters in the uncorrected working blade characteristic diagram.

[0088] The formula for calculating the correction factor for intake components is as follows:

[0089]

[0090] σ act σ represents the total pressure recovery coefficient obtained from experimental data. ref This represents the uncorrected total pressure recovery coefficient.

[0091] The formula for calculating the correction factor for combustion chamber components is as follows:

[0092]

[0093] In the formula, R η η represents the combustion efficiency of the combustion chamber. act η represents the combustion efficiency obtained from experimental data. ref Indicates the uncorrected combustion efficiency. R σ σ represents the correction factor for the total pressure recovery coefficient of the combustion chamber. act σ represents the total pressure recovery coefficient obtained from experimental data. ref This represents the uncorrected total pressure recovery coefficient.

[0094] Furthermore, in step S9, several engine measurement parameters are selected, including but not limited to thrust F, fuel consumption rate sfc, and fan pressure ratio π. F Turbine inlet total temperature T4, compressor pressure ratio π C Based on the selected measurement parameters, the optimization function is constructed as follows:

[0095] e I =(X ref / X act ) I -1.0 (I = 1, 2, ..., n)

[0096]

[0097] F = F(C) π1 C πn C η1 C ηn ,I,R η η σ )

[0098] X actX is the selected measurement parameter obtained from the faulty engine test. ref The parameters are calculated using the uncorrected blade stage and component characteristic curves; F is the constructed deviation optimization function, a I These are weighting factors. The final optimization objective function can be expressed as a function of blade-level and component correction factors.

[0099] Furthermore, such as Figure 3 As shown, the process of the reinforcement learning-based aero-engine fault feature extraction method is as follows:

[0100] T1. Establishing an intelligent agent model and fault environment for the aero-engine: The blade-level performance simulation model is used as the intelligent agent. For the engine, the fault environment includes throttle position, flight Mach number, flight altitude, atmospheric conditions, and regulation patterns.

[0101] T2. Construct a reinforcement learning interactive environment in the fault environment, including the state parameter space and action space of the reinforcement learning algorithm. The state parameter space consists of performance parameters calculated by the aero-engine intelligent agent model, including thrust and fuel consumption rate. The action space is a parameter vector composed of correction factors for aero-engine blades and components.

[0102] [I, F] π1 F η1 C π1 C η1 , ..., R η R σ T π1 T η1 ,...]

[0103] I represents the correction factor for the total pressure recovery coefficient of the intake components, F π1 F η1 C represents the boost ratio of the first-stage fan blades and the efficiency correction factor, respectively. π1 C η1 R represents the pressure ratio and efficiency correction factor of the first stage blades of the high-pressure compressor, respectively. η R σ T represents the combustion efficiency of the combustion chamber, the total pressure recovery coefficient correction factor, and T, respectively. π1 T η1 These represent the pressure ratio and efficiency correction factor of the first-stage turbine blades, respectively; the value range of each correction factor is [0,1], where 0 indicates complete damage and 1 indicates a healthy state.

[0104] T3. Based on prior knowledge of aero-engines, construct an evaluation function for aero-engine intelligent agents.

[0105] Based on the thrust performance parameter of aero-engines, the evaluation function is as follows:

[0106]

[0107] Among them, F ratio F represents the thrust performance parameter of the engine under the current condition. limit This indicates the threshold value for engine thrust performance parameters set under fault conditions.

[0108] Based on the fuel consumption rate, a performance parameter of aero engines, the evaluation function is as follows:

[0109]

[0110] Among them, sfc ratio This represents the engine's fuel consumption performance parameter under the current condition, sfc limit This indicates the threshold value for the engine fuel consumption performance parameter under fault conditions.

[0111] The final evaluation function is obtained by linear summation: reward = r1 + r2.

[0112] T4. The aero-engine intelligent agent is trained by interacting with the fault environment to obtain fault characteristics that meet given fault conditions.

[0113] T5. Determine whether the reinforcement learning algorithm has converged. If it has converged, write the fault feature vector and the corresponding performance parameter threshold obtained in step T4 into the database to construct the aero-engine fault feature library. If it has not converged, adjust the performance parameter threshold under the aero-engine fault conditions and proceed to step T4.

[0114] Furthermore, step T4 includes:

[0115] T41. Initialize the aircraft engine fault environment. For the engine, the fault environment includes throttle position, flight Mach number, flight altitude, atmospheric conditions, and control rules.

[0116] T42. Initialize state parameters. The state parameters are the performance parameters calculated by the blade-level performance simulation model under healthy conditions, and are input into the neural network.

[0117] T43. Neural network output action. That is, output the vector corresponding to the blade and component-level correction factors.

[0118] T44. Actions executed by the aero-engine intelligent agent model: The inputs to the aero-engine intelligent agent model include aero-engine fault environment parameters and neural network action parameters.

[0119] T45. State Update: The aero-engine intelligent agent model calculates new state parameters and obtains a reward value based on the evaluation function.

[0120] T46. Store the experience pool. Store <current state, updated state, action parameters, reward value> as a set of data into the experience pool.

[0121] T47. Neural Network Parameter Update: Randomly obtain this set of data from the experience pool and update the neural network parameters according to the data update strategy.

[0122] T48. Save Action Parameters: When the algorithm converges, save the action parameters, i.e., the fault feature vector. Convergence condition: The reward value obtained is 0, that is, the engine performance parameters obtained from the action space parameters meet the set performance parameter threshold under the fault condition.

[0123] This invention also discloses an electronic device, comprising: one or more processors; a storage device for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the fault feature reinforcement learning extraction method based on the mechanistic model. The electronic device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers.

[0124] This invention also discloses a computer-readable medium storing a computer program that, when executed by a processor, implements the fault feature reinforcement learning extraction method based on the mechanistic model as described above. Embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing execution flow... Figure 1-5 The program code for the method shown.

[0125] It should be noted that the computer-readable medium of this disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0126] In this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0127] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0128] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0129] Example 1: Fault Feature Extraction Based on Aero-engine Blade-Level Performance Simulation Model

[0130] Taking the WS-15 twin-shaft turbofan hybrid engine as an example, it includes 3-stage fan blades, 6-stage high-pressure compressor blades, a single-stage high-pressure turbine blade, and a single-stage low-pressure turbine blade.

[0131] Step 1: Obtain the single-stage blade operating characteristic diagram. In this embodiment, both the high-pressure turbine and low-pressure turbine components are single-stage blades, so it is not necessary to replace the component characteristic diagram with a single-stage blade operating characteristic diagram. Based on engine blade tests, the performance parameters of the blades under different operating conditions such as different speeds and flow rates are obtained. Based on the test discrete points, an interpolation algorithm is used to obtain continuous performance parameter data of the blades under different operating conditions, forming a single-stage blade operating characteristic diagram, and finally obtaining the single-stage blade operating characteristic diagrams of the fan and high-pressure turbine.

[0132] Step 2: Correct the single-stage blade characteristic diagrams of the fan and high-pressure compressor using component test data. Based on the principle that each stage of the blades in the fan and high-pressure compressor follows equal rotational speed and balanced flow rate, the inlet and outlet aerodynamic parameters and single-stage performance parameters of each stage of the blades can be obtained from the single-stage blade characteristic diagrams obtained in Step 1. The single-stage performance parameters are then calculated using the inlet and outlet aerodynamic parameters of each stage of the blades obtained from the component test data, and the single-stage blade characteristic diagrams are subsequently corrected.

[0133] Step 3: Construct a simulation model of engine blade-level performance. In this embodiment, the engine components include the intake duct, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, and exhaust nozzle.

[0134] Step 4: Initialization Calculation. Obtain the corrected characteristic curves of the fan and high-pressure compressor single-stage working blades, as well as the general characteristic curves of other components; specify the throttle position, flight Mach number, flight altitude, and atmospheric conditions; select the engine regulation law.

[0135] Step 5: Construct correction factors for component and blade characteristics, and initialize all correction factors to 1.

[0136] The formula for calculating the correction factor for intake components is as follows:

[0137]

[0138] In equation (1), I represents the correction factor for the total pressure recovery coefficient of the intake duct components, and σ act σ is the total pressure recovery coefficient obtained from experimental data; ref This is the uncorrected total pressure recovery coefficient.

[0139] The formula for calculating the fan blade-level correction factor is as follows:

[0140]

[0141] In equation (2), F πI This represents the pressure ratio correction factor for the first stage blade of the fan, π. Iact π represents the pressure ratio of the first-stage fan blades; Iref This represents the uncorrected pressure ratio of the first-stage fan blades. F ηI η represents the efficiency correction factor for the first-stage blades of the fan. Iact η represents the efficiency of the first-stage fan blades; Iref This represents the uncorrected efficiency of the first-stage fan blades.

[0142] The formula for calculating the correction factor for the high-pressure compressor blade stage is as follows:

[0143]

[0144] In equation (3), C πI π represents the pressure ratio correction factor for the first stage blades of a high-pressure compressor. Iact π represents the pressure ratio of the blades in the first-stage high-pressure compressor. Iref This indicates the uncorrected pressure ratio of the first-stage high-pressure compressor blades. (C) ηI η represents the efficiency correction factor for the first stage blades of the high-pressure compressor. Iact η represents the efficiency of the first-stage high-pressure compressor blades; Iref This represents the uncorrected blade efficiency of the first-stage high-pressure compressor blades.

[0145] The formula for calculating the correction factor for combustion chamber components is as follows:

[0146]

[0147] In equation (4), R η η represents the combustion efficiency correction factor for the combustion chamber. act This represents the combustion chamber efficiency parameter obtained based on experimental data; η ref This represents the uncorrected combustion chamber efficiency parameter. R σ σ represents the correction factor for the total pressure recovery coefficient of the combustion chamber. act σ represents the total pressure recovery coefficient of the combustion chamber obtained from experimental data. ref This represents the uncorrected total pressure recovery coefficient of the combustion chamber.

[0148] The formulas for calculating the correction factors for high-pressure and low-pressure turbine components are as follows:

[0149]

[0150] In equation (5), T π π represents the turbine pressure ratio correction factor. Iact This represents the turbine pressure ratio parameter obtained from experimental data; π Iref This represents the uncorrected turbine pressure ratio parameter. T η η represents the turbine efficiency correction factor. act This represents the turbine efficiency parameter obtained from experimental data; η ref This indicates the uncorrected turbine efficiency parameter.

[0151] Step 6: Perform thermodynamic calculations from the undisturbed section at the far front of the engine to the engine outlet section. Specifically, for the thermodynamic calculations of rotating components, starting with the first-stage blades, obtain the outlet thermodynamic parameters of the first-stage blades based on their inlet parameters, trial performance parameters, and the characteristic curves of the first-stage working blades. The calculation method for subsequent stages is similar, until the outlet section thermodynamic parameters of the last stage blade of the rotating component are calculated; these are the outlet section thermodynamic parameters of that component. For other components, the outlet section thermodynamic parameters are calculated from the inlet section thermodynamic parameters based on trial parameters and the general component characteristic diagram.

[0152] Step 7: Construct the common working equations. In this embodiment, the parameters used are the first-stage fan blade pressure ratio, the total temperature at the inlet of the high-pressure turbine, the first-stage high-pressure compressor blade pressure ratio, the relative equivalent speed of the high-pressure compressor, the equivalent flow rate at the inlet of the high-pressure turbine, and the equivalent flow rate at the inlet of the low-pressure turbine. The common working equations specifically include:

[0153] (1) Low-pressure turbine / fan flow balance.

[0154]

[0155] Among them, W 4.5cor This represents the initial calculated flow rate of the low-pressure turbine. This represents the calculated low-pressure turbine equivalent flow rate.

[0156] (2) Low-pressure turbine / fan power balance.

[0157] E2=(N TL -N CL ) / N CL

[0158] Where, N TL N CL These represent the low-pressure turbine power and the fan power, respectively.

[0159] (3) High pressure turbine / high pressure compressor flow balance.

[0160]

[0161] Among them, W 4cor This represents the initial value of the converted flow rate at the high-pressure turbine inlet. This represents the calculated equivalent flow rate at the high-pressure turbine inlet.

[0162] (4) Power balance of high pressure turbine / high pressure compressor.

[0163] E4=(N TH -N CH ) / N CH

[0164] Where, N TH N CH These represent the power of the high-pressure turbine and the power of the high-pressure compressor, respectively.

[0165] (5) Static pressure balance at the inlet of the mixing chamber.

[0166] E5=(p5-p 5II ) / p5

[0167] Among them, p5,p 5II These represent the static pressure at the low-pressure turbine outlet and the static pressure at the bypass duct outlet, respectively.

[0168] (6) The mass flow rate of the tail nozzle airflow is continuous.

[0169]

[0170] Among them, W8, These represent the mass flow rates of the exhaust gas in the tailpipe and the mass flow rates of the afterburner outlet gas, respectively.

[0171] The above contains 6 equations and 6 trial parameters, which satisfy the closure property of the equations.

[0172] Step 8: Solve the common working equations to obtain the performance parameters of the component cross section and the blade cross section.

[0173] Step 9: Construct the function to be optimized based on the performance parameters of each section calculated in Step 8 and the test data obtained from the faulty engine. Select several engine measurement parameters, including thrust F, fuel consumption rate sfc, and fan pressure ratio π. F Turbine inlet total temperature T4, high-pressure compressor pressure ratio π C And so on. An optimization function is constructed based on the selected measurement parameters. The constructed optimization objective function is as follows:

[0174] e I =(X ref / X act ) I -1.0 (I = 1, 2, ..., 5)

[0175]

[0176] F = F(I, F) π1 F η1 F π3 F η3 C π1 C η1 C π6 C η6 R η R σ T π T η )

[0177] Among them, X act X is the selected measurement parameter obtained from the faulty engine test. ref The parameters are calculated using the uncorrected blade stage and component characteristic curves; F is the constructed deviation optimization function, and the weighting factor is a. I When set to 1, the final optimization function can be expressed as a function of the blade-level and component correction factors.

[0178] Step 10: Solve the optimization function using the simplex method to obtain the component and blade characteristic correction factors.

[0179] Step 11: Determine whether the optimization function has converged. If it has not converged, repeat steps 6-10. If it has converged, obtain the blade and component correction factor parameters. The vector composed of these parameters is the fault characteristic.

[0180] Example 2: A Method for Extracting Fault Features of Aero-engines Based on Reinforcement Learning

[0181] like Figure 5 As shown, reinforcement learning can be described as the problem of maximizing rewards or achieving specific goals through continuous exploration and learning by an aero-engine intelligent agent in the process of interacting with the environment. This embodiment is also based on the WS-15 dual-shaft turbofan hybrid engine, and the specific implementation process will be described in detail below.

[0182] Step 1: Build an intelligent agent model and fault environment for the aero-engine. Use the model obtained in Module 1 as the intelligent agent. The fault environment includes throttle position, flight Mach number, flight altitude, atmospheric conditions, and control rules.

[0183] Step 2: Construct a reinforcement learning interactive environment within the fault environment, including the state parameter space and action space of the reinforcement learning algorithm. The state parameter space consists of performance parameters calculated by the aero-engine intelligent agent model, including thrust and fuel consumption rate. The action space can be represented as a parameter vector composed of engine blade and component correction factors, specifically:

[0184] [I, F] π1 F η1 F π3 F η3 C π1 C η1 C π6 C η6 R η R σ T hπ T hη T lπ T lη ]

[0185] Where I represents the total pressure recovery coefficient correction factor for the intake duct components, and F π1 F η1 C represents the boost ratio of the first-stage fan blades and the efficiency correction factor, respectively. π1 C η1 R represents the pressure ratio and efficiency correction factor of the first stage blades of the high-pressure compressor, respectively. η R σ T represents the combustion efficiency of the combustion chamber, the total pressure recovery coefficient correction factor, and T, respectively. hπ T hη T represents the high-pressure turbine pressure ratio and efficiency correction factor, respectively. lπ T lη These represent the low-pressure turbine pressure ratio and efficiency correction factor, respectively. The value range of each correction factor is [0,1], where 0 represents complete failure and 1 represents a healthy state.

[0186] Step 3: Based on prior knowledge of aero-engines, design the evaluation function for the aero-engine intelligent agent. Starting from the aero-engine performance parameter thrust, the evaluation function is as follows:

[0187]

[0188] Among them, F ratio F represents the thrust performance parameter of the engine under the current condition. limit This represents the threshold value for engine thrust performance parameters set manually under fault conditions.

[0189] Based on the fuel consumption rate, a performance parameter of aero engines, the evaluation function is as follows:

[0190]

[0191] The final evaluation function is obtained by linear summation: reward = r1 + r2.

[0192] Step 4: The aero-engine intelligent agent is trained by interacting with the fault environment, ultimately obtaining fault characteristics that satisfy given fault conditions. This includes:

[0193] Step 41: Initialize the aircraft engine fault environment, including throttle position, flight Mach number, flight altitude, atmospheric conditions, and adjustment rules.

[0194] Step 42: Initialize state parameters. State parameters are the performance parameters calculated in the engine simulation under healthy conditions, and are input into the neural network.

[0195] Step 43: The neural network outputs the action. That is, it outputs the vectors corresponding to the blade and component-level correction factors.

[0196] Step 44: The aero-engine intelligent agent model executes actions. The inputs to the aero-engine intelligent agent model include aero-engine fault environment parameters and neural network action parameters.

[0197] Step 45: State Update. The aero-engine intelligent agent model calculates new state parameters and simultaneously obtains a reward value based on the evaluation function.

[0198] Step 46: Store the experience pool. Store <current state, updated state, action parameters, reward value> as a set of data into the experience pool.

[0199] Step 46: Neural Network Parameter Update. Randomly select a set of data from the experience pool and update the neural network parameters according to the data update strategy.

[0200] Step 47: Save the action parameters. When the algorithm converges, save the action parameters, i.e., the fault feature vector. Convergence condition: The reward value obtained is 0, that is, the engine performance parameters obtained from the action space parameters meet the performance parameter thresholds set manually under the fault conditions.

[0201] Step S5: Determine whether the algorithm has converged. If it has converged, write the fault feature vector and the corresponding performance parameter threshold obtained in step S4 into the database to construct the aero-engine fault feature library. If it has not converged, modify the performance parameter threshold under the engine fault conditions and repeat step S4.

[0202] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, substitutions, or variations made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for mechanism model-based fault feature reinforcement learning extraction, characterized in that: A high-precision aero-engine blade-level performance simulation model is constructed. This model is used as a reinforcement learning agent. A reinforcement learning environment is built, and the reinforcement learning model is trained in a specific aero-engine fault environment. When the reinforcement learning model converges, aero-engine fault features are extracted, and an aero-engine fault feature library is constructed. The reinforcement learning environment includes a state space and an action space. The method for constructing the high-precision aero-engine blade-level performance simulation model is as follows: S1. Obtain the working characteristic diagram of a single-stage blade: Obtain the performance parameters of the blade under different working conditions, and use the interpolation algorithm based on the experimental discrete points to obtain the continuous performance parameters of the blade under different working conditions, and form a working characteristic diagram of a single-stage blade. S2. Correcting the single-stage working blade characteristic diagram using component test data: Based on the principle that the blades of each stage in the rotating component follow the same rotation speed and flow balance, the inlet and outlet aerodynamic parameters of each stage blade in the component are obtained from the single-stage blade working characteristic diagram. The single-stage performance parameters are calculated using the inlet and outlet aerodynamic parameters of each stage blade, and then the single-stage working blade characteristic diagram is corrected using component test data. S3. Construct a high-precision simulation model of aero-engine blade-level performance: Determine the calculation components according to the type of aero-engine, including the air intake, fan, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, and tail nozzle; S4. Calculation Initialization: Obtain the corrected characteristic curves of the fan and high-pressure compressor single-stage working blades, and the general characteristic curves of other components; for the engine, give the throttle position, flight Mach number, flight altitude, and atmospheric environment; select the aero-engine regulation law; S5, correction factors for structural components and blade characteristics, and initialization; S6. Perform thermodynamic calculations from the undisturbed section at the far front of the aero-engine to the outlet section. For rotating components, start with the first-stage blade. Obtain the outlet thermodynamic parameters of the first-stage blade based on the inlet parameters, trial performance parameters, and characteristic curves of the first-stage working blade. The calculation method for subsequent stages is the same as for the first-stage blade, until the outlet section thermodynamic parameters of the last stage blade of the rotating component are calculated. These are the outlet section thermodynamic parameters of the rotating component. For other components, calculate the outlet section thermodynamic parameters from the inlet section thermodynamic parameters based on trial parameters and general component characteristic diagrams. S7. Based on the continuous flow and power balance within the aero-engine, construct a common working equation; The number of working equations is the same as the number of trial parameters, in order to satisfy the equation closure condition. The trial parameters are the unknowns encountered in the thermodynamic calculations in step S6. S8. Solve the common working equations to obtain the performance parameters of the component cross section and the blade cross section; S9. Based on the performance parameters of the component cross section and blade cross section in step S8, and the test data obtained from the faulty aero-engine, construct an optimization objective function; S10. Solve the optimization objective function to obtain the component and blade characteristic correction factors; S11. Determine whether the objective function has converged. If it has not converged, repeat steps S6-S10. If convergence is achieved, the component and blade characteristic correction factor parameters are obtained, and the fault characteristics are constructed. 2.The mechanism model-based fault feature reinforcement learning extraction method according to claim 1, characterized in that: In S5, the formula for calculating the blade-level correction factor is as follows: ； This represents the pressure ratio / pressure drop ratio correction factor for the first-stage working blade. π represents the efficiency correction factor for the first-stage working blade. Iact This represents the boost ratio / drop ratio parameter obtained from experimental data, π. Iref This indicates the pressure ratio / pressure drop ratio parameter in the uncorrected working blade characteristic diagram, η. Iact η represents the efficiency parameter obtained from experimental data. Iref This indicates the efficiency parameters in the uncorrected working blade characteristic diagram; And / or, the formula for calculating the correction factor for intake duct components is as follows: ； This represents the total pressure recovery coefficient obtained from experimental data. This represents the uncorrected total pressure recovery factor; And / or, the formula for calculating the correction factor for combustion chamber components is as follows: ； This indicates the combustion efficiency obtained based on experimental data. This indicates the uncorrected combustion efficiency. This represents the total pressure recovery coefficient obtained from experimental data. This represents the uncorrected total pressure recovery coefficient.

3. The fault feature extraction method based on a mechanism model according to claim 2, characterized in that: In step S9, the objective function is optimized as follows: ； ； ； These are the selected measurement parameters obtained from the faulty engine test. The parameters are calculated using the uncorrected blade stage and component characteristic curves; F is the constructed deviation optimization function. As a weighting factor; The final optimization objective function can be expressed as a function of the blade stage and component correction factors.

4. The fault feature extraction method based on a mechanism model according to claim 3, characterized in that: The reinforcement learning process is as follows: T1. Building an intelligent agent model and fault environment for an aero-engine: The high-precision aero-engine blade-level performance simulation model is used as the intelligent agent; for the engine, the fault environment includes throttle position, flight Mach number, flight altitude, atmospheric conditions, and regulation rules; T2. Construct a reinforcement learning interactive environment in a fault environment, including the state parameter space and action space of the reinforcement learning algorithm; T3. Based on prior knowledge of aero-engines, construct an evaluation function for aero-engine intelligent agents; T4. The aero-engine intelligent agent is trained by interacting with the fault environment to obtain fault characteristics that meet given fault conditions. T5. Determine whether the reinforcement learning algorithm has converged. If it has converged, write the fault feature vector and the corresponding performance parameter threshold obtained in step T4 into the database to construct the aero-engine fault feature library. If convergence is not achieved, adjust the performance parameter thresholds under the aircraft engine failure condition and proceed to step T4.

5. The fault feature extraction method based on a mechanism model according to claim 4, characterized in that: In T2, the state parameter space consists of performance parameters calculated by the aero-engine intelligent agent model, including thrust and fuel consumption rate. And / or, the action space is a parameter vector composed of correction factors for aero-engine blades and components: ； This represents the correction factor for the total pressure recovery coefficient of the intake duct components. , These represent the boost ratio and efficiency correction factor of the first-stage fan blades, respectively. , These represent the pressure ratio and efficiency correction factor of the first-stage blades of the high-pressure compressor, respectively. , These represent the combustion chamber efficiency and the total pressure recovery coefficient correction factor, respectively. , These represent the pressure ratio and efficiency correction factor of the first-stage turbine blades, respectively; the value range of each correction factor is [0,1], where 0 indicates complete damage and 1 indicates a healthy state.

6. The fault feature extraction method based on a mechanism model according to claim 4, characterized in that: The T4 includes: T41. Initialize the aircraft engine fault environment; T42. Initialize the state parameters, which are the performance parameters calculated by the high-precision aero-engine blade-level performance simulation model under healthy conditions, and input them into the neural network; T43. Neural network output action: Output vectors corresponding to blade and component-level correction factors; T44. Actions executed by the aero-engine intelligent agent model: The inputs to the aero-engine intelligent agent model include aero-engine fault environment parameters and neural network action parameters; T45, State Update: The aero-engine intelligent agent model calculates new state parameters and obtains a reward value based on the evaluation function; T46, Experience Pool; T47. Neural network parameter update: Randomly obtain this set of data from the experience pool and update the neural network parameters according to the data update strategy; T48. Save action parameters: When the algorithm converges, save the action parameters, i.e., the fault feature vector.

7. An electronic device, characterized in that: include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a fault feature extraction method based on a mechanism model as described in any one of claims 1-6.

8. A computer-readable medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements a fault feature reinforcement learning extraction method based on a mechanism model as described in any one of claims 1-6.

Images