Aero-engine inter-turbine temperature dimension increasing method based on physical information neural network

Abstract

The invention discloses an aero-engine high-temperature turbine component temperature parameter dimension increasing method and system based on a physical information neural network (PINN). The internal physical state of the turbine is deduced through the residual power of the rotor, thermodynamic and dynamic constraints are embedded in combination with the PINN, and intelligent reconstruction of unmeasurable parameters such as the outlet temperature and the enthalpy value of the high-pressure turbine is achieved. The method breaks through the dependence of traditional data driving on large-scale samples, is remarkable in overfitting resistance in a small sample scene, is suitable for aero-engine gas path fault diagnosis and all-condition performance monitoring, and has the characteristics of high physical consistency and wide engineering adaptability.

Description

1. Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for aerospace power systems. Specifically, it relates to an intelligent dimensionality enhancement method for key parameters of high-temperature turbine components of aero-engines based on a Physical Information Neural Network (PINN). This method is suitable for solving the problem of real-time reconstruction of unmeasurable thermodynamic parameters between high-pressure and low-pressure turbines in turbofan engines and can be widely applied to aero-engine gas path fault diagnosis, performance monitoring, and health management systems. 2. Background Technology

[0002] As complex thermomechanical systems, aero-engines have high-temperature components (such as the cross-section between the high-pressure turbine (HPT) and the low-pressure turbine (LPT)) operating in extreme environments exceeding 2000K. Sensors are difficult to install directly due to limitations in high-temperature resistant materials and structural constraints (such as narrow turbine clearances and harsh vibration environments), resulting in the inability to measure key parameters (such as turbine inter-turbine temperature T43 and enthalpy H43) in real time. Existing gas path fault diagnosis relies on incomplete sensor data, leading to insufficient diagnostic accuracy, especially under non-steady-state conditions such as acceleration and deceleration, where traditional models fail due to the lack of dynamic parameter compensation mechanisms.

[0003] Existing technologies have the following limitations: bottlenecks in sensor solutions: traditional thermocouples, infrared temperature measurement and other technologies are limited by response speed and lifespan, and cannot cover high-frequency dynamic conditions; defects of pure data-driven approaches: deep learning models (such as LSTM and CNN) require tens of thousands of labeled samples and lack physical constraints, are prone to overfitting, and have poor generalization ability in small sample (such as <500 samples) scenarios; and shortcomings of traditional dynamic models: component-level models based on lumped parameters rely on simplification assumptions (such as quasi-steady-state flow) and cannot accurately describe the non-equilibrium thermodynamic processes inside the turbine.

[0004] How to integrate prior physical knowledge and data-driven models to achieve intelligent dimensionality enhancement of high-temperature component parameters—namely, physical consistency, robustness to small samples, and adaptability across operating conditions—under limited measurable parameters (such as rotational speed and pressure) is a core technical challenge in the field of aerospace propulsion. 3. Summary of the Invention

[0005] To address the problem of insufficient diagnostic accuracy caused by missing parameters of high-temperature components in existing technologies, an intelligent dimension-enhancing method based on turbine power difference inverse thrust dynamics modeling and physical information neural network constraint optimization is proposed. This method overcomes sensor installation limitations, enables high-precision prediction of unmeasurable parameters, and improves the engine's full-condition monitoring capabilities.

[0006] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0007] A method for increasing the dimensionality of inter-turbine temperature in aero-engines based on a physical information neural network, characterized by the following steps:

[0008] Step A) Establish a dynamic modeling module for turbine power difference drive.

[0009] Step B) Establish a physical information neural network intelligent dimension enhancement module.

[0010] Step C) Embed physical constraints and form a complete method.

[0011] Furthermore, the specific steps for establishing the dynamic modeling module driven by the turbine power difference in step A) are as follows:

[0012] Step A1) clearly defines the measurable parameter system. A typical dual-rotor hybrid exhaust turbofan engine includes major components such as the intake, fan, compressor, combustion chamber, high-pressure turbine, and low-pressure turbine. However, the number of measurable parameters is limited. This patent selects the fan outlet parameter P... 22 T 22 Compressor outlet parameters P3 and T3, low-pressure turbine outlet temperatures P5 and T5, and high and low pressure speeds N L N H As measurable parameters. In addition, the inlet outlet parameters can be determined by the engine's flight altitude Mach number, and can also be considered as known parameters.

[0013] Step A2) Utilize measurable parameters and, based on component characteristics, perform thermodynamic derivations such as cross-sectional temperature and enthalpy according to the flow path calculation.

[0014] The temperature and flow rate of the rotating components of the engine can be determined by characteristic diagrams, expressed as follows.

[0015]

[0016] Where α is the guide vane angle of the rotating component, π is the pressure ratio, n is the rotational speed of the shaft, and f w (·) and f e (·) represent the mapping relationships between flow rate and efficiency obtained from the characteristic diagram and guide vane angle, rotational speed, and pressure ratio, respectively. The pressure ratio π can be obtained from the measured parameters of inlet and outlet pressures, and the rotational speed can also be measured. The flow rate of the engine and the corresponding efficiency of the components can be obtained through the above formula.

[0017] Taking the temperature calculation of compressed rotating components such as fans and compressors as an example, it is expressed as follows:

[0018]

[0019] Among them, f a For the oil-gas ratio, H in and H outFor the specific enthalpy of the component's inlet and outlet, H outd For the ideal specific enthalpy of export, H cool To introduce the enthalpy change of the gas, S in and S out f is the ratio of the inlet and outlet entropy of the component. T2S (·) is the temperature-entropy conversion function, f S2H (·) is the entropy-enthalpy conversion function, f T2H (·) represents the temperature-enthalpy conversion function. From the measured parameters, the inlet and outlet temperatures of the rotating component are known. The above formula can be used to determine the characteristics of the venting gas, improving the accuracy of subsequent turbine component calculations.

[0020] Step A3) The rotor's remaining power is reversed using the rotational speed measurement parameters. Based on the rotor dynamics equation, the remaining power of the high and low pressure turbines is calculated through rotational speed acceleration.

[0021] In a conventional aerodynamic-thermodynamic component-level dynamic model, the rotational speed is calculated using rotor dynamics, and the remaining power is calculated using rotational acceleration, as shown below.

[0022]

[0023] Where, variable k represents time k, η L and η H J represents the mechanical efficiency of the low-pressure shaft and the high-pressure shaft. L and J H N represents the rotational inertia of the low-pressure shaft and the high-pressure shaft, N is the power, and Δτ is the simulation step size of the component-level model. Fan and N Comp The power of the high and low pressure turbines can be uniquely determined by calculating the enthalpy and flow rate of the rotating components, and the change in rotational speed can be obtained by the change in the rotational speed measurement signal.

[0024] The power output of the high-pressure and low-pressure turbines is calculated based on the engine shaft dynamics model, considering changes in rotational speed and moment of inertia. Taking the output power of the high-pressure turbine as an example...

[0025]

[0026] By combining the known parameters of the compressor and fan sections, and based on the flow rate and enthalpy difference, the power of the compression components can be calculated, thereby estimating the output power of the high-pressure turbine (HPT) and the low-pressure turbine (LPT).

[0027] The combustion chamber outlet enthalpy (H) can be obtained from the combustion chamber efficiency parameters and the combustion chamber inlet temperature. 41 The flow parameter W inside the high-pressure turbine 41 The turbine interphase temperature can be obtained from the compressor flow rate and the bleed gas flow rate and their set ratio; therefore, it is expressed as follows.

[0028]

[0029] T 43 =f H2T (f a H 43 )

[0030] Furthermore, the specific steps for establishing the physical information neural network intelligent dimensionality enhancement module in step B) are as follows:

[0031] Step B1) Design the network architecture and construct a 5-layer fully connected neural network (1 input layer, 3 hidden layers, and 1 output layer). The input layer is a 9-dimensional vector, which represents the fan outlet parameter P. 22 T 22 Compressor outlet parameters P3, T3, high and low pressure speeds N L N H Flight conditions: altitude, Mach number (H, Ma), fuel flow rate (W) f The input layer is denoted by ; each hidden layer has 64 neurons, with the activation function being Swish, and Dropout (rate 0.1) is added between layers to prevent overfitting; the output layer is a 2D vector, representing the explicit output T. 43 and stealth output H 43 .

[0032] Step B2), design the physical constraint loss function, and design a hybrid loss function to balance data fitting and physical consistency. The loss function is expressed as follows: L total =L data +λ·L physics Where λ is the weighting factor for the physical loss term, and the data loss term is... T represents the mean square error between the PINN predicted values ​​and the simulated data, where M is the training sample size and T is the mean square error. 43,i For the network's predicted output, The data is for the training samples; the physical loss term is... H is used to strengthen the network's satisfaction of thermodynamic conservation relations and ensure that the prediction results have physical interpretability. 43,i This is the network prediction value. This was obtained by inverse calculation based on the turbine power difference.

[0033] Step C) Embedding physical constraints and forming a complete method: The specific steps are as follows:

[0034] Step C1) embeds the thermodynamic balance equation (energy conservation) and the rotor dynamics equation (power balance) into the network in residual form. Taking a high-pressure shaft as an example,

[0035]

[0036] By automatically differentiating and calculating the gradient of the residual with respect to the network parameters, joint optimization based on physical laws and data-driven approaches is achieved.

[0037] Beneficial Effects: The present invention provides a method for enhancing the temperature between aero-engine turbines based on a physical information neural network. Compared with existing technologies, the above technical solution has the following technical advantages:

[0038] (1) Breakthrough in physical consistency: By back-engineering the turbine power difference, strict dynamic constraints are established to ensure that the predicted parameters satisfy energy conservation and rotor dynamics laws, avoiding the "physically meaningless solution" of pure data models; simulations show that the T predicted by the method of this invention is consistent with the actual dynamics of the rotor. 43 The deviation from the theoretical value is <1.5%, which is significantly better than traditional deep learning methods (deviation >5%).

[0039] (2. Improved robustness with small samples: Cross-condition prediction can be achieved with only 200 training samples, and the root mean square error (RMSE) of prediction is as low as 0.0025 (RMSE of traditional methods is >0.01); under acceleration and deceleration conditions (speed change rate >10% / s), the prediction error is reduced by 68% compared with traditional methods.

[0040] (3) Engineering application advantages: Reduce the number of sensors in high-temperature areas (2-3 high-temperature thermocouples can be eliminated), reducing hardware costs by more than 30%; support online real-time calculation (latency <50ms) to meet closed-loop control requirements; can be extended to high-temperature rotating machinery such as industrial gas turbines and aerospace propulsion systems. 4. Description of the attached drawings

[0041] Figure 1 This is a structural diagram of a method for increasing the temperature between aero-engine turbines based on a physical information neural network, as described in this invention.

[0042] Figure 2 This is a schematic diagram of the cross-section and measurable parameters of a turbofan engine component.

[0043] Figure 3 This is a diagram of the PINN dimensionality enhancement technology architecture based on the inverse calculation of turbine power difference.

[0044] Figure 4 This is a simulation curve of the overfitting performance of traditional deep learning.

[0045] Figure 5 These are simulation curves showing the anti-overfitting performance of the method of this invention.

[0046] Figure 6 This is a surface plot showing the impact of physical constraint weights λ on model performance. 5. Detailed Implementation

[0047] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0048] The invention describes a method for increasing the dimensionality of inter-turbine temperature in aero-engines based on a physical information neural network, characterized by comprising the following steps:

[0049] Step A) Data acquisition and preprocessing.

[0050] Step B) Turbine power difference reverse thrust dynamics modeling.

[0051] Step C) Training the physical information neural network.

[0052] Step D) Online prediction and verification of total outlet temperature of high-pressure turbine.

[0053] Furthermore, the specific steps for data acquisition and preprocessing in step A) are as follows:

[0054] Step A1), collect data, including the main components of the turbine engine and measurable parameters such as... Figure 2 As shown, HPT and LPT represent the high-pressure turbine and low-pressure turbine, respectively. The outlet temperature of the turbofan engine combustion chamber is extremely high; in foreign military engines, this cross-section temperature reaches 2000K, posing a significant challenge to temperature measurement. Besides this cross-section, the temperature between the high-pressure and low-pressure turbines is also difficult to measure during actual maneuvering flight due to the harsh environment. Temperature measurement between the high-pressure and low-pressure turbines is crucial for monitoring turbine (especially high-pressure turbine) operation. Apart from these two cross-sections, the measurement environment at other locations is relatively comfortable, resulting in more stable and relatively reliable signals, such as the fan outlet parameter P. 22 T 22 Compressor outlet parameters P3 and T3, low-pressure turbine outlet temperatures P5 and T5, and high and low pressure speeds N L N H Furthermore, the inlet exit parameters can be determined by the engine's Mach number at flight altitude, can be calculated by an atmospheric data computer, or can be treated as known parameters;

[0055] Step A2) involves preprocessing the data, applying median filtering to the measurement parameters to eliminate noise interference, normalizing all types of collected data, and ensuring that the normalized parameters of the training and test sets are consistent. The normalization expression is as follows: x′=(x-μ train ) / σ train , where μ train Let σ be the mean of the signal. train Let x be the standard deviation of the signal, and x and x′ be the values ​​before and after normalization.

[0056] Furthermore, the specific steps for turbine power difference reverse thrust dynamics modeling in step B) are as follows:

[0057] Step B1), rotor dynamics calculation: the temperature and flow rate of the engine's rotating components can be determined from the characteristic diagram, expressed as follows:

[0058]

[0059] Where α is the guide vane angle of the rotating component, π is the pressure ratio, n is the rotational speed of the shaft, and f w (·) and f e (·) represent the mapping relationships between flow rate and efficiency obtained from the characteristic diagram and guide vane angle, rotational speed, and pressure ratio, respectively. The pressure ratio π can be obtained from the measured parameters of inlet and outlet pressures, and the rotational speed can also be measured. The flow rate of the engine and the corresponding efficiency of the components can be obtained through the above formula.

[0060] Taking compressor power calculation as an example, it is expressed as follows:

[0061] N Comp =W air (h3-h 22 )

[0062] The flow rate can be calculated from the compressor characteristics, and the enthalpy is calculated as follows.

[0063] Taking the temperature calculation of compressed rotating components such as fans and compressors as an example, it is expressed as follows:

[0064]

[0065] Among them, f a For the oil-gas ratio, H in and H out For the specific enthalpy of the component's inlet and outlet, H outd For the ideal specific enthalpy of export, H cool To introduce the enthalpy change of the gas, S n and S out f is the ratio of the inlet and outlet entropy of the component. T2S (·) is the temperature-entropy conversion function, f S2H (·) is the entropy-enthalpy conversion function, f T2H (·) represents the temperature-enthalpy conversion function. From the measured parameters, the inlet and outlet temperatures of the rotating component are known. The above formula can be used to determine the characteristics of the venting gas, improving the accuracy of subsequent turbine component calculations.

[0066] In a conventional aerodynamic-thermodynamic component-level dynamic model, the rotational speed is calculated using rotor dynamics, and the remaining power is calculated using rotational acceleration, as shown below.

[0067]

[0068] Where, variable k represents time k, η L and η HJ represents the mechanical efficiency of the low-pressure shaft and the high-pressure shaft. L and J H N represents the rotational inertia of the low-pressure shaft and the high-pressure shaft, N is the power, and Δτ is the simulation step size of the component-level model. Fan and N Comp The power of the high and low pressure turbines can be uniquely determined by calculating the enthalpy and flow rate of the rotating components, and the change in rotational speed can be obtained by the change in the rotational speed measurement signal.

[0069] Step B2), derive the thermodynamic parameters.

[0070] The combustion chamber outlet enthalpy (H) can be obtained from the combustion chamber efficiency parameters and the combustion chamber inlet temperature. 41 The flow parameter W inside the high-pressure turbine 41 The turbine interphase temperature can be obtained from the compressor flow rate and the bleed gas flow rate and their set ratio, therefore it is expressed as follows:

[0071]

[0072] Step C) The specific steps for training the physical information neural network are as follows:

[0073] Step C1) Design the network architecture, integrating turbine information intelligent dimension enhancement technology based on turbine power difference backpropagation, such as... Figure 3 As shown, a system is constructed with rotational speed N L N H Fuel flow rate W f Intake conditions (height H, Mach number Ma), fan outlet parameters (P) 22 T 22 The compressor outlet parameters (P3, T3) are used as inputs, with the high-pressure turbine outlet temperature T as the input. 43 As an explicit output, this output can be supervised learning using measurement data or high-precision simulation data to form a data loss term, such as the high-pressure turbine outlet enthalpy H. 43 As a hidden output, and as a variable that is not directly observable but can be verified through thermodynamic calculations, a physical constraint loss term is introduced. During training, in addition to utilizing existing sensor data, nonlinear partial differential constraint terms are embedded to ensure that the prediction results conform to thermodynamic and kinetic laws.

[0074] Step C2), design the loss function, which is expressed as follows:

[0075] L total =L data +λ·L physics

[0076] Where λ is the weighting factor for the physical loss term, and the data loss term is... T represents the mean square error between the PINN predicted values ​​and the simulated data, where M is the training sample size and T is the mean square error. 43,i For the network's predicted output, The data is for the training samples; the physical loss term is... H is used to strengthen the network's satisfaction of thermodynamic conservation relations and ensure that the prediction results have physical interpretability. 43,i This is the network prediction value. This was obtained by inverse calculation based on the turbine power difference.

[0077] Step D) Online prediction and verification of high-pressure turbine outlet total temperature: The specific steps are as follows:

[0078] Step D1): Real-time prediction process. Preprocessed measurable parameters are acquired via the bus. The power difference inverse calculation algorithm from step B is then called to obtain H. 43 Input 9-dimensional parameters into the trained model and output T 43 .

[0079] Step D2) verifies the overfitting performance of this invention compared to traditional deep learning methods. Traditional deep learning methods typically rely on massive amounts of labeled data for training. By constructing complex neural network structures and using optimization algorithms such as gradient descent to continuously adjust model parameters, they learn the latent patterns and features in the data. While this data-driven training model has achieved excellent results in many tasks, it is also prone to falling into the trap of data oversaturation and model overfitting.

[0080] In the experiment, the first 200 samples were used as training data, and the last 200 were used as prediction data. In the initial stage, through optimized training, the model accuracy quickly reached the 1e-8 order of magnitude, demonstrating good fitting performance. However, as... Figure 4 As shown, as the number of iterations increased from 15,000 to 20,000 and then 25,000, the prediction error exhibited a significant change, first decreasing and then increasing. In the early stages of training, the model fully explored the features of the training data and continuously optimized the parameters, effectively reducing the prediction error. However, as training progressed, the model gradually exhibited overfitting, excessively memorizing noise and special samples in the training data, leading to a decline in its generalization ability on the prediction data. Ultimately, this caused the prediction error to rise continuously. This trend is also common in traditional deep learning methods, reflecting the common challenges faced by both types of methods during the training process.

[0081] To address the unavoidable error fluctuations and overfitting issues inherent in traditional methods during training, Physical Information Neural Networks (PINNs) offer a novel approach. By integrating physical laws into the model architecture, they provide additional constraints to the optimization process, potentially breaking this deadlock and achieving more stable and efficient predictive performance. PINNs embed physical laws such as differential equations and conservation laws as regularization terms into the neural network architecture. Through the reconstruction of the loss function, the model satisfies physical constraints while fitting data, providing an additional regularization mechanism for the optimization process. Figure 5 The simulation experiment demonstrating the intelligent dimensionality enhancement of turbine information based on turbine power difference back-calculation provides strong empirical evidence for this advantage. Maintaining the same data partitioning, network architecture, and initial training configuration as previous experiments, the prediction error of the PINN model is similar to that of traditional methods after 15,000 iterations, showing their comparability in initial learning capabilities. However, as training progresses to 20,000, 25,000, and even 30,000 iterations, the prediction error curve of the PINN model exhibits a drastically different trend, not only avoiding divergence but also steadily decreasing. This significant difference stems from the deep integration of physical laws: the principles of energy conservation and fluid mechanics followed during turbine system operation provide a reliable constraint framework for the PINN model, enabling it to accurately capture the dynamic relationship between turbine power difference and operating parameters even with limited training data. Even when the last 200 prediction data points differ significantly from the first 200 training data points in terms of operating conditions and environmental parameters, the PINN model can still achieve high-precision predictions across operating conditions based on its deep understanding of the turbine system's operating mechanism through physical constraints. This robust performance under small sample conditions not only breaks through the bottleneck of traditional deep learning's dependence on large-scale data, but also provides innovative solutions for complex engineering scenarios such as aero-engine performance monitoring and industrial turbine fault diagnosis, demonstrating the core value and broad application prospects of physical information-driven modeling in solving practical problems.

[0082] Step D3) verifies the quantitative impact of physical constraint weights (λ) on model accuracy and training efficiency. In the performance evaluation of the intelligent dimensionality augmentation model for turbine information fusion based on turbine power difference, the physical constraint weights λ, as a core adjustment parameter, have a significant impact on model prediction accuracy and training efficiency. Figure 6As shown in (a), as λ gradually increases from 0 to 0.01, the root mean square error (RMSE) of prediction continuously decreases from 0.0175 to 0.0025, a reduction of 85.7%, indicating that the introduction of physical constraints effectively suppresses model overfitting and significantly improves prediction accuracy. When λ is further increased to 0.1 and 1, the RMSE slightly rebounds to 0.0035 and 0.0030, respectively. This is because excessively high physical constraint weights cause the model to rely too heavily on prior knowledge, limiting the data-driven adaptive capability and leading to a slight performance degradation. Therefore, under the training sample conditions in this case, when λ is 0.01, the model achieves an efficient balance between physical constraints and data learning, reaching optimal performance.

[0083] It is worth noting that the optimal value of λ is not fixed, but closely related to the quantity and quality of the training samples. When the training samples are of high quality and accurately reflect the true operating rules of the system, moderately increasing λ can further enhance the model's ability to capture the inherent laws through physical constraints, accelerate convergence, and improve prediction accuracy. However, if the training samples have problems such as noise interference, missing data, or labeling errors, an excessively high λ will exacerbate the difficulty of model training. At this time, the contradiction between physical constraints and low-quality samples becomes prominent—the model must satisfy strict physical laws while adapting to incorrect or incomplete data features. This will force the model to spend more iterations to reconcile the relationship between the two, leading to a decrease in training efficiency and even potentially getting trapped in a local optimum. Figure 6 (b) shows the evolution of the number of network training iterations with λ. When λ exceeds 0.01, the number of iterations increases sharply, which is a direct reflection of the model's difficulty in adapting to the characteristics of the training samples under the high-weight physical constraints.

[0084] Therefore, in practical applications, λ needs to be set according to the specific circumstances of the training samples. For high-quality samples, physical constraints can be appropriately strengthened to optimize performance; while when the sample quality is poor, the weight of λ should be reduced to prioritize the model's learning ability from the data and avoid physical constraints becoming an obstacle to training. This finding provides more targeted guidance for the accurate inverse estimation of the turbine system's operating state and parameter optimization, helping to maximize model performance under different data conditions.

[0085] In this implementation case, by integrating turbine power difference backpropagation with a physical information neural network (PINN) intelligent dimensionality enhancement method, high-precision prediction of turbine inter-turbine temperature (T43) and enthalpy (H43) of aero-engines is achieved using only measurable parameters such as fan outlet and compressor outlet and 200 small samples for training. The steady-state error is less than 0.22%, and the non-steady-state response delay is less than 0.5 seconds. Furthermore, overfitting is effectively suppressed through a physical constraint mechanism (the error continues to decrease after 30,000 iterations). Compared with traditional deep learning methods, this invention reduces the deployment of high-temperature sensors by 50%, lowers hardware costs by more than 30%, and improves prediction accuracy by 70% in small sample scenarios. It significantly overcomes the diagnostic bottleneck caused by the lack of parameters for high-temperature components of aero-engines, providing an innovative solution with both physical reliability and engineering practicality for intelligent monitoring of complex thermodynamic systems under all operating conditions.

[0086] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for increasing the dimensionality of inter-turbine temperature in aero-engines based on a physical information neural network, characterized in that, Includes the following steps: Step A), Measurable parameter acquisition: Collect pressure / temperature parameters, high and low pressure speeds, fuel flow rate and flight environment parameters at the engine fan outlet, compressor outlet and low-pressure turbine outlet; Step B), Turbine power reverse calculation: Based on the rotor dynamics equations, the remaining power of the high and low pressure turbines is calculated through rotational acceleration and moment of inertia, and the turbine output power is derived. Step C), Thermodynamic parameter calculation: Combining the combustion chamber outlet enthalpy and turbine flow rate, the theoretical enthalpy and temperature between the turbines are derived using the energy conservation equation: Step D], Physical Information Neural Network Modeling: Construct a PINN model with measurable parameters as input, turbine inter-temperature as explicit output, and turbine inter-enthalpy as implicit output. Optimize network parameters through a hybrid loss function that includes data loss terms and physical loss terms. The physical loss terms are constructed based on the theoretical enthalpy value derived from the turbine power difference.

2. The intelligent dimensionality enhancement method based on turbine power difference inverse dynamics modeling and physical information neural network constraint optimization according to claim 1, characterized in that, The rotor dynamics equation mentioned in step B) is: Where, variable k represents time k, η L and η H J represents the mechanical efficiency of the low-pressure shaft and the high-pressure shaft. L and J H N represents the rotational inertia of the low-pressure shaft and the high-pressure shaft, N is the power, and Δτ is the simulation step size of the component-level model. Fan and N Comp The power output of the high and low pressure turbines can be uniquely determined by calculating the enthalpy and flow rate of the rotating components, while the change in rotational speed can be obtained from the change in the rotational speed measurement signal. The power output of the high and low pressure turbines is calculated based on the engine shaft dynamics model, the change in rotational speed, the calculated power of the compressor, and the moment of inertia.

3. The intelligent dimensionality enhancement method based on turbine power difference inverse dynamics modeling and physical information neural network constraint optimization according to claim 1, characterized in that, The loss function of the physical information neural network described in step D) is: L total =L data +λ·L physics Where λ is the weighting factor for the physical loss term, and the data loss term is... T represents the mean square error between the PINN predicted values ​​and the simulated data, where M is the training sample size and T is the mean square error. 43,i For the network's predicted output, The data is for the training samples; the physical loss term is... H is used to strengthen the network's satisfaction of thermodynamic conservation relations and ensure that the prediction results have physical interpretability. 43,i This is the network prediction value. This was obtained by inverse calculation based on the turbine power difference.

4. The method according to claim 3, characterized in that, The theoretical enthalpy value is calculated using constructed physical constraints. A turbine component information augmentation method driven by turbine power difference is introduced as the physical constraint for the neural network. Based on other measurable parameters, such as pressure and temperature, it can progressively deduce the component's flow rate and power, ultimately deriving the turbine inter-component temperature, which is T. 43 =f(W,T,P,n), introducing the thermodynamic equilibrium equation as the physical constraint of the neural network, the construction form is as follows: The turbine power residual obtained from the power difference calculation is used as a physical residual loss term in PINN and embedded into the total loss function to enhance the physical consistency of the model.

5. A parameter enhancement system for high-temperature components of an aero-engine based on the method of any one of claims 1-4, characterized in that, include: The data acquisition module consists of a sensor group and signal conditioning circuitry used to acquire measurable parameters of the engine. The dynamics calculation unit is an embedded processor that integrates a rotor power inverse calculation algorithm and a thermodynamic parameter calculation module; the physical information neural network unit uses the constructed PINN model to realize intelligent prediction of turbine parameters. The output interface provides enhanced temperature and enthalpy parameters, supporting data interaction with the engine health management system.

Images