CN-121980272-A - Method and system for constructing aeroengine state prediction model based on physical constraint

CN121980272ACN 121980272 ACN121980272 ACN 121980272ACN-121980272-A

Abstract

The application belongs to the technical field of aero-engine performance test, in particular relates to an aero-engine state prediction model construction method and system based on physical constraint, and aims to solve the problems of high calculation complexity and poor data quality of the existing physical information model. The method comprises the steps of obtaining historical operation data of the aeroengine and an engineering simplified physical constraint rule set, predicting performance parameters by adopting a deep learning model, constructing a total loss function formed by weighting data loss items and physical loss items to train the model, wherein the physical loss items are generated based on the deviation degree of the predicted performance parameters and the engineering simplified physical constraint rule set and are used for replacing complex partial differential equation constraint. According to the application, by introducing engineering simplified physical rules, the calculation cost of model training is obviously reduced, the model is effectively guided to learn the characteristics conforming to the physical rules, and under the condition of limited data, the accuracy and generalization capability of the prediction model are obviously improved.

Inventors

SHEN YAFENG
ZHOU MINGYANG
Miao Keqiang
TANG ZHENCHAO
SUN JIAXIAN
HU CHUNYAN

Assignees

中国科学院工程热物理研究所

Dates

Publication Date: 20260505
Application Date: 20260209

Claims (10)

1. The method for constructing the state prediction model of the aeroengine based on the physical constraint is characterized by comprising the following steps of: Acquiring historical operation data representing the operation state of the aeroengine and an engineering simplified physical constraint rule set representing the physical operation mechanism of the aeroengine; a deep learning model is adopted, and according to the historical operation data, the performance parameters of the aeroengine are predicted, so that predicted performance parameters are obtained; constructing a total loss function, and training a deep learning model by utilizing the total loss function to optimize internal parameters of the deep learning model so as to obtain the state prediction model of the aeroengine; the total loss function is formed by weighting a data loss term and a physical loss term, the data loss term is generated based on deviation of the predicted performance parameter and a corresponding performance parameter true value, and the physical loss term is generated based on deviation degree between the predicted performance parameter and the engineering simplified physical constraint rule set and is used for applying punishment loss to the predicted performance parameter violating the rule set.
2. The method for constructing an aircraft engine state prediction model based on physical constraints according to claim 1, wherein the engineering reduced physical constraint rule set comprises at least one of the following types: A monotonicity constraint rule for imposing a penalty loss when the predicted performance parameter or an intermediate physical quantity derived therefrom does not satisfy a monotonic trend of change to be followed with a change in engine operating state; and a boundary range constraint rule for applying a penalty loss when the predicted performance parameter or its derived intermediate physical quantity exceeds a preset reasonable physical range.
3. The method for constructing an aircraft engine state prediction model based on physical constraints according to claim 2, wherein the performance parameter is a remaining service life of an aircraft engine, and the engineering simplified physical constraint rule set comprises: a component temperature monotonicity constraint rule which represents unidirectional variation of component temperature along with deepening of degradation degree of an engine; A component pressure monotonicity constraint rule which represents unidirectional variation of component pressure along with deepening of degradation degree of an engine; a monotonicity constraint rule of the cooling air flow, which represents unidirectional variation of the cooling air flow along with the deepening of the degradation degree of the engine; a rotational speed deviation constraint rule for representing the variation of deviation among different rotational speeds along with the deepening of the degradation degree of the engine; and limiting the predicted value range constraint rule that the predicted value of the residual service life is in a reasonable numerical range.
4. The method for constructing an aeroengine state prediction model based on physical constraints according to claim 3, wherein the constraint rule specifically comprises: ; ; ; ; ; ; ; Wherein, the Representing non-normalized residual life predictions, For the maximum remaining useful life to be present, For the sample Average deviation between the fan speed and the core unit speed, 、、、、 Respectively represent samples The average rate of change of the high pressure compressor outlet temperature, the low pressure turbine outlet temperature, the high pressure compressor outlet pressure, the high pressure turbine cooling air flow rate, and the low pressure turbine cooling air flow rate, characterizes a trend of the corresponding physical quantity as the engine is degraded.
5. The method for constructing an aeroengine state prediction model based on physical constraints according to claim 2, wherein the performance parameter is thrust of the aeroengine, and the engineering simplified physical constraint rule set comprises: a temperature consistency constraint rule for ensuring physical consistency between measurement data of a plurality of temperature sensors at different positions; A temperature-pressure association constraint rule for representing association relations between different physical quantities; And defining a temperature range constraint rule that the temperature characteristic value is in a preset reasonable physical range interval.
6. The method for constructing an aeroengine state prediction model based on physical constraints according to claim 1, wherein the deep learning model is a long-term and short-term memory network model or a transformer network model.
7. The method for constructing the state prediction model of the aeroengine based on the physical constraint according to claim 1, wherein before training the deep learning model by utilizing the total loss function, the method further comprises the step of optimizing super parameters of the deep learning model by adopting a quantum genetic algorithm improved by multiple strategies, and the optimization specifically comprises the following steps: Generating a uniform initial population covering the whole solution space by using a Latin hypercube sampling method, and transforming the codes of key parameters by using a concave mapping function so as to realize denser solution distribution in a preset experience optimal area; Determining the rotation direction of the quantum bit according to the adaptability difference between the current individual and the historical optimal individual, and dynamically adjusting the rotation step length according to the angle difference between the current individual and the historical optimal individual; The lowest fitness individuals in the population are periodically removed and new individuals are generated by adding random perturbations near the optimal individuals to maintain population diversity.
8. The method of constructing a physical constraint based aeroengine state prediction model of claim 1, further comprising data enhancing the historical operating data prior to training a deep learning model using the total loss function, the data enhancing comprising: And performing data enhancement on the historical operating data to generate an extended data set, wherein the data enhancement process utilizes the engineering simplified physical constraint rule set to apply constraint on the data generation process so as to ensure that the generated data accords with physical characteristics, and the prediction and training of the deep learning model are performed based on the extended data set.
9. The method for constructing a state prediction model of an aircraft engine based on physical constraints of claim 8, further comprising, prior to training a deep learning model using the total loss function: And optimizing the super parameters of the deep learning model by adopting a quantum genetic algorithm improved by multiple strategies, wherein sample data used in the optimization process is a data set enhanced and expanded by the data.
10. An aeroengine state prediction model building system based on physical constraints, comprising: The data and rule acquisition module is configured to acquire historical operation data representing the operation state of the aeroengine and an engineering simplified physical constraint rule set representing the operation physical characteristics of the aeroengine; The performance prediction module is configured to predict the performance parameters of the aeroengine according to the historical operation data by adopting a deep learning model to obtain predicted performance parameters; the model optimization module is configured to construct a total loss function, and train a deep learning model by utilizing the total loss function so as to optimize internal parameters of the deep learning model and obtain the state prediction model of the aeroengine; the total loss function is formed by weighting a data loss term and a physical loss term, the data loss term is generated based on deviation of the predicted performance parameter and a corresponding performance parameter true value, and the physical loss term is generated based on deviation degree between the predicted performance parameter and the engineering simplified physical constraint rule set and is used for applying punishment loss to the predicted performance parameter violating the rule set.

Description

Method and system for constructing aeroengine state prediction model based on physical constraint Technical Field The application belongs to the technical field of aero-engine performance test, and particularly relates to a method and a system for constructing an aero-engine state prediction model based on physical constraint. Background Aeroengines are used as the core power units of modern aircraft, and the health status of the aeroengines is directly related to the safety of flight and the operation economy of airlines. Therefore, the development of accurate and efficient predictive maintenance technology is of great importance to the guarantee of the stable and high-quality development of the aviation industry. Currently, the main stream of aero-engine predictive maintenance methods in the industry can be mainly divided into two main categories, namely a model based on a physical mechanism and a deep learning model based on data. However, in carrying out the present invention, the inventors have found that the above prior art has at least the following challenges in engineering applications: On the one hand, although the method based on the physical mechanism, such as a finite element analysis model, can describe the running process of the engine theoretically, in order to achieve the calculability, the physical processes such as a complex flow field, a temperature field, a stress field and the like in the engine are often required to be excessively simplified, so that a difference exists between a model prediction result and the actual running condition of the engine, and the prediction precision is limited. On the other hand, deep learning models represented by Long Short-Term Memory (LSTM), transformer, and the like, by virtue of their strong nonlinear data fitting capability, show great potential in processing engine sensor monitoring data. However, the high-precision predictive performance of such models often depends on large-scale, high-quality sensor monitoring data. In the field of aeronautics, obtaining high-fidelity data covering the full life cycle and diversified failure modes of an engine generally requires a large number of and expensive bench tests, which results in a sharp contradiction between the data requirements of high-precision models and the high data acquisition costs. The contradiction makes model training insufficient and generalization capability insufficient due to data scarcity in many practical application scenes, and an ideal prediction effect is difficult to achieve. To alleviate the problem of data scarcity, physical information neural networks (Physics-Informed Neural Networks, PINN) and their derivatives have been proposed in the academia. Such methods attempt to incorporate Partial Differential Equations (PDEs) describing the physical process as constraint terms into the loss function of the neural network, thereby incorporating physical prior knowledge into the model training. However, the existing physical information model faces a bottleneck in engineering application, on one hand, the constraint solving process based on partial differential equation has extremely high computational complexity, which results in long model training period and poor instantaneity, and is difficult to meet the requirements of rapid deployment and iteration in engineering practice, and on the other hand, the traditional data enhancement technology, such as generation of a countermeasure network (GAN), is difficult to ensure physical consistency of generated samples when high-dimensional time sequence data is generated, and artifacts are possibly introduced, so that model performance is damaged. Therefore, how to effectively utilize the physical operation mechanism of the engine under the limited data condition and improve the accuracy and generalization capability of the prediction model in a computationally efficient manner is a technical problem to be solved in the field of predictive maintenance of the current aeroengine. Disclosure of Invention In order to solve the above problems in the prior art, that is, the problem that the existing physical information model has high computational complexity and poor data quality, the first aspect of the present application proposes an aeroengine state prediction model construction method based on physical constraints, the method comprising the following steps: Acquiring historical operation data representing the operation state of the aeroengine and an engineering simplified physical constraint rule set representing the physical operation mechanism of the aeroengine; a deep learning model is adopted, and according to the historical operation data, the performance parameters of the aeroengine are predicted, so that predicted performance parameters are obtained; constructing a total loss function, and training a deep learning model by utilizing the total loss function to optimize internal parameters of the deep learning model so as to obtain th