Aero-engine strong transient on-board adaptive modeling method

Abstract

The application relates to the technical field of aero-engine modeling and health monitoring, in particular to an aero-engine strong transient on-board adaptive modeling method. The method comprises the following steps: constructing an LPV mechanism model and obtaining simulation values of engine gas path parameters; constructing a data set, obtaining a deep error compensation model, and correcting the LPV mechanism model to obtain an on-board hybrid model. The application improves the prediction and error compensation precision of the model, considers the physical explanation and generalization ability, and provides an effective technical scheme for aero-engine steady state and transition state working condition gas path health monitoring.

Description

Technical Field

[0001] This invention relates to the field of aero-engine modeling and health monitoring technology, specifically to an airborne adaptive modeling method for aero-engines with strong transients. Background Technology

[0002] Aero-engines are the core power units of aviation equipment, belonging to complex nonlinear thermodynamic systems that integrate aerodynamics, engineering thermodynamics, combustion science, structural mechanics, and control theory. In the aero-engine lifecycle management system, high-precision modeling, digital twins, and gas path condition monitoring technologies are the core technical support for conducting engine control law design verification, performance optimization, fault diagnosis, and condition-based maintenance (CBM), and have significant engineering application value.

[0003] Currently, aero-engine modeling methods are mainly divided into two categories: physical mechanism-based modeling methods (white-box models) and data-driven modeling methods (black-box models). Physical mechanism-based methods typically employ component-level models (0-dimensional or 1-dimensional aero-thermodynamic models). This method simulates the joint working process of various engine components by solving the mass, momentum, and energy conservation equations, and features clear physical meaning and interpretable model structure.

[0004] However, in practical engineering applications, physical models face the challenge of balancing real-time performance and accuracy. First, to ensure real-time computation, physical models often introduce numerous simplifying assumptions, such as neglecting the spatial distribution effects of the flow field and simplifying the combustion efficiency characteristics of the combustion chamber. These assumptions prevent the model from reflecting the true characteristics of the engine. Second, during long-term service, engine components undergo nonlinear degradation due to factors such as blade fouling, corrosion, and wear, causing the "nominal model" to fail to accurately describe the true performance throughout its "life cycle." These factors collectively lead to steady-state deviations, dynamic lags, and transient errors between the simulated values ​​of the physical model's airflow parameters and the measured values ​​of the engine's airflow parameters.

[0005] To balance real-time computation with the ability to describe nonlinear operating characteristics, linear parameter-variable (LPV) models based on the fusion of local linear models under multiple operating conditions are applied to the dynamic modeling of aero-engines across a wide range of operating conditions. This type of method typically establishes local linear state-space models around a series of selected operating points and selects, interpolates, or weights and fuses these local models based on engine speed or other scheduling parameters.

[0006] However, existing LPV modeling methods still have certain limitations. On the one hand, the accuracy of local linear models is affected by factors such as the distribution of operating points, the coverage of identified data, the selection of state variables, and linear approximation errors, making it difficult to maintain high accuracy over a wide range of operating conditions. On the other hand, when there is a lack of smooth fusion between different local models, or when the steady-state reference quantities of each local model are not updated synchronously with the scheduling parameters, the model output may show discontinuous changes near the boundaries of adjacent operating conditions. In addition, during engine acceleration and deceleration, dynamic effects and unmodeled errors are more significant, easily causing errors in the amplitude and phase of air path parameters such as temperature, pressure, and speed.

[0007] To overcome the accuracy bottleneck of physical models, researchers began exploring data-driven error compensation techniques, which utilize machine learning algorithms to predict and correct the residuals of physical models. Early compensation methods often employed multinomial regression, support vector machines (SVR), or shallow artificial neural networks (ANN). While these methods can achieve some correction under single steady-state conditions (such as cruise), they reveal the following serious limitations when dealing with the multi-condition and highly transient nature of aero-engines: First, it lacks the ability to adaptively perceive complex operating states. The operating conditions of aero-engines are extremely complex, encompassing various modes such as steady-state, quasi-steady-state, rapid acceleration, and rapid deceleration. The physical mechanisms and statistical distribution characteristics of errors are completely different in different modes. For example, during rapid acceleration, errors mainly originate from response lag caused by the dynamic effects of components; while during steady-state cruise, errors stem more from differences in sensor noise and measurement distribution. Existing traditional methods often ignore this "state heterogeneity," simply mixing data from all operating conditions together to train a single model. This leads to the model attempting to fit all patterns with a single set of parameters, ultimately resulting in a "mistakenly focusing on one aspect while neglecting another," and extremely poor prediction accuracy at transition state boundaries.

[0008] Second, the ability to extract temporal features and capture long-range dependencies is insufficient. An engine is a typical thermally inertial system. While changes in fuel flow are instantaneous, the heat absorption and release processes of metal components such as the turbine disc and casing require a considerable amount of time. This results in significant lag and integral effects in the response of parameters such as exhaust temperature. Traditional regression models or simple feedforward neural networks typically only focus on the input within the current moment or a very short time window, lacking the ability to remember long-term historical information and thus failing to effectively model this "long-short-term dependency" determined by thermodynamic principles. This often leads to phase lag or amplitude deviation when the model predicts temperature peaks during dynamic processes.

[0009] Third, the key feature selection capability and noise robustness are insufficient. Deep learning models are often considered "black boxes," and traditional neural networks cannot intuitively reflect the contribution of different input parameters and historical time points to the residual prediction results. Meanwhile, the measured values ​​of engine air path parameters are affected by a combination of environmental conditions, sensor noise, and operating conditions. If existing methods lack an effective feature weight allocation mechanism, non-critical features and local measurement disturbances can easily participate in residual prediction, reducing the stability and generalization ability of the model output. Therefore, existing error compensation methods still need to improve their adaptive selection capability for key time steps and key parameter channels, and reduce the impact of non-critical features and measurement disturbances on the residual prediction results.

[0010] Furthermore, in the process of airflow health monitoring of aero-engines based on Kalman observers, the observation residuals between measured values ​​and model estimates of engine airflow parameters are typically used to recursively estimate the flow and efficiency health factors of components such as the compressor and turbine. When the mechanistic model itself has large steady-state or transient state errors, these errors can easily overlap with residuals caused by component performance degradation, thus affecting the accuracy and stability of airflow health factor estimation. Therefore, improving the accuracy of airflow parameter estimation by airborne models under wide operating conditions and transient processes is a crucial foundation for conducting online airflow health monitoring.

[0011] Therefore, there is a need to provide an adaptive airborne modeling method for strong transients in aero-engines. Summary of the Invention

[0012] To address the problems of low error compensation accuracy, weak ability to capture transient dynamic effects, and poor model generalization ability in existing aero-engine physical models, this invention provides an airborne adaptive modeling method for aero-engines with strong transients. This method characterizes the engine's dynamic characteristics across a wide operating range using a continuously scheduled LPV (Limited-Vehicle Phase) mechanism model. It combines this with a deep error compensation model that possesses operational status awareness, joint extraction of long and short time-series features, and adaptive weighting of key features to perform online correction of the LPV mechanism model output. This improves the online estimation accuracy and model output continuity of engine gas path parameters across a wide operating range and during transient processes, reduces the interference of mechanism model errors on observation residuals, and performs online estimation of gas path health factors based on a Kalman observer.

[0013] This invention provides a method for adaptive airborne modeling of aero-engines under strong transient conditions, employing the following technical solution: A linear state deviation benchmark model of an aero-engine at different operating conditions is constructed, and a Gaussian membership degree weighted fusion algorithm is used to weight and fuse the linear state deviation benchmark models at different operating conditions to obtain a continuously adjustable LPV mechanism model from idle condition to maximum condition. The LPV mechanism model is used to obtain the simulation values ​​of engine gas path parameters. A hybrid feature vector is constructed based on the measured values ​​of engine air path parameters, the simulated values ​​of engine air path parameters, and the transient change characteristics of fuel flow at each benchmark operating point under different operating conditions of the aero-engine. A Gaussian mixture clustering model is constructed, and the probability density of the mixture feature vectors is modeled using the Gaussian mixture clustering model. The Gaussian component parameters of the Gaussian mixture clustering model are iteratively solved using the expectation-maximization algorithm until the log-likelihood function converges to obtain the target Gaussian mixture clustering model. The mixture feature vector at each time step is used as a sample to construct a sample set. The sample set is divided into acceleration state sub-sample set, deceleration state sub-sample set, and steady state sub-sample set according to the direction of the rate of change of fuel flow. The state label corresponding to the sample in each sub-sample set is obtained based on the posterior probability of the Gaussian component parameters of the target Gaussian mixture clustering model. The mixture feature vector is used as the input vector, and the state label corresponding to the mixture feature vector and the engine gas path parameter residual are used as the output vector to construct the dataset. A multi-scale spatiotemporal attention residual network is constructed to perform convolution operations on the mixed feature vectors to capture the local transient waveform features of the mixed feature vectors in the high-frequency domain. It also captures long-term time-dependent features in the mixed feature vectors caused by the dynamic effects of the aero-engine, and weights and fuses the spatial features and long-term time-series features to obtain fused features. Based on the fused features and combined with a multi-head self-attention mechanism, attention features are obtained. Residual predictions of engine gas path parameters are obtained based on the attention feature mapping. A loss function is constructed based on the mean square error function, the mean absolute error function, and the L2 norm of the model parameters. The multi-scale spatiotemporal attention residual network is trained on the dataset until the loss function converges to obtain a deep error compensation model. The real-time air path parameters of the engine are input into the trained deep error compensation model to obtain the predicted value of the engine air path parameter residual at the current moment. Based on the predicted value of the engine air path parameter residual, the LPV mechanism model is corrected in real time to obtain the airborne hybrid model.

[0014] A further technical solution of the present invention is that the steps for constructing a linear state deviation benchmark model of an aero-engine at different operating conditions are as follows: Acquire dynamic simulation data of aero-engine component-level nonlinear models under different operating conditions; Using each steady-state operating point as a benchmark, the dynamic simulation data is converted into input deviation, state deviation and output deviation; Based on the input deviation, state deviation and output deviation, the least squares method is used to identify the state space matrix at different working points and establish a linear state deviation benchmark model for each working point. The dynamic simulation data includes input parameters, state parameters, and output parameters. The input parameters include fuel flow rate, compressor flow rate health factor, compressor efficiency health factor, turbine flow rate health factor, and turbine efficiency health factor. The state parameters include high-pressure rotor speed. The output parameters include high-pressure rotor speed, compressor outlet pressure, compressor outlet temperature, turbine outlet pressure, and turbine outlet temperature.

[0015] A further technical solution of the present invention is that the transient change characteristic of fuel flow rate is the difference between the fuel flow rate at the current moment and the fuel flow rate at the previous moment under the same reference operating condition.

[0016] A further technical solution of the present invention is to use the difference between the measured value of the engine air path parameter and the simulated value of the engine air path parameter as the engine air path parameter residual.

[0017] A further technical solution of the present invention is that the multi-scale spatiotemporal attention residual network includes: The input layer is used to input the mixed feature vector; The spatial feature extraction layer is used to perform convolution operations on the mixed feature vector to capture the local transient waveform features of the mixed feature vector in the high-frequency domain. The temporal feature extraction layer is used to capture long-term temporal features in the mixed feature vector that are subject to long-range time dependence caused by dynamic effects. An adaptive feature fusion layer is used to weightedly fuse spatial features and long-term temporal features to obtain fused features; A multi-head self-attention mechanism layer is used to obtain attention features based on fused features and combined with the multi-head self-attention mechanism. And a residual regression output layer, which is used to obtain the predicted values ​​of the engine air path parameters residuals based on attention feature mapping, and output the predicted values ​​of the engine air path parameters residuals.

[0018] A further technical solution of the present invention is to use a one-dimensional causal dilated convolutional neural network to perform convolution operations on the mixed feature vectors. The expression of the one-dimensional causal dilated convolutional neural network is as follows:

[0019] In the formula, It is a local transient waveform characteristic; The kernel size; Void ratio; This is the weight matrix; It is the bias vector; It is a non-linear activation function; For network layer indexing; For time step index; This is the index of the convolution kernel position.

[0020] A further technical solution of the present invention is that the expression of the loss function is:

[0021] In the formula, The value of the loss function; To utilize The residual of engine air path parameters obtained by comparing the measured values ​​of engine air path parameters with the simulated values ​​of engine air path parameters at any given time; For multi-scale spatiotemporal attention residual networks in The predicted residual values ​​of the engine air path parameters output at time t; Model parameters for multi-scale spatiotemporal attention residual networks The L2 norm; These are the weighting coefficients of the mean squared error function; These are the weighting coefficients for the mean absolute error function; These are the weighting coefficients of the L2 norm of the model parameters.

[0022] A further technical solution of the present invention is that, during the training of a multi-scale spatiotemporal attention residual network, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to the network parameters, and the AdamW optimization algorithm combined with the cosine annealing learning rate scheduling strategy is used to update the network weights until the loss function converges to obtain a deep error compensation model with state awareness.

[0023] A further technical solution of the present invention is that the engine air circuit parameters include: compressor outlet pressure, turbine outlet pressure, compressor outlet temperature, turbine outlet temperature, and high-pressure rotor speed.

[0024] A further technical solution of the present invention is that the step of obtaining an airborne hybrid model by real-time correction of the LPV mechanism model based on the predicted residual values ​​of engine air path parameters is as follows: The estimated values ​​of engine air path parameters are obtained by correcting the simulated values ​​of engine air path parameters output by the LPV mechanism model using the predicted residual values ​​of engine air path parameters. The difference between the measured values ​​and the estimated values ​​of the engine air path parameters is used as the observation residual after compensation. The compensated observation residuals are input into the Kalman observer for recursive estimation to obtain the real-time gas path health factor. The real-time gas path health factor is then fed back to the LPV mechanism model to correct the LPV mechanism model and obtain the airborne hybrid model.

[0025] The beneficial effects of this invention are: 1. This invention constructs a linear state deviation benchmark model corresponding to different operating conditions of an aero-engine. Simultaneously, using the high-pressure rotor speed as a scheduling parameter, it employs a Gaussian membership function to smoothly weight and fuse each local linear state deviation benchmark model. Furthermore, it updates the operating condition weights and steady-state benchmark quantities with weights to obtain a continuously adjustable LPV mechanism model ranging from idle to maximum operating conditions. This achieves continuous and stable scheduling of model parameters for different operating conditions, reducing the risk of output jumps during operating condition switching in traditional piecewise linear models. Secondly, this invention adopts a data-driven and physical mechanism fusion method. Based on a Gaussian mixture clustering model, it adaptively partitions the operating data of the aero-engine across operating conditions, effectively solving the problem of inconsistent error distribution of a single model under multiple operating conditions. Then, it constructs a multi-scale spatiotemporal attention residual network, performing joint abstraction of spatial features, temporal features, and multi-head attention features on the mixed feature vectors to achieve adaptive prediction of the nonlinear residuals of the LPV mechanism model. Meanwhile, by constraining state consistency, suppressing non-physical fluctuations, and model complexity, the robustness, generalization ability, and airborne deployment adaptability of the deep error compensation model are improved. Finally, this invention further integrates the LPV mechanism model, which is continuously scheduled from slow to maximum operating conditions, with the data-driven deep error compensation model to form an airborne adaptive model that complements the advantages of the mechanism model and the data-driven model.

[0026] 2. The LPV (Limited-Pipe Valve) mechanism model is used to describe the mechanistic mapping relationship between engine control inputs, air path health factors, and measurable engine parameters. A data-driven deep error compensation model is used to correct the error between the LPV mechanism model and the actual engine, thereby improving the accuracy and reliability of the model output and avoiding the propagation of modeling errors to health factor estimation. Based on this, a Kalman observer is introduced to construct a closed-loop air path health monitoring system. The residual between the measured values ​​of engine air path parameters and the estimated values ​​of engine air path parameters corrected by the airborne hybrid model is used to recursively correct the estimated values ​​of air path health factors. The corrected health factors are then fed back to the LPV mechanism model to achieve closed-loop adaptive estimation of air path health factors, thereby improving the dynamic tracking capability, estimation accuracy, and anti-interference capability of the aero-engine air path health monitoring system. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1This is a flowchart illustrating an airborne adaptive modeling method for strong transients in aero-engines according to the present invention. Figure 2 This is a schematic diagram of step S1 of the present invention, which involves constructing the LPV mechanism model. Figure 3 This is a schematic diagram of the process for constructing a depth error compensation model in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the working condition adaptive clustering principle based on Gaussian mixture model (GMM) in an embodiment of the present invention; Figure 5 This is an architecture diagram of the multi-scale spatiotemporal attention residual network in an embodiment of the present invention; Figure 6 This is a schematic diagram of the gas path health monitoring system in an embodiment of the present invention. Detailed Implementation

[0029] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] An embodiment of the present invention provides a method for adaptive airborne modeling of aero-engines under strong transient conditions, such as... Figure 1 As shown, it includes: S1. Construct an LPV mechanism model and obtain simulation values ​​of engine airflow parameters; Specifically, a linear state deviation benchmark model of the aero-engine is constructed for different operating conditions. Based on the high-pressure rotor speed scheduling, the linear state deviation benchmark models of each operating condition are weighted and fused using a Gaussian membership degree weighted fusion algorithm to obtain a continuously adjustable LPV mechanism model from the idle condition to the maximum operating condition. The simulated values ​​of the engine gas path parameters are obtained using the LPV mechanism model.

[0031] S11. Construct a linear state deviation benchmark model: For example, in this embodiment, the steps for constructing a linear state deviation benchmark model for an aero-engine at different operating points are as follows: acquiring dynamic simulation data of the aero-engine component-level nonlinear model at different operating points; using the steady-state operating point corresponding to each operating point as a benchmark, converting the dynamic simulation data into input deviation, state deviation, and output deviation; based on the input deviation, state deviation, and output deviation, identifying the state space matrix at each operating point using the least squares method, establishing a linear state deviation benchmark model corresponding to each operating point, thereby forming a set of multi-operating-point linear state deviation models. This linear state deviation benchmark model retains the dynamic mechanism characteristics of the aero-engine component-level nonlinear model while reducing online computational complexity. It should be noted that each operating point is a preset operating point, and the selection of operating points is based on the actual operating state, mission spectrum, speed variation range, fuel flow variation range, and steady-state operating point distribution of the real aero-engine, representing typical engine operating states. Typical operating points include those covered by the engine in idle state, intermediate state, acceleration transition state, deceleration transition state, and near the maximum state. In this embodiment, taking the idle to maximum operating condition range of an aero-engine as an example, multiple steady-state operating points covering the percentage speed of the high-pressure rotor from 36.6% to 99.0% are used as typical operating points. It should be noted that each operating condition in this embodiment is used to define the effective operating condition range for model construction and scheduling, while the dynamic simulation data of the component-level nonlinear model is used to generate the data samples required for model identification within the range from idle to maximum operating condition. Compared with the component-level nonlinear model of an aero-engine, the linear state deviation model in this embodiment does not require iterative solution of the common working equations of components during online calculation, which helps to reduce the complexity of online calculation and improve the real-time calculation capability of the model.

[0032] For example, in one specific embodiment, the nonlinear component level model (NCLM) of an aero-engine is a classic model in the field of aero-engines, and its model structure is existing technology, which will not be described in detail in this embodiment. In this embodiment, the NCLM of the aero-engine has 14 operating points. The percentage speeds at these 14 operating points, from smallest to largest, are 36.6%, 40.8%, 45.8%, 50.6%, 55.6%, 60.6%, 65.4%, 70.4%, 75.2%, 80.2%, 85.2%, 90.0%, 95.0%, and 99.0%. The altitude is based on standard sea level. The NCLM of the aero-engine has 5 input parameters and 5 output parameters. The input parameters include: fuel flow rate. Compressor efficiency and health factors Compressor flow health factors Turbine efficiency health factors and turbine flow health factors Output parameters include high-pressure rotor speed. Compressor outlet pressure Compressor outlet temperature Turbine outlet pressure and turbine outlet temperature It should be noted that compressor efficiency health factors... Compressor flow health factors Turbine efficiency health factors and turbine flow health factors This is not the actual control input of the engine, but rather an equivalent health factor input used to characterize the performance deviations of engine components. It describes the impact of changes in the flow capacity and efficiency of the compressor and turbine components on the engine state and output. Specifically, in this example, the dynamic simulation data includes input parameters, state parameters, and output parameters. The input parameters include fuel flow rate. Compressor flow health factors Compressor efficiency and health factors Turbine flow health factors and turbine efficiency health factors State parameters include high-voltage rotor speed. .

[0033] Specifically, in this embodiment, for the first Using a preset operating point as a benchmark, and taking the corresponding steady-state input, steady-state state, and steady-state output as the reference, the steps for converting dynamic simulation data into deviation data and establishing a linear state deviation model are as follows: Setting compressor efficiency health factors Compressor flow health factors Turbine efficiency health factors and turbine flow health factors The initial values ​​of all parameters are 1, allowing the aero-engine component-level nonlinear model to iterate to a stable state at this operating point, thus obtaining the steady-state reference input corresponding to this operating point. Steady-state reference state and steady-state reference output Dynamic simulation was performed starting from the steady-state reference point. The five input parameters were sequentially perturbed according to the process of "increase by 2%, restore, decrease by 2%, restore," resulting in dynamic simulation time-series data for the input, state, and output parameters. The high-voltage rotor speed was selected. As state variables, the corresponding steady-state reference values ​​are subtracted from the collected dynamic simulation time series data to obtain the deviation time series dataset.

[0034] For the A preset operating point is defined, and the input deviation is defined. for:

[0035] In the formula, Indicates fuel flow deviation; This indicates a deviation in the compressor efficiency health factor. This indicates a deviation in the compressor flow health factor; This indicates a deviation in the turbine efficiency health factor; This indicates the turbine flow health factor deviation; in this embodiment, the health factor deviation is used to characterize the performance deviation of the engine air passage component's flow capacity and efficiency relative to the baseline state.

[0036] Define state deviation for:

[0037] In the formula, Indicates the deviation in high-voltage rotor speed; Define output deviation for:

[0038] In the formula, Indicates the deviation in high-voltage rotor speed; This indicates the compressor outlet pressure deviation; This indicates the deviation of the compressor outlet temperature; Indicates turbine outlet pressure deviation; This represents the turbine outlet temperature deviation; where the high-pressure rotor speed deviation ΔN in the output deviation is directly mapped from the state deviation through the output matrix.

[0039] Then for the first The first preset operating point, the first The discrete linear state deviation model for each preset working point is as follows:

[0040] In the formula, express State deviation at any given time; express Input deviation at any given time; express State deviation at any given time; express Output deviation at any given time; Represents the state matrix; Represents the input matrix; Indicates the output matrix; This represents a directly passed matrix.

[0041] Furthermore, the time-series deviation data obtained from dynamic simulation are subjected to least squares identification to obtain the state matrix, input matrix, and output matrix, and the direct transfer matrix is ​​then used. Following step S11, the 14 operating points are identified sequentially to obtain the linear state deviation models corresponding to the 14 operating points, forming a set of linear state deviation models for multiple operating points.

[0042] S12. Constructing the LPV mechanism model: like Figure 2 As shown, for the continuous scheduling problem of the linear state deviation model corresponding to the 14 operating points obtained in step S11, the high-voltage rotor speed is selected. As scheduling parameters, the scheduling weights corresponding to each working point are calculated using the Gaussian membership function, and the scheduling weights are normalized so that the sum of the weights corresponding to each working point under any scheduling parameter is 1. Based on the normalized scheduling weights, the linear state deviation model matrices of each working point are weighted and fused to obtain the LPV mechanism model for continuous scheduling.

[0043] For example, in one specific embodiment, the high-pressure rotor speed of the aero-engine is selected. As scheduling parameters, a Gaussian membership function is used to calculate the membership degree of the current scheduling parameter relative to each working point, and each membership degree is normalized so that the sum of the weights corresponding to all working points under any scheduling parameter is 1. The Gaussian membership function corresponding to each working point can be expressed as:

[0044] In the formula, Indicates the current high-voltage rotor speed Compared to the first Membership degree of each working condition point Indicates the first The reference speed of the high-voltage rotor corresponding to each operating point To indicate the first The standard deviation of the Gaussian weighted distribution corresponding to each operating point.

[0045] To ensure that the sum of the weights of all operating conditions is 1, the membership degree of each operating condition point is normalized:

[0046] In the formula, Indicates the first The normalized scheduling weights corresponding to each working condition point satisfy the following:

[0047] In one specific embodiment, The speed is determined based on the high-voltage rotor reference speed interval between adjacent operating points; in this embodiment, the speeds at each operating point are in an arithmetic progression relationship. A preset multiple of the average speed interval between two adjacent operating points is taken. Through this embodiment, the Gaussian weights of adjacent operating points can smoothly transition within the speed range, avoiding output discontinuity issues caused by hard switching of discrete models. Then, based on the normalized scheduling weights, the matrices of the linear state deviation models corresponding to the 14 operating points are weighted and fused to obtain the matrix of the LPV mechanism model under the current scheduling parameters:

[0048] In the formula, This represents the state matrix of the LPV mechanism model under scheduling parameters; This represents the input matrix of the LPV mechanism model under scheduling parameters; This represents the output matrix of the LPV mechanism model under scheduling parameters; This represents the direct transfer matrix of the LPV mechanism model under scheduling parameters; Represents the state matrix; Represents the input matrix; Indicates the output matrix; This represents a directly passed matrix.

[0049] Since the models obtained in step S11 are all deviation models established based on the steady-state references of each operating point, in order to ensure that the conversion relationship between actual quantities and deviation quantities remains consistent during the scheduling process, it is also necessary to perform weighted updates on the steady-state reference input, steady-state reference state, and steady-state reference output corresponding to each operating point. Specifically, the current scheduling parameters (in this embodiment, the high-pressure rotor speed of the aero-engine) The steady-state reference input, steady-state reference state, and steady-state reference output (as scheduling parameters) are respectively expressed as:

[0050] In the formula, Indicates the first Steady-state reference input corresponding to each operating point; Indicates the first Steady-state reference state corresponding to each operating point; Indicates the first Steady-state reference output corresponding to each operating point; This represents the steady-state reference input under the current scheduling parameters; This represents the steady-state baseline state under the current scheduling parameters; This represents the steady-state baseline output under the current scheduling parameters. Therefore, the actual input under the current scheduling parameters... Deviation from actual input Actual state Deviation from actual state Actual output Deviation from actual output The relationship between them is:

[0051] The final LPV mechanism model for continuous scheduling from slow condition to maximum condition is as follows:

[0052] In the formula, express The state deviation of the LPV mechanism model at time; express The state deviation of the LPV mechanism model at time; express The output deviation of the LPV mechanism model at time; express The input deviation of the LPV mechanism model at time; , , and All parameters change continuously with the scheduling parameters. Through Gaussian membership weight normalization and steady-state benchmark weighted update, the LPV mechanism model can achieve smooth scheduling between multiple local linear models within the preset speed range corresponding to the actual aero-engine test bench conditions. This improves the continuity of model switching between adjacent test points and reduces the risk of output jump caused by hard switching of discrete test models.

[0053] S2. Construct the dataset; Specifically, a hybrid feature vector is constructed based on the measured values ​​and simulated values ​​of engine gas path parameters under different operating conditions, as well as the transient variation characteristics of fuel flow. A Gaussian mixture clustering model is constructed, and the probability density of the hybrid feature vector is modeled using the Gaussian mixture clustering model. The Gaussian component parameters of the Gaussian mixture clustering model are iteratively solved using the expectation-maximization algorithm until the log-likelihood function converges to obtain the target Gaussian mixture clustering model. The hybrid feature vector at each time step is used as a sample to construct a sample set, which is divided into acceleration state sub-sample set, deceleration state sub-sample set, and steady state sub-sample set according to the direction of the rate of change of fuel flow. The state label corresponding to the sample in each sub-sample set is obtained based on the posterior probability of the Gaussian component parameters of the target Gaussian mixture clustering model. The hybrid feature vector is used as the input vector, and the state label corresponding to the hybrid feature vector and the engine gas path parameter residual are used as the output vector to construct the dataset.

[0054] S21. Constructing a hybrid feature vector: For example, in one specific embodiment, the step of constructing a hybrid feature vector based on the measured values ​​of engine air path parameters, simulated values ​​of engine air path parameters, and transient variation characteristics of fuel flow under different operating conditions of the aero-engine is as follows: The measured values ​​of engine air path parameters are collected through the airborne sensors of the aero-engine, and the simulated values ​​of engine air path parameters output by the LPV mechanism model in step S1 under the corresponding operating conditions are acquired simultaneously; outlier cleaning and time alignment processing are performed on the measured values ​​and simulated values ​​of engine air path parameters, and the fuel flow rate at the current time relative to the previous time is calculated using a differential operator. The difference in fuel flow rate is used as a transient characteristic of fuel flow rate, which represents the engine's dynamic control intent. The difference between the measured and simulated values ​​of engine air path parameters at the same operating point is used as the engine air path parameter residual. Outlier cleaning, time alignment, and standardization are performed on the measured, simulated, and residual values ​​of engine air path parameters, as well as the transient characteristics of fuel flow rate. The engine air path parameter residual is the difference between the measured and simulated values ​​of engine air path parameters at the same moment. A sliding window technique is used to concatenate the measured, simulated, and transient characteristics of engine air path parameters at the current and historical moments in chronological order, and then perform Z-score standardization to obtain a hybrid feature vector containing long-term temporal dependency information. At the same time, the difference between the measured and simulated values ​​of engine air path parameters at the same moment is used as the label for the engine air path parameter residual for subsequent supervised training. It should be noted that in this embodiment, Z-Score standardization is used to eliminate numerical differences between different physical dimensions (such as pressure, temperature, and rotational speed) to facilitate the convergence of the subsequent model. The engine air path parameters include: compressor outlet pressure, turbine outlet pressure, compressor outlet temperature, turbine outlet temperature, and high-pressure rotor speed. In this embodiment, the engine air path parameter measurements obtained by the sensors of the aero-engine are derived from actual engine measurement data. The hybrid feature vector designed in this embodiment not only includes the thermodynamic state at the current moment but also introduces the dynamic change trend of the control quantity, thereby effectively solving the prediction lag problem caused by ignoring the rate of change of control commands in traditional methods.

[0055] S22. Construct a Gaussian mixture clustering model: For example, in one specific embodiment, to address the problems of complex operating conditions and inconsistent error distribution during aero-engine operation, this embodiment proposes unsupervised state clustering based on a Gaussian Mixture Model (GMM). This state clustering step, based on the mixed feature vector constructed in step S21, aims to automatically identify and discover potential data distribution patterns under different operating phases of the engine (such as steady state, rapid acceleration, and rapid deceleration). By introducing probability density modeling, GMM can describe the posterior probability of a data point belonging to a specific operating condition cluster in a soft clustering manner, thereby avoiding the boundary effects caused by traditional hard threshold-based partitioning methods. Specifically, in this embodiment, the construction process of the Gaussian mixture clustering model includes: First, constructing a sample set by taking the mixture feature vector at each time step as a sample, and initially dividing the sample set into an acceleration state sub-sample set, a deceleration state sub-sample set, and a steady state sub-sample set according to the direction of the rate of change of fuel flow (positive, negative, or equal to zero); then, for each macroscopic sub-sample set, setting the number of clusters (for example, setting 3 clusters for the steady state, and 9 clusters for both the acceleration and deceleration states), and initializing the mean, covariance matrix, and mixing coefficients of the Gaussian components; then, using the expectation-maximization (EM) algorithm for iterative optimization until the log-likelihood function converges; finally, calculating the response of each sample to each Gaussian component parameter, and taking the state of the sample set corresponding to the Gaussian component parameter with the maximum response as the state label of that sample.

[0056] In this embodiment, the probability density function expression of the Gaussian mixture clustering model is:

[0057] In the formula, For mixed feature vectors The probability density; The number of clusters; For the first The mixing coefficient of the Gaussian components, For the first The multivariate Gaussian distribution density function of each component, where It is the mean vector. Let be the covariance matrix.

[0058] To accurately describe the distribution of engine multi-dimensional sensor data in the feature space, the first... Multivariate Gaussian distribution density function of Gaussian component parameters The specific expression is:

[0059] In the formula, Indicates the length of the mixed feature vector; The determinant of the covariance matrix; The inverse matrix of the covariance matrix; This represents the transpose of a vector. The multivariate Gaussian distribution density function quantifies the sample... and Cluster center The Mahalanobis distance can accommodate strong coupling correlations between different sensor parameters (e.g., the correlation between rotational speed and pressure). Furthermore, during the model training phase, the Expectation-Maximization (EM) algorithm is used for iterative parameter estimation; the E-step of the Expectation-Maximization (EM) algorithm calculates the... Sample Belongs to the The posterior probability of each Gaussian component, i.e., the posterior probability in this embodiment, is used as the response degree. Response The calculation formula is:

[0060] In the formula, This represents the responsiveness, i.e., the response level within a given range of features composed of mixed feature vectors. Sample data Given the current Gaussian mixture model parameters, this data is generated by the first... The posterior probability is generated by a Gaussian component. This probability value embodies the idea of ​​"soft clustering," that is, a working point can belong to multiple state clusters with different probabilities at the same time. This has important physical significance for describing the gradual transition process of an engine from steady state to transient state.

[0061] In the M-step of the Expectation Maximization (EM) algorithm, based on the calculated response degree... Then, re-estimate the parameters of each Gaussian component. The update formula for the Gaussian component parameters is as follows:

[0062]

[0063]

[0064] In the formula, The total number of samples; For the first The number of valid samples in each cluster; This is the updated mean vector; This is the updated covariance matrix; The updated mixing coefficients are used. By alternately executing the Expectation Maximization (EM) algorithm in E and M steps until the log-likelihood function converges, the optimal Gaussian mixture clustering model is obtained. The optimal Gaussian mixture clustering model is used as the target Gaussian mixture clustering model in this embodiment.

[0065] Specifically, by iteratively solving the expectation-maximization (EM) algorithm, soft clustering for complex working conditions is achieved. That is, in this embodiment, the state label corresponding to the sample in each sub-sample set is obtained based on the posterior probability of the Gaussian component parameters of the target Gaussian mixture clustering model. For example... Figure 4 The diagram shown is a schematic representation of the principle of GMM state perception in this embodiment. Figure 4 The horizontal axis represents the rate of change of fuel flow, and the vertical axis represents the characteristic dimension of the engine's high-pressure rotor speed. It can be seen that the data points are automatically clustered into three Gaussian distribution regions: the red region represents the acceleration condition cluster (high fuel flow rate and high engine speed), the blue region represents the deceleration condition cluster, and the green region represents the stable condition cluster. Each cluster is surrounded by elliptical contour lines, representing the covariance range of the Gaussian distribution. This embodiment utilizes this probability density distribution to accurately distinguish transitional data with ambiguous boundaries.

[0066] It should be noted that the GMM obtained in step S22 is a state label, which is only used as a state prior or a basis for hierarchical training; that is, in this embodiment, the mixed feature vector is used as the input vector, and the state label corresponding to the mixed feature vector and the engine air path parameter residual are used as the output vector to construct the dataset.

[0067] S3. Obtain the depth error compensation model; Specifically, a multi-scale spatiotemporal attention residual network is constructed. This network is used to perform convolution operations on the mixed feature vectors to capture the local transient waveform features of the mixed feature vectors in the high-frequency domain. It also captures the long-term time-dependent long-series features in the mixed feature vectors caused by the dynamic effects of the aero-engine. The spatial features and long-series features are weighted and fused to obtain fused features. Based on the fused features and combined with a multi-head self-attention mechanism, attention features are obtained. Based on the attention feature mapping, the residual prediction values ​​of the engine gas path parameters are obtained. A loss function is constructed based on the mean square error function, the mean absolute error function, and the L2 norm of the model parameters. The multi-scale spatiotemporal attention residual network is trained on the dataset until the loss function converges to obtain a deep error compensation model.

[0068] S31. Construct a multi-scale spatiotemporal attention residual network; For example, in one specific embodiment, the multi-scale spatiotemporal attention residual network includes an input layer, a spatial feature extraction layer, a temporal feature extraction layer, an adaptive feature fusion layer, a multi-head self-attention mechanism layer, and a residual regression output layer.

[0069] It should be noted that the multi-scale spatiotemporal attention residual network in this embodiment corresponds to the MST-Attention-ResNet in the attached figure, but is not limited to a fixed network topology. Instead, it adopts a modular spatiotemporal representation framework for dynamic residual compensation of aero-engines. This framework uses the output of the physical mechanism model as a priori and the operating condition label as the state constraint. Through the coordinated expression of spatial features, temporal features, and multi-head attention features, it enhances the model's adaptability to changes in residual distribution at transition states, steady states, and operating condition switching boundaries. Figure 3 In this system, GMM is a Gaussian mixture model, 1D-CNN is a one-dimensional convolutional neural network, Bi-LSTM is a bidirectional long short-term memory network, and MST-Attention-ResNet is a multi-scale spatiotemporal attention residual network. Among them, GMM is used for state-aware clustering, 1D-CNN is used for local transient feature extraction, Bi-LSTM is used for long temporal feature extraction, and MST-Attention-ResNet is used to fuse spatial features, temporal features, and attention features and output the predicted residual values ​​of engine gas path parameters.

[0070] The spatial feature extraction layer is used to perform convolution operations on the mixed feature vectors to capture the local transient waveform features of the mixed feature vectors in the high-frequency domain. Specifically, in this embodiment, the spatial feature extraction layer is used to extract short-term dynamic changes in the mixed feature vectors under causal constraints. Causal constraints ensure that the feature abstraction at the current moment depends only on the input information at the current moment and previous moments, avoiding the use of future information in the airborne real-time prediction process. The spatial feature extraction layer can adopt multi-scale temporal filtering, causal dilated convolutional neural networks, local window coding, or equivalent lightweight temporal feature extraction structures. Its scale parameters, receptive field range, and channel mapping relationship can be adaptively configured according to the engine operating conditions and residual distribution characteristics. In this embodiment, a one-dimensional causal dilated convolutional neural network is used to perform convolution operations on the mixed feature vectors. The expression of the one-dimensional causal dilated convolutional neural network is:

[0071] In the formula, It is a local transient waveform characteristic; The kernel size; Void ratio; This is the weight matrix; It is the bias vector; It is a non-linear activation function; For network layer indexing; For time step index; This is the index of the convolution kernel position.

[0072] The spatial feature extraction layer is used to perform convolution operations on the mixed feature vectors to capture the local transient waveform features of the mixed feature vectors in the high-frequency domain. The spatial feature extraction layer employs a temporal modeling structure with gated memory or long context encoding capabilities to preserve the gradual changes in parameters such as pressure, temperature, and speed over multiple sampling periods after fuel disturbance. Compared to compensation methods that only utilize features at the current moment, this layer improves the deep error compensation model's ability to characterize temperature lag, transition state peak shift, and residual phase deviation.

[0073] The adaptive feature fusion layer is used to weightedly fuse spatial features and long-term temporal features to obtain fused features. The multi-head self-attention mechanism layer is used to obtain attention features based on the fused features and combined with the multi-head self-attention mechanism. The multi-head self-attention mechanism adaptively recalibrates the fused features according to the contribution of different time steps, different parameter channels, and different state clusters to residual prediction, thereby highlighting historical segments and key gas path parameters that play a dominant role in current residual compensation. This attention recalibration process can be implemented using multi-branch attention, gated weight allocation, state-conditional weight generation, or equivalent feature recalibration methods; the specific weight calculation form and attention head configuration are not limited.

[0074] It should be noted that, Figure 5 The hybrid feature vector is the feature vector input to the multi-scale spatiotemporal attention residual network; spatial feature extraction is used to extract local transient waveform features from the hybrid feature vector; temporal feature extraction is used to extract long temporal dependent features caused by the dynamic effects of aero-engines; and the multi-head self-attention mechanism is used to perform adaptive weight allocation and feature recalibration for spatial and temporal features. Figure 5 In This indicates the fusion and convergence of multiple feature streams; Figure 5 Below This indicates the predicted value of the turbine outlet pressure residual. This indicates the predicted value of the turbine outlet temperature residual. This indicates the predicted residual value of the high-voltage rotor speed. This indicates the predicted value of the compressor outlet temperature residual. This represents the predicted value of the compressor outlet pressure residual.

[0075] The residual regression output layer is used to generate predicted residual values ​​for engine airflow parameters based on attention characteristics. To maintain the physical continuity of the mechanistic model output, the residual regression output layer can be combined with residual connectivity, normalization constraints, amplitude boundary constraints, or non-physical fluctuation suppression constraints. This allows the predicted residual values ​​to improve compensation accuracy while avoiding abrupt changes that do not conform to the dynamic laws of the engine.

[0076] S32. Train the multi-scale spatiotemporal attention residual network to obtain a trained deep error compensation model: For example, in one specific embodiment, during the training phase, the residual engine airflow parameters at the current prediction time are used as supervision labels, and a hybrid feature vector and state labels are used as input information to train a multi-scale spatiotemporal attention residual network. The loss function during training is:

[0077] In the formula, The value of the loss function; To utilize The residual of engine air path parameters obtained by comparing the measured values ​​of engine air path parameters with the simulated values ​​of engine air path parameters at any given time; For multi-scale spatiotemporal attention residual networks in The residual predicted value of the engine air path parameters output at time t; Model parameters for multi-scale spatiotemporal attention residual networks The L2 norm; The weighting coefficients of the mean squared error function range from 0.3 to 0.8. The weighting coefficients for the mean absolute error function range from 0.2 to 0.7. These are the weighting coefficients for the L2 norm of the model parameters, with values ​​ranging from 10. -6 ~10 -3 In a preferred embodiment, , , .

[0078] For example, in one specific embodiment, during the training of a multi-scale spatiotemporal attention residual network, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to the network parameters, and the AdamW optimization algorithm combined with the cosine annealing learning rate scheduling strategy is used to update the network weights until the loss function converges to obtain a deep error compensation model with state awareness.

[0079] At this point, the multi-scale spatiotemporal attention residual network can be iteratively trained until the preset convergence condition is met, resulting in a well-trained deep error compensation model.

[0080] S4. The LPV mechanism model is modified to obtain the airborne hybrid model; Specifically, the real-time air path parameters of the engine are input into the trained deep error compensation model to obtain the predicted residual values ​​of the engine air path parameters at the current moment. Based on the predicted residual values ​​of the engine air path parameters, the LPV mechanism model is corrected in real time to obtain the airborne hybrid model.

[0081] For example, in one specific embodiment, the measured air path parameters collected from the aero-engine, after preprocessing and status label identification in step S2, are input into the deep error compensation model trained in step S3. The deep error compensation model outputs the residual prediction value of the simulated engine air path parameters output by the LPV mechanism model at the current moment. Finally, the engine air path parameters output by the LPV mechanism model are corrected in real time based on the residual prediction value of the engine air path parameters to obtain the airborne hybrid model. This process realizes the online fusion of "physical model + data-driven compensation" and improves the credibility of the model simulation output.

[0082] This embodiment also includes the step of obtaining an airborne hybrid model by real-time correction of the LPV mechanism model based on the predicted residual values ​​of engine air path parameters: The predicted residual values ​​of engine air path parameters obtained in step S4 of this invention are used to correct the simulated engine air path parameters output by the LPV mechanism model to obtain estimated engine air path parameters; the difference between the measured engine air path parameters and the estimated engine air path parameters is used as the compensated observation residual; the compensated observation residual is input into a Kalman observer for recursive estimation to obtain a real-time air path health factor, and the real-time air path health factor is fed back to the LPV mechanism model to form an airborne hybrid model, such as... Figure 6 As shown, the airborne hybrid model and the Kalman observer recursion form a closed-loop gas path health monitoring system. It should be noted that... Figure 6 The solid blue line represents the physical signal transmission path of the LPV mechanism model, the dashed red line represents the feature input and residual compensation path of the depth error compensation model, the area enclosed by the dashed blue line is the airborne hybrid model established in this embodiment, and KF represents the Kalman observer. Further, Figure 6 The simulated values ​​of the air path parameters are corrected by the predicted values ​​of the engine air path parameter residuals to obtain the estimated values ​​of the engine air path parameters.

[0083] In this embodiment, the Kalman observer constructs an augmented state vector using the high-pressure rotor speed deviation and four types of airway health factors. These four airway health factors include compressor flow rate health factor deviation, compressor efficiency health factor deviation, turbine flow rate health factor deviation, and turbine efficiency health factor deviation, which characterize the performance deviations in the flow capacity and efficiency of engine airway components as they change with service conditions. It should be noted that during the identification phase of the LPV mechanism model, the airway health factors serve as equivalent inputs characterizing component performance deviations in the model identification process. During the airborne health monitoring phase, the airway health factors are used as slowly varying state variables to be estimated by the Kalman observer and are recursively updated using observation residuals before being fed back to the LPV mechanism model.

[0084] In this embodiment, the four types of gas pathway health factors are determined using a state recursion method combining random walk and bounded process noise. The health factor deviation vector is defined as follows:

[0085] In the formula, express Health factor deviation vector at time point; express The deviation of the compressor flow health factor at any given moment; express The deviation of compressor efficiency health factors at any given time; express Deviation of turbine flow health factor at any given time; express Deviation in turbine efficiency health factor at any given moment; Incorporating the health factor deviation vector into the LPV mechanism model, the augmented state vector, including high-pressure rotor speed deviation, compressor flow rate health factor deviation, compressor efficiency health factor deviation, turbine flow rate health factor deviation, and turbine efficiency health factor deviation, can be expressed as:

[0086] In the formula, express The augmented state vector at time step; express State deviation at any given time; express Health factor deviation vector at time point.

[0087] Based on the augmented state vector, augmented state prediction equations and observation equations incorporating health factor states are established. Through Kalman prediction, observation residual calculation, Kalman gain update, and state correction steps, the estimated values ​​of gas path health factors are recursively obtained as follows:

[0088]

[0089]

[0090] In the formula, Kalman gain; This represents a state matrix that combines health factor states; Represents the input matrix that incorporates health factor states; This represents the output matrix that incorporates health factor states; This represents the direct transfer matrix that incorporates health factor states; According to the first The information at time 1 is used to predict the first The predicted value of the augmented state vector at time step; According to the first The information correction at time 1 obtained The estimated value of the augmented state vector at time step; This represents the predicted residual values ​​of the gas path parameters output by the error-compensated hybrid model. For the first The estimated values ​​of engine air path parameters after correction by the predicted values ​​of air path parameter residuals at any given time; Indicates the first The fuel flow rate at any time; where the compensated observation residual is the difference between the measured value of the engine parameters and the estimated value of the corrected engine air path parameters; the compensated observation residual is input into the Kalman observer to recursively estimate the real-time air path health factor, and the real-time air path health factor is fed back to the LPV mechanism model to correct the LPV mechanism model to obtain the airborne hybrid model.

[0091] This invention also provides an embodiment of an error compensation hybrid system for aero-engines, comprising: an LPV mechanism model module, a dataset construction module, a deep error compensation model, and an airborne hybrid module; the LPV mechanism model module is used to construct a linear state deviation benchmark model of the aero-engine at different operating conditions, and uses a Gaussian membership weighted fusion algorithm to weightedly fuse the linear state deviation benchmark models at different operating conditions to obtain a continuously adjustable LPV mechanism model within the range from idle to maximum operating conditions, and uses the LPV mechanism model to obtain simulated values ​​of engine gas path parameters; the dataset construction module is used to construct a linear state deviation benchmark model of the aero-engine at different operating conditions based on various benchmarks. A mixed feature vector is constructed using the measured values ​​of engine airflow parameters, simulated values ​​of engine airflow parameters, and transient changes in fuel flow rate at each condition point. A Gaussian mixture clustering model is then constructed, and the probability density of the mixed feature vector is modeled using this model. The Gaussian component parameters of the Gaussian mixture clustering model are iteratively solved using the expectation-maximization algorithm until the log-likelihood function converges to obtain the target Gaussian mixture clustering model. The mixed feature vector at each time point is used as a sample to construct a sample set, which is then divided into acceleration state sub-sample sets, deceleration state sub-sample sets, and steady state sub-sample sets according to the direction of the rate of change of fuel flow rate. Based on the target Gaussian mixture clustering model... The posterior probability of the Gaussian component parameters of the model is used to obtain the state label corresponding to the sample in each subsample set; the mixed feature vector is used as the input vector, and the state label corresponding to the mixed feature vector and the residual of the engine gas path parameter are used as the output vector to construct the dataset; a deep error compensation model is used to construct a multi-scale spatiotemporal attention residual network, which is used to perform convolution operation on the mixed feature vector to capture the local transient waveform features of the mixed feature vector in the high frequency domain; long-term time-dependent long-series features caused by the dynamic effects of the aero-engine are captured in the mixed feature vector, and the spatial features and long-series features are weighted and fused to obtain the fused features. The system employs a multi-head self-attention mechanism to obtain attention features based on fused features. Based on these attention features, it maps the residual values ​​of engine airflow parameters to obtain predicted values. A loss function is constructed using the mean squared error function, mean absolute error function, and the L2 norm of the model parameters. A multi-scale spatiotemporal attention residual network is trained on the dataset until the loss function converges to obtain a deep error compensation model. An airborne hybrid module is used to input the real-time airflow parameters of the engine into the trained deep error compensation model to obtain the predicted residual values ​​of the engine airflow parameters at the current moment. Based on these predicted residual values, the LPV mechanism model is corrected in real time to obtain the airborne hybrid model.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for adaptive airborne modeling of aero-engines under strong transient conditions, characterized in that, include: A linear state deviation benchmark model of an aero-engine at different operating conditions is constructed, and a Gaussian membership degree weighted fusion algorithm is used to weight and fuse the linear state deviation benchmark models at different operating conditions to obtain a continuously adjustable LPV mechanism model from idle condition to maximum condition. The LPV mechanism model is used to obtain the simulation values ​​of engine gas path parameters. A hybrid feature vector is constructed based on the measured values ​​of engine air path parameters, the simulated values ​​of engine air path parameters, and the transient change characteristics of fuel flow at each benchmark operating point under different operating conditions of the aero-engine. A Gaussian mixture clustering model is constructed, and the probability density of the mixture feature vectors is modeled using the Gaussian mixture clustering model. The Gaussian component parameters of the Gaussian mixture clustering model are iteratively solved using the expectation-maximization algorithm until the log-likelihood function converges to obtain the target Gaussian mixture clustering model. The mixture feature vector at each time step is used as a sample to construct a sample set. The sample set is divided into acceleration state sub-sample set, deceleration state sub-sample set, and steady state sub-sample set according to the direction of the rate of change of fuel flow. The state label corresponding to the sample in each sub-sample set is obtained based on the posterior probability of the Gaussian component parameters of the target Gaussian mixture clustering model. The mixture feature vector is used as the input vector, and the state label corresponding to the mixture feature vector and the engine gas path parameter residual are used as the output vector to construct the dataset. A multi-scale spatiotemporal attention residual network is constructed to perform convolution operations on the mixed feature vectors to capture the local transient waveform features of the mixed feature vectors in the high-frequency domain. This method captures long-term time-dependent features caused by the dynamic effects of aero-engines in the mixed feature vector, and weights and fuses spatial features and long-term time-dependent features to obtain fused features. Based on the fused features and combined with a multi-head self-attention mechanism, attention features are obtained. Based on the attention feature mapping, the predicted values ​​of engine gas path parameters residuals are obtained. A loss function is constructed based on the mean square error function, the mean absolute error function, and the L2 norm of the model parameters. The multi-scale spatiotemporal attention residual network is trained on the dataset until the loss function converges to obtain a deep error compensation model. The real-time air path parameters of the engine are input into the trained deep error compensation model to obtain the predicted value of the engine air path parameter residual at the current moment. Based on the predicted value of the engine air path parameter residual, the LPV mechanism model is corrected in real time to obtain the airborne hybrid model.

2. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, The steps for constructing a linear state deviation benchmark model of an aero-engine at different operating conditions are as follows: Acquire dynamic simulation data of aero-engine component-level nonlinear models under different operating conditions; Using the steady-state operating points corresponding to different operating points as a benchmark, the dynamic simulation data is converted into input deviation, state deviation and output deviation. Based on the input deviation, state deviation and output deviation, the least squares method is used to identify the state space matrix at different operating points and establish a linear state deviation benchmark model corresponding to different operating points. The dynamic simulation data includes input parameters, state parameters, and output parameters. The input parameters include fuel flow rate, compressor flow rate health factor, compressor efficiency health factor, turbine flow rate health factor, and turbine efficiency health factor. The state parameters include high-pressure rotor speed. The output parameters include high-pressure rotor speed, compressor outlet pressure, compressor outlet temperature, turbine outlet pressure, and turbine outlet temperature.

3. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, The transient change characteristic of fuel flow rate is the difference between the fuel flow rate at the current moment and the fuel flow rate at the previous moment under the same reference operating condition.

4. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, The difference between the measured values ​​and the simulated values ​​of the engine air path parameters is taken as the engine air path parameter residual.

5. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, Multi-scale spatiotemporal attention residual networks include: The input layer is used to input the mixed feature vector; The spatial feature extraction layer is used to perform convolution operations on the mixed feature vector to capture the local transient waveform features of the mixed feature vector in the high-frequency domain. The temporal feature extraction layer is used to capture long-term temporal features in the mixed feature vector that are subject to long-range time dependence caused by dynamic effects. An adaptive feature fusion layer is used to weightedly fuse spatial features and long-term temporal features to obtain fused features; A multi-head self-attention mechanism layer is used to obtain attention features based on fused features and combined with the multi-head self-attention mechanism. And a residual regression output layer, which is used to obtain the predicted values ​​of the engine air path parameters residuals based on attention feature mapping, and output the predicted values ​​of the engine air path parameters residuals.

6. The airborne adaptive modeling method for strong transients in aero-engines according to claim 5, characterized in that, A one-dimensional causal dilated convolutional neural network is used to perform convolution operations on the mixed feature vectors. The expression for the one-dimensional causal dilated convolutional neural network is as follows: In the formula, It is a local transient waveform characteristic; The kernel size; Void ratio; This is the weight matrix; It is the bias vector; It is a non-linear activation function; For network layer indexing; For time step index; This is the index of the convolution kernel position.

7. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, The expression for the loss function is: In the formula, The value of the loss function; To utilize The residual of engine air path parameters obtained by comparing the measured values ​​of engine air path parameters with the simulated values ​​of engine air path parameters at any given time; For multi-scale spatiotemporal attention residual networks in The predicted residual values ​​of the engine air path parameters output at time t; Model parameters for multi-scale spatiotemporal attention residual networks The L2 norm; These are the weighting coefficients of the mean squared error function; These are the weighting coefficients of the mean absolute error function; These are the weighting coefficients of the L2 norm of the model parameters.

8. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, During the training of the multi-scale spatiotemporal attention residual network, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to the network parameters, and the AdamW optimization algorithm combined with the cosine annealing learning rate scheduling strategy is used to update the network weights until the loss function converges to obtain a deep error compensation model with state awareness.

9. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, The measured values ​​of engine air circuit parameters include: compressor outlet pressure, turbine outlet pressure, compressor outlet temperature, turbine outlet temperature, and high-pressure rotor speed.

10. The airborne adaptive modeling method for strong transients in aero-engines according to claim 1, characterized in that, The steps for obtaining the airborne hybrid model by real-time correction of the LPV mechanism model based on the predicted residual values ​​of engine airflow parameters are as follows: The estimated values ​​of engine air path parameters are obtained by correcting the simulated values ​​of engine air path parameters output by the LPV mechanism model using the predicted residual values ​​of engine air path parameters. The difference between the measured values ​​and the estimated values ​​of the engine air path parameters is used as the observation residual after compensation. The compensated observation residuals are input into the Kalman observer for recursive estimation to obtain the real-time gas path health factor. The real-time gas path health factor is then fed back to the LPV mechanism model to correct the LPV mechanism model and obtain the airborne hybrid model.

Images