Method and system for automatically adjusting rotation angle of engine blade

Abstract

The invention discloses a method and a system for automatically adjusting the rotation angle of an engine blade. The method for automatically adjusting the rotation angle of the engine blade comprises the following steps: collecting historical parameters of an engine, collecting real-time data of the engine, carrying out multi-dimensional data fusion on the data to extract key information, screening remarkable influence factors, and constructing a parameter influence network diagram; a four-channel long-short-term memory prediction model is established, four channels output corresponding prediction results respectively, and constraint and error correction are carried out on the prediction results; a corner angle is calculated according to a four-channel output prediction result, constraint limitation and a parameter influence network diagram, and an output corner instruction is directly mapped into actuator displacement; through compensation of a nonlinear actuator, the adjustment precision is guaranteed, and it is ensured that the actuator accurately executes a rotation angle instruction; and the model is lightened, so that the finally obtained decision tree model can efficiently run in an embedded system. The rotating angle of the engine blade can be automatically adjusted under different working conditions, and the adjusting efficiency of the rotating angle of the engine blade is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of engines, and particularly to a method and system for automatically adjusting the blade angle of an engine. Background Art

[0002] As the core component of various power equipment, the performance of an engine is crucial for the operating efficiency and stability of the entire system. The blade angle of the engine has a significant impact on the intake air volume, combustion efficiency, and power output of the engine. Traditional methods for adjusting the blade angle of an engine are mostly manual adjustment or automatic adjustment based on simple fixed logic, which are difficult to adapt to the complex and changeable operating conditions of the engine in real time. During engine startup, acceleration, deceleration, and different load conditions, the demands for intake air volume and combustion efficiency of the engine vary greatly. Traditional adjustment methods cannot accurately match these demands, resulting in increased fuel consumption, unstable power output under certain operating conditions, and even affecting the service life of the engine. Therefore, there is an urgent need for a method that can automatically adjust the blade angle according to the actual operating conditions of the engine in real time and accurately to improve the comprehensive performance of the engine. Based on this, the present invention proposes a method and system for automatically adjusting the blade angle of an engine. Summary of the Invention

[0003] The present invention provides a method for automatically adjusting the blade angle of an engine, including:

[0004] S10. Collect the historical parameters of the engine and acquire the real-time data of the engine, perform multi-dimensional data fusion on the data to extract key information, screen out the significant influencing factors, and construct a parameter influence network diagram;

[0005] S20. Establish a four-channel long short-term memory prediction model, the four channels respectively output corresponding prediction results, and perform constraint and error correction on the prediction results;

[0006] S30. Calculate the rotation angle according to the prediction results output by the four channels, constraint limitations, and the parameter influence network diagram, and directly map the output rotation angle command to the actuator displacement;

[0007] S40. Through non-linear actuator compensation, ensure the adjustment accuracy and ensure that the actuator accurately executes the rotation angle command;

[0008] S50. Lightweight the model so that the finally obtained decision tree model can operate efficiently in an embedded system.

[0009] In the method for automatically adjusting the blade angle of an engine as described above, four types of parameters need to be acquired when acquiring the real-time data of the engine, including aerodynamic parameters, mechanical parameters, control parameters, and environmental parameters.

[0010] An automatic adjustment method for the engine blade angle as described above, wherein the multi-dimensional data fusion step is divided into multi-dimensional data standardization, calculating the cross-modal covariance matrix, and projecting and mapping the high-dimensional cross-modal covariance matrix into a low-dimensional space.

[0011] An automatic adjustment method for the engine blade angle as described above, wherein Granger causality test is used to screen significant influencing factors, which are divided into a restricted model and an unrestricted model. The restricted model can capture the linear correlation characteristics of the engine operating conditions, and the unrestricted model can capture the non-linear correlation characteristics of the engine operating conditions.

[0012] An automatic adjustment method for the engine blade angle as described above, wherein the four channels of the four-channel long short-term memory prediction model are the aerodynamic channel, the structural channel, the control channel, and the environmental channel. The hidden layer of the long short-term memory prediction model performs deep feature extraction on the parameters of the four channels, and the model output result is used to generate the engine rotation angle.

[0013] An automatic adjustment method for the engine blade angle as described above, wherein during the model training process, equation constraints are introduced into the loss function to ensure that the model prediction conforms to the laws of fluid mechanics, avoid incorrect predictions that deviate from physical reality, and correct the prediction error of the model. When the predicted pressure deviation exceeds the threshold, the adjoint equation is triggered to backpropagate and correct the network weights.

[0014] An automatic adjustment method for the engine blade angle as described above, wherein to correct the displacement error of the hydraulic actuator, a hysteresis operator needs to be defined, and a displacement output model is established according to the hysteresis operator to correct the displacement error of the hydraulic actuator caused by non-linear factors such as hysteresis.

[0015] The present invention also provides an automatic adjustment system for the engine blade angle, including:

[0016] The acquisition and processing module: collects the historical parameters of the engine and acquires the real-time data of the engine, performs multi-dimensional data fusion on the data to extract key information, screens significant influencing factors, and constructs a parameter influence network diagram;

[0017] The model construction module: establishes a four-channel long short-term memory prediction model, the four channels respectively output corresponding prediction results, and performs constraint and error correction on the prediction results;

[0018] The rotation angle generation module: calculates the rotation angle according to the prediction results output by the four channels, constraint limitations, and the parameter influence network diagram, and directly maps the output rotation angle command to the actuator displacement;

[0019] The compensation and adjustment module: compensates through a non-linear actuator to ensure the adjustment accuracy and ensure that the actuator accurately executes the rotation angle command;

[0020] Lightweight module: Lightweight the model so that the resulting decision tree model can operate efficiently in an embedded system.

[0021] The beneficial effects achieved by the present invention are as follows: By establishing a four-channel long short-term memory prediction model, the present invention enables the engine blade angle to be automatically adjusted under different working conditions, improving the adjustment efficiency of the engine blade angle. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0023] Figure 1 It is a flowchart of a method for automatically adjusting the engine blade angle provided in Embodiment 1 of the present application.

[0024] Figure 2 It is a schematic diagram of a system for automatically adjusting the engine blade angle provided in Embodiment 2 of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0026] Embodiment 1

[0027] As Figure 1 shown, Embodiment 1 of the present application provides a method for automatically adjusting the engine blade angle, including:

[0028] S10. Collect the historical parameters of the engine and collect the real-time data of the engine, perform multi-dimensional data fusion on the data to extract key information, screen out the significant influencing factors, and construct a parameter influence network diagram.

[0029] S11. Collect the historical parameters of the engine and collect the real-time data of the engine.

[0030] Collect the historical parameters of the engine.

[0031] Deploy sensors at the key parts of the engine to collect four types of parameters in real time, namely, pneumatic parameters, mechanical parameters, control parameters, and environmental parameters. The engine working conditions have strong time-varying characteristics, and the sampling frequencies and dimensions of different sensors will cause time-axis mismatch, so the sampling frequency is unified to 10 kHz.

[0032] Aerodynamic parameters reflect the state of the engine flow field and the distribution of aerodynamic loads, including Mach number, angle of attack, dynamic pressure, engine gas velocity, and flow rate data. The Mach number calculation formula is The total pressure P is measured by an on-board pitot tube t and the static pressure P s , P t is measured through the central hole of the pitot tube, and P s is measured through the side wall hole of the pitot tube. γ is the specific heat ratio of air. α is the angle of attack, which is the angle between the oncoming flow direction and the blade chord line, used to predict the risk of flow separation and optimize the blade camber. The dynamic pressure q is jointly derived from the Mach number and the total temperature T total , including the engine inlet pressure, combustion chamber pressure, etc.

[0033] Mechanical parameters describe the structural state of the blade and the actuator system, providing key inputs such as material stress and vibration modes for the control algorithm, including the surface stress and strain of the engine blade, the engine rotor speed, vibration parameters, clearance data, actuator displacement, and spindle torque. 40 groups of strain gauges are arranged on the surface of the engine blade, covering key positions on the leading edge, pressure surface, and suction surface, to measure the circumferential strain distribution ε i (t), with a range of ±5000 με. The actuator displacement is the linear displacement of the hydraulic actuator push rod, directly controlling the blade rotation angle. The output of the actuator sensor is x(t), with an accuracy range of ±0.1 mm. The measurement result of the spindle torque sensor is τ(t), with a range of 0 - 5000 N·m, used to monitor sudden changes in load and prevent structural damage caused by overload.

[0034] Control parameters reflect the commands and feedback signals of the control system state and historical actions, including the command voltage of the full-authority digital controller, throttle opening, historical rotation angle trajectory, fuel injection quantity, valve timing, and control mode status code. The output command voltage of the full-authority digital controller is V cmd ∈[0, 10] V, and the historical rotation angle trajectory buffer stores the rotation angle data θ hist (t - Δt) in the last 60 seconds, with a time resolution of 1 ms. The control mode state is the basis for dynamically adjusting the multi-objective optimization weight. The state code 0 represents the manual mode, 1 represents the efficiency priority mode, 2 represents the safety protection mode, and 3 represents the fault recovery mode.

[0035] Environmental parameters reflect the environmental variables of the external working conditions of the engine, including total temperature, atmospheric pressure, humidity, icing detection signal, barometric altitude, wind direction and wind speed. The total temperature sensor measures the intake total temperature T total , using a K-type thermocouple, with a range of -50°C to 600°C. The output signal I of the millimeter-wave icing detector ice ∈[0, 1], and 1 represents the severe icing state. The barometric altitude is calculated by the formula ​Calculated, T0 represents the standard temperature at sea level, L represents the temperature lapse rate, P s represents the static pressure, P0 represents the standard atmospheric pressure at sea level, R represents the specific gas constant of dry air, and g represents the acceleration due to gravity.

[0036] S12. Establish a cross-modal correlation subspace and perform multi-dimensional data fusion to extract key feature information.

[0037] In the engine control system, the coupled interference of multi-source sensor data causes distortion in feature extraction. When the compressor surges, the mechanical vibration signal will contaminate the measurement of aerodynamic parameters, and the existing Kalman filtering method is difficult to handle the data association across physical domains. By constructing a cross-modal joint subspace, the cross-interference of sensors is eliminated.

[0038] Normalize the multi-dimensional data. The parameter data is transformed into matrix form through X ∈ A a×b way, and the parameter matrix includes the aerodynamic parameter matrix mechanical parameter matrix control parameter matrix environmental parameter matrix The normalization formula is Normalize the four types of parameter data respectively. A represents the set of real numbers, a represents the number of parameter samples, b represents the number of features in the sample, b1 represents the number of aerodynamic parameter features, b2 represents the number of mechanical parameter features, b3 represents the number of control parameter features, and b4 represents the number of environmental parameter features. X i represents the i-th column feature of the parameter matrix, μ X represents the mean vector of parameter X, and σ X represents the standard deviation vector of parameter X.

[0039] Calculate the cross-modal covariance matrix based on the normalized parameter matrix to discover the potential structure between parameters, so as to better fuse the information of different parameters. Concatenate the four normalized parameter matrices by column into The specific formula for the cross-modal covariance matrix is where n is the number of samples in the time window, X' is the deviation matrix of matrix X total and X' T is the transpose of the deviation matrix.

[0040] Project the high-dimensional cross-modal covariance matrix onto a low-dimensional space to eliminate redundant information. Solve the maximum correlation projection vector through singular value decomposition, and the formula is where Z is the left singular vector matrix, which is an orthogonal matrix; Λ is a diagonal matrix, and the elements on the diagonal are singular values; V is the right singular vector matrix, which is an orthogonal matrix. a * - * ​=Z(:,1) means to take the first column of Z to extract the most crucial feature information. b * represents the maximum correlation projection vector obtained from the right singular vector matrix V, b * =V(:,1) means to take the first column of V. Sort the singular values, select the top k larger singular values and their corresponding singular vectors, and generate a 2k-dimensional feature vector F = [z1, z2,..., z k , v1, v2,..., v k .

[0041] S13. Screen significant influencing factors and construct a parameter influence network diagram.

[0042] Use Granger causality test to screen significant influencing factors. For each element F in the feature vector F k Construct an autoregressive model. The restricted model can capture the linear correlation characteristics of the engine operating conditions. The formula is where p is the past p time steps for constructing the model, represents the weight coefficient corresponding to the i-th past time step, which determines the influence degree of the eigenvalue F k (t - iΔt) on the output F k1 (t) at the current moment. F k (t - iΔt) represents the value of the k-th element in the feature vector F at the moment t - iΔt, that is, the eigenvalue at the i-th past time step, and Δt is the time interval. is the residual term, representing the random noise not explained by the model; the unrestricted model can capture the non-linear correlation characteristics of the engine operating conditions. The formula is where, β i is the weight coefficient corresponding to the i-th past time step of the k-th element in the feature vector F. j is used to traverse other elements in the feature vector F except the k-th element. is the weight coefficient corresponding to the j-th past time step of other elements in the feature vector F except the k-th element.

[0043] Use the formula to compare the differences in the sum of squared residuals between the restricted model and the unrestricted model to determine whether one variable is the Granger cause of another variable. If the F Granger value is large, it indicates that the fitting effect of the unrestricted model on the data is significantly better than that of the restricted model, meaning that including the lag terms of other variables can significantly improve the prediction ability, that is, there is a Granger causal relationship. Among them, RSS r is the sum of squared residuals of the restricted model, and RSS u is the sum of squared residuals of the unrestricted model. p is the lag order, selected according to the Akaike criterion, and T is the length of the time series. Screen the causal relationships with p < 0.01 and construct a parameter influence network diagram Used for subsequent control strategy optimization.

[0044] S20. Establish a four-channel long-short-term memory prediction model, where the four channels output corresponding prediction results respectively, and constrain and correct the prediction results.

[0045] S21. Construction of four-channel long short-term memory prediction model.

[0046] A four-channel long-short-term memory prediction model is established to capture the relationship between flow field structure, blade stress state, and the influence of engine environment and operating conditions, providing comprehensive information for generating blade angle commands. The network structure consists of four parallel branches: an aerodynamic channel, a structural channel, a control channel, and an environmental channel. Information flow is controlled by input gates, forget gates, and output gates.

[0047] When processing the data at each time step, the input gate incorporates the changes in key parameters, and the memory unit continuously stores and updates long-term information that is useful for predicting blade angles. The aerodynamic channel inputs historical aerodynamic parameters through the input gate, the structural channel inputs historical mechanical parameters, the control channel inputs historical control parameters, and the environmental channel inputs historical environmental parameters. The specific formula is i t =σ(W i ·[h t-1 ,x t ]+b i ), where x t is the current input information, σ is the activation function, W i is the input gate weight matrix, b i is the input gate bias vector; the forget gate decides to discard the early irrelevant parameter information from the memory unit. The specific formula is f t =σ(W f ·[h t-1 ,x t ]+b f ), where W f is the forget gate weight matrix, [h t-1 ,x t ] indicates that the previous moment hidden state h t-1 and the current input x t Splice by dimension, b f is the forget gate bias vector; the output gate determines the output value based on the memory unit state, and the specific formula is o t =σ(W o ·[h t-1 ,x t ]+b o ), where W o is the output gate weight matrix, b o is the output gate bias vector. The hidden state update formula is Current hidden state h tOutput o is output by the output gate t and the memory cell C processed by the tanh activation function t are element-wise multiplied. The hidden state contains the current input and historical information and is used to be passed to the next time step and as an output.

[0048] As data is sequentially input in time series, the hidden layer of the long short-term memory prediction model performs deep feature extraction on the four-channel parameters. Through the hierarchical processing of multiple long short-term memory prediction units, the complex non-linear relationships between the parameters are mined and memorized. After training, the input gate of the model inputs the real-time parameter data of the engine, and the output gate of the pneumatic channel outputs the pressure distribution P pred 、shock position S sep 、flow separation region (x shock , y shock ); the structural channel outputs the first-order torsional mode frequency f modal 、damping ratio ζ、predicted stress distribution σ pred ; the control channel outputs the historical rotation angle trajectory θ hist 、actuator state feedback displacement x act 、delay τ delay ; the environmental channel outputs the icing thickness d ice 、temperature gradient ΔT / Δt、air density correction coefficient ρ corr , which are used to generate the engine rotation angle, rather than passively respond.

[0049] S22. Introduce equation constraints and error correction.

[0050] Introduce equation constraints in the loss function to ensure that the model prediction conforms to the laws of fluid mechanics and avoid incorrect predictions that deviate from physical reality. The physical residual term is defined as where u is the velocity field vector, p is the pressure field scalar, ρ is the fluid density, and v is the kinematic viscosity. A composite loss function L loss is constructed based on the physical residual term. The composite loss function combines the deviations of multiple physical quantities, measures the difference between the model prediction and the actual physical situation and reference data, and is used for model training optimization.

[0051]

[0052] Pressure field, P CFD is the benchmark value of the fluid dynamics simulation, Ω is the fluid flow space region under study, is the predicted blade modal frequency, is the reference value of the blade modal frequency from finite element analysis, and h1, h2, h3 are the weighting coefficients.

[0053] Perform model prediction error correction and calculate the relative error of the pressure field P predFor predicting the pressure field, P sensor is the numerical value of the pressure field actually measured by the sensor, P max is the maximum value in the pressure field data, P min is the minimum value in the pressure field data. When the predicted pressure deviation exceeds the threshold, the adjoint equation backpropagation is triggered to correct the network weights. Based on the backpropagation of the adjoint equation, the adjoint variable λ is constructed and solved Update the network weights where η is the learning rate is the loss function L loss The partial derivative of the network weight W, Φ is the physical correction coefficient

[0054] S30. Calculate the corner angle according to the four-channel output prediction result, constraint limit, and parameter influence network diagram, and directly map the output corner command to the actuator displacement

[0055] Perform actual constraint operation limitations based on the results obtained from the model in step S20 to obtain the actual operable corner command. The specific steps are as follows

[0056] S31. Construct multi-channel constraints

[0057] S311. Aerodynamic channel constraints

[0058] The aerodynamic channel constraints include flow separation limit and shock oscillation limit. The flow separation limit means that the predicted separation area A sep is greater than the separation area threshold A th , then the angle of attack adjustment range is restricted: α max =α nominal -d1A sep , α nominal is the angle of attack value under normal conditions, d1 is used to quantify the influence degree of the predicted separation area on the maximum angle of attack; the shock oscillation limit means that the shock wave intensity ΔP shock constrains the maximum corner rate, and the formula is

[0059] S312. Structural channel constraints

[0060] The structural channel constraints include resonance avoidance and stress limit. Resonance avoidance means that when the modal frequency f modal is close to the rotational speed harmonic n×RPM / 60, a specific corner is prohibited. The specific formula is n is the harmonic order, RPM is the engine rotor speed per minute, and the stress limit means dynamically adjusting the allowable corner according to the predicted stress θ design refers to the rotation angle of components such as blades set during the engine design stage, which is used as the reference value for calculating the allowable corner. σ pred is the predicted stress, σ yieldis the yield stress.

[0061] S313. Control channel constraints.

[0062] The control channel constraints include actuator delay compensation and displacement tracking error. Actuator delay compensation refers to the use of the formula Corrected command phase, θ cmd is the original instruction angle, e is a natural constant, τ delay is the actuator delayed displacement, s is the Laplace operator, which is used to convert the time domain signal to the complex frequency domain for analysis and processing; the displacement tracking error is expressed using the formula Real-time gain adjustment, K p is the proportional gain, K i is the integral gain, K d is the differential gain, f CMA-ES Indicates the error e and error change rate and error integral ∫e to determine K p , K i , K d function of θ. cmd is the angle command, θ act is the actual rotation angle of the blade.

[0063] S314, environmental channel constraints.

[0064] Environmental channel constraints include icing stiffness correction and temperature-density compensation. Icing stiffness correction refers to the ice thickness d ice Increase the blade equivalent stiffness, the formula is K eff =K0(1+0.2d ice ), K0 represents the initial stiffness of the blade; temperature-density compensation refers to the air density correction affecting the aerodynamic load calculation, and the formula is ρ corr is the corrected air density, u air is the gas flow rate, ρ nominal is the air density under standard conditions.

[0065] S32. Calculate the turning angle according to the four-channel constraints and the parameter influence network diagram.

[0066] The engine blade angle is calculated by integrating the four-channel constraints and parameter influence network diagram. The specific formula is: Among them, θ hist The historical corner trajectory generated by the model, w aero is the weight of the aerodynamic channel constraint, Y aero is the pneumatic channel constraint, w mech is the weight of the structural channel constraint, Y mech is the structural channel constraint, w ctrlTo control the weight of the channel constraint, Y ctrl To control the channel constraint condition, w env To control the weight of the environmental channel constraint, Y env To control the environmental channel constraint condition represents the parameter influence network diagram. The actuator displacement command is generated according to the difference between the corner angle and the engine's real-time corner angle, so as to realize the adjustment of the engine blade corner.

[0067] S40. Through non-linear actuator compensation, the adjustment accuracy is guaranteed, ensuring that the actuator accurately executes the corner command and eliminating the non-linear error of the mechanical system.

[0068] The displacement error of the hydraulic actuator is corrected through the Preisach model to ensure the linearization of the command-displacement. For any input voltage V(t), the hysteresis operator is defined m, n are threshold parameters, satisfying m > n. By comparing the hysteresis operator with the thresholds m and n, the hysteresis behavior between the input voltage and the output state of the hydraulic actuator is characterized. According to the hysteresis operator, the displacement output model x(t) = ∫∫ m≥n μ(m,n)C mn [V(t)]dmdn is established to correct the displacement error of the hydraulic actuator caused by non-linear factors such as hysteresis, improve the control accuracy of the actuator, and ensure the consistency between the actual corner of the blade and the target value. Among them, μ(m,n) represents the weight function, which assigns weights to different threshold combinations (m,n) and is used to characterize the contribution degree of different hysteresis characteristics to the final displacement output.

[0069] S50. Lightweight the model so that the finally obtained decision tree model can run efficiently in the embedded system.

[0070] Conduct strategy lightweighting, compress the complex model into a decision tree that can be run by the embedded system, and on the premise of ensuring the real-time performance of the model, retain the performance of the complex model as much as possible, so that the finally obtained decision tree model can run efficiently in the embedded system and accurately provide decision-making strategies for blade corner control.

[0071] The specific formula is where S is the state set, H() represents the cross-entropy function, which is used to measure the difference degree between two probability distributions and guide the migration model to learn from the original model, π origina (a origina |s origina ) is the original model, indicating the probability distribution of taking action a origina under the state s origina , and π migrate (a migrate |s migrate) It is the lightweight model after migration. By calculating the difference in the action probability distributions of the two in the same state, the migrated model learns the decision-making mode of the original model. χ is a hyperparameter used to balance the weights of the cross-entropy term and the decision tree depth penalty term. Depth(B) represents the depth of the decision tree. During the process of compressing a complex model into a decision tree, the depth of the decision tree affects the complexity and computational efficiency of the model. This term is introduced as a penalty term to prevent the decision tree from being too deep, resulting in an overly complex model that affects real-time performance and generalization ability.

[0072] Embodiment 2

[0073] As Figure 2 shown, Embodiment 2 of the present application provides an automatic engine blade angle adjustment system, including:

[0074] Acquisition and processing module: including a collection sub-module, a fusion sub-module, and a screening sub-module.

[0075] Collection sub-module: used to collect historical engine parameters and acquire real-time engine data.

[0076] Collect historical engine parameters.

[0077] Deploy sensors at key parts of the engine to collect four types of parameters in real time, namely aerodynamic parameters, mechanical parameters, control parameters, and environmental parameters. The engine operating conditions have strong time-varying characteristics. The sampling frequencies and dimensions of different sensors will cause time-axis mismatch. Therefore, the sampling frequency is unified to 10 kHz.

[0078] Aerodynamic parameters reflect the engine flow field state and aerodynamic load distribution, including Mach number, angle of attack, dynamic pressure, engine gas velocity, and flow rate data. The Mach number calculation formula is Measure the total pressure P through the on-board pitot tube t and the static pressure P s , P t is measured through the center hole of the pitot tube, P s is measured through the side wall hole of the pitot tube, γ is the specific heat ratio of air. α is the angle of attack, which is the angle between the oncoming flow direction and the blade chord line, used to predict the risk of flow separation and optimize the blade camber. The dynamic pressure q is jointly derived from the Mach number and the total temperature T total , including engine inlet pressure, combustion chamber pressure, etc.

[0079] Mechanical parameters describe the structural state of the blade and the actuator system, providing key inputs such as material stress and vibration modes for the control algorithm, including engine blade surface stress and strain, engine rotor speed, vibration parameters, clearance data, actuator displacement, and spindle torque. Arrange 40 groups of strain gauges on the engine blade surface, covering key positions on the leading edge, pressure surface, and suction surface, and measure the circumferential strain distribution ε i(t), with a measurement range of ±5000 με. The displacement of the actuator is the linear displacement of the hydraulic actuator push rod, which directly controls the blade rotation angle. The output of the actuator sensor is x(t), with an accuracy range of ±0.1 mm. The measurement result of the main shaft torque sensor is τ(t), with a measurement range of 0 - 5000 N·m, which is used to monitor sudden load changes and prevent structural damage caused by overload.

[0080] Control parameters reflect the instructions and feedback signals of the control system state and historical actions, including the command voltage of the full-authority digital controller, throttle opening, historical rotation angle trajectory, fuel injection quantity, valve timing, and control mode status code. The output command voltage of the full-authority digital controller is V cmd ∈[0, 10] V. The historical rotation angle trajectory buffer stores the rotation angle data θ hist (t - Δt) with a time resolution of 1 ms. The control mode state is the basis for dynamically adjusting the multi-objective optimization weight. The state code 0 represents the manual mode, 1 represents the efficiency priority mode, 2 represents the safety protection mode, and 3 represents the fault recovery mode.

[0081] Environmental parameters reflect the environmental variables of the external working conditions of the engine, including total temperature, atmospheric pressure, humidity, icing detection signal, barometric altitude, wind direction and speed. The total temperature sensor measures the intake total temperature T total , using a K-type thermocouple with a measurement range of -50°C to 600°C. The output signal of the millimeter-wave icing detector is I ice ∈[0, 1], where 1 represents the severe icing state. The barometric altitude is calculated by the formula , where T0 represents the standard temperature at sea level, L represents the temperature lapse rate, P s represents the static pressure, P0 represents the standard atmospheric pressure at sea level, R represents the specific gas constant of dry air, and g represents the acceleration due to gravity.

[0082] Fusion sub-module: used to establish a cross-modal correlation subspace and perform multi-dimensional data fusion to extract key feature information.

[0083] In the engine control system, the coupled interference of multi-source sensor data causes distortion in feature extraction. When the compressor surges, the mechanical vibration signal will contaminate the measurement of aerodynamic parameters, and the existing Kalman filtering method is difficult to handle the data association across physical domains. By constructing a cross-modal joint subspace, the cross-interference of sensors is eliminated.

[0084] Standardize the multi-dimensional data. The parameter data is converted into matrix form through X ∈ A a×b way. The parameter matrix includes the aerodynamic parameter matrix mechanical parameter matrix control parameter matrix environmental parameter matrix The standardization formula is Standardize the four types of parameter data respectively. Let \(A\) represent the set of real numbers, \(a\) represent the number of parameter samples, \(b\) represent the number of features in the sample, \(b_1\) represent the number of aerodynamic parameter features, \(b_2\) represent the number of mechanical parameter features, \(b_3\) represent the number of control parameter features, and \(b_4\) represent the number of environmental parameter features. \(X\) i represents the \(i\)-th column feature of the parameter matrix, and \(\mu\) X represents the mean vector of parameter \(X\), and \(\sigma\) X represents the standard deviation vector of parameter \(X\).

[0085] Calculate the cross-modal covariance matrix based on the standardized parameter matrix to discover the potential structure between parameters, so as to better fuse the information of different parameters. Concatenate the four standardized parameter matrices column by column into The specific formula for the cross-modal covariance matrix is where \(n\) is the number of samples in the time window, and \(X'\) is the deviation matrix of matrix \(X\) total , and \(X'\) T is the transpose of the deviation matrix.

[0086] Project the high-dimensional cross-modal covariance matrix into a low-dimensional space to eliminate redundant information. Solve the maximum correlation projection vector through singular value decomposition, and the formula is where \(Z\) is the left singular vector matrix and is an orthogonal matrix; \(\Lambda\) is a diagonal matrix, and the elements on the diagonal are singular values; \(V\) is the right singular vector matrix and is an orthogonal matrix. \(a\) * [[ID=2nd]]represents the maximum correlation projection vector obtained through the left singular vector matrix \(Z\), and \(a\) * = \(Z(:,1)\) means taking the first column of \(Z\) to extract the most critical feature information. \(b\) * represents the maximum correlation projection vector obtained through the right singular vector matrix \(V\), and \(b\) * = \(V(:,1)\) means taking the first column of \(V\). Sort the singular values, select the top \(k\) larger singular values and their corresponding singular vectors, and generate a \(2k\)-dimensional feature vector \(F = [z_1, z_2,..., z\) k , \(v_1, v_2,..., v\) k .

[0087] Screening sub-module: used to screen significant influencing factors and construct a parameter influence network diagram.

[0088] Use Granger causality test to screen significant influencing factors. For each element \(F\) in the feature vector \(F\) k Construct an autoregressive model. The restricted model can capture the linear correlation characteristics of the engine operating conditions, and the formula is where \(p\) is the past \(p\) time steps for constructing the model, represents the weight coefficient corresponding to the \(i\)-th past time step, which determines the eigenvalue \(F\) at the \(i\)-th past moment k(t - iΔt)'s influence on the output F at the current moment k1 on F(t), k (t - iΔt) represents the value of the k-th element in the feature vector F at the moment t - iΔt, that is, the eigenvalue at the i-th past time step. Δt is the time interval. is the residual term, representing the random noise not explained by the model; the unrestricted model can capture the non-linear related features of the engine operating conditions. The formula is where β i is the weight coefficient corresponding to the k-th element in the feature vector F at the i-th past time step. j is used to traverse the other elements in the feature vector F except the k-th element. is the weight coefficient corresponding to the j-th past time step for the elements in the feature vector F except the k-th element.

[0089] Using the formula to compare the differences in the sum of squared residuals between the restricted model and the unrestricted model to determine whether one variable is the Granger cause of another variable. If the F Granger value is large, it indicates that the unrestricted model fits the data significantly better than the restricted model, meaning that incorporating the lag terms of other variables can significantly improve the prediction ability, that is, there is a Granger causal relationship. Among them, RSS r is the sum of squared residuals of the restricted model, and RSS u is the sum of squared residuals of the unrestricted model. p is the lag order, selected according to the Akaike criterion. T is the length of the time series. Screen the causal relationships with p < 0.01 to construct a parameter influence network diagram for subsequent control strategy optimization.

[0090] Construct the model module: including the construction sub-module and the constraint correction sub-module.

[0091] Construction sub-module: used to establish a four-channel long short-term memory prediction model.

[0092] Establish a four-channel long short-term memory prediction model to capture the correlation between the flow field structure and the blade force state, and the influence of the engine environment and operating conditions, providing comprehensive information for generating the blade angle command. The network structure includes four parallel branches: the aerodynamic channel, the structural channel, the control channel, and the environment channel. The information flow is controlled by the input gate, the forget gate, and the output gate.

[0093] When processing the data at each time step, the input gate incorporates the changes in key parameters, and the memory unit continuously stores and updates the long-term information useful for predicting the blade angle. The aerodynamic channel inputs the historical aerodynamic parameters through the input gate, the structural channel inputs the historical mechanical parameters, the control channel inputs the historical control parameters, and the environment channel inputs the historical environment parameters. The specific formula is i t = σ(W i · [ht-1 , x t + b i ), where x t is the current input information, σ is the activation function, and W i is the input gate weight matrix, and b i is the input gate bias vector; the forget gate determines which earlier irrelevant parameter information to discard from the memory cell. The specific formula is f t = σ(W f · [h t-1 , x t + b f ), where W f is the forget gate weight matrix, and [h t-1 , x t represents concatenating the previous hidden state h t-1 and the current input x t along the dimension, and b f is the forget gate bias vector; the output gate determines the output value based on the memory cell state. The specific formula is o t = σ(W o · [h t-1 , x t + b o ), where W o is the output gate weight matrix, and b o is the output gate bias vector. The hidden state update formula is The current hidden state h t is obtained by element-wise multiplication of the output o t from the output gate and the memory cell C t processed by the tanh activation function. The hidden state contains the current input and historical information and is used to be passed to the next time step and as an output.

[0094] As data is input sequentially in a time series, the hidden layer of the long short-term memory prediction model performs deep feature extraction on the four-channel parameters. Through the layer-by-layer processing of multiple long short-term memory prediction units, the complex non-linear relationships between the parameters are mined and memorized. After training, the input gate of the model inputs the real-time parameter data of the engine, and the output gate of the aerodynamic channel outputs the future 500 ms pressure distribution P pred , shock position S sep , flow separation region (x shock , y shock ); the structural channel outputs the first-order torsional mode frequency f modal , damping ratio ζ, and predicted stress distribution σ pred ; the control channel outputs the historical rotation angle trajectory θ hist , actuator state feedback displacement x act , and delay τ delay; The icing thickness d of the environmental channel output ice , the temperature gradient ΔT / Δt, and the air density correction coefficient ρ corr , which is used to generate the engine rotation angle, rather than passively respond.

[0095] Constraint correction sub-module: used to introduce equation constraints and error correction.

[0096] Introduce equation constraints in the loss function to ensure that the model prediction conforms to the laws of fluid mechanics and avoid incorrect predictions that deviate from physical reality. The physical residual term is defined as where u is the velocity field vector, p is the pressure field scalar, ρ is the fluid density, and v is the kinematic viscosity. Construct the composite loss function L based on the physical residual term loss , and the composite loss function synthesizes the deviations of multiple physical quantities to measure the difference between the model prediction and the actual physical situation and reference data, and is used for model training optimization.

[0097] The specific formula is P pred is the predicted pressure field, P CFD is the benchmark value of the fluid dynamics simulation, Ω is the fluid flow space region under study, is the predicted blade modal frequency, [[ID=...]] (the content continues as in the original, with proper line breaks and tags preserved) is the reference value of the blade modal frequency from the finite element analysis, and h1, h2, h3 are the weighting coefficients.

[0098] Perform model prediction error correction and calculate the relative error of the pressure field P pred is the predicted pressure field, P sensor is the pressure field value actually measured by the sensor, P max is the maximum value in the pressure field data, P min is the minimum value in the pressure field data. When the predicted pressure deviation exceeds the threshold, trigger the adjoint equation backpropagation to correct the network weights. Based on the backpropagation of the adjoint equation, construct the adjoint variable λ and solve Update the network weights where η is the learning rate, is the loss function L loss is the partial derivative of the loss function L with respect to the network weight W, and Φ is the physical correction coefficient.

[0099] Generate rotation angle module: construct a multi-channel constraint sub-module and a rotation angle calculation sub-module.

[0100] Construct a multi-channel constraint sub-module: including an aerodynamic channel constraint sub-module, a structural channel constraint sub-module, a control channel constraint sub-module, and an environmental channel constraint sub-module.

[0101] Aerodynamic channel constraint sub-module: used to generate aerodynamic channel constraints.

[0102] The pneumatic channel constraint includes flow separation limitation and shock oscillation limitation. The flow separation limitation means predicting the separation area A sep is greater than the separation area threshold A th , then the angle of attack adjustment range is restricted: α max = α nominal - d1A sep , α nominal is the angle of attack value under normal conditions, and d1 is used to quantify the influence degree of the predicted separation area on the maximum angle of attack; the shock oscillation limitation means that the shock intensity ΔP shock restricts the maximum rotational speed rate, and the formula is

[0103] Structural channel constraint sub-module: used to generate structural channel constraints.

[0104] The structural channel constraint includes resonance avoidance and stress limitation. Resonance avoidance means that when the modal frequency f modal is close to the rotational speed harmonic n×RPM / 60, a specific rotation angle is prohibited. The specific formula is n is the harmonic order, RPM is the engine rotor speed per minute, and the stress limitation means dynamically adjusting the allowable rotation angle according to the predicted stress, θ design refers to the rotation angle of components such as blades set during the engine design stage and serves as the reference value for calculating the allowable rotation angle. σ pred is the predicted stress, and σ yield is the yield stress.

[0105] Control channel constraint sub-module: used to generate control channel constraints.

[0106] The control channel constraint includes actuator delay compensation and displacement tracking error. Actuator delay compensation means using the formula to correct the command phase, θ cmd is the original command rotation angle, e is the natural constant, τ delay is the actuator delay displacement, and s is the Laplace operator used to transform the signal in the time domain to the complex frequency domain for analysis and processing; the displacement tracking error means using the formula to adjust the gain in real time. K p is the proportional gain, K i is the integral gain, K<> d is the derivative gain, and f CMA-ES represents a function to determine K , K p , and K i based on the error e, the error change rate d and the error integral ∫e. θ cmd is the rotation angle command, and θ act is the actual rotation angle of the blade.

[0107] Environmental channel constraint sub-module: used to generate environmental channel constraints.

[0108] Environmental channel constraints include icing stiffness correction and temperature-density compensation. Icing stiffness correction refers to the icing thickness d ice increasing the equivalent stiffness of the blade, and the formula is K eff = K0(1 + 0.2d ice ), where K0 represents the initial stiffness of the blade; temperature-density compensation refers to the influence of air density correction on the calculation of aerodynamic loads, and the formula is ρ corr is the corrected air density, u air is the gas flow velocity, ρ nominal is the air density under standard environmental conditions.

[0109] Calculation of the rotation angle sub-module: used to calculate the rotation angle according to the four-channel constraints and the parameter influence network diagram.

[0110] Combining the four-channel constraints and the parameter influence network diagram, calculate the rotation angle of the engine blade. The specific formula is where θ hist is the historical rotation angle trajectory generated by the model, w aero is the weight of the aerodynamic channel constraint, Y aero is the aerodynamic channel constraint condition, w mech is the weight of the structural channel constraint, Y mech is the structural channel constraint condition, w ctrl is the weight of the control channel constraint, Y ctrl is the control channel constraint condition, w env is the weight of the environmental channel constraint, Y env is the environmental channel constraint condition, represents the parameter influence network diagram. Generate the actuator displacement command according to the difference between the rotation angle and the real-time rotation angle of the engine, so as to realize the adjustment of the rotation angle of the engine blade.

[0111] Compensation adjustment module: Ensure the adjustment accuracy through non-linear actuator compensation, and ensure that the actuator accurately executes the rotation angle command.

[0112] Correct the displacement error of the hydraulic actuator through the Preisach model to ensure the linearization of the command-displacement. For any input voltage V(t), define the hysteresis operator m, n are threshold parameters, satisfying m > n. By comparing the hysteresis operator with the thresholds m and n, the hysteresis behavior between the input voltage and the output state of the hydraulic actuator is characterized. Establish the displacement output model x(t) = ∫∫ m≥n μ(m,n)C mn[V(t)]dmdn corrects the displacement error of the hydraulic actuator caused by non-linear factors such as hysteresis, improves the control accuracy of the actuator, and ensures the consistency between the actual blade rotation angle and the target value. Among them, μ(m,n) represents the weight function, which assigns weights to different threshold combinations (m,n) to describe the contribution degree of different hysteresis characteristics to the final displacement output.

[0113] Lightweight module: Lightweight the model so that the finally obtained decision tree model can run efficiently in the embedded system.

[0114] Perform strategy lightweighting, compress the complex model into a decision tree that can be run by the embedded system, and while ensuring the real-time performance of the model, retain the performance of the complex model as much as possible, so that the finally obtained decision tree model can run efficiently in the embedded system and accurately provide decision-making strategies for blade rotation angle control.

[0115] The specific formula is where S is the state set, H() represents the cross-entropy function, which is used to measure the difference degree between two probability distributions, guide the migration model to learn from the original model, π origina (a origina |s origina ) is the original model, indicating the probability distribution of taking action a origina under the state s origina , and π migrate (a migrate |s migrate ) is the migrated lightweight model. By calculating the difference in the action probability distributions of the two under the same state, the migration model learns the decision-making mode of the original model. χ is a hyperparameter used to balance the weights of the cross-entropy term and the decision tree depth penalty term. Depth(B) represents the depth of the decision tree. During the process of compressing the complex model into a decision tree, the decision tree depth will affect the complexity and computational efficiency of the model. Introduce this term as a penalty term to prevent the decision tree from being too deep, resulting in an overly complex model that affects real-time performance and generalization ability.

[0116] The specific implementation manners described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above is only the specific implementation manners of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of the present invention shall be included in the protection scope of the present invention.

Claims

1. An automatic adjustment method for the engine blade angle, characterized in that Including: S10. Collect the historical parameters of the engine and acquire the real-time data of the engine, perform multi-dimensional data fusion on the data to extract key information, screen out the significant influencing factors, and construct a parameter influence network diagram; S20. Establish a four-channel long short-term memory prediction model. The four channels respectively output corresponding prediction results, and perform constraint and error correction on the prediction results; S30. Calculate the rotation angle according to the prediction results output by the four channels, the constraint limits, and the parameter influence network diagram, and directly map the output rotation angle command to the actuator displacement; S40. Through non-linear actuator compensation, ensure the adjustment accuracy and ensure that the actuator accurately executes the rotation angle command; S50. Lightweight the model so that the finally obtained decision tree model can operate efficiently in the embedded system.

2. The automatic adjustment method for the engine blade rotation angle according to claim 1, characterized in that When acquiring the real-time data of the engine, four types of parameters need to be acquired, including aerodynamic parameters, mechanical parameters, control parameters, and environmental parameters.

3. The automatic adjustment method for the engine blade rotation angle according to claim 1, characterized in that, The multi-dimensional data fusion steps are divided into multi-dimensional data standardization, calculating the cross-modal covariance matrix, and projecting and mapping the high-dimensional cross-modal covariance matrix into a low-dimensional space.

4. The automatic adjustment method for the engine blade rotation angle according to claim 1, wherein, Use Granger causality test to screen out the significant influencing factors, which are divided into a restricted model and an unrestricted model. The restricted model can capture the linear correlation characteristics of the engine operating conditions, and the unrestricted model can capture the non-linear correlation characteristics of the engine operating conditions.

5. The automatic adjustment method for the engine blade rotation angle according to claim 1, characterized in that The four channels of the four-channel long short-term memory prediction model are the aerodynamic channel, the structural channel, the control channel, and the environmental channel. The hidden layer of the long short-term memory prediction model performs deep feature extraction on the parameters of the four channels, and the model output result is used to generate the engine rotation angle.

6. The automatic adjustment method for the engine blade rotation angle according to claim 1, characterized in that, During the model training process, introduce equation constraints into the loss function to ensure that the model prediction conforms to the laws of fluid mechanics, avoid incorrect predictions that deviate from physical reality, and perform prediction error correction on the model. When the predicted pressure deviation exceeds the threshold, trigger the adjoint equation backpropagation to correct the network weights.

7. The automatic adjustment method for the engine blade rotation angle according to claim 1, characterized in that, To correct the displacement error of the hydraulic actuator, a hysteresis operator needs to be defined, and a displacement output model is established according to the hysteresis operator to correct the displacement error of the hydraulic actuator caused by non-linear factors such as hysteresis.

8. An automatic engine blade angle adjustment system, characterized in that, Including: Acquisition and processing module: Collect the historical parameters of the engine and acquire the real-time data of the engine, perform multi-dimensional data fusion on the data to extract key information, screen out the significant influencing factors, and construct a parameter influence network diagram; Model construction module: Establish a four-channel long short-term memory prediction model. The four channels respectively output corresponding prediction results, and perform constraint and error correction on the prediction results; Rotation angle generation module: Calculate the rotation angle according to the prediction results output by the four channels, the constraint limits, and the parameter influence network diagram, and directly map the output rotation angle command to the actuator displacement; Compensation and adjustment module: Through non-linear actuator compensation, ensure the adjustment accuracy and ensure that the actuator accurately executes the rotation angle command; Lightweight module: Lightweight the model so that the finally obtained decision tree model can operate efficiently in the embedded system.

Images