Protection and control methods for hypersonic aircraft that do not require air intake start-up

Abstract

This invention belongs to the field of hypersonic vehicle control technology, specifically relating to a protection control method for hypersonic vehicles that do not require inlet operation. Specifically, it includes: establishing a quasi-one-dimensional nominal mechanism model to describe the flow field changes in the isolation section and combustion chamber; constructing a reduced-order model using a physical information neural network as a real-time proxy for the quasi-one-dimensional model; designing a dual extended Kalman filter algorithm to achieve joint online estimation of model parameters and blind zone states, and introducing pseudo-measurements of shock wave positions to form multi-loop feedback; constructing a dual-channel fusion early warning system of model channels and intelligent sensor channels, achieving robust early warning through adaptive weights; and designing a continuous adaptive model predictive controller to dynamically adjust the control strategy based on risk indicators and parameter uncertainties to achieve active protection control.

Description

Technical Field

[0001] This invention belongs to the field of hypersonic vehicle control technology, specifically relating to a protection and control method for hypersonic vehicles that takes into account the inlet not starting. Background Technology

[0002] Hypersonic vehicles are crucial equipment for the future aerospace field. As the core power plant of hypersonic vehicles, the scramjet engine exhibits extremely complex internal flow characteristics, including strong nonlinear phenomena such as shock wave-boundary layer interference, strong coupling between combustion and flow, and unsteady shock train motion. Intake failure is a significant threat to flight safety. When the combustion chamber back pressure is too high, the shock train is pushed into the intake, causing a sharp deterioration or even complete failure of the intake performance, resulting in a sudden drop in engine thrust or engine shutdown, and potentially leading to a flight accident. Traditional control and monitoring technologies face significant challenges in addressing this problem.

[0003] The existing technology has the following main shortcomings:

[0004] First, there is the problem of state estimation caused by model uncertainty. While traditional quasi-one-dimensional mechanistic models are computationally efficient, their empirical parameters (such as wall friction coefficient, shock train length coefficient, and combustion efficiency) are usually calibrated based on specific operating conditions, making them difficult to adapt to a wide range of flight envelopes. The inaccuracy of model parameters and the structural simplification errors introduced by the quasi-one-dimensional assumptions lead to insufficient accuracy in blind zone state estimation. Traditional state estimation methods often employ fixed-parameter Kalman filtering, failing to consider the time-varying characteristics of model parameters. This results in the inability to correct for large model errors or unforeseen operating conditions online, leading to estimation divergence.

[0005] Secondly, the robustness of early warning systems is insufficient. Existing non-activation early warning methods rely excessively on a single information source, either entirely based on mechanistic model predictions or solely on sensor data analysis. Pure model-driven methods may issue false warnings when the model is inaccurate, while pure data-driven methods lack predictive power and suffer from latency. Early warning systems based on a single information source are prone to missed or false alarms when facing complex and ever-changing flight environments.

[0006] Secondly, the adaptive capability of protection control strategies is insufficient. Traditional control methods are mostly based on fixed safety margins and control weights, failing to dynamically adjust the control strategy according to the system state and model uncertainties. When model errors increase or the system enters a dangerous region, fixed control parameters may lead to untimely or overly conservative responses, making it difficult to achieve an optimal balance between performance and safety.

[0007] Existing related patent technologies also have certain limitations:

[0008] The patent "A Method for Controlling the Outlet Flow of an Aircraft Inlet" (CN118051072B) proposes a flow control method based on CFD numerical simulation, which achieves flow matching by dynamically adjusting the outlet back pressure. This method employs a fixed adjustment strategy, and its efficiency and accuracy may decrease when model parameters deviate or when encountering complex operating conditions. This patent focuses on boundary flow control and lacks monitoring methods for the internal flow field state of the inlet, making it difficult to identify abnormal flow field characteristics that may lead to inaction and provide early warnings.

[0009] The patent "A Design Method for a Two-Dimensional Adjustable Inlet" (CN113076610B) proposes a variable geometry inlet design method based on the throat Mach number, which presets geometric adjustment schemes for different operating conditions using empirical formulas. This method is based on nominal operating condition calibration adjustment rules; however, the applicability of the preset scheme decreases when actual flight conditions deviate from the design point. Furthermore, theoretical prediction models are difficult to correct based on actual measurement information when model parameters drift or unforeseen operating conditions are encountered, potentially leading to performance prediction deviations.

[0010] The patent "A Flow Control Method for Improving Inlet Starting Performance Using Depressurization Slits" (CN115544646B) proposes a passive flow control method based on depressurization slits, which improves starting performance by arranging slits on the compression profile. This method optimizes the slit arrangement through offline simulation during the design phase; however, once the slit position and size are determined, they cannot be adjusted, making it a fixed-configuration passive control. When flight conditions change or abnormal flow fields occur, the fixed depressurization scheme is difficult to adapt dynamically and lacks real-time monitoring capabilities of the inlet's operating status, failing to identify and warn of starting risks.

[0011] In summary, the core problems of existing technologies are: the accuracy of state estimation is limited by fixed-parameter models, making it unable to adapt to wide envelopes; the early warning system relies on a single information source, resulting in insufficient robustness; and the protection control strategy lacks an adaptive adjustment mechanism and fails to consider the impact of model uncertainty. Therefore, a novel integrated method is urgently needed that can correct model parameters online, integrate model-driven and data-driven information for robust early warning, and dynamically adjust the control strategy based on real-time quantified uncertainty, achieving closed-loop control for predicting, warning, and actively protecting against automatic risks. Summary of the Invention

[0012] The purpose of this invention is to provide a protection and control method for hypersonic vehicles that do not require inlet start-up. This method utilizes a reduced-order model based on a physical information neural network and an online model correction estimator based on a dual extended Kalman filter to jointly estimate model parameters and blind zone states. Furthermore, it incorporates adaptive fusion of model and sensor channels to achieve robust inlet start-up warning and protection control. Specifically, the method includes: establishing a quasi-one-dimensional nominal mechanism model to describe the flow field changes in the isolation section and combustion chamber; constructing a reduced-order model using a physical information neural network as a real-time proxy for the quasi-one-dimensional model; designing a dual extended Kalman filter algorithm to achieve joint online estimation of model parameters and blind zone states, and introducing a shock wave position pseudo-measurement to form multi-loop feedback; constructing a dual-channel fusion warning system using model and intelligent sensor channels, achieving robust warning through adaptive weights; and designing a continuous adaptive model predictive controller to dynamically adjust the control strategy based on risk indicators and parameter uncertainties to achieve active protection control.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] The protection and control method for hypersonic vehicles that do not require air intake operation is as follows:

[0015] Step (1) Establish a quasi-one-dimensional mechanism model and define measurement constraints

[0016] Define the location of the critical section along the axial direction of the scramjet engine: Let the coordinate along the engine body axis be... The initial contact section position between the aircraft and the incoming flow is defined as... The location of the isolation section inlet section is defined as follows: The location of the cross-section in the middle of the isolation section is The location of the combustion chamber inlet section is .section The corresponding isolation section inlet is the boundary between the compression section and the isolation section; cross-section Corresponding to the combustion chamber inlet, this is the boundary between the isolator section and the combustion chamber. The total geometric length of the isolator section is defined as... The isolation section height is defined as The cross-sectional area of ​​the flow channel is denoted as... The hydraulic diameter is denoted as Both are axial coordinates. The function is determined by the engine geometry.

[0017] Sensor placement is defined as follows: Due to the high-temperature environment of the combustion chamber region, the wall pressure sensor array can only be placed in the front to middle section of the isolation section. The observable area is defined as the interval from the isolation section inlet to the middle of the isolation section, i.e.

[0018]

[0019] in, The observable area. The blind zone is defined as the area from the middle of the isolation section to the combustion chamber inlet, i.e.

[0020]

[0021] in, This is a blind spot. Deploy [equipment] within the observable area. The pressure sensor, the first The axial position of each sensor is denoted as ,in ,and .

[0022] A quasi-one-dimensional nominal mechanism model describing the flow field changes in the isolation section and combustion chamber is established. This model is a multi-modal hybrid model that automatically switches operating modes based on combustion chamber congestion. The gas specific heat ratio is defined as... The Mach number of the airflow is Ma static pressure is Total temperature is Introducing auxiliary variables In the superheated, shock-free mode, the flow field is described by the Rayleigh-Fanno flow governing equations:

[0023]

[0024]

[0025] In the formula, Let be the wall friction coefficient. This set of equations describes the one-dimensional flow evolution under conditions considering area change, friction effect, and heat release.

[0026] In the superheated oblique shock wave mode and the subheated shock wave mode, a shock train exists within the isolation section. The shock train is a complex flow structure composed of a shock system, an expansion wave system, and a boundary layer separation region. The position of the shock train leading edge is defined as... The physical length of the shock train is The pressure at the leading edge of the shock wave is The combustion chamber inlet back pressure is The shock train length was calculated using an empirical model based on the Crocco number.

[0027]

[0028]

[0029] In the formula, To normalize the shock train length, These are empirical constants for the diffusion model. and Cross sections and cross section The Crocco number is a dimensionless parameter characterizing the flow state, determined by the local Mach number and pressure. The shock train length varies with the upstream Mach number, downstream back pressure, and the geometry of the isolation section.

[0030] The pressure distribution within the shock train is described by an empirical fitting equation, which represents the pressure jump caused by the shock system.

[0031]

[0032] In the formula, Let be the normalized axial coordinate within the shock train, with values ​​ranging from [0,1]. Corresponding to the leading edge of the shock train, This corresponds to the combustion chamber inlet. The cubic polynomial form reflects the pressure within the shock train from its leading edge value. Smooth transition to back pressure Spatial distribution characteristics.

[0033] The aforementioned quasi-one-dimensional nominal mechanism model includes several empirical parameters, which are typically calibrated based on specific operating conditions and may deviate from their nominal values ​​over a wide flight envelope. Let the set of parameters to be corrected in the model be denoted as . Key physical parameters affecting flow field prediction include wall friction coefficient, combustion efficiency, and shock wave diffusion coefficient. Uncertainties in these parameters lead to discrepancies between model predictions and actual flow fields, making them the primary targets for online model correction.

[0034] time The observation vector is defined as the set of all sensor measurements within the observable area:

[0035]

[0036] In the formula, For the observation vector, For position time The measured value of static pressure on the wall. Zero-mean Gaussian measurement noise, satisfying , This is the measurement noise covariance matrix. Measurement noise reflects the combined effects of sensor accuracy limitations, signal acquisition errors, and environmental interference.

[0037] The core challenge lies in the fact that the estimator must rely solely on measurements within the observable region to infer critical states and model parameters within the blind zone. Critical states within the blind zone include the combustion chamber inlet back pressure. and shock train length These two states directly determine the operating state of the air intake and the risk of it not starting. Since sensors cannot directly measure the blind zone state, it is necessary to establish a state-space model and observation equations, and use state estimation methods such as extended Kalman filtering to infer the blind zone state from the pressure distribution in the observable area.

[0038] Step (2) Design the line model correction estimator

[0039] Define the blind zone state vector as This includes the combustion chamber inlet back pressure and shock train length. The known input vector is defined as... Including the Mach number of the incoming flow Flight altitude Angle of attack Fuel equivalence ratio and the Mach number at the entrance of the isolation section Measurable or calculable flight status and control inputs.

[0040] The physical significance of this state-space design lies in: parameter vector The uncertainty and time-varying characteristics in the model require online identification and tracking of their true values; state vector This represents blind spot information that cannot be directly measured but is crucial for motion warning; the input vector is known. It provides flow field boundary conditions and control effects. The parameter set design enables the algorithm to automatically focus on the correction of the friction coefficient in the shock-free mode, and naturally unlock the identification of all parameters in the shock mode, ensuring convergence and robustness.

[0041] Quasi-one-dimensional nominal mechanism models involve multimodal determination, piecewise integration, and iterative solutions, resulting in long computation times per calculation and failing to meet the real-time requirements of high-frequency online estimation. A Physics-Informed Neural Network (PINN-ROM) is employed to construct a reduced-order model as a fast proxy for the quasi-one-dimensional nominal mechanism model. The PINN-ROM embeds physical equation residuals into the loss function, ensuring adherence to physical laws while driving data learning. The reduced-order model establishes a direct mapping from flight state and model parameters to blind zone states.

[0042]

[0043] In the formula, This is a reduced-order model of a physical information neural network, with the input being a flight state vector. and the parameter vector to be identified The output is a blind zone state vector. The network is trained using a hybrid loss function, simultaneously optimizing data fitting accuracy and physical equation residuals. After training, the time taken for a single forward propagation is reduced to the millisecond level, and it implicitly learns the multimodal switching logic and nonlinear coupling characteristics of the quasi-one-dimensional model, meeting the real-time requirements of online estimation.

[0044] The measurement model establishes a mapping relationship between the blind zone state estimate, the parameters to be identified, and sensor measurements in the observable area. A piecewise hybrid approach is adopted, automatically switching calculation methods based on the shock train position. The estimated shock train leading edge position is defined as:

[0045]

[0046] In the formula, The model predicts the location of the shock wave front. This is an estimate of the shock train length. For the _th The pressure prediction function of a sensor is defined as follows:

[0047]

[0048] In the formula, For the first Predicted pressure values ​​at each sensor location For the pressure distribution model within the shock wave train, For the Van Noor flow pressure calculation model, This is the estimated back pressure value. This is an estimate of the pressure at the shock wave front. This is an estimate of the friction coefficient. When the sensor is located inside the shock train, the shock train pressure distribution model is used for calculation; when the sensor is located upstream of the shock train, the Van Noor flow model is used. The piecewise nature of this hybrid measurement model causes the sensitivity of the measurement residuals to the state vector and parameter vector to dynamically change with the operating mode, which is the physical basis for the dual extended Kalman filter to decouple parameter and state estimates.

[0049] To fully utilize the spatial information of the sensor array, a pseudo-measurement of the shock train length is introduced based on standard pressure measurement. The position of the shock train leading edge is estimated by the maximum spatial gradient of the pressure array.

[0050]

[0051] In the formula, This represents the position of the shock wave front estimated based on the pressure gradient. The shock wave front corresponds to the peak position of the spatial pressure gradient, providing a direct physical indication of the shock wave's location. Based on the estimated shock wave front position, the estimated shock wave length is:

[0052]

[0053] In the formula, The shock train length is calculated by the sensor. The augmented measurement vector and augmented measurement model are defined as follows:

[0054]

[0055]

[0056] In the formula, To augment the measurement vector, To augment the measurement function, the preceding Each component Let be the pressure prediction function at each sensor location, the th Each component For shock train length prediction, the augmented state-space equation is expressed as:

[0057]

[0058]

[0059]

[0060] In the formula, for The parameter vector at time step, for The state vector at time t, For parameter process noise, satisfy , For state-process noise, satisfying , To enhance measurement noise and meet , The parameter process noise covariance matrix, The noise covariance matrix of the state process is... To augment the measurement noise covariance matrix. Parametric process noise reflects the slow time-varying characteristics of physical parameters, while state process noise reflects model structural errors and unmodeled dynamics.

[0061] The dual extended Kalman filter consists of two stages: a parameter filter and a state filter. The parameter filter updates the parameter estimates using the state estimate from the previous time step and the augmented measurement from the current time step. In the prediction step, the covariance matrix of the predicted parameters and the prior parameters is:

[0062]

[0063]

[0064] In the formula, Based on Time information Prediction of time parameters, for Posterior parameter estimation at time t. for The prior parameter covariance matrix at time t. for The posterior parameter covariance matrix at time step [time]. Parameter prediction assumes that the parameters change slowly between adjacent time steps; the prior covariance tracks the time-varying nature of the parameters by superimposing process noise. In the update step, the Jacobian matrix, augmented measurement residuals, and Kalman gain matrix of the augmented measurement model for the parameters are calculated as follows:

[0065]

[0066]

[0067]

[0068] In the formula, To augment the Jacobian matrix of the measurement model for the parameters, To enhance measurement residuals, Here is the Kalman gain matrix. The parameter update and posterior parameter covariance matrices are:

[0069]

[0070]

[0071] In the formula, To estimate posterior parameters after incorporating augmented measurement information, The covariance matrix of the posterior parameters. This is the identity matrix with the same dimension as the parameter vector. The state filter estimates the blind zone state using the corrected parameters and the augmented measurement at the current time step. In the prediction step, the state prediction and the prior state covariance matrix are:

[0072]

[0073]

[0074] In the formula, To predict the state based on the updated parameters and the current input, Let be the prior covariance matrix of the state. The Jacobian matrix with respect to the parameters of the reduced-order model is defined as follows:

[0075]

[0076] The state prediction covariance takes into account the impact of parameter uncertainties propagating to the state prediction through the reduced-order model. In the update step, the Jacobian matrix of the augmented measurement model with respect to the state, the augmented measurement residuals, and the Kalman gain matrix are calculated as follows:

[0077]

[0078]

[0079]

[0080] In the formula, To augment the Jacobian matrix of the measurement model with respect to the state, To enhance measurement residuals, Let be the Kalman gain matrix. The state update and posterior state covariance matrices are:

[0081]

[0082]

[0083] In the formula, To estimate the posterior state after incorporating augmented measurement information, Let be the posterior state covariance matrix. It is the identity matrix with the same dimension as the state vector.

[0084] The augmented measurement framework provides two independent and complementary information flow paths for extended Kalman filtering. The pressure distribution loop indirectly infers the shock train length based on the curve shape of the pressure distribution within the shock train, fully utilizing the distributed information from multiple sensors and exhibiting strong robustness to single-point measurement noise. The position direct loop directly locates the shock train leading edge based on the peak position of the pressure gradient, independent of the specific functional form of the pressure distribution, providing redundancy constraints based on physical characteristics. The two loops are automatically fused through the Kalman gain matrix, with their weights dynamically adjusted by their respective uncertainties. When structural errors exist in the pressure distribution model, the independent constraints provided by the position direct loop prevent the estimate from diverging. When sensor placement is sparse or local sensor failures occur, the pressure distribution loop utilizes the overall distribution information of the remaining sensors to maintain estimation stability. This redundancy design improves the algorithm's tolerance to modeling uncertainties and hardware imperfections.

[0085] Step (3) Construct a dual-channel fusion early warning system

[0086] The model channel constructs a risk index based on the posterior state estimate of the model correction estimator. The normal shock wave ultimate pressure is calculated from the inlet conditions of the isolation section using the normal shock wave relation:

[0087]

[0088] In the formula, The ultimate pressure of the normal shock wave. cross section static pressure, for k The Mach number at the inlet of the isolator section. This ultimate pressure represents the maximum back pressure that the isolator section inlet can withstand. When the combustion chamber back pressure exceeds this value, the shock train will be pushed out of the isolator section, causing it to stop. The pressure margin and length margin are defined as follows:

[0089]

[0090]

[0091] In the formula, For pressure margin, This is the estimated back pressure value. For length margin, This is an estimate of the shock train length. The pressure margin reflects the relative distance between the combustion chamber back pressure and the inertia limit, while the length margin reflects the relative distance between the leading edge of the shock train and the inlet of the isolation section. The model risk index is defined as:

[0092]

[0093] In the formula, The model risk index ranges from [0,1], with larger values ​​indicating a higher risk of non-startup. A minimum value operator is used to ensure the risk index is dominated by the tightest constraint. The model channel offers predictive and smoothing advantages, anticipating blind zone states and effectively filtering sensor noise; however, its performance depends on the convergence of equivalent parameters and the structural accuracy of the quasi-one-dimensional model.

[0094] The intelligent sensor channel uses only the time series data from the sensors in the observation area, without relying on any model or extended Kalman filter information. It extracts three features in parallel: spatial gradient, pressure ratio, and temporal fluctuation to characterize the shock wave series. The spatial gradient feature quantifies the shock wave steepness.

[0095]

[0096] In the formula, for The maximum spatial pressure gradient at any given time. The peak pressure gradient corresponds to the leading edge of the shock train, and this characteristic increases as the shock train moves upstream. The pressure ratio characteristic quantifies the shock wave location:

[0097]

[0098] In the formula, This represents the pressure ratio between the downstream and upstream sensors. This characteristic reflects the overall pressure rise level in the observable area, increasing as the shock train moves upstream. Temporal fluctuation characteristics quantify shock wave jitter.

[0099]

[0100] In the formula, The root mean square of the short-term fluctuations in pressure at the downstream sensor. The time window length, The average pressure within the window represents the shock train's unsteady motion intensity; shock jitter significantly increases near the critical state of inertia. The sensor risk index is obtained by weighted fusion of the three features after normalization to the [0,1] interval.

[0101]

[0102]

[0103] In the formula, For normalized eigenvalues, These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature within the normal working range, respectively. As a sensor risk indicator, , and The weighting coefficients satisfy the normalization condition. Intelligent sensor channels have the advantages of realism and robustness, being based entirely on actual sensor data and utilizing the spatial and temporal characteristics of all sensors. However, they suffer from the disadvantage of hysteresis, meaning they can only effectively detect risks after the shock train has moved into the observable area.

[0104] The final risk indicator is achieved by merging the two channels using adaptive weighting:

[0105]

[0106] In the formula, As the final risk indicator, For model channel weights, satisfying The model channel weights are adaptively calculated based on the posterior state covariance matrix of the dual extended Kalman filter state filter. The model uncertainty index is defined as:

[0107]

[0108] In the formula, As an indicator of model uncertainty, The trace of the matrix is ​​represented. The trace of the posterior state covariance matrix reflects the overall uncertainty level of the state estimate. The adaptive weights employ an exponential decay function:

[0109]

[0110] In the formula, The reference uncertainty level is a threshold parameter pre-calibrated based on system characteristics. When the extended Kalman filter converges and the measurement residual is small, the posterior state covariance is small, the model uncertainty index is small, the adaptive weights approach 1, and the system mainly trusts the model risk, gaining predictive advantage. When the extended Kalman filter fails to converge, encounters severe disturbances, or the model error is large, the measurement residual increases, the posterior state covariance automatically expands, the model uncertainty index increases, the adaptive weights approach 0, the system automatically degrades, mainly trusts the sensor risk, and ensures robustness.

[0111] A two-tiered threshold early warning mechanism is defined based on the final risk indicator. The first-tier warning is triggered when the risk indicator exceeds the warning threshold:

[0112]

[0113] In the formula, The warning threshold is set at Level 1. A Level 1 warning indicates that the system is approaching the shutdown threshold, requiring operator attention and preparation for protective measures. A Level 2 warning is triggered when a risk indicator exceeds a critical threshold.

[0114]

[0115] In the formula, This is the critical threshold. A level-two judgment indicates that inaction is imminent, and the system should immediately take protective measures. The selection of the threshold should comprehensively consider the warning lead time, false alarm rate, and missed alarm rate. A typical value is... , The specific values ​​can be adjusted according to the characteristics of the aircraft and mission requirements.

[0116] Step (4) Design a continuous adaptive model predictive control protection strategy

[0117] A continuous adaptive model predictive controller is established, whose decision logic is deeply integrated with an online model correction estimator. The controller utilizes posterior parameter estimates and the covariance matrix to continuously and adaptively adjust based on risk indicators. Model predictive control optimizes performance indicators by rolling optimization of the control sequence within a finite time domain, while satisfying safety constraints. The adaptive mechanism is reflected in the dynamic adjustment of weighting coefficients, constraint boundaries, and safety buffers according to the system state and uncertainty level.

[0118] In each control cycle Solve the finite-time optimization problem. Define the prediction time domain length as... The expected fuel equivalence ratio is No margin of failure Performance metrics are defined as follows:

[0119]

[0120] In the formula, For performance indicators, To predict the fuel equivalence ratio, Indicates the prediction time domain from time [time]. to The fuel equivalence ratio sequence, For performance weighting, For smoothness weights. The first term on the right-hand side of the equation. The penalty for a deviation in fuel equivalence ratio from the expected value reflects thrust performance requirements; the second item The penalty rate of change in fuel equivalence ratio reflects control smoothness and actuator protection requirements. Constraints include safety constraints.

[0121]

[0122] In the formula, The predicted shock train length is obtained using a reduced-order model based on a physical information neural network. As a safety buffer zone, To constrain the geometric boundaries. This constraint ensures that the shock train length remains within a safe range while considering parameter uncertainties. The control amplitude and rate of change constraints are as follows:

[0123]

[0124]

[0125] In the formula, and These are the lower and upper limits of the fuel equivalence ratio, determined by the ignition boundary and the fuel-rich boundary, respectively. This is the upper limit of the rate of change of fuel equivalence ratio, determined by the dynamic response capability of the fuel supply system.

[0126] Performance weights are dynamically adjusted based on the margin. The upper and lower thresholds of the margin are defined as follows: and The performance weights for high and low margins are respectively and The performance-weighted scheduling rule is as follows:

[0127]

[0128] In the formula, it is usually set This allows for performance to be prioritized when there is a high margin, while performance requirements are reduced to prioritize safety when there is a low margin. Smoothness-weighted reverse scheduling:

[0129]

[0130] In the formula, and These represent the smoothness weights for high and low margin conditions, respectively. With high margin, performance tracking is prioritized, allowing for larger control variations; with low margin, control smoothness is prioritized to avoid flow field instability caused by drastic fuel changes.

[0131] The constrained geometric boundaries are dynamically tightened based on the margin. The constrained boundaries for high and low margins are defined as follows: and The constraint boundary scheduling rules are as follows:

[0132]

[0133] In the formula, the typical value is , With a high margin, the shock train is allowed to approach the isolator outlet, making full use of the isolator length; with a low margin, the shock train is forced to remain at a more upstream position, leaving a larger safety margin. The safety buffer zone based on parameter uncertainty is calculated as follows: The Jacobian matrix of the reduced-order model is...

[0134]

[0135] In the formula, This represents the second component of the reduced-order model output, namely the shock train length. The shock train length prediction variance is:

[0136]

[0137] In the formula, The variance for shock train length prediction is calculated using linearized propagation of the parameter covariance matrix on a reduced-order model. This variance reflects the impact of parameter uncertainties on shock train length prediction. The safety buffer zone is defined as:

[0138]

[0139] In the formula, As a safety buffer zone, This is the confidence interval coefficient. Typical value. Corresponding to the 99.7% confidence interval, the probability of ensuring that the shock train length remains within the safe boundary within the range of parameter uncertainty is very high. As parameter uncertainty increases, the buffer band thickens, forcing the controller to make more conservative decisions.

[0140] The system automatically switches between different operating modes based on margin and uncertainty. In safety mode, with high margin, weights and constraint boundaries are actively configured, and low parameter uncertainty results in a thin buffer band. The controller prioritizes performance, fully utilizing the isolation section to pursue high thrust. In margin recovery mode, as the margin decreases, weights and boundaries smoothly transition to a conservative configuration, optimization objectives and feasible regions automatically tighten, and controller decisions shift from pursuing performance to risk avoidance. If parameter uncertainty increases simultaneously, the safety buffer band thickens to provide double protection, ensuring safety is maintained even with model errors. This framework, by integrating online model correction and uncertainty quantification information, achieves proactive, continuous, and robust protection against auto-risk, confining the system within safe and transitional regions and preventing it from entering dangerous states.

[0141] The beneficial effects of this invention are:

[0142] 1. Regarding online model correction, a dual extended Kalman filter is used to jointly estimate model parameters and blind zone states, overcoming the limitation of fixed parameters in traditional mechanistic models. The hybrid piecewise characteristics of the measurement model enable the algorithm to automatically adjust the parameter identification strategy according to the working mode, avoiding divergence problems caused by forced identification when observability is insufficient. A multi-loop feedback mechanism is constructed by introducing a pseudo-measurement of shock wave position, which can effectively correct for structural errors in the pressure distribution model using position information. The physical information neural network reduced-order model is trained using a hybrid loss function, ensuring data fitting accuracy while satisfying physical constraints. This transforms the complex multimodal quasi-one-dimensional model into a real-time computational neural network agent, providing a technical foundation for high-frequency online estimation.

[0143] 2. Regarding dual-channel fusion early warning, the system combines the predictive power of the model channel with the realism of the sensor channel, achieving optimal fusion through adaptive weights based on posterior state covariance. When the extended Kalman filter confidence level is high, the system primarily relies on the model channel for predictive advantage; when the confidence level is low, it automatically degrades to the sensor channel to ensure robustness. This mechanism overcomes the problem of missed or false alarms that single-source early warning systems are prone to when facing model uncertainties and complex flight environments.

[0144] 3. In terms of continuous adaptive protection control, proactive protection against automatic risks is achieved by integrating online model correction and uncertainty quantification information. The controller dynamically adjusts and optimizes target weights, constraint boundaries, and safety buffers based on risk indicators. It prioritizes performance when there is a high margin and transitions to a conservative configuration when there is a low margin. The safety buffer is calculated based on the parameter covariance matrix, ensuring that even when parameter uncertainty increases, the increased buffer can force the controller to make conservative decisions, limiting the system to the safe and transitional regions and preventing it from entering a dangerous state. Attached Figure Description

[0145] Figure 1 This is the overall system architecture flowchart;

[0146] Figure 2 This is a schematic diagram of the engine flow path and sensor arrangement;

[0147] Figure 3 This is a flowchart of the dual extended Kalman filter algorithm;

[0148] Figure 4 This is a comparison chart of the effects of online parameter correction;

[0149] Figure 5 It is a time-history diagram of fuel equivalence ratio control;

[0150] Figure 6 It is a time history diagram of the shock train length evolution;

[0151] Figure 7 It is a timeline of no activation margin. Detailed Implementation

[0152] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0153] 1. System Architecture and Measurement Configuration

[0154] Typical parameters were set based on the X-51A hypersonic vehicle. Cruise Mach number 6.0, cruise altitude 25 km. Total length of the isolation section. It is 900 mm high. The length is 50 mm. Eight pressure sensors are arranged from the front to the middle of the isolation section, covering the observable area (450 mm in length). The rear half of the isolation section and the combustion chamber are blind zones (450 mm in length). The nominal parameters of the quasi-one-dimensional model include the wall friction coefficient. Shock wave diffusion coefficient And combustion efficiency. Actual parameters deviate from nominal values ​​by up to 30% due to simulation model mismatch. Vector of parameters to be identified. Include and The initial estimated value is the nominal value. The filter update frequency is 100 Hz, and the parameters, state process noise, and measurement noise are set according to the standard Kalman filtering method.

[0155] The reference uncertainty level, sensor feature weights, and warning threshold of the dual-channel warning system are empirically calibrated. The adaptive model predictive controller performs a 10-step prediction in the time domain, with a control frequency of 10 Hz and a fuel equivalence ratio of... The range is 0.4 to 0.9, with a maximum rate of change of 0.05 per second and a target value of 0.8. The high and low margin thresholds are 0.7 and 0.3, respectively, corresponding to inverse weighting of performance and smoothness; high margin prioritizes performance, while low margin prioritizes safety. The constraint boundaries are 0.95 and 0.85 for high and low margins, respectively. A confidence coefficient of 3 corresponds to a 99.7% confidence interval.

[0156] like Figure 1 As shown, the system architecture consists of five layers. The quasi-one-dimensional mechanistic model layer provides physical constraints, the reduced-order model establishes a fast mapping from input to output, and the dual extended Kalman filter layer implements parameter... and state The joint estimation, the dual-channel early warning layer integrates the model channel and the sensor channel, and the adaptive model predicts the output fuel equivalence ratio of the control layer. Instructions. For example... Figure 2 As shown, the flow path includes the intake duct, isolation section, combustion chamber, and exhaust nozzle. The sensor array (circles distributed along the isolation section) is located in the observable area and the blind zone (…). arrive The section covers the latter half of the isolator and the combustion chamber. The leading edge of the shock wave train (the X-shaped shock wave in the isolator) is located at... Corresponding pressure gradient peak value, length Characterizes the distance from the front edge to the combustion chamber inlet. When back pressure... When it rises Increase Approaching or exceeding Sometimes it fails to start.

[0157] 2. Physical Information Neural Network Order Reduction Model

[0158] A quasi-one-dimensional model takes approximately 50 milliseconds to solve in a single iteration, which is insufficient for real-time estimation at 100 Hz. The Physical Information Neural Network embeds the residuals of the physical equations into the loss function, constructing a model from flight state and model parameters to back pressure. and shock train length The network employs a 3-layer fully connected structure with 8, 32, 16, and 2 neurons in each layer, and uses hyperbolic tangent activation functions. The input includes eight variables: incoming Mach number, flight altitude, angle of attack, fuel equivalence ratio, isolation section inlet Mach number, and parameters to be identified. 10,000 training samples are generated within the flight envelope, and the loss function includes a data fitting term and a physical residual term. After 500 training epochs using the Adam optimizer, the forward propagation time is reduced to 0.8 milliseconds, a 60-fold improvement over the original model. The reduced-order model propagates parameter uncertainties to the state space using the Jacobian matrix.

[0159] 3. Online estimation using dual extended Kalman filter

[0160] like Figure 3As shown, the algorithm employs a cascaded structure of a parameter filter and a state filter, sharing augmented measurement information and processing the parameters separately. Correction and Status Estimation. The parameter filter first updates the prior parameter covariance to reflect the time-varying nature of the parameters, and then calculates the measured predicted values ​​through a hybrid measurement model. The hybrid model automatically switches according to the estimated shock front position: a shock pressure distribution model is used within the shock front, and a Fanno flow model is used upstream. The measurement residuals serve as the driving force for parameter correction, and the Kalman gain automatically balances the reliability of the model predictions and measurement information. The posterior parameter estimates are obtained by adding weighted residuals to the predicted values, and the covariance matrix is ​​updated synchronously to reflect the level of uncertainty.

[0161] The state filter uses the corrected parameters to predict the state through a reduced-order model. The prior state covariance propagates parameter uncertainties and is superimposed with process noise via the Jacobian matrix. The measurement residual calculates the deviation between the current measurement and the predicted measurement, and the Kalman gain is calculated based on the measurement model's sensitivity to the state and its uncertainty. The posterior state estimate and covariance are obtained using the standard Kalman filter update formula.

[0162] The augmented measurement framework introduces a pseudo-measurement of the shock train length, using the pressure gradient peak to locate the leading edge, forming a dual-loop feedback. The pressure distribution loop indirectly infers the length based on the shape of the multi-sensor curves. The direct-loop positioning system directly locates the leading edge based on the gradient peak value. The two loops are automatically fused using Kalman gain, with their weights dynamically adjusted based on their respective uncertainties, providing complementary constraints.

[0163] like Figure 4 As shown, the measurement of residual convergence effect verifies the online correction capability. With fixed parameters (gray line), the residual fluctuates continuously between 0.10 and 0.14, failing to converge. With online correction (black line), the residual rapidly decreases from 0.16, converging to below 0.02 within 5 seconds, a reduction of over 80% compared to the fixed parameter case, verifying the convergence capability of the dual extended Kalman filter when the parameters deviate by 30%.

[0164] 4. Dual-channel fusion early warning system

[0165] The model channel constructs risk indicators based on posterior state estimates. The normal shock wave ultimate pressure represents the maximum back pressure that the isolator inlet can withstand, calculated from the inlet conditions using the normal shock wave equation. The pressure margin reflects the estimated back pressure. The relative distance to the distance limit, and the length margin, reflect the estimated shock train length. The relative distance to the entrance of the isolation section. The model risk index takes the smaller of the two margins and then takes the complement, dominated by the tightest constraint. The model channel has predictive and smooth properties, but depends on parameter convergence and model accuracy.

[0166] The intelligent sensor channel uses only the pressure time series of the observable area to extract three features to characterize the shock wave train. Spatial gradient features quantify the shock wave steepness, obtained by calculating the maximum ratio of the pressure difference between adjacent sensors to their distance. Pressure ratio features quantify the shock wave position, defined as the pressure ratio between the downstream and upstream sensors. Temporal fluctuation features quantify the shock wave jitter, defined as the root mean square of the short-term pressure fluctuations at the downstream sensor. The sensor risk index is obtained by weighted fusion after normalization of these three features. The sensor channel exhibits realism and robustness but suffers from hysteresis, only able to detect shock waves after they enter the observable area.

[0167] The final risk indicator is fused from two channels using adaptive weights. The weights are dynamically adjusted based on model uncertainty, defined as the square root of the trace of the posterior state covariance matrix. When the covariance is small, the weights approach 1, indicating that the model channel is trusted to provide predictive power; when the covariance is large, the weights approach 0, downgrading to the sensor channel to ensure robustness. A two-tiered threshold warning system is implemented: a risk indicator exceeding 0.6 triggers a level 1 warning, and exceeding 0.9 triggers a level 2 critical judgment.

[0168] 5. Adaptive Model Predictive Control Protection Strategy

[0169] The controller performs rolling optimization on a finite-time problem, predicting 10 steps in the time domain. Performance metrics include performance and smoothing terms; the performance term penalizes deviations of the fuel equivalence ratio from the target value, while the smoothing term penalizes the rate of change. Constraints include safety constraints, amplitude constraints, and rate of change constraints. Safety constraints ensure that the shock train length remains within a safe range after considering parameter uncertainties, while amplitude constraints limit... Between 0.4 and 0.9, the rate of change constraint limits the change to no more than 0.05 per second.

[0170] A triple adaptive mechanism achieves a balance between security and performance. Weighted scheduling is based on margin. SM Dynamic adjustments are made: performance weights are set to 100 for high margins, prioritizing performance; and 1 for low margins, prioritizing safety. Smoothness weights are reversed. Boundary adjustment sets the constraint boundary to 0.95 for high margins, allowing the shock train to approach the exit; and tightens it to 0.85 for low margins, providing a larger safety margin. Robustness of the buffer band is achieved by calculating the shock train length prediction variance using the Jacobian matrix. The safety buffer band is set to three times the prediction standard deviation, corresponding to a 99.7% confidence interval. The buffer band automatically thickens as parameter uncertainty increases, forcing conservative decision-making.

[0171] The system automatically switches operating modes based on margin and uncertainty. In safety mode, weights and boundaries are aggressively configured with a thin buffer band, prioritizing performance. In margin recovery mode, weights and boundaries smoothly transition to a conservative configuration, and the optimization objective and feasible region are automatically tightened. When parameter uncertainty increases, the buffer band provides dual protection, ensuring safety even when the model has errors.

[0172] 6. Simulation Verification

[0173] The simulation scenario involves a sudden and abrupt disturbance during cruise, leading to increased back pressure, forward movement of the shock train, and increased risk. Upon identifying the risk, the control system proactively reduces fuel and pressure to push the shock train back. Once the risk is mitigated, fuel is gradually restored, eventually returning to normal operation.

[0174] like Figure 5 As shown, fuel equivalence ratio The timeline shows the adaptive adjustment process. 0 to 10 seconds. The value has stabilized at the target value of 0.8, and no risk warning has been triggered. If the risk exceeds the 0.6 threshold within 15 seconds... The price rapidly dropped to a low of 0.48, a decrease of approximately 40%, with the active protection phase lasting up to 25 seconds. A margin of 25 to 40 seconds was then established for recovery. Ascend at a gentle slope. After 40 seconds Reverted to 0.8. This timeline reflects the adaptive shift from a performance-oriented to a security-oriented approach.

[0175] like Figure 6 As shown, shock train length The timeline displays the dynamic response to the blind spot status. 0 to 10 seconds. Fluctuations between 500 and 520 mm. A 10- to 15-second perturbation results in... The rainfall increased sharply to 850 mm, an increase of over 60%, approaching the first-level warning line of 855 mm, but not exceeding the total length of 900 mm. Fuel reduction takes effect in 15 to 25 seconds. Rapid pullback. 25 to 40 seconds. The temperature dropped slowly and stabilized, reaching approximately 500 mm after 40 seconds. The peak value did not exceed the total length, verifying the effectiveness of the protection.

[0176] like Figure 7 As shown, margin SM The timeline displays real-time changes in safety metrics. From 0 to 10 seconds, the SM remains stable between 0.76 and 0.79, indicating a safe mode. After 15 seconds, it drops below the 0.7 threshold, switching to margin recovery mode. From 15 to 20 seconds... SM It dropped to a low of 0.34, close to the low margin threshold of 0.3, but did not enter the danger zone. (25 to 35 seconds) SM The value rose smoothly from 0.52 to 0.70. After 35 seconds, it exceeded 0.7 again, switching back to safe mode, and stabilized at the initial state after 40 seconds. The minimum margin of 0.34 remained above the 0.3 threshold, indicating that the protection strategy successfully controlled the risk.

[0177] Comprehensive analysis of the three charts: 15 seconds SM Falling below 0.7 When approaching the warning line Descend immediately, with timely warning response. 15 to 25 seconds. Reduced to the minimum, pullback SM The descent stopped and resumed, with the three curves showing a highly synchronized causal relationship, verifying the effective regulation of the blind zone state by the control. A gradual strategy was adopted from 25 to 40 seconds to avoid secondary risks. Verification results show that the dual extended Kalman filter converges within 5 seconds when the parameters deviate by 30%, reducing the residual by more than 80%; the dual-channel early warning is triggered in a timely manner, and the minimum margin of 0.34 remains above the threshold; the adaptive model predictive control achieves a maximum fuel reduction of 40%, with peak... The total length was 850 mm, not exceeding 900 mm, and it did not stop during the entire process.

Claims

1. A method for protecting a hypersonic vehicle against inlet unstart, characterized in that, The steps are as follows: Step (1) Establish a quasi-one-dimensional mechanism model and define measurement constraints; Step (2) Design a line model correction estimator; Step (3) Construct a dual-channel fusion early warning system; Step (4) Design a continuous adaptive model predictive control protection strategy; Step (2) is as follows: The blind zone state vector is defined as , containing the combustor inlet back pressure and the shock train length; the known input vector is defined as , containing the flight conditions and control inputs of the incoming Mach number , the flight altitude , the angle of attack , the fuel equivalence ratio , and the isolator inlet Mach number . Parameter vector to be identified Representing the uncertainty and time-varying characteristics in the model, the true values need to be tracked by online identification; state vector represents blind zone information that cannot be directly measured but is crucial for the no-start warning; known input vector provides flow field boundary conditions and control action; A reduced-order model PINN-ROM is constructed using a physical information neural network as a fast proxy for the quasi-one-dimensional nominal mechanism model; the reduced-order model establishes a direct mapping from the flight state and model parameters to the blind area state: ； In the formula, is a reduced order model of physical information neural network, the input is the flight state vector and the parameter vector to be identified , the output is the blind area state vector ; the network is trained through a hybrid loss function, and the data fitting accuracy and the physical equation residual are optimized simultaneously; The measurement model establishes the mapping relationship between the estimated blind area state, the to-be-identified parameters and the observable area sensor measurements; a segmented mixed form is used, and the calculation method is automatically switched according to the shock wave string position; the estimated shock wave string front position is defined as: ； wherein is the model predicted shock train front position, is the shock train length estimate; for the th sensor, the pressure prediction function is defined as: ； wherein Ppred is the pressure prediction value for the Pshock is the shock train pressure distribution model, Pvan is the van der waals flow pressure calculation model, Pback is the back pressure estimate, Pfront is the shock train front pressure estimate, Cf is the friction coefficient estimate; calculated using the shock train pressure distribution model when the sensor is located inside the shock train; calculated using the van der waals model when the sensor is located upstream of the shock train.​ To make full use of the spatial information of the sensor array, a shock wave string length pseudo-measurement is introduced based on standard pressure measurement; the shock wave string front position is estimated by the maximum spatial gradient of the pressure array: ； In the formula, is the position of the shock train front based on the pressure gradient estimation; based on the estimated position of the shock train front, the shock train length is calculated as: ； wherein is the sensor-estimated shock train length; define the augmented measurement vector and the augmented measurement model as: ； ； where is the augmented measurement vector, is the augmented measurement function, where is the first component is the pressure prediction function for each sensor location, where is the first component is the shock train length prediction; the augmented state space equation is expressed as: ； ； ； wherein is the parameter vector at time instant, is the state vector at time instant, is the parameter process noise, satisfying , is the state process noise, satisfying , is the augmented measurement noise, satisfying , is the parameter process noise covariance matrix, is the state process noise covariance matrix, is the augmented measurement noise covariance matrix; The dual extended Kalman filter includes a parameter filter and a state filter; the parameter filter updates the parameter estimate using the state estimate at the previous time and the augmented measurement at the current time; in the prediction step, the parameter prediction value and the prior parameter covariance matrix are: ； ； wherein is based on the time information pair the prediction of the time parameter, is the a posteriori parameter estimate of the time, is the a priori parameter covariance matrix of the time, is the a posteriori parameter covariance matrix of the time; the parameter prediction assumes that the parameter changes slowly from one time to the next, and the a priori covariance enables tracking of the time variation of the parameter by superimposing process noise; in the update step, the Jacobian matrix of the augmented measurement model with respect to the parameter, the augmented measurement residual, and the Kalman gain matrix are calculated as follows: ； ； ； wherein is the Jacobian matrix of the augmented measurement model with respect to the parameters, is the augmented measurement residual, is the Kalman gain matrix; the parameter update and the posterior parameter covariance matrix are: ； ； wherein is the posterior parameter estimate after fusion of the augmented measurement information, is the posterior parameter covariance matrix, is an identity matrix of the same dimension as the parameter vector; the state filter estimates the blind state using the corrected parameter and the augmented measurement at the current time; in the prediction step, the state prediction value and the prior state covariance matrix are: ； ； wherein is a prediction of the state based on the updated parameters and the current input, is a prior covariance matrix of the state, is the Jacobian of the reduced model output with respect to the parameters, defined as ； The state prediction covariance considers the influence of parameter uncertainty propagated to the state prediction through the reduced-order model; in the update step, the augmented measurement model is calculated to obtain the state Jacobian matrix, the augmented measurement residual and the Kalman gain matrix: ； ； ； wherein is the Jacobian matrix of the augmented measurement model with respect to the state, is the augmented measurement residual, is the Kalman gain matrix; the state update and the posterior state covariance matrix are: ； ； wherein is the augmented state estimate after fusion of the measurement information, is the posterior state covariance matrix, is the identity matrix of the same dimension as the state vector. Step (3) is as follows: The model channel constructs a risk index based on the posterior state estimate of the model correction estimator; the positive shock limit pressure is calculated from the isolator inlet condition through the positive shock relationship: ； wherein is the normal shock limit pressure, is the cross-sectional static pressure, is k the Mach number at the inlet of the isolation section at the time instant; the pressure margin and the length margin are defined as: ； ； wherein is the pressure margin, is the back pressure estimate, is the length margin, is the shock train length estimate; the model risk indicator is defined as: ； In the formula, is a model risk indicator, with a value range of [0, 1]; The minimum value operator is used to ensure that the risk index is dominated by the tightest constraint; The intelligent sensor channel only uses the time series of the observation area sensors and does not rely on any model or extended Kalman filter information; three feature fingerprints are extracted in parallel to characterize the shock wave string: spatial gradient features, pressure ratio features and time fluctuation features; the spatial gradient feature quantifies the shock steepness: ； wherein is maximum pressure spatial gradient at the moment; pressure ratio characterizes shock position: ； wherein is the pressure ratio of the most downstream sensor to the most upstream sensor; time fluctuation characterizes the shock oscillation: ； wherein RMS is the root mean square of short term fluctuations of the pressure at the most downstream sensor, T is the length of the time window, P is the average pressure within the window. After normalization to the [0, 1] interval, the three features are weighted and fused to obtain the sensor risk index: ； ； wherein, is a normalized characteristic value, is an original characteristic value, and are minimum and maximum values of the characteristic in a normal operating range, respectively, is a sensor risk index, , and are weight coefficients satisfying a normalization condition ; The final risk index is fused by adaptive weight: ； wherein is the final risk indicator, is the model channel weight, satisfying The model channel weight is adaptively calculated based on the posterior state covariance matrix of the dual extended Kalman filter state filter. The model uncertainty indicator is defined as: ； wherein is a model uncertainty indicator, denotes the trace of a matrix; The adaptive weight uses an exponential decay function: ； In the formula, is a reference uncertainty level, which is a threshold parameter pre-calibrated according to system characteristics; Based on the final risk index, a two-level threshold early warning mechanism is defined.

2. The method of claim 1, wherein, Step (1) is as follows: The key axial cross-section positions of the scramjet engine are defined as follows: the coordinate along the engine body axis is , the position of the starting cross-section of the aircraft contacting the incoming flow is , the position of the inlet cross-section of the isolator is , the position of the middle cross-section of the isolator is , the position of the inlet cross-section of the combustion chamber is ; the cross-section corresponds to the inlet of the isolator, which is the junction of the compression section and the isolator; the cross-section corresponds to the inlet of the combustion chamber, which is the junction of the isolator and the combustion chamber; the total length of the isolator is defined as , and the height of the isolator is defined as ; the cross-sectional area of the flow passage is denoted as , and the hydraulic diameter is denoted as , both of which are functions of the axial coordinate and are determined by the engine geometry. Define the sensor arrangement position: the observable region is defined as the interval from the isolator inlet to the middle of the isolator, i.e. ； wherein is the observable region; the blind region is defined as the interval from the middle of the isolation section to the entrance of the combustion chamber, i.e. ； wherein blind area; arranged within the observable area pressure sensor, the first axial position of the first sensor is denoted as wherein , and ; A quasi-one-dimensional nominal mechanism model is established to describe the variation of the flow field in the isolation section and the combustion chamber: the model is a multi-modal mixing model, which automatically switches the working mode according to the choking condition of the combustion chamber; the specific heat ratio of the gas is defined as , the Mach number of the airflow is Ma , the static pressure is , and the total temperature is ; an auxiliary variable is introduced; in the super-combustion shock-free mode, the flow field is described by the Rayleigh-Fanno flow control equation set: ； ； In the formula, is the wall friction coefficient; In the supercritical oblique shock mode and the subcritical mode, there are shock trains in the isolator; the shock train front position is defined as , the shock train physical length is , the shock train front pressure is , and the combustion chamber inlet back pressure is ; the shock train length is calculated by an empirical model based on the Crocco number: ； ； where is the normalized shock train length, is the shock train spread coefficient, which is an empirical constant for the spread model, and are the Crocco numbers for the cross-section and the cross-section respectively; the Crocco number is a dimensionless parameter that characterizes the flow state and is determined by the local Mach number and pressure; The pressure distribution within the shock wave string is described by an empirical fitting relationship to describe the pressure jump process caused by the shock wave system: ； wherein is the normalized axial coordinate within the shock train, which takes values in the range [0, 1], corresponds to the shock train front, corresponds to the combustion chamber inlet; The wall friction coefficient, the combustion efficiency and the shock train diffusion coefficient are denoted as the parameter vector to be identified ; Time The observation vector at time t is defined as the collection of all sensor measurements within the observable region: ； wherein is the observation vector, is the position is the time is the wall static pressure measurement, is the zero-mean Gaussian measurement noise satisfying , is the measurement noise covariance matrix.

3. The method of claim 1, wherein, In step (2): The augmented measurement framework provides two independent and complementary information flow paths for the extended Kalman filter; the pressure distribution loop indirectly infers the shock train length based on the curve shape of the pressure distribution within the shock train, and has strong robustness to single-point measurement noise using the distributed information of multiple sensors; the position direct loop directly locates the shock train front based on the position of the pressure gradient peak, and does not depend on the specific function form of the pressure distribution, providing a redundant constraint based on physical characteristics; the two loops are automatically fused through the Kalman gain matrix, and the weights are dynamically adjusted by the respective uncertainties; when there is a structural error in the pressure distribution model, the independent constraint provided by the position direct loop prevents the estimated value from diverging; when the sensor arrangement is sparse or local sensor failure occurs, the pressure distribution loop uses the overall distributed information of the remaining sensors to maintain the estimation stability.

4. The method of claim 1, wherein, The two-level threshold early warning mechanism based on the final risk index in step (3) is defined as follows: The first-level warning is triggered when the risk index exceeds the warning threshold: ； In the formula, is a warning threshold; a first level of warning indicates that the system is approaching the non-starting boundary and should attract the attention of the operator and prepare for protective measures; The second-level warning is triggered when the risk index exceeds the critical threshold: ； In the formula, is a critical threshold; the secondary determination indicates that the start is about to occur, and the system should immediately take protective measures; the selection of the threshold should consider the early warning lead time, false alarm rate and false negative rate.

5. The method of claim 1, wherein, Step (4) is as follows: At each control cycle , a finite horizon optimization problem is solved; define the prediction horizon length as , the desired fuel equivalence ratio as , the no-start margin as ; the performance index is defined as: ； wherein is a performance index, is a predicted fuel equivalence ratio, denotes a sequence of fuel equivalence ratios from time to in the prediction horizon, is a performance weight, is a smoothness weight; the first term on the right-hand side penalizes the fuel equivalence ratio from the desired value, reflecting the thrust performance requirement; the second term penalizes the rate of change of the fuel equivalence ratio, reflecting the control smoothness and actuator protection requirement; the constraint conditions include safety constraints: ； wherein is the shock train length prediction value, predicted by the physical information neural network reduced order model, is the safety buffer, is the constraint geometry boundary; control amplitude constraint and rate of change constraint are: ； ； wherein and are lower and upper limits of fuel equivalence ratio, respectively, determined by the ignition boundary and the rich boundary, is an upper limit of fuel equivalence ratio change rate, determined by the dynamic response capability of the fuel supply system; The performance weight is dynamically adjusted according to the margin; the upper and lower thresholds of the margin are defined as and The performance weights when the margin is high and low are and respectively; and the performance weight scheduling rule is: ； In the formula, the setting such that at high margins performance is pursued preferentially, and at low margins performance requirements are reduced to give way to safety; smoothness weight reverse scheduling: ； wherein and are the smoothness weights for high and low margins, respectively; high margin prioritizes tracking performance, allowing larger control changes; low margin prioritizes control smoothness, avoiding flow field instability from large fuel changes. The constraint geometry boundary is dynamically tightened according to the margin; the constraint boundaries when the high margin and the low margin are defined are respectively and ; the constraint boundary scheduling rule is that ； When the margin is high, the shock train is allowed to approach the outlet of the isolation section, making full use of the length of the isolation section; when the margin is low, the shock train is forced to remain in a more upstream position, leaving a larger safety margin; the safety buffer based on parameter uncertainty is calculated as follows: the reduced-order model Jacobian matrix is ； where denotes the second component of the reduced order model output, i.e. the shock train length; the shock train length prediction variance is ； wherein is the shock train length prediction variance; the safety buffer band is defined as: wherein is a safety buffer, is a confidence interval coefficient.

6. The method of claim 4, wherein, The system automatically switches between different working modes according to the margin and uncertainty; in the safety mode, the weight and constraint boundary are in an active configuration when the margin is high, and the parameter uncertainty is small, resulting in a thin buffer; the controller is performance-oriented, making full use of the isolation section's ability to pursue high thrust; in the margin recovery mode, the weight and boundary transition smoothly to a conservative configuration when the margin decreases, and the optimization objective and feasible region are automatically tightened; if the parameter uncertainty also increases, the safety buffer provides double protection, ensuring that the model can still maintain safety even if there is an error.

Images