A method and system for intelligent control of fuel multi-point injection of a scramjet engine

Abstract

The application discloses a kind of super-combustion ramjet engine fuel multi-point injection intelligent control method and system, method includes: using multi-core parallel computing architecture to build simulation model, constructs multi-core parallel computing one-dimensional model;Safety constraints are integrated into the learning process, to build Safe SAC agent considering safety constraints;Carry out the pre-training of reinforcement learning multi-point intelligent injection based on the interaction of multi-core parallel computing one-dimensional model, obtain the pre-trained Safe SAC reinforcement learning agent;Design the indirect control law of thrust based on wall pressure: calculate the engine thrust according to wall pressure;Pressure sensor is deployed on the wall of isolation section, and the position of shock train front edge in the isolation section is calculated using pressure ratio method;Engine thrust and shock train front edge position are input into pre-trained Safe SAC reinforcement learning agent, and the fuel flow of multi-injection point is controlled.The application realizes the intelligent regulation and control of combustion chamber fuel multi-point injection under the safety constraint condition of aircraft.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent control of combustion performance of scramjet, and particularly relates to an intelligent control method and system for fuel multi-point injection of a scramjet. BACKGROUND

[0002] As the core propulsion system of hypersonic flight, the performance advantage of scramjet is derived from its unique atmospheric air-breathing combustion mode. Compared with traditional rocket power, it can realize revolutionary specific impulse leap and has become a strategic commanding point for aerospace technology development. The combustion chamber, as the energy conversion center of the power device, its combustion stability and heat release intensity directly determine the thrust output efficiency of the engine. A slight improvement in energy release efficiency in the combustion chamber can significantly increase the specific impulse of the whole machine. This strong correlation highlights the core position of the combustion chamber in the power system.

[0003] To overcome the mixing difficulty caused by the extremely short residence time of fuel in the supersonic combustion flow field (usually only milliseconds), scramjet generally adopts a multi-point fuel injection architecture. By arranging multiple independently controlled fuel injection points in the axial direction of the combustion chamber, a gradient equivalence ratio distribution is constructed in the flow field, and the interaction between shock waves and flames is used to enhance the intensity of turbulent mixing, which can greatly improve the overall heat release uniformity of the combustion chamber.

[0004] However, this distributed injection strategy, while improving performance, also introduces high-dimensional strong nonlinear coupling problems: the pressure disturbance caused by the flow regulation of each injection point will be quickly transmitted through the supersonic flow field, causing complex fluid dynamics interference between the combustion chamber and the upstream isolation section. The traditional PID control strategy is limited by its linear control nature and is difficult to handle the deep coupling characteristics of the multi-input multi-output system. During the wide operating condition of the engine, local optimization of the control command often leads to global flow field instability, and when the combustion chamber pressure fluctuation is transmitted reversely to the inlet, it may cause a catastrophic inlet unstart phenomenon.

[0005] The emerging unconstrained reinforcement learning control has the potential to handle complex systems, but its exploration mechanism has fundamental defects in the context of clear safety constraints. For example, the agent may generate fuel injection actions that cause the combustion chamber pressure to rise sharply during the training process, inducing the shock wave string in the isolation section to move sharply; or produce fuel injection that is too lean, causing the thrust to decline sharply. More importantly, existing data-driven methods are difficult to adapt to the performance degradation of the engine in the long-term work (such as the drift of the flow coefficient caused by the coking of the injector), leading to a continuous deterioration of control accuracy with the increase of task time.

[0006] Furthermore, the strong vibrations and electromagnetic interference in hypersonic flight environments significantly increase sensor measurement noise, making traditional control algorithms based on precise models prone to misjudgments and further exacerbating the difficulty of fuel injection control. Therefore, developing a multi-point injection intelligent control method and system that can balance control accuracy, system stability, and operational safety has become a core technological bottleneck that urgently needs to be overcome in the practical application of scramjet engines. Summary of the Invention

[0007] To address the aforementioned shortcomings in the prior art, the present invention provides an intelligent control method and system for multi-point fuel injection in scramjet engines, which solves the problem of the difficulty in accurately controlling fuel injection in scramjet engines using existing methods.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a smart control method for multi-point fuel injection in a scramjet engine, comprising the following steps:

[0009] S1. A simulation model is built using a multi-core parallel computing architecture. While retaining the core coupling relationship between fuel injection, combustion heat release, and thrust output, a one-dimensional multi-core parallel computing model for controlling scramjet engines is constructed.

[0010] S2. In the SAC reinforcement learning framework, safety constraints are integrated into the learning process to construct a Safe SAC agent that considers safety constraints.

[0011] S3. Based on the constructed Safe SAC agent, with multi-point fuel flow as the control variable, carry out reinforcement learning multi-point intelligent injection pre-training based on multi-core parallel computing one-dimensional model interaction to obtain the pre-trained Safe SAC reinforcement learning agent.

[0012] S4. Design a thrust indirect control law based on wall pressure: Establish a nonlinear mapping model between the rear wall pressure of the combustion chamber and the engine thrust through ground tests, and calculate the engine thrust based on the wall pressure;

[0013] S5. Deploy pressure sensors on the wall of the isolation section and use the pressure ratio method to calculate the position of the leading edge of the shock train in the isolation section;

[0014] S6. Build a ground-based direct-connection test rig covering specified inflow conditions and high-precision fuel supply. Input the engine thrust and shock wave leading edge position into a pre-trained Safe SAC reinforcement learning agent to control the fuel flow at multiple injection points.

[0015] Furthermore: In S1, the one-dimensional model of multi-core parallel computing includes a first serial region, a parallel region, and a second serial region;

[0016] The specific steps for initializing the first serial region are as follows:

[0017] A1, according to the selected chemical kinetics model, set the quasi-one-dimensional Euler equation dimension, read the basic parameters required for chemical reaction calculation, set the maximum iteration step number and inflow condition;

[0018] A2, using the finite volume method, according to the corresponding engine type, draw the grid required for subsequent calculation;

[0019] A3, the fuel is added to simulate the exponential distribution: for four oil injection points, the one-dimensional distribution of fuel is simulated using an exponential distribution, and the sum of the one-dimensional distributions of the four oil injection points is used as the total mass of the fuel at the current grid point;

[0020] A4, initialize the conservation matrix Q and its differential form dQ matrix;

[0021] A5, set the discrete time and the update rate of the conservation matrix Q;

[0022] The parallel region continuously iterates and updates the conservation matrix Q through forward scanning, backward scanning and conservation matrix Q update formula. After each update, a residual error is calculated. When the residual error is less than the convergence threshold or the iteration reaches the maximum number of steps, the calculation is ended and the final conservation matrix Q is output.

[0023] The second serial region obtains the performance parameters of the combustion chamber according to the final conservation matrix Q, outputs the engine thrust and the position of the shock wave string front edge, and completes the construction of the multi-core parallel calculation one-dimensional model of the controlled scramjet engine.

[0024] Further, in S2, the Safe SAC agent includes a Critic network, a Safety network and an Actor network, the Critic network and the Safety network use TD error to construct a loss function gradient descent to update parameters; the Actor network introduces a Lagrange multiplier, uses the Lagrange relaxation method to convert the original constrained optimization problem into an unconstrained optimization problem, for the unconstrained min-max problem, the primal-dual optimization uses an iterative primal-dual method, in each iteration, the original strategy and the dual variable are updated in turn, and the updating method is specifically: fixing the Lagrange multiplier, updating the strategy of the Actor network by using the strategy gradient ascent, fixing the strategy parameters of the Actor network, and updating the dual variable.

[0025] Further, S3 is specifically:

[0026] The multi-core parallel computing one-dimensional model is closed-loop interacted with the Safe SAC agent, fuel flow of four fuel injection points output by the Safe SAC agent is input to the multi-core parallel computing one-dimensional model, engine thrust and shock train front edge position output by the multi-core parallel computing one-dimensional model under the current working condition are input to the Safe SAC agent, and the multi-point intelligent injection pre-training is completed through the closed-loop interaction process.

[0027] Further, S4 is specifically:

[0028] According to the direct-connected ground test or numerical simulation of the scramjet engine, the axial position interval corresponding to the monotonic wall pressure and engine thrust is determined, the wall pressure sensor is arranged in the interval, the direct-connected test or numerical simulation under different fuel injection equivalence ratios is carried out, the discrete wall pressure-thrust sample data is collected, the continuous mapping relationship from the wall pressure to the engine thrust is established by using the Gaussian process regression, the nonlinear mapping model of the combustion chamber rear wall pressure and the engine thrust is obtained, and the engine thrust is calculated by using the wall pressure fed back by the wall pressure sensor.

[0029] Further, S5 is specifically:

[0030] The pressure sensor is arranged on the wall of the isolation section, the upper and lower wall pressure sensors at the inlet of the isolation section and the upper and lower wall pressure sensors arranged at equal intervals in the middle and rear part of the isolation section are used to form a feedback array, and the shock train front edge position in the isolation section is calculated by using the pressure ratio method.

[0031] Further, S6 is specifically:

[0032] A ground direct-connected test bench including an isolation section, a combustion chamber, a four-channel independent fuel supply system and corresponding sensors is built, under the determined incoming flow condition, the fuel flow of four fuel injection points is controlled by the pre-trained Safe SAC reinforcement learning agent under the feedback of different shock train front edge positions and engine thrusts.

[0033] A fuel multi-point injection intelligent control system of a scramjet engine, comprising a main controller, an execution / feedback module and a scramjet engine;

[0034] The main controller comprises a reinforcement learning agent unit and an input preprocessing unit, the input preprocessing unit is used for receiving discrete wall pressure data collected by the pressure sensor arranged on the inner wall of the isolation section, and calculating a real-time estimated value of the shock train front edge position in the current isolation section;

[0035] The execution / feedback module comprises a fuel control module, a pressure monitoring module and a driving circuit; the fuel control module comprises four fuel valves, which are respectively installed in independent fuel supply pipelines connecting the oil tank and four oil injection points in the combustion chamber, the driving circuit is used for matching the voltage and current requirements between the output of the reinforcement learning agent unit and the input of the fuel valve, the fuel valve receives the output electric control signal of the reinforcement learning agent unit amplified by the driving circuit, and converts it into a corresponding valve opening adjustment signal, so that the fuel flow of the oil injection point is controlled by adjusting the valve opening of each valve; the pressure monitoring module comprises an isolation section wall surface pressure sensor array and a combustion chamber wall surface pressure sensor, the isolation section wall surface pressure sensor array specifically comprises: a pair of sensors symmetrically arranged on the upper and lower walls of the inlet cross section of the isolation section, and a plurality of groups of sensors symmetrically arranged on the upper and lower walls of the middle and rear regions of the isolation section along the flow direction, which are used for collecting discrete wall surface pressure data, and the combustion chamber wall surface pressure sensor is located at the lower wall of the rear part of the combustion chamber.

[0036] The scramjet engine comprises four oil injection points, which are respectively arranged at specific positions on the upper and lower walls of the combustion chamber.

[0037] The beneficial effects of the present application are:

[0038] (1) The present application provides a kind of supercritical ramjet engine fuel multi-point injection intelligent control method and system, uses multiple injection points to carry out fuel injection, realizes the reasonable control of the fuel flow of multiple injection points under the feedback of different shock wave string front edge positions and thrust size by pre-training Safe SAC reinforcement learning agent, can effectively avoid that fuel forms local rich oil in combustion chamber, fuel is easier to ignite and burn, and combustion intensity is higher. Under the condition of consuming the same fuel, greater thrust benefit can be obtained by multi-point injection.

[0039] (2) In the SAC reinforcement learning framework of Actor-Critic architecture, the safety network is introduced to prospectively evaluate the action, the safety constraint is integrated into the learning process, the Safe SAC agent considering multiple safety state constraints is constructed, so as to realize the intelligent regulation and control of the fuel multi-point injection of the combustion chamber under the safety constraint condition of the aircraft.

[0040] (3) The present application uses pressure ratio method to complete the identification of shock wave string front edge position in isolation section, controls the front edge position measurement error within 5% of the length of isolation section, provides real-time state perception for shock wave string stable control, and realizes accurate measurement of shock wave string front edge position. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 It is a supercritical ramjet engine fuel multi-point injection intelligent control method flow chart of the present application.

[0042] Figure 2 It is a parallel computing interaction diagram of multi-core parallel computing one-dimensional model.

[0043] Figure 3 This is a structural diagram of the Safe SAC agent.

[0044] Figure 4 This is a schematic diagram for solving the position of the leading edge of the shock train using the pressure ratio method.

[0045] Figure 5 This is a schematic diagram of a multi-point fuel injection intelligent control system for a scramjet engine according to the present invention.

[0046] Figure 6 This is a control block diagram of an intelligent control system for multi-point fuel injection in a scramjet engine according to the present invention. Detailed Implementation

[0047] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0048] like Figure 1 As shown, in one embodiment of the present invention, a smart control method for multi-point fuel injection in a scramjet engine includes the following steps:

[0049] S1. A simulation model is built using a multi-core parallel computing architecture. While retaining the core coupling relationship between fuel injection, combustion heat release, and thrust output, a one-dimensional multi-core parallel computing model for controlling scramjet engines is constructed.

[0050] S2. In the SAC reinforcement learning framework, safety constraints are integrated into the learning process to construct a Safe SAC agent that considers safety constraints.

[0051] S3. Based on the constructed Safe SAC agent, with multi-point fuel flow as the control variable, carry out reinforcement learning multi-point intelligent injection pre-training based on multi-core parallel computing one-dimensional model interaction to obtain the pre-trained Safe SAC reinforcement learning agent.

[0052] S4. Design a thrust indirect control law based on wall pressure: Establish a nonlinear mapping model between the rear wall pressure of the combustion chamber and the engine thrust through ground tests, and calculate the engine thrust based on the wall pressure;

[0053] S5. Deploy pressure sensors on the wall of the isolation section and use the pressure ratio method to calculate the position of the leading edge of the shock train in the isolation section;

[0054] S6, build a ground direct connection test bench covering the specified inflow conditions and high-precision fuel supply, input the engine thrust and shock wave train front edge position into the pre-trained Safe SAC reinforcement learning agent, and control the fuel flow of multiple injection points.

[0055] In S1, based on the predefined inflow conditions and engine configuration, the computational domain is spatially discretized using the finite volume method (FVM), the generalized quasi-one-dimensional Euler equations with chemical reaction source terms are solved by implicit format strong coupling to simulate the combustion chamber flow field, and the key parameters of engine thrust are derived according to the flow field solution. To meet the millisecond-level control cycle requirement, the total calculation domain is divided into several sub-regions and distributed to the CPU multi-logical processors for parallel computing using the OpenMP parallel architecture. The sub-region data synchronization and interaction are realized through the virtual grid mechanism, and a multi-core parallel computing one-dimensional model with a "series-parallel-series" three-level computing architecture is constructed. The parallel computing interaction diagram is shown in Figure 2 The derivation steps of the multi-core parallel computing one-dimensional model are as follows:

[0056] S11, establish a one-dimensional flow control equation based on partial differential:

[0057] The generalized quasi-one-dimensional Euler equations with chemical reaction source terms are solved by implicit format strong coupling to simulate the combustion chamber flow field, and the establishment of the one-dimensional flow control equation is completed; in the Cartesian coordinate system, the strong coupling form of the generalized one-dimensional Euler equation and the component continuity equation is:

[0058] (1)

[0059] In the formula, is a conservative variable, is a non-viscous flux term, is a source term, is a partial differential symbol, t is time, x is a spatial coordinate;

[0060] (2)

[0061] In the formula, A is the engine cross-sectional area, d is the differential symbol, is the additive term, is the friction coefficient, is the equivalent hydraulic diameter, is the heating per unit mass, u sx is the component of the additive term along the axial direction, is the additive term per unit mass of the stagnation enthalpy, is the component of the first component, is the component of the second component, is the mass generation rate of the first component in unit time per unit volume, N is the component of the first component, is the mass generation rate of the first component in unit time per unit volume, is the mass generation rate of the second component in unit time per unit volume, is the mass generation rate of the first component in unit time per unit volume, N is the mass generation rate of the first component in unit time per unit volume, N is the number of components, is the density of the mixture gas, u is the gas velocity, p is the pressure, E is the specific internal energy of the mixture gas per unit mass; considering the area change, mass addition, friction and chemical reaction, the heating term is zero. When considering mass addition, the component continuity equation will become:

[0062] (3)

[0063] where, is the mass generation rate of the component in unit time per unit volume, i is the component of the component; i The enthalpy, energy, pressure and density are related as follows:

[0064]

[0065] (4)

[0066] where, is the specific enthalpy of the component at temperature i is the specific heat capacity at constant pressure of the component, T is the temperature, is the reference enthalpy at the reference temperature, i 0 is the reference temperature; T T (5)

[0067] (6)

[0068] where, c i is the mass fraction of the component; i

[0069]

[0070] where, is the average molar mass of the mixture gas, is the molar mass of the component; i

[0071] ​​​​​​​ (7)

[0072] wherein, is the static pressure of the component i , is the universal gas constant;

[0073] (8)

[0074] In the embodiment, one-dimensional Euler equations (mass continuity equation, momentum equation and energy equation) containing chemical reaction source terms are coupled with component continuity equations to solve the distribution of key physical quantity parameters along the axial direction in the combustion chamber, so as to obtain key control parameters such as engine thrust and shock wave train front edge position, and further realize the construction of a one-dimensional model of a multi-core parallel computing hypersonic scramjet engine facing control.

[0075] S12, a chemical kinetics model is established:

[0076] For the chemical kinetics model, taking ten-component ten-equation ethylene combustion as an example: the main components of the gas are: oxygen , acetylene , water vapor , hydrogen , carbon monoxide , carbon dioxide , oxygen element , hydrogen element , free radicals and nitrogen , and the main reactions are:

[0077] (9)

[0078] wherein, is the third body collision particle participating in the reaction.

[0079] The above reaction formula can be expressed as an equivalent formula of the general form:

[0080] (10)

[0081] wherein, r is the serial number of the elementary reaction equation, o is the sum of the component number and the third body collision particle number, nj is the sum of the component number and the third body collision particle number in the reaction equation, X o represents the component and the catalyst, and are the equivalent coefficients of the reactants and products, is the forward reaction rate constant, is the reverse reaction rate constant.

[0082] The reaction rate of each elementary reaction can be expressed in terms of a forward reaction rate constant and a reverse reaction rate constant, the forward reaction rate constant and the reverse reaction rate constant are given by the equations:

[0083] (11)

[0084] (12)

[0085] where is the frequency factor for the forward reaction of the r th elementary reaction, is the temperature factor for the forward reaction of the r th elementary reaction, is the activation energy for the forward reaction of the r th elementary reaction, is the frequency factor for the reverse reaction of the r th elementary reaction, is the temperature factor for the reverse reaction of the r th elementary reaction, is the activation energy for the reverse reaction of the r th elementary reaction, R is the universal gas constant;

[0086] The forward and reverse reactions of each elementary reaction are coupled by the chemical equilibrium constant of the reaction For the th elementary reaction, the following equilibrium relationship exists:

[0087] (13)

[0088] where R is the universal gas constant, is the stoichiometric number change for the r th elementary reaction;

[0089] (14)

[0090] where is the equivalent coefficient for the r th elementary reaction component i as a product, is the equivalent coefficient for the r th elementary reaction component i as a reactant;

[0091] (15)

[0092] where and is the component calculated at the reference pressure i Unit molar ratio enthalpy and unit molar ratio entropy.

[0093] The molar concentration of the component i is calculated as follows:

[0094] (16)

[0095] The molar density change rate of the component r in the nth reaction equation is: i

[0096] (17)

[0097] (18)

[0098] Considering the case where the third body collides with the particles:

[0099] (19)

[0100] where, L j is the control coefficient of whether the third body collision reaction occurs, and is set to 1 when the third body collision participates in the reaction, and is set to 0 when the third body collision does not participate in the reaction, and the collision probability of the corresponding nth component as the third body collision body in the nth elementary reaction is obtained by j k kj C j is the unit volume molar number of the component X The formula expression is simplified as follows: j

[0101] (20)

[0102] (21)

[0103] The mass generation rate of the component in the unit volume per unit time has the following relationship with the molar concentration of the component:

[0104] (22)

[0105] The generation rate of the component i is obtained by summing all the reaction equations as follows:

[0106] ​​​​​​​​ (23)

[0107] wherein, is the forward reaction rate, is the reverse reaction rate;

[0108] In the case of high temperature and considering chemical reactions, the thermodynamic parameters such as constant-pressure specific heat capacity , entropy and enthalpy are no longer constants independent of temperature, and in actual engineering calculations, polynomial fitting can be used to obtain the thermodynamic parameters of each reaction component at different temperatures. In the present application, the constant-pressure specific heat capacity , entropy and enthalpy are fitted by the following formula:

[0109]

[0110]

[0111] (24)

[0112] wherein, a 1, a 2, a 3, a 4, a 5, a 6 and a 7 are fitting coefficients;

[0113] The constant-pressure specific heat capacity of the mixed gas C p and the enthalpy of the mixed gas h are calculated by the following formula:

[0114] (25)

[0115] S13, the one-dimensional flow control equation considering chemical reactions is solved:

[0116] Because the equation contains chemical reactions, the equation is rigid, and using an explicit format to advance the solution of the equation requires taking a very small step to ensure that the final convergence of the equation calculation result is guaranteed, which seriously affects the calculation efficiency. Therefore, in the present application, an implicit format is used to solve the partial differential equation considering chemical reactions. Deriving the equation (1) on both sides with respect to time, we obtain:

[0117] (26)

[0118] The following formula is used to transform the inviscid flux term from the Cartesian coordinate system to the curvilinear coordinate system:

[0119] (27)

[0120] where is the intermediate parameter defined by the curvilinear coordinate ;

[0121] Then equation (26) can be written as

[0122] (28)

[0123] where D is the matrix of partial derivatives of the source term with respect to the conserved quantities; let be represented by the matrix A , then equation (28) can be written as

[0124] (29)

[0125] The partial derivative of the first term in the above equation with respect to time is expanded as

[0126] (30)

[0127] where n represents the n time;

[0128] After rearrangement, we have

[0129] (31)

[0130] Define the perturbation variable , where is the n +1 time Q , is the n time Q , and let , where is the n time RHS , then the equation is

[0131] (32)

[0132] where is the n +1 time heating per unit mass

[0133] After rearrangement, we have

[0134]

[0135] (33)

[0136] where I is the identity matrix

[0137] The matrix A is split into and , and the arrangement can be obtained:

[0138] (34)

[0139] In the formula, y is the grid, is n +1 moment y The heating of the unit mass at the grid point;

[0140] Discretize the spatial derivative term, for Use backward difference, for Use forward difference, can be obtained:

[0141] (35)

[0142] In the formula, is y The positive eigenvalue at the grid point A Matrix, is y The positive eigenvalue at the -1 grid point A Matrix, is y The negative eigenvalue at the grid point A Matrix, is y The negative eigenvalue at the +1 grid point A Matrix;

[0143] The arrangement can be obtained:

[0144] (36)

[0145] In the formula, is the interval time;

[0146] Assume Q The matrix changes slowly in space, and formula (36) can be written as:

[0147] (37)

[0148] Since , is A The maximum eigenvalue of the matrix, write the above formula as:

[0149] (38)

[0150] In the formula, represents the absolute value;

[0151] Let , be a diagonal matrix, the above equation can be written as:

[0152] (39)

[0153] (40)

[0154] Neglecting the higher order infinitesimal, the above equation can be equivalent to:

[0155] (41)

[0156] That is:

[0157] (42)

[0158] Equation (42) can be written as:

[0159] (43)

[0160] In which, is a placeholder for the intermediate quantity;

[0161] Equation (43) can be written as:

[0162] (44)

[0163] (45)

[0164] Using the grid point forward iteration to calculate the update of , then using the grid point backward iteration to calculate the update of , and then using the following equation (46), set a reasonable matrix update rate , update the matrix, and after it meets the convergence condition or reaches the maximum iteration step, output the matrix as the result.

[0165] (46)

[0166] S14, solve the inviscid flux term and the source term:

[0167] In equation (44), the value of is still unknown, and the values of and need to be solved respectively, and the value of is obtained by summing up using the formula . Using the AUSM+ format, the inviscid flux term is split into the convective flux term and pressure flux term P The processing is performed:

[0168] (47)

[0169] When the MUSCL interpolation is extended to high order accuracy, the following can be obtained:

[0170] (48)

[0171] In the formula, is the current position sound speed, The following formula is used for calculation:

[0172] (49)

[0173] (50)

[0174] (51)

[0175] (52)

[0176] Pressure flux term The expression of the pressure flux term is as follows:

[0177] (53)

[0178] In the formula, is the Mach number on the left side of the interface, is the pressure on the left side of the interface, is the Mach number on the right side of the interface, is the pressure on the right side of the interface, is the positive pressure flux term parameter, is the negative pressure flux term parameter, and the positive and negative pressure flux term parameters are calculated by the following formula.

[0179] (54)

[0180] (55)

[0181] In the formula, is a sign function;

[0182] According to the formula (46) to the formula (54), the inviscid flux term can be split as follows:

[0183] (56)

[0184] In the formula above Matrix parameter term and , the following formula is used for calculation under the consideration of grid non-uniformity.

[0185] (57)

[0186] wherein, s y and k y are parameters in the interpolation formula, and , , and are the corresponding parameters at the grid point, y , and are the corresponding parameters at the +1 grid point, y , and can be obtained by the following formulae:

[0187] (58)

[0188] wherein, represents at the grid point, represents at the grid point, represents at the grid point, represents at the grid point;

[0189] The Van Albada limiter is adopted:

[0190] (59)

[0191] wherein, , , .

[0192] After the inviscid flux term at the current grid point is obtained by using formulae (46) to (58), the value of the chemical reaction source term at the current grid point is obtained according to formula (2), and the two are vector summed to obtain the value of at the current grid point.

[0193] S15, parallelization calculation:

[0194] In the process of solving the one-dimensional flow control equation of the simulated combustion chamber flow field, most of the time is consumed in the iterative solution process of formulae (44)-(46). The calculation program needs to calculate from the first grid point to the last grid point by formula (44), then calculate from the last grid point to the first grid point by formula (45), and finally calculate from the first grid point to the last grid point by formula (46) to complete​​​ One-time update of the matrix. Optimization of this part of the program can greatly reduce the calculation time of the entire program.

[0195] Based on this, the formulas before formula (44)-(46) are integrated in this embodiment, and a first serial area is programmed. The first serial area completes the initialization of the calculation, which is specifically composed of the following 5 steps:

[0196] A1, according to the selected chemical kinetics model, set the quasi-one-dimensional Euler equation dimension, read the basic parameters required for chemical reaction calculation, set the maximum iteration step number and inflow condition;

[0197] A2, using the finite volume method, according to the corresponding engine type, draw the grid required for subsequent calculation;

[0198] A3, the fuel is added by using exponential distribution to simulate: for four oil injection points, the one-dimensional distribution of fuel is simulated by using exponential distribution, and the sum of one-dimensional distribution of four oil injection points is used as the total mass of fuel at the current grid point;

[0199] A4, initialize the conservation quantity Q matrix and its differential form dQ matrix;

[0200] A5, set the discrete time and the update rate of the conservation quantity Q matrix;

[0201] Formulas (44)-(46) are programmed as a parallel area: using multi-core parallel computing to optimize formulas (44)-(46), as shown in Figure 2 The entire calculation area from the inlet to the outlet is equally divided into a sub-area, each sub-area represents a sub-module, and each sub-module contains N grid points. Using the multi-core and multi-thread architecture of modern CPUs, multi-thread computing tasks are constructed, and sub-modules 1-a are distributed to a threads for parallel operation. The calculation process includes two time scales, the time scale to ensure the convergence of the partial differential equation calculation, and the total calculation scale to ensure the continuous and stable operation of the control system. Considering factors such as chemical reaction, gas flow rate and grid size, the time scale is 1e-05s, that is, all time scales in the partial differential equation solving process are 1e-05s. Considering the continuity of the control system, the total calculation scale The setting is 50 ms, that is, the entire calculation program takes 50 ms to complete the entire partial differential equation solving calculation and output the final result after accepting external control amount (K1-K4 fuel change amount). At the beginning of the first calculation period, all grids in sub-module 1-a are initialized according to the set conditions. After initialization, enter the main calculation loop. At the beginning of each main calculation loop, update the grid with fuel distribution according to the total fuel mass, and the remaining grids inherit the output results after the last calculation period. Then, sub-module 1-a respectively performs forward iteration calculation on the N grids it contains according to formula (44). After completing the forward iteration calculation, sub-module 1-a respectively performs backward iteration calculation on the N grids it contains according to formula (45). The virtual grid is used for data exchange between adjacent grids in different sub-modules. After completing the forward iteration and backward iteration, sub-module 1-a updates the N grids it contains using formula (46). After completing the Q matrix update of the conservation quantity, the residual is calculated. When the residual is less than the convergence threshold or the iteration reaches the maximum number of steps, the calculation is ended and the final result of the Q matrix is output.

[0202] The formulas after formulas (44)-(46) are integrated and compiled into a second serial area. After the second serial area completes the calculation, the output results of the Q matrix are used to calculate the engine thrust, along the wall pressure, and shock string front edge position according to the related formulas. The results are output and the S1 is ended.

[0203] The final output of the conservation quantity Q matrix in the parallel area has the following relationship:

[0204] (60)

[0205] In the formula, the first term of the conservation quantity Q matrix gives the density of the mixture gas at the current grid. The second term divided by the first term can be used to calculate the velocity of the mixture gas at the current grid. The pressure is solved by the energy equation , and the formula about pressure is constructed :

[0206] (61)

[0207] (62)

[0208] The pressure at the current grid is calculated by Newton iteration method :

[0209] (63)

[0210] In the formula, p0 is the initial value in the Newton iteration method;

[0211] After the engine along the line pressure is calculated using one-dimensional program, the Billig experience formula correction model is applied to more accurately fit the actual flow phenomenon:

[0212] (64)

[0213] In the formula, L is the length of the shock train in the isolation section, is the Mach number at the inlet of the isolation section, is the Reynolds number, which is taken as 10000, is the momentum thickness of the boundary layer at the inlet, H is the height of the isolation section, generally , is the static pressure behind the shock train, is the static pressure in front of the shock train.

[0214] After the length of the shock train in the isolation section L is obtained, X the length of the isolation section L is subtracted, and the position of the leading edge of the shock train L is obtained. X If this physical quantity is greater than the length of the isolation section , it is considered that the inlet does not start phenomenon occurs.

[0215] (65)

[0216] The engine thrust is calculated by the following formula:

[0217] (66)

[0218] In the formula, and are the mass flow rates at the outlet and inlet of the engine, and are the gas velocities at the outlet and inlet of the engine, in this embodiment, only one-dimensional flow of gas is considered, i.e. , , and are the pressures at the outlet and inlet of the engine, and are the cross-sectional areas at the outlet and inlet of the engine.

[0219] After the conservation quantity Q matrix is obtained by solving the partial differential equation by parallel computing, the engine thrust and the position of the leading edge of the shock train are finally obtained by using the formulas (60)-(66), and the multi-core parallel computing one-dimensional model of the scramjet engine oriented to control is constructed.

[0220] In the present embodiment, the scramjet engine needs to meet multiple safety constraints in actual operation, such as inlet unstart protection, thrust lower limit guarantee and the like. Therefore, the present application adopts a Safe SAC algorithm to solve the intelligent control problem of fuel multi-point injection of the scramjet engine. After modeling the target problem by using a constrained Markov decision process (CMDP), based on the Actor-Critic architecture, a safety network (Safety Network) is introduced into the reinforcement learning SAC algorithm. Similar to the classical reward function, a safety function is defined to calculate the constraint violation value, and then the value is used to form the TD (Temporal-Difference) target of the safety network to realize the parameter update of the safety network. Then, by using the outputs of the Critic network and the Safety network, the constrained CMDP problem is converted into an unconstrained optimization problem by using the Lagrange relaxation method for solving. Through the Safe SAC algorithm, the agent learns a safe policy, so as to maximize the future cumulative reward while ensuring that the cost is within the set safety limit d.

[0221] In S2, the Safe SAC agent includes a Critic network, a Safety network and an Actor network, and the structure is as shown in Figure 3 The Safe SAC agent continuously interacts with the environment to obtain a sequence and stores it in the experience replay pool. After storing enough samples in the experience replay pool, training is started. At the beginning of each training, sequences B are randomly extracted from the experience replay pool to update and train the network parameters. The Critic network and the Safety network use the TD error to construct the loss function gradient descent to update their own parameters. The Actor network introduces a Lagrange multiplier and uses the Lagrange relaxation method to convert the original constrained optimization problem into an unconstrained optimization problem. For the unconstrained min-max problem, the primal-dual optimization uses an iterative primal-dual method. In each iteration, the primal policy and the dual variable are updated in turn. First, the dual variable is fixed, and the policy gradient is used to update the Actor network. Then, the Actor network parameters are fixed, and the dual variable is updated.

[0222] The derivation steps of the Safe SAC agent are as follows:

[0223] SAC is an advanced reinforcement learning algorithm that can be used to deal with problems in continuous state space and action space. It introduces an entropy regularization mechanism, which increases the exploration degree of the reinforcement learning algorithm, helps to speed up policy learning, and reduces the possibility of the policy falling into local optimum. The objective function of SAC is written as:

[0224] (67)

[0225] where the randomness of the policy output is represented by the entropy term , is the discount factor, is the coefficient of regularization, which is used to determine the relative importance of entropy and reward, is the state , is the reward obtained by taking action , is the state of the agent, is the action of the agent, is the sampled trajectory that obeys the policy , is the mathematical expectation of the sampled trajectory , is the action probability distribution in state . After the current state and action

[0226] are determined, the state value function s and the action value function a are given by the following formula:

[0227] (68)

[0228] (69)

[0229] where is the reward obtained by the agent; SAC algorithm uses Actor network to approximate the policy function, and uses Critic network to approximate the value function. SAC algorithm updates the parameters by minimizing the loss function :

[0230] (70)

[0231] where is the target state value in state , which is calculated by the target Critic network in formula (68).

[0232] The Actor network parameters ​​​​by minimizing the loss function is updated as:

[0233] (71)

[0234] where, is the KL divergence of and is a normalization term, SAC agent action can be expressed as:

[0235] (72)

[0236] where, is the policy function characterized by policy parameters is the noise vector in standard normal distribution sampling, the Actor network outputs a normal distribution with mean and variance is the Hadamard product. Therefore, equation (71) can be written as:

[0237] (73)

[0238] In the standard SAC algorithm, the agent learns the control policy through exploratory interaction without explicit safety constraints. However, in the actual flight mission of the scramjet engine, such unconstrained exploration will lead to unacceptable safety risks: the agent may generate actions that induce the inlet not to start, the thrust to be below the safety threshold, and other faults, which may cause catastrophic consequences. Therefore, the Safety network is introduced into the traditional SAC algorithm to realize the quantitative evaluation and active protection of system safety constraints, which is similar to the Critic network in the AC framework. During the training process of the Safe SAC algorithm, the output of the Critic network and the output of the Safety network together constitute the loss function of the Actor network, which is used to update its parameters so that the agent maximizes the future cumulative reward while meeting the constraint conditions.

[0239] Similar to equations (68) and (69), after the current state s and action a are determined, the action value function and the state value function of the Safety network can be expressed as:

[0240] (74)

[0241] (75)​​​

[0242] where, c t the cost obtained by the agent;

[0243] The Safety network also uses a deep neural network for approximation, and its update method is similar to the Critic network, as shown in the following formula:

[0244] (76)

[0245] where, is the cost of the state and action ;

[0246] Introducing the Lagrange multiplier , using the Lagrange relaxation technique to convert the constrained Markov decision process into an unconstrained dual problem, which is more conducive to the exploration of the solution space by the agent, without considering the complex constraints during training. Using the primal-dual algorithm, the Safe SAC algorithm optimizes the Actor network and the Lagrange multiplier at the same time in each iteration, speeding up the training convergence speed. The invention considers the case of two safety constraints of shock train front edge position and thrust lower limit, and different values can be set for different safety constraints. The primal-dual optimization method is used to solve the min-max problem in formula (76), and in each iteration, the primal policy and the Lagrange multiplier are updated in turn. First, fix the Lagrange multiplier , and use the gradient ascent method to update the policy :

[0247] (77)

[0248] (78)

[0249] where, is the constructed objective function, is the Lagrange multiplier, is the policy parameter of the k w th training, is the policy parameter of the k w th training, is the gradient of , and is the regularization coefficient of the k w th training, is the safety threshold, is the regularization coefficient of the kw the value of the Lagrange multiplier in the next training;

[0250] After fixing the policy parameters , the update of the Lagrange multiplier is performed:

[0251] (79)

[0252] wherein, is the step size, is the gradient of , is the dual variable, is the projection of the dual variable , is the projection of the dual variable , is the projection of the dual variable , is the projection of the dual variable k w the value of the Lagrange multiplier in the next training.

[0253] The overall block diagram of Safe SAC is shown in Figure 3 The basic environment required for configuring the Safe SAC algorithm in Anaconda is configured, the object-oriented programming method is adopted, and the Safe SAC reinforcement learning intelligent agent is written using the python language.

[0254] S3 is specifically:

[0255] The basic training parameters of the safe reinforcement learning are configured, including the learning rate, the discount factor, the target network update rate and the safety threshold. The multi-core parallel computing one-dimensional model is closed-loop interacted with the Safe SAC intelligent agent. The fuel flow of the four injection points output by the Safe SAC intelligent agent is input to the multi-core parallel computing one-dimensional model. The engine thrust and the shock wave string front edge position output by the multi-core parallel computing one-dimensional model under the current working condition are input to the Safe SAC intelligent agent. Through the closed-loop interaction process, the multi-point intelligent injection pre-training is completed. The Safe SAC intelligent agent obtains the initial strategy capability of realizing the accurate control of the K1-K4 flow of the multi-injection point of the scramjet engine combustion chamber under the condition of meeting the safety constraint condition (intake not starting protection, thrust lower limit guarantee). In this embodiment, the Critic network and the Safety network in the Safe SAC intelligent agent are both six-input one-output black box models, and the Actor network is a two-input four-output black box model.

[0256] S4 is specifically:

[0257] According to the direct-connected ground test or numerical simulation of the scramjet engine, the axial position interval corresponding to the monotonic wall pressure and engine thrust is determined, the wall pressure sensor is arranged in the interval, the direct-connected test or numerical simulation under different fuel injection equivalence ratios is carried out, the discrete wall pressure-thrust sample data is collected, the continuous mapping relationship of the wall pressure to the engine thrust is established by using the Gaussian process regression, the nonlinear mapping model of the wall pressure at the rear part of the combustion chamber to the engine thrust is obtained, and the engine thrust is calculated by the wall pressure fed back by the wall pressure sensor. The indirect control law of the thrust is arranged on the edge AI computing platform, the engine thrust estimation value is calculated in real time through the feedback data of the wall pressure sensor, and a closed-loop indirect control loop of the thrust is formed.

[0258] In this embodiment, the engine thrust of the scramjet engine is taken as a core performance parameter. In the interaction between the reinforcement learning agent and the environment in step S3, the engine thrust can be obtained by analytical calculation through the inlet / outlet mass flow, flow velocity, pressure and flow passage cross-sectional area. However, the parameter cannot be directly measured in the actual flight task. In order to effectively control the engine thrust, a measurable indirect representation parameter needs to be selected. A large number of direct-connected ground tests and CFD numerical simulations have proved that the wall static pressure at a specific axial position interval at the rear part of the combustion chamber has a strict monotonic mapping relationship with the engine thrust. The wall pressure is easy to measure and responds quickly, and is an excellent choice for the controlled parameter of the engine. Therefore, for the scramjet engine structure proposed in the application, direct-connected ground tests or numerical simulations are carried out, the change relationship between the wall pressure at the rear part of the combustion chamber and the engine thrust is studied, the continuous mapping relationship of the wall pressure to the thrust is established by using the Gaussian process regression, and the nonlinear mapping model of the wall pressure at the rear part of the combustion chamber to the engine thrust is established.

[0259] S5 is specifically:

[0260] The pressure sensors are arranged on the wall of the isolation section, the upper and lower wall pressure sensors at the inlet of the isolation section and the upper and lower wall pressure sensors arranged at equal intervals at the rear part of the isolation section are used to form a feedback array, and the shock wave train front position in the isolation section is calculated by using the pressure ratio method.

[0261] In this embodiment, the pressure ratio method is used to identify the shock wave train front position in the isolation section. The pressure ratio method needs to obtain a reference pressure for subsequent calculation. The pressure at the upstream of the shock wave train in the isolation section is basically unchanged, and one pressure sensor is arranged on the upper wall and one pressure sensor is arranged on the lower wall near the inlet of the isolation section. The pressures measured by the pressure sensors are taken as the upper wall reference pressure and the lower wall reference pressure . There are 7 wall pressure sensors arranged at equal intervals along the axis at the rear part of the isolation section, and there are 14 pressure sensors on the upper and lower walls, so that 14 pressure sensor feedback data and , static pressure measured by the 7 sensors on the upper wall, static pressure measured by the 7 sensors on the lower wall. Let the total running time of the scramjet engine be T s , the interval time be , and the constant be k s represent the first k s sampling, and the total number of samplings in the total running time be n s The following relationships exist among the three:

[0262] (80)

[0263] and are two groups of variables that change with the change of , which are written as and . Then, the creation of the pressure ratio vector on the upper wall and the pressure ratio vector on the lower wall is completed through equation (81) and equation (82):

[0264] (81)

[0265] (82)

[0266] A line is constructed through the above pressure ratio points by linear interpolation. After the construction of the pressure ratio line is completed, the determination of the position of the front edge of the shock train on the upper and lower walls in the isolation section is completed through equation (83) and equation (84):

[0267] (83)

[0268] (84)

[0269] In the equations, and are the vectors of the installation positions of the wall pressure sensors on the upper wall and the lower wall, respectively. According to the experimental conclusion, when , the position of the front edge of the shock train solved by the pressure ratio method has the smallest difference with the experimental result, and are the positions corresponding to the pressure rise ratio of 1.5 with respect to the reference pressure point on the upper wall and the lower wall at each sampling time, and the continuous front edge of the shock train is solved by the pressure ratio method and the discrete wall pressure sensors as Figure 4The corresponding abscissa value at a pressure ratio of 1.5 is considered as the current time shock wave train front position in the isolation section. In this embodiment, the pressure ratio method is used to identify the characteristic mutation of the shock wave train, and the front position measurement error is controlled within 5% of the length of the isolation section, which provides real-time state perception for shock wave train stability control and realizes accurate measurement of the shock wave train front position.

[0270] S6 is specifically:

[0271] A ground direct connection test bench including an isolation section, a combustion chamber, a four-channel independent fuel supply system and corresponding sensors is built. Under the determined flow conditions, the pre-trained Safe SAC reinforcement learning agent controls the fuel flow of the four injection points under different shock wave train front positions and engine thrust feedback. Under the determined flow conditions, step, sine, ramp and other target thrust commands are applied respectively, and the fuel valve opening, shock wave train front position and engine thrust estimated value are recorded in real time. The engine thrust tracking error size, shock wave train moving range and incidence of intake duct non-starting are calculated to verify the effectiveness of the fuel multi-point injection intelligent control system and method in the present application by quantitative indicators.

[0272] In this embodiment, as shown in Figure 5 and Figure 6 , a multi-core parallel computing one-dimensional model corresponding to the actual configuration of the scramjet engine and the oil injection position is established in the computer. The flow conditions are set as: Mach number 2.1, total pressure 0.8Mpa, total temperature 937K, and flow rate 3.09kg / s. The four injection points are set as: one before the upper wall cavity of the combustion chamber, one after the cavity, and two evenly distributed in the middle of the lower wall of the combustion chamber. The pre-training of the Safe SAC agent is completed in the computer using the multi-core parallel computing one-dimensional model. After pre-training, the agent can reasonably control the fuel flow of the four injection points under different shock wave train front positions and thrust size feedback. Under the same flow conditions, the wind tunnel test data is used to construct a nonlinear mapping model of the combustion chamber rear wall pressure and thrust based on the Gaussian process regression algorithm, and an indirect thrust control law based on the wall pressure is constructed. The pre-trained Safe SAC reinforcement learning agent, nonlinear mapping model and related preprocessing algorithm are downloaded to the NVIDIA Jetson Orin NX development board. Through the cross-compilation tool chain, an executable file is generated and deployed to the edge AI computing development board described in the present application. In the hardware layer, the development board outputs 4-way fuel valve PWM control signals to the driving circuit through the GPIO interface, and the ADC interface collects wall pressure sensor array data at a sampling rate of more than 20KHz. In the software layer, the Safe SAC agent makes online decisions with millisecond-level time consumption, and the preprocessing algorithm performs real-time calculation with microsecond-level delay.

[0273] A smart control system for multi-point fuel injection in a scramjet engine includes a main controller, an execution / feedback module, and a scramjet engine.

[0274] The main controller includes a reinforcement learning agent unit and an input preprocessing unit. The input preprocessing unit receives discrete wall pressure data collected by pressure sensors arranged on the inner wall of the isolation section and calculates a real-time estimate of the position of the leading edge of the shock train within the isolation section. In this embodiment, the reinforcement learning agent unit is configured with a pre-trained Safe SAC reinforcement learning agent, and the input preprocessing unit includes a noise processing algorithm, a normalization algorithm, and a shock train leading edge position detection algorithm. The Safe SAC reinforcement learning agent in the main controller is as follows: Figure 5 As shown.

[0275] The execution / feedback module includes a fuel control module, a pressure monitoring module, and a drive circuit; the fuel control module includes four fuel valves, such as... Figure 5 Fuel valves V1, V2, V3, and V4 are shown and are installed in independent fuel supply pipes connecting the fuel tank and four fuel injection points in the combustion chamber. The drive circuit is used to match the voltage and current requirements between the output of the reinforcement learning agent unit and the input of the fuel valve. The fuel valve receives the electronic control signal output by the reinforcement learning agent unit after being amplified by the drive circuit, and converts it into a corresponding valve opening adjustment signal. The fuel flow at the fuel injection point is controlled by adjusting the opening of each valve. The pressure monitoring module includes an isolation section wall pressure sensor array and a combustion chamber wall pressure sensor. Specifically, the isolation section wall pressure sensor array consists of a pair of sensors symmetrically arranged on the upper and lower walls of the isolation section inlet section, and several groups of sensors symmetrically arranged at equal intervals along the flow direction on the upper and lower walls of the middle and rear areas of the isolation section, used to collect discrete wall pressure data. The combustion chamber wall pressure sensor is located on the lower wall of the rear part of the combustion chamber. In this embodiment, the wall pressure sensor is a wall pressure sensor manufactured by Rosemount.

[0276] like Figure 5 As shown, the scramjet engine includes four fuel injection points, namely fuel injection point K1, fuel injection point K2, fuel injection point K3 and fuel injection point K4, which are respectively located at specific positions on the upper and lower walls of the combustion chamber.

[0277] Engine thrust during aircraft flight cannot be directly obtained through measurement. Therefore, in engine thrust closed-loop control, a monotonic, stable, and measurable physical quantity that can characterize thrust is needed. The wall pressure in a specific region of the lower part of the combustor has a monotonic relationship with engine thrust, and is easy to measure and responds quickly, making it a preferred controlled parameter for engine thrust control. Therefore, a sensor is placed at a specific location on the lower rear wall of the combustor to measure the static pressure of the gas at that location. This static pressure value serves as a characterizing parameter of engine thrust, supporting the indirect thrust closed-loop control based on safety reinforcement learning.

[0278] In the description of the application, it needs to be understood that the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implying the number of technical features indicated. Therefore, the features defined by "first", "second", "third" can explicitly or implicitly include one or more of the features.

Claims

1. A method for intelligent control of multi-point fuel injection in a scramjet engine, characterized in that, Includes the following steps: S1. A simulation model is built using a multi-core parallel computing architecture. While retaining the core coupling relationship between fuel injection, combustion heat release, and thrust output, a one-dimensional multi-core parallel computing model for controlling scramjet engines is constructed. S2. In the SAC reinforcement learning framework, safety constraints are integrated into the learning process to construct a SafeSAC agent that considers safety constraints. S3. Based on the constructed Safe SAC agent, with multi-point fuel flow as the control variable, carry out reinforcement learning multi-point intelligent injection pre-training based on multi-core parallel computing one-dimensional model interaction to obtain the pre-trained Safe SAC reinforcement learning agent. S4. Design a thrust indirect control law based on wall pressure: Establish a nonlinear mapping model between the rear wall pressure of the combustion chamber and the engine thrust through ground tests, and calculate the engine thrust based on the wall pressure; S5. Deploy pressure sensors on the wall of the isolation section and use the pressure ratio method to calculate the position of the leading edge of the shock train in the isolation section; S6. Build a ground-based direct-connection test bench covering specified inflow conditions and high-precision fuel supply, input the engine thrust and shock wave leading edge position into the pre-trained Safe SAC reinforcement learning agent, and control the fuel flow at multiple injection points. In S2, the Safe SAC agent includes a Critic network, a Safety network, and an Actor network. The Critic and Safety networks utilize the TD error to construct a loss function gradient descent for parameter updates. The Actor network introduces Lagrange multipliers and uses the Lagrange relaxation method to transform the original constrained optimization problem into an unconstrained optimization problem. For the unconstrained mini-maximum problem, the primal-dual optimization adopts an iterative primal-dual method. In each iteration, the primal policy and dual variables are updated sequentially. The specific update method is as follows: fix the Lagrange multipliers and use policy gradient ascent to update the Actor network's policy; fix the Actor network's policy parameters and update the dual variables.

2. The intelligent control method for multi-point fuel injection in a scramjet engine according to claim 1, characterized in that, In S1, the one-dimensional model of multi-core parallel computing includes a first serial region, a parallel region, and a second serial region. The specific steps for initializing the first serial region are as follows: A1. Based on the selected chemical kinetic model, set the dimension of the quasi-one-dimensional Euler equation, read the basic parameters required for chemical reaction calculation, and set the maximum number of iterations and inflow conditions; A2. Using the finite volume method, draw the mesh required for subsequent calculations based on the corresponding engine type; A3. The fuel addition is simulated using an exponential distribution: For the four fuel injection points, the fuel one-dimensional distribution is simulated using an exponential distribution, and the sum of the one-dimensional distributions of the four fuel injection points is taken as the total fuel addition mass at the current grid point. A4. Initialize the conserved quantity Q matrix and its differential form dQ matrix; A5. Set the update rate for discrete time and the conserved quantity Q matrix; The parallel region iteratively updates the conserved Q matrix through forward scanning, backward scanning, and the conserved Q matrix update formula. After each update, a residual is calculated. The calculation ends when the residual is less than the convergence threshold or the maximum number of iterations is reached, and the final conserved Q matrix is ​​output. The second serial region obtains the combustion chamber performance parameters by solving the final conserved quantity Q matrix, outputs the engine thrust and shock wave train leading edge position, and completes the construction of a one-dimensional model of multi-core parallel computing for controlled scramjet engines.

3. The intelligent control method for multi-point fuel injection in a scramjet engine according to claim 1, characterized in that, S3 specifically refers to: The multi-core parallel computing one-dimensional model is interacted with the Safe SAC agent in a closed loop. The fuel flow rate of the four injection points output by the Safe SAC agent is input to the multi-core parallel computing one-dimensional model. The engine thrust and shock wave leading edge position under the current operating conditions output by the multi-core parallel computing one-dimensional model are input to the Safe SAC agent. Multi-point intelligent injection pre-training is completed through this closed-loop interaction process.

4. The intelligent control method for multi-point fuel injection in a scramjet engine according to claim 1, characterized in that, S4 specifically refers to: Based on the scramjet engine configuration, direct ground tests or numerical simulations were conducted to determine the axial position range where wall pressure and engine thrust monotonically correspond. Within this range, wall pressure sensors were deployed, and direct tests or numerical simulations were carried out under different fuel injection equivalence ratios. Discrete wall pressure-thrust sample data were collected, and a continuous mapping relationship between wall pressure and engine thrust was established using Gaussian process regression. A nonlinear mapping model between the rear wall pressure of the combustion chamber and engine thrust was obtained, and the engine thrust was calculated using the wall pressure feedback from the wall pressure sensors.

5. The intelligent control method for multi-point fuel injection in a scramjet engine according to claim 4, characterized in that, S5 specifically refers to: Pressure sensors are deployed on the walls of the isolation section. A feedback array is formed by the pressure sensors on the upper and lower walls at the entrance of the isolation section and the pressure sensors on the upper and lower walls at equal intervals in the middle and rear of the isolation section. The position of the leading edge of the shock train in the isolation section is calculated by the pressure ratio method.

6. The intelligent control method for multi-point fuel injection in a scramjet engine according to claim 5, characterized in that, S6 specifically refers to: A ground-based direct-connection test rig was built, including an isolation section, combustion chamber, four-channel independent fuel supply system, and corresponding sensors. Under certain incoming flow conditions, the fuel flow at four injection points was controlled by a pre-trained Safe SAC reinforcement learning agent at different shock train leading edge positions and engine thrust feedback.

7. A smart control system for multi-point fuel injection in a scramjet engine, applied to the smart control method for multi-point fuel injection in a scramjet engine as described in any one of claims 1 to 6, characterized in that, Includes the main controller, execution / feedback module, and scramjet engine; The main controller includes a reinforcement learning agent unit and an input preprocessing unit. The input preprocessing unit is used to receive discrete wall pressure data collected by pressure sensors arranged on the inner wall of the isolation section and calculate the real-time estimate of the position of the leading edge of the shock train in the current isolation section. The execution / feedback module includes a fuel control module, a pressure monitoring module, and a drive circuit. The fuel control module includes four fuel valves, which are installed in independent fuel supply pipes connecting the fuel tank and four fuel injection points in the combustion chamber. The drive circuit is used to match the voltage and current requirements between the output of the reinforcement learning agent unit and the input of the fuel valve. The fuel valve receives the electronic control signal output by the reinforcement learning agent unit after being amplified by the drive circuit, and converts it into a corresponding valve opening adjustment signal. The fuel flow at the fuel injection point is controlled by adjusting the opening of each valve. The pressure monitoring module includes an isolation section wall pressure sensor array and a combustion chamber wall pressure sensor. Specifically, the isolation section wall pressure sensor array consists of a pair of sensors symmetrically arranged on the upper and lower walls of the isolation section inlet section, and several groups of sensors symmetrically arranged at equal intervals along the flow direction on the upper and lower walls of the middle and rear areas of the isolation section, used to collect discrete wall pressure data. The combustion chamber wall pressure sensor is located on the lower wall of the rear part of the combustion chamber. The scramjet engine includes four oil injection points, which are located at specific positions on the upper and lower walls of the combustion chamber.

Images