Air inlet end shock wave control method and system based on BP neural network

Abstract

The invention provides an air inlet end shock wave control method and system based on a BP neural network, and belongs to the field of supersonic aircraft air inlets. The method comprises the following steps: firstly, combining CFD numerical simulation data and wind tunnel test data to establish a mapping relation database among inlet outlet static pressure, bypass deflation flow and end shock wave position; then taking the outlet static pressure and the deflation flow as input and the shock wave position as output for training to obtain a supersonic air inlet end shock wave position model, and utilizing the model to realize high-precision online estimation of the shock wave position; and then combining with an ADRC controller to form a closed-loop control system of the ending shock wave. According to the method, data-driven modeling and active-disturbance-rejection control technologies are fused, the control precision, the response speed and the anti-jamming capability of the supersonic air inlet duct end shock wave position are remarkably improved, and a high-reliability intelligent control solution is provided for stable and efficient work of a supersonic aircraft air inlet duct.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of supersonic aircraft inlet, and particularly relates to an inlet terminal shock control method and system based on a BP neural network. BACKGROUND

[0002] The supersonic inlet is a core component of the air-breathing hypersonic aircraft, and its performance directly affects the thrust efficiency and working stability of the engine. The position of the terminal shock is a key parameter for determining the aerodynamic performance of the inlet: the forward movement of the shock position may cause the inlet to be unstarted, and the backward movement of the shock position may cause flow separation and total pressure loss. Therefore, accurate and rapid closed-loop control of the position of the terminal shock is a core challenge to ensure the efficient and safe operation of the ramjet engine.

[0003] Traditional terminal shock control methods mainly rely on geometric mechanism adjustment (such as a ramp) or fuel flow adjustment to suppress shock oscillation. However, such methods are difficult to effectively deal with complex working conditions such as sudden changes in flight Mach number and angle of attack. In recent years, the bypass bleed technology has been introduced into the field of shock control due to its simple structure and rapid response. However, existing research still has significant limitations: first, system modeling is difficult, there is a complex nonlinear coupling relationship between the position of the terminal shock, the outlet static pressure and the bleed flow, and it is difficult to establish a physical model that accurately reflects the dynamic characteristics of the system; second, the accuracy of state perception is insufficient, in wind tunnel tests, the shock position is usually obtained by manually interpreting the schlieren image, which is poor in real-time and cannot be directly used for closed-loop control, and the method based on pressure sensor data to calculate the shock position usually has a large error; third, there is a lack of real-time control strategies, although CFD (Computational Fluid Dynamics) can construct a parameter mapping database, its calculation is time-consuming and difficult to meet the online control requirements, and the open-loop control strategy based on a simplified model is prone to regulation lag and overshoot due to the failure to fully consider dynamic factors such as airflow disturbance and boundary layer separation, which cannot achieve precise and stable control of the shock position. Therefore, there is an urgent need to develop a new closed-loop control system that can accurately predict the actual position of the terminal shock and adaptively maintain the terminal shock at the target position. SUMMARY

[0004] The application provides a terminal shock control method and system based on a BP neural network to solve the problems in the prior art, which can provide a high-reliability intelligent control solution for stable and efficient operation of the supersonic aircraft inlet.

[0005] To solve the above technical problems, the application provides the following technical scheme: a terminal shock control method for an inlet based on a BP neural network, comprising the following steps:

[0006] S1, construct an ultrasonic inlet end shock wave position dataset, including the inlet outlet static pressure, bypass bleed flow, and the end shock wave position corresponding to the inlet outlet static pressure and bypass bleed flow; divide the dataset into a training set, a validation set, and a test set;

[0007] S2, construct an ultrasonic inlet end shock wave position neural network, with the inlet outlet static pressure and bypass bleed flow in the training set as input and the ultrasonic inlet end shock wave position as output to train the neural network, to obtain an ultrasonic inlet end shock wave position model;

[0008] S3, input the end shock wave position obtained by the ultrasonic inlet end shock wave position model into an ADRC self-anti-disturbance controller, to calculate and output the control amount instruction required to drive the bypass bleed actuator in real time, to drive the end shock wave position to move to the target position, to realize directional control of the shock wave position.

[0009] Further, the aforementioned step S1 includes the following sub-steps:

[0010] S1.1, for a specific configuration of the ultrasonic inlet, collect the inlet outlet static pressure, bypass bleed flow, and end shock wave position, and construct an ultrasonic inlet end shock wave position simulation dataset; at the same time, conduct a ground wind tunnel test, collect the inlet outlet static pressure, bypass bleed flow, and obtain the schlieren image through the optical glass observation window of the expansion section, accurately extract the end shock wave position through image processing, and form an ultrasonic inlet end shock wave position measured dataset,

[0011] S1.2, implement multi-stage joint preprocessing on the simulation dataset and the measured dataset: perform data cleaning and alignment, based on the time sequence label of the wind tunnel test, remove abnormal frames caused by sensor transient failure or image blur from the test collected shock wave position, outlet static pressure, and bleed flow three-channel data, at the same time, match the simulation data closest to the test working point in the simulation dataset, establish simulation-test data pairs; then perform physical quantity dimension normalization, perform Z-score standardization processing on the outlet static pressure, bleed flow, and shock wave position respectively; finally, use the Savitzky-Golay filter to smooth the shock wave position time sequence signal, retain the true motion trend while suppressing the random jitter introduced by the schlieren image extraction.

[0012] Further, in the aforementioned step S2, the ultrasonic inlet end shock wave position neural network is constructed using a BP neural network architecture, and the training of the neural network includes the following sub-steps:

[0013] S2.1, initialize the BP neural network model, and randomly generate the weight matrix and bias vector;

[0014] S2.2 During the neural network training process, the training set is used to directly optimize the model parameters, namely weights and biases, in the forward and backward propagation algorithms, so that it learns the data features and patterns related to the task.

[0015] S2.3 During training, the validation set is used to periodically evaluate the model’s generalization performance on unseen data. Its core function is to guide hyperparameter tuning, select the optimal model, or implement early stopping strategies to prevent the model from overfitting the training set.

[0016] S2.4 After the entire model training and tuning process is completed, a final, independent and unbiased evaluation of the model's generalization ability is performed using a test set, reflecting the model's expected performance when handling new and unknown data in real-world application scenarios.

[0017] Furthermore, the aforementioned step S3 includes the following sub-steps:

[0018] S3.1, Set the target position command x target As input, and after smoothing the step command, a transition trajectory x is generated. ref and its differential signal To prevent the system from overshooting or oscillating due to sudden changes in instructions;

[0019] S3.2 Real-time reception of the actual position x of the terminal shock wave estimated online by the supersonic inlet terminal shock wave position model. real The current position and state of the final shock wave are estimated by combining the control quantity u output by the ADRC controller itself. and its rate of change The model error of the shock wave position at the end of the supersonic inlet, the actuator hysteresis, and the airflow disturbance are all unified into a total disturbance.

[0020] S3.3 Calculate tracking error and differential error Combined with total disturbance Generate control quantity Drive the bypass venting actuator to operate, where k p k d b0 is the proportional-derivative gain, and b0 is the control input gain.

[0021] S3.4 The bypass venting actuator dynamically adjusts the valve opening based on the control quantity u output by the ADRC controller. By changing the effective flow area of ​​the bypass channel, it dynamically adjusts the venting flow rate, reconstructs the pressure distribution of the flow field in the intake duct, and drives the actual position x of the final shock wave. real Approaching target position x target This enables directional control of the position of the final shock wave.

[0022] The application also provides a BP neural network-based inlet termination shock wave control system, comprising: a specific configuration supersonic inlet, a BP neural network training module, and a termination shock wave front position control module.

[0023] The specific configuration supersonic inlet is integrated with a bypass air release actuator, which can realize active control of the termination shock wave position in the inlet by precisely adjusting the air release flow rate.

[0024] The BP neural network training module comprises a pressure sensor, an air release flow rate monitoring system, a schlieren optical system, and a computer.

[0025] The pressure sensor is used to collect wall surface pressure data and calculate the inlet outlet static pressure.

[0026] The air release flow rate monitoring system is used to monitor and feed back the air release flow rate data of the bypass channel in real time.

[0027] The schlieren optical system can make the termination shock wave structure present light and dark stripes in the schlieren image by collecting the light deflection caused by the air flow density gradient, output the termination shock wave position, and visualize the shock wave form and position.

[0028] The computer can construct and train a supersonic inlet termination shock wave position model based on the inlet outlet static pressure, the air release flow rate data of the bypass channel, and the termination shock wave position, and output the termination shock wave position.

[0029] The termination shock wave front position control module comprises an ADRC active disturbance rejection controller and a bypass air release actuator.

[0030] The bypass air release actuator comprises a bypass channel arranged in the inlet expansion section, an air release valve, and a driving device, which can quickly and accurately convert the control amount instruction output by the ADRC controller into a change in the physical air release area, directly adjust the bypass air release flow rate through the air release area change, change the mass flow rate distribution of the inlet flow field, affect the local pressure distribution in the expansion section, drive the termination shock wave position to move to the target position, and realize shock wave position control.

[0031] Further, the ADRC active disturbance rejection controller comprises a tracking differentiator TD, an extended state observer ESO, and a nonlinear error feedback control law module NLSEF.

[0032] The tracking differentiator TD is used to receive the terminal shock wave target position instruction, smooth the terminal shock wave target position instruction and extract a differential signal of the terminal shock wave target position instruction, and generate a reference signal of a transition process which is smooth and contains a target change trend.

[0033] The extended state observer ESO is used to observe the actual position and the change rate of the terminal shock wave in real time, and estimate the total disturbance of the system in real time; the nonlinear error feedback control law NLSEF is calculated through nonlinear combination based on the reference signal provided by the tracking differentiator TD and the state estimation value provided by the extended state observer ESO, and the total disturbance estimated by the extended state observer ESO is introduced in real time to dynamically compensate, so as to enhance the robustness of the system and generate a control amount instruction with strong anti-interference ability and output to the bypass bleeding actuator.

[0034] Further, the total disturbance of the system includes airflow disturbance, neural network model error and downstream pressure fluctuation, and the state estimation value provided by the extended state observer ESO includes the actual position and the change rate of the terminal shock wave and the total disturbance.

[0035] Further, the pressure sensors are arranged on the wall surface of the inlet duct outlet in equal intervals in the circumferential direction.

[0036] Further, the inlet duct terminal shock wave control system based on the BP neural network is based on the real-time change of the inlet duct outlet static pressure and the bypass channel bleeding flow data, the supersonic inlet duct terminal shock wave position neural network continuously feeds back the estimation value of the actual position of the terminal shock wave to the ADRC controller, updates the control amount in real time and drives the bypass bleeding actuator to make a dynamic response; at the same time, the change of the bleeding flow of the actuator and the change of the outlet static pressure collected by the pressure sensor on the wall surface of the inlet duct outlet are fed back to the input end of the neural network again.

[0037] Compared with the prior art, the beneficial technical effects of the above technical solutions are as follows:

[0038] The neural network is trained and the parameters are optimized through a large amount of simulation data and test measured data, a high-reliability BP neural network system is obtained, a dynamic coupling model of the shock wave position-bleeding flow-outlet static pressure is formed, the shock wave position detection accuracy and response speed are significantly improved, and the problem that the terminal shock wave cannot be accurately and stably controlled under complex working conditions by the traditional method and the existing bypass bleeding technology is effectively solved; the airflow disturbance and other dynamic disturbances are adaptively compensated by combining the ADRC self-disturbance rejection control algorithm, the regulation lag and overshoot are greatly reduced, the terminal shock wave can still quickly converge and stably maintain at the target position under complex working conditions, and the reliability and adaptability of the control system are greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1The figure is a schematic diagram of the inlet termination shock control system based on BP neural network of the present application.

[0040] Figure 2 The figure is a two-dimensional schematic diagram of the supersonic inlet of a certain specific configuration of the present application.

[0041] Figure 3 The figure is a schematic diagram of the BP neural network training system of the present application.

[0042] Figure 4 The figure is a schematic diagram of the three-layer BP neural network structure of the present application.

[0043] Figure 5 The figure is a flowchart of the BP neural network algorithm of the present application.

[0044] Figure 6 The figure is a working principle diagram of the ADRC self-disturbance control controller of the present application. DETAILED DESCRIPTION

[0045] In order to better understand the technical content of the present application, specific embodiments are described below with reference to the accompanying drawings.

[0046] Aspects of the present application are described in this detailed description and illustrated in the accompanying drawings by a number of illustrative embodiments. The embodiments of the present application are not limited to the drawings described. It should be understood that the present application is realized by any one of the above-mentioned concepts and embodiments, and the concepts and embodiments described in detail below, since the concepts and embodiments disclosed by the present application are not limited to any embodiment. In addition, some aspects disclosed by the present application can be used alone, or in any appropriate combination with other aspects disclosed by the present application.

[0047] As Figure 1 shown, the present application provides an inlet termination shock control system based on BP neural network, characterized in that it comprises a supersonic inlet of a certain specific configuration, a BP neural network training module, and a termination shock front position control module.

[0048] The two-dimensional schematic diagram of the supersonic inlet of a certain specific configuration is shown in Figure 2 The core feature is that a bypass air release actuator is integrated, which can accurately adjust the air release flow rate to achieve active control of the termination shock position in the inlet. At the same time, to meet the needs of visual observation of the flow field structure in wind tunnel tests, part of the wall surface of the expansion section of the inlet will be replaced by an optical glass observation window with high light transmittance in the test state, which facilitates the use of schlieren imaging technology to clearly and real-time capture and record the motion state of the termination shock in the expansion section, providing direct experimental basis for the verification of the control system and the study of the flow mechanism.

[0049] The end shock wave front position control system comprises an ADRC self-disturbance rejection controller and a bypass air exhaust actuator;

[0050] The ADRC self-disturbance rejection controller is composed of a tracking differentiator (TD), an extended state observer (ESO) and a nonlinear error feedback control law (NLSEF), and can calculate and output the control quantity required for driving the bypass air exhaust actuator in real time according to the set end shock wave target position; the tracking differentiator (TD) is responsible for receiving the end shock wave target position instruction, smoothing the instruction and extracting the differential signal thereof to generate a reference signal which is smooth in the transition process and contains the target change trend; the extended state observer (ESO) can observe the system state (mainly the actual position of the end shock wave and the change rate thereof) in real time and estimate the total disturbance (including air flow disturbance, neural network model error, downstream pressure fluctuation and the like) in real time; the nonlinear error feedback control law (NLSEF) can calculate the preliminary control quantity based on the reference signal provided by the TD and the state estimation value (including the actual position of the end shock wave, the change rate and the total disturbance) provided by the ESO, and can introduce the total disturbance estimated by the ESO for dynamic compensation in real time to enhance the robustness of the system and generate a control quantity instruction with strong anti-disturbance capability and output to the actuator;

[0051] The bypass air exhaust actuator is composed of a bypass channel arranged in the inlet passage expansion section, an air exhaust valve and a driving device thereof, and can quickly and accurately convert the control quantity instruction output by the ADRC controller into the change of the physical air exhaust area (achieved by adjusting the opening degree of the air exhaust valve), and then directly adjust the bypass air exhaust flow through the change of the air exhaust area, change the mass flow distribution of the flow field in the inlet passage, affect the local pressure distribution in the expansion section (especially the static pressure behind the end shock wave), and finally drive the end shock wave position to move to the target position to realize the shock wave position control.

[0052] The BP neural network training system is shown in Figure 3 and specifically comprises a pressure sensor, an air exhaust flow monitoring system, a schlieren optical system and a computer;

[0053] The pressure sensor is arranged on the wall surface of the inlet passage outlet in the circumferential direction at equal intervals, and is used for collecting wall surface pressure data and then calculating the inlet passage outlet static pressure;

[0054] The air exhaust flow monitoring system is used for monitoring and feeding back the air exhaust flow data of the bypass channel in real time;

[0055] The schlieren optical system comprises a point light source, a collimating concave mirror, a focusing concave mirror, a knife edge and a high-speed camera, and this system can make the end shock wave structure present light and dark stripes in the schlieren image by capturing the light deflection caused by the air flow density gradient, so as to visually show the shock wave form and position.

[0056] The computer is firstly used for developing inlet CFD numerical simulation research to obtain simulation data of internal flow field, then pre-processes simulation data and measured data collected in the inlet wind tunnel test to provide high-quality data set for the neural network, and finally constructs a BP neural network model and completes neural network training aiming at the requirement of controlling the shock wave at the end of the inlet.

[0057] The application further provides an inlet end shock wave control method based on a BP neural network, comprising the following steps: S1, constructing an inlet end shock wave position data set of a supersonic inlet, including an inlet outlet static pressure, a bypass air release flow rate, and an inlet end shock wave position corresponding to the inlet outlet static pressure and the bypass air release flow rate; dividing the data set into a training set, a verification set and a test set;

[0058] S2, constructing a supersonic inlet end shock wave position neural network, taking the inlet outlet static pressure and the bypass air release flow rate in the training set as input and taking the supersonic inlet end shock wave position as output to train the neural network, so as to obtain a supersonic inlet end shock wave position model;

[0059] S3, inputting the end shock wave position obtained by the supersonic inlet end shock wave position model into an ADRC self-disturbance control device, calculating and outputting a control amount instruction required for driving a bypass air release actuator in real time, driving the end shock wave position to move to a target position, and realizing directional control of the shock wave position.

[0060] As a preferred embodiment of the method, step S1 is specifically: developing CFD numerical simulation research on a supersonic inlet with a specific configuration, systematically researching the change law of the end shock wave position with the inlet outlet static pressure and the bypass air release flow rate, establishing a parameterized mapping relationship between the two and the shock wave position, and forming a simulation data set. At the same time, ground wind tunnel test is carried out on the supersonic inlet with the specific configuration, and full-condition flow field data are obtained through the coordinated action of the bypass air release actuator and the throttle cone: fixing the bypass air valve opening degree (covering 0% to 100% full stroke), slowly moving the throttle cone through the stepping motor, continuously increasing the inlet outlet static pressure from the initial value, driving the end shock wave to gradually transmit forward from the inlet outlet position, until the throat is completely pushed out, and the system traverses the whole dynamic evolution process in which the end shock wave stably exists.

[0061] In order to accurately capture the movement of the end shock wave, a schlieren optical system is used for flow field visualization: a high-brightness narrow-slit light source or a laser point light source is used as a transmitting end, the light is converted into a parallel light beam by a collimating concave mirror and vertically penetrates the optical glass observation window of the test section, the light beam after penetrating the flow field is focused and imaged by a focusing concave mirror, a knife edge with a precisely adjustable cutting depth (usually 50% to 80% light spot coverage) is installed at the focal plane, the contrast of the shock wave density gradient is enhanced through selective light filtering, and finally time sequence schlieren images are recorded on the imaging plane by a high-speed camera.

[0062] The position of the terminal shock wave is determined based on the schlieren image sequence, the instantaneous static pressure data is collected by the pressure sensor arranged on the wall surface of the inlet channel outlet, the static pressure value of the inlet channel outlet at the corresponding time is obtained through dynamic calibration calculation, the bleed flow rate data fed back by the bleed flow rate monitoring system in real time is recorded, the three-channel synchronous data is integrated to form a structured test measured data set, and the complete mapping relationship of the bleed flow rate-outlet static pressure-shock wave position is included.

[0063] The simulation data set and the test measured data set are subjected to multi-stage joint preprocessing; first, data cleaning and alignment are performed, based on the time sequence label of the wind tunnel test, the three-channel data of the shock wave position, the outlet static pressure and the bleed flow rate collected by the test are strictly synchronized, the abnormal frames caused by the transient failure of the sensor or the image blur are removed, the simulation data most adjacent to the test working condition point in the simulation database are matched, the simulation-test data pair is established; then, the dimensionless normalization of physical quantities is performed, the input features (outlet static pressure and bleed flow rate) and the output label (shock wave position) are subjected to Z-score standardization processing, that is,

[0064]

[0065] Wherein, μ x is the mean value of each physical quantity, and σ x is the standard deviation of each physical quantity, so as to eliminate the bias of network training caused by the difference in magnitude; for the high-frequency noise specific to the test data, the Savitzky-Golay filter is used for smoothing processing of the shock wave position time sequence signal, so as to retain the true motion trend while suppressing the random jitter introduced by the schlieren image extraction; finally, the integrated unified feature-label data set is divided into a training set, a verification set and a test set in a ratio of 7:2:1 through hierarchical random sampling, so as to ensure that each subset covers the full working condition state and avoid the decline of generalization ability caused by the lack of local working condition.

[0066] As a preferred embodiment of the method, in step S2, the neural network for constructing the position of the terminal shock wave of the supersonic inlet channel adopts a BP neural network architecture,

[0067] Initialize the BP neural network model and randomly generate the weight matrix and bias vector. During neural network training, the training set is used to directly optimize the model's parameters (weights and biases) through forward and backpropagation algorithms, enabling it to learn task-related data features and patterns. The validation set is used to periodically evaluate the model's generalization performance on unseen data during training. Its core role is to guide hyperparameter tuning, select the optimal model, or implement early stopping strategies (such as stopping training when the validation loss no longer decreases) to prevent the model from overfitting the training set. The test set is used only once after the entire model development and tuning process is completely completed to provide a final, independent, and unbiased evaluation of the model's generalization ability, reflecting the model's expected performance when handling new and unknown data in real-world application scenarios.

[0068] The parameter optimization process of the BP neural network model is divided into two stages: The first stage is the forward propagation process, where the input information (outlet static pressure, venting flow rate) is processed layer by layer through the input layer and hidden layers to calculate the actual output value (end shock wave position) of each unit; The second stage is the backward propagation process, where if the expected output value is not obtained in the output layer, the difference between the actual output and the expected output (i.e., the error) is calculated recursively layer by layer, and the weights are modified by the gradient descent method to minimize the total error function and achieve the preset accuracy of the estimation of the actual position of the end shock wave, thus obtaining a highly reliable BP neural network model.

[0069] by Figure 4 The algorithm is briefly described using a three-layer BP neural network as an example. The network has 2 input nodes and 1 output node. Let q be the number of nodes in the hidden layer and P be the total number of input samples. pi Let v represent the i-th input value of the p-th sample. ki θ represents the weights from the i-th node in the input layer to the k-th node in the hidden layer. k ω represents the threshold of the k-th node in the hidden layer. k Let θ be the weight from the k-th node in the hidden layer to the output layer. ′ If the threshold of the output layer is given, then the output of the k-th node in the hidden layer is...

[0070]

[0071] The output of the output layer is

[0072]

[0073] The excitation function f is selected as the standard Sigmoid function.

[0074]

[0075] Its derivative satisfies

[0076] f′ = f(l-f)

[0077] Define the global error function as

[0078]

[0079] where E p is the error of the pth sample, t p is the ideal output.

[0080] According to the gradient descent method, the following weight adjustment formula is derived:

[0081] (1) Adjustment of output layer weights

[0082]

[0083] where η is called the learning rate, generally in the range of 0.1-0.3.

[0084] Define the error signal as

[0085]

[0086] The first term:

[0087]

[0088] The second term:

[0089]

[0090] Thus

[0091] δ p = (t p -y p )·y p (1-y p )

[0092] Therefore, the weight adjustment formula for each neuron in the output layer is

[0093]

[0094] (2) Adjustment of hidden layer weights

[0095]

[0096] Define the error signal as

[0097]

[0098] where

[0099]

[0100] then

[0101] δ pk =(δ p ω k )z pk (1-z pk )

[0102] Therefore, the formula for adjusting the weights of each neuron in the hidden layer is:

[0103]

[0104] Figure 5 This demonstrates the overall process of the BP neural network algorithm. First, the neural network architecture is built, and the connection weights and biases (thresholds) are initialized. Then, the training loop begins, sequentially reading training samples from the training set. For each sample, forward propagation is performed, calculating the weighted inputs and activation outputs of each neuron layer by layer, and calculating the error between the network's final output layer and the expected target value of the sample. Then, backpropagation is performed, calculating the error terms of each hidden layer layer by layer based on the output error, and adjusting the network's connection weights and biases accordingly. After all samples have been trained (i.e., one training cycle is completed), the average error E of the network on the entire training set is calculated. If the average error E reaches the preset accuracy requirement, the algorithm terminates. If the accuracy requirement is not met, it checks whether the current number of iterations (number of training cycles) has reached the preset maximum limit. If the limit has not been reached, the training loop is returned to continue reading samples for the next iteration. If the limit has been reached, the algorithm terminates.

[0105] In a preferred embodiment of the present invention, given the target position x of the trailing shock wave target The actual position x of the terminal shock wave is estimated in real time online using a neural network model. real The ADRC controller calculates and outputs the optimal control command u to the bypass venting actuator based on the error between the target position and the actual position, the dynamic characteristics of the system, and the estimated total disturbance. Its core working principle is to achieve robust control through three working parts: the tracking differentiator (TD), the extended state observer (ESO), and the nonlinear state error feedback (NLSEF).

[0106] Figure 6 This demonstrates the working principle of the ADRC controller. Step S3 includes the following sub-steps:

[0107] S3.1, The tracking differentiator (TD) will transfer the target position command x target As input, and after smoothing the step command, a transition trajectory x is generated. ref and its differential signal To prevent the system from overshooting or oscillating due to sudden changes in instructions;

[0108] S3.2, the ESO receives the actual shock position x estimated by the shock position model online real and the control variable u output by the ADRC controller itself, estimates the current shock position state and its rate of change and unifies the shock position model error, the actuator hysteresis, and the airflow disturbance into a total disturbance

[0109] S3.3, the NLSEF calculates the tracking error and the differential error combined with the total disturbance to generate the control variable to drive the bypass bleed actuator, where k p , k d is the proportional-differential gain, and b0 is the control input gain

[0110] S3.4, the bypass bleed actuator adjusts the valve opening according to the control variable u output by the ADRC controller, dynamically adjusts the bleed flow by changing the effective flow area of the bypass channel, reconfigures the pressure distribution of the inlet channel flow field, and drives the actual shock position x real to approach the target position x target , achieving directional control of the shock position (related to the response characteristics of the actuator).

[0111] Based on the real-time changes in outlet static pressure and bleed flow, the BP neural network model continuously feeds back the estimated value of the actual shock position to the ADRC controller, updates the control variable in real time, and drives the bypass bleed actuator to respond dynamically; at the same time, the action effect (bleed flow change) of the actuator and the dynamic data (outlet static pressure change) collected by the inlet channel outlet wall pressure sensor are again perceived and fed back to the input end of the BP neural network model.

[0112] Thus, the perception and estimation of the BP neural network model, the decision and control of the ADRC controller, the dynamic response of the bypass bleed actuator, and the real-time feedback of environmental parameters together form a self-adaptive shock wave closed-loop control system, realizing the "perception-decision-execution-feedback" closed-loop control logic; under this closed-loop circuit, the system can continuously correct and compensate in real time according to the actual deviation of the shock position, and through continuous iterative optimization, ultimately ensures that the shock wave converges stably to the target position.

[0113] While the application has been described by way of example with reference to certain embodiments thereof, it is to be understood that the application is not limited to the embodiments described above, but intrinsically extends to various modifications and changes in form and details. Therefore, the scope of the application should be determined not by the embodiments described above, but by the scope of the appended claims and their equivalents.

Claims

1. A method for controlling the shock wave at the end of the inlet channel based on BP neural network, characterized in that, The method comprises the following steps: S1, constructing an ultrasonic inlet shock wave position data set, including the inlet outlet static pressure, bypass air flow, and the corresponding shock wave position of the inlet outlet static pressure and bypass air flow; the data set is divided into a training set, a validation set and a test set; S2, constructing an ultrasonic inlet shock wave position neural network, using the inlet outlet static pressure and bypass air flow in the training set as input, and the ultrasonic inlet shock wave position as output to train the neural network, to obtain an ultrasonic inlet shock wave position model; S3, inputting the shock wave position obtained by the ultrasonic inlet shock wave position model into an ADRC self-anti-disturbance controller, to calculate and output the control quantity instruction required to drive the bypass air exhaust actuator in real time, to drive the shock wave position to move to the target position, and to realize directional control of the shock wave position.

2. The method of claim 1, wherein, Step S1 includes the following sub-steps: S1.1, collecting the inlet outlet static pressure, bypass air flow and shock wave position of a specific ultrasonic inlet, constructing an ultrasonic inlet shock wave position simulation data set; at the same time, carrying out a ground wind tunnel test, collecting the inlet outlet static pressure and bypass air flow, and obtaining a schlieren image through an optical glass observation window of the expansion section, and accurately extracting the shock wave position through image processing, to form an ultrasonic inlet shock wave position measured data set, S1.2, implementing multi-stage joint preprocessing on the simulation data set and the measured data set: performing data cleaning and alignment, based on the time sequence label of the wind tunnel test, removing abnormal frames of the shock wave position, outlet static pressure and air flow three-channel data collected in the test caused by sensor transient failure or image blur, at the same time, matching the simulation data closest to the test working point in the simulation data set, establishing simulation-test data pairs; then performing physical quantity dimension normalization, performing Z-score standardization processing on the outlet static pressure, air flow and shock wave position respectively; finally, using a Savitzky-Golay filter to smooth the shock wave position time sequence signal, retaining the true motion trend while suppressing the random jitter introduced by the schlieren image extraction.

3. The method of claim 1, wherein the method is based on a BP neural network. In step S2, the BP neural network architecture is used to construct the ultrasonic inlet shock wave position neural network, and the training of the neural network includes the following sub-steps: S2.1, initializing the BP neural network model, and randomly generating a weight matrix and a bias vector; S2.2, in the neural network training process, the training set is used for forward propagation and backward propagation algorithm to directly optimize the parameters of the model, i.e. the weight and bias, so as to learn the data features and patterns related to the task; S2.3, during the training process, the validation set is used to periodically evaluate the generalization performance of the model on unseen data, which plays a key role in guiding the optimization of hyperparameters, selecting the optimal model or implementing the early stopping strategy to prevent overfitting of the model to the training set. S2.4, after the completion of the entire model training and optimization process, the test set is used to evaluate the final, independent and unbiased generalization ability of the model, reflecting the expected performance of the model when processing new and unknown data in real application scenarios.

4. The method of claim 1, wherein, Step S3 includes the following sub-steps: S3.1, the target position instruction x target as input, and smooth the step instruction to generate a transition trajectory x ref and its differential signal Avoid overshoot or oscillation of the system due to instruction mutation; S3.2, Real-time receive supersonic inlet shock wave position model online estimation of the actual position of the shock wave x real And the control amount u output by the ADRC controller itself, estimate the current shock wave position state And its rate of change And the supersonic inlet shock wave position model error, the actuator hysteresis, the airflow disturbance is unified as the total disturbance S3.3, compute tracking error and differential error combined total disturbance generate control quantity drive bypass bleed actuator action, where k p , k d is a proportional-differential gain, b0 is a control input gain; S3.4, bypass gas exhaust actuator dynamically adjusts the valve opening according to the control quantity u output by the ADRC controller, dynamically adjusts the gas exhaust flow by changing the effective flow area of the bypass channel, reconfigures the pressure distribution of the flow field in the intake duct, and drives the actual position x of the trailing shock wave real Approaches the target position x target , realizes directional control of the trailing shock wave position.

5. A system for controlling the shock wave at the end of the inlet channel based on the BP neural network, characterized in that, It comprises: The specific configuration of the supersonic inlet, the BP neural network training module, and the trailing shock wave front position control module; The specific configuration of the supersonic inlet, the integrated bypass bleed actuator, by accurately adjusting the bleed flow rate, the trailing shock wave position in the inlet is actively controlled, part of the wall surface of the inlet expansion section is a high-transmittance optical glass observation window, and the Schlieren imaging technology is used to capture and record the movement state of the trailing shock wave in the expansion section; The BP neural network training module includes a pressure sensor, a bleed flow rate monitoring system, a Schlieren optical system, and a computer; wherein the pressure sensor is used to collect wall surface pressure data and calculate the static pressure at the inlet outlet; The bleed flow rate monitoring system is used to monitor and feed back the bleed flow rate data of the bypass channel in real time; The Schlieren optical system makes the trailing shock wave structure present light and dark stripes in the Schlieren image by collecting the light deflection caused by the air flow density gradient, outputs the trailing shock wave position, and visualizes the shock wave form and position; The computer constructs and trains the trailing shock wave position model of the supersonic inlet based on the static pressure at the inlet outlet, the bleed flow rate data of the bypass channel, and the trailing shock wave position, and outputs the trailing shock wave position; The trailing shock wave front position control module includes an ADRC active disturbance rejection controller and a bypass bleed actuator; wherein the ADRC active disturbance rejection controller calculates and outputs the control amount command required to drive the bypass bleed actuator in real time according to the target position of the trailing shock wave; The bypass bleed actuator includes a bypass channel arranged in the inlet expansion section, a bleed valve, and a driving device; which is used to quickly and accurately convert the control amount command output by the ADRC controller into a change in the physical bleed area, directly adjust the bypass bleed flow rate through the bleed area change, change the mass flow distribution of the flow field in the inlet, affect the local pressure distribution in the expansion section, drive the trailing shock wave position to move to the target position, and realize the shock wave position control.

6. The inlet lip shockwave control system based on BP neural network according to claim 5, characterized in that, The ADRC active disturbance rejection controller includes a tracking differentiator TD, an extended state observer ESO, and a nonlinear error feedback control law module NLSEF; The tracking differentiator TD is used to receive the trailing shock wave target position command, smooth it, and extract its differential signal to generate a reference signal that is smooth in the transition process and contains the target change trend; The extended state observer ESO observes the actual position of the trailing shock wave and its rate of change in real time, and estimates the total disturbance of the system in real time; the nonlinear error feedback control law NLSEF calculates the preliminary control amount based on the reference signal provided by the tracking differentiator TD and the state estimation value provided by the extended state observer ESO, and dynamically compensates the total disturbance estimated by the extended state observer ESO in real time to enhance the robustness of the system and generate a control amount command with strong anti-disturbance ability to output to the bypass bleed actuator.

7. The inlet lip shockwave control system based on BP neural network according to claim 6, characterized in that, The system total disturbance includes airflow disturbance, neural network model error, and downstream pressure fluctuation; the state estimation value provided by the extended state observer ESO includes the actual position of the shock wave, the rate of change, and the total disturbance.

8. The inlet lip shockwave control system based on BP neural network according to claim 5, characterized in that, The pressure sensor is arranged equidistantly along the circumference on the wall surface of the inlet outlet.

9. The inlet lip shockwave control system based on BP neural network according to claim 5, characterized in that, Based on the real-time changes of the bypass passage discharge flow data and the inlet passage outlet static pressure, the neural network continuously feeds back the estimation of the actual position of the shock wave at the end of the inlet to the ADRC controller, which updates the control variable in real time and drives the bypass discharge actuator to make a dynamic response. At the same time, the changes of the discharge flow of the actuator and the changes of the outlet static pressure collected by the outlet wall pressure sensor are fed back to the input end of the neural network again.

Images