Water-to-air cross-medium self-adaptive cooperative control method

Abstract

The invention discloses a water-to-air cross-medium adaptive cooperative control method, and belongs to the technical field of cross-medium aircraft control. In order to overcome the hysteresis defect of cross-domain control of a cross-medium aircraft, through the steps of multi-source sensing and disturbance prediction, cross-modal strategy preloading and power excitation, structure unfolding and transient energy coupling, pneumatic steady-state control and fractal recursive optimization and the like, microsecond-level fluid disturbance pre-judgment is achieved through a chaos phase synchronization engine; and in combination with fractal recursion reinforcement learning compression attitude control errors, the millisecond delay problem of nonlinear disturbance real-time prediction and compensation in the cross-medium process is solved. According to the method, the control precision and stability of the aircraft in cross-medium motion can be improved, smooth transition of a water-air dynamic model can be realized, and the reliability of the aircraft in a complex sea-air interaction scene is enhanced.

Description

Technical Field

[0001] The present invention relates to a water-to-air trans-medium adaptive collaborative control method, belonging to the field of ships. Background Art

[0002] In the context where traditional vehicles are limited to a single medium and lack cross-domain maneuverability, the technology of trans-medium vehicles, as a cutting-edge field of multidisciplinary intersection of extreme environment fluid mechanics, materials science, intelligent control, and artificial intelligence, integrates knowledge in multiple fields such as multiphase flow dynamics optimization, research and development of anti-impact composite materials, design of multimodal power systems, and integration of high-precision sensors. With intelligent adaptive control algorithms and multi-physical field coupling simulation technology as the core breakthrough directions, it also faces key challenges such as insufficient accuracy of multiphase flow modeling, material failure in extreme environments, poor real-time performance of cross-domain control, and low energy and power efficiency. These together constitute the background art of trans-medium vehicles. Summary of the Invention

[0003] Object of the Invention: In order to overcome the deficiencies in the prior art, the present invention provides a (to be supplemented after the solution is determined)

[0004] Technical Solution: To solve the above technical problems, a water-to-air trans-medium adaptive collaborative control method of the present invention includes the following steps:

[0005] (1) Multi-source perception and disturbance prediction: Through real-time sensor data, combined with a chaotic phase synchronization engine, high-frequency sampling of fluid pressure pulsation is carried out to predict in advance the mutation direction, amplitude, and action time of the hydrodynamic moment at the moment of water emergence;

[0006] (2) Input the current attitude and speed of the vehicle, as well as the disturbance data of the mutation direction, amplitude, and action time of the hydrodynamic moment predicted by the synchronization engine in step (1) into the fractal recursive reinforcement learning algorithm. Use the fractal recursive reinforcement learning algorithm to map the above input data to the interface transition layer of the Sierpinski carpet fractal space. In this special fractal space, the algorithm deeply analyzes and calculates the data according to its unique rules and structures. After normalizing the state space, it is mapped to the levels of the fractal structure:

[0007]

[0008] In the formula, 1 represents the underwater state, 2 represents the transition state, i.e., water-air switching, 3 represents the air state, T water , T air represents the layering threshold between the underwater and air states, ||ν PCA || represents the norm of the eigenvector after principal component analysis;

[0009] If Level(s) = 1, proceed to step (3); if Level(s) = 2, proceed to step (5); if Level(s) = 3;

[0010] (3) Based on the current state and historical information, use the DDPG algorithm to generate and update the policy, obtain the optimal control parameters, and repeat steps (1) and (2);

[0011] (4) At the moment of breaking through the water surface, eject and deploy the folding wing and rotor structures, synchronously trigger the chaotic phase folding algorithm, compress the predicted perturbation energy from the phase space and inject it into the propulsion system in a directional manner to increase the initial lift, and switch the fractal controller to the air level;

[0012] (5) After emerging from the water, the traditional PID controller takes over the flight attitude stabilization, maintains the angle of attack and roll angle stability through aileron deflection and thrust vector adjustment. The fractal recursive reinforcement learning module analyzes the aerodynamic parameters in real time, dynamically adjusts the control weights, and at the same time the chaotic oscillator continuously monitors the flow field perturbation characteristics and updates the prediction model in an incremental learning manner.

[0013] Preferably, in step (1), the prediction method includes the following steps:

[0014] (11) Initialize the chaotic oscillator, construct the initial state of the chaotic system, and use the Lorenz system as the chaotic oscillator. Its differential equation is:

[0015]

[0016] In the formula, the initial values of each parameter are: σ = 10, ρ = 28;

[0017] (12) Map the pressure amplitude p(t) to the chaotic variable range and perform signal normalization; inject the normalized pressure signal p′(t) into the Lorenz system and modify the first equation to:

[0018]

[0019] where ε = 0.3 is the coupling strength;

[0020] (13) Force the phase alignment of the chaotic oscillator and the external perturbation signal to establish a perturbation-chaos correlation model;

[0021] The present invention adopts Coupling protocol: Introduce a coupling term between three groups of parallel Lorenz oscillators:

[0022]

[0023] In the formula, γ = 0.5 is the coupling coefficient between oscillators, and i, j are oscillator numbers;

[0024] (14) Calculate the synchronization error:

[0025]

[0026] Where x and y are the means of the states of the three groups of oscillators. By continuously adjusting the parameters, the synchronization error E sync < 0.01;

[0027] (15) Map the synchronized chaotic trajectory to a high-dimensional phase space, capture the time-evolution characteristics of the perturbation, and perform phase-space reconstruction;

[0028] (16) Generate a feedforward control command, specifically generating a command, which refers to the torque direction, amplitude, and action time.

[0029] Preferably, in step (15), the delay embedding method is used for phase-space reconstruction, and the delay time τ = 5 ms and the embedding dimension m = 3 are selected to reconstruct the phase-space trajectory: X(t) = [x(t), x(t - τ), x(t - 2τ)].

[0030] Preferably, step (16) extrapolates future torque mutations based on the phase-space trajectory to generate a feedforward control command, including the following steps:

[0031] (161) Determine the torque direction according to the sign of the derivative of the phase-space trajectory and calculate the torque amplitude:

[0032]

[0033] Where k = 0.1 is a calibration coefficient;

[0034] (162) Predict the duration of the torque:

[0035]

[0036] (163) Finally, output the prediction results of the torque direction, amplitude, and action time.

[0037] Furthermore, to improve the long-term prediction accuracy, the present invention performs feedback and parameter adjustment, that is, dynamically optimizing the chaotic parameters according to the synchronization error; (which step is this mainly for above)

[0038] Parameter adaptation rule:

[0039] ρ(t + 1) = ρ(t) + K p ·E sync (t), K p = 0.1

[0040] Where ρ(t) represents the value of the chaotic parameter ρ at the current moment, which controls the nonlinear strength of the Lorenz system, and E sync (t) represents the synchronization error at the current moment, which measures the degree of phase alignment between the chaotic oscillator group and the external disturbance signal. Restricting ρ ∈ [20, 45] can prevent the system from becoming unstable.

[0041] Coupling strength adjustment:

[0042]

[0043] ε(t) represents the value of the coupling strength ε at the current moment, which controls the influence degree of the external disturbance signal on the chaotic system. K i represents the integration coefficient, which controls the cumulative effect of ε adjustment;

[0044] By using real-time error feedback to adjust the chaotic parameters, a closed-loop optimization is formed, making the phase alignment between the chaotic oscillator group and the external signal more accurate; the vehicle can better adapt to environmental changes such as different flow velocities and water depths, maintaining the system stability; and reducing the synchronization error, improving the accuracy of torque mutation prediction.

[0045] Preferably, the step (2) is specifically as follows:

[0046] (21) Input the current attitude and speed of the vehicle, as well as the disturbance data of the flow force moment mutation direction, amplitude, and action time predicted by the chaotic phase synchronization engine in step one, into the fractal recursive reinforcement learning algorithm together;

[0047] (22) Use the fractal recursive reinforcement learning algorithm to map the above input data to the interface transition layer of the Sierpinski carpet fractal space;

[0048] (23) Perform data preprocessing. The data of the pressure sensor is low-pass filtered through the FPGA module with a cut-off frequency of 5 kHz; normalized to the range of [-1, 1]; the inertial navigation data uses a moving average filter to smooth the acceleration and angular velocity data, eliminating instantaneous jitter. The preprocessed multi-dimensional feature matrix is subjected to principal component analysis, and the original data is projected into the three-dimensional principal component space to obtain a three-dimensional principal component score vector as the input for judging the structural hierarchy: υ PCA =[υ1, υ2, υ3];

[0049] After normalizing the state space, map it to the hierarchy of the fractal structure:

[0050]

[0051] Where 1 represents the underwater state, 2 represents the transition state, that is, the water-air switch, 3 represents the air state, T water , T air represents the layering threshold between the underwater and air states; ||νPCA || represents the norm of the eigenvector after principal component analysis, which is used to measure the significance of the state in a specific medium, and the threshold represents the preset hierarchical threshold for distinguishing different medium states.

[0052] Preferably, in step (4), according to the predicted torque amplitude and the acting time, the total energy requirement is calculated: E = τ·Δt, where E represents the total energy requirement, that is, the total energy that the propulsion system needs to output; τ represents the predicted disturbing torque; Δt represents the torque acting time; this energy needs to be output by the propulsion system; the thrust arm of the tail propeller is L1, and the thrust arm of the vector nozzle is L2, and the thrusts are F1 and F2 respectively, then the total reverse torque is:

[0053] τ 反向 = F1·L1 + F2·L2

[0054] Under the goal of minimizing energy consumption:

[0055]

[0056] Optimize and solve for F1 and F2;

[0057] According to the thrust demand, energy is allocated to the thrusters in real time, and the relationship between energy and thrust is: where E i represents the energy required for a single thruster; F i represents the thrust of the corresponding thruster; t represents the energy acting time; substituting the data can obtain the energy ratio of the tail propeller and the vector nozzle, and the total energy is input to each thruster according to this ratio.

[0058] Beneficial effects: The present invention has achieved a leap in prediction accuracy and response speed: In the prior art for cross-medium motion control, when using traditional control algorithms, due to the lack of accuracy in fluid disturbance modeling, the response is delayed. This not only causes attitude instability problems but also significantly increases energy loss. The chaotic phase synchronization engine of the present invention collects data from a Doppler velocimeter, a nine-axis MEMS inertial navigation unit, etc. with the help of multi-source sensors, and uses the Lorenz chaotic oscillator model to be able to predict fluid disturbances at the μs level. Its time resolution is 3 orders of magnitude higher than that of traditional Kalman filtering, and it can predict in advance the sudden change of the hydrodynamic torque at the moment of water entry, giving a precise disturbance model for dynamic control and making the vehicle respond more timely and accurately to environmental changes.

[0059] Significant improvement in control precision: In the prior art, the attitude control error is high during cross-media transitions because the linear or quasi-static dynamic models designed for a single medium are difficult to adapt to the non-linear mutations during cross-media processes, resulting in a mismatch between control instructions and real dynamic characteristics. The present invention uses fractal recursive reinforcement learning to map the state information such as the attitude and speed of the vehicle into the Sierpinski carpet fractal space, combines the DDPG algorithm and the GRU recursive neural network, and generates optimal control parameters based on historical states. In this way, the attitude control error can be controlled within ±2°, enabling the vehicle to operate stably along the predetermined trajectory even in complex sea-air interaction environments.

[0060] Smooth transition of model switching: In the prior art, during the water-air medium switch, the dynamic model is difficult to adapt quickly, affecting the stability and safety of the vehicle. At the moment of breaking through the water surface, the present invention uses an ejection device to deploy the folding wing and rotor structures, synchronously triggers the chaotic phase folding algorithm, injects the energy after phase space compression into the propulsion system, and enhances the initial lift. At the same time, the fractal controller quickly switches from the interface transition layer to the air layer and loads the pre-trained sliding mode control parameters to achieve a millisecond-level smooth transition of the water-air dynamic model, ensuring the stable operation of the vehicle during the cross-media process. Brief Description of the Drawings

[0061] Figure 1 It is the overall cross-media flow chart of the cross-media vehicle;

[0062] Figure 2 It is the working flow chart of the chaotic phase synchronization engine of the cross-media vehicle;

[0063] Figure 3 It is the principle flow chart of the energy distribution stage during the transition stage of the cross-media vehicle. Detailed Embodiment

[0064] The present invention will be further described in detail below with reference to the accompanying drawings.

[0065] The present invention provides a water-to-air cross-media adaptive cooperative control method based on chaotic phase and fractal recursive reinforcement learning, including,

[0066] Step 1: Multi-source perception and disturbance prediction. Through the real-time data of sensors and combined with the chaotic phase synchronization engine, high-frequency sampling of fluid pressure pulsations is carried out to predict in advance the mutation direction, amplitude and action time of the fluid moment at the moment of breaking through the water surface, providing an accurate disturbance model for dynamic control.

[0067] Step 2: Cross-modal strategy preloading and power excitation. Based on the fractal recursive reinforcement learning algorithm, map the current attitude, speed, and predicted disturbances of the vehicle to the interface transition layer of the Sierpinski carpet fractal space to generate an initial control strategy. Start the tail propeller and pre-charged propulsion motor, and accelerate the vehicle to the threshold speed to ensure the kinetic energy reserve required to break through the water surface.

[0068] Step 3: Structure deployment and transient energy coupling. At the moment of breaking through the water surface, eject and deploy the folding wing and rotor structures, and synchronously trigger the chaotic phase folding algorithm. Compress and directionally inject the predicted disturbance energy from the phase space into the propulsion system to increase the initial lift, and switch the fractal controller to the air level to complete the smooth transition of the water-air dynamics model.

[0069] Step 4: Aerodynamic steady-state control and fractal recursive optimization. After emerging from the water, the traditional PID controller takes over the flight attitude stabilization, and maintains the angle of attack and roll angle stability through aileron deflection and thrust vector adjustment; the fractal recursive reinforcement learning module analyzes the aerodynamic parameters in real time, dynamically adjusts the control weights, and at the same time the chaotic oscillator continuously monitors the flow field disturbance characteristics, and updates the prediction model in an incremental learning manner to form a closed-loop adaptive control.

[0070] Furthermore, as shown in Figure 1 The method of multi-source perception and disturbance prediction in Step 1 includes:

[0071] Install and calibrate various sensors for the cross-media vehicle, including: Doppler velocimeter, nine-axis MEMS inertial navigation unit, capacitive liquid level sensor array, micro pressure pulsation sensor;

[0072] Specifically, the Doppler velocimeter is used to measure the three-dimensional velocity of the vehicle relative to the water body, with an accuracy of ±0.1 m / s. It is deployed in the streamlined cabin at the tail to avoid turbulence interference and finally outputs 100 velocity vector updates per second;

[0073] The nine-axis MEMS inertial navigation unit is used to collect three-axis acceleration, three-axis angular velocity, and three-axis magnetic field data in real time, and finally outputs a sampling rate of 400 Hz;

[0074] The capacitive liquid level sensor array is used to detect the water surface contact state and local liquid level height changes, with a resolution of ±1 mm, and finally outputs the 0 / 1 contact state and the analog liquid level height;

[0075] The micro pressure pulsation sensor is used to capture the high-frequency pulsation of the fluid pressure on the vehicle surface in the 0-10 kHz frequency band. After final AD acquisition, it is preprocessed by FPGA and transmitted through Ethernet.

[0076] Furthermore, the present invention uses the principal component analysis dimensionality reduction and simplification method for data preprocessing, inputs the pressure sensor data collected under typical working conditions, collects 10 - 20 groups of data for each working condition, with each group lasting for 1 second, and transmits the 3D principal component scores after dimensionality reduction to the chaotic phase synchronization engine as the input of the perturbation characteristics;

[0077] Even further, as shown in Figure 2 the present invention utilizes the nonlinear dynamic characteristics of the Lorenz chaotic oscillator to predict in advance the sudden change direction, amplitude, and action time of the fluid moment at the moment of water emergence, providing an accurate perturbation model for dynamic control.

[0078] First, initialize the chaotic oscillator to construct the initial state of the chaotic system, providing a dynamic model basis for subsequent synchronization and prediction.

[0079] Use the Lorenz system as the chaotic oscillator, and its differential equation is:

[0080]

[0081] In the formula, the initial values of each parameter are: σ = 10, ρ = 28;

[0082] The initialized chaotic oscillator is in a free evolution state, preparing for subsequent injection of external perturbation signals;

[0083] Furthermore, map the pressure amplitude p(t) to the chaotic variable range for signal normalization processing; inject the normalized pressure signal p′(t) into the Lorenz system and modify the first equation to:

[0084]

[0085] where ε = 0.3 is the coupling strength.

[0086] The external perturbation signal drives the chaotic oscillator to deviate from the free evolution trajectory, thereby initiating the synchronization process.

[0087] Furthermore, force the phase alignment of the chaotic oscillator and the external perturbation signal to establish a perturbation - chaos correlation model;

[0088] The present invention adopts a coupling protocol: introduce a coupling term among three groups of parallel Lorenz oscillators:

[0089]

[0090] In the formula, γ = 0.5 is the coupling coefficient between oscillators, and i, j are oscillator numbers.

[0091] Furthermore, calculate the synchronization error:

[0092]

[0093] wherein is the mean value of the states of three groups of oscillators.

[0094] By continuously adjusting the parameters, the synchronization error E sync < 0.01.

[0095] The chaotic oscillator group is made to be in phase with the external signal through the coupling protocol, and the perturbation drives the oscillator group into the synchronous state, providing a stable dynamic trajectory after synchronization for subsequent prediction.

[0096] Furthermore, the synchronized chaotic trajectory is mapped to a high-dimensional phase space to capture the time-evolution characteristics of the perturbation and perform phase-space reconstruction;

[0097] Specifically, the present invention adopts the delay embedding method, selects the delay time τ = 5 ms and the embedding dimension m = 3, and reconstructs the phase-space trajectory:

[0098] X(t) = [x(t), x(t - τ), x(t - 2τ)]

[0099] In the formula, τ represents the delay time, which is determined by the mutual information method or the autocorrelation function to ensure that the information between adjacent delay points is independent and the dynamic characteristics are retained; x(t - τ) and x(t - 2τ) respectively represent the observed values at the previous and two delay time points.

[0100] Thus, the present invention adopts the weighted local linear fitting algorithm to predict the trajectory X(t + Δt) in the future Δt = 1 ms based on the current phase point X(t). The chaotic trajectory processed by the phase-space reconstruction is converted into an analyzable dynamic pattern, providing a data basis for torque prediction.

[0101] Furthermore, based on the phase-space trajectory, the future torque mutation is extrapolated to generate a feedforward control instruction;

[0102] First, the direction of the torque is determined according to the sign of the derivative of the phase-space trajectory, and then the torque amplitude is calculated:

[0103]

[0104] In the formula, k = 0.1, which is a calibration coefficient;

[0105] Furthermore, the duration of the predicted torque is predicted:

[0106]

[0107] Finally, the prediction results of the torque direction, amplitude and acting time are output.

[0108] Furthermore, in combination with Figure 1As shown, in Step 2, cross-modal strategy preloading and power excitation prepare the vehicle for the smooth completion of the cross-medium movement from water to air, generating a reasonable control strategy to ensure the stability and efficiency of the entire cross-medium process;

[0109] Specifically, the method of cross-modal strategy preloading and power excitation includes

[0110] Data input and algorithm startup: Input the current attitude and speed of the vehicle, as well as the perturbation data such as the direction, amplitude, and action time of the fluid moment mutation predicted by the chaotic phase synchronization engine in Step 1, into the fractal recursive reinforcement learning algorithm together;

[0111] Furthermore, use the fractal recursive reinforcement learning algorithm to map the above input data to the interface transition layer of the Sierpinski carpet fractal space. In this special fractal space, the algorithm deeply analyzes and calculates the data according to its unique rules and structures;

[0112] Still further, perform data preprocessing: Pass the pressure sensor through the FPGA module for low-pass filtering (cutoff frequency 5 kHz) to eliminate high-frequency noise; normalize it to the range [-1, 1]; use sliding average filtering for inertial navigation data to smooth the acceleration and angular velocity data and eliminate instantaneous jitter;

[0113] Perform principal component analysis on the preprocessed multi-dimensional feature matrix, project the original data to the three-dimensional principal component space, and obtain the three-dimensional principal component score vector as the input for judging the structural level:

[0114] υ PCA =[υ1,υ2,υ3]

[0115] Specifically, after normalizing the state space, map it to the levels of the fractal structure:

[0116]

[0117] In the formula, 1 represents the underwater state, 2 represents the transition state, i.e., water-air switching, 3 represents the air state, T water , T air represents the stratification threshold between the underwater and air states; T water =1.4, T air =2.2; ||ν PCA || represents the norm of the eigenvector after principal component analysis, used to measure the significance of the state in a specific medium, and the threshold represents the preset stratification threshold for distinguishing different medium states;

[0118] T water 、T airParameter determination method: When the vehicle is completely submerged, record the data of the pressure sensor, inertial navigation unit, and liquid level sensor; after the vehicle completely leaves the water surface, collect the data of the Doppler velocimeter, pressure sensor, and attitude. For each state, 100 sets of typical working condition data need to be collected, covering different environmental conditions such as flow velocity, water depth, and airspeed to ensure data diversity.

[0119] Furthermore, perform data preprocessing: denoise the pressure signal, normalize to eliminate the dimension difference, and perform time synchronization to ensure the alignment of sensor data. Input the preprocessed multi-dimensional feature matrix; perform standardization processing on the data, calculate the covariance matrix, extract the first 3 principal components, and project the data into the principal component space to obtain the 3-dimensional principal component score vector ν of each sample. PCA ; For the ν of each sample PCA Calculate the Euclidean norm ν PCA ;

[0120] Still further, perform statistical analysis and threshold calibration, and analyze the underwater state: After calculating the norm mean and standard deviation of all underwater state samples, it is obtained that:

[0121] T water = μ water + 2σ water = 1.4

[0122] In the formula, μ water , σ water represent the norm mean and standard deviation respectively;

[0123] Similarly, after calculating the mean and standard deviation of the air state samples, it is obtained that:

[0124] T air = μ air + 2σ air = 2.2

[0125] Furthermore, to improve the time series prediction ability of the strategy, perform time series modeling of the recurrent neural network, input the state sequence of the past 10 moments: s t-9 , s t-8 , …, s t , and perform training in the hidden layer, and calculate the hidden state of each layer of GRU in turn. The loss function is:

[0126]

[0127] In the formula, h t is the hidden state predicted by the RNN, is the true value;

[0128] Finally, output the true hidden state h t of the current moment, encode the historical information, and use it for the generation of the reinforcement learning strategy;

[0129] Furthermore, based on the current state and historical information, the DDPG algorithm is used to generate and update the policy to obtain the optimal control parameters;

[0130] Specifically, according to the DDPG algorithm, the policy is generated, and the reward value at time t is output to evaluate the quality of the action;

[0131] In the Actor network, the true hidden state h is input t , and the action control parameter a is output t ;

[0132] In the Critic network, the true hidden state h t is concatenated with the action generated by the Actor network as its input, and the value of the state-action pair is evaluated through the output Q value to guide policy optimization;

[0133] Based on the parameters output by the two networks, the policy is updated based on DDPG:

[0134]

[0135] In the formula, represents the policy gradient, which is the partial derivative of the objective function J with respect to the policy parameter θ μ , θ μ represents the parameters in the Actor network, Q(s,a|θ Q ) represents the Q-value function of the Critic network, which evaluates the value of the state s and the action a, and θ Q represents the parameters of the Critic network, μ(s|θ μ ) represents the policy function of the Actor network, which outputs the action a;

[0136] Furthermore, after the policy optimization is completed, the present invention adopts fractal hierarchical policy transfer to achieve fast policy adaptation in cross-media scenarios and avoid repeated training;

[0137] Specifically, the control policies between adjacent levels are pre-trained, and according to the similarity between the current state and the target level in terms of attitude angle, speed, etc., the attention weight α is calculated:

[0138]

[0139] In the formula, represents the cosine similarity between x and y;

[0140] The current level policy and the adjacent level policy are fused according to the weight:

[0141] α fused =α·α current +(1 - α)·αneighbor

[0142] In the formula, α current is the current layer strategy, and α neighbor is the adjacent layer strategy;

[0143] Furthermore, the control parameter dynamic output module maps the abstract policy actions to the instructions of the actual actuator, ensuring that the vehicle can accurately execute the control strategy;

[0144] Specifically, the output policy, that is, the normalized action, is mapped to the physical range. For the pitch angle:

[0145]

[0146] In the formula, θ cmd represents the decoded pitch angle instruction;

[0147] After obtaining the instruction, time series smoothing and instruction distribution processing are performed, that is, the sudden change instruction is subjected to first-order low-pass filtering with a cut-off frequency of 10 Hz to prevent mechanical shock; it is sent to actuators such as the steering gear and propulsion motor through the CAN bus, and the response delay < 1 ms;

[0148] Furthermore, start the tail propeller to make it start running and provide propulsion force for the vehicle. At the same time, start the pre-charged propulsion motor, which has completed the charging preparation before and can quickly respond and output power at this time. Under the combined action of the two power components, the vehicle gradually accelerates in the water;

[0149] During the acceleration process of the vehicle, its speed is monitored in real time. When the speed of the vehicle reaches the pre-set threshold speed, it indicates that the vehicle has stored the kinetic energy required to break through the water surface and meets the conditions for entering the next control step.

[0150] Furthermore, as shown in Figure 1 In step 3, the structure deployment and transient energy coupling during the transition from water to air include,

[0151] The pre-charged propulsion system accelerates the vehicle to break through the threshold. When the capacitive liquid level sensor array detects a sudden change in the water surface contact state, it triggers the shape memory alloy hinge to quickly deploy the folding wing and rotor structure within 80 ms. This structure is made of carbon fiber composite material, and the built-in micro strain sensor can provide real-time feedback on the deployment angle to ensure that the aerodynamic surface is in place accurately.

[0152] Furthermore, while deploying the structure, trigger the chaotic phase folding algorithm, convert the fluid perturbation energy predicted in step 1 into controllable propulsion energy through the chaotic phase folding technology, process the fluid moment data predicted in step 1 in real time, and inject the energy pulse after phase space compression into the propulsion system;

[0153] Specifically, the surface fluid pressure pulsation of the vehicle monitored by the input pressure sensor in real time and the spatio-temporal distribution of the disturbance energy predicted by Lorenz oscillator synchronization. When CPSE locks the peak moment of the disturbance energy, extracts its amplitude, frequency and acting direction, and then uses wavelet transform to separate the high-frequency disturbance component in the range of 1 - 3 kHz as the target frequency band for energy capture;

[0154] Further, based on the parameters in step one, map this frequency band to the three-dimensional phase space trajectory, filter out the noise through principal component analysis and retain the core energy characteristics, and finally convert this energy characteristic to the working frequency band of the propulsion system. According to the predicted torque direction, distribute the energy to the corresponding thrusters.

[0155] Specifically, combined with Figure 3 As shown, input the torque direction, amplitude and acting time predicted by the chaotic phase synchronization engine in step one to generate the reverse torque demand. For example, to counteract the upward trend of the aircraft nose, a reverse torque with the nose pointing down in the pitch direction needs to be generated. According to the torque balance equation, the amplitude of the reverse torque should be equal to the predicted torque and in the opposite direction.

[0156] Further, filter out the noise of the 1 - 3 kHz high-frequency disturbance component captured in step one through principal component analysis, extract the core energy characteristics, and convert them to the working frequency band of the propulsion system;

[0157] Calculate the total energy demand according to the predicted torque amplitude and acting time:

[0158] E = τ·Δt

[0159] In the formula, E represents the total energy demand, that is, the total energy that the propulsion system needs to output; τ represents the predicted disturbing torque; Δt represents the torque acting time; this energy needs to be output through the propulsion system.

[0160] Further, based on the fractal recursive reinforcement learning algorithm, combined with the current attitude and dynamic constraints of the vehicle, dynamically optimize the thrust distribution ratio;

[0161] Specifically, the thrust arm of the tail propeller is L1, the thrust arm of the vector nozzle is L2, and the thrusts are F1 and F2 respectively. Then the total reverse torque is:

[0162] τ 反向 = F1·L1 + F2·L2

[0163] Under the goal of minimizing energy consumption:

[0164]

[0165] Optimize and solve for F1 and F2;

[0166] Allocate energy to the thrusters according to the thrust requirement. The relationship between energy and thrust is as follows:

[0167]

[0168] In the formula, E i represents the required energy of a single thruster; F i represents the thrust of the corresponding thruster; t represents the energy application time. Substituting the data, the energy ratio of the tail propeller and the vector nozzle can be obtained, and the total energy is input to each thruster according to this ratio;

[0169] Finally, smooth the command mutation through first-order low-pass filtering and send it to the actuator through the CAN bus with a 1ms delay.

[0170] For example, when the moment direction predicted by CPSE is the pitch direction with the nose up, the moment amplitude is 50 N·m, and the action time is 10 ms, different proportions of energy are given to the tail propeller and the vector nozzle respectively, so that the downward thrust can offset the upward trend of the nose;

[0171] After this step, the surface tension of the vehicle is offset and the initial lift is increased, ensuring the dynamic stability of the medium switching process;

[0172] Furthermore, within 5 ms after the wing deployment is completed, the fractal controller switches from the interface transition layer to the air level through the CAN bus and loads the pre-trained sliding mode control parameters, which are stored in the non-volatile memory, to achieve a millisecond-level smooth transition of the water-air dynamics model.

[0173] Furthermore, as shown in Figure 1 the pneumatic steady-state control and fractal recursive optimization in step four include

[0174] After the amphibious vehicle emerges from the water, the traditional PID controller takes over the attitude stabilization task and maintains the angle of attack and roll angle stability through aileron deflection and thrust vector adjustment;

[0175] Furthermore, the fractal recursive reinforcement learning module analyzes the aerodynamic parameters and environmental data in real time based on the DDPG algorithm and the policy update algorithm in step two, dynamically adjusts and optimizes the control weights to optimize the energy consumption and anti-turbulence performance. When encountering airflow disturbances, the system will automatically increase the aileron control authority to improve the anti-interference ability;

[0176] Furthermore, the chaotic oscillator continuously monitors the characteristics of the flow field disturbance, updates the prediction model in an incremental learning manner, and adjusts the chaotic parameters in real time according to the synchronization error E sync < 0.01, and the integral value of the past N prediction errors:

[0177]

[0178] where K p represents the proportional gain coefficient, which controls the direct impact of the current error on parameter adjustment, and K i represents the integral gain coefficient, which controls the long-term impact of the historical error accumulation on parameter adjustment;

[0179] In this way, the prediction error can be reduced, the system robustness can be enhanced, and a closed-loop optimization mechanism of "prediction - control - feedback" can be formed.

[0180] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for cross-medium adaptive collaborative control of water entering air, characterized in that Including the following steps: (1) Multi-source perception and disturbance prediction: Through real-time sensor data, combined with a chaotic phase synchronization engine, high-frequency sampling of fluid pressure pulsations is carried out to predict in advance the mutation direction, amplitude, and action time of the hydrodynamic moment at the moment of water emergence; (2) Input the current attitude and speed of the vehicle, as well as the disturbance data of the mutation direction, amplitude, and action time of the hydrodynamic moment predicted by the synchronization engine in step (1) into the fractal recursive reinforcement learning algorithm. Use the fractal recursive reinforcement learning algorithm to map the above input data to the interface transition layer of the Sierpinski carpet fractal space. In this special fractal space, the algorithm deeply analyzes and calculates the data according to its unique rules and structures. After normalizing the state space, it is mapped to the levels of the fractal structure: Wherein, 1 represents the underwater state, 2 represents the transition state, i.e., water-air switching, 3 represents the air state, T water , T air represents the hierarchical threshold between the underwater and air states, ||ν PCA || represents the norm of the eigenvector after principal component analysis; If Level(s) = 1, go to step (3); if Level(s) = 2, go to step (5); if Level(s) = 3, go to step (4); (3) Based on the current state and historical information, use the DDPG algorithm for policy generation and update to obtain the optimal control parameters, and repeat steps (1) and (2); (4) At the moment of breaking through the water surface, eject and deploy the folding wing and rotor structures, synchronously trigger the chaotic phase folding algorithm, use the predicted disturbance moment direction, amplitude, and action time to calculate the total energy requirement according to the formula. Subsequently, through the fractal recursive reinforcement learning algorithm, dynamically optimize the thrust distribution under the constraint of minimizing energy consumption: establish a moment balance equation, solve the thrust ratio and convert it into an energy distribution ratio. Finally, after smoothing the energy proportionally through low-pass filtering, it is sent to the thruster through the CAN bus with a <1ms delay, and switch the fractal controller to the air level; (5) After emerging from the water, the traditional PID controller takes over the flight attitude stabilization, maintains the angle of attack and roll angle stability through aileron deflection and thrust vector adjustment. The fractal recursive reinforcement learning module analyzes the aerodynamic parameters in real time, dynamically adjusts the control weights, and at the same time the chaotic oscillator continuously monitors the flow field disturbance characteristics and updates the prediction model in an incremental learning manner.

2. The cross-medium adaptive cooperative control method from water to air according to claim 1, wherein In step (1), the sensors include a Doppler velocimeter, a nine-axis MEMS inertial navigation unit, a capacitive liquid level sensor array, and a micro pressure pulsation sensor; The Doppler velocimeter is used to measure the three-dimensional velocity of the vehicle relative to the water body with an accuracy of ±0.1m / s. It is deployed in the streamlined cabin at the tail to avoid turbulence interference and finally outputs 100 velocity vector updates per second; The nine-axis MEMS inertial navigation unit is used to collect three-axis acceleration, three-axis angular velocity, and three-axis magnetic field data in real time, and finally outputs a sampling rate of 400Hz; The capacitive liquid level sensor array is used to detect the water surface contact state and local liquid level height changes with a resolution of ±1mm, and finally outputs the 0 / 1 contact state and the analog liquid level height; The micro pressure pulsation sensor is used to capture the high-frequency pulsations in the 0-10kHz frequency band of the fluid pressure on the surface of the vehicle. After final AD acquisition, it is preprocessed by the FPGA and transmitted through the Ethernet.

3. The water-into-air cross-medium adaptive cooperative control method according to claim 1, characterized in that, In step (1), the prediction method includes the following steps: (11) Initialize the chaotic oscillator to construct the initial state of the chaotic system. The Lorenz system is used as the chaotic oscillator, and its differential equation is: wherein, the values of the respective parameters are: σ = 10, ρ = 28; (12) Map the pressure amplitude p(t) to the chaotic variable range for signal normalization processing; inject the normalized pressure signal p′(t) into the Lorenz system and modify the first equation as follows: where ε = 0.3 is the coupling strength; (13) Align the phase of the forced chaotic oscillator with the external disturbance signal to establish a disturbance-chaos correlation model, and adopt Coupling protocol: Introduce a coupling term among three groups of parallel Lorenz oscillators: where γ = 0.5 is the coupling coefficient between oscillators, and i, j are the oscillator numbers; (14) Calculate the synchronization error: wherein is the mean value of the states of three groups of oscillators. If the synchronization error E sync < 0.01, go to step (15); if E sync ≥ 0.01, adjust the coupling coefficient γ, and repeat steps (13)-(14); (15) Map the synchronized chaotic trajectory to a high-dimensional phase space, capture the time-evolution characteristics of the perturbation, and perform phase-space reconstruction; (16) Generate a feedforward control command based on the result of the phase-space reconstruction.

4. The water-into-air cross-medium adaptive cooperative control method according to claim 2, characterized in that In step (15), the delay embedding method is used for phase-space reconstruction. The delay time τ = 5 ms and the embedding dimension m = 3 are selected to reconstruct the phase-space trajectory: X(t) = [x(t), x(t - τ), x(t - 2τ)]; where τ represents the delay time, which is determined by the mutual information method or the autocorrelation function to ensure that the information between adjacent delay points is independent and the dynamic characteristics are retained; x(t - τ) and x(t - 2τ) represent the observed values at the previous and two delay time points respectively.

5. The method for water-into-air cross-media adaptive collaborative control according to claim 2, characterized in that Step (16) extrapolates the future torque mutation based on the phase-space trajectory to generate a feedforward control command, that is, the torque direction, amplitude, and action time, including the following steps: (161) Determine the direction of the torque according to the sign of the derivative of the phase-space trajectory and calculate the torque amplitude: X is the phase space trajectory X(t) where k = 0.1 is the calibration coefficient; (162) Predict the duration of the torque: (163) Finally, output the predicted results of the torque direction, amplitude, and action time.

6. The water-into-air cross-media adaptive cooperative control method according to claim 2, characterized in that, The specific content of step (2) is as follows: (21) Input the current attitude and speed of the vehicle, as well as the perturbation data of the fluid torque mutation direction, amplitude, and action time predicted by the chaotic phase synchronization engine in step one, into the fractal recursive reinforcement learning algorithm; (22) Use the fractal recursive reinforcement learning algorithm to map the above input data to the interface transition layer of the Sierpinski carpet fractal space; (23) Perform data preprocessing. The data from the pressure sensor is low-pass filtered through the FPGA module with a cut-off frequency of 5 kHz, and normalized to the range [-1, 1]. The inertial navigation data uses a moving average filter to smooth the acceleration and angular velocity data and eliminate instantaneous jitter. The preprocessed multi-dimensional feature matrix is subjected to principal component analysis, and the original data is projected onto a three-dimensional principal component space to obtain a three-dimensional principal component score vector as the input for judging the structural hierarchy: υ PCA = [υ1, υ2, υ3], where v1, v2, and v3 are the data scores on the three principal components obtained after principal component analysis; After normalizing the state space, map it to the level of the fractal structure: In the formula, 1 represents the underwater state, 2 represents the transition state, i.e., water-air switching, 3 represents the air state, T water , T air represents the stratification threshold between the underwater and air states; ||ν PCA || represents the norm of the eigenvector after principal component analysis, which is used to measure the significance of the state in a specific medium. The threshold represents the preset stratification threshold for distinguishing different medium states.

7. The water-into-air cross-medium adaptive collaborative control method according to claim 3, characterized in that In the said step (4), the thrust arm of the tail propeller is L1, the thrust arm of the vector nozzle is L2, and the thrusts are F1 and F2 respectively. Then the total reverse moment is: τ 反向 = F1·L1 + F2·L2, τ 反向 That is the moment amplitude and direction predicted in step (16). Under the goal of minimizing energy consumption: Optimize and solve for F1 and F2; According to the thrust requirement, energy is allocated to the thrusters in real time. The relationship between energy and thrust is as follows: In the formula, E m represents the required energy of a single thruster m; F m represents the thrust of the corresponding thruster; t represents the energy application time. Substituting the data, the energy ratio of the tail propeller and the vector nozzle can be obtained, and the total energy is input to the corresponding two thrusters according to this ratio.

Images