An experimental device for reducing the wake characteristics of underwater vehicle appendages through jet excitation.

Abstract

This invention discloses an experimental apparatus for reducing the wake characteristics of underwater vehicle appendages through jet excitation. The apparatus includes a water tank system, a model motion system, a PIV flow field measurement system, a blow-suction jet excitation system, and an intelligent controller. The appendage model is assembled from 3D-printed blocks, and the jet nozzle layout is adjustable. The PIV system acquires the velocity field in the wake region in real time. The intelligent controller, based on deep reinforcement learning, uses the velocity field as the state variable and the jet intensity as the action variable to generate control commands online to drive the jet excitation, forming a real-time closed loop of "measurement-solution-decision-control". This invention achieves adaptive suppression of the velocity deficit and turbulent kinetic energy of the underwater vehicle appendage wake, and has the advantages of high control accuracy, fast response, and adaptability to various jet layouts.

Description

Technical Field

[0001] This invention belongs to the fields of fluid mechanics, flow control and underwater vehicles, and specifically relates to an experimental device for reducing the wake characteristics of underwater vehicle appendages through jet excitation. Background Technology

[0002] During navigation, the appendages of underwater vehicles (such as rudders, fins, and supports) generate complex wake structures at their tails, accompanied by significant velocity losses and high turbulent kinetic energy characteristics. These wakes not only increase the vehicle's hydrodynamic drag and reduce propulsion efficiency but can also generate hydrodynamic noise, affecting stealth capabilities. Furthermore, the unsteady characteristics of the wakes can cause appendage vibrations, reducing structural reliability. Therefore, effectively suppressing the wakes of underwater vehicle appendages is one of the key technologies for improving the overall performance of underwater vehicles.

[0003] Existing wake control technologies mainly fall into two categories: passive control and active control. Passive control intervenes in the flow field by optimizing the shape of appendages and setting up vortex breaking devices, but it suffers from a narrow range of applicable operating conditions and poor adaptability to complex flow fields. Active control, on the other hand, intervenes in the flow field by adjusting excitation in real time, offering greater flexibility. Among these, blow-suction jets, as an efficient active flow control method, have been proven to effectively modulate shear layer development and suppress wake vortex shedding.

[0004] However, existing active control technologies mostly employ preset control strategies, making it difficult to adjust control parameters in real time according to the dynamic changes in the wake, resulting in insufficient control accuracy and robustness. Meanwhile, real-time measurement of the wake field and rapid response of control signals are the core challenges in achieving efficient wake suppression: traditional flow field measurement methods (such as hot-wire anemometers) have limited measurement ranges and cannot obtain velocity information across the entire wake region; conventional control algorithms have low computational efficiency and cannot meet the requirements of high-frequency real-time control.

[0005] Furthermore, the structures of underwater vehicle appendages vary, and different jet nozzle arrangements significantly affect wake control. Existing experimental devices struggle to quickly adapt to different jet layouts, resulting in low experimental efficiency. Therefore, there is an urgent need to develop an experimental device capable of real-time, high-precision measurement of the wake field, high-frequency intelligent control, and adaptability to various jet layouts, in order to achieve efficient suppression of underwater vehicle appendage wakes. Summary of the Invention

[0006] The technical problem to be solved: To overcome the shortcomings of existing technologies, this invention provides an experimental device for reducing the wake characteristics of underwater vehicle appendages through jet excitation. The device acquires real-time global velocity field information of the wake region using a PIV flow field measurement system. After high-frequency computation by a GPU acceleration module, the information is input into a deep reinforcement learning algorithm module. The algorithm module optimizes the control strategy online based on the reward function, generating a jet intensity control signal to drive the blow-suction jet excitation system to intervene in the wake structure, forming a high-speed closed-loop control link of 'measurement-computation-decision-control'. The control frequency can reach 20Hz, achieving adaptive and efficient suppression of the wake of underwater vehicle appendages.

[0007] The technical solution of this invention is: an experimental device for reducing the wake characteristics of underwater vehicles by jet excitation, comprising: A water tank system is used to provide a controlled fluid experimental environment; The model motion system includes a reconfigurable appendage model and a drive unit, used to drive the appendage model to simulate the motion of an underwater vehicle in the water tank system; The flow field measurement system is used to collect global velocity field information in the wake region of the attached model in real time; A jet excitation system, arranged on the surface of the attached model, is used to apply adjustable blow-suction jet excitation to the wake region; The intelligent controller is connected to both the flow field measurement system and the jet excitation system. It uses the velocity field information fed back by the flow field measurement system in real time as the state input, generates jet control commands online based on a deep reinforcement learning model, and drives the jet excitation system to execute the corresponding jet actions, forming a real-time closed-loop control to achieve adaptive suppression of velocity deficit and turbulent kinetic energy in the wake of the underwater vehicle's appendages.

[0008] A further technical solution of the present invention is: the flow field measurement system is a PIV flow field measurement system, including a laser and an underwater high-speed camera, wherein the frame rate of the underwater high-speed camera is not less than 200fps; the intelligent controller includes a GPU acceleration module, runs a PIV flow field calculation program based on the LK optical flow algorithm, calculates the velocity components of the wake region measurement point array in real time, and the calculation frequency can reach up to 50Hz, while the control frequency is maintained at 20Hz.

[0009] A further technical solution of the present invention is: the attachment model adopts a 3D printed modular assembly structure, and the surface of the model is reserved with jet port installation interfaces. The number, position, angle, shape and longitudinal spacing of the jet ports can be flexibly configured. The jet ports adopt a symmetrical arrangement, which supports the switching of jet configurations with different nozzle spacing, different axis angles, different elliptical nozzle shapes and multiple rows of longitudinal layouts, so as to adapt to the wake control experimental requirements of different types of attachments.

[0010] A further technical solution of the present invention is that the jet outlet on the surface of the attached model has at least one of the following configurable layouts: Adjustable spanwise spacing: Multiple jet nozzles are arranged along the spanwise direction of the airfoil. By changing the spacing between adjacent jet nozzles, the mutual interference and merging effects of the jets are studied, and the minimum effective nozzle density to suppress flow separation is determined. Adjustable axis angle: Jet outlets are set at symmetrical positions on both sides of the airfoil. By adjusting the axis angle of the jet outlets on both sides, a converging jet or a scattering jet can be formed. The effects of different angles on flow reattachment and stall delay under conditions of large angle of attack and strong adverse pressure gradient are tested to achieve directional momentum injection. Adjustable nozzle shape: Elliptical nozzles with different major and minor axis ratios are used to change the jet diffusivity and turbulence while keeping the nozzle area constant, so as to meet the fine control requirements of boundary layers with different thicknesses. Adjustable longitudinal spacing: Multiple rows of jet nozzles are arranged along the airfoil chord. By adjusting the longitudinal spacing between the front and rear rows, a multi-level control layout with front row excitation and rear row compensation is formed, and the spatial collaborative control effect of multi-level jets is studied.

[0011] A further technical solution of the present invention is as follows: the intelligent controller adopts the Proximal Policy Optimization (PPO) algorithm as a deep reinforcement learning framework and adopts an online incremental training method, using the flow field data collected in real time by the flow field measurement system as continuous input; in the policy update, the change amplitude of adjacent steps is limited, and the amplitude of policy update is limited by clip operation, so as to improve control stability and convergence speed.

[0012] A further technical solution of the present invention is as follows: the deep reinforcement learning model uses a one-dimensional array formed by integrating the velocity components of each measurement point calculated by PIV as the state variable, and uses the jet intensity, direction, and action sequence as the action variables, and embeds the physical constraints of jet flow rate and action direction into the output layer of the policy network; the reward function of the deep reinforcement learning model takes minimizing the average velocity loss as the optimization objective, and the reward function is defined as:

[0013] in, i Indicates the measurement point number. n Indicates the total number of measuring points. This indicates a loss in the flow rate. This represents the intensity of longitudinal velocity fluctuations; a larger reward function value indicates a better wake suppression effect. Meanwhile, turbulent kinetic energy is used as an auxiliary evaluation index to monitor the stability of the control process and as a reference condition for training convergence, thereby improving control stability and suppressing wake pulsation and structural vibration.

[0014] A further technical solution of the present invention is: the model motion system further includes a PLC control cabinet, which is communicatively connected to the intelligent controller and supports manual parameter input mode and automatic synchronization command mode; in automatic mode, the PLC control cabinet receives the synchronization motion command sent by the intelligent controller to realize the coordinated linkage of model motion, PIV measurement and jet excitation.

[0015] A wake suppression method based on the experimental setup includes the following steps: Step 1: Configure the jet port layout of the attached model according to the experimental plan, adjust the PIV measurement area to cover the wake core area, and initialize the model motion parameters and deep reinforcement learning hyperparameters. Step 2: Start the model motion system to move the attached model at a constant speed in the water tank, and at the same time start the PIV flow field measurement system to collect particle images in the wake region at a set frequency. Step 3: Real-time processing of PIV images using GPU acceleration module to obtain instantaneous velocity components at each measurement point, which are then used as state inputs to the deep reinforcement learning model; Step 4: The deep reinforcement learning model generates the optimal jet intensity control signal online based on the current state variables and the reward function, driving the jet excitation system to apply blowing and suction jet excitation to the wake region; Step 5: Repeat steps 3 to 4 to form a continuous closed-loop control until the wake suppression effect meets the preset convergence condition.

[0016] A further technical solution of the present invention is: in step 4, the deep reinforcement learning model adopts an online incremental training method, with PIV real-time flow field data as continuous input, limits the change range of adjacent steps in policy update, and limits the range of policy update through clip operation, so as to improve control stability and convergence speed. A further technical solution of the present invention is: the convergence condition in step 5 is: the change in wake parameters is less than 5% within 500 consecutive control cycles, and the wake parameters include average velocity deficit and turbulent kinetic energy; after the experiment, the system automatically generates an analysis report, including wake parameter change curves over time and control effect comparison charts.

[0017] Beneficial effects The beneficial effects of this invention are as follows: This invention forms a complete wake suppression closed loop through the organic coupling of a PIV flow field measurement system, a 3D-printed assemblable attached model, a deep reinforcement learning intelligent control strategy, and a PLC collaborative control system. The integration of deep reinforcement learning with each system is key to achieving precise dynamic wake suppression, as detailed below: 1. Purpose of the integrated design: In response to the problem that existing technologies for wake suppression often rely on preset fixed parameters, which cannot adapt to the dynamic changes in the wake of underwater vehicles (such as velocity deficit fluctuations and uneven distribution of turbulent kinetic energy), this invention deeply integrates deep reinforcement learning with PIV flow field measurement and jet control systems. Using real-time wake data as input and wake suppression effect as the target, it achieves adaptive adjustment of the control strategy, breaks through the limitations of "disconnect between measurement and control and fixed control parameters", and constructs a dynamic response closed-loop control system.

[0018] 2. Integrating the advantages of design and the coupling effect of various features: (1) Coupling of measurement and control: PIV flow field measurement technology combined with GPU-accelerated LK optical flow algorithm provides real-time, high-precision wake data input for deep reinforcement learning, ensuring that the reinforcement learning model can accurately perceive the dynamic changes of the wake (such as velocity components, velocity deficit, and real-time fluctuations of turbulent kinetic energy); at the same time, the control decision results of deep reinforcement learning can guide the PIV measurement system to focus on the measurement points, prioritize the collection of data in key areas of wake suppression, further improve measurement efficiency and pertinence, and form a virtuous cycle of "measurement-decision-feedback-optimized measurement".

[0019] (2) Coupling of control strategy and experimental carrier: The flexible arrangement of jet ports in the 3D printed modular assembly attachment model provides adaptability support for the implementation of deep reinforcement learning control strategy. The reinforcement learning model can adaptively adjust the blowing and suction intensity ratio of each jet port according to different attachment layouts and different wake distributions. The flexible jet arrangement design is just right to meet the needs of reinforcement learning for multi-parameter and dynamic control, solves the problem that the traditional fixed jet layout cannot adapt to dynamic control strategy, and greatly improves the adaptability range of closed-loop control.

[0020] (3) Coordinated coupling of the overall system: The PLC control cabinet and the computer realize the coordinated linkage of PIV measurement, reinforcement learning decision-making and jet control, and switch between manual and automatic modes. This ensures the stability of the automatic control of deep reinforcement learning, and allows for manual intervention in special experimental scenarios to ensure the reliability of closed-loop control. At the same time, the integrated design of data acquisition, processing and storage provides sufficient historical data support for the training and optimization of deep reinforcement learning models, which can continuously improve the control accuracy of the model and form a long-term mechanism of "data accumulation - model optimization - control effect improvement". Attached Figure Description Figure 1 This is a schematic diagram of the overall layout of the experimental setup; Figure 2 These are detailed schematic diagrams of each subsystem; Figure 3 It represents the expected trend of the reward function curve obtained during training under experimental conditions.

[0021] Figure labeling: 1. PLC control cabinet, 2. Main unit, 3. Experimental water tank, 31. Observation window, 4. Towable platform, 5. Underwater camera, 6. Linear motor guide rail, 7. Laser, 8. Jet pump set, 9. Model. Detailed Implementation

[0022] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0023] In recent years, researchers have begun to explore the application of intelligent control algorithms in the field of active flow control. For example, patent CN115482706A discloses an experimental device for achieving hydrodynamic stealth of bluff bodies through active flow control. It employs deep reinforcement learning (DRL) to generate control strategies and modulates the wake vortex structure through jet extraction / absorption to eliminate the velocity deficit caused by bluff body motion. Patent CN117951995A further proposes a cylindrical active flow control method based on deep reinforcement learning. It utilizes dynamic feature enhancement technology to extract flow field states from sparse sensor data, achieving drag reduction control of the flow around a cylinder. However, these studies are mainly based on numerical simulations or applied to simple bluff bodies (cylinders, square cylinders). Their control strategies rely on a pre-defined sensor layout (such as sparsely arranged pressure or velocity probes in the wake), making it difficult to directly extend to complex underwater vehicle attachment scenarios with strong three-dimensional wake fields.

[0024] For the specific engineering problem of suppressing the wake of underwater vehicle appendages, existing active control technologies still have the following shortcomings: First, the control strategies lack real-time performance and adaptive capabilities. Existing active control systems mostly employ preset control parameters or offline-trained control models, making it difficult to adjust jet excitation parameters in real time based on dynamic changes in the wake (such as instantaneous fluctuations in velocity deficit and non-uniform distribution of turbulent kinetic energy). This is especially true at Reynolds numbers of 10. 5 Under large-scale operating conditions, the wake field exhibits strong nonlinear and transient characteristics, and fixed control strategies cannot adapt to changes in the flow regime, making it difficult to guarantee control accuracy and robustness.

[0025] Second, the coordination between flow field measurement and control response is poor. A prerequisite for efficient wake suppression is the ability to acquire wake field information in real-time and across the entire region, and to quickly generate control commands based on this information. Traditional flow field measurement methods (such as hot-wire anemometers and single-point pressure probes) can only acquire information from a limited number of measurement points, failing to fully characterize the velocity deficit distribution and vortex structure evolution in the wake region. Meanwhile, conventional control algorithms (such as PID and fuzzy control) have low computational efficiency, making it difficult to meet the demands of high-frequency real-time control. In existing research, although PIV (Particle Image Velocimetry) has been used for flow field measurement (e.g., CN114878137A), it is typically used only as an offline testing method and does not form a closed-loop linkage with the control algorithm.

[0026] Third, the experimental setup has poor adaptability to jet layout. The structures of underwater vehicle appendages (such as rudders, fins, and supports) vary, and different jet port arrangements (number, position, angle, and spacing) significantly affect wake control performance. Existing experimental setups typically have fixed jet port designs, making rapid reconfiguration to adapt to different jet layout schemes difficult. This results in low experimental efficiency and fails to provide a diverse control action space for deep reinforcement learning.

[0027] To address the aforementioned problems, this invention proposes an experimental apparatus for reducing the wake characteristics of underwater vehicle appendages through jet excitation, comprising: A water tank system is used to provide a controlled fluid experimental environment; The model motion system includes a reconfigurable appendage model and a drive unit, used to drive the appendage model to simulate the motion of an underwater vehicle in the water tank system; The flow field measurement system is used to collect global velocity field information in the wake region of the attached model in real time; A jet excitation system, arranged on the surface of the attached model, is used to apply adjustable blow-suction jet excitation to the wake region; The intelligent controller is connected to both the flow field measurement system and the jet excitation system. It uses the velocity field information fed back by the flow field measurement system in real time as the state input, generates jet control commands online based on a deep reinforcement learning model, and drives the jet excitation system to execute the corresponding jet actions, forming a real-time closed-loop control to achieve adaptive suppression of velocity deficit and turbulent kinetic energy in the wake of the underwater vehicle's appendages.

[0028] The above technical solution will be further analyzed below with reference to examples and accompanying figures: In one embodiment, taking an underwater vehicle control surface appendage model as an example, at a Reynolds number of 10... 5 Under high-speed operating conditions, the experimental apparatus and method described in this invention are verified to suppress wake velocity deficit and turbulent kinetic energy.

[0029] I. Specific Composition of the Experimental Apparatus The underwater vehicle appendage wake suppression experimental device of this embodiment includes a water tank system, a model motion system, a PIV flow field measurement system, a blow-suction jet excitation system, and an intelligent control and data processing system. These systems work together to achieve real-time wake suppression. The specific scheme is as follows: 1. Water Tank System: A long, straight, transparent water tank is used, with a length of no less than 10m, a width of no less than 1.5m, and a depth of no less than 1.2m, to ensure that the flow field develops fully and without boundary effect interference during the movement of the attached model. The inner wall of the water tank is made of optical-grade transparent material, and a shock-absorbing support device is installed at the bottom to avoid the impact of external vibration on the accuracy of PIV measurements. The water tank is filled with deionized water, and non-toxic, highly reflective tracer particles (particle size 5-50μm, density close to that of water) are added to the water to ensure that the PIV system can clearly capture the particle motion trajectory.

[0030] 2. Model Motion System: This system includes a support frame, a linear motor, a PLC control cabinet, and the attached model. The support frame is constructed from high-strength aluminum profiles, with sliding guide rails at the bottom to fit the water tank, ensuring smooth movement. The linear motor is fixed to one end of the water tank and connected to the support frame via a coupling. It drives the frame to move the attached model along the water tank axis at a uniform linear speed, ranging from 0.1 to 3 m / s, satisfying a Reynolds number of 10. 5 Requirements for high-volume operating conditions.

[0031] The appendage model employs a 3D-printed, modular assembly structure. The main body of the model is made of high-strength photosensitive resin, and the surface is polished to ensure smooth flow. The model surface has pre-drilled interfaces for jet nozzle installation, and the modules are connected by bolts for easy assembly and disassembly. The nozzle placement strategy involves symmetrically arranging jet nozzles on both sides of the airfoil appendage. By adjusting the spacing, angle, shape, and longitudinal spacing of the jet nozzles, diverse jet excitation geometries can be constructed. Each placement method corresponds to specific and targeted experimental requirements.

[0032] The jet inlets on the surface of the attached model have at least one of the following configurable layouts: When the jet nozzle is fixed, the spacing along the span of the airfoil is changed. The corresponding experimental requirements are to study the mutual interference and jet merging effect between jets under different spacings, determine the minimum effective nozzle density to suppress flow separation, and compare the appendage flow control effect corresponding to different spacings. This will provide low-energy and high-control-efficiency nozzle arrangement parameters for engineering applications. b. When the jet nozzle position is fixed, adjust the angle between the jet nozzles on both sides and the mid-arc line to form a converging jet or a scattering flow. The corresponding experimental requirements are to explore the characteristics of jet vector synthesis and the wall jet adhesion effect, test the influence of different angles on flow reattachment and stall delay under large angle of attack and strong adverse pressure gradient conditions, and achieve directional momentum injection to meet the experimental objectives of lift enhancement and drag suppression. c. Using elliptical nozzles with different semi-axis lengths, while keeping the nozzle area approximately constant, the ratio of the major and minor axes of the ellipse is changed. The corresponding experimental requirements are to study the influence of nozzle shape on jet diffusivity, turbulence intensity and near-wall jet thickness, compare the efficiency differences of different elliptical nozzles in boundary layer momentum injection and mixing layer control, and adapt to the fine flow control requirements of boundary layers with different thicknesses. Multiple jet nozzles are arranged along the chord of the airfoil. The longitudinal spacing of the jet nozzles is adjusted to form a multi-stage control layout with front-row excitation and rear-row compensation. The corresponding experimental requirement is to study the spatial cooperative control effect of multi-stage jets, compare the differences between single-row and multi-row longitudinal arrangements in continuous flow control at large angles of attack, explore the suppression mechanism of secondary separation after flow reattachment, and provide a layout basis for flow control in a wide range of operating conditions and wide angles of attack.

[0033] The PLC control cabinet has a built-in motion control module that supports two control modes: one is to manually input motion parameters (speed, stroke, start and stop time) through the built-in touch panel; the other is to communicate with the computer of the intelligent control and data processing system through the RS485 serial port to receive synchronous motion commands and realize coordinated linkage with PIV measurement and jet excitation.

[0034] 3. PIV Flow Field Measurement System: This system includes a laser, an underwater high-definition camera, a laser emitter, and a camera mount. The laser outputs a 532nm green laser beam, which is formed into a sheet of light by an optical lens group. The sheet thickness is controlled to be less than 1mm. This sheet illuminates the wake region of the attached model (0.5-3 times the characteristic length of the model downstream), covering the wake measurement area. The underwater high-definition camera is a high-speed camera with a frame rate of no less than 200fps. The lens is equipped with a waterproof optical lens. The camera is fixed on a three-dimensionally adjustable mount, with the shooting direction perpendicular to the laser sheet to ensure clear capture of the movement images of the tracer particles.

[0035] The measurement area is set up with a measuring point array, and different density of measuring points can be selected according to experimental requirements, such as a 7×7 measuring point distribution of velocity components. u - Along the direction of motion, v The vertical motion direction information is integrated into a one-dimensional array of length 98 as the state input for deep reinforcement learning. The measurement point array covers the core region of the wake (width not less than 2 times the model feature length, and length not less than 3 times the model feature length).

[0036] 4. Blowing and Suction Jet Excitation System: This system includes a high-precision, high-power servo motor, a magnetically coupled gear pump, a jet pipeline, jet nozzles, and a flow sensor. The servo motor is connected to the magnetically coupled gear pump. The pump's outlet / inlet is connected to the jet pipeline via a corrosion-resistant hose. The jet pipeline is embedded inside the attached model and connects to the jet nozzles on the model's surface. The jet nozzles are miniature precision nozzles, supporting both blowing and suction actions. The pump speed range is 0-3000 rpm.

[0037] The servo motor communicates with the computer via an RS485 serial port, receives control signals to adjust its speed, and then controls the output flow and pressure of the gear pump to achieve continuous adjustment of the jet intensity; the magnetic coupling structure avoids pump leakage and ensures the safety of underwater experiments.

[0038] 5. Intelligent Control and Data Processing System: This includes a computer, a GPU acceleration module, and a deep reinforcement learning algorithm module. The computer establishes communication connections with the PLC control cabinet and servo motors via a USB interface, and with the PIV camera via a 10 Gigabit Ethernet card. The GPU acceleration module uses a high-performance NVIDIA graphics card to run a PIV flow field calculation program based on the LK optical flow algorithm, with a maximum calculation frequency of 50Hz, processing particle images acquired by the PIV camera in real time.

[0039] 6. The deep reinforcement learning algorithm module adopts the proximal policy optimization (PPO) algorithm and has been customized as follows for real-time control of the experimental flow field: A deep reinforcement learning control kernel was built using Python and TensorFlow. Employing the Proximal Policy Optimization (PPO) algorithm, it deeply couples real-time flow field measurement, jet excitation, and intelligent decision-making to form a complete closed-loop adaptive control framework. The intelligent decision-making process involves determining the current jet velocity given the measured velocity in the flow field.

[0040] The wake velocity component calculated in real time using PIV is used as the state variable, and the jet intensity, direction, and actuation sequence are used as the action variables. Physical constraints such as jet flow rate and actuation direction are embedded into the output layer of the strategy network to ensure deep integration of intelligent decision-making and jet excitation. Minimizing the average velocity deficit is used as the reward function to optimize navigation resistance and hydrodynamic characteristics. At the same time, turbulent kinetic energy is used as an auxiliary evaluation index for the flow field to improve control stability and suppress wake pulsation and structural vibration.

[0041] In this embodiment, the physical constraints such as jet flow rate and actuation direction are embedded into the output layer of the strategy network as follows: the output layer of the strategy network directly outputs two control parameters: jet flow rate and jet direction angle. The jet flow rate corresponds to the drive signal of the water pump, and the jet direction angle corresponds to the deflection angle of the servo motor.

[0042] A physical constraint module is embedded after the policy network output layer, specifically including: Flow constraint: The jet flow rate shall not exceed 80% of the current rated flow rate of the gear pump.

[0043] Angle constraint: The jet direction angle is limited to ±15°. This constraint is based on the structural strength and sealing requirements of the jet holes on the attached model surface, while ensuring that the jet can effectively act on the shear layer.

[0044] The constrained flow rate and angle values ​​are sent to the pump control board and the servo control board respectively via RS485 serial port: the pump control board adjusts the motor speed to control the jet flow rate; the servo control board drives the servo to deflect the jet nozzle, thereby achieving continuous adjustment of the jet direction.

[0045] This invention presents a training design for real-time control of experimental flow fields based on the existing PPO algorithm: it adopts an online incremental training method, using real-time PIV flow field data as continuous input, without relying on offline datasets or pre-training processes; it introduces flow field temporal sequence into the policy update to limit the change amplitude of adjacent steps, thus avoiding flow field instability caused by violent jet fluctuations; at the same time, it limits the amplitude of policy updates through clip operations, thereby improving the stability and convergence speed of wake control.

[0046] In this embodiment, the online incremental training method consists of the following stages in each training round: Acceleration phase: The linear motor drives the attached model to accelerate from a standstill to the preset speed; Uniform velocity phase: The model maintains uniform motion. During this phase, the PIV system, laser source, power supply, and camera move synchronously with the model under the drive of linear motors, implementing complete closed-loop control: the PIV collects the wake velocity field in real time, and the intelligent controller outputs jet excitation signals according to the current strategy network, driving the water pump and servo motor to execute, forming a "measurement-solution-decision-control" closed loop.

[0047] Deceleration phase: The model decelerates to a stop.

[0048] After completing one round, the reinforcement learning agent updates its policy parameters according to the PPO algorithm: all experience tuples collected during the uniform speed segment are stored in the experience buffer, and PPO is calculated by random sampling to update the policy.

[0049] Subsequently, the linear motor drives the attached model and PIV system back to the starting position, and it is left to stand still for about 2 minutes to allow the flow field in the water tank to return to calm before starting the next round of training.

[0050] During the control process, the PIV system collects and calculates the velocity components at each measuring point in real time. u , v The data is input into the agent's policy network. The network outputs jet control actions that conform to physical constraints (removing excessively large and unreasonable jet velocities) based on the current flow field state, driving the blowing and suction jet excitation system to act on the flow field. After the flow field responds, the PIV system obtains the new flow field state again and calculates the reward value. The PPO algorithm then completes the online update and iterative optimization of the policy parameters. Through the above continuous closed-loop control, the agent continuously learns the optimal control law autonomously, ultimately achieving efficient, real-time, and adaptive suppression of the underwater vehicle's appendage wake.

[0051] The solution process is as follows: First, the particle image is preprocessed (noise reduction, enhancement, matching), and then the instantaneous velocity components of each measurement point are calculated using the LK optical flow algorithm. u - Along the direction of motion, v -Vertical motion direction), and then calculate the velocity deficit (the difference between the velocity at a measuring point and the incoming flow velocity) and turbulent kinetic energy in the wake region. Turbulent kinetic energy TKE It is calculated based on velocity fluctuation components, and the formula is: (1) in, , They are respectively u , v The time-averaged values ​​of directional velocity fluctuations represent the instantaneous and time-averaged velocities of the fluid, respectively. x , y The time average of the square of the directional difference.

[0052] The velocity components at each measuring point obtained by PIV calculation ( u , v The jet intensity (flow rate, direction) is used as the state variable; the jet intensity (flow rate, direction) is used as the action variable; and the reward function is defined. r : (2) in, i Represents the measurement point number. n The total number of measurement points. To reduce the flow rate loss, This represents the intensity of longitudinal velocity pulsation; a larger reward function value indicates a better wake suppression effect.

[0053] II. Experimental Preparation 1. Water tank system: Water is poured into a transparent water tank, and tracer particles (concentration approximately 10 g / m³) are added. 3 Start the built-in stirring device in the water tank to stir thoroughly and ensure that the particles are evenly distributed; check the shock absorption support device of the water tank to ensure that it is not loose.

[0054] 2. Model Motion System: The underwater vehicle control surface is attached to the model. The chord length of the model section is the characteristic length. It is assembled in blocks by 3D printing. Several rows of jet nozzles are arranged along the chord direction on the surface, with several nozzles in each row. The model is fixed to an aluminum profile frame. The connection accuracy between the linear motor and the frame is adjusted to ensure the accuracy of the linearity of the motion.

[0055] 3. PIV flow field measurement system: Adjust the position of the laser so that the laser sheet covers a suitable area downstream of the model, and control the thickness of the sheet to a small level; fix the underwater camera on the downstream side of the model, adjust the lens focal length to make the laser surface particles clear, set the shooting frame rate, and use a matrix array layout for the measurement points, with the spacing between the measurement points gradually increasing according to the distance from the model.

[0056] 4. Intelligent Control and Data Processing System: The computer is equipped with an NVIDIA RTX 4090 GPU, and has a Python-based PIV solver and a deep reinforcement learning framework (Tensorflow) installed; the learning rate and frequency are set.

[0057] III. Experimental Procedure Step 1: Collaborative Startup The PLC control cabinet and PIV system are started, receiving motion commands from the computer via RS485 serial port. Based on the design Reynolds number and water temperature, the linear motor drive model's motion speed, the length of the reusable water tank, and the appropriate start-up delay are calculated and set to ensure flow field stability. Simultaneously, the PIV laser continuously emits laser light, and the underwater camera begins acquiring particle images (20Hz). Image data is transmitted to the computer in real-time via a 10 Gigabit Ethernet interface. The blow-suction jet excitation system remains in standby mode.

[0058] Step 2: Initial Flow Measurement The flow field fully develops after 5 seconds of model motion. 100 consecutive PIV images are acquired to calculate the average velocity deficit and turbulent kinetic energy of the initial wake field.

[0059] Step 3: Online Training and Closed-Loop Control The GPU acceleration module preprocesses the particle image (Gaussian filtering for noise reduction, histogram equalization enhancement, etc.) and calculates the instantaneous velocity components of the measurement point array using the LK optical flow algorithm. u , v The reward function is calculated, and the above parameters are used to calculate the reward function. Input deep reinforcement learning module.

[0060] The action network of the reinforcement learning module adjusts its state variables in real time. u , v The jet intensity control signal is updated and sent to the servo motor of the blow-suction jet excitation system via RS485 serial port; the servo motor responds and adjusts its speed, drives the magnetic coupling gear pump to output the corresponding flow rate, and the jet nozzle applies a blow-suction jet to the wake region.

[0061] Continuous closed-loop control is performed, completing one measurement-solution-decision-regulation cycle according to the control frequency; data is recorded in real time during the experiment, including model motion speed, velocity data at each measuring point, jet intensity, etc.

[0062] IV. Experimental Results The experiment lasted for several seconds and was divided into a start-up phase, a training phase, and a deceleration phase. The data available for processing can be divided into two categories: one is real-time training data based on the measured point velocities, such as state (measured point velocity components), rewards, and action quantities; the other is data analysis performed after training, specifically including: 1. Calculating the instantaneous velocity field and vorticity field of the entire field; 2. Calculating the time-averaged flow field, Reynolds stress in the wake, and turbulent kinetic energy distribution field, and performing further analysis based on these; 3. Calculating the time-averaged values ​​of jet signals in the training flow field for each round to explore the changes learned during the deep reinforcement learning agent's training process. Figure 3 The expected reward function curve converges with the training rounds.

[0063] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. An experimental apparatus for reducing the wake characteristics of underwater vehicle appendages through jet excitation, characterized in that, include: A water tank system is used to provide a controlled fluid experimental environment; The model motion system includes a reconfigurable appendage model and a drive unit, used to drive the appendage model to simulate the motion of an underwater vehicle in the water tank system; The flow field measurement system is used to collect global velocity field information in the wake region of the attached model in real time; A jet excitation system, arranged on the surface of the attached model, is used to apply adjustable blow-suction jet excitation to the wake region; The intelligent controller is communicatively connected to both the flow field measurement system and the jet excitation system. Using the velocity field information fed back in real time by the flow field measurement system as the state input, jet control commands are generated online based on a deep reinforcement learning model, and the jet excitation system is driven to execute the corresponding jet actions, forming a real-time closed-loop control to achieve adaptive suppression of velocity deficit and turbulent kinetic energy in the wake of the underwater vehicle's appendages.

2. The experimental apparatus for reducing the wake characteristics of underwater vehicles by jet excitation according to claim 1, characterized in that: The flow field measurement system is a PIV flow field measurement system, which includes a laser and an underwater high-speed camera. The frame rate of the underwater high-speed camera is not less than 200fps. The intelligent controller includes a GPU acceleration module, which runs a PIV flow field calculation program based on the LK optical flow algorithm to calculate the velocity components of the wake region measurement point array in real time. The calculation frequency can reach up to 50Hz, and the control frequency is maintained at 20Hz.

3. The experimental apparatus for reducing the wake characteristics of underwater vehicles by jet excitation according to claim 1, characterized in that: The attachment model adopts a 3D printed modular assembly structure. The model surface has reserved jet port installation interfaces. The number, position, angle, shape and longitudinal spacing of the jet ports can be flexibly configured. The jet ports adopt a symmetrical arrangement, which supports the switching of jet configurations with different nozzle spacing, different axis angles, different elliptical nozzle shapes and multiple rows of longitudinal layouts, so as to adapt to the wake control experimental requirements of different types of attachments.

4. The experimental apparatus for reducing the wake characteristics of underwater vehicles by jet excitation according to claim 3, characterized in that: The jet inlets on the surface of the attached model have at least one of the following configurable layouts: Adjustable spanwise spacing: Multiple jet nozzles are arranged along the spanwise direction of the airfoil. By changing the spacing between adjacent jet nozzles, the mutual interference and merging effects of the jets are studied, and the minimum effective nozzle density to suppress flow separation is determined. Adjustable axis angle: Jet outlets are set at symmetrical positions on both sides of the airfoil. By adjusting the axis angle of the jet outlets on both sides, a converging jet or a scattering jet can be formed. The effects of different angles on flow reattachment and stall delay under conditions of large angle of attack and strong adverse pressure gradient are tested to achieve directional momentum injection. Adjustable nozzle shape: Elliptical nozzles with different major and minor axis ratios are used to change the jet diffusivity and turbulence while keeping the nozzle area constant, so as to meet the fine control requirements of boundary layers with different thicknesses. Adjustable longitudinal spacing: Multiple rows of jet nozzles are arranged along the airfoil chord. By adjusting the longitudinal spacing between the front and rear rows, a multi-level control layout with front row excitation and rear row compensation is formed, and the spatial collaborative control effect of multi-level jets is studied.

5. The experimental apparatus for reducing the wake characteristics of underwater vehicles by jet excitation according to claim 1, characterized in that: The intelligent controller employs the Proximal Policy Optimization (PPO) algorithm as a deep reinforcement learning framework and uses an online incremental training method, with real-time flow field data collected by the flow field measurement system as continuous input. In policy updates, the magnitude of action changes in adjacent steps is limited, and the magnitude of policy updates is limited through clip operations to improve control stability and convergence speed.

6. The experimental apparatus for reducing the wake characteristics of underwater vehicle appendages through jet excitation according to claim 5, characterized in that: The deep reinforcement learning model uses a one-dimensional array formed by integrating the velocity components of each measurement point calculated by PIV as the state variables, and jet intensity, direction, and actuation timing as action variables. The physical constraints of jet flow rate and actuation direction are embedded into the output layer of the policy network. The reward function of the deep reinforcement learning model aims to minimize the average velocity loss, and the reward function is defined as follows: in, i Indicates the measurement point number. n Indicates the total number of measuring points. This indicates a loss in the flow rate. This represents the intensity of longitudinal velocity fluctuations; a larger reward function value indicates a better wake suppression effect. Meanwhile, turbulent kinetic energy is used as an auxiliary evaluation index to monitor the stability of the control process and as a reference condition for training convergence, thereby improving control stability and suppressing wake pulsation and structural vibration.

7. The experimental apparatus for reducing the wake characteristics of underwater vehicles by jet excitation according to claim 1, characterized in that: The model motion system also includes a PLC control cabinet, which is communicatively connected to the intelligent controller and supports manual parameter input mode and automatic synchronization command mode. In automatic mode, the PLC control cabinet receives synchronization motion commands sent by the intelligent controller to realize the coordinated linkage of model motion, PIV measurement and jet excitation.

8. A wake suppression method based on the experimental apparatus according to any one of claims 1-7, characterized in that, Includes the following steps: Step 1: Configure the jet port layout of the attached model according to the experimental plan, adjust the PIV measurement area to cover the wake core area, and initialize the model motion parameters and deep reinforcement learning hyperparameters. Step 2: Start the model motion system to move the attached model at a constant speed in the water tank, and at the same time start the PIV flow field measurement system to collect particle images in the wake region at a set frequency. Step 3: Real-time processing of PIV images using GPU acceleration module to obtain instantaneous velocity components at each measurement point, which are then used as state inputs to the deep reinforcement learning model; Step 4: The deep reinforcement learning model generates the optimal jet intensity control signal online based on the current state variables and the reward function, driving the jet excitation system to apply blowing and suction jet excitation to the wake region; Step 5: Repeat steps 3 to 4 to form a continuous closed-loop control until the wake suppression effect meets the preset convergence condition.

9. The wake suppression method based on the experimental apparatus according to claim 8, characterized in that: In step 4, the deep reinforcement learning model adopts an online incremental training method, using real-time PIV flow field data as continuous input. In policy updates, the range of action changes in adjacent steps is limited, and the range of policy updates is limited through clip operations to improve control stability and convergence speed.

10. The wake suppression method based on the experimental apparatus according to claim 8, characterized in that: The convergence condition in step 5 is: the change in wake parameters is less than 5% within 500 consecutive control cycles. The wake parameters include average velocity deficit and turbulent kinetic energy. After the experiment, the system automatically generates an analysis report, including wake parameter change curves over time and a comparison chart of control effects.

Images