Underwater vehicle cavitation drag reduction optimization method and system based on AI deep reinforcement learning

Abstract

The application provides an underwater vehicle cavitation drag reduction optimization method and system based on AI deep reinforcement learning, and belongs to the technical field of deep reinforcement learning. The method comprises the following steps: determining a first drag reduction scheme according to a preset path, a shape parameter and a deep reinforcement learning model, correcting the first drag reduction scheme to obtain a second scheme in combination with an actual driving path, arranging an image sensor according to the shape parameter, collecting and analyzing images to obtain real-time cavitation drag reduction parameters, and generating an optimized scheme by fusing multiple parameters and the second scheme. The method realizes dynamic self-adaptive optimization of the drag reduction scheme, and improves the drag reduction efficiency and the navigation stability.

Description

Technical Field

[0001] This invention relates to the field of deep reinforcement learning technology, and in particular to a method and system for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning. Background Technology

[0002] The speed, endurance, and drag characteristics of underwater vehicles directly determine their operational efficiency, and frictional drag at high speeds is a key bottleneck restricting performance improvement. Cavitation, as an efficient drag reduction method, transforms the solid-liquid friction interface into a liquid-gas friction interface by forming cavitation bubbles in low-pressure areas on the surface of the vehicle, thus significantly reducing frictional drag. Therefore, cavitation drag reduction technology has become a core research direction in the high-performance design of underwater vehicles.

[0003] In existing technologies, achieving drag reduction through cavitation in underwater vehicles requires pre-designing cavitation cavitation devices of fixed shapes, setting constant airflow rates, and determining control parameters based on experience or offline simulations, while maintaining the control strategy unchanged during navigation. However, the fixed-parameter strategies in existing technologies lack online optimization capabilities, making it difficult to adapt to the actual operating scenarios of underwater vehicles in practical applications, thus affecting the drag reduction effect.

[0004] Therefore, this invention proposes an optimization method and system for cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning. Summary of the Invention

[0005] This invention provides a method and system for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning, in order to solve the aforementioned technical problems.

[0006] This invention provides a method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning, comprising:

[0007] Step 1: Based on the preset travel path of the target underwater vehicle, the morphological parameters of the target underwater vehicle, and the preset deep reinforcement learning model, determine the first cavitation drag reduction scheme of the target underwater vehicle;

[0008] Step 2: Collect the actual travel path of the target underwater vehicle in real time, and modify the first cavitation drag reduction scheme according to the actual travel path to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

[0009] Step 3: Based on the morphological parameters of the target underwater vehicle, determine multiple image acquisition locations and install an image sensor at each image acquisition location;

[0010] Step 4: Based on the image sensor, acquire real-time operating environment images of the target underwater vehicle in real time, and perform image analysis on the real-time operating environment images to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle;

[0011] Step 5: Based on the real-time cavitation drag reduction parameters at each acquisition time, the trajectory coordinates and navigation attitude parameters of the target underwater vehicle, and in conjunction with the second cavitation drag reduction scheme, determine the cavitation drag reduction optimization scheme for the target underwater vehicle.

[0012] Preferably, step 1 includes:

[0013] Determine the preset navigation parameters for each segment of the preset travel path of the target underwater vehicle, wherein the preset navigation parameters include: travel speed, travel depth, and angle of attack range;

[0014] The preset navigation condition parameters of the target underwater vehicle in each segment and the morphological parameters of the target underwater vehicle are input into a preset deep reinforcement learning model to obtain the sub-drag reduction scheme of the target underwater vehicle in each segment.

[0015] Based on the connection information between each segment and adjacent segments, the sub-drag reduction schemes of each segment are merged to obtain the first cavitation drag reduction scheme of the target underwater vehicle.

[0016] Preferably, step 2 includes:

[0017] Based on the position sensors and navigation module carried by the target underwater vehicle, the trajectory coordinates and navigation direction of the target underwater vehicle at each data acquisition time are determined;

[0018] By integrating the trajectory coordinates and navigation direction at each acquisition moment, the actual travel path of the target underwater vehicle is obtained;

[0019] Based on the actual travel path, the future travel path of the target underwater vehicle is predicted to obtain the predicted travel path.

[0020] Based on the path deviation between the predicted travel path and the preset travel path, the first cavitation drag reduction scheme is modified to obtain the second cavitation drag reduction scheme for the target underwater vehicle.

[0021] Preferably, the first cavitation drag reduction scheme is corrected based on the path deviation between the predicted driving path and the preset driving path, including:

[0022] The predicted driving path is compared with the preset driving path time by time to obtain the path deviation parameters of the target underwater vehicle at each future time within the preset time length. The path deviation parameters include: offset direction, horizontal offset and vertical offset.

[0023] The path deviation parameters are normalized. At the same time, based on the environmental information of the preset travel path and the environmental information of the predicted travel path of the target underwater vehicle at each future time, the environmental deviation parameters for the corresponding future time are determined, and the weight of each path deviation parameter at the corresponding future time is determined.

[0024] The deviation weight of the target underwater vehicle at the corresponding future time is calculated based on the weight of the path deviation parameter corresponding to each future time and the normalized path deviation parameter.

[0025] The preset time length of the first future moment after the actual driving path is divided according to the preset sliding window, and the occurrence position of each future moment in each divided time window is determined. The deviation degree of the corresponding divided time window is determined by combining the deviation weight of the corresponding future moment.

[0026] The average deviation of all time windows divided under the first future moment is taken as the deviation value of the corresponding first future moment. The next future moment is taken as the first future moment for further analysis to obtain the deviation value of each future moment in the preset time length of the first future moment after the actual driving path.

[0027] If the number of future moments whose deviation value is within the first deviation interval after the actual driving path is greater than or equal to the first preset number, then the first cavitation drag reduction scheme of the target underwater vehicle is determined as the second cavitation drag reduction scheme of the target underwater vehicle.

[0028] If the number of future moments whose deviation values ​​fall within the second deviation interval after the first future moment following the actual driving path is greater than or equal to the second preset number, then the deviation value of each future moment in the corresponding preset time length is combined with the first cavitation drag reduction scheme and input into the preset first deviation adjustment model to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

[0029] Preferably, step 3 includes:

[0030] Key structural parameters are extracted from the morphological parameters of the target underwater vehicle, including: surface curvature distribution, cavitation topology, and bow and stern contour gradient.

[0031] Under preset underwater environmental conditions, the spatial influence characteristics of the key structural parameters on the cavitation flow field are simulated to obtain the spatial distribution characteristics of the cavitation sensitive field corresponding to each structural parameter.

[0032] Based on the spatial distribution characteristics of the cavitation sensitive field, and combined with the pressure and velocity distribution data of the cavitation flow field obtained by fluid dynamics simulation, the cavitation sensitive level region on the surface of the target underwater vehicle is divided.

[0033] The quantitative mapping relationship between cavitation sensitivity level and flow field parameters is fitted to determine the image acquisition resolution, acquisition angle and coverage requirements for each cavitation sensitivity level region, and a multi-dimensional acquisition requirement matrix is ​​constructed.

[0034] Based on the multi-dimensional acquisition requirement matrix, the performance parameters of the image sensor, and the key morphological feature points of each cavitation sensitivity level region, a spatial solution model based on surface curvature constraints is adopted to calculate the three-dimensional spatial coordinates of each image acquisition location and install the image sensor.

[0035] Preferably, step 4 includes:

[0036] The image sensor is controlled to acquire real-time images of the target underwater vehicle's operating environment at a preset sampling frequency;

[0037] The real-time operating environment image is subjected to underwater noise filtering, illumination compensation and contrast enhancement to obtain the processed real-time operating environment image;

[0038] The processed real-time operating environment image is input into the image processing model to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle. The real-time cavitation drag reduction parameters include: cavitation profile, cavitation distribution characteristics, and cavitation encapsulation rate of the target underwater vehicle.

[0039] Preferably, step 5 includes:

[0040] Extract the drag reduction path and drag reduction target for each drag reduction direction of the second cavitation drag reduction scheme, and perform a first correlation analysis on the drag reduction paths of any two drag reduction directions and a second correlation analysis on the drag reduction targets of any two drag reduction directions to determine the first priority of each drag reduction direction based on the drag reduction path and the second priority based on the drag reduction target.

[0041] All drag reduction directions are sorted in the first priority order and assigned a first number to each drag reduction direction. At the same time, all drag reduction directions are sorted in the second priority order and assigned a second number to each drag reduction direction.

[0042] Based on the first number and the second number, and according to the drag reduction standard, determine the optimization direction of each drag reduction target, and construct the basic control strategy for the second cavitation drag reduction scheme.

[0043] Extract the spatiotemporal distribution characteristics and coupling correlation characteristics of each parameter in the real-time cavitation drag reduction parameters, trajectory coordinates, and flight attitude parameters at the corresponding acquisition time. At the same time, analyze the flight segment control threshold and cavitation configuration benchmark in the second cavitation drag reduction scheme to construct a full-dimensional coupling feature set for cavitation drag reduction.

[0044] Based on the full-dimensional coupling feature set of cavitation drag reduction, combined with the pressure and velocity distribution parameters of the real-time underwater flow field environment, the quantitative deviation value and state classification level of the current cavitation drag reduction state relative to the optimal state are calculated.

[0045] Based on the state attribution level and quantization deviation value, a quantization mapping relationship between cavitation drag reduction control requirements and execution parameters is constructed. The baseline control parameters of the second cavitation drag reduction scheme, the full-dimensional coupling feature set of cavitation drag reduction, and the quantization deviation value are input into the deep reinforcement learning model. With drag reduction efficiency improvement, navigation attitude stability, and trajectory deviation correction as multi-objective optimization directions, a model reward function with a penalty term is constructed to obtain the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters of the cavitation device.

[0046] By sequentially calculating the cavitation flow field adaptability and removing parameters from the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters that are mismatched with the current underwater flow field environment or exceed the dynamic execution threshold of the vehicle body, an effective control parameter set is obtained.

[0047] The effective set of control parameters is integrated with the basic control strategy of the second cavitation drag reduction scheme to generate an optimized cavitation drag reduction scheme.

[0048] This invention provides an underwater vehicle cavitation drag reduction optimization system based on AI deep reinforcement learning, comprising:

[0049] The scheme determination module is used to determine the first cavitation drag reduction scheme of the target underwater vehicle based on the preset travel path of the target underwater vehicle, the morphological parameters of the target underwater vehicle, and the preset deep reinforcement learning model.

[0050] The scheme correction module is used to collect the actual travel path of the target underwater vehicle in real time, and correct the first cavitation drag reduction scheme according to the actual travel path to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

[0051] The location determination module is used to determine multiple image acquisition locations based on the morphological parameters of the target underwater vehicle, and to install an image sensor at each image acquisition location;

[0052] The image analysis module is used to acquire real-time operating environment images of the target underwater vehicle based on the image sensor, and to perform image analysis on the real-time operating environment images to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle.

[0053] The scheme optimization module is used to determine the cavitation drag reduction optimization scheme for the target underwater vehicle based on the real-time cavitation drag reduction parameters at each acquisition time, the trajectory coordinates and navigation attitude parameters of the target underwater vehicle, and in combination with the second cavitation drag reduction scheme.

[0054] Compared with the prior art, the beneficial effects of this application are as follows:

[0055] An initial drag reduction scheme is generated by combining a deep reinforcement learning model with preset path and morphological parameters, and dynamically corrected according to the actual driving path. Real-time cavitation drag reduction parameters are then obtained through targeted image sensors. An optimized scheme is generated by integrating multi-dimensional navigation parameters and the corrected scheme. This achieves dynamic adaptive optimization of the cavitation drag reduction scheme, effectively solving the problem that traditional fixed drag reduction parameters are difficult to adapt to actual underwater operating scenarios. This significantly improves the cavitation drag reduction efficiency of underwater vehicles while ensuring their navigation stability.

[0056] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0057] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of an underwater vehicle cavitation drag reduction optimization method based on AI deep reinforcement learning in an embodiment of the present invention;

[0060] Figure 2 This is a structural diagram of an underwater vehicle cavitation drag reduction optimization system based on AI deep reinforcement learning, as described in an embodiment of the present invention. Detailed Implementation

[0061] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0062] In this invention, all three-dimensional spatial coordinates (trajectory coordinates, image acquisition position coordinates) are established with the bow apex of the target underwater vehicle as the origin, the hull axis along the forward direction as the positive X-axis, the horizontal direction perpendicular to the X-axis to the right as the positive Y-axis, and the vertical downward direction as the positive Z-axis, establishing a right-handed spatial rectangular coordinate system; the horizontal direction of the trajectory deviation is the XOY plane, and the vertical direction is the Z-axis direction.

[0063] In this invention, the preset underwater environmental conditions for CFD simulation and cavitation-sensitive field analysis are standard marine environmental reference parameters: water temperature 20℃, water pressure 0.6MPa, water flow velocity 0.3m / s, and water density. Dynamic viscosity In practical applications, the data can be proportionally adjusted based on environmental monitoring data of the operating sea area.

[0064] This invention provides an optimization method for cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning, such as... Figure 1 As shown, it includes:

[0065] Step 1: Based on the preset travel path of the target underwater vehicle, the morphological parameters of the target underwater vehicle, and the preset deep reinforcement learning model, determine the first cavitation drag reduction scheme of the target underwater vehicle;

[0066] Step 2: Collect the actual travel path of the target underwater vehicle in real time, and modify the first cavitation drag reduction scheme according to the actual travel path to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

[0067] Step 3: Based on the morphological parameters of the target underwater vehicle, determine multiple image acquisition locations and install an image sensor at each image acquisition location;

[0068] Step 4: Based on the image sensor, acquire real-time operating environment images of the target underwater vehicle in real time, and perform image analysis on the real-time operating environment images to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle;

[0069] Step 5: Based on the real-time cavitation drag reduction parameters at each acquisition time, the trajectory coordinates and navigation attitude parameters of the target underwater vehicle, and in conjunction with the second cavitation drag reduction scheme, determine the cavitation drag reduction optimization scheme for the target underwater vehicle.

[0070] In this embodiment, the preset travel path is a pre-planned navigation trajectory based on the underwater vehicle's operational tasks. It includes segments such as straight lines, turns, and depth changes, and is a three-dimensional spatial trajectory. It is generated by underwater trajectory planning software in combination with nautical chart data of the operational sea area.

[0071] In this embodiment, the morphological parameters are a collective term for the geometric structure of the underwater vehicle and the parameters related to the cavitation configuration, including at least: the length of the vessel, the maximum diameter, the surface curvature distribution of the Gaussian curvature / average curvature at various points on the surface, the cavitation layout topology of the number / position / arrangement of cavitation devices, the bow-stern profile gradient of the diameter change rate from the bow to the stern, and the initial angle / vent size of the cavitation device.

[0072] The beneficial effects of the above technical solution are as follows: by using a deep reinforcement learning model to generate an initial drag reduction scheme in combination with preset path and morphological parameters, and dynamically correcting it according to the actual driving path, and then using targeted image sensors to obtain real-time cavitation drag reduction parameters, and integrating multi-dimensional navigation parameters with the corrected scheme to generate an optimized scheme, dynamic adaptive optimization of the cavitation drag reduction scheme is achieved. This effectively solves the problem that traditional fixed drag reduction parameters are difficult to adapt to actual underwater operating scenarios, greatly improves the cavitation drag reduction efficiency of underwater vehicles, and ensures their navigation stability.

[0073] This invention provides a method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning. Step 1 includes:

[0074] Determine the preset navigation parameters for each segment of the preset travel path of the target underwater vehicle, wherein the preset navigation parameters include: travel speed, travel depth, and angle of attack range;

[0075] The preset navigation condition parameters of the target underwater vehicle in each segment and the morphological parameters of the target underwater vehicle are input into a preset deep reinforcement learning model to obtain the sub-drag reduction scheme of the target underwater vehicle in each segment.

[0076] Based on the connection information between each segment and adjacent segments, the sub-drag reduction schemes of each segment are merged to obtain the first cavitation drag reduction scheme of the target underwater vehicle.

[0077] In this embodiment, the preset navigation operating parameters are the navigation operation parameters of the vehicle under each segment, including navigation speed, navigation depth, and angle of attack range, i.e., the angle between the axis of the vehicle and the direction of water flow. The criterion for dividing the preset travel path into different segments is the abrupt change threshold of the navigation operating parameters. When the change in navigation speed of adjacent path nodes is ≥5kn, the change in navigation depth is ≥10m, and the change in angle of attack range is ≥3°, they are determined to be different segments. The minimum length of the segment is not less than twice the length of the target underwater vehicle to avoid the segment division being too fine, which would lead to frequent switching of sub-drag reduction schemes.

[0078] The sub-drag reduction scheme is a dedicated cavitation drag reduction scheme generated for the operating and morphological parameters of a single flight segment, including parameters such as cavitation airflow, cavitation on / off status, and cavitation angle.

[0079] In this embodiment, the quantitative indicators of flight segment connection information include the flight segment transition position (X, Y, Z), the gradual range of operating parameters, the transition distance L, and the operating condition change amplitude K, wherein the calculation formula for the operating condition change amplitude K is:

[0080] ,in, , The speed of the adjacent flight segment; Maximum sailing speed; The navigation depth for adjacent segments; Maximum sailing depth; , The angle of attack range for adjacent flight segments; The maximum angle of attack range; the value of K ranges from 0 to 1, and the larger the value, the more drastic the change in operating conditions.

[0081] In this embodiment, the deep reinforcement learning model adopts an improved PPO model, which is optimized for cavitation drag reduction scenarios based on the classic PPO model. It is an actor-critic dual-network structure, specifically:

[0082] Input layer: The number of neurons is N, where N = the dimension of flight segment condition parameters + the dimension of morphological parameters; where the dimension of condition parameters is 3 (sailing speed, sailing depth, angle of attack range), and the dimension of morphological parameters is extracted to be 6~12 dimensions according to the actual flight body. The activation function of the input layer is ReLU.

[0083] Hidden layers: There are 3 fully connected layers in total. The first layer has 256 neurons, the second layer has 128 neurons, and the third layer has 64 neurons. The activation function of all hidden layers is ReLU. Dropout layers (dropout rate 0.2) are added between layers to prevent overfitting.

[0084] Actor output layer: The number of neurons is M, where M is the dimension of the cavitation control parameters (including ventilation flow rate, angle, on / off state, etc.). The output layer activation function is Sigmoid, and the output value is mapped to the effective value range of each control parameter.

[0085] Critic output layer: The number of neurons is 1, the output is the state value function, and the activation function is a linear function.

[0086] Training dataset: Composed of CFD simulation data and experimental data on cavitation drag reduction of different underwater vehicles (torpedoes, unmanned underwater vehicles, underwater robots), with a data volume ratio of 7:3; The simulation data covers the full operating range of speed 5~30kn, depth 10~100m, and angle of attack range -5°~+5°. At least 100 sets of corresponding data on morphological parameters, cavitation control parameters and drag reduction efficiency are collected for each operating condition.

[0087] In this embodiment, the connection information between adjacent flight segments is quantified based on the analytic hierarchy process (AHP). According to the change in operating conditions during the transition of flight segments, the merging rules of sub-drag reduction schemes are determined, including the gradual adjustment step size of the control parameters and the transition distance.

[0088] Using a high-speed underwater unmanned vehicle as the target, its morphological parameters are: length 12m, maximum diameter 1.2m, and maximum Gaussian curvature of the bow surface. The system consists of one forward-facing main cavitation unit and two side-mounted auxiliary cavitation units arranged in an equilateral triangle. Its preset route is from sea area A (latitude and longitude 118°E, 30°N, depth 50m) to sea area B (latitude and longitude 118.01°E, 30°N, depth 80m), broken down into a straight segment 1 and a turning segment 2 based on operating conditions.

[0089] Segment 1 (straight line): speed 20 knots, depth 50 meters, angle of attack range 0° ± 2°;

[0090] Segment 2 (turn): speed 15 knots, depth 60-80 m, angle of attack range -3° to +3°.

[0091] By inputting the operating parameters of the two flight segments and the underwater vehicle's morphological parameters into the improved PPO model, a sub-drag reduction scheme is obtained:

[0092] Segment 1 is the forward cavitation airflow rate. Side-mounted cavitation unit shut off;

[0093] Segment 2 is the forward cavitation airflow rate. Side-mounted cavitation unit airflow .

[0094] Based on the connection information between segment 1 and segment 2 (turn start point coordinates (118.005°E, 30°N), operating condition transition distance 50m), the two sub-drag reduction schemes are merged, and the first cavitation drag reduction scheme is determined as follows: Segment 1 is executed according to the sub-scheme; after reaching the turn start point, the forward cavitation airflow rate is 0.004... The gradient gradually decreases to The side-mounted cavitation unit has an airflow rate of 0.004 The gradient gradually increases to The transition distance is 50m, and segment 2 will be executed according to the sub-scheme.

[0095] In this embodiment, the sub-drag reduction schemes for each flight segment are merged, including:

[0096] When the transition distance L≥50m and the operating condition change range K≤0.2, the sub-drag reduction schemes are merged in a linear gradual manner, and the gradual change step size of the control parameters =|parameter1-parameter2| / L;

[0097] When 30m≤L<50m and 0.2<K≤0.5, the sub-drag reduction schemes are merged using an exponential gradual change method, and the gradual change step size of the control parameters increases exponentially with the increase of the transition distance.

[0098] When L < 30m and K > 0.5, the sub-drag reduction schemes are merged using a step switching method, and the sub-drag reduction scheme for the next segment is directly switched at the segment transition position.

[0099] The beneficial effects of the above technical solution are as follows: by segmenting the preset driving path according to the differences in operating conditions, extracting the operating condition parameters of each segment and combining them with the morphological parameters of the aircraft body, inputting them into a deep reinforcement learning model to generate sub-drag reduction schemes, and then merging them into the first airflow drag reduction scheme based on the segment connection information, the initial drag reduction scheme can fit the differences in operating conditions of different segments, avoiding the one-size-fits-all problem of fixed parameter strategies for different flight conditions in the existing technology, improving the operating condition adaptability of the initial scheme, and laying a precise foundation for subsequent scheme correction and optimization.

[0100] This invention provides a method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning. Step 2 includes:

[0101] Based on the position sensors and navigation module carried by the target underwater vehicle, the trajectory coordinates and navigation direction of the target underwater vehicle at each data acquisition time are determined;

[0102] By integrating the trajectory coordinates and navigation direction at each acquisition moment, the actual travel path of the target underwater vehicle is obtained;

[0103] Based on the actual travel path, the future travel path of the target underwater vehicle is predicted to obtain the predicted travel path.

[0104] Based on the path deviation between the predicted travel path and the preset travel path, the first cavitation drag reduction scheme is modified to obtain the second cavitation drag reduction scheme for the target underwater vehicle.

[0105] Preferably, the first cavitation drag reduction scheme is corrected based on the path deviation between the predicted driving path and the preset driving path, including:

[0106] The predicted driving path is compared with the preset driving path time by time to obtain the path deviation parameters of the target underwater vehicle at each future time within the preset time length. The path deviation parameters include: offset direction, horizontal offset and vertical offset.

[0107] The path deviation parameters are normalized. At the same time, based on the environmental information of the preset travel path and the environmental information of the predicted travel path of the target underwater vehicle at each future time, the environmental deviation parameters for the corresponding future time are determined, and the weight of each path deviation parameter at the corresponding future time is determined.

[0108] The deviation weight of the target underwater vehicle at the corresponding future time is calculated based on the weight of the path deviation parameter corresponding to each future time and the normalized path deviation parameter.

[0109] The preset time length of the first future moment after the actual driving path is divided according to the preset sliding window, and the occurrence position of each future moment in each divided time window is determined. The deviation degree of the corresponding divided time window is determined by combining the deviation weight of the corresponding future moment.

[0110] The average deviation of all time windows divided under the first future moment is taken as the deviation value of the corresponding first future moment. The next future moment is taken as the first future moment for further analysis to obtain the deviation value of each future moment in the preset time length of the first future moment after the actual driving path.

[0111] If the number of future moments whose deviation value is within the first deviation interval after the actual driving path is greater than or equal to the first preset number, then the first cavitation drag reduction scheme of the target underwater vehicle is determined as the second cavitation drag reduction scheme of the target underwater vehicle.

[0112] If the number of future moments whose deviation values ​​fall within the second deviation interval after the first future moment following the actual driving path is greater than or equal to the second preset number, then the deviation value of each future moment in the corresponding preset time length is combined with the first cavitation drag reduction scheme and input into the preset first deviation adjustment model to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

[0113] In this embodiment, the position sensor is a sensor used to collect the real-time position of the vehicle, including underwater GPS and inertial navigation sensors;

[0114] The navigation module is a hardware and software integrated module that provides positioning and navigation for the vehicle. This implementation adopts the Beidou underwater navigation system.

[0115] The trajectory coordinates are the real-time position coordinates (X, Y, Z) of the vehicle in three-dimensional space, and a spatial rectangular coordinate system is established with the starting point of the preset travel path as the origin;

[0116] The predicted driving path is based on historical actual driving path data. It is the trajectory of the vehicle over a period of time in the future obtained by the prediction model. The model is trained on the neural network model with historical actual driving path data as the training set until the trajectory coordinate prediction error is ≤0.5m before it is put into use. During prediction, the latest trajectory data at the T times is input in real time, and the predicted trajectory at the next T1 times is output. The training sample exceeds 1000 cases.

[0117] Path deviation parameters are the differences between the predicted driving path and the preset driving path, including the offset direction (horizontal / vertical offset direction), horizontal offset (trajectory deviation distance in the horizontal plane), and vertical offset (trajectory deviation distance in the depth direction).

[0118] Environmental deviation parameters are the differences in the underwater environment between the preset driving path and the predicted driving path at the corresponding locations, including water flow velocity, water flow direction, water pressure, water temperature, etc.

[0119] Deviation weights are importance coefficients for each path deviation parameter, determined by the degree of influence of environmental deviation parameters on path deviations.

[0120] The sliding window is a time window used to segment and analyze the trajectory over a future preset time period. It includes the window length and step size. The dynamic response time of the cavitation flow field of the underwater vehicle is about 1 to 3 seconds. The window length of 5 seconds can cover a complete dynamic change cycle of the cavitation flow field, ensuring the comprehensiveness of the deviation analysis. The step size of 1 second can ensure the real-time nature of the scheme correction, adapt to the high-speed navigation characteristics of the underwater vehicle (speed 5 to 30 knots, travel distance of about 2.5 to 15 meters in 1 second), and avoid correction lag due to excessive step size.

[0121] In this embodiment, the industry-standard quantitative grading based on the fuzzy comprehensive evaluation method is determined by CFD simulation based on the correlation between trajectory deviation and drag reduction efficiency. The first deviation interval is [0, 0.2], the second deviation interval is (0.2, 0.5], and the third deviation interval is (0.5, 1]. These are all pre-stored and can be used directly.

[0122] In this embodiment, the statistical significance test is used to determine the value of the preset quantity as follows: the first preset quantity = 0.8 × T1, the second preset quantity = 0.5 × T1; when the number of future moments in the third deviation interval is ≥ 0.3 × T1, the correction of the first scheme is abandoned, and a new drag reduction scheme is generated by calling the improved PPO model based on the predicted driving path.

[0123] In this embodiment, the first deviation adjustment model is used to implement the BP neural network model for drag reduction scheme correction, and its network structure includes:

[0124] Input layer: The number of neurons is T1+M, where T1 is the number of future time steps and M is the dimension of the control parameters of the first cavitation drag reduction scheme; the input consists of the deviation values ​​at the next T1 time steps and the M control parameters of the first scheme.

[0125] Hidden layers: 2 fully connected layers, the first layer has 128 neurons and the second layer has 64 neurons, the activation function is ReLU;

[0126] Output layer: The number of neurons is M, and the output consists of M control parameters of the modified second cavitation drag reduction scheme. The activation function is Sigmoid.

[0127] Its training set consists of simulation data of path deviation value and drag reduction scheme correction amount, covering the entire range of deviation value [0, 0.5].

[0128] At this point, the inputs are the deviation value and the first cavitation drag reduction scheme, and the output is the corrected drag reduction scheme.

[0129] In this embodiment, the three-dimensional coordinates of the predicted driving path and the preset driving path are compared at each time step using 0.1s as the time unit, the path deviation parameter is extracted, and the offset is mapped to the [0,1] interval using min-max normalization to eliminate the influence of dimensions.

[0130] Environmental data for preset / predicted paths is collected using underwater environmental monitoring sensors. Environmental deviation parameters are calculated, and the analytic hierarchy process (AHP) is used to determine the deviation weights of each path deviation parameter based on the degree of influence of these parameters. Specifically, this includes:

[0131] Using offset direction, horizontal offset, and vertical offset as three evaluation indicators, a judgment matrix with a scale of 1-9 is established based on the influence of environmental deviation parameters on each indicator. ,in This indicates the importance of indicator i relative to indicator j. , ;

[0132] Among the environmental deviation parameters, the water flow direction deviation mainly affects the offset direction, the water flow velocity deviation mainly affects the horizontal offset, and the water pressure deviation mainly affects the vertical offset. Based on this, the basic judgment matrix is ​​determined as follows: In practical applications, the judgment matrix can be corrected based on the specific values ​​of environmental deviations.

[0133] Calculate the eigenvalues ​​and eigenvectors of the judgment matrix A, and normalize the eigenvectors to obtain the initial weights w1, w2, and w3 for each indicator;

[0134] Calculate the consistency index ,in, To determine the largest eigenvalue of the matrix, n=3;

[0135] Find the average random consistency index RI; when n=3, RI=0.58; calculate the consistency ratio. ,when When ≤0.1, the judgment matrix satisfies the consistency requirement, and the initial weights are valid; if Adjust the elements of the judgment matrix until ≤0.1.

[0136] Set the sliding window parameters (window length 5s, step size 1s), and calculate the deviation degree of each window according to the deviation degree calculation formula of this invention. It should be noted that, after multiple simulation verifications, a sensitivity adjustment coefficient of 0.8 can balance the sensitivity and stability of deviation analysis to outliers, avoiding over-response to small deviations while effectively identifying significant deviations. To avoid mathematical singularities with a denominator of 0 in the formula and to ensure the accuracy of the calculation results, a small positive value smoothing constant commonly used in the industry is selected with a value of 0.001. The historical benchmark value is 0.2, which is the average deviation weight obtained based on more than 100 historical cavitation drag reduction simulations.

[0137] Based on the degree of deviation, the deviation value for each future moment is obtained, and a first deviation range [0, 0.2] (small deviation) and a first preset quantity of 24 (80% of the next 30 seconds) are set, and a second deviation range [0.2, 0.5] (medium deviation) and a second preset quantity of 15 (50% of the next 30 seconds) are set.

[0138] If the number of small deviation moments is greater than or equal to the first preset number, the first scheme is the second scheme; if the number of medium deviation moments is greater than or equal to the second preset number, the deviation value and the first scheme are input into the trained BP neural network model to obtain the corrected second cavitation drag reduction scheme; in this embodiment, the BP neural network model uses the deviation value as the input layer, the parameters of the first scheme as the hidden layer, and the parameters of the corrected scheme as the output layer, and the training set is the simulation data of path deviation-scheme correction.

[0139] For example, during the navigation of a high-speed underwater unmanned vehicle, the actual trajectory coordinates for the first 10 seconds are (10,2,50), (20,3,50)...(100,11,50), while the preset trajectory coordinates are (10,0,50), (20,0,50)...(100,0,50), showing a horizontal rightward veer. A prediction model is used to obtain the predicted travel path for the next 30 seconds. Path deviation parameters are extracted at 30 time points: horizontal rightward veer, horizontal offset 0.5~2.0m, and vertical offset 0~0.3m. After min-max normalization, these parameters are mapped to the interval [0.05,0.4]. The water flow velocity at the preset path is 0.5m / s, and the direction is due east. The predicted water flow velocity at the path is 0.8 m / s, heading southeast. Environmental deviation parameters are a water flow velocity deviation of 0.3 m / s and a water flow direction deviation of 45°. Using AHP (Adaptive High Voltage) analysis, the horizontal offset weight is determined to be 0.6, the vertical offset weight to be 0.3, and the offset direction weight to be 0.1. The deviation weights at each time point are calculated. After calculating the deviation degree using a 5-second sliding window, deviation values ​​at 30 time points are obtained. The deviation values ​​at 20 time points fall within the second deviation interval (0.2, 0.5], reaching the second preset quantity of 15. Therefore, the deviation values ​​at these 30 time points and the first cavitation drag reduction scheme are input into the BP neural network model to obtain the second cavitation drag reduction scheme: the forward cavitation venting flow rate for segment 1 is adjusted to 0.75. The transition distance at the segment connection point has been shortened from 50m to 30m, and the forward cavitation flow gradient has been adjusted to 0.005. The flow gradient of the side-mounted cavitation unit was adjusted to 0.0067. .

[0140] In this embodiment, the formula for calculating the degree of deviation is: ;

[0141] in, For the first The degree of deviation in the division of time windows, For the first One time window is used for segmentation. 0 is located at the th The first time window within the segmented time window 0 future moments, For the first Deviation weights for 0 future moments, For the first The start time of a time window. To preset the window length of the sliding window, This is a historical benchmark value. This is the sensitivity adjustment coefficient, used to control the formula's sensitivity to abnormal fluctuations. As a smoothing constant, it should be noted that its position is used to determine the difference within the window at that future moment.

[0142] The beneficial effects of the above technical solution are: by collecting the actual driving path of the underwater vehicle in real time and integrating multiple models to predict the future driving path, and by quantitatively analyzing the path deviation and environmental deviation, the first cavitation drag reduction scheme can be accurately corrected, so that the drag reduction scheme can be dynamically adjusted according to the actual path change of the underwater vehicle. This solves the problem that fixed parameters cannot cope with path deviation in the existing technology, improves the adaptability of the scheme to the actual driving state of the underwater vehicle, and avoids the decrease in cavitation drag reduction efficiency caused by path deviation.

[0143] This invention provides a method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning. Step 3 includes:

[0144] Key structural parameters are extracted from the morphological parameters of the target underwater vehicle, including: surface curvature distribution, cavitation topology, and bow and stern contour gradient.

[0145] Under preset underwater environmental conditions, the spatial influence characteristics of the key structural parameters on the cavitation flow field are simulated to obtain the spatial distribution characteristics of the cavitation sensitive field corresponding to each structural parameter.

[0146] Based on the spatial distribution characteristics of the cavitation sensitive field, and combined with the pressure and velocity distribution data of the cavitation flow field obtained by fluid dynamics simulation, the cavitation sensitive level region on the surface of the target underwater vehicle is divided.

[0147] The quantitative mapping relationship between cavitation sensitivity level and flow field parameters is fitted to determine the image acquisition resolution, acquisition angle and coverage requirements for each cavitation sensitivity level region, and a multi-dimensional acquisition requirement matrix is ​​constructed.

[0148] Based on the multi-dimensional acquisition requirement matrix, the performance parameters of the image sensor, and the key morphological feature points of each cavitation sensitivity level region, a spatial solution model based on surface curvature constraints is adopted to calculate the three-dimensional spatial coordinates of each image acquisition location and install the image sensor.

[0149] In this embodiment, curvature analysis is performed on the 3D model of the underwater vehicle using 3D modeling software (SolidWorks / UG). The Gaussian curvature and average curvature of discrete points on the model surface are extracted. A sampling point is taken every 0.1m along the X-axis to form a discrete dataset of surface curvature distribution. The center coordinates, number, angle with the hull, and snorkel diameter of the cavitation devices are extracted from the 3D model and organized into a topology parameter set according to the classification of main cavitation devices + auxiliary cavitation devices. The hull diameter of each sampling point along the X-axis is extracted, and the rate of change of diameter between adjacent sampling points is calculated using the following formula: ,in, The diameter of adjacent sampling points. The X-axis coordinates of adjacent sampling points.

[0150] In this embodiment, the spatial distribution characteristics of the cavitation sensitive field are the distribution characteristics of the sensitivity of different locations on the surface of the vehicle to cavitation phenomena. The higher the sensitivity, the easier it is for cavitation phenomena to occur and the more significant the changes.

[0151] The cavitation flow field pressure / velocity distribution data are the pressure and water velocity distribution parameters of the cavitation flow field around the vehicle, obtained by computational fluid dynamics simulation.

[0152] Cavitation sensitivity level zones are areas on the surface of a vehicle that are divided into three sensitivity levels: high, medium, and low, based on the degree of cavitation sensitivity.

[0153] The multi-dimensional acquisition requirement matrix is ​​arranged with cavitation sensitivity level as the row and image acquisition resolution, acquisition angle, and coverage as the column, forming an acquisition parameter matrix. The values ​​are the acquisition parameters corresponding to each sensitivity level.

[0154] The spatial solution model of surface curvature constraint uses the surface curvature of the vehicle body as a constraint condition to solve the three-dimensional spatial coordinates of the image acquisition position.

[0155] In this embodiment, CFD simulation software (Fluent / Star-CCm+) is used to simulate the cavitation flow field. The simulation settings are as follows:

[0156] Computational Domain and Mesh Generation: A cuboid computational domain is established centered on the underwater vehicle, with dimensions of (10L×5L×5L), where L is the length of the vehicle. The surface of the vehicle is meshed with a structured grid with a mesh size of 0.01m, while the far field of the computational domain is meshed with an unstructured grid with a mesh size of 0.5m and a mesh quality ≥0.8.

[0157] Physical model: The Schnerr-Sauer cavitation model is adopted, which is suitable for cavitation simulation of underwater vehicles. The turbulence model is... The model considers the effects of gravity and buoyancy, and the water body is an incompressible fluid.

[0158] Boundary conditions: The inlet is a velocity inlet (set according to the preset sailing speed), the outlet is a pressure outlet (set according to the water pressure at the preset sailing depth), the surface of the vehicle body is a non-slip wall, and the cavitation vent is a mass flow inlet;

[0159] Solution setup: A pressure-based solver is used, with SIMPLEC as the coupling method. The discrete schemes for the momentum equation, energy equation, and cavitation volume fraction are all second-order upwind schemes. The convergence residual is ≤1e-6, and the number of iterations is ≥1000.

[0160] In this embodiment, the K-means clustering algorithm is used to divide the cavitation sensitivity level regions, specifically including:

[0161] Cavitation sensitive field intensity, flow field pressure, and water flow velocity were selected as three clustering features. The three feature values ​​of discrete sampling points on the surface of the vehicle were normalized to form a clustered sample set.

[0162] The elbow rule is used to determine the number of clusters k, and the sum of squared clustering errors (SSE) is calculated when k=2~5. When the rate of decrease of SSE slows down significantly, k=3 is determined (high, medium and low cavitation sensitivity levels).

[0163] The K-means++ algorithm is used for clustering. The initial cluster centers are selected by random sampling, and the iteration stops when the cluster centers no longer change. The sampling points of the same cluster are connected to form a continuous region, which is the region corresponding to the cavitation sensitivity level.

[0164] Morphological smoothing is applied to the clustered regions to eliminate isolated small regions, making the regions of each sensitivity level continuous and non-overlapping.

[0165] In this embodiment, the spatial solution model uses the surface curvature of the vehicle surface as a constraint, combined with the performance parameters of the image sensor, to calculate the three-dimensional coordinates (x, y, z) of the image acquisition position. The core formula is:

[0166] ,in, To calculate the Gaussian curvature at the position; To correspond to the curvature threshold of the cavitation sensitivity level, the high-sensitivity region medium sensitive area Low-sensitivity area ; The coordinates of key morphological feature points in the cavitation-sensitive region; This refers to the sensor's acquisition perspective; The maximum acquisition angle of the sensor is defined as 120°, 90°, and 60°, respectively, for the high-sensitivity, medium-sensitivity, and low-sensitivity areas. The distance from the sensor to the feature point; The effective imaging distance of the sensor is [0.3m, 5m] in this invention; The coverage area of ​​the sensor; This represents the minimum coverage area corresponding to the sensitivity level, with high-sensitivity areas, medium-sensitivity areas, and low-sensitivity areas being respectively... .

[0167] By solving the above set of constraint equations, the three-dimensional coordinates (x, y, z) that satisfy all conditions are obtained, which is the image acquisition position;

[0168] If multiple solutions exist, the location with the largest coverage area and the most uniform curvature is selected as the final acquisition location.

[0169] In this embodiment, the general construction rules for the multi-dimensional collection requirement matrix are as follows:

[0170] The behavior cavitation sensitivity level (high / medium / low) of the multi-dimensional data acquisition requirement matrix is ​​represented by columns for acquisition resolution, acquisition viewpoint, coverage area, and sampling frequency. The value rules for each column are as follows:

[0171] Acquisition resolution: High sensitivity area ≥3840×2160 (4K), medium sensitivity area ≥2560×1440 (2K), low sensitivity area ≥1920×1080 (1080P);

[0172] Acquisition angle: High sensitivity area ≥120°, medium sensitivity area ≥90°, low sensitivity area ≥60°;

[0173] Coverage area: High-sensitivity area ≥ 0.5m × 0.5m, medium-sensitivity area ≥ 1m × 1m, low-sensitivity area ≥ 2m × 2m;

[0174] Sampling frequency: High sensitivity area ≥20fps, medium sensitivity area ≥15fps, low sensitivity area ≥10fps;

[0175] The acquisition parameters can be adjusted upwards according to the actual performance of the image sensor to ensure real-time acquisition and recognition of cavitation state.

[0176] In this embodiment, such as a high-speed underwater unmanned vehicle, key structural parameters are extracted using SolidWorks: the surface curvature distribution is a Gaussian curvature of 5-8 at the bow. 2-5 cm of the middle section of the hull Stern <2 The cavitation cavitation topology consists of one forward-facing main cavitation cavitation and two side-mounted auxiliary cavitation cavitation cavitations arranged in an equilateral triangle. The bow and stern profile gradient is as follows: the diameter of the bow section (0-2m) increases from 0 to 1.2m, with a gradient of 0.6m / m. CFD simulation using Fluent yielded the spatial distribution characteristics of the cavitation-sensitive field: around the bow cavitation ... The cavitation sensitivity field intensity was highest in the bow region, followed by the midsection, and lowest in the stern. Cavitation flow field parameters were also obtained: pressure 0.05 MPa and current velocity 25 m / s in the bow low-pressure zone; pressure 0.1–0.2 MPa and current velocity 15–20 m / s in the midsection; and pressure >0.2 MPa and current velocity <15 m / s in the stern. K-means clustering was used to divide the submersible surface into three sensitivity levels, and a quantitative mapping relationship was obtained: high sensitivity zone corresponds to flow field pressure <0.1 MPa and current velocity >20 m / s; medium sensitivity zone corresponds to 0.1–0.2 MPa and 15–20 m / s; and low sensitivity zone corresponds to >0.2 MPa and <15 m / s.

[0177] In this embodiment, a multi-dimensional data collection requirement matrix is ​​constructed based on this relationship, as shown in Table 1:

[0178] Table 1 Multi-dimensional data collection requirements matrix

[0179]

[0180] It should be noted that a spatial rectangular coordinate system was established with the bow apex as the origin. Combining the image sensor performance parameters (lens focal length 8mm, imaging distance 0.3~5m, waterproof rating IP68) and key morphological feature points of sensitive areas, the three-dimensional spatial coordinates of six image acquisition positions were calculated using a spatial solution model constrained by surface curvature: three high-sensitivity areas (bow): (0.2,0,0), (0.2,0.2,0), (0.2,-0.2,0); two medium-sensitivity areas (midships): (4,0,0), (5,0,0); and one low-sensitivity area (stern): (10,0,0). Underwater high-definition image sensors were installed at the above coordinate positions, and the sampling frequency was set according to the sensitivity level (high 20fps, medium 15fps, low 10fps).

[0181] The beneficial effects of the above technical solution are as follows: by extracting key structural parameters that significantly affect the cavitation flow field of the vehicle, combining CFD simulation analysis of the cavitation sensitive field distribution and dividing the sensitive level region, and then obtaining the precise image acquisition position through spatial calculation, the image sensor can selectively acquire image data of the cavitation sensitive area, avoiding the problem of inaccurate and incomplete cavitation state data caused by unreasonable monitoring position, providing a precise and effective data source for subsequent real-time cavitation parameter extraction, and solving the problem of no effective cavitation state monitoring means in the existing technology.

[0182] This invention provides a method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning, step 4 of which includes:

[0183] The image sensor is controlled to acquire real-time images of the target underwater vehicle's operating environment at a preset sampling frequency;

[0184] The real-time operating environment image is subjected to underwater noise filtering, illumination compensation and contrast enhancement to obtain the processed real-time operating environment image;

[0185] The processed real-time operating environment image is input into the image processing model to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle. The real-time cavitation drag reduction parameters include: cavitation profile, cavitation distribution characteristics, and cavitation encapsulation rate of the target underwater vehicle.

[0186] In this embodiment, the preset sampling frequency is the image acquisition frequency set by the image sensor for different sensitivity levels based on the cavitation flow field change rate. The higher the sensitivity level, the higher the sampling frequency.

[0187] In this embodiment, a combination of median filtering and wavelet transform (db4 wavelet, decomposed into 3 layers) is used for underwater noise filtering, histogram equalization and adaptive brightness adjustment are used for illumination compensation, and gamma correction (gamma value 1.2-1.5) is used for contrast enhancement to obtain the processed real-time operating environment image.

[0188] An image processing model fusion of YOLOv8 and maskR-CNN was used to analyze the processed images. YOLOv8 enabled fast target detection in cavitation regions, while maskR-CNN enabled accurate semantic segmentation of these regions. The training set consisted of underwater cavitation flow field image datasets, with precise annotation of cavitation regions. The LabelMe annotation tool was used to annotate the underwater cavitation flow field images to obtain samples. The annotations included: bounding boxes (bboxes) of cavitation regions, cavitation masks, and cavitation categories (primary cavitation / secondary cavitation). The number of annotated images was ≥10,000, covering cavitation states with different cavitation encapsulation rates and different flow field environments. The neural network model was trained based on the obtained samples to obtain the YOLOv8+maskR-CNN fusion image processing model.

[0189] The image processing model outputs real-time cavitation drag reduction parameters, where the cavitation encapsulation rate = the coverage area of ​​the cavitation region in the image acquired by each sensor / the area of ​​the monitored area on the surface of the ship × 100%.

[0190] For example, the six image sensors of a high-speed underwater unmanned vehicle acquire images at a preset sampling frequency. Taking the sensor in the high-sensitivity area at the bow (4K, 120°, 20fps) as an example, the original images acquired have problems such as salt-and-pepper noise, dim lighting, and low contrast between cavitation and water. The original images are subjected to 3×3 window median filtering + db4 wavelet decomposition 3-layer noise filtering to remove noise caused by suspended particles. Histogram equalization is used to increase the average brightness of the image from 80 to 120 to achieve lighting compensation. The gamma value is set to 1.3 for gamma correction to improve the contrast between cavitation and water, resulting in a clear image after processing. The processed image is input into the YOLOv8+maskR-CNN fusion model. The model quickly identifies and segments the cavitation region, outputting real-time cavitation drag reduction parameters: the cavitation outline is an irregular circle with edge pixel coordinates of (100,100), (120,90), (130,110), and (110,130); the cavitation distribution characteristics are: there is one main cavitation around the main cavitation at the bow, with a diameter of approximately 0.2 m, no secondary cavitation, and no cavitation offset; the cavitation coverage rate is 35%. The parameters from the six sensors are integrated to obtain the overall real-time cavitation drag reduction parameters of the submersible.

[0191] The beneficial effects of the above technical solution are as follows: by performing multi-stage preprocessing on the acquired underwater images, including noise filtering, illumination compensation, and contrast enhancement, image interference caused by the complex underwater environment is removed. Combined with a deep learning model that integrates YOLOv8 and maskR-CNN, real-time and accurate extraction of cavitation drag reduction parameters is achieved. This enables dynamic monitoring of the actual cavitation state of the vehicle, solving the problems of existing technologies that cannot obtain cavitation state in real time and are difficult to adjust drag reduction strategies according to actual cavitation conditions. This provides a real-time and accurate data analysis foundation for the online optimization of subsequent drag reduction schemes.

[0192] This invention provides a method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning, step 5 of which includes:

[0193] Extract the drag reduction path and drag reduction target for each drag reduction direction of the second cavitation drag reduction scheme, and perform a first correlation analysis on the drag reduction paths of any two drag reduction directions and a second correlation analysis on the drag reduction targets of any two drag reduction directions to determine the first priority of each drag reduction direction based on the drag reduction path and the second priority based on the drag reduction target.

[0194] All drag reduction directions are sorted in the first priority order and assigned a first number to each drag reduction direction. At the same time, all drag reduction directions are sorted in the second priority order and assigned a second number to each drag reduction direction.

[0195] Based on the first number and the second number, and according to the drag reduction standard, determine the optimization direction of each drag reduction target, and construct the basic control strategy for the second cavitation drag reduction scheme.

[0196] Extract the spatiotemporal distribution characteristics and coupling correlation characteristics of each parameter in the real-time cavitation drag reduction parameters, trajectory coordinates, and flight attitude parameters at the corresponding acquisition time. At the same time, analyze the flight segment control threshold and cavitation configuration benchmark in the second cavitation drag reduction scheme to construct a full-dimensional coupling feature set for cavitation drag reduction.

[0197] Based on the full-dimensional coupling feature set of cavitation drag reduction, combined with the pressure and velocity distribution parameters of the real-time underwater flow field environment, the quantitative deviation value and state classification level of the current cavitation drag reduction state relative to the optimal state are calculated.

[0198] Based on the state attribution level and quantization deviation value, a quantization mapping relationship between cavitation drag reduction control requirements and execution parameters is constructed. The baseline control parameters of the second cavitation drag reduction scheme, the full-dimensional coupling feature set of cavitation drag reduction, and the quantization deviation value are input into the deep reinforcement learning model. With drag reduction efficiency improvement, navigation attitude stability, and trajectory deviation correction as multi-objective optimization directions, a model reward function with a penalty term is constructed to obtain the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters of the cavitation device.

[0199] By sequentially calculating the cavitation flow field adaptability and removing parameters from the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters that are mismatched with the current underwater flow field environment or exceed the dynamic execution threshold of the vehicle body, an effective control parameter set is obtained.

[0200] The effective set of control parameters is integrated with the basic control strategy of the second cavitation drag reduction scheme to generate an optimized cavitation drag reduction scheme.

[0201] In this embodiment, the drag reduction direction is the control direction for achieving cavitation drag reduction, including cavitation airflow adjustment, cavitation angle adjustment, and fine adjustment of the aircraft's angle of attack.

[0202] The drag reduction path is the operation steps corresponding to the drag reduction direction, and the drag reduction target is the expected effect of the operation corresponding to the drag reduction direction. For example, the path of ventilation flow adjustment is gradual adjustment, and the target is to control the cavitation encapsulation rate within the optimal range.

[0203] The first and second correlation analyses are correlation analyses of the drag reduction paths and drag reduction targets in each drag reduction direction, using Pearson correlation coefficient and Spearman correlation coefficient respectively.

[0204] In this embodiment, the criteria for determining the first priority (drag reduction path) are as follows: the smaller the Pearson correlation coefficient |r|, the higher the priority; when |r|≤0.2, it is the highest priority; when 0.2<|r|≤0.5, it is the medium priority; when |r|>0.5, it is the low priority.

[0205] The criteria for determining the second priority (drag reduction target) are: the smaller the Spearman rank correlation coefficient |p|, the higher the priority; the determination threshold is the same as that for the first priority.

[0206] Overall priority of drag reduction direction: When the first priority and the second priority are the same, they are directly sorted according to that priority; when they are different, the first priority (drag reduction path) shall prevail, because the correlation of drag reduction path directly affects the efficiency of control execution.

[0207] The basic control strategy is the basic control direction and operating rules determined based on the priority numbering of drag reduction direction and cavitation drag reduction standard;

[0208] The cavitation drag reduction full-dimensional coupling feature set is a feature set that integrates the spatiotemporal distribution characteristics and coupling correlation characteristics of real-time cavitation drag reduction parameters, trajectory coordinates, and flight attitude parameters, combined with the segment control threshold and cavitation configuration benchmark of the second scheme.

[0209] In this embodiment, the quantization deviation value is calculated as follows:

[0210] Construct an evaluation factor set U: U={u1,u2,u3,u4,u5}, where u1 is the cavitation encapsulation rate, u2 is the uniformity of cavitation distribution, u3 is the horizontal offset, u4 is the vertical offset, and u5 is the angle of attack.

[0211] Construct a comment set V: V={v1,v2,v3,v4,v5}, corresponding to the degree of deviation as {minimal, small, medium, large, extremely large}, quantized as {0,0.25,0.5,0.75,1}, and determined using the equal interval quantization method.

[0212] Weight set W: The weights of each evaluation factor are determined using the AHP method, W={0.4,0.2,0.15,0.1,0.15}. The void encapsulation rate has the highest weight because it directly affects drag reduction efficiency. This weight is determined using the analytic hierarchy process.

[0213] In this embodiment, the state affiliation level is calculated as follows:

[0214] The membership degree of each evaluation factor to the comment set is determined by using a trapezoidal membership function, and a fuzzy evaluation matrix is ​​constructed. ,in, Indicator Factors Comments Membership degree;

[0215] Calculate the fuzzy comprehensive evaluation results The weighted average method is used to defuzzify B and obtain the quantization deviation value D1, where D1∈[0,1].

[0216] Status classification level: Based on the value of D1, Level 1 (Excellent): D1∈[0,0.2], Level 2 (Medium): D1∈(0.2,0.5], Level 3 (Poor): D1∈(0.5,1).

[0217] In this embodiment, the multi-objective reward function J with a penalty term designed using the improved PPO model is a weighted sum of drag reduction efficiency reward, flight attitude reward, trajectory deviation reward, and penalty term. ,in, , , , Rewards for drag reduction efficiency Flight attitude reward Trajectory Deviation Reward Penalty items The weighting coefficients, and Adjust proportionally according to operational requirements (such as high-speed navigation / precise trajectory).

[0218] ,in, The optimal value for cavitation encapsulation rate is 35%. The optimal interval deviation is 5%. This represents the maximum cavitation encapsulation rate (100%).

[0219] ,in, The optimal angle of attack is 0°. The maximum threshold for angle of attack (5°);

[0220] ,in, , These are the horizontal and vertical offsets; This is the optimal value for the offset; , The maximum offset threshold;

[0221] ;

[0222] It is worth noting that the reward function J ranges from -0.1 to 1. A larger R value indicates a better current state. The model achieves multi-objective optimization by maximizing the cumulative reward.

[0223] Multi-dimensional dynamic control parameters and adaptive fine-tuning parameters are control parameters output by the deep reinforcement learning model, including core parameters such as cavitation flow rate and angle, as well as fine-tuning parameters such as adjustment step size and adjustment time.

[0224] Cavitation flow field adaptability calculation and dynamic execution threshold are quickly calculated using CFD to determine the matching of control parameters with the current flow field and the maximum execution capability of the vehicle's power system, which serve as the selection criteria for control parameters;

[0225] The effective control parameter set is the set of executable control parameters obtained after removing parameters that do not match the flow field or exceed the dynamic execution threshold.

[0226] In this embodiment, each drag reduction direction is double-numbered and combined with cavitation drag reduction standards (such as cavitation encapsulation rate of 30%-40% as the optimal range) to determine the optimization direction of each drag reduction target and construct a basic control strategy.

[0227] Time series analysis (ARImA) was used to extract the spatiotemporal distribution characteristics of real-time cavitation drag reduction parameters, trajectory coordinates, and flight attitude parameters:

[0228] Perform a stationarity test (ADF test) on the time series of parameters. If the series is non-stationary, perform d-order differencing (d=1 or 2) to make the series stationary.

[0229] The values ​​of p1 and q1 are determined using the autocorrelation function (ACF) and the partial autocorrelation function (PACF), where p1 is the autoregression order and q1 is the moving average order, both ranging from 0 to 3.

[0230] Establish an ARIMA(p1,d,q1) model, fit the time series of the parameters, and obtain the autoregressive coefficients and moving average coefficients of the model, which are the time distribution characteristics of the parameters.

[0231] By combining the spatial coordinates (X, Y, Z) of the parameters, the temporal distribution characteristics are correlated with the spatial location to obtain the spatiotemporal distribution feature set.

[0232] It should be noted that the optimal order of the ARIMA model for cavitation encapsulation rate, horizontal offset, and angle of attack is ARIMA(1,1,1), which can meet the accuracy requirements of feature extraction.

[0233] In this embodiment, the coupling correlation characteristics between the parameters are pre-defined. The core coupling correlation is: angle of attack change → cavitation encapsulation rate change → drag change, horizontal offset change → angle of attack fine-tuning → cavitation distribution characteristic change.

[0234] The second scheme analyzes the segment control threshold (optimal range of cavitation parameters for each segment) and cavitation configuration benchmark (basic operating parameters of each cavitation device), and integrates them to construct a full-dimensional coupled feature set for cavitation drag reduction.

[0235] Combining the pressure and velocity distribution parameters of the underwater real-time flow field, the fuzzy comprehensive evaluation method is used to calculate the quantitative deviation value of the current cavitation drag reduction state relative to the optimal state. At the same time, the deviation value is divided into three state classification levels (Level 1: 0-0.2, excellent; Level 2: 0.2-0.5, medium; Level 3: >0.5, poor). Multiple linear regression is used to construct a quantitative mapping relationship between the state classification level, the quantitative deviation value and the cavitation drag reduction control requirements.

[0236] The baseline control parameters, full-dimensional coupled feature set, and quantized deviation value of the second scheme are input into the improved PPO deep reinforcement learning model. With the optimization directions of improving drag reduction efficiency, stabilizing the navigation attitude, and correcting the trajectory deviation as multiple objectives, a model reward function with a penalty term is designed (reward term weights: drag reduction efficiency 0.5, navigation attitude 0.3, trajectory deviation 0.2, which can be flexibly adjusted according to actual operation requirements (such as precise trajectory operation, high-speed drag reduction operation), and the weight sum is 1; penalty term: exceeding the power threshold / flow field mismatch, penalty coefficient -0.1). The model outputs multi-dimensional dynamic control parameters and adaptive fine-tuning parameters.

[0237] A CFD fast solution algorithm is used to perform cavitation flow field adaptation calculations on the control parameters. Combined with the dynamic execution threshold of the vehicle, mismatched and threshold-exceeding parameters are eliminated to obtain an effective set of control parameters. The CFD fast solution algorithm for cavitation flow field adaptation calculations on the control parameters includes:

[0238] Extract the current real-time underwater flow field parameters (pressure, velocity, water temperature, water pressure) and establish a simplified CFD simulation model (the number of grids is 1 / 10 of that of the fine model).

[0239] The control parameters output by the deep reinforcement learning model are input into the simplified model for fast CFD simulation. The simulation iteration steps are 200 steps, and the convergence residual is ≤1e-4.

[0240] Extract the cavitation encapsulation rate, cavitation distribution characteristics, and surface drag coefficient of the vehicle body from the simulation results and compare them with the optimal state;

[0241] The control parameters are considered to match the current flow field when all of the following conditions are met; otherwise, they are considered to be mismatched: The conditions are: the simulated cavitation encapsulation rate ∈ [30%, 40%]; the uniformity of cavitation distribution ≥ 0.8 (the variance of cavitation distribution on the surface of the vehicle body ≤ 0.2); the drag coefficient of the vehicle body surface ≤ 1.1 times the optimal drag coefficient; and there is no abnormal cavitation (such as cavitation bursting or cavitation shifting to the stern).

[0242] It should be noted that the power execution threshold is: cavitation venting flow rate threshold: main cavitation ≤ 1.0 The auxiliary cavitation unit is ≤0.5. ;

[0243] Cavitation angle adjustment threshold: -10° to +10°, relative to the hull axis; hull angle of attack fine-tuning threshold: -5° to +5°; Adjustment step threshold for control parameters: single adjustment of ventilation flow rate ≤ 0.05. The angle adjustment is ≤0.5° per cycle; the power execution threshold can be corrected according to the actual power system parameters of different underwater vehicles, and the corrected threshold needs to be entered into the system before model training.

[0244] In this embodiment, the fusion rules between the effective control parameter set and the basic control strategy are as follows:

[0245] Priority fusion: Based on the overall priority in the direction of drag reduction, the highest priority control parameters are fused first, followed by the medium and low priority parameters in turn;

[0246] Step size fusion: The adjustment step size of the effective control parameter must be less than or equal to the maximum step size set by the basic control strategy. If it exceeds the limit, it will be corrected according to the maximum step size.

[0247] Time-series fusion: For continuously adjustable parameters, the adjustment amount of the effective control parameter is distributed to each acquisition time according to the time series to achieve gradual adjustment and avoid abrupt changes in cavitation state caused by step adjustment;

[0248] Feedback fusion: The fused control parameters are incorporated into a real-time feedback mechanism. If the collected cavitation parameters exceed the optimal range, the current adjustment is immediately stopped and the effective control parameters are recalculated.

[0249] In this embodiment, for example, a high-speed underwater unmanned vehicle, the drag reduction directions of the second cavitation drag reduction scheme are: forward / side-mounted cavitation ventilator flow rate adjustment, cavitation ventilator angle adjustment, and body angle of attack fine-tuning. The corresponding drag reduction paths are: gradual flow rate adjustment, step angle adjustment, and continuous angle of attack adjustment. The drag reduction targets are to control the cavitation envelopment rate at 30%-40%, maintain symmetrical cavitation profiles, and stabilize the angle of attack at 0°±0.5°. A first correlation analysis is performed on the drag reduction paths, yielding a Pearson correlation coefficient of 0.1 between the ventilator flow rate adjustment and other paths, indicating it is the first priority. A second correlation analysis is performed on the drag reduction targets, yielding a Spearman correlation coefficient of 0.2 between the cavitation envelopment rate control and other targets, indicating it is the second priority. A basic control strategy is determined using dual numbering: priority is given to adjusting the forward cavitation ventilator flow rate, with the optimization direction being a slight increase in flow rate and an adjustment step not exceeding 0.05. The spatiotemporal distribution characteristics (cavitation envelope rate decreases by 1% every 5 seconds, horizontal offset increases by 0.1m every 5 seconds) and coupling correlation characteristics (an increase in 0.5° in angle of attack leads to an increase in cavitation envelope rate of 0.3%) of real-time cavitation drag reduction parameters (cavitation envelope rate decreases by 1% every 5 seconds, horizontal offset increases by 0.1m every 5 seconds) of flight attitude parameters (0.5° increase in angle of attack leads to an increase in cavitation envelope rate of 0.3%) are extracted. The segment control threshold (cavitation envelope rate 30%-40%) and cavitation configuration benchmark (forward cavitation 0.75) of the second scheme are analyzed. After integration, a full-dimensional coupled feature set for cavitation drag reduction is constructed. The current underwater real-time flow field parameters are pressure 0.06 MPa and flow velocity 24 m / s. Using the fuzzy comprehensive evaluation method, a quantization deviation value of 0.18 is calculated, and the state classification is Level 1 (excellent). A quantization mapping relationship is constructed through multiple linear regression, clarifying that the control requirement corresponding to a quantization deviation value of 0.1-0.2 under Level 1 state is a small-scale adjustment of the ventilation flow rate to maintain a stable angle of attack. The baseline control parameters for the second cavitation drag reduction scheme (forward cavitation 0.75) are then used. Side-mounted cavitation unit 0.2 The improved PPO deep reinforcement learning model is input with a cavitation drag reduction full-dimensional coupling feature set and a quantization deviation value of 0.18. A reward function with a penalty term is designed: drag, attitude, and trajectory. Here, drag is the drag reduction efficiency reward (1 when the cavitation encapsulation rate is 30%-40%, otherwise it decreases linearly according to the deviation), attitude is the navigation attitude reward (1 when the angle of attack is 0°±0.5°), trajectory is the trajectory deviation reward (1 when the offset is <1m), and P is the penalty term (1 when the parameter exceeds the threshold / flow field mismatch, otherwise it is 0). The model outputs multi-dimensional dynamic adjustment parameters: the forward cavitation venting flow rate is adjusted to 0.78. The side-mounted cavitation unit is maintained at 0.2. The cavitation angle remains unchanged, and the angle of attack is maintained at 0.5°; adaptive fine-tuning parameters: flow rate adjustment step size 0.01. The adjustment time is 10 seconds. Verification using a CFD fast solution algorithm shows that the control parameters perfectly match the current cavitation flow field (pressure 0.06 MPa, water flow velocity 24 m / s); simultaneously, the vehicle's power execution threshold is checked, with an airflow rate of 0.78... <1.0 The value did not exceed the threshold, therefore this parameter set is an effective control parameter set. The effective control parameter set is then compared with the basic control strategy (prioritizing the adjustment of the forward cavitation flow rate, with a step size ≤ 0.05). The fusion process generates the final optimized cavitation drag reduction scheme: Within the next 10 seconds, the forward cavitation airflow rate increases by 0.01... The step size starts from 0.75. Gradually increased to 0.78 The side-mounted cavitation unit maintains an airflow rate of 0.2 kW. The cavitation angle and the angle of attack of the aircraft remain unchanged. The cavitation encapsulation rate is monitored in real time. If it exceeds the 30%-40% range, the adjustment is stopped immediately and the system is re-optimized.

[0250] The beneficial effects of the above technical solution are as follows: By constructing a basic control strategy through dual-dimensional priority analysis of the drag reduction direction, extracting a full-dimensional coupled feature set to achieve a comprehensive representation of the cavitation, trajectory, and attitude state of the vehicle, and designing a multi-objective reward function with a penalty term using a deep reinforcement learning model, dynamic optimization of cavitation drag reduction parameters is achieved. At the same time, the executability of the control parameters is ensured through flow field adaptability calculation and dynamic threshold elimination. Finally, the optimized scheme generated by the fusion of the basic strategy can adapt to the actual cavitation state, trajectory deviation, and flow field environment of the underwater vehicle in real time. This solves the core problems of existing technologies where fixed drag reduction parameters cannot be adjusted online and are difficult to adapt to complex underwater dynamic environments. It realizes adaptive online optimization of the cavitation drag reduction scheme, improving drag reduction efficiency while ensuring the attitude stability and trajectory correction of the vehicle, forming a closed-loop optimization control for cavitation drag reduction.

[0251] This invention provides an underwater vehicle cavitation drag reduction optimization system based on AI deep reinforcement learning, such as... Figure 2 As shown, it includes:

[0252] The scheme determination module is used to determine the first cavitation drag reduction scheme of the target underwater vehicle based on the preset travel path of the target underwater vehicle, the morphological parameters of the target underwater vehicle, and the preset deep reinforcement learning model.

[0253] The scheme correction module is used to collect the actual travel path of the target underwater vehicle in real time, and correct the first cavitation drag reduction scheme according to the actual travel path to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

[0254] The location determination module is used to determine multiple image acquisition locations based on the morphological parameters of the target underwater vehicle, and to install an image sensor at each image acquisition location;

[0255] The image analysis module is used to acquire real-time operating environment images of the target underwater vehicle based on the image sensor, and to perform image analysis on the real-time operating environment images to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle.

[0256] The scheme optimization module is used to determine the cavitation drag reduction optimization scheme for the target underwater vehicle based on the real-time cavitation drag reduction parameters at each acquisition time, the trajectory coordinates and navigation attitude parameters of the target underwater vehicle, and in combination with the second cavitation drag reduction scheme.

[0257] The beneficial effects of the above technical solution are as follows: by using a deep reinforcement learning model to generate an initial drag reduction scheme in combination with preset path and morphological parameters, and dynamically correcting it according to the actual driving path, and then using targeted image sensors to obtain real-time cavitation drag reduction parameters, and integrating multi-dimensional navigation parameters with the corrected scheme to generate an optimized scheme, dynamic adaptive optimization of the cavitation drag reduction scheme is achieved. This effectively solves the problem that traditional fixed drag reduction parameters are difficult to adapt to actual underwater operating scenarios, greatly improves the cavitation drag reduction efficiency of underwater vehicles, and ensures their navigation stability.

[0258] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimizing cavitation drag reduction of underwater vehicles based on AI deep reinforcement learning, characterized in that, include: Step 1: Based on the preset travel path of the target underwater vehicle, the morphological parameters of the target underwater vehicle, and the preset deep reinforcement learning model, determine the first cavitation drag reduction scheme of the target underwater vehicle; Step 2: Collect the actual travel path of the target underwater vehicle in real time, and modify the first cavitation drag reduction scheme according to the actual travel path to obtain the second cavitation drag reduction scheme of the target underwater vehicle. Step 3: Based on the morphological parameters of the target underwater vehicle, determine multiple image acquisition locations and install an image sensor at each image acquisition location; Step 4: Based on the image sensor, acquire real-time operating environment images of the target underwater vehicle in real time, and perform image analysis on the real-time operating environment images to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle; Step 5: Based on the real-time cavitation drag reduction parameters at each acquisition time, the trajectory coordinates and navigation attitude parameters of the target underwater vehicle, and in conjunction with the second cavitation drag reduction scheme, determine the cavitation drag reduction optimization scheme for the target underwater vehicle. Step 5 includes: Extract the drag reduction path and drag reduction target for each drag reduction direction of the second cavitation drag reduction scheme, and perform a first correlation analysis on the drag reduction paths of any two drag reduction directions and a second correlation analysis on the drag reduction targets of any two drag reduction directions to determine the first priority of each drag reduction direction based on the drag reduction path and the second priority based on the drag reduction target. All drag reduction directions are sorted in the first priority order and assigned a first number to each drag reduction direction. At the same time, all drag reduction directions are sorted in the second priority order and assigned a second number to each drag reduction direction. Based on the first number and the second number, and according to the drag reduction standard, determine the optimization direction of each drag reduction target, and construct the basic control strategy for the second cavitation drag reduction scheme. Extract the spatiotemporal distribution characteristics and coupling correlation characteristics of each parameter in the real-time cavitation drag reduction parameters, trajectory coordinates, and flight attitude parameters at the corresponding acquisition time. At the same time, analyze the flight segment control threshold and cavitation configuration benchmark in the second cavitation drag reduction scheme to construct a full-dimensional coupling feature set for cavitation drag reduction. Based on the full-dimensional coupling feature set of cavitation drag reduction, combined with the pressure and velocity distribution parameters of the real-time underwater flow field environment, the quantitative deviation value and state classification level of the current cavitation drag reduction state relative to the optimal state are calculated. Based on the state attribution level and quantization deviation value, a quantization mapping relationship between cavitation drag reduction control requirements and execution parameters is constructed. The baseline control parameters of the second cavitation drag reduction scheme, the full-dimensional coupling feature set of cavitation drag reduction, and the quantization deviation value are input into the deep reinforcement learning model. With drag reduction efficiency improvement, navigation attitude stability, and trajectory deviation correction as multi-objective optimization directions, a model reward function with a penalty term is constructed to obtain the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters of the cavitation device. By sequentially calculating the cavitation flow field adaptability and removing parameters from the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters that are mismatched with the current underwater flow field environment or exceed the dynamic execution threshold of the vehicle body, an effective control parameter set is obtained. The effective set of control parameters is integrated with the basic control strategy of the second cavitation drag reduction scheme to generate an optimized cavitation drag reduction scheme.

2. The underwater vehicle cavitation drag reduction optimization method based on AI deep reinforcement learning according to claim 1, characterized in that, Step 1 includes: Determine the preset navigation parameters for each segment of the preset travel path of the target underwater vehicle, wherein the preset navigation parameters include: travel speed, travel depth, and angle of attack range; The preset navigation condition parameters of the target underwater vehicle in each segment and the morphological parameters of the target underwater vehicle are input into a preset deep reinforcement learning model to obtain the sub-drag reduction scheme of the target underwater vehicle in each segment. Based on the connection information between each segment and adjacent segments, the sub-drag reduction schemes of each segment are merged to obtain the first cavitation drag reduction scheme of the target underwater vehicle.

3. The underwater vehicle cavitation drag reduction optimization method based on AI deep reinforcement learning according to claim 1, characterized in that, Step 2 includes: Based on the position sensors and navigation module carried by the target underwater vehicle, the trajectory coordinates and navigation direction of the target underwater vehicle at each data acquisition time are determined; By integrating the trajectory coordinates and navigation direction at each acquisition moment, the actual travel path of the target underwater vehicle is obtained; Based on the actual travel path, the future travel path of the target underwater vehicle is predicted to obtain the predicted travel path. Based on the path deviation between the predicted travel path and the preset travel path, the first cavitation drag reduction scheme is modified to obtain the second cavitation drag reduction scheme for the target underwater vehicle.

4. The underwater vehicle cavitation drag reduction optimization method based on AI deep reinforcement learning according to claim 3, characterized in that, Based on the path deviation between the predicted driving path and the preset driving path, the first cavitation drag reduction scheme is corrected, including: The predicted driving path is compared with the preset driving path time by time to obtain the path deviation parameters of the target underwater vehicle at each future time within the preset time length. The path deviation parameters include: offset direction, horizontal offset and vertical offset. The path deviation parameters are normalized. At the same time, based on the environmental information of the preset travel path and the environmental information of the predicted travel path of the target underwater vehicle at each future time, the environmental deviation parameters for the corresponding future time are determined, and the weight of each path deviation parameter at the corresponding future time is determined. The deviation weight of the target underwater vehicle at the corresponding future time is calculated based on the weight of the path deviation parameter corresponding to each future time and the normalized path deviation parameter. The preset time length of the first future moment after the actual driving path is divided according to the preset sliding window, and the occurrence position of each future moment in each divided time window is determined. The deviation degree of the corresponding divided time window is determined by combining the deviation weight of the corresponding future moment. The average deviation of all time windows divided under the first future moment is taken as the deviation value of the corresponding first future moment. The next future moment is taken as the first future moment for further analysis to obtain the deviation value of each future moment in the preset time length of the first future moment after the actual driving path. If the number of future moments whose deviation value is within the first deviation interval after the actual driving path is greater than or equal to the first preset number, then the first cavitation drag reduction scheme of the target underwater vehicle is determined as the second cavitation drag reduction scheme of the target underwater vehicle. If the number of future moments whose deviation values ​​fall within the second deviation interval after the first future moment following the actual driving path is greater than or equal to the second preset number, then the deviation value of each future moment in the corresponding preset time length is combined with the first cavitation drag reduction scheme and input into the preset first deviation adjustment model to obtain the second cavitation drag reduction scheme of the target underwater vehicle.

5. The underwater vehicle cavitation drag reduction optimization method based on AI deep reinforcement learning according to claim 1, characterized in that, Step 3 includes: Key structural parameters are extracted from the morphological parameters of the target underwater vehicle, including: surface curvature distribution, cavitation topology, and bow and stern contour gradient. Under preset underwater environmental conditions, the spatial influence characteristics of the key structural parameters on the cavitation flow field are simulated to obtain the spatial distribution characteristics of the cavitation sensitive field corresponding to each structural parameter. Based on the spatial distribution characteristics of the cavitation sensitive field, and combined with the pressure and velocity distribution data of the cavitation flow field obtained by fluid dynamics simulation, the cavitation sensitive level region on the surface of the target underwater vehicle is divided. The quantitative mapping relationship between cavitation sensitivity level and flow field parameters is fitted to determine the image acquisition resolution, acquisition angle and coverage requirements for each cavitation sensitivity level region, and a multi-dimensional acquisition requirement matrix is ​​constructed. Based on the multi-dimensional acquisition requirement matrix, the performance parameters of the image sensor, and the key morphological feature points of each cavitation sensitivity level region, a spatial solution model based on surface curvature constraints is adopted to calculate the three-dimensional spatial coordinates of each image acquisition location and install the image sensor.

6. The underwater vehicle cavitation drag reduction optimization method based on AI deep reinforcement learning according to claim 1, characterized in that, Step 4 includes: The image sensor is controlled to acquire real-time images of the target underwater vehicle's operating environment at a preset sampling frequency; The real-time operating environment image is subjected to underwater noise filtering, illumination compensation and contrast enhancement to obtain the processed real-time operating environment image; The processed real-time operating environment image is input into the image processing model to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle. The real-time cavitation drag reduction parameters include: cavitation profile, cavitation distribution characteristics, and cavitation encapsulation rate of the target underwater vehicle.

7. A cavitation drag reduction optimization system for underwater vehicles based on AI deep reinforcement learning, characterized in that, include: The scheme determination module is used to determine the first cavitation drag reduction scheme of the target underwater vehicle based on the preset travel path of the target underwater vehicle, the morphological parameters of the target underwater vehicle, and the preset deep reinforcement learning model. The scheme correction module is used to collect the actual travel path of the target underwater vehicle in real time, and correct the first cavitation drag reduction scheme according to the actual travel path to obtain the second cavitation drag reduction scheme of the target underwater vehicle. The location determination module is used to determine multiple image acquisition locations based on the morphological parameters of the target underwater vehicle, and to install an image sensor at each image acquisition location; The image analysis module is used to acquire real-time operating environment images of the target underwater vehicle based on the image sensor, and to perform image analysis on the real-time operating environment images to obtain the real-time cavitation drag reduction parameters of the target underwater vehicle. The scheme optimization module is used to determine the cavitation drag reduction optimization scheme for the target underwater vehicle based on the real-time cavitation drag reduction parameters at each acquisition time, the trajectory coordinates and navigation attitude parameters of the target underwater vehicle, and in combination with the second cavitation drag reduction scheme. The scheme optimization module is used for: Extract the drag reduction path and drag reduction target for each drag reduction direction of the second cavitation drag reduction scheme, and perform a first correlation analysis on the drag reduction paths of any two drag reduction directions and a second correlation analysis on the drag reduction targets of any two drag reduction directions to determine the first priority of each drag reduction direction based on the drag reduction path and the second priority based on the drag reduction target. All drag reduction directions are sorted in the first priority order and assigned a first number to each drag reduction direction. At the same time, all drag reduction directions are sorted in the second priority order and assigned a second number to each drag reduction direction. Based on the first number and the second number, and according to the drag reduction standard, determine the optimization direction of each drag reduction target, and construct the basic control strategy for the second cavitation drag reduction scheme. Extract the spatiotemporal distribution characteristics and coupling correlation characteristics of each parameter in the real-time cavitation drag reduction parameters, trajectory coordinates, and flight attitude parameters at the corresponding acquisition time. At the same time, analyze the flight segment control threshold and cavitation configuration benchmark in the second cavitation drag reduction scheme to construct a full-dimensional coupling feature set for cavitation drag reduction. Based on the full-dimensional coupling feature set of cavitation drag reduction, combined with the pressure and velocity distribution parameters of the real-time underwater flow field environment, the quantitative deviation value and state classification level of the current cavitation drag reduction state relative to the optimal state are calculated. Based on the state attribution level and quantization deviation value, a quantization mapping relationship between cavitation drag reduction control requirements and execution parameters is constructed. The baseline control parameters of the second cavitation drag reduction scheme, the full-dimensional coupling feature set of cavitation drag reduction, and the quantization deviation value are input into the deep reinforcement learning model. With drag reduction efficiency improvement, navigation attitude stability, and trajectory deviation correction as multi-objective optimization directions, a model reward function with a penalty term is constructed to obtain the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters of the cavitation device. By sequentially calculating the cavitation flow field adaptability and removing parameters from the multi-dimensional dynamic control parameters and adaptive fine-tuning parameters that are mismatched with the current underwater flow field environment or exceed the dynamic execution threshold of the vehicle body, an effective control parameter set is obtained. The effective set of control parameters is integrated with the basic control strategy of the second cavitation drag reduction scheme to generate an optimized cavitation drag reduction scheme.

Images