A method for reconstructing a near-surface current profile of a buoy based on deep learning

Abstract

The application provides a kind of based on deep learning's buoy near surface current velocity profile reconstruction method, belong to marine hydrological observation technical field. Offline stage, construct the paired flow field of having buoy and no buoy working condition by using computational fluid dynamics simulation, with the disturbance profile of superimposed buoy attitude and blind area mask as input, no buoy bottom profile as label, train deep learning network; Online stage, rigid body motion compensation is carried out to the data collected by buoy-mounted acoustic Doppler current profiler and dynamic blind area mask is generated, the disturbance profile, buoy attitude and blind area mask are spliced and input into the model, and the continuous bottom flow velocity profile that compensates hydrodynamic disturbance and completes near surface blind area is output. The application simultaneously solves the distortion problem of flow velocity profile caused by buoy hydrodynamic interference and acoustic top blind area, has strong physical consistency, high calculation efficiency, and can be deployed in embedded processing unit for real-time processing.

Description

Technical Field

[0001] This invention relates to the field of marine hydrological observation technology, and in particular to a method for reconstructing near-surface current profiles of buoys based on deep learning. Background Technology

[0002] Current velocity profiles are core parameters for oceanographic observation, directly supporting port and waterway design, vessel scheduling, marine engineering construction, and marine disaster early warning. In-situ vertical current velocity profile observation using an acoustic Doppler current profiler mounted on a buoy is currently the mainstream technique for acquiring current velocity profile data.

[0003] When a buoy is equipped with an acoustic Doppler current profiler, a top blind zone exists between the transducer end face and the first effective measurement depth unit due to the limitations of acoustic propagation physics, resulting in missing near-surface velocity data. Simultaneously, the buoy's underwater structure exerts hydrodynamic interference on the surrounding natural flow field, causing the measured velocity within the effective measurement depth range to deviate from the natural flow field. The superposition of the top blind zone and flow field distortion causes systematic distortion in the near-surface velocity profile.

[0004] In existing technologies, the problem of the top blind zone is often addressed by adding a central vertical beam to reduce its range. However, this approach cannot overcome the inherent characteristics of the physical components and cannot solve the flow field disturbance problem caused by the buoy structure. Regarding the flow field disturbance problem, methods based on potential flow theory do not consider turbulence effects and lack accuracy in complex flow field regions; methods based on computational fluid dynamics numerical simulation are time-consuming per calculation and cannot meet real-time processing requirements; data-driven machine learning methods lack fluid dynamics physical constraints, and the prediction results are prone to non-physical oscillations, resulting in insufficient generalization ability. Therefore, existing technologies have not yet formed a technical solution that can simultaneously achieve near-surface blind zone completion and flow field disturbance correction, while balancing physical consistency and computational efficiency. Summary of the Invention

[0005] In view of the above problems, a deep learning-based method for reconstructing near-surface velocity profiles of buoys is proposed to overcome or at least partially solve these problems, including: Acquire the data stream collected by the acoustic Doppler current profiler on the observation buoy and the motion and attitude data of the observation buoy; Rigid body motion compensation is performed on the original velocity profile in the data stream using motion data to obtain the disturbed velocity profile; A dynamic blind spot mask is generated using device configuration parameters and data quality identifiers in the data stream; The disturbed flow velocity profile, attitude data, and dynamic blind zone mask are concatenated into a vertical feature tensor. The vertical feature tensor is input into the velocity profile reconstruction model, and the output is a continuous background velocity profile at full depth. Among them, the velocity profile reconstruction model is a deep learning network, which is trained offline using training samples constructed by computational fluid dynamics simulation. The input features of the training samples are the buoy-performed disturbance profile with superimposed buoy attitude data and blind zone mask. The labels of the training samples are the background velocity profile without buoy under the same boundary conditions.

[0006] Optionally, generate a dynamic blind zone mask, including: Obtain the blind zone distance included in the device configuration parameters; Extract real-time pressure depth, pitch angle, and roll angle from the data stream; Based on the real-time pressure depth, the real-time vertical distance from the transducer end face to the free water surface is calculated. Based on the blind spot distance and the pitch and roll angles, calculate the vertical projection distance of the blind spot distance; Depth cells whose depth is within the range of the vertical distance from the transducer end face to the free water surface and the blind zone distance projected in the vertical direction are identified as hard blind zones. The goodness rate of each depth cell is obtained from the data stream, and depth cells with a goodness rate lower than a preset threshold are identified as soft blind spots. Merge hard and soft blind zones to generate a dynamic blind zone mask.

[0007] Optionally, training samples are obtained by constructing a flow field sample library. The steps for constructing the flow field sample library include: A three-dimensional model is established based on the underwater geometric features of the observation buoy, and a numerical computation domain is constructed based on the three-dimensional model and the unsteady Reynolds time-averaged Navier-Stokes equations. In the numerical computation domain, overlapping grid technology is used to divide the buoy motion subdomain and the background flow field domain, and a turbulence model is used to solve the flow field of the buoy motion subdomain and the background flow field domain. Multiple sets of incoming flow velocity, flow direction, water depth, and waves are set, and boundary conditions corresponding to the incoming flow velocity, flow direction, water depth, and waves are applied to the numerical computation domain; Simultaneous simulation of three-dimensional flow fields under the same boundary conditions, with and without buoys; Extract the vertical velocity profile directly below the buoy in the buoy-containing working condition, and perform spatial integration averaging according to the depth unit size of the acoustic Doppler current profiler to obtain a disturbed velocity profile with the same resolution as the actual measurement. Using perturbation velocity profiles superimposed with attitude data and blind zone masks as input features, and buoy-free operating condition velocity profiles as labels, mapping sample pairs are formed to create a flow field sample library.

[0008] Optionally, the deep learning network includes a feature embedding layer, a feature extraction layer, and a physical information decoding layer; The feature embedding layer performs a linear projection on the vertical feature tensor, maps each depth unit to a unified dimension token, and superimposes the depth position encoding. The feature extraction layer extracts flow field perturbation features from the tokens; The physical information decoding layer maps the flow field disturbance characteristics to the background velocity profile of the full depth layer.

[0009] Optionally, the total loss function during deep learning network training is composed of the blind zone prediction loss, the effective observation layer consistency loss, and the vertical profile smoothing loss, weighted and summed. Blind zone prediction loss measures the deviation between the predicted velocity of the blind zone depth cell and the computational fluid dynamics label; Effective observation layer consistency loss measures the deviation between predicted and measured flow velocities in effective observation depth cells; Vertical profile smoothing loss measures the degree of change in predicted flow velocity between adjacent depth cells.

[0010] Optionally, the feature extraction layer employs a shift-window multi-head self-attention mechanism to extract flow field perturbation features, including: The token sequence is divided into several non-overlapping local windows along the depth direction, and each local window contains a preset number of depth units; Within each local window, multi-head self-attention is calculated on the tokens within the window to obtain the self-attention output within the window, and the flow field feature correlation between depth cells within the window is extracted; The local window is shifted along the depth direction by a preset step size. Multi-head self-attention is calculated in the new window obtained by the shift, and the self-attention output of the shifted window is obtained, so as to realize the interaction of feature information between adjacent local windows. The self-attention output within the window and the self-attention output of the shifted window are stacked alternately to output a token sequence containing multi-scale flow field perturbation features.

[0011] Optionally, the velocity profile reconstruction model can output a continuous background velocity profile across the entire depth layer in the following way: The measured flow velocity within the effective measurement depth range is processed by the flow profile reconstruction model to eliminate the hydrodynamic disturbance increment caused by the underwater structure of the observation buoy to the surrounding flow field, and the corrected flow velocity value of the lower boundary depth unit of the near-surface blind zone is obtained. Using the corrected flow velocity value of the lower boundary depth unit of the near-surface blind zone as the boundary condition for extrapolation, and combining the vertical distribution law of flow velocity learned by the deep learning network from the computational fluid dynamics simulation samples during the training process, the flow velocity value of each depth unit of the near-surface blind zone is extrapolated layer by layer along the vertical direction to complete the vertical upward extrapolation reconstruction. Among them, the velocity profile after correction of the effective measurement depth range and the velocity profile obtained by extrapolation and reconstruction of the near-surface blind zone together constitute the continuous background velocity profile of the full depth layer, which is uniformly output by the velocity profile reconstruction model.

[0012] Optionally, when simultaneously simulating the three-dimensional flow field under the same boundary conditions with and without buoys, the boundary conditions applied to the numerical computation domain include: The upstream boundary along the direction of the incoming flow is set as the velocity inlet boundary, specifying the inlet flow velocity and direction, and the inlet wave generation method is used to simulate wave propagation; The downstream boundary along the incoming flow direction is set as a zero-gradient free outflow boundary; The two lateral boundaries perpendicular to the direction of the incoming flow are set as symmetrical boundaries; The bottom of the computational domain is set to a non-slip wall boundary; In the case of buoy operation, the surface of the underwater structure of the buoy is set as a non-slip wall boundary.

[0013] This invention, by organically coupling computational fluid dynamics simulation with a deep learning network, achieves for the first time an integrated process for correcting hydrodynamic disturbances and completing near-surface blind zones in the reconstruction of near-surface velocity profiles of buoys. In the offline phase, computational fluid dynamics is used to perform paired simulations of various operating conditions with and without buoys. The background velocity profile of the buoy-free condition is used as a label to provide training samples with embedded fluid dynamic physical laws for the deep learning network. In the online application phase, only the disturbed velocity profile compensated for rigid body motion, attitude data, and dynamic blind zone mask need to be input into the model. A single forward inference can then output a continuous background velocity profile that compensates for hydrodynamic disturbances and completes the near-surface blind zone. This method fundamentally overcomes the systematic distortion of near-surface velocity profiles caused by the superposition of underwater structural flow field distortion and acoustic top blind zone, significantly improving the integrity and accuracy of the velocity profile within the near-surface range of 0 to 1.5 meters.

[0014] Compared to existing correction methods based on potential flow theory, this invention constructs training samples using high-fidelity flow field data from computational fluid dynamics (CFD) simulations, fully considering turbulence effects in the flow field and avoiding the accuracy limitations of potential flow theory in complex flow field regions such as the buoy wake. Compared to existing purely data-driven deep learning methods, this invention uses the physical results of CFD simulations as supervision signals. The mapping relationships learned by the model embed fluid dynamics physical laws, avoiding the unnatural oscillations and insufficient generalization capabilities that arise from the lack of physical constraints in purely data-driven methods, ensuring physical consistency in the reconstructed profiles. In terms of computational efficiency, this invention limits time-consuming CFD numerical simulations to the offline training stage. Online inference only requires a lightweight deep learning model to complete a single forward computation, with a single profile reconstruction taking less than one second. It can be directly deployed in the buoy's embedded processing unit for in-situ real-time processing, solving the bottleneck of traditional CFD methods failing to meet real-time observation requirements.

[0015] Furthermore, this invention requires no modification to existing buoy hardware; it can be adapted to various deployed observation buoy systems simply by upgrading the software algorithm. The modification cost is low, and it is easy to scale up and promote. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a buoy near-surface velocity profile reconstruction method based on deep learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the training data organization format provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a near-surface velocity profile reconstruction model provided in an embodiment of the present invention; Figure 4 This is a comparison chart of simulation results of a method provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 This invention provides a method for reconstructing near-surface current profiles of buoys based on deep learning, which may specifically include: Acquire the data stream collected by the acoustic Doppler current profiler on the observation buoy and the motion and attitude data of the observation buoy; Rigid body motion compensation is performed on the original velocity profile in the data stream using motion data to obtain the disturbed velocity profile; A dynamic blind spot mask is generated using device configuration parameters and data quality identifiers in the data stream; The disturbed flow velocity profile, attitude data, and dynamic blind zone mask are concatenated into a vertical feature tensor. The vertical feature tensor is input into the velocity profile reconstruction model, and the output is a continuous background velocity profile at full depth. Among them, the velocity profile reconstruction model is a deep learning network, which is trained offline using training samples constructed by computational fluid dynamics simulation. The input features of the training samples are the buoy-performed disturbance profile with superimposed buoy attitude data and blind zone mask. The labels of the training samples are the background velocity profile without buoy under the same boundary conditions.

[0020] In actual operation, the observation buoy is equipped with an acoustic Doppler current profiler and a combination of a global navigation satellite system and an inertial measurement unit. The acoustic Doppler current profiler is used to collect current velocity data at various depths in the water. Its working principle is that the transducer emits acoustic pulses downwards. When the sound waves encounter scattering objects such as suspended particles and plankton in the water, they produce scattered echoes. The transducer receives the scattered echoes and calculates the radial velocity along each beam direction according to the Doppler frequency shift principle. Then, the three-dimensional velocity components in the geodetic coordinate system are obtained through vector synthesis.

[0021] The acoustic Doppler current profiler divides the water body into multiple depth units along the depth direction, with each depth unit corresponding to a three-dimensional velocity vector. The data stream acquired by the acoustic Doppler current profiler includes the eastward, northward, and vertical current velocities in the geodetic coordinate system for each depth unit, as well as real-time pressure depth, echo intensity, beam correlation coefficient, and data quality rate for each depth unit. The data stream also carries the configuration parameters of the acoustic Doppler current profiler, including blind zone distance. A combined Global Navigation Satellite System (GNSS) and Inertial Measurement Unit (INS) system is used to synchronously acquire motion and attitude data of the observation buoy. Motion data includes the buoy's three-dimensional translational velocity and three-axis angular velocity. Attitude data includes the buoy's pitch, roll, and rotation angles. The three-dimensional translational velocity describes the speed of the observation buoy in the east-west, north-south, and vertical directions. The three-axis angular velocities correspond to the angular rates of the buoy's roll, pitch, and bow movements, respectively. The pitch angle is the angle at which the observation buoy rotates around the horizontal axis, the roll angle is the angle at which the observation buoy rotates around the vertical axis, and the gyration angle is the azimuth of the observation buoy's horizontal orientation relative to the north.

[0022] The first step in the data preprocessing stage is time synchronization and alignment. The timestamps recorded by the acoustic Doppler current profiler and the combined GPS and inertial measurement unit (IMU) may have slight clock discrepancies. Therefore, before rigid body motion compensation, the two types of data need to be matched by timestamps. Using Coordinated Universal Time (UTC) output by the GPS as the reference time, the time recorded by each measurement profile of the acoustic Doppler current profiler is paired with the nearest time in the attitude data, controlling the time deviation to within one sampling period of the acoustic Doppler current profiler.

[0023] After time alignment, rigid body motion compensation is performed. The observation buoy experiences six degrees of freedom motion on the sea surface due to wave excitation: translation along three orthogonal directions and rotation around three orthogonal axes. The translational motion directly causes a relative velocity between the acoustic Doppler current profiler transducer and the water body, resulting in the measured current velocity being superimposed with the buoy's own motion velocity component. The purpose of rigid body motion compensation is to eliminate this superposition.

[0024] Specifically, for each depth cell of the acoustic Doppler current profiler, the velocity vector measured by the acoustic Doppler current profiler is subtracted from the three-dimensional translational velocity vector of the observed buoy, synchronously output by the combined global navigation satellite system and inertial measurement unit. This vector difference calculation is performed cell by cell to obtain the velocity profile after deducting the buoy's own motion component. The compensated velocity profile is called the disturbed velocity profile. The transducer tilt caused by the pitch and roll angles has been corrected during the beam coordinate to geodetic coordinate conversion process within the acoustic Doppler current profiler. The effective measurement depth range covered by the disturbed velocity profile extends from the depth of the first effective measurement depth cell to the maximum effective operating depth of the acoustic Doppler current profiler. Within this range, the velocity data still contains hydrodynamic disturbances caused by the buoy's underwater structure to the surrounding water flow.

[0025] Next, a near-surface dynamic blind zone mask is generated. The near-surface blind zone is the depth range in the top depth region near the free surface where the acoustic Doppler current profiler cannot obtain effective velocity data. Its causes include installation limitations, the inherent blind zone distance of the transducer, and the shielding of acoustic signals by bubble entrainment under conditions of high winds and waves. Furthermore, the blind zone range dynamically changes with sea state. When generating the dynamic blind zone mask, the pressure depth, equipment configuration parameters, and data quality identifiers in the data stream are used to identify and mark the locations of near-surface invalid depth units where velocity data is missing due to installation limitations and acoustic blind zones. A one-dimensional binary vector with the same number of depth layers as the disturbed velocity profile is generated, where effective measurement depth units are marked with a value of 1, and invalid blind zone depth units are marked with a value of 0.

[0026] The obtained perturbation velocity profile, attitude data, and dynamic blind zone mask are concatenated into a vertical feature tensor. The perturbation velocity profile is numerically a two-dimensional matrix, with the number of rows representing the total number of depth cells and the number of columns representing the number of velocity components. Each depth cell's corresponding feature vector includes eastward, northward, and vertical velocities. The pitch, roll, and yaw angles in the attitude data are single numerical values, which can be expanded in the row direction to the same length as the total number of depth cells. The dynamic blind zone mask is a one-dimensional vector with the same length as the total number of depth cells. The concatenation operation is performed in the channel dimension, adding the pitch, roll, yaw, and dynamic blind zone mask as new columns to the channel dimension of the perturbation velocity profile matrix. This ensures that the feature vector of each depth cell includes eastward, northward, and vertical velocities, as well as the pitch, roll, yaw, and yaw angles, and a validity flag for that depth cell. The vertical feature tensors are arranged along the depth direction, and the deep learning network uses validity labels to distinguish between valid observation areas and blind areas that need to be reconstructed and completed.

[0027] The vertical feature tensor is input into the velocity profile reconstruction model. The velocity profile reconstruction model is a pre-trained deep learning network trained offline, with training samples constructed from computational fluid dynamics simulations. The input features of the training samples are the buoy-capped perturbation profile and attitude data with a superimposed blind zone mask, and the labels of the training samples are the background velocity profiles under the same boundary conditions without a buoy. The buoy-capped scenario simulates the flow field state after buoy deployment in actual ocean observations. The underwater structure of the buoy impedes and bypasses the water flow, causing a deviation between the velocity profile directly below the buoy and the natural flow field. The buoy-free scenario simulates the natural flow field state without any structural interference, and its velocity profile represents the true vertical velocity distribution. The deep learning network establishes a nonlinear mapping relationship from the perturbation velocity profile with a superimposed blind zone mask to the background velocity profile without a buoy through iterative learning on a large number of paired simulation samples.

[0028] After offline training, the deep learning network is deployed in the embedded processing unit of the observation buoy. The embedded processing unit is an industrial-grade single-board computer within the buoy's instrument compartment, equipped with a central processing unit and hardware acceleration resources for deep learning inference. In online applications, after each acoustic Doppler current profiler measurement cycle, the data processing flow sequentially executes time alignment, rigid body motion compensation, dynamic blind zone mask generation, vertical feature tensor construction, and model inference. After receiving the vertical feature tensor, the current profile reconstruction model outputs a continuous background current profile across the entire depth layer through internal calculations within the deep learning network. This background current profile covers the entire depth range from near the free surface to the maximum effective measurement depth of the acoustic Doppler current profiler. Within the effective measurement depth range, the hydrodynamic disturbance increments caused by the buoy's underwater structure have been eliminated, and current velocity completion has been completed within the near-surface blind zone. The execution time of the entire data preprocessing and model inference process is controlled within one second, meeting the time cycle requirements for continuous in-situ real-time monitoring of the observation buoy.

[0029] In one or more embodiments of the present invention, generating a dynamic blind zone mask includes: Obtain the blind zone distance included in the device configuration parameters; Extract real-time pressure depth, pitch angle, and roll angle from the data stream; Based on the real-time pressure depth, the real-time vertical distance from the transducer end face to the free water surface is calculated. Based on the blind spot distance and the pitch and roll angles, calculate the vertical projection distance of the blind spot distance; Depth cells whose depth is within the range of the vertical distance from the transducer end face to the free water surface and the blind zone distance projected in the vertical direction are identified as hard blind zones. The goodness rate of each depth cell is obtained from the data stream, and depth cells with a goodness rate lower than a preset threshold are identified as soft blind spots. Merge hard and soft blind zones to generate a dynamic blind zone mask.

[0030] Specifically, the equipment configuration parameters of the acoustic Doppler current profiler can be obtained first. These parameters are written into the device firmware at the factory and output along with the data stream each time data is acquired, including the blind zone distance. The blind zone distance is a fixed distance, measured in meters, between the transducer end face and the center of the first measurable depth cell. The blind zone distance is determined by the transducer's inherent blind zone distance and near-field disturbance factors. The inherent blind zone distance occurs because the transducer requires a recovery time after emitting an acoustic pulse to receive the echo normally. During this period, the acoustic signal arriving at the transducer end face cannot be effectively processed, therefore depth cells within this range do not have the capability to measure current velocity. In three-dimensional open ocean areas, when fluid flows around a buoy, local acceleration and boundary layer separation due to streamline compression cause near-field disturbances below the buoy, typically ranging from 0.1D to 0.3D (where D is the buoy diameter). In practical applications, the near-field disturbance layer is usually covered by the blind zone distance configured in the equipment.

[0031] Real-time pressure depth, pitch angle, and roll angle are extracted from the data stream. Real-time pressure depth is measured by the pressure sensor built into the acoustic Doppler current profiler. The pressure sensor senses the hydrostatic pressure generated by the water column above the transducer location, and the real-time vertical distance from the transducer end face to the free water surface is calculated using the linear relationship between hydrostatic pressure and water depth. Under wave conditions, the observation buoy moves with the water surface, and the real-time vertical distance changes periodically accordingly; therefore, the real-time pressure depth needs to be read separately for each measurement profile. Pitch and roll angles are read from the data stream acquired by the acoustic Doppler current profiler. The pitch angle describes the pitch amplitude of the observation buoy's rotation about its transverse axis in the bow direction, and the roll angle describes the lateral sway amplitude of the observation buoy's rotation about its longitudinal axis; both are in degrees. Under calm water conditions, both pitch and roll angles are close to 0 degrees; under high wind and wave conditions, their variation increases.

[0032] The real-time vertical distance from the transducer end face to the free water surface is calculated based on real-time pressure depth. The pressure sensor measurement directly corresponds to the hydrostatic pressure of the water column above the transducer location. The water density and atmospheric pressure at sea level are known quantities or can be obtained through on-site calibration. The real-time vertical distance from the transducer end face to the free water surface is calculated according to the hydrostatic pressure formula.

[0033] The projected distance of the blind zone in the vertical direction is calculated based on the blind zone distance, pitch angle, and roll angle. The tilt of the observation buoy under wave action is the result of the combined rotation of two orthogonal axes in the horizontal plane. The rotation of the buoy around the horizontal axis produces the pitch angle, and the rotation around the vertical axis produces the roll angle. These two rotational components together determine the degree of tilt of the buoy's deck plane relative to the horizontal plane. By combining the pitch angle and the roll angle into a single spatial tilt angle, the instantaneous tilt angle of the buoy relative to the water surface normal is obtained. The water surface normal is perpendicular to the free water surface, and under still water conditions, the water surface normal coincides with the direction of gravity. The instantaneous tilt angle of the buoy reflects the deflection angle of the transducer beam emission direction relative to the water surface normal. When the observation buoy tilts to one side, the angle between the beam towards the tilted side and the water surface normal increases, while the angle between the beam away from the tilted side and the water surface normal decreases. The blind zone distance is perpendicular to the transducer end face. When the transducer tilts due to the influence of the observed buoy's movement, the projected distance of the blind zone in the vertical direction is calculated using the instantaneous tilt angle of the buoy.

[0034] The combined vertical distance from the transducer end face to the free water surface and the projected distance of the blind zone in the vertical direction constitute a physically unobservable depth range. All depth cells within this depth range are identified as hard blind zones. Critical depth cells within this depth range are also identified as hard blind zones.

[0035] The data goodness rate for each depth cell is obtained from the data stream. The data goodness rate is the percentage of valid measurements successfully obtained by the acoustic Doppler current profiler in each depth cell out of the total number of measurements, expressed as a percentage. A high data goodness rate indicates good acoustic propagation conditions, sufficient scattering bodies, and no significant interference within the depth cell, resulting in reliable velocity measurements. A low data goodness rate indicates problems such as acoustic signal attenuation, insufficient scattering bodies, or bubble shielding within the depth cell, leading to unreliable velocity measurements. A preset threshold is a critical value for determining data usability; this value is determined based on the characteristics of the acoustic Doppler current profiler and long-term observation experience. All depth cells are traversed, and depth cells with a data goodness rate below the preset threshold are identified as soft blind zones.

[0036] The echo intensity of each depth cell is obtained from the data stream. Echo intensity is the intensity of the scattered acoustic wave signal received by the acoustic Doppler current profiler, measured in decibels (dB). In normal water bodies, echo intensity gradually decreases with increasing depth, and the attenuation pattern is related to the sound wave propagation distance and the water's acoustic absorption coefficient. When a large number of bubbles are entrained in the near-surface water due to wave breaking, the bubbles strongly scatter the sound waves, causing the echo intensity of the corresponding depth cell to abnormally increase, exceeding the background echo intensity of the normal water body by several times. Depth cells with echo intensities higher than the normal background echo intensity by a preset multiple are identified as soft blind zones. The normal background echo intensity can be obtained through statistical analysis of the echo intensity values ​​of deeper depth cells with good data quality in this profile.

[0037] Hard and soft blind zones are combined using a logical OR operation. Hard blind zones are depth cells identified by projecting the real-time vertical distance from the transducer end face to the free water surface and the blind zone distance in the vertical direction. Soft blind zones are depth cells marked based on data goodness rate or echo intensity triggering threshold conditions. The logical OR operation means that if a depth cell meets the conditions of any of the three types of blind zones, it is determined as an invalid depth cell. The combined result generates a one-dimensional binary vector with the same number of depth layers as the disturbed velocity profile, and the vector length is equal to the total number of depth cells in the current measurement profile of the acoustic Doppler current profiler. Each element in the vector takes a value of 1 or 0; elements corresponding to valid measured depth cells have a value of 1, and elements corresponding to invalid depth cells in the blind zone have a value of 0. This binary vector is the near-surface dynamic blind zone mask.

[0038] The term "dynamic" in "dynamic blind zone mask" has two meanings. First, the dynamic blind zone mask for the same observation buoy differs under different sea states and at different measurement times. Under normal sea states, the bubble content in the near-surface water is low, the soft blind zone is extremely small or even nonexistent, and the number of invalid depth cells in the dynamic blind zone mask is small, typically only including depth cells corresponding to the hard blind zone. Under rough sea states, waves break and engulf a large number of bubbles, expanding the soft blind zone and significantly increasing the number of invalid depth cells in the dynamic blind zone mask, extending to deeper locations. Second, the dynamic blind zone mask is not a pre-set fixed value, but is independently calculated and generated based on the real-time data stream acquired in each measurement cycle. It is highly coupled with the instantaneous state of the observation buoy and the marine environmental conditions, truly reflecting the effective data coverage of the current measurement profile.

[0039] The dynamic blind zone mask corresponds one-to-one with the depth cells of the perturbation velocity profile. By using the dynamic blind zone mask in conjunction with the perturbation velocity profile, the subsequent deep learning network can accurately determine whether the velocity data of each depth cell is a valid observation or a missing value that needs to be filled in. This allows for hydrodynamic perturbation correction in the valid observation area and velocity completion in the blind zone area by combining the surrounding valid data and the physical laws of the flow field.

[0040] In one or more embodiments of the present invention, training samples are obtained by constructing a flow field sample library, the step of constructing the flow field sample library including: A three-dimensional model is established based on the underwater geometric features of the observation buoy, and a numerical computation domain is constructed based on the three-dimensional model and the unsteady Reynolds time-averaged Navier-Stokes equations. In the numerical computation domain, overlapping grid technology is used to divide the buoy motion subdomain and the background flow field domain, and a turbulence model is used to solve the flow field of the buoy motion subdomain and the background flow field domain. Set multiple sets of incoming flow velocity, direction, water depth, and waves, and apply the corresponding boundary conditions to the numerical computation domain. Simultaneous simulation of three-dimensional flow fields under the same boundary conditions, with and without buoys; Extract the vertical velocity profile directly below the buoy in the buoy-containing working condition, and perform spatial integration averaging according to the depth unit size of the acoustic Doppler current profiler to obtain a disturbed velocity profile with the same resolution as the actual measurement. The perturbation velocity profile superimposed with attitude data and blind zone mask is used as input features, and the velocity profile under buoyless conditions is used as a label to form a mapping sample pair, thus forming a flow field sample library.

[0041] The flow field sample library serves as the data foundation for offline training of deep learning networks. Since it is impossible to simultaneously obtain paired velocity profile data with and without buoys under the same sea state in a real ocean environment, computational fluid dynamics numerical simulation is used to reproduce the flow field state of the buoy under different incoming flow conditions in a virtual numerical computation domain. The three-dimensional flow fields with and without buoys are calculated separately, and paired velocity profiles are extracted as training samples.

[0042] The first step in constructing the flow field sample library is to establish a three-dimensional parametric geometric model based on the underwater geometric features of the observation buoy. The underwater geometric features of the observation buoy include the buoy's main body shape, the mounting base of the acoustic Doppler current profiler transducer, the biofouling shield, and other sensors and structural components protruding from the buoy's hull surface. These underwater structures collectively constitute obstructions to the surrounding water flow. The establishment of the three-dimensional parametric geometric model can be completed using 3D modeling software, completely reproducing the geometric dimensions, relative positions, and assembly relationships of each underwater component of the observation buoy. Geometric features that significantly affect the water flow need to be retained in the model, including the shape of the upstream face of structural components, the chamfer radius of edges, and the gaps between components. Minor geometric features with little impact on the water flow can be appropriately simplified to reduce the complexity of subsequent mesh generation and computational resource consumption. After the three-dimensional parametric geometric model is constructed, it is placed in a predetermined position within the numerical computation domain.

[0043] A numerical computational domain is constructed based on a three-dimensional model and the unsteady Reynolds-averaged Navier-Stokes equations. The unsteady Reynolds-averaged Navier-Stokes equations are the fundamental governing equations describing the motion of viscous incompressible fluids. These equations decompose the instantaneous flow field into two parts: a time-averaged quantity and a fluctuating quantity. By averaging over time, the high-frequency components of turbulent fluctuations are eliminated, while the evolution of the average flow motion is preserved. The numerical computational domain is a cuboid spatial region surrounding the observation buoy. Its spatial dimensions need to satisfy the condition of eliminating the influence of the boundary on the flow field simulation results. The length of the computational domain along the direction of the incoming flow is typically several tens of times the characteristic size of the buoy. The length of the upstream region of the buoy is used to ensure that the incoming flow develops sufficiently before reaching the buoy, while the length of the downstream region is used to fully capture the flow structure of the buoy's wake region. The width of the computational domain perpendicular to the direction of the incoming flow is more than ten times the characteristic size of the buoy, and the height of the computational domain is consistent with the operating water depth. The buoy is located at a distance downstream of the upstream inlet along the transverse centerline of the computational domain.

[0044] In the numerical computation domain, an overlapping mesh technique is used to divide the buoy motion subdomain and the background flow field domain. This technique decomposes the entire computational domain into two independent mesh systems. The first mesh system, representing the background flow field, covers most of the numerical computation domain. It uses a structured hexahedral mesh with mesh lines arranged along the direction of the incoming flow and perpendicular to the direction of the current, facilitating the application of boundary conditions and improving solution accuracy. The second mesh system, representing the buoy motion subdomain, covers only the buoy and its surrounding space. Mesh cells are closely fitted to the underwater structural surface of the buoy, and mesh refinement is applied to the near-wall region of the buoy to capture velocity gradient changes within the boundary layer. The two mesh systems overlap in the space surrounding the buoy, but the mesh nodes in the overlapping region are not completely coincident. Data is transferred between the two mesh systems in the overlapping region using an interpolation algorithm. When simulating the heave and pitch motion of the buoy under wave action, the buoy motion subdomain mesh moves as a whole with the rigid body motion of the buoy, while the background flow field mesh remains stationary. The relative positions of the two mesh systems change over time, and the data interpolation between the mesh nodes in the overlapping region is updated accordingly. This approach avoids the regeneration of the entire mesh due to buoy movement, significantly improving the computational efficiency of dynamic boundary flow field simulation.

[0045] A turbulence model is used to solve the flow field in the buoy motion subdomain and the background flow field. The unsteady Reynolds-averaged Navier-Stokes equations introduce a Reynolds stress term, which originates from the nonlinear contribution of fluctuating velocity-related terms during the time-averaging process of the Navier-Stokes equations and needs to be closed using a turbulence model. The role of the turbulence model is to establish the constitutive relationship between Reynolds stress and the mean flow field velocity gradient, making the governing equations closed and solvable. The turbulence model adopts the eddy viscosity assumption, expressing Reynolds stress as the product of the turbulent viscosity coefficient and the mean velocity gradient. The turbulent viscosity coefficient is obtained by solving the additional turbulent variable transport equations. Considering the turbulence characteristics of streamline curvature, flow separation, and fully developed wake region in the buoy flow problem, a turbulence model variant with good predictive ability for swirling flow and streamline curvature is selected. The flow field solution uses the finite volume method to discretize the governing equations. The pressure-velocity coupling is solved iteratively using a semi-implicit method for the pressure coupling equations, and a second-order implicit scheme is used for time propagation to ensure the numerical stability of the unsteady calculation.

[0046] Multiple sets of combined conditions for incoming current velocity, direction, water depth, and waves are set, and corresponding boundary conditions are applied to the numerical computation domain. The range of incoming current velocity values ​​is determined based on the current velocity characteristics of the sea area where the observation buoy is deployed, covering multiple operating conditions from weak to strong currents. The current direction covers an all-around angular range, taking 12 directions in 30-degree increments. The water depth is taken from several representative water depth values ​​that may occur in the sea area where the observation buoy is deployed. The waves are taken from typical wave conditions occurring in the sea area where the observation buoy is deployed. The above four sets of parameters—incoming current velocity, direction, water depth, and waves—are orthogonally combined to obtain an operating condition matrix covering a wide range of marine environmental conditions.

[0047] For each operating condition, corresponding boundary conditions are applied to the numerical computational domain boundary. The upstream end face of the computational domain along the incoming flow direction is set as the velocity inlet boundary, where the magnitude and direction of the incoming flow velocity are specified. The vertical velocity distribution along the water depth adopts a logarithmic distribution to simulate the velocity profile of the natural near-bottom boundary layer. Simultaneously, the inlet wave generation method is used to simulate wave propagation at the inlet boundary, generating Stokes waves or stream function waves by giving wave height, period, and water depth parameters. The downstream end face of the computational domain along the incoming flow direction is set as a zero-gradient free outflow boundary, where the normal gradient of the fluid and turbulence variables is zero, allowing the flow field structure to flow freely out of the computational domain without generating non-physical reflections. The two lateral end faces perpendicular to the incoming flow direction are set as symmetrical boundaries. Symmetrical boundaries mean that the normal velocity component and the normal gradient of the tangential velocity component are zero on the boundary, used to simulate the flow state in open water without lateral wall constraints.

[0048] Different boundary conditions are set at the top of the computational domain depending on whether the free water surface is considered. When simulating free water surface and wave propagation, the fluid volume method is used to capture the water-air interface, and the top boundary is set as a pressure outlet boundary with atmospheric pressure. The bottom of the computational domain is set as a no-slip wall boundary to simulate the obstruction of water flow by the seabed, and the near-wall velocity is handled using the wall function method. In the case with buoys, the surface of the underwater structure of the buoy is also set as a no-slip wall boundary.

[0049] Simultaneous simulation of the three-dimensional flow field under the same boundary conditions, with and without buoys. For each combination of parameters including inflow velocity, direction, water depth, and waves, two computational fluid dynamics simulations are performed sequentially.

[0050] The first calculation includes a 3D geometric model of the buoy in the numerical computation domain, simulating the flow field state when the buoy's deployment disturbs the flow field; this is the buoy-containing case. The second calculation, under the exact same numerical computation domain, mesh system, and boundary conditions, removes the 3D geometric model of the buoy from the computation domain, replacing the removed mesh region with a water region, and solves for the natural flow field state undisturbed by the buoy structure; this is the buoy-free case. The two simulations are identical except for the presence or absence of the buoy geometric model, ensuring that the difference in the simulation results stems solely from the disturbance effect of the buoy structure on the flow field. Simulations with and without buoy conditions are paired for each set of boundary conditions, forming a complete set of simulation examples. This paired simulation process is executed iteratively for all parameter combinations in the case matrix, accumulating a corresponding order of magnitude of simulation example sets.

[0051] Buoy attitude data and a vertical velocity profile directly below the buoy are extracted from the simulation results of the buoy-operated condition. The buoy attitude data includes pitch, roll, and yaw angles, which are expanded vertically to the same number as the depth cells. The velocity profile is extracted on a vertical line directly below the geometric center of the buoy's vertical projection, corresponding to the vertically downward direction of the acoustic Doppler current profiler transducer beam. The three-dimensional velocity components are sampled along the vertical line from near the free surface to the bottom of the computational domain at a fixed spatial step size. Computational fluid dynamics simulations yield a velocity distribution over continuous space, with a spatial sampling step size much smaller than the actual depth cell size of the acoustic Doppler current profiler. To match the spatial resolution of the simulation data with the measured data, the simulated velocity profile needs to be spatially integrated and averaged according to the depth cell size of the acoustic Doppler current profiler. Spatial integral averaging involves assigning simulated sampling points to various depth cells of the acoustic Doppler current profiler according to their depth location. The velocity values ​​of several sampling points within each depth cell are then taken as an arithmetic mean, and this average value is used as the representative velocity for that depth cell. After spatial integral averaging, the number of depth cells in the simulated velocity profile under buoy conditions is the same as the number of depth cells within the effective measurement depth range of the acoustic Doppler current profiler. Each depth cell corresponds to a set of three-dimensional velocity components.

[0052] The simulation uses velocity profiles superimposed with buoy attitude data and blind zone masks as input features, and velocity profiles without buoys as labels. Vertical velocity profiles at the same plumb line position are extracted from the simulation results without buoys and spatially integrated at the same depth to obtain the velocity profile without buoys. This profile represents the natural vertical velocity distribution without buoy interference, i.e., the background velocity profile. For the velocity profile with buoys, after spatial integration and averaging, a corresponding blind zone mask is superimposed according to the aforementioned dynamic blind zone mask generation method. The blind zone mask obscures the velocity components of the corresponding depth units, simulating the state of missing near-surface data in actual ocean observations. The velocity profile with buoys, after superimposing the blind zone mask and buoy attitude data, serves as the input feature for training samples, and the velocity profile without buoys under the same boundary conditions serves as the label for the training samples. Each simulation case group generates a mapped sample pair, and the mapped sample pairs generated by all simulation case groups together constitute the flow field sample library.

[0053] The number of samples in the flow field sample library depends on the number of combinations of four parameters: inflow velocity, flow direction, waves, and water depth. The wider the operating conditions covered by the sample library, the richer the flow field states the deep learning network encounters during training, and the stronger its generalization ability to unseen operating conditions after training. The flow field sample library is divided into a training set, a validation set, and a test set according to a preset ratio, and these three subsets are mutually exclusive. The training set is used for iterative updates of the deep learning network parameters, the validation set is used to monitor the model's convergence status and adjust hyperparameters during training, and the test set is used to independently evaluate the model's performance after training.

[0054] In one or more embodiments of the present invention, the deep learning network includes a feature embedding layer, a feature extraction layer, and a physical information decoding layer; The feature embedding layer performs a linear projection on the vertical feature tensor, maps each depth unit to a unified dimension token, and superimposes the depth position encoding. The feature extraction layer extracts flow field perturbation features from the tokens; The physical information decoding layer maps the flow field disturbance characteristics to the background velocity profile of the full depth layer.

[0055] Deep learning networks are the core component of the velocity profile reconstruction model. Their role is to map the input vertical feature tensor into a continuous background velocity profile across the entire depth layer. The vertical feature tensor simultaneously contains velocity information affected by buoy hydrodynamic disturbances, buoy attitude data, and blind zone mask information identifying the locations of invalid depth cells near the surface. The deep learning network needs to extract flow field disturbance features from this input and, based on these features, perform hydrodynamic disturbance correction and near-surface blind zone velocity completion. To accomplish these tasks, the deep learning network adopts a three-layer sequential architecture: a feature embedding layer, a feature extraction layer, and a physical information decoding layer.

[0056] The feature embedding layer receives a vertical feature tensor as input. The vertical feature tensor is a two-dimensional data structure, with one dimension corresponding to the depth direction and the other dimension corresponding to the feature channels of each depth unit. The feature vector of each depth unit contains multiple elements: the eastward flow velocity component, the northward flow velocity component, the vertical flow velocity component, the pitch angle, the roll angle, the rotation angle, and the label value of the dynamic blind zone mask for that depth unit. The depth units are arranged vertically from the near-surface layer to the deeper layers.

[0057] The feature embedding layer first performs a linear projection on the vertical feature tensor. The principle behind linear projection is to map the original feature vector of each deep unit to a higher-dimensional feature space using a trainable linear transformation matrix. The number of rows in this linear transformation matrix equals the dimension of the mapped feature vector, and the number of columns equals the number of elements in the original feature vector. After linear projection, each deep unit is mapped to a vector of a uniform dimension, which is called a token in deep learning. The dimension of the token is a fixed value throughout the network, for example, it can be set to 64 or 128 dimensions. The choice of dimension comprehensively considers the balance between model expressive power and computational efficiency. Each deep unit generates a token, and the token sequence corresponding to the total number of deep units is arranged along the depth direction.

[0058] After linear projection, the feature embedding layer superimposes depth position encoding onto each token. The purpose of depth position encoding is to provide the deep learning network with the absolute positional information and relative positional relationship of each token in the depth direction. Without depth position encoding, the network treats each depth unit as a set of position-independent features when processing token sequences, failing to distinguish the physical differences between shallow and deep depth units. Depth position encoding generates a unique position encoding vector for each depth position, adding this vector element-wise to the token of the corresponding depth unit, allowing the token to carry information about the depth of that depth unit throughout the entire profile. The dimension of the depth position encoding vector is the same as the dimension of the token. Depth position encoding can be generated by combining sine and cosine functions; different depth positions correspond to different frequencies of sine and cosine function values. This encoding method ensures that position encoding vectors of adjacent depths have high similarity in vector space, while position encoding vectors of distant depths differ significantly, enabling the network to learn the continuity and local correlation of flow velocity in the vertical direction. Another approach is to set up trainable deep positional encoding vectors, with each deep position corresponding to a set of trainable parameters. During training, the optimal positional encoding representation is automatically learned through backpropagation. The token sequence resulting from the superposition of deep positional encodings is the output of the feature embedding layer.

[0059] The feature extraction layer receives the position-encoded token sequence output from the feature embedding layer as input. The design goal of the feature extraction layer is to extract flow field disturbance features related to buoy hydrodynamic disturbances and near-surface velocity distribution from the token sequence. These flow field disturbance features contain two aspects of information: firstly, the intensity differences and spatial distribution patterns of the influence of the buoy's underwater structure on the velocity at different depth units; and secondly, the vertical velocity profile variation trend implied by the velocity distribution near the near-surface blind zone. This trend information is crucial for subsequent completion of the near-surface blind zone velocity data.

[0060] The feature extraction layer extracts the aforementioned flow field perturbation features by stacking multiple layers of self-attention modules. The core operation of the self-attention module is to calculate the attention weights between depth units in the token sequence. These attention weights reflect the strength of the correlation between the velocity characteristics of two depth units under the influence of buoy perturbation. For hydrodynamic perturbation problems, the buoy wake region typically has a spatial scale spanning multiple depth units in the vertical direction. The perturbation intensity at a single depth unit is related to the flow state of several adjacent depth units; therefore, it is necessary to capture long-range dependencies in the depth direction. The self-attention module captures these long-range dependencies through global attention calculation.

[0061] In the self-attention module, each token generates a query vector, key vector, and value vector through three different sets of linear projections. The query vector of a depth unit is inner-producted with the key vectors of all depth units. The inner-product result is normalized to form an attention weight distribution; a higher weight value indicates a greater contribution of the corresponding depth unit to the feature representation of that depth unit. The value vectors of all depth units are weighted and summed according to their attention weights to obtain the output representation of that depth unit after self-attention update. Stacking multiple layers of self-attention modules enables the network to progressively extract multi-scale flow field perturbation features from local to global perspectives. Lower-layer modules capture local velocity variation features between adjacent depth units, while higher-layer modules capture macroscopic perturbation patterns across the entire profile scale.

[0062] Feedforward neural network layers are inserted between the multi-layer self-attention modules. These layers act independently on each deep unit, enhancing the network's non-linear expressive power through non-linear activation functions. Residual connections and layer normalization operations are also implemented between layers. Residual connections sum the input and output of the current layer before feeding them into the next layer, avoiding the vanishing gradient problem during deep network training. Layer normalization normalizes the feature vector of each deep unit to a distribution with zero mean and one variance, accelerating network training convergence and improving numerical stability. The output of the feature extraction layer is a sequence of tokens with the same number and dimensions as the input, where each token incorporates the flow field perturbation feature information of that deep unit within the entire vertical profile context.

[0063] The Physical Information Decoding (PID) layer receives the high-level feature token sequence output by the Feature Extraction layer as input. The task of the PID layer is to map the high-level features of each depth unit to flow velocity values ​​in physical space. Unlike the Feature Extraction layer, the PID layer does not require spatial information interaction between depth units because the flow field disturbance features have been fully extracted and fused in the Feature Extraction layer. This layer only needs to perform a regression mapping from high-level feature vectors to flow velocity values ​​on a per-depth-unit basis.

[0064] The physical information decoding layer consists of multiple fully connected layers stacked sequentially. Each fully connected layer maps the output vector of the previous layer to the input vector of the next layer through a linear transformation and a nonlinear activation function. The number of parameters in the fully connected layer is independent of the number of depth units; the same set of parameters is applied to all depth units. This parameter-sharing mechanism ensures the consistency of velocity profile predictions across all depth layers. The output dimension of the last fully connected layer corresponds to the number of output components in the baseline velocity profile, outputting the corrected and completed eastward and northward velocities for each depth unit.

[0065] The velocity profile output by the physical information decoding layer contains velocity values ​​in both the effective measurement depth range and the near-surface blind zone. The velocity values ​​in the effective measurement depth range are baseline velocity values ​​obtained by eliminating the hydrodynamic disturbance increments caused by buoys in the input disturbance velocity profile. The velocity values ​​in the near-surface blind zone are baseline velocity values ​​obtained by extrapolating and reconstructing the network based on the inherent laws governing the distribution of effective velocity and vertical velocity at the lower boundary of the blind zone. The two velocity values ​​are naturally connected in the depth direction, forming a continuous baseline velocity profile across the entire depth layer from near the free surface to the maximum effective measurement depth.

[0066] The forward computation flow of the entire deep learning network is as follows: Vertical feature tensors are input into the feature embedding layer, and after linear projection and depth position encoding, a token sequence is generated. This token sequence is then fed into the feature extraction layer, where multiple self-attention modules and feedforward neural network layers extract flow field perturbation features. The output token sequence from the feature extraction layer is then fed into the physical information decoding layer, where multiple fully connected layers map high-level features to flow velocity values ​​in physical space unit by unit, ultimately outputting a continuous baseline flow velocity profile across the entire depth layer. All trainable parameters of the network are updated via backpropagation during offline training. After training, the network parameters are fixed, and only forward computation is performed during the online inference phase.

[0067] In one or more embodiments of the present invention, the total loss function during deep learning network training is composed of the blind zone prediction loss, the effective observation layer consistency loss, and the vertical profile smoothing loss summed according to weights. Blind zone prediction loss measures the deviation between the predicted velocity of the blind zone depth cell and the computational fluid dynamics label; Effective observation layer consistency loss measures the deviation between predicted and measured flow velocities in effective observation depth cells; Vertical profile smoothing loss measures the degree of change in predicted flow velocity between adjacent depth cells.

[0068] The total loss function expression used in the velocity profile reconstruction model is: L=L blind +αL obs +βL smooth Among them, L blind The near-surface blind zone prediction loss is used to constrain the model to approximate the computational fluid dynamics label corrected by the acoustic Doppler velocity profiler in the missing measurement region; L obs Ls represents the consistency loss of the effective observation layer, used to constrain the model from deviating from the measured flow velocity within the effective observation area of ​​the acoustic Doppler current profiler; moothThe vertical profile smoothing loss is used to suppress unreasonable local oscillations in the predicted profile. α and β are weighting coefficients, both positive, used to adjust the influence of the effective observation layer consistency constraint and the vertical smoothing constraint on the total loss function. Their values ​​are determined based on the training sample size, flow field variation characteristics, and model convergence. Preferably, α is between 0.1 and 1, and β is between 0.01 and 0.5. When high near-blind zone reconstruction accuracy is required, the value of α is increased to enhance the observation consistency constraint; when local oscillations exist in the velocity profile, the value of β is increased to enhance the profile smoothing constraint.

[0069] In the offline training phase, deep learning networks require a loss function to measure the difference between the network's predicted output and the expected result. The magnitude of the loss function reflects the quality of the velocity profile reconstruction under the current parameter state of the network. The training process involves continuously adjusting all trainable parameters of the network through optimization algorithms, gradually reducing the value of the loss function until the deviation between the network output and the expected result converges to an acceptable range. Since the velocity profile reconstruction task involves two sub-tasks simultaneously: correction of hydrodynamic disturbances within the effective measurement depth range and velocity completion in the near-surface blind zone, and these two sub-tasks differ in terms of physical constraints and data availability, it is difficult to simultaneously address the optimization needs of both sub-tasks using a single loss function. Therefore, a total loss function is designed, consisting of the weighted sum of three sub-loss functions: blind zone prediction loss, effective observation layer consistency loss, and vertical profile smoothing loss.

[0070] Blind zone prediction loss measures the deviation between the velocity prediction by the deep learning network and the computational fluid dynamics (CFD) simulation label within the near-surface blind zone. In actual ocean observations, ground-value velocity data is lacking in the near-surface blind zone, making it impossible to obtain supervisory signals from measured data. However, during offline training, CFD simulations can simultaneously provide paired velocity profiles for both buoy-equipped and buoy-free scenarios. The velocity profile for the buoy-free scenario has actual velocity values ​​within the corresponding depth range of the blind zone; these values ​​serve as the labels for the supervised deep learning network's blind zone prediction. The calculation of blind zone prediction loss is limited to depth cells marked as invalid by the dynamic blind zone mask; valid measured depth cells are not included in this loss calculation. The calculation method compares the predicted velocity value of each depth cell within the blind zone with the velocity value of the depth cell corresponding to the CFD simulation label. The smaller the blind zone prediction loss value, the closer the deep learning network's completion result for the near-surface blind zone velocity is to the actual velocity given by the flow field physical simulation. The physical significance of this loss is that it forces deep learning networks to extrapolate the flow velocity in the blind zone based on the physical laws of the flow field, rather than relying solely on statistical correlation of data for interpolation.

[0071] The effective observation layer consistency loss measures the deviation between the current velocity predicted by the deep learning network and the actual current velocity measured by the acoustic Doppler current profiler within the effective measurement depth range. Within this range, the acoustic Doppler current profiler provides velocity values ​​affected by buoy hydrodynamic disturbances. While these values ​​contain disturbance components, they still represent accurate measurements under the current sea conditions. The deep learning network's task is to correct these velocity values ​​for hydrodynamic disturbances and output the background velocity values ​​for the corresponding depth cells.

[0072] However, if the network is trained using only the blind zone prediction loss, it may over-correct the velocity values ​​within the effective measurement depth range, causing the corrected velocity to deviate too much from the measured velocity, thus losing fidelity to the measured data. The effective observation layer consistency loss suppresses this tendency by constraining the network's output in the effective observation region to not deviate from the measured velocity. It quantifies the correction magnitude by comparing the predicted velocity with the measured velocity at the effective measurement depth unit. The calculation of this loss is limited to depth units marked as effective by the dynamic blind zone mask. The value of the effective observation layer consistency loss is maintained within a moderate range, allowing the deep learning network to make necessary corrections to hydrodynamic disturbances while avoiding excessive deviation from the measured values. This loss, together with the blind zone prediction loss, forms a differentiated constraint on the deep learning network in two completely different regions, enabling the network to rely on computational fluid dynamics physical labels for completion in the blind zone and to follow the constraints of measured data for correction in the effective observation region.

[0073] Vertical profile smoothing loss measures the abruptness of predicted velocity changes between adjacent depth cells. In natural flow fields, the velocity distribution in the vertical direction exhibits physical continuity; velocity changes between adjacent depth cells are constrained by fluid motion equations and turbulent mixing processes, typically without isolated abrupt changes. Deep learning networks, in the early stages of training or with insufficient data coverage, may introduce unreasonable local oscillations into the predicted velocity profile. This manifests as a significant deviation between the predicted velocity of a certain depth cell and its adjacent depth cells, forming a jagged profile shape. Such local oscillations lack physical basis and reduce the reliability of the reconstructed profile. Vertical profile smoothing loss suppresses these oscillations by penalizing the difference in predicted velocities between adjacent depth cells. The smaller the value of this loss, the smoother the vertical change in the predicted velocity profile. The constraint strength applied by the smoothing loss is controlled by a weighting coefficient. An excessively large weighting coefficient leads to overly smoothed predicted velocity profiles, masking velocity shear and stratification structures present in the actual flow field; an excessively small weighting coefficient fails to effectively suppress oscillations. The three sub-loss functions are summed by weighted summation using corresponding weighting coefficients, which are used to adjust the relative importance of each loss in the total loss.

[0074] During training, the total loss function is calculated by forward propagation of the deep learning network to obtain the predicted velocity profile, which is then compared with computational fluid dynamics labels and measured velocities. After obtaining the total loss value, the gradient of the total loss with respect to each trainable parameter in the network is calculated using the backpropagation algorithm, and then the network parameters are updated using a gradient-based optimization algorithm. After multiple iterations with a large number of samples in the training set, the total loss function gradually decreases and tends to stabilize. The deep learning network gradually learns the mapping relationship from the perturbed velocity profile with superimposed blind zone masks to the continuous baseline velocity profile across the entire depth layer.

[0075] In one or more embodiments of the present invention, the feature extraction layer employs a shift-window multi-head self-attention mechanism to extract flow field disturbance features, including: The token sequence is divided into several non-overlapping local windows along the depth direction, and each local window contains a preset number of depth units; Within each local window, multi-head self-attention is calculated on the tokens within the window to extract the flow field feature correlation between depth cells within the window; The local window is shifted along the depth direction by a preset step size, and multi-head self-attention is calculated in the new window obtained by the shift, so as to realize the interaction of feature information between adjacent local windows; The self-attention within the window and the self-attention of the shifted window are stacked alternately, and the output is a token sequence containing multi-scale flow field perturbation features.

[0076] The feature extraction layer employs a shift-window multi-head self-attention mechanism to extract flow field perturbation features. The velocity profile along the depth direction is a one-dimensional sequential data, with each element in the sequence corresponding to a feature token for a depth unit. The buoy hydrodynamic perturbation has a continuous spatial scale spanning multiple depth units in the vertical direction. The perturbation mechanism at a single depth unit is correlated with the flow field state of its neighboring depth units and more distant depth units. Therefore, feature extraction needs to simultaneously capture fine-grained perturbation features within a local range and long-range dependencies across the global scope.

[0077] While traditional self-attention mechanisms can capture the correlation between any two depth cells through global attention computation, their computational complexity increases quadratically with sequence length. For profiles with a large number of depth cells, the computational and storage overhead of global self-attention is significant. Furthermore, the physical correlation in the vertical direction of velocity profiles exhibits local dominance; a perturbation in one depth cell has the most significant impact on its immediate neighboring depth cells, while its impact on more distant depth cells weakens with increasing distance. The shifted-window multi-head self-attention mechanism, by restricting global attention computation to a local window and achieving information exchange between adjacent windows through window shifting, retains the ability to model long-range dependencies while controlling computational complexity.

[0078] The specific operation process of the shift window multi-head self-attention mechanism is as follows.

[0079] The first step is to divide the token sequence output from the feature embedding layer into multiple non-overlapping local windows along the depth direction. The division method is to start from the first depth unit along the depth direction, with each window consisting of several consecutive depth units, and the windows are connected end-to-end but do not overlap. Each local window contains a preset number of depth unit tokens, and the window size is determined based on the total number of depth units in the profile and the typical scale of the flow velocity perturbation in the vertical direction. If the total number of depth units is not exactly divisible by the window size, zeros are padded at the end of the token sequence to ensure that all windows have a consistent size.

[0080] The second step is to compute multi-head self-attention within each local window. Multi-head self-attention is an extension of the self-attention mechanism, mapping the query vector, key vector, and value vector to multiple different subspaces through multiple independent linear projections. Self-attention is computed separately in each subspace, and finally, the outputs of each subspace are concatenated and fused using linear projection. The multi-head design allows deep learning networks to focus on different types of vertical feature association patterns in different subspaces, improving the diversity and richness of feature extraction. Computing self-attention within a window means that for each deep unit token within the window, only its attention weights with other deep unit tokens within the same window are calculated, without considering tokens outside the window. This constraint reduces the computational cost of attention from being proportional to the square of the total sequence length to being proportional to the square of the window size, and multiplying by the number of windows results in only linear complexity. After completing the multi-head self-attention computation within each window, the tokens within each window receive an updated representation that incorporates the contextual information within that window.

[0081] The third step is to shift the local window along the depth direction by a preset step size, and then calculate multi-head self-attention within the new shifted window. While intra-window self-attention captures the local relationships between depth units within the window, the lack of direct information exchange channels between depth units on either side of the boundary of adjacent windows due to the non-overlapping nature of the window partitioning process. Window shifting addresses this by offsetting the starting position of the window partitioning along the depth direction by a preset step size. The shifted window covers the boundary region between the original two adjacent windows, allowing depth units that originally belonged to two separate windows to enter the same new window. Multi-head self-attention is then calculated again within the new shifted window, establishing direct attention connections between depth units in these boundary regions, thus enabling cross-window information transfer. The shift step size is typically set to half the window size, ensuring that the shifted window exactly covers the boundary region between the original windows.

[0082] The feature extraction layer is constructed by alternating stacks of in-window and shift-window self-attention modules. An alternating computational unit is formed by placing one in-window self-attention module immediately after another, and multiple alternating computational units are stacked sequentially. This alternating stacking method allows the receptive field of each depth unit token to expand layer by layer as the network depth increases, gradually expanding from the initial local range within the window to a global range covering the entire depth profile, ultimately outputting a token sequence containing multi-scale flow field perturbation features. Lower-level modules capture small-scale local perturbation features, while higher-level modules capture large-scale macroscopic perturbation patterns and the overall trend of vertical velocity distribution.

[0083] Furthermore, a relative position encoding mechanism can be introduced within each self-attention module to consider the relative depth distance between two depth units when calculating attention weights. Depth units with closer relative depth distances receive higher attention weight biases, while those with farther distances receive lower biases. This bias aligns with the law that the physical correlation of flow velocity in the vertical direction decays with distance, which helps to further improve the physical rationality of feature extraction.

[0084] In one or more embodiments of the present invention, the velocity profile reconstruction model outputs a continuous background velocity profile across the entire depth layer in the following manner: The measured flow velocity within the effective measurement depth range is processed by the flow profile reconstruction model to eliminate the hydrodynamic disturbance increment caused by the underwater structure of the observation buoy to the surrounding flow field, and the corrected flow velocity value of the lower boundary depth unit of the near-surface blind zone is obtained. Using the corrected flow velocity value of the lower boundary depth unit of the near-surface blind zone as the boundary condition for extrapolation, and combining the vertical distribution law of flow velocity learned by the deep learning network from the computational fluid dynamics simulation samples during the training process, the flow velocity value of each depth unit of the near-surface blind zone is extrapolated layer by layer along the vertical direction to complete the vertical upward extrapolation reconstruction. Among them, the velocity profile after correction of the effective measurement depth range and the velocity profile obtained by extrapolation and reconstruction of the near-surface blind zone together constitute the continuous background velocity profile of the full depth layer, which is uniformly output by the velocity profile reconstruction model.

[0085] After receiving the vertical feature tensor, the velocity profile reconstruction model performs an end-to-end mapping from a perturbed and incomplete velocity profile to a complete background velocity profile within a deep learning network. This mapping process can be physically decomposed into two interconnected subtasks: correcting for hydrodynamic perturbations within the effective measurement depth range, and reconstructing the velocity profile vertically by extrapolation in the near-surface blind zone. These two subtasks share the flow field perturbation features extracted by the deep learning network and are collaboratively completed within a unified computational framework.

[0086] First, the deep learning network processes the measured flow velocity within the effective measurement depth range to eliminate the hydrodynamic disturbance increment caused by the underwater structure of the observation buoy to the surrounding flow field. The effective measurement depth range refers to the depth interval from the first effective depth cell of the acoustic Doppler current profiler to the maximum effective measurement depth. Although the flow velocity data within this interval has been compensated for by rigid body motion to remove the overall motion component of the buoy, it still contains the flow field distortion effect caused by the underwater structure of the buoy. The obstruction effect of the underwater structure of the buoy on the incoming flow causes the water flow in front of the structure to slow down and generate upward flow around the structure, forming a velocity-deficient wake region behind the structure, and the water flow on both sides of the structure to accelerate around the structure. These flow field distortions are manifested as a systematic deviation in the magnitude and direction of the flow velocity within a specific depth range on the vertical velocity profile directly below the buoy. This deviation is the hydrodynamic disturbance increment. The spatial distribution characteristics of the hydrodynamic disturbance increment depend on factors such as the geometry of the underwater structure of the buoy, the incoming flow velocity, the incoming flow direction, and the water depth. The disturbance amount is different in each depth cell, forming a continuous variation curve in the vertical direction.

[0087] During offline training, the deep learning network has learned the vertical distribution patterns of hydrodynamic disturbance increments under different inflow conditions from a large number of computational fluid dynamics simulation samples. The input features in the training samples are disturbance profiles and buoy attitude data under buoy-containing conditions, labeled as background velocity profiles under buoy-free conditions. The network learns the transformation relationship from the disturbance profile to the background profile by minimizing the loss function. In this learning process, the network implicitly establishes functional mappings between disturbance increments and parameters such as depth, velocity, and buoy attitude under several typical inflow conditions. During online inference, after receiving the measured disturbance velocity profile, the network combines the velocity distribution characteristics of the current profile, buoy attitude, the effective observation range indicated by the dynamic blind zone mask, and the disturbance pattern knowledge implicit in the training data to correct the velocity value of each depth unit within the effective measurement depth range, thus obtaining the corresponding background velocity value. After correction, the velocity value at the lower boundary depth unit of the near-surface blind zone changes from the measured value containing disturbances to the background velocity value reflecting the natural inflow state. This velocity value is the corrected velocity value of the lower boundary depth unit of the near-surface blind zone.

[0088] Next, the deep learning network uses the corrected velocity value of the lower boundary depth unit of the near-surface blind zone as the starting boundary condition for vertical upward extrapolation reconstruction. The near-surface blind zone is the depth interval below the free water surface and above the effective measurement depth range, where measured velocity data cannot be obtained. In the natural flow field, the velocity distribution within this interval is not arbitrary but constrained by the fluid motion equations and boundary conditions. The boundary condition at the free water surface is the surface shear stress condition; under wind-driven flow conditions, the surface shear stress is given by the wind stress, and under windless conditions, the surface shear stress is zero. The boundary condition at the lower boundary of the blind zone is the corrected velocity value. Between the two boundary conditions, the vertical distribution of water velocity is controlled by turbulent mixing and momentum transfer processes, forming a velocity profile with a certain curvature.

[0089] Deep learning networks learn the intrinsic laws governing the vertical velocity distribution from computational fluid dynamics (CFD) simulation samples during training. CFD simulations, by solving unsteady Reynolds-averaged Navier-Stokes equations, fully simulate the flow field evolution from the free surface to near the bottom layer. The vertical velocity distribution within the depth range corresponding to the blind zone in the simulation results represents the numerical solution of the fundamental fluid dynamics equations under given boundary conditions, embedding physical mechanisms including turbulent diffusion, surface shear stress constraints, and convection term balance. The network learns not a single, fixed profile shape during training with a large number of samples, but rather the mapping relationship between the vertical velocity distribution and factors such as inflow conditions, water depth, and buoy attitude. This mapping relationship reflects the physical laws governing fluid motion.

[0090] During online inference, the network uses the corrected velocity values ​​of the lower boundary depth cells of the near-surface blind zone as the starting boundary condition for extrapolation. Combined with the learned vertical velocity distribution, it extrapolates the velocity values ​​of each depth cell in the near-surface blind zone layer by layer upwards. The extrapolation direction is from the lower boundary of the blind zone upwards to the free surface, and the extrapolation step size corresponds to the depth cell size of the acoustic Doppler current profiler. In each extrapolation step, the velocity values ​​already extrapolated in the next layer provide a reference for the extrapolation of the current layer's velocity values. Simultaneously, the network's internal feature representation encodes the overall flow field perturbation characteristics and vertical distribution trend information of the current profile, providing constraints for the layer-by-layer extrapolation. After extrapolating to the uppermost depth cell, the velocity values ​​of all depth cells in the near-surface blind zone are obtained, completing the vertical upward extrapolation reconstruction.

[0091] The corrected velocity profile within the effective measurement depth range and the velocity profile reconstructed by extrapolation from the near-surface blind zone naturally connect in the depth direction. The shallowest depth unit of the velocity profile within the effective measurement depth range is adjacent to the deepest depth unit of the near-surface blind zone, and the velocity values ​​at this connection point are kept consistent by the same corrected velocity value. Together, the two velocity profiles constitute a continuous background velocity profile across the entire depth layer from near the free surface to the maximum effective measurement depth, which is output as a whole by the velocity profile reconstruction model. This background velocity profile reflects the natural vertical distribution of the water velocity at this location assuming the absence of the observation buoy, and can be used in applications requiring real flow field data, such as port and waterway flow velocity monitoring, marine engineering environmental assessment, and underwater vehicle path planning.

[0092] In one or more embodiments of the present invention, the real-time vertical distance from the transducer end face to the free water surface is calculated as follows: The depth is calculated by extracting real-time pressure from the data stream. The acoustic Doppler current profiler incorporates a pressure sensor, which is integrated with the transducer and mounted on the underwater portion of the observation buoy. The pressure sensor contains a pressure-sensitive element; one side of the element contacts the water, bearing the hydrostatic pressure at the transducer's location, while the other side contacts an internal reference pressure chamber. The hydrostatic pressure causes a slight deformation on the surface of the pressure-sensitive element, which converts this deformation into an electrical signal output. Since the pressure sensor is located at the depth of the transducer's end face, the hydrostatic pressure measured by the sensor directly corresponds to the vertical water column height from the transducer's end face to the free water surface. To calculate the water depth at the transducer's location, the pressure value output by the pressure sensor is divided by the product of the water density and gravitational acceleration, yielding a depth value in meters. Water density varies with temperature, salinity, and pressure; typically, the standard seawater density value is used in open sea areas, while in estuaries or nearshore areas, the density value is determined based on on-site hydrological survey data. The acceleration due to gravity varies with latitude; when the latitude is known, the corresponding standard value is taken. The water depth where the transducer is located, obtained through the above conversion, is the vertical distance from the transducer end face to the free water surface.

[0093] It should be noted that under wave conditions, the position of the free water surface changes periodically over time, the observation buoy moves with the waves, and the water depth where the transducer is located also changes dynamically. The real-time pressure depth reflects the instantaneous height of the water column above the instantaneous position of the pressure sensor. Therefore, the water depth where the transducer is located, calculated based on the real-time pressure depth, directly corresponds to the vertical distance from the transducer end face to the instantaneous free water surface at the current measurement moment, without the need for separate correction for the water surface fluctuations caused by waves.

[0094] Furthermore, the blind zone distance is perpendicular to the transducer end face. When the observation buoy tilts, the depth affected by the blind zone distance has a geometric projection relationship with the blind zone distance itself. When the buoy tilt angle is small, the difference between the two is negligible. When the buoy tilt angle is large, a tilt angle cosine correction can be further introduced.

[0095] In one or more embodiments of the present invention, the depth of influence of the blind zone distance is the projected distance of the blind zone distance in the vertical direction, and the calculation method is as follows:

[0096] In the formula, Let be the vertical projection of the blind zone distance at time t. Configure blind spot distance for the equipment. Let t be the instantaneous tilt angle of the buoy, P be the pitch angle, and R be the roll angle.

[0097] In one or more embodiments of the present invention, when simultaneously simulating three-dimensional flow fields under the same boundary conditions with and without buoys, the boundary conditions applied to the numerical computation domain include: The upstream boundary along the direction of the incoming flow is set as the velocity inlet boundary, specifying the inlet flow velocity and direction, and the inlet wave generation method is used to simulate wave propagation; The downstream boundary along the incoming flow direction is set as a zero-gradient free outflow boundary; The two lateral boundaries perpendicular to the direction of the incoming flow are set as symmetrical boundaries; The bottom of the computational domain is set to a non-slip wall boundary; In the case of buoy operation, the surface of the underwater structure of the buoy is set as a non-slip wall boundary.

[0098] Boundary conditions define the physical behavior of the fluid at the boundary of the computational domain and directly determine whether the simulation results can truly reflect the flow field characteristics in the actual marine environment. For each set of parameters including inflow velocity, flow direction, water depth, and waves, identical boundary conditions are applied to two simulations with and without buoys, ensuring that the difference in flow field between the two simulations arises solely from the presence or absence of the buoy geometry model.

[0099] The upstream boundary along the direction of the incoming flow is set as the velocity inlet boundary. The velocity inlet boundary is a boundary condition type used in computational fluid dynamics simulations to specify the magnitude and direction of the velocity of the fluid entering the computational domain from this boundary. On the upstream velocity inlet boundary, a velocity vector is assigned to each boundary grid cell. The magnitude of the velocity vector is taken from the incoming flow velocity value under the current conditions, and the direction of the velocity vector is taken from the direction of the incoming flow under the current conditions. The direction angle is based on true north as zero degrees and increases clockwise. The vertical distribution of the incoming flow velocity along the water depth is not a uniform constant, but rather exhibits a gradient distribution with decreasing velocity at depth near the bottom due to seabed friction. To simulate the distribution characteristics of the natural vertical velocity profile, velocity values ​​at each depth are set according to a logarithmic distribution along the depth direction on the upstream velocity inlet boundary. The velocity value is smaller near the bottom, gradually increasing with decreasing depth from near the bottom upwards, reaching its maximum value near the free surface, forming a common vertical velocity distribution pattern in natural marine environments. The specific parameters of the logarithmic distribution are determined based on measured velocity profile data or empirical parameters from the target sea area.

[0100] Wave propagation is simulated simultaneously at the velocity inlet boundary using the inlet wave generation method. The inlet wave generation method is a numerical method that generates target waves by providing time-varying velocity or wave height boundary conditions at the inlet boundary. Based on the wave height, period, and water depth parameters of the target sea state, the velocity components of water particles caused by the waves are superimposed at the velocity inlet boundary. The wave-induced water particle velocities are distributed along the depth direction according to wave theory. For Stokes waves, the water particle velocity decreases exponentially with depth, being higher near the surface and gradually decreasing towards deeper layers. The direction of the wave-induced water particle velocity components is consistent with the wave propagation direction at wave crests and opposite to the wave propagation direction at wave troughs. Through the synthesis of the wave-induced water particle velocities superimposed at the velocity inlet boundary and the basic incoming flow velocity, the inlet boundary simultaneously completes the input of both steady and unsteady waves. The waves propagate downstream from the inlet boundary into the computational domain, forming a composite flow field in the computational domain that simultaneously includes mean flow and wave motion.

[0101] The downstream boundary along the flow direction is set as a zero-gradient free outflow boundary. A zero-gradient free outflow boundary means that the gradient of the fluid variables in the boundary normal direction is zero; that is, the values ​​of the fluid variables on the boundary are the same as those of the grid cells inside the computational domain immediately adjacent to the boundary. For velocity variables, the normal gradient of the velocity component on the outflow boundary is zero. For turbulent variables, the normal gradients of turbulent kinetic energy and turbulent dissipation rate on the outflow boundary are zero. The zero-gradient outflow boundary allows the flow field structure to flow freely out of the computational domain without generating non-physical reflections or blockage effects at the outflow boundary, making it suitable for situations where the outflow boundary is far from the main flow structure in the flow field. In the computational domain design, the distance between the downstream boundary and the buoy needs to ensure that the wake structure generated by the buoy has fully developed and tends to level off by the time it reaches the downstream boundary, and the gradients of the velocity and turbulent variables along the flow direction have been reduced to a low level.

[0102] Two lateral boundaries perpendicular to the incoming flow direction are set as symmetrical boundaries. A symmetrical boundary is an idealized reflective boundary condition, which physically means that the flow at the boundary is mirror-symmetric about the boundary plane. On a symmetrical boundary, the normal velocity component is zero, the normal gradient of the tangential velocity component is zero, and the normal gradients of other scalar variables are also zero. The zero normal velocity component means that the fluid cannot penetrate the symmetrical boundary plane, and the water flow always flows parallel to the plane containing the symmetrical boundary. The zero normal gradient of the tangential velocity means that there is no fluid shear stress on the symmetrical boundary, and there is no friction between the water flow and the boundary. Setting the two lateral boundaries perpendicular to the incoming flow direction as symmetrical boundaries is used to simulate the flow conditions on both sides of the target observation position in open sea areas where there are no lateral walls constraining the flow. In actual marine environments, there are usually no lateral physical boundaries within several kilometers horizontally of the buoy's location, and the water flow can flow freely in the horizontal direction. Using symmetrical boundary conditions to approximate this open water condition avoids the non-physical sidewall effects that may be introduced by using lateral wall boundaries.

[0103] The bottom of the computational domain is set as a no-slip wall boundary. A no-slip wall boundary means that the fluid velocity close to the wall is the same as the wall's velocity; for a stationary seabed wall, the fluid velocity at the wall is zero. The frictional resistance of the seabed to near-bottom flow is the fundamental reason for the formation of the natural vertical velocity distribution near the bottom gradient. At the no-slip wall boundary, the velocity at the wall is zero, and the velocity gradually increases from the wall towards the interior of the water body, forming a velocity boundary layer. Due to computational resource limitations, the computational mesh is usually not fine enough to directly analyze the velocity distribution of the viscous sublayer within the boundary layer. Therefore, the wall function method is used in conjunction with the no-slip wall boundary. The wall function method establishes the relationship between the velocity at the first mesh node near the wall and the wall shear stress through a semi-empirical formula, avoiding the need for direct mesh analysis of the viscous sublayer. The applicability of the wall function method requires that the first mesh node is located in the logarithmic law region of the turbulent boundary layer; this requirement is met by controlling the height of the first mesh layer at the wall during mesh generation.

[0104] In buoy operation, the underwater structural surfaces of the buoy are designed as no-slip wall boundaries. These surfaces include the buoy's main hull surface, the surface of the acoustic Doppler current profiler transducer mounting base, the surface of the biofouling shield, and the exposed surfaces of other underwater sensors. All these surfaces are solid walls, and the water flow on them satisfies the no-slip condition. At the no-slip wall boundaries of the buoy structure, the fluid velocity close to the wall is zero, and boundary layer flow forms near the wall through fluid viscosity and turbulent diffusion. The surface roughness of the buoy structure is set according to the surface material and the degree of marine biofouling growth; roughness affects the wall friction coefficient and the boundary layer transition position. At the edges and corners of the buoy structure, water flow separation occurs, forming backflow and wake zones downstream. These flow characteristics directly affect the disturbance morphology of the velocity profile directly below the buoy. The no-slip wall boundary design allows these flow characteristics to be fully simulated. In the buoy-free operation, the buoy's three-dimensional geometric model is removed from the computational domain, and the non-slip wall boundary on the buoy's structural surface is correspondingly canceled. This area is replaced by the water body area, and the corresponding grid cells participate in the flow field calculation.

[0105] The above five boundary conditions collectively define the complete boundary condition system of the numerical computation domain. In simultaneous simulations with and without buoys, except for the buoy underwater structure surface boundary condition, which exists in one case as a no-slip wall and has been removed in the other, the other four boundary conditions—upstream velocity inlet boundary, downstream zero-gradient free outflow boundary, laterally symmetric boundary, and bottom no-slip wall boundary—remain completely consistent between the two cases. This setting ensures that there are no flow field differences between the simulation results with and without buoys, except for the buoy structure disturbance effect. Therefore, the difference between the paired velocity profiles extracted from both accurately reflects the contribution of the buoy hydrodynamic disturbance.

[0106] This invention, by organically coupling computational fluid dynamics simulation with a deep learning network, achieves for the first time an integrated process for correcting hydrodynamic disturbances and completing near-surface blind zones in the reconstruction of near-surface velocity profiles of buoys. In the offline phase, computational fluid dynamics is used to perform paired simulations of various operating conditions with and without buoys. The background velocity profile of the buoy-free condition is used as a label to provide training samples with embedded fluid dynamic physical laws for the deep learning network. In the online application phase, only the disturbed velocity profile compensated for rigid body motion, buoy attitude data, and dynamic blind zone mask need to be input into the model. A single forward inference can then output a continuous background velocity profile that compensates for hydrodynamic disturbances and completes the near-surface blind zone. This method fundamentally overcomes the systematic distortion of near-surface velocity profiles caused by the superposition of underwater structural flow field distortion and acoustic top blind zone, significantly improving the integrity and accuracy of the velocity profile within the near-surface range of 0 to 1.5 meters.

[0107] Compared to existing correction methods based on potential flow theory, this invention constructs training samples using high-fidelity flow field data from computational fluid dynamics (CFD) simulations, fully considering turbulence effects in the flow field and avoiding the accuracy limitations of potential flow theory in complex flow field regions such as the buoy wake. Compared to existing purely data-driven deep learning methods, this invention uses the physical results of CFD simulations as supervision signals. The mapping relationships learned by the model embed fluid dynamics physical laws, avoiding the unnatural oscillations and insufficient generalization capabilities that arise from the lack of physical constraints in purely data-driven methods, ensuring physical consistency in the reconstructed profiles. In terms of computational efficiency, this invention limits time-consuming CFD numerical simulations to the offline training stage. Online inference only requires a lightweight deep learning model to complete a single forward computation, with a single profile reconstruction taking less than one second. It can be directly deployed in the buoy's embedded processing unit for in-situ real-time processing, solving the bottleneck of traditional CFD methods failing to meet real-time observation requirements.

[0108] Furthermore, this invention requires no modification to existing buoy hardware; it can be adapted to various deployed observation buoy systems simply by upgrading the software algorithm. The modification cost is low, and it is easy to scale up and promote.

[0109] The above is the overall concept of the present invention. For ease of understanding, the present invention also provides the following embodiments: This embodiment is applied to a multi-functional observation buoy in the open waters of a coastal port. The main body of the observation buoy has a diameter of 1m, a height of 1.658m, a freeboard height of 1.025m, and a draft of 0.695m. The observation buoy is equipped with a 1200kHz downward-looking acoustic Doppler current profiler combined with a high-precision global navigation satellite system and inertial measurement unit for current velocity profile observation.

[0110] The acoustic Doppler current profiler divides the water body into multiple depth cells along the depth direction, with a cell size of 0.25m. Considering the inherent blind zone and near-field disturbances, the blind zone distance is configured to 0.45m, resulting in an effective measurement depth range of 0.5m to 20m, totaling 78 depth cells. The data stream acquired by the acoustic Doppler current profiler includes eastward flow velocity, northward flow velocity, vertical flow velocity, real-time pressure depth, echo intensity, beam correlation coefficient, data goodness rate, and equipment configuration parameters in the geodetic coordinate system. Equipment configuration parameters include blind zone distance and depth cell size. A combined GPS and inertial measurement unit (INS) system synchronously acquires motion and attitude data, including three-dimensional translational velocity, three-axis angular velocity, pitch angle P, roll angle R, and rotation angle Y, with the sampling frequency consistent with that of the acoustic Doppler current profiler.

[0111] The data preprocessing stage begins with time synchronization alignment. The acoustic Doppler current profiler data stream and the attitude data from the combined GNSS and inertial measurement unit (INS) equipment are matched by timestamps, using Coordinated Universal Time (UTC) output by the GNSS as the reference time, with time deviation controlled within one sampling period. After time alignment, rigid body motion compensation is performed. For each depth cell, the velocity vector in the original velocity profile is subtracted from the synchronously acquired buoy three-dimensional translational velocity vector. This vector difference calculation is performed cell-by-cell, deducting the velocity component introduced by the observed buoy's wave-following motion to obtain the disturbed velocity profile.

[0112] Next, a near-surface dynamic blind zone mask is generated. The causes of near-surface blind zones include three aspects: the transducer's inherent blind zone distance, near-field disturbances in the flow field caused by buoys, and bubble entrainment.

[0113] Obtain device configuration parameters from the data stream and extract the blind zone distance, including inherent blind zone and near-field disturbances. Extract real-time pressure depth, pitch angle P, and roll angle R from the data stream. Based on the real-time pressure depth, calculate the real-time vertical distance from the transducer end face to the free water surface using the hydrostatic pressure measured by the pressure sensor. The instantaneous tilt angle of the buoy relative to the water surface normal is calculated using the following formula. :

[0114] The projected distance of the blind zone in the vertical direction at time t is calculated using the following formula. :

[0115] in The blind zone distance configured for the equipment. Depth cells whose depth falls within the vertical distance from the transducer end face to the free water surface and whose blind zone distance is projected vertically are identified as hard blind zones; critical depth cells whose depth falls within the above depth range are also identified as hard blind zones.

[0116] When identifying soft blind zones, the data goodness rate and echo intensity of each depth cell are obtained from the data stream. If the echo intensity of a depth cell is more than three times the background value or the data goodness rate is less than 50%, that depth cell is marked as a soft blind zone. Combining hard and soft blind zones, a binary dynamic blind zone mask is generated, with valid depth cells marked as 1 and invalid depth cells marked as 0. Under normal sea conditions, the total blind zone depth is approximately 1.0m, corresponding to the first four depth cells being marked as invalid; under conditions of strong winds, waves, and bubble interference, the blind zone can expand to 1.5m to 2.5m, corresponding to the first six to ten depth cells being marked as invalid.

[0117] The perturbation velocity profile, attitude data, and dynamic blind zone mask are concatenated into a vertical feature tensor. Figure 2The data organization is illustrated. For each profile, the input data includes the number of completed profile layers and channels, and can be represented as follows: Where N is the number of profile layers after completion, and C is the number of channels. The variables and meanings represented by each channel are shown in Table 1.

[0118] Table 1

[0119] The vertical feature tensor is input into the pre-trained offline flow profile reconstruction model. Combined with... Figure 3 As shown, the velocity profile reconstruction model is a deep learning network, including a feature embedding layer, a feature extraction layer, and a physical information decoding layer.

[0120] The feature embedding layer linearly projects the input vertical feature tensor, mapping the feature vectors of each depth unit into 64-dimensional tokens. These tokens are then arranged in depth order and superimposed with depth position codes to generate a token sequence with positional information. The feature extraction layer employs a shift-window multi-head self-attention mechanism to model the long-range dependency of the vertical velocity distribution and extract multi-scale flow field perturbation features. The feature extraction layer uses two sets of one-dimensional shift-window multi-head self-attention modules with a window size of 8, a shift step size of 4, and 4 attention heads. Window self-attention and shift-window self-attention calculations are performed alternately, enhancing information interaction between adjacent depth windows while controlling computational load. The physical information decoding layer maps high-level features to a full-depth layer background velocity profile through two fully connected layers.

[0121] Training samples were obtained by constructing a flow field sample library. When constructing the flow field sample library, a three-dimensional parametric geometric model was first established based on the observed underwater geometric characteristics of the buoy, fully encompassing the buoy body, the acoustic Doppler current profiler mounting base, and the cylindrical anti-biofouling cover. A cuboid numerical computation domain was established, with a length along the incoming flow direction equal to 40 times the buoy diameter and a width perpendicular to the incoming flow direction equal to 10 times the buoy diameter. The depth of the computation domain was consistent with the operating water depth, covering 8m to 15m. The computation domain was partitioned using a structured grid, with finer meshes applied to the buoy's near-wall region, resulting in a total grid size of approximately 1.2 million.

[0122] Overlapping meshes were used to divide the buoy motion subdomain and the background flow field in the numerical computation domain, simulating the heave and pitch motion of the buoy under wave action. The governing equations were the unsteady Reynolds-averaged Navier-Stokes equations, which were solved numerically using a turbulence model. The fluid volume method was used to capture the water-air free interface, and wave propagation was simulated by combining the inlet velocity wave boundary.

[0123] The following are the combined operating conditions with multiple characteristic water depths, inflow velocities, flow directions, and waves: (1) Selection of characteristic water depth Based on the characteristics of extreme low tide, mean tide, and storm surge high tide levels from hydrological observation stations in the target sea area over many years, and taking into account the range of elevation profile variations in the calculation domain, four sets of typical characteristic water depths were selected, with values ​​of 8.00m (extreme lowest tide), 10.33m (normal median low tide), 12.67m (normal median high tide), and 15.00m (extreme design flood tide level). (2) Selection of incoming flow velocity and direction Incoming current velocity: Based on the long-term ocean current and tidal current velocity spectrum observed in the target operating area, the range of incoming current velocity was selected as 0.1 m / s to 2.5 m / s, and the sampling interval step size was set to 0.5 m / s. Incoming flow direction: Different incoming flow directions are simulated by rotating the three-dimensional model. Starting from 0° (due north incoming flow), a sampling direction is set every 30° (i.e., 0°~330°), for a total of 12 groups.

[0124] (3) Selection of typical wave parameters 1) Normal wave conditions The effective wave height H=1.20m and the effective wave period T=7.0 s correspond to the mid-wave characteristics that occur most frequently throughout the year in the target sea area, representing the most typical and longest-lasting conventional flow field in the working area; 2) High wave and steep working conditions The effective wave height H = 3.20 m and the effective wave period T = 57.0 s (corresponding to wave steepness H / L ≈ 1 / 8 under wave velocity phase length, approaching the physical wave breaking limit of 1 / 7). Under extreme wave steepness conditions, the measured floating body experiences a sudden change in transverse / pitch transient angle and a high-frequency local water flow around it due to the sudden drop in short waves. 3) Long-period extreme surge conditions With an effective wave height of H=5.0m and an effective wave period of T=10.0 s, the flow field at the maximum dynamic submerged draft interface reached by the buoy body under long-period conditions.

[0125] Simultaneous simulation of the three-dimensional flow field under the same boundary conditions, with and without buoys. The boundary conditions are set as follows: upstream along the incoming flow direction is set as the velocity inlet boundary, specifying the inlet velocity and direction, and using the inlet wave generation method to simulate wave propagation; downstream along the incoming flow direction is set as the zero-gradient free outflow boundary; both sides perpendicular to the incoming flow direction are set as symmetrical boundaries; the bottom of the computational domain is set as a no-slip wall boundary; in the buoy case, the surface of the underwater structure of the buoy is set as a no-slip wall boundary.

[0126] For each set of boundary conditions, simultaneous simulations were performed with and without buoys to extract the velocity profile directly below the buoy along the vertical direction. The specific steps are as follows: To avoid flow field interference caused by energy dissipation during initial wave propagation, data sampling begins after at least six wave cycles have elapsed since the wave arrived at the buoy and the flow field and buoy motion have entered a quasi-steady-state cycle. Within two wave cycles, instantaneous velocity profiles and buoy attitude data samples along the vertical direction directly below the buoy are derived at a sampling step size Δt = T / 16. A single computational fluid dynamics simulation task corresponds to the extraction of 2 × 16 = 32 samples.

[0127] Spatial integration averaging was performed using an acoustic Doppler current profiler with a depth element size of 0.25m to obtain a velocity profile consistent with the measured spatial resolution. Using the buoy-based disturbance profile with superimposed attitude data and a blind zone mask as input features, and the buoy-free background velocity profile as labels, 23,040 mapped sample pairs were constructed based on 720 operating conditions to form a flow field sample library. The flow field sample library was then divided into training, validation, and test sets in a 7:2:1 ratio.

[0128] The expression for the total loss function during deep learning network training is: L=L blind +αL obs +βL smooth Among them, L blind The loss represents the prediction loss for the surface blind zone, used to constrain the model to approximate the computational fluid dynamics label corrected by the acoustic Doppler velocity profiler in the missing measurement region; L obs Ls represents the consistency loss of the effective observation layer, used to constrain the model from deviating from the measured flow velocity within the effective observation area of ​​the acoustic Doppler current profiler; mooth α represents the vertical profile smoothing loss, used to suppress unreasonable local oscillations in the predicted profile; α and β are weighting coefficients, both positive, used to adjust the influence of the effective observation layer consistency constraint and the vertical smoothing constraint on the total loss function. Their values ​​are determined based on the training sample size, flow field change characteristics, and model convergence.

[0129] An adaptive moment estimation optimizer was used with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. A validation set was used to monitor the training process, and an early stopping mechanism was employed. Training was stopped when the validation set loss showed no decrease for 15 consecutive epochs. The trainable parameters of the entire network were iteratively updated using the backpropagation algorithm, ultimately yielding a converged velocity profile reconstruction model.

[0130] The velocity profile reconstruction model is deployed within the embedded processing unit of the observation buoy. During online application, after each measurement cycle, time alignment, rigid body motion compensation, dynamic blind zone mask generation, vertical feature tensor construction, and model inference are executed sequentially. After receiving the vertical feature tensor, the model corrects for the hydrodynamic disturbance increment caused by the buoy body within the effective measurement depth range. For the near-surface blind zone, it uses the effective flow velocity at the lower boundary of the blind zone as the boundary condition, combined with the training process, to complete the vertical upward extrapolation reconstruction, outputting a continuous background velocity profile across the entire depth layer.

[0131] A single profile reconstruction takes approximately 0.85 seconds. Validated by an independent test set, the relative mean square error of the velocity at a single point 0.5m from the surface is 5.8%, and the average relative mean square error of the velocity profile within the range of 0.5m to 1.5m from the surface is 2.8%. No non-physical velocity abrupt changes were observed in the reconstructed velocity profile.

[0132] Reference Figure 4 This invention provides a comparison between the reconstructed velocity profiles obtained by this method and the measured velocity profiles at different water depths, including a depth of 0.5m. The measured velocity profiles were obtained simultaneously using a bottom-mounted acoustic Doppler current profiler. As shown in the figures, the reconstructed profiles and the measured profiles agree well. At water depths of 10.2m, 12.0m, and 15.3m, the root mean square errors of the reconstructed and measured velocity profiles are 0.051m / s, 0.02m / s, and 0.03m / s, respectively. At a depth of 0.5m, the average velocity error is 0.04m / s. Table 2 compares the reconstructed velocity profiles obtained by this method with the measured velocity profiles at a depth of 0.5m.

[0133] Table 2 Serial Number Observation time Bottom observations This method reconstructs the value. Difference (Reconstructed value - Observed value) 53 2025 / 3 / 19 14:00 0.22 0.24 0.02 54 2025 / 3 / 19 15:00 0.28 0.31 0.02 55 2025 / 3 / 19 16:00 0.59 0.65 0.06 56 2025 / 3 / 19 17:00 0.69 0.72 0.03 57 2025 / 3 / 19 18:00 0.56 0.55 -0.01 58 2025 / 3 / 19 19:00 0.24 0.24 0.00 59 2025 / 3 / 19 20:00 0.12 0.17 0.05 60 2025 / 3 / 19 21:00 0.30 0.35 0.06 61 2025 / 3 / 19 22:00 0.42 0.42 0.00 62 2025 / 3 / 19 23:00 0.53 0.56 0.03 63 2025 / 3 / 20 0:00 0.57 0.61 0.05 64 2025 / 3 / 20 1:00 0.65 0.59 -0.06 65 2025 / 3 / 20 2:00 0.40 0.35 -0.05 66 2025 / 3 / 20 3:00 0.23 0.21 -0.03 67 2025 / 3 / 20 4:00 0.58 0.56 -0.02 68 2025 / 3 / 20 5:00 0.83 0.85 0.02 69 2025 / 3 / 20 6:00 0.89 0.85 -0.04 70 2025 / 3 / 20 7:00 0.65 0.66 0.02 71 2025 / 3 / 20 8:00 0.23 0.19 -0.05 72 2025 / 3 / 20 9:00 0.49 0.45 -0.04 73 2025 / 3 / 20 10:00 0.49 0.44 -0.05 74 2025 / 3 / 20 11:00 0.41 0.36 -0.06 75 2025 / 3 / 20 12:00 0.63 0.57 -0.06 76 2025 / 3 / 20 13:00 0.71 0.74 0.04 The above provides a detailed description of a deep learning-based method for reconstructing near-surface velocity profiles of buoys. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for reconstructing a near-surface current profile of a buoy based on deep learning, characterized in that, The method includes: Acquire the data stream collected by the acoustic Doppler current profiler on the observation buoy and the motion and attitude data of the observation buoy; The motion data is used to perform rigid body motion compensation on the original velocity profile in the data stream to obtain the disturbed velocity profile. A dynamic blind spot mask is generated using the device configuration parameters and data quality identifiers in the data stream; The disturbed flow velocity profile, the attitude data, and the dynamic blind zone mask are concatenated into a vertical feature tensor; The vertical feature tensor is input into the velocity profile reconstruction model to output a continuous background velocity profile across the entire depth layer. The velocity profile reconstruction model is a deep learning network, which is trained offline using training samples constructed by computational fluid dynamics simulation. The input features of the training samples are buoy-performed disturbance profiles with superimposed buoy attitude data and blind zone masks. The labels of the training samples are background velocity profiles without buoys under the same boundary conditions.

2. The method of claim 1, wherein, The generation of the dynamic blind zone mask includes: Obtain the blind zone distance included in the device configuration parameters; Extract real-time pressure depth, pitch angle, and roll angle from the data stream; Based on the real-time pressure depth, calculate the real-time vertical distance from the transducer end face to the free water surface; Based on the blind zone distance and the pitch and roll angles, calculate the vertical projection distance of the blind zone distance; Depth cells whose depth is within the range of the vertical distance from the transducer end face to the free water surface and the blind zone distance projected in the vertical direction are identified as hard blind zones. The goodness rate of each depth unit is obtained from the data stream, and depth units with a goodness rate lower than a preset threshold are identified as soft blind spots; The hard blind zone and the soft blind zone are merged to generate the dynamic blind zone mask.

3. The method of claim 2, wherein, The training samples are obtained by constructing a flow field sample library. The steps for constructing the flow field sample library include: A three-dimensional model is established based on the underwater geometric features of the observation buoy, and a numerical computation domain is constructed based on the three-dimensional model and the unsteady Reynolds time-averaged Navier-Stokes equations. In the numerical computation domain, an overlapping grid technique is used to divide the buoy motion subdomain and the background flow field domain, and a turbulence model is used to solve the flow field of the buoy motion subdomain and the background flow field domain. Multiple sets of incoming flow velocity, flow direction, water depth, and waves are set, and boundary conditions corresponding to the incoming flow velocity, flow direction, water depth, and waves are applied to the numerical computation domain; Simultaneous simulation of three-dimensional flow fields under the same boundary conditions, with and without buoys; Extract the vertical velocity profile directly below the buoy in the buoy-operating condition, and perform spatial integration averaging according to the depth unit size of the acoustic Doppler current profiler to obtain a disturbed velocity profile with the same resolution as the actual measurement. Using the disturbance velocity profile superimposed with attitude data and blind zone mask as input features, and the velocity profile under buoyless conditions as labels, mapping sample pairs are formed to create the flow field sample library.

4. The method according to claim 3, characterized in that, The deep learning network includes a feature embedding layer, a feature extraction layer, and a physical information decoding layer; The feature embedding layer performs a linear projection on the vertical feature tensor, maps each depth unit to a unified dimension token, and superimposes a depth position code. The feature extraction layer extracts flow field perturbation features from the token; The physical information decoding layer maps the flow field disturbance features into a full-depth layer background velocity profile.

5. The method according to claim 4, characterized in that, The total loss function during the training of the deep learning network is composed of the blind zone prediction loss, the effective observation layer consistency loss, and the vertical profile smoothing loss, which are summed according to their weights. The blind zone prediction loss measures the deviation between the predicted flow rate of the blind zone depth cell and the computational fluid dynamics label. The effective observation layer consistency loss measures the deviation between the predicted flow velocity and the measured flow velocity in the effective observation depth cell. The vertical profile smoothing loss measures the degree of change in predicted flow velocity between adjacent depth cells.

6. The method according to claim 4, characterized in that, The feature extraction layer employs a shift-window multi-head self-attention mechanism to extract flow field perturbation features, including: The token sequence is divided into several non-overlapping local windows along the depth direction, and each local window contains a preset number of depth units; Within each local window, multi-head self-attention is calculated on the tokens within the window to obtain the self-attention output within the window, and the flow field feature correlation between depth cells within the window is extracted; The local window is shifted along the depth direction by a preset step size. Multi-head self-attention is calculated in the new window obtained by the shift, and the self-attention output of the shifted window is obtained, so as to realize the interaction of feature information between adjacent local windows. The self-attention output within the window and the self-attention output of the shifted window are stacked alternately to output a token sequence containing multi-scale flow field perturbation features.

7. The method according to claim 1, characterized in that, The velocity profile reconstruction model outputs a continuous background velocity profile across the entire depth layer in the following manner: The measured flow velocity within the effective measurement depth range is processed by the flow velocity profile reconstruction model to eliminate the hydrodynamic disturbance increment caused by the underwater structure of the observation buoy to the surrounding flow field, and the corrected flow velocity value of the lower boundary depth unit of the near-surface blind zone is obtained. Using the corrected flow velocity value of the lower boundary depth unit of the near-surface blind zone as the boundary condition for extrapolation, and combining the vertical distribution law of flow velocity learned by the deep learning network from computational fluid dynamics simulation samples during the training process, the flow velocity value of each depth unit of the near-surface blind zone is extrapolated layer by layer along the vertical direction to complete the vertical upward extrapolation reconstruction. The velocity profile obtained by correcting the effective measurement depth range and the velocity profile obtained by extrapolating and reconstructing the near-surface blind zone together constitute the continuous background velocity profile of the full depth layer, which is uniformly output by the velocity profile reconstruction model.

8. The method according to claim 3, characterized in that, When simulating the three-dimensional flow field under the same boundary conditions, including both buoy-based and buoy-free conditions, the boundary conditions applied to the numerical computation domain include: The upstream boundary along the direction of the incoming flow is set as the velocity inlet boundary, specifying the inlet flow velocity and direction, and the inlet wave generation method is used to simulate wave propagation; The downstream boundary along the incoming flow direction is set as a zero-gradient free outflow boundary; The two lateral boundaries perpendicular to the direction of the incoming flow are set as symmetrical boundaries; The bottom of the computational domain is set to a non-slip wall boundary; In the case of buoy operation, the surface of the underwater structure of the buoy is set as a non-slip wall boundary.

Images