A method for constructing a flow velocity prediction model based on a thermocline adaptive loss

Abstract

The present application relates to the technical field of flow velocity prediction model construction, and particularly relates to a method for constructing a flow velocity prediction model based on a thermocline adaptive loss. The method comprises the following steps: in the process of shipborne sailing, low-frequency full-water-depth flow velocity background field data and medium-high-frequency target water layer high-resolution flow velocity data are obtained by using a multi-frequency acoustic Doppler current profiler, sound velocity profile data and unified space-time reference data are synchronously obtained, and a multi-source original observation data set is formed; sound velocity gradient data is calculated based on the sound velocity profile data, a thermocline interval is identified, and thermocline structure identification information is generated; under the constraint of the thermocline structure identification information, adaptive layered three-dimensional sound ray tracking calculation is performed on each wave beam measurement unit of the preset multi-frequency ADCP; by means of adaptive layered sound ray tracking and high-frequency detail reservation, the present application overcomes the measurement distortion caused by the dramatic change of sound velocity in the thermocline in the traditional method, and obtains a high-precision flow velocity field of the key water layer.

Description

Technical Field

[0001] This invention relates to the field of flow velocity prediction model construction technology, and in particular to a method for constructing a flow velocity prediction model based on thermocline adaptive loss. Background Technology

[0002] In marine and lake environments, flow velocity variations are influenced by a combination of factors, including water depth, temperature, and water column structure. The presence of a thermocline leads to significant differences in flow velocity across different depths. Traditional flow velocity prediction methods often rely on single-frequency observation data or overall average velocity fields, failing to adequately reflect the high-frequency fluctuations within the thermocline region. Furthermore, they introduce excessive noise outside the thermocline, resulting in low prediction accuracy. In addition, existing numerical simulation or statistical prediction methods typically ignore resolution differences and spatial non-uniformity of multi-source data, making it difficult to accurately construct continuous flow velocity profiles across the entire water depth. In complex aquatic environments, the impact of the thermocline on flow velocity variations is particularly pronounced. Without targeted treatment, flow velocity prediction models are prone to error concentration and instability, thus limiting the reliability of applications such as hydrological analysis, navigation planning, and environmental monitoring. Summary of the Invention

[0003] Therefore, it is necessary to provide a method for constructing a flow velocity prediction model based on thermocline adaptive loss to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for constructing a velocity prediction model based on thermocline adaptive loss includes the following steps: Step S1: During the shipboard navigation process, the low-frequency full-depth current velocity background field data and the medium- and high-frequency target water layer high-resolution current velocity data are acquired in a coordinated manner using a multi-frequency acoustic Doppler current profiler, and the sound velocity profile data and unified spatiotemporal reference data are acquired simultaneously to form a multi-source raw observation dataset. Step S2: Calculate sound velocity gradient data based on sound velocity profile data, identify thermocline intervals, and generate thermocline structure identification information including the top and bottom boundaries of the thermocline and the location of gradient extrema. Step S3: Under the constraint of the thermocline structure identification information, perform adaptive layered three-dimensional acoustic ray tracking calculation on each beam measurement unit of the preset multi-frequency ADCP to obtain the real spatial coordinates and the corrected beam direction vector, and correct the pre-collected original radial velocity to form a low-frequency background velocity field and a high-frequency detail velocity field. Step S4: Construct a fused reference velocity field based on the low-frequency background velocity field, the high-frequency detail velocity field, and the thermocline structure identification information. High-frequency details are preserved in the thermocline region, and low-frequency continuity is maintained in the non-thermocline region to obtain a full-depth reference velocity profile.

[0005] The beneficial effects of this invention are as follows: First, by acquiring low-frequency full-depth current velocity background field and mid-to-high frequency high-resolution current velocity data through shipboard navigation, and combining this with sound velocity profiles and a unified spatiotemporal reference, high-precision alignment of current velocity measurement data in space and time is achieved, providing a reliable raw data foundation for subsequent processing. Second, based on the sound velocity gradient data calculated from the sound velocity profile, thermocline intervals are accurately identified, and thermocline structure identification information including the top boundary, bottom boundary, and gradient extreme value locations is generated. This allows current velocity fusion processing to adopt different strategies for thermoclines and non-thermoclines respectively. To achieve zoned adaptation, adaptive layered three-dimensional acoustic ray tracking using a multi-frequency ADCP beam measurement unit is employed. This allows for the acquisition of accurate spatial coordinates and corrected beam direction vectors, while also correcting the original radial velocity, ensuring the measurement accuracy of both low-frequency background velocity and high-frequency detail velocity. Finally, based on the fusion processing of low-frequency background velocity, high-frequency detail velocity, and thermocline structure information, high-frequency details are preserved in the thermocline region, while low-frequency continuity is maintained in the non-thermocline region. Boundary transitions ensure smooth and continuous velocity changes, resulting in a full-depth reference velocity profile. This full-depth reference velocity profile significantly improves the input accuracy and stability of the velocity prediction model, helps reduce the impact of the thermocline on prediction errors, and enhances the model's reliability and adaptability under complex water conditions. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating the steps involved in constructing a flow velocity prediction model based on thermocline adaptive loss. Figure 2 To identify the thermocline in the sound velocity profile and generate a structural identification diagram; Figure 3 This is a schematic diagram of the reference velocity profile across the entire water depth. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0007] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. 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 inventive effort are within the scope of protection of the present invention.

[0008] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0009] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0010] To achieve the above objectives, please refer to Figures 1 to 3 A method for constructing a flow velocity prediction model based on thermocline adaptive loss includes the following steps: All specific values ​​involved in this embodiment are exemplary parameters used to clearly illustrate the technical operation process and are not the only limitation of the present invention.

[0011] Step S1: During the shipboard navigation process, the low-frequency full-depth current velocity background field data and the medium- and high-frequency target water layer high-resolution current velocity data are acquired in a coordinated manner using a multi-frequency acoustic Doppler current profiler, and the sound velocity profile data and unified spatiotemporal reference data are acquired simultaneously to form a multi-source raw observation dataset. Step S2: Calculate sound velocity gradient data based on sound velocity profile data, identify thermocline intervals, and generate thermocline structure identification information including the top and bottom boundaries of the thermocline and the location of gradient extrema. Step S3: Under the constraint of the thermocline structure identification information, perform adaptive layered three-dimensional acoustic ray tracking calculation on each beam measurement unit of the preset multi-frequency ADCP to obtain the real spatial coordinates and the corrected beam direction vector, and correct the pre-collected original radial velocity to form a low-frequency background velocity field and a high-frequency detail velocity field. Step S4: Construct a fused reference velocity field based on the low-frequency background velocity field, the high-frequency detail velocity field, and the thermocline structure identification information. High-frequency details are preserved in the thermocline region, and low-frequency continuity is maintained in the non-thermocline region to obtain a full-depth reference velocity profile.

[0012] In one embodiment, during shipboard navigation, the primary ADCP (low frequency, e.g., 150 kHz) continuously scans the background current field across the entire water depth, while the secondary ADCP (mid-to-high frequency, e.g., 600 kHz–1.2 MHz) focuses on the target water layer for high-resolution current velocity acquisition. All multi-frequency ADCPs transmit and receive synchronously with a uniform sampling period (e.g., 1 Hz). A sound velocity profiler (SVP) acquires depth-based sound velocity data sequences, and GNSS and ship attitude sensors obtain unified spatiotemporal reference information. After acquisition, the low / high frequency current velocity data and sound velocity profile data are aligned by timestamps, and spatial calibration and attitude correction are performed to form a complete multi-source raw observation dataset.

[0013] The sound velocity profile data is sorted by depth from shallowest to deep and smoothed to obtain a continuous sound velocity profile sequence. Based on this sequence, the sound velocity gradient at adjacent depths is calculated, generating gradient data for each depth point. According to a preset threshold (e.g., 0.2 m⁻¹), depth intervals that continuously meet the conditions are identified as candidate thermoclines. For each candidate interval, the point with the maximum absolute gradient value is identified as the gradient extremum location. Using this location, a search is performed upwards and downwards until the gradient decreases to a threshold percentage (e.g., 50%) to determine the top and bottom boundaries of the thermocline. Thermocline structure identification information, including the top and bottom boundaries and the gradient extremum location, is generated.

[0014] Based on the thermocline structure identification information, an adaptive layered structure is generated: a smaller layer thickness (e.g., 0.2–0.5 m) is used in the thermocline zone, and a larger layer thickness (e.g., 1–2 m) is used in the non-thermocline zone. For each beam measurement unit of the multi-frequency ADCP, the acoustic ray path is traced layer by layer, the horizontal and vertical displacements of each layer are calculated, and the summation is used to obtain the true position coordinates of the beam measurement unit in three-dimensional space. Combining the ADCP attitude and installation parameters, the coordinates are unified to a reference frame to generate a corrected beam direction vector. Finally, the original radial velocity is matched with the direction vector and corrected to obtain the low-frequency background velocity field and the high-frequency detail velocity field.

[0015] The low-frequency background velocity field and high-frequency detail velocity field obtained from S3 are spatially aligned and divided into intervals based on thermocline structure identification information. Within the thermocline interval, high-frequency detail velocities dominate, while low-frequency data is interpolated to ensure continuity. In non-thermocline intervals, low-frequency background velocities dominate, with high-frequency data weighted and fused or smoothed. Boundary transition processing is applied at the boundary between the thermocline and non-thermocline to ensure continuous velocity across the entire depth and locally high resolution. The output is a fused reference velocity profile across the entire depth, which can be used as input for training supervision or prediction models.

[0016] It should be added that the system adopts a working mode that combines hardware triggering and logical coordination. The main ADCP serves as the reference device for continuous scanning across the entire water depth, and its echo signal is processed in real time to detect the sound velocity gradient. The vertical distribution. When detected Exceeding a preset threshold (e.g., 0.5s) -1 When the thickness is ≥2 meters, the central controller sends a hardware trigger signal to the auxiliary ADCP. It represents the change in sound speed between adjacent depths, reflecting the magnitude of the change in sound speed in water with depth. The depth interval (in meters) corresponding to the change in sound velocity is usually determined by the resolution of the profile measurement, such as 0.5m or 1m.

[0017] The sound velocity gradient, representing the rate of change of sound velocity per unit depth, reflects the degree of drastic change in water density or temperature, and initiates high-resolution focusing observations within the corresponding water layer. In homogeneous water layer regions, the auxiliary ADCP can enter low-power standby or downsampling mode. Simultaneously, both the primary and auxiliary ADCPs synchronously receive GNSS pulse-per-second (1PPS) signals to achieve time synchronization, ensuring strict alignment of the two sets of beam data in the time dimension.

[0018] Spatially, the primary and secondary ADCPs are mounted on a shared frame and their relative positions and attitudes are pre-calibrated. Precise overlap of beam footprints is achieved through spatial coordinate transformation. Finally, all raw data enters the post-processing system after acquisition for joint acoustic correction and variational fusion. This mode realizes a multi-level observation closed loop of "primary ADCP scanning the entire domain to discover targets, secondary ADCP focusing on acquiring details, and unified spatiotemporal reference for collaborative output".

[0019] The system acquires high-resolution sound velocity profile c(z) in real time and automatically identifies the vertical stratification structure of the entire water body using a comprehensive algorithm of "gradual threshold-extreme value locking". The specific process is as follows: Data preprocessing and gradient calculation: Calculating the normalized sound velocity gradient ( (For surface sound velocity), and perform smoothing filtering to suppress high-frequency noise. For a certain depth The sound velocity value (in m / s) at a certain point is measured by a sound velocity profiler (SVP) and represents the distribution of sound wave propagation speed with depth in the water. After normalization, the gradient unit is 1 / m, which physically represents the proportion of the relative change in sound velocity per meter of depth.

[0020] Initial screening of tiered gradients (threshold method): setting a basic gradient threshold. (like Layer segments within a continuous depth range that meet this threshold are marked as potential skip regions.

[0021] Precise localization and layer boundary delineation (extreme value-inflection point method): For each potential skip region, further search for local maxima of the absolute value of the gradient. This point represents the strongest core position of the multi-level structure. Subsequently, with... Centered on the target layer, search for depths where the gradient drops to half the threshold (or the gradient curvature changes significantly) to the upper and lower sides, and define these depths as the top boundaries of the strata. and bottom boundary .

[0022] Multi-layer processing and priority ranking: If multiple thermoclines are identified across the entire water body (e.g., two thermoclines exist in deep water), the system adds each thermocline object and its characteristic parameters (core strength, thickness, top and bottom depth) to the scheduling queue. The focused observation resources of the auxiliary ADCP are dynamically allocated according to a "priority strategy," with priorities... The following formula is used for comprehensive evaluation: ; in, For the core strength of the layer, This is a typical strong-jump gradient reference value. For the thickness of the interlayer, Given the current hull / platform depth, For reference to the thickness of the mezzanine, the weights It can be preset according to the observation task objective (e.g., biased towards the strongest or shallowest jump layer). It is a local maximum point of the absolute value of the gradient.

[0023] Adaptive Scheduling: The central controller sends commands to the auxiliary ADCP based on priority ranking. If there is only one hop layer, the system focuses on observing that layer; if multiple high-priority hop layers exist, a "time-division multiplexing" mode is adopted: within a single pulse transmission cycle, multiple receiving windows are configured using beamforming technology to sequentially sample hop layers at different depths, or the focusing depth is alternately switched between consecutive pings to achieve time-division coverage of multiple targets. The system simultaneously ensures that the main ADCP continuously performs full-depth background field scanning, unaffected by scheduling.

[0024] Please refer to [link / reference needed] for further information. Figure 2 The shipborne multi-frequency ADCP in the figure transmits sound beams underwater (indicated by the blue arrows) to collaboratively collect raw data such as background current velocity across the entire water depth and high-resolution current velocity of the target water layer, providing a multi-source observation dataset for subsequent thermocline identification, sound ray tracking correction, and model construction.

[0025] Please refer to [link / reference needed] for further information. Figure 3The vertical axis (left side) represents water depth (m), gradually increasing from top to bottom. The horizontal axis (bottom) represents flow velocity (m / s), gradually increasing from left to right. The cluster of curves represents flow velocity profiles at different depths, illustrating the variation of flow velocity with depth. The dashed outline represents the low-frequency background velocity field, reflecting the overall continuous trend of flow velocity across the entire water depth. The fine solid lines represent the high-frequency detailed velocity field, preserving subtle variations in velocity within the thermocline region.

[0026] Of particular importance, step S1 includes: During the shipboard navigation process, low-frequency and mid-to-high-frequency acoustic Doppler current profilers are deployed in a coordinated manner and synchronously transmitted and received according to a unified sampling period to obtain echo data in different frequency bands. Low-frequency echo data is processed to obtain low-frequency velocity background field data covering the entire water depth, and mid-to-high frequency echo data is processed to obtain high-resolution velocity data within the target water layer. Simultaneously, sound velocity data sequences distributed along the depth are acquired using sound velocity profile measurement equipment and recorded in chronological order; Based on the unified time reference and spatial location information obtained by positioning and attitude measurement equipment, the above-mentioned data are processed for time alignment and spatial matching to form a multi-source raw observation dataset.

[0027] In one embodiment, during shipboard navigation, low-frequency and mid-to-high-frequency acoustic Doppler current profilers are deployed on both sides of the hull and synchronously transmitted and received at a unified sampling period to ensure that echo data from different frequency bands are synchronized in time. The low-frequency echo data undergoes preprocessing, including noise reduction, backpropagation correction, and Doppler frequency shift calculation, to obtain low-frequency current velocity background field data covering the entire water depth. The mid-to-high-frequency echo data undergoes preprocessing, including high-resolution beamforming and noise filtering, to obtain high-resolution current velocity data within the target water layer.

[0028] Simultaneously, sound velocity data sequences distributed along the depth are acquired using sound velocity profile measurement equipment (such as CTD sound velocity profiler), and the sound velocity value of each sampling point is recorded in chronological order. The ship's position and attitude information are acquired through positioning and attitude measurement equipment (GNSS+IMU), and the low-frequency flow velocity field, high-frequency flow velocity field, and sound velocity profile data are processed with time alignment and spatial matching using a unified timestamp to form a multi-source original observation dataset, providing unified and comparable basic data for thermocline analysis and adaptive flow velocity prediction model training.

[0029] In another embodiment, it is assumed that the low-frequency ADCP sampling frequency is 1Hz, covering the entire water depth from 0 to 100m, with sampling every 2m along the water depth, for a total of 50 layers; the mid-to-high frequency ADCP sampling frequency is 10Hz, covering only the target water layer from 10 to 40m, with a water depth resolution of 0.5m, for a total of 60 layers; the sound velocity profiler measures the sound velocity from the surface to the bottom of the water body in the range of 1475–1485. Records data once per second. The hull positioning accuracy is approximately 0.5m, and the attitude angle accuracy is approximately... Through time synchronization and spatial alignment processing, the low-frequency background velocity field and high-frequency detail velocity field, along with the corresponding sound velocity profile sequence, are obtained at one frame per second. This constitutes a multi-source raw observation dataset containing 100 frames, which is used for subsequent thermocline detection, layered velocity correction, and deep learning prediction model construction.

[0030] Preferably, step S4 further includes: A flow velocity prediction model was constructed using multi-source raw observation datasets and full-depth reference flow velocity profiles as training samples. During the training process, a depth-weighted loss function based on sound velocity gradient data and thermocline structure identification information was introduced. The flow velocity prediction model is optimized by using a depth-weighted loss function, so that the prediction error weight in the thermocline region is higher than that in the non-thermocline region. This results in a flow velocity prediction model that maintains background consistency throughout the water depth and has high-resolution representation capability in the thermocline region.

[0031] In one embodiment, an existing multi-source raw observation dataset is used, including low-frequency ADCP full-depth background velocity data, high-frequency ADCP target water layer high-resolution velocity data, and sound velocity profile sequences. Simultaneously, a full-depth reference velocity profile (e.g., obtained from CTD observations or historical profile averaging) is acquired as training samples. When constructing the deep learning velocity prediction model, the data is divided into equally spaced depth blocks according to water depth (e.g., one depth layer per 1m), and corresponding sound velocity gradient information and thermocline identification information are appended to each depth layer.

[0032] The constructed prediction model comprises convolutional neural network (CNN) layers and recurrent neural network (RNN) layers: the CNN layers are used to extract local spatial velocity features at each depth level; the RNN layers (such as bidirectional LSTM) are used to capture the dynamic features of water changes with depth and time. During training, a depth-weighted loss function is employed. This function assigns higher weights to the prediction errors in the thermocline region, prioritizing the prediction accuracy in the thermocline region over that in non-thermocline regions, while ensuring the consistency of the background velocity field across the entire water depth. The model parameters are optimized using a backpropagation algorithm, resulting in a velocity prediction model that maintains velocity consistency across the entire water depth and possesses high-resolution representation capabilities in the thermocline region.

[0033] In another embodiment, it is assumed that the training samples include 100 frames of low-frequency ADCP velocity profiles across the entire water depth from 0 to 100 m, and 200 frames of high-frequency ADCP velocity profiles at the target water layer from 10 to 40 m, with each layer spaced 1 m apart; the sound velocity gradient averages 0.25 in the thermocline range (assumed to be 20–30 m). The average value for the non-thermocline region is 0.05. Thermochromatic layer identifiers are represented in binary form (1 represents a thermochromatic layer, 0 represents a non-thermochromatic layer). The model is constructed using a three-layer convolutional network: Conv1 (3×3 convolution, 32 channels) → → Conv2 (3×3 convolution, 64 channels) → → Conv3 (3×3 convolution, 128 channels) → → Output feature vectors for each depth layer The depth feature sequence is then input into a two-layer bidirectional LSTM (256 hidden units), and the forward and reverse hidden states are concatenated to obtain the full-depth feature vector. .

[0034] The depth-weighted loss function is defined as: ; in, For reference flow rate, To predict flow rate, The weighting coefficient is taken within the thermocline range. =5, non-thermal spring section is taken =1. During training, the model iterated 100 times, with a batch size of B=16, a learning rate of 0.001, and used Adam as the optimizer. After training, the model's average prediction error in the thermocline region decreased to 0.03 m / s, and the background error across the entire water depth was controlled within 0.05 m / s. It was able to achieve high-resolution flow velocity representation in the thermocline region while maintaining consistency across the entire water depth.

[0035] Preferably, the depth-weighted loss function constructed based on the sound velocity gradient and thermocline structure identification information includes: The flow velocity observation data is extracted from the multi-source raw observation dataset and used as the model input. The full-depth reference flow velocity profile is used as the supervision label to form training samples. A flow rate prediction model is built based on training samples, and a basic loss function is constructed between the prediction results and the supervision labels. Based on the sound velocity gradient data and thermocline structure identification information, determine whether each depth location is located within the thermocline interval, and construct weight coefficients corresponding to the depth. By incorporating weighting coefficients into the basic loss function, the prediction errors at different depth locations are weighted to form a depth-weighted loss function.

[0036] In one embodiment, a multi-source raw observation dataset is used, including low-frequency ADCP full-depth velocity profiles, high-frequency ADCP target layer high-resolution velocity data, and sound velocity profile sequences. Velocity observation data is used as model input, and the full-depth reference velocity profile is used as a supervision label to form training samples. Next, a velocity prediction model based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) is constructed. CNNs are used to extract local spatial velocity features in the depth direction, while RNNs (such as bidirectional LSTMs) are used to capture the dynamic features of velocity changes with time and depth.

[0037] Subsequently, a depth-weighted loss function is constructed during training: the basic mean squared error loss between the predicted result and the supervision label is calculated; combining the sound velocity gradient data and the thermocline structure identification information, it is determined whether each depth location is located within the thermocline interval, and a corresponding weight coefficient is assigned to each depth location. (Thermoclimate intervals have higher weights than non-thermoclimate intervals); finally, the weighting coefficients are introduced into the basic loss function to weight the prediction errors at different depth locations, forming a depth-weighted loss function. The model is trained using this depth-weighted loss function, enabling it to maintain consistency of background current velocity across the entire water depth while possessing high-resolution current velocity prediction capabilities within the thermocline region.

[0038] In another embodiment, it is assumed that the training samples include 100 frames of low-frequency ADCP full-depth profiles (0–100m, with each depth layer in 1m increments) and 200 frames of mid-to-high-frequency ADCP target water layer profiles (10–40m, with each depth layer in 0.5m increments), and the sound velocity gradient averages 0.25 in the thermocline range (assumed to be 20–30m). The average value of the non-thermocline zone is The thermocline identifier is represented in binary form (1 represents a thermocline, 0 represents a non-thermocline).

[0039] When building the model, the CNN part uses a three-layer convolutional network: Conv1 (3×3 convolution, 32 channels) → → Conv2 (3×3 convolution, 64 channels) → Conv3 (convolution, 128 channels) → Output feature vectors for each depth layer The depth feature sequence is then input into a two-layer bidirectional LSTM (256 hidden units), and the forward and reverse hidden states are concatenated to obtain the full-depth feature vector. A depth-weighted loss function L was used for model training, employing the Adam optimizer with a learning rate of 0.001, a batch size of 16, and 100 training iterations. After training, the average prediction error for the thermocline interval was 0.03. The background error across the entire water depth was controlled within 0.05 m / s, achieving consistent background across the entire water depth and high-resolution velocity prediction of the thermocline.

[0040] Preferably, step S2 includes: The sound velocity profile data is sorted by depth and then smoothed and filtered to obtain a continuous sound velocity profile sequence. The sound velocity gradient data is obtained by calculating the sound velocity gradient at adjacent depths based on a continuous sound velocity profile sequence. Based on the preset gradient threshold, the depth intervals that continuously meet the conditions are determined as candidate thermocline intervals, and the gradient extremum position with the largest absolute value of the sound velocity gradient is determined in each interval. Based on the extreme value of the gradient, the depth positions where the gradient drops to a preset proportional threshold are determined upwards and downwards as the top and bottom boundaries of the thermocline. Based on the top and bottom boundaries of the thermocline and the sound velocity gradient data, generate thermocline structure identification information corresponding to the depth.

[0041] In one embodiment, sound velocity profile data of the target water area is acquired, and the profile data is sorted in ascending order of water depth to obtain a depth-sound velocity mapping sequence. This sequence is then smoothed using a filtering process (e.g., using a five-point weighted moving average filter) to eliminate local noise and outliers, resulting in a continuous sound velocity profile sequence. Subsequently, the sound velocity gradient at adjacent depth points is calculated to form a sound velocity gradient data sequence.

[0042] Based on a preset gradient threshold, depth intervals where the sound velocity gradient data continuously exceeds the threshold are identified and designated as candidate thermocline intervals. Within each candidate interval, the depth location with the largest absolute value of the sound velocity gradient is identified as the gradient extremum location. Using this extremum location as a baseline, the search continues upwards and downwards until the absolute value of the gradient decreases to a preset proportion (e.g., 50%), which are denoted as the top and bottom boundaries of the thermocline, respectively. The top and bottom boundaries of each thermocline interval, along with the corresponding depth's sound velocity gradient information, are integrated to generate a sequence of thermocline structure identification information corresponding to the depth, providing a reference for the subsequent depth-weighted loss function.

[0043] In another embodiment, it is assumed that the extracted sound velocity profile data covers a water depth of 0–100 m, with sampling once per meter, for a total of 101 depth points. A continuous sequence is obtained after applying a five-point weighted smoothing filter to the original profile data; the calculated sound velocity gradient between adjacent depths ranges from -0.12 to... Let the gradient threshold be... Through continuous screening, it was found that the gradient in the 20–30m depth range consistently exceeded the threshold, thus identifying it as a candidate thermocline range. Within this range, the maximum sound velocity gradient is... The location is at a depth of 25m. Using 25m as a baseline, search upwards until the gradient decreases to 50% of its maximum value. The location is the top boundary of the thermocline, at a depth of approximately 22m; the search proceeds downwards until the gradient decreases. The location is at the bottom boundary of the thermocline, at a depth of approximately 28m. The final generated thermocline structure identification information is marked as "1" in the 22–28m range and "0" for depth, which is used to construct the depth-weighted loss of the subsequent velocity prediction model.

[0044] Preferably, the sound velocity gradient is calculated based on a continuous sound velocity profile sequence to obtain sound velocity gradient data including: Number each depth position according to a preset depth interval and obtain the corresponding sound velocity value; Calculate the sound velocity difference and depth difference between adjacent depth positions, and obtain the sound velocity gradient by ratio calculation; When both upper and lower neighbor points exist, the sound velocity gradients in the two directions are calculated separately and weighted to obtain the sound velocity gradient at that depth.

[0045] In one embodiment, the continuous sound velocity profile sequence is numbered according to a preset depth interval, with each depth location corresponding to a sound velocity value, forming a depth-sound velocity mapping list. Then, the sound velocity values ​​at adjacent depth locations are compared, and the ratio of the sound velocity difference to the depth difference is calculated to obtain the preliminary sound velocity gradient at that depth location. During the calculation process, when a depth point has both upper and lower neighbors, the sound velocity gradients in the directions of the upper and lower neighbors are calculated separately, and weighted according to distance or depth intervals to obtain the comprehensive sound velocity gradient at that depth location. Finally, the sound velocity gradients of all depth points are integrated to form a complete sound velocity gradient data sequence, providing input for subsequent thermocline determination and depth-weighted loss function construction.

[0046] In another embodiment, it is assumed that continuous sound velocity profile data covers water depths of 0–100 m, with sampling occurring once per meter, for a total of 101 depth points. The sound velocity value at each depth location is obtained sequentially according to a preset number. For each depth point, the sound velocity difference with the upper neighboring depth point and the depth difference are calculated to obtain the upward sound velocity gradient; simultaneously, the sound velocity difference with the lower neighboring depth point and the depth difference are calculated to obtain the downward sound velocity gradient. It is assumed that at a depth of 50 m, the upward gradient is 0.15. The downward gradient is 0.12. The weighted average gradient is then approximately 0.135. After processing all depth points sequentially, a sound velocity gradient sequence is generated for the entire water depth from 0 to 100 m, with a gradient range of approximately -0.10 to 0.30. This sound velocity gradient sequence can be used to identify thermocline intervals and construct depth-weighted loss functions, providing refined depth weight information for flow velocity prediction models.

[0047] Preferably, using the gradient extremum location as a reference, determining the depth positions upward and downward where the gradient descends to a preset proportional threshold as the top and bottom boundaries of the thermocline includes: Within the thermocline zone, the location of the gradient extreme value with the largest absolute value of the sound velocity gradient is determined as the reference depth location. Starting from this location, search along the direction of decreasing depth. When the absolute value of the sound velocity gradient drops to a preset proportional threshold, determine the top boundary of the thermocline. Starting from this location, search along the direction of increasing depth. When the absolute value of the sound velocity gradient decreases to a preset proportional threshold, determine the bottom boundary of the thermocline.

[0048] In one embodiment, within the identified candidate thermocline intervals, the sound velocity gradient data at all depth points are scanned, and the gradient point with the largest absolute value is determined as the reference depth position, i.e., the center position of the thermocline. Starting from this center depth position, the search proceeds point by point upwards along the depth decreasing direction, monitoring the change in the absolute value of the sound velocity gradient. When the absolute value of the gradient drops to a preset percentage threshold (e.g., 50%), the corresponding depth position is identified as the top boundary of the thermocline. Subsequently, starting from the same center depth position, the search proceeds point by point downwards along the depth increasing direction. When the absolute value of the sound velocity gradient drops to the same percentage threshold, the corresponding depth position is identified as the bottom boundary of the thermocline. Each candidate thermocline interval is precisely limited to its actual upper and lower bounds, obtaining complete thermocline depth information.

[0049] In another embodiment, assuming a candidate thermocline range of 40–60 m is detected in a water profile with a depth of 0–100 m, and the absolute value of the sound velocity gradient reaches its maximum value at 50 m within this range, this is taken as the baseline depth. Starting from 50 m, the search proceeds upwards. When the absolute value of the sound velocity gradient decreases to 50% of its maximum value, the corresponding depth is 45 m, thus the upper boundary of the thermocline is identified as 45 m. Then, starting from 50 m, the search proceeds downwards. When the absolute value of the gradient decreases to 50%, the corresponding depth is 58 m, thus the lower boundary of the thermocline is identified as 58 m. Finally, the complete depth range of the thermocline is obtained as 45–58 m, which can be used to generate thermocline structure identification information and to weight the contribution of the thermocline at each depth in the velocity prediction model.

[0050] Preferably, step S3 includes: Based on the thermocline structure identification information, the layering rules along the depth direction are determined. A smaller layer thickness is used in the thermocline interval and a larger layer thickness is used in the non-thermocline interval to generate an adaptive layered structure corresponding to the depth. Under the adaptive hierarchical structure constraint, the sound ray propagation path is calculated layer by layer for each beam measurement unit of the multi-frequency ADCP, the refraction direction of each layer is determined and the horizontal and vertical displacements are accumulated. Based on the accumulated horizontal and vertical displacements, the position coordinates of each beam measurement unit in space are determined, and coordinate transformation is performed in combination with the equipment attitude and installation parameters to obtain the beam direction vector. The beam direction vector is matched with the original radial velocity, and the radial velocity is decomposed and corrected based on the beam direction vector to generate a low-frequency background velocity field and a high-frequency detail velocity field.

[0051] In one embodiment, an adaptive stratification rule is determined along the depth direction based on the acquired thermocline structure identification information. Within the thermocline region, due to rapid changes in sound velocity and water structure, a smaller layer thickness is used for stratification; in non-thermocline regions, water parameters change more slowly, resulting in a larger layer thickness, thus generating an adaptive stratification structure corresponding to the depth. Subsequently, under this adaptive stratification constraint, for each beam measurement unit of the multi-frequency ADCP, the propagation path of the sound rays is calculated layer by layer to determine the sound ray refraction direction of each layer, and the horizontal and vertical displacements are accumulated accordingly. Based on the accumulated displacement information, the three-dimensional coordinates of each beam measurement unit in space can be determined, and coordinate transformation is performed in conjunction with equipment attitude information and installation parameters to finally obtain the direction vector of each beam. The beam direction vector is matched with the original radial flow velocity, and direction decomposition and correction are performed on this basis to generate a low-frequency background velocity field and a high-frequency detail velocity field, providing spatially refined input for subsequent velocity prediction and adaptive loss training.

[0052] In another embodiment, it is assumed that the thermocline range is 30–60m, with a layer thickness of 1m; and the non-thermocline ranges are 0–30m and 60–100m, with a layer thickness of 5m. For a 4-beam measurement unit of a certain ADCP, the refraction angle and horizontal / vertical displacement of each layer are obtained after calculating the sound ray propagation path layer by layer. For example, at a position 40m from the center of the thermocline, the upward refraction angle is... The downward refraction angle is -1.5°, with a cumulative horizontal displacement of approximately 0.12m and a vertical displacement of approximately 0.05m; the refraction angle of a certain layer in the non-thermocline layer is... The cumulative horizontal displacement is approximately 0.03 m, and the vertical displacement is approximately 0.01 m. After obtaining the beam direction vector through coordinate transformation, it is matched with the original radial velocity, and direction decomposition and correction are performed to generate a low-frequency background velocity field (e.g., average velocity of 0.45 m / s) and a high-frequency detail velocity field (e.g., fluctuation range ±0.08 m / s). This method can accurately reflect the layered velocity characteristics of the thermocline and non-thermocline layers, providing reliable support for subsequent adaptive loss training and model optimization.

[0053] It should be added that multi-frequency data synchronous acquisition and three-dimensional dynamic ray tracking correction are performed, with the main and auxiliary ADCPs acquiring data synchronously. For each distance unit of each beam, an adaptive hierarchical three-dimensional ray tracking algorithm is applied for independent correction. Adaptive hierarchical: based on the sound velocity gradient... The computational layers are dynamically partitioned. This allows for adaptive adjustment, using thicker layers in regions of gentle gradients and thinner layers in regions of abrupt gradient changes. Specifically, a continuous thickness function is introduced: ; in, , It is divided into preset maximum layer thickness (e.g., 2m) and minimum layer thickness (e.g., 0.2m). This is a dynamic threshold (automatically updated based on recent gradient statistics).

[0054] Acoustic ray path tracking: Based on the constant gradient assumption, the bending path of the acoustic wave is tracked layer by layer using Snell's law. For the i-th layer, the horizontal displacement of the beam... and vertical displacement It is given by the following formula: ; ; in, and These are the sound velocity at the top of the layer and the gradient within the layer, respectively. and The grazing angles are the top and bottom of the layer. The true three-dimensional spatial coordinates of the measurement unit are determined by tracing the acoustic path length until it equals the slant distance measured by ADCP.

[0055] Data correction: The true direction vector of each beam is recalculated based on the corrected coordinates and used to correct the original radial velocity observations. The final output consists of primary and secondary ADCP velocity profiles that have undergone dual correction based on both spatial location and velocity geometry. and .

[0056] Preferably, under the constraint of an adaptive layered structure, the sound ray propagation path is calculated layer by layer for each beam measurement unit of the multi-frequency ADCP, the refraction direction of each layer is determined, and the horizontal and vertical displacements are accumulated, including: Under the constraint of adaptive layered structure, the refraction angle is calculated based on the incident angle and sound speed in each layer to determine the sound ray propagation direction of the current layer; Calculate the horizontal and vertical displacements corresponding to the current layer based on the propagation direction of the current layer. The horizontal and vertical displacements of each layer are accumulated layer by layer along the propagation path to obtain the cumulative displacement of the sound ray from the emission point to the target measurement unit. The sound ray propagation path data is generated based on the cumulative displacement and used as input for the spatial positioning of the measurement unit.

[0057] In one embodiment, based on the generated adaptive hierarchical structure, for each ADCP beam measurement unit, the refraction angle of the sound ray in each layer is calculated layer by layer according to the incident angle and sound velocity information to determine the sound ray propagation direction of the current layer. Subsequently, based on the propagation direction of the current layer and the thickness of the layer, the corresponding horizontal and vertical displacements are calculated. The horizontal and vertical displacements of each layer are accumulated layer by layer along the sound ray propagation path to obtain the cumulative displacement information of the sound ray from the exit point to the target measurement unit. Through these cumulative displacements, sound ray propagation path data can be generated, including the three-dimensional coordinates and refraction direction of each layer, providing refined input for subsequent spatial positioning of measurement units, velocity direction decomposition, and adaptive loss training.

[0058] In another embodiment, it is assumed that the thermocline range is 40–70m, and the non-thermocline range is 0–40m and 70–100m. The thermocline is divided into layers with a thickness of 1m, and the non-thermocline is divided into layers with a thickness of 5m. A certain beam incidence angle... For example: at the 3m level of the thermocline, the speed of sound is 1480m / s, and the calculated angle of refraction is approximately... The horizontal displacement is approximately 0.035m, and the vertical displacement is approximately 0.015m; at the 4th layer of the thermocline, the refraction angle is approximately... The horizontal displacement is approximately 0.034m, and the vertical displacement is approximately 0.016m; the refraction angle of the first layer of the non-thermocline is approximately... The horizontal displacement is approximately 0.02m, and the vertical displacement is approximately 0.01m. After accumulating the data layer by layer along the sound ray, the cumulative horizontal displacement from the ADCP emission point to the measurement unit is approximately 0.45m, and the cumulative vertical displacement is approximately 0.21m. Based on this cumulative displacement, complete sound ray propagation path data is generated, including the three-dimensional position and propagation direction corresponding to each layer.

[0059] Preferably, based on the accumulated horizontal and vertical displacements, the position coordinates of each beam measurement unit in space are determined, and coordinate transformation is performed in conjunction with the equipment attitude and installation parameters to obtain the beam direction vector, including: Based on the accumulated horizontal and vertical displacements, the position coordinates of each beam measurement unit in three-dimensional space are calculated; The spatial coordinates of the measurement unit are transformed with the equipment attitude and installation parameters to be unified to the reference coordinate system; Under a unified coordinate system, a spatial direction vector is constructed using the coordinates of the beam exit point and the spatial coordinates of the measurement unit, and the spatial direction vector is normalized to generate the beam direction vector.

[0060] In one embodiment, the cumulative horizontal and vertical displacements calculated in step S3 are used to determine the position coordinates of each ADCP beam measurement unit in three-dimensional space. For each measurement unit, the cumulative horizontal displacement along the sound ray propagation path is used as the increment of the X / Y coordinates, and the vertical displacement is used as the increment of the Z coordinate, to obtain the three-dimensional position of the beam measurement unit in the local sensor coordinate system. Subsequently, the three-dimensional coordinates are transformed with the ADCP device attitude (including pitch, roll, and yaw angles) and installation parameters (such as installation height, tilt angle, and installation position offset) to uniformly map each beam measurement unit to the global reference coordinate system. In the unified coordinate system, a spatial direction vector is constructed using the beam exit point coordinates and the spatial coordinates of the measurement unit, and then normalized to generate a standardized direction vector for each beam, providing basic data for subsequent radial velocity direction decomposition and flow field prediction.

[0061] Preferably, under a unified coordinate system, a spatial direction vector is constructed using the coordinates of the beam exit point and the spatial coordinates of the measurement unit, and the spatial direction vector is normalized to generate the beam direction vector, including: Obtain the coordinates of the beam exit point and the spatial coordinates of the measurement unit under a unified coordinate system; Based on the coordinate difference between the spatial coordinates of the measurement unit and the coordinates of the beam exit point, the corresponding spatial displacement components are calculated. The spatial displacement components are combined according to the coordinate axis directions to construct a spatial direction vector pointing to the measurement unit; The spatial direction vector is calculated in terms of magnitude, and then normalized based on the magnitude to generate the beam direction vector.

[0062] In one embodiment, the exit point position of each beam and the corresponding measurement unit's position in space are obtained under a unified reference coordinate system. The differences between the measurement unit's position and the exit point position along three directions are calculated, namely, the displacement components in the horizontal and vertical directions. These displacement components are then combined according to spatial directions to form a spatial direction vector pointing towards the measurement unit. Subsequently, the length of this vector is measured, and the vector is normalized based on the length to obtain a standardized beam direction vector. The generated direction vector is directly used for subsequent flow velocity prediction, ray propagation path correction, or input feature construction for thermocline adaptive loss models.

[0063] In another embodiment, assuming the exit point of a beam in a unified coordinate system is located on a reference plane, and the measurement unit is located approximately half a meter in front of the exit point, slightly to the left and slightly higher, the horizontal displacement between the measurement unit and the exit point is approximately 0.47 meters, the vertical displacement is approximately 0.21 meters, and the slight lateral offset is approximately 0.01 meters. Combining these displacements to form a spatial vector pointing towards the measurement unit, and then normalizing it according to the vector length, yields the final standardized beam direction vector. This vector clarifies the direction of each beam in three-dimensional space, providing an accurate directional reference for the spatial correction of the multi-frequency ADCP measurement unit and velocity prediction based on thermocline adaptive loss.

[0064] Of particular importance, step S4 includes: Spatial alignment processing is performed on the low-frequency background velocity field and the high-frequency detail velocity field, and the thermocline interval and non-thermocline interval are determined based on the thermocline structure identification information. Within the thermocline range, high-frequency detailed flow velocity data are used as the dominant data, and low-frequency data are interpolated to ensure data continuity. In the non-thermocline zone, low-frequency background velocity data is the main component, and high-frequency data is weighted, fused, or smoothed to avoid the influence of noise. The boundary transition processing of the fusion results of each interval is carried out to make the velocity change at the junction of the thermocline and the non-thermocline continuous, so as to obtain the full-depth reference velocity profile.

[0065] In one embodiment, low-frequency background velocity field and high-frequency detail velocity field are acquired, and the two types of data are spatially aligned so that each measurement point corresponds to the same location in a unified coordinate system. Then, based on thermocline structure identification information, the water body is divided into thermocline and non-thermocline intervals. Within the thermocline interval, high-frequency detail velocity data is used as the primary data source to ensure that rapid changes in velocity within the thermocline are accurately reflected, while interpolation compensation is applied to the low-frequency background velocity data to fill in missing or discontinuous points in the high-frequency data. Within the non-thermocline interval, low-frequency background velocity data is used as the primary data source, while high-frequency detail velocity data is smoothed or weighted and fused to reduce the impact of noise on the overall velocity field. Next, boundary transition processing is performed at the boundary between the thermocline and non-thermocline intervals. Through smooth transition or weighted mixing, the velocity changes continuously between different intervals, avoiding abrupt or discontinuous phenomena. Finally, a full-depth reference velocity profile is obtained, which can be used for subsequent adaptive loss function training and velocity prediction model construction.

[0066] In another embodiment, it is assumed that the thermocline of a water body is located between 20 and 35 meters deep. Within the thermocline range, the average high-frequency detail velocity is 0.85 m / s, and the average low-frequency background velocity is 0.72 m / s. Therefore, during fusion processing, high-frequency data is prioritized, while low-frequency data undergoes linear interpolation compensation to maintain continuity. In non-thermocline ranges (such as water depths of 0–20 meters and 35–50 meters), the average low-frequency background velocity is 0.65 m / s, and high-frequency data fluctuates significantly. Therefore, low-frequency data is prioritized, and high-frequency data undergoes weighted averaging or smoothing to effectively suppress noise. At water depths of 20 meters and 35 meters, where the thermocline and non-thermocline meet, a weighted smoothing strategy is employed to smoothly transition velocity changes, ultimately generating a complete full-depth reference velocity profile. This provides accurate and continuous input data for the velocity prediction model based on thermocline adaptive loss.

[0067] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for constructing a velocity prediction model based on thermocline adaptive loss, characterized in that, Includes the following steps: Step S1: During the shipboard navigation process, the low-frequency full-depth current velocity background field data and the medium- and high-frequency target water layer high-resolution current velocity data are acquired in a coordinated manner using a multi-frequency acoustic Doppler current profiler, and the sound velocity profile data and unified spatiotemporal reference data are acquired simultaneously to form a multi-source raw observation dataset. Step S2: Calculate sound velocity gradient data based on sound velocity profile data, identify thermocline intervals, and generate thermocline structure identification information including the top and bottom boundaries of the thermocline and the location of gradient extrema. Step S3: Under the constraint of the thermocline structure identification information, perform adaptive layered three-dimensional acoustic ray tracking calculation on each beam measurement unit of the preset multi-frequency ADCP to obtain the real spatial coordinates and the corrected beam direction vector, and correct the pre-collected original radial velocity to form a low-frequency background velocity field and a high-frequency detail velocity field. Step S4: Construct a fused reference velocity field based on the low-frequency background velocity field, the high-frequency detail velocity field, and the thermocline structure identification information. High-frequency details are preserved in the thermocline region, and low-frequency continuity is maintained in the non-thermocline region to obtain a full-depth reference velocity profile.

2. The method for constructing a velocity prediction model based on adaptive loss of a thermocline as described in claim 1, characterized in that, Step S4 also includes: A flow velocity prediction model was constructed using multi-source raw observation datasets and full-depth reference flow velocity profiles as training samples. During the training process, a depth-weighted loss function based on sound velocity gradient data and thermocline structure identification information was introduced. The flow velocity prediction model is optimized by using a depth-weighted loss function, so that the prediction error weight in the thermocline region is higher than that in the non-thermocline region. This results in a flow velocity prediction model that maintains background consistency throughout the water depth and has high-resolution representation capability in the thermocline region.

3. The method for constructing a velocity prediction model based on thermocline adaptive loss according to claim 2, characterized in that, The depth-weighted loss function constructed based on sound velocity gradient and thermocline structure identification information includes: The flow velocity observation data is extracted from the multi-source raw observation dataset and used as the model input. The full-depth reference flow velocity profile is used as the supervision label to form training samples. A flow rate prediction model is built based on training samples, and a basic loss function is constructed between the prediction results and the supervision labels. Based on the sound velocity gradient data and thermocline structure identification information, determine whether each depth location is located within the thermocline interval, and construct weight coefficients corresponding to the depth. By incorporating weighting coefficients into the basic loss function, the prediction errors at different depth locations are weighted to form a depth-weighted loss function.

4. The method for constructing a velocity prediction model based on thermocline adaptive loss according to claim 1, characterized in that, Step S2 includes: The sound velocity profile data is sorted by depth and then smoothed and filtered to obtain a continuous sound velocity profile sequence. The sound velocity gradient data is obtained by calculating the sound velocity gradient at adjacent depths based on a continuous sound velocity profile sequence. Based on the preset gradient threshold, the depth intervals that continuously meet the conditions are determined as candidate thermocline intervals, and the gradient extremum position with the largest absolute value of the sound velocity gradient is determined in each interval. Based on the extreme value of the gradient, the depth positions where the gradient drops to a preset proportional threshold are determined upwards and downwards as the top and bottom boundaries of the thermocline. Based on the top and bottom boundaries of the thermocline and the sound velocity gradient data, generate thermocline structure identification information corresponding to the depth.

5. The method for constructing a velocity prediction model based on thermocline adaptive loss according to claim 4, characterized in that, The sound velocity gradient at adjacent depths is calculated based on a continuous sequence of sound velocity profiles, yielding sound velocity gradient data including: Number each depth position according to a preset depth interval and obtain the corresponding sound velocity value; Calculate the sound velocity difference and depth difference between adjacent depth positions, and obtain the sound velocity gradient by ratio calculation; When both upper and lower neighbor points exist, the sound velocity gradients in the two directions are calculated separately and weighted to obtain the sound velocity gradient at that depth.

6. The method for constructing a velocity prediction model based on thermocline adaptive loss according to claim 4, characterized in that, Based on the gradient extremum location, the depth positions where the gradient descends to a preset proportional threshold are determined upwards and downwards, serving as the top and bottom boundaries of the thermocline, including: Within the thermocline zone, the location of the gradient extreme value with the largest absolute value of the sound velocity gradient is determined as the reference depth location. Starting from this location, search along the direction of decreasing depth. When the absolute value of the sound velocity gradient drops to a preset proportional threshold, determine the top boundary of the thermocline. Starting from this location, search along the direction of increasing depth. When the absolute value of the sound velocity gradient decreases to a preset proportional threshold, determine the bottom boundary of the thermocline.

7. The method for constructing a velocity prediction model based on adaptive loss of a thermocline according to claim 1, characterized in that, Step S3 includes: Based on the thermocline structure identification information, the layering rules along the depth direction are determined. A smaller layer thickness is used in the thermocline interval and a larger layer thickness is used in the non-thermocline interval to generate an adaptive layered structure corresponding to the depth. Under the adaptive hierarchical structure constraint, the sound ray propagation path is calculated layer by layer for each beam measurement unit of the multi-frequency ADCP, the refraction direction of each layer is determined and the horizontal and vertical displacements are accumulated. Based on the accumulated horizontal and vertical displacements, the position coordinates of each beam measurement unit in space are determined, and coordinate transformation is performed in combination with the equipment attitude and installation parameters to obtain the beam direction vector. The beam direction vector is matched with the original radial velocity, and the radial velocity is decomposed and corrected based on the beam direction vector to generate a low-frequency background velocity field and a high-frequency detail velocity field.

8. The method for constructing a velocity prediction model based on adaptive loss of a thermocline according to claim 7, characterized in that, Under the adaptive hierarchical structure constraint, the sound ray propagation path is calculated layer by layer for each beam measurement unit of the multi-frequency ADCP, the refraction direction of each layer is determined, and the horizontal and vertical displacements are accumulated, including: Under the constraint of adaptive layered structure, the refraction angle is calculated based on the incident angle and sound speed in each layer to determine the sound ray propagation direction of the current layer; Calculate the horizontal and vertical displacements corresponding to the current layer based on the propagation direction of the current layer. The horizontal and vertical displacements of each layer are accumulated layer by layer along the propagation path to obtain the cumulative displacement of the sound ray from the emission point to the target measurement unit. The sound ray propagation path data is generated based on the cumulative displacement and used as input for the spatial positioning of the measurement unit.

9. The method for constructing a velocity prediction model based on thermocline adaptive loss according to claim 7, characterized in that, Based on the accumulated horizontal and vertical displacements, the position coordinates of each beam measurement unit in space are determined. Combined with the equipment attitude and installation parameters, coordinate transformation is performed to obtain the beam direction vector, which includes: Based on the accumulated horizontal and vertical displacements, the position coordinates of each beam measurement unit in three-dimensional space are calculated; The spatial coordinates of the measurement unit are transformed with the equipment attitude and installation parameters to be unified to the reference coordinate system; Under a unified coordinate system, a spatial direction vector is constructed using the coordinates of the beam exit point and the spatial coordinates of the measurement unit, and the spatial direction vector is normalized to generate the beam direction vector.

10. The method for constructing a velocity prediction model based on adaptive loss of a thermocline according to claim 9, characterized in that, Under a unified coordinate system, a spatial direction vector is constructed using the coordinates of the beam exit point and the spatial coordinates of the measurement unit. This spatial direction vector is then normalized to generate the beam direction vector, which includes: Obtain the beam exit point coordinates and the spatial coordinates of the measurement unit under a unified coordinate system; Based on the coordinate difference between the spatial coordinates of the measurement unit and the coordinates of the beam exit point, the corresponding spatial displacement components are calculated. The spatial displacement components are combined according to the coordinate axis directions to construct a spatial direction vector pointing to the measurement unit; The spatial direction vector is calculated in terms of magnitude, and then normalized based on the magnitude to generate the beam direction vector.

Images