Wave parameter regression method and system based on structured features and physical constraints

Abstract

The application discloses a wave parameter regression method and system based on structured features and physical constraints, comprising: obtaining wave video data and preprocessing to obtain wave surface height program data; identifying wave crest and trough elevation values, extracting candidate wave height and candidate period to form a structured feature data set and generating a feature quality score; calling a structured template protocol to fill in each field and normalize to obtain a model input data sequence; inputting a wave parameter regression large model to infer an initial prediction result; loading a physical constraint correction model with a built-in wave dispersion relationship constraint term, and correcting according to the weight ratio of the data fidelity term and the physical residual term adjusted adaptively according to the feature quality score to obtain corrected wave height and period; and generating final wave parameters through confidence weighted fusion based on the feature quality score and the model confidence, so that the engineering generalization ability and the physical consistency of the regression result are improved through structured feature packaging and physical constraint adaptive correction.

Description

Technical Field

[0001] This invention relates to the field of wave monitoring and computer vision data processing technology in marine engineering, and more specifically, to a wave parameter regression method and system based on structured features and physical constraints. Background Technology

[0002] Wave parameters, especially wave height and period, are crucial foundational data for determining operational windows and assessing construction risks in marine engineering construction. Accurate and real-time acquisition of wave parameters for the target sea area is of significant engineering importance for ensuring construction safety and optimizing project scheduling.

[0003] Currently, the main technical means for acquiring wave parameters include three categories: contact measurement, remote sensing measurement, and vision-based non-contact measurement. Contact measurement, such as using buoys and pressure sensors, offers high accuracy, but its deployment and maintenance costs are high, and it is significantly limited by adverse sea conditions and deployment requirements, making it difficult to achieve large-scale, rapid, real-time coverage. Remote sensing measurements, such as radar and microwave wavemetry, have high equipment costs and high engineering integration barriers, and their effectiveness is limited in complex near-shore environments, under obstructions, or in specific sea conditions. Vision-based non-contact measurement acquires wave video using image acquisition devices, combines it with algorithms such as stereo matching and 3D reconstruction to obtain wave surface geometry information, and then extracts wave parameters. Due to its advantages such as flexible deployment, relatively low cost, and continuous monitoring capability, it is gradually becoming an important technical approach for wave monitoring on nearshore and offshore platforms.

[0004] In the field of intelligent regression of wave parameters, existing research has introduced large language models into wave prediction, utilizing the model's massive parameters and semantic understanding capabilities to infer wave height. This type of method typically inputs historical wave data or observed features into a large language model after multi-layer encoding, and the model outputs the prediction results. However, existing technical approaches have the following shortcomings: Firstly, the various features extracted from wave videos by front-end algorithms lack standardized and unified interfaces in terms of organization and expression. Feature data from different sources differ in format and semantics, making it difficult for regression models to be effectively transferred between different sea conditions and different acquisition devices, thus limiting the versatility and reusability of engineering applications. Secondly, existing methods generally adopt a purely data-driven parameter fitting paradigm, relying solely on statistical learning of training data to establish the mapping relationship between features and wave parameters, lacking the introduction and verification of the basic physical laws of wave propagation. This can easily lead to results that appear numerically reasonable but are physically invalid under complex sea conditions or changing data distribution. Furthermore, existing methods typically only output single estimates of wave height and period, lacking quantitative assessment of the reliability of regression results and means of quality traceability throughout the regression process. This leaves insufficient reliable evidence for safety decisions in engineering settings. The aforementioned shortcomings make it difficult for existing technologies to meet the high reliability and interpretability requirements of wave parameter regression in marine engineering construction. Summary of the Invention

[0005] The purpose of this invention is to provide a wave parameter regression method and system based on structured features and physical constraints to solve the above-mentioned problems existing in the prior art.

[0006] The application is as follows: On the one hand, this invention provides a wave parameter regression method based on structured features and physical constraints, including: Acquire wave video data of the target sea area, preprocess the wave video data to obtain wave surface height sequence data; The wavefront elevation sequence data is used to identify peak and trough elevation values. Within a preset time window, candidate wave height values ​​and candidate period values ​​are extracted based on the identified peak and trough elevation values ​​to obtain a structured feature data set. A feature quality score associated with the structured feature data set is then generated. The structured feature data set includes at least the peak elevation value, the trough elevation value, the candidate wave height value, and the candidate period value. A preset structured template protocol is invoked. The structured template protocol has predefined fields, including at least a wave feature value field, a unit field, a context parameter field, a physical prior cue field, and an uncertainty field. The values ​​contained in the structured feature data group are filled into the wave feature value field, the feature quality score is filled into the uncertainty field, and the values ​​filled into the wave feature value field are normalized to obtain a normalized model input data sequence. The normalized model input data sequence is input into a pre-trained wave parameter regression model to perform regression inference. The inference results are then structured and parsed according to a preset output field format to obtain initial prediction result data. The initial prediction result data includes at least the initial wave height value, the initial period value, and the model confidence. A pre-built physical constraint correction model is loaded, and a physical consistency correction operation is performed based on the initial prediction result data and the feature quality score to obtain the corrected wave height value and the corrected period value. The physical constraint correction model has a built-in wave dispersion relation constraint term, and when performing the physical consistency correction operation, the weight ratio of the data fidelity term and the physical residual term is adaptively adjusted according to the feature quality score. Based on the feature quality score and the model confidence level, the initial prediction result data, the corrected wave height value, and the corrected period value are subjected to confidence-weighted fusion to generate the final wave parameter data.

[0007] Furthermore, the preprocessing of the wave video data to obtain wavefront height sequence data includes: The wave video data is analyzed frame by frame, and the pixels in each frame are physically mapped according to the acquisition parameters of the wave video data to generate wave surface height sequence data arranged in chronological order; the acquisition parameters include at least one of the installation height of the image acquisition device, pitch angle and focal length. Perform at least one of image stabilization and noise reduction operations on the wavefront high-program sequence data.

[0008] Furthermore, the structured feature data set includes: Within a preset time window, the wavefront elevation sequence data at the same spatial location or in the same cross-sectional direction are acquired. The maximum value of the wavefront elevation sequence data within the preset time window is determined as the wave crest elevation value, and the minimum value of the wavefront elevation sequence data within the preset time window is determined as the wave trough elevation value. The difference between the peak elevation value and the trough elevation value is determined as the candidate wave height value; Perform a Fast Fourier Transform (FFT) on the wavefront high-order sequence data to obtain the spectrum. Determine the frequency corresponding to the main peak in the spectrum as the FFT main frequency, and determine the reciprocal of the FFT main frequency as the candidate period value.

[0009] Furthermore, generating feature quality scores associated with the structured feature data set specifically includes: The feature quality score is obtained by weighting and summing at least two of the following quality indicators: effective pixel ratio of wavefront, consistency index between wave crest elevation and wave trough elevation, signal-to-noise ratio, occlusion ratio, and window stability. The weighted sum is then subjected to Sigmoid function mapping or truncation normalization processing. The value range of the feature quality score is a preset interval.

[0010] Furthermore, obtaining the normalized model input data sequence includes: For each value filled into the wave feature value field, a normalization mapping is performed according to the preset mean and preset standard deviation corresponding to each value to obtain the normalized value corresponding to each value. The original values ​​of each item in the wave feature value field, the normalized values ​​corresponding to each item, the unit of measurement corresponding to the unit field, the context parameters corresponding to the context parameter field, the physical prior information corresponding to the physical prior cue field, and the feature quality score corresponding to the uncertainty field are organized according to the structured template protocol to obtain the normalized model input data sequence.

[0011] Furthermore, the initial prediction result data also includes output basis information, which is used to record the key field identifiers in the structured feature data group referenced in this inference.

[0012] Furthermore, the wave dispersion relation constraint term Constructed based on the following physical residuals:

[0013] Where ω is the angular frequency, determined by the period value during the correction process; k is the wave number, determined by the wave wavelength, k = 2π / λ; h is the water depth of the target sea area; g is the gravitational acceleration; tanh is the hyperbolic tangent function; The wave wavelength λ is determined based on the spatial distance between adjacent wave crests in the wavefront height sequence data, or the wave number k is obtained as a latent variable by the physical constraint correction model during the training process.

[0014] Furthermore, the adaptive adjustment of the weight ratio between the data fidelity term and the physical residual term based on the feature quality score includes: Construct the corrective loss function:

[0015] Among them, L data For data fidelity items, , and These are the corrected wave height value and the corrected period value, respectively. and The initial wave height and the initial period value are respectively, λ data The weighting coefficient for the data fidelity item is set to the feature quality score q. L phys For physical residuals, , λ phys The weighting coefficient for the physical residual term is 1 - q; L reg λ is the regularization term; reg The weight coefficients for the regularization term; When the feature quality score q is higher, the data fidelity item L... data The larger the weight of the feature quality score q, the lower the physical residual term L. phys The greater the weight, the better.

[0016] Furthermore, the generation of the final wave parameter data includes: The fusion coefficient is calculated based on the feature quality score and the model confidence score, wherein the fusion coefficient α is specifically:

[0017] in, LLM The confidence level of the model is defined by η1 and η2, which are preset weight coefficients, and clip indicates that the calculation results are restricted to the range of 0 to 1. The initial prediction data, the corrected wave height value, and the corrected period value are weighted and summed using the fusion coefficients to obtain the final wave height value. final =α· LLM +(1-α)· The final period value is final =α· LLM +(1-α)· ; The final wave parameter data includes at least the final wave height value and the final period value.

[0018] On the other hand, the present invention provides a wave parameter regression system based on structured features and physical constraints, comprising: The data preprocessing unit is used to preprocess the acquired wave video data of the target sea area and output wave surface height sequence data. The feature extraction and quality assessment unit is used to identify the peak elevation and trough elevation values ​​from the wavefront height sequence data, extract candidate wave height values ​​and candidate period values ​​to form a structured feature data set, and generate a feature quality score associated with the structured feature data set. The prompt word encapsulation unit has a built-in preset structured template protocol, which is used to fill and normalize the structured feature data group and the feature quality score according to the structured template protocol, and output a normalized model input data sequence. The large model regression unit has a built-in pre-trained wave parameter regression large model, which is used to receive the normalized model input data sequence, perform regression inference and output initial prediction result data including initial wave height value, initial period value and model confidence. The physical constraint correction unit has a built-in wave dispersion relation constraint term. It is used to receive the initial prediction result data and the feature quality score, perform physical consistency correction on the initial prediction result data under the wave dispersion relation constraint term, and adaptively adjust the weight ratio of the data fidelity term and the physical residual term during the correction process according to the feature quality score, and output the corrected wave height value and the corrected period value. The confidence fusion unit is used to perform weighted fusion of the initial prediction result data, the corrected wave height value, and the corrected period value based on the feature quality score and the model confidence, and output the final wave parameter data.

[0019] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects: By encapsulating wave feature values, units of measurement, context parameters, physical prior information, and feature quality scores into a standardized model input data sequence using a pre-defined structured template protocol, the differences in format and semantics between feature data from different sources are overcome, significantly improving the transferability and engineering reusability of the wave parameter regression method across different application scenarios. A physical constraint correction model with built-in wave dispersion relation constraints is introduced based on the initial prediction results of the large wave parameter regression model. The weight ratio of the data fidelity term and the physical residual term is adaptively adjusted according to the feature quality score. When the observed feature quality is high, the data-driven inference advantage of the large model is preserved; when the observed feature quality is low, physical laws are used for fallback correction. This effectively overcomes the problem of purely data-driven methods easily producing physically inconsistent results under complex sea conditions, improving the physical rationality and engineering credibility of the regression results. By generating feature quality scores associated with the structured feature data set and using them together with the model confidence score for confidence-weighted fusion, the final output wave parameters comprehensively assess both the input data quality and the output model confidence, providing a more robust and reliable reference for engineering decisions. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a wave parameter regression method based on structured features and physical constraints provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a wave parameter regression system based on structured features and physical constraints provided in an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] It should be noted that many specific details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention may have other embodiments, and therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0023] refer to Figure 1 As shown, this embodiment provides a wave parameter regression method based on structured features and physical constraints. The steps of the method are described in detail below.

[0024] Step 1: Acquire wave video data and preprocess it to obtain wavefront height sequence data. First, wave video data of the target sea area is acquired. In this embodiment, the wave video data can be time-series image frames continuously acquired from the target sea area using an image acquisition device. For example, a binocular image acquisition device can be used to acquire video of the target sea area. The binocular image acquisition device can simultaneously acquire wave images from two different perspectives, providing richer geometric information for subsequent three-dimensional wave surface reconstruction and physical space mapping. The wave video data includes at least multiple consecutive wave images and associated acquisition parameters. The acquisition parameters may include at least one of the following: the installation height of the image acquisition device, the pitch angle, and the focal length.

[0025] After acquiring wave video data, the data is preprocessed. The preprocessing process specifically includes: image analysis of each wave image in the multi-frame wave image, physical spatial mapping of the pixels in each wave image according to the acquisition parameters, and generating wave surface height sequence data arranged in chronological order.

[0026] Specifically, image analysis refers to reading the pixel data of each frame of a wave image and obtaining the position and grayscale value of each pixel in the image coordinate system. Based on the grayscale or color characteristics of the pixels in the wave image, as well as the texture and brightness variation patterns of the wave surface in the image, pixels belonging to the wave surface region can be identified and distinguished.

[0027] Physical space mapping refers to converting the pixel coordinates of pixels belonging to the wavefront region in a two-dimensional image coordinate system into their actual coordinates in three-dimensional physical space. The installation height of the image acquisition device determines its vertical distance relative to the horizontal plane; the pitch angle determines the tilt angle of the optical axis of the image acquisition device relative to the horizontal plane; and the focal length determines the correspondence between pixel scale and physical scale. Based on these acquisition parameters, using the imaging geometry model in photogrammetry, pixels belonging to the wavefront region in each frame of a wave image can be mapped to coordinate points in three-dimensional physical space, where the coordinate components perpendicular to the horizontal plane are the physical elevation values ​​of the wavefront. Arranging the physical elevation values ​​at the same spatial location or along the same cross-sectional direction in multiple wave images in chronological order forms wavefront elevation sequence data.

[0028] Through the aforementioned physical space mapping process, the pixel information of the wave surface region in each frame of the wave image is converted into the spatial distribution of wave surface elevation at the corresponding time. Multiple consecutive frames of wave surface elevation spatial distribution constitute wave surface elevation sequence data. This wave surface elevation sequence data reflects the change in wave elevation over time within the observation period and serves as the data foundation for subsequent structured feature extraction.

[0029] Preferably, the preprocessing process may further include performing at least one of image stabilization and denoising operations on the wavefront height sequence data. Image stabilization is used to eliminate the influence of image jitter caused by image acquisition device shaking or platform vibration on the wavefront height measurement results. Commonly used image stabilization methods include electronic image stabilization algorithms based on feature point matching, which estimate inter-frame motion parameters and perform motion compensation by matching the positions of feature points between adjacent frames. Denoising is used to remove high-frequency noise from the wavefront height sequence data. Commonly used denoising methods include Gaussian filtering, median filtering, or wavelet thresholding. Through image stabilization and denoising operations, wavefront height sequence data with higher signal-to-noise ratio and more stable sequences can be obtained, providing a reliable data foundation for subsequent structured feature extraction.

[0030] Step 2: Extract structured feature data sets from the wavefront height sequence data and generate feature quality scores. After obtaining the wavefront height sequence data, the peak elevation value and trough elevation value are identified in the wavefront height sequence data. Within a preset time window, candidate wave height values ​​and candidate period values ​​are extracted based on the identified peak elevation values ​​and trough elevation values ​​to obtain a structured feature data set, and a feature quality score associated with the structured feature data set is generated.

[0031] In this embodiment, the specific method for obtaining the structured feature data group is as follows: First, within a preset time window, wave surface height sequence data are acquired at the same spatial location or along the same cross-sectional direction. The length of the preset time window can be set according to the typical wave cycle range; for example, it can be set to 30 to 120 seconds in nearshore waters to ensure that a sufficient number of complete wave cycles are included within the window.

[0032] Then, the maximum value of the wavefront elevation sequence data within a preset time window is determined as the peak elevation value, and the minimum value of the wavefront elevation sequence data within the preset time window is determined as the trough elevation value. It should be noted that multiple peaks and multiple troughs may be identified within a single time window. In this embodiment, the peak elevation value refers to the elevation value corresponding to the global maximum value of the wavefront elevation sequence data within the window, and the trough elevation value refers to the elevation value corresponding to the global minimum value of the wavefront elevation sequence data within the window.

[0033] Next, the difference between the wave crest elevation and the wave trough elevation is determined as the candidate wave height value. The candidate wave height value reflects the maximum variation in wave surface elevation within this time window.

[0034] Simultaneously, a Fast Fourier Transform (FFT) is performed on the wavefront elevation sequence data to obtain the spectrum. The FFT converts the time-domain wavefront elevation signal into a frequency-domain signal, with each peak in the spectrum corresponding to a different frequency component of the wave. The frequency corresponding to the main peak in the spectrum is determined as the FFT dominant frequency, which represents the main frequency component of the wave within that time window. The reciprocal of the FFT dominant frequency is determined as a candidate period value. The candidate period value reflects the main periodic characteristics of the wave.

[0035] Therefore, the structured feature data set includes at least peak elevation, trough elevation, candidate wave height, candidate period, and FFT dominant frequency. These features describe the morphology and motion characteristics of waves from different dimensions.

[0036] While extracting the structured feature data set, this implementation also generates a feature quality score associated with the structured feature data set. The feature quality score is used to quantify the reliability of the extracted structured feature data set, providing a basis for subsequent physical constraint correction and confidence fusion.

[0037] In this embodiment, generating a feature quality score specifically includes: weighting and summing at least two quality indicators among the effective pixel ratio of the wavefront, the consistency index between the peak elevation value and the valley elevation value, the signal-to-noise ratio, the occlusion ratio, and the window stability, and performing Sigmoid function mapping or truncation normalization on the weighted summation result to obtain the feature quality score. The value range of the feature quality score is a preset interval, usually a closed interval between 0 and 1.

[0038] The methods for obtaining the above-mentioned quality indicators are described below. It should be noted that the following methods are all conventional technical means in this field, and this embodiment will not elaborate on them.

[0039] The effective pixel ratio of the wavefront is obtained through statistical analysis of pixels within the wavefront region. Specifically, for each frame of wave image with an identified wavefront region, the number of pixels whose grayscale values ​​fall within a reasonable dynamic range is counted. This number is then compared to the total number of pixels within the wavefront region; the resulting ratio is the effective pixel ratio of the wavefront. The reasonable dynamic range of grayscale values ​​can be preset, for example, excluding grayscale ranges that are close to pure white (reflective saturation) and close to pure black (shadow occlusion). A higher effective pixel ratio indicates better image quality in the wavefront region and less impact from reflections or occlusions.

[0040] The consistency index for peak and trough elevation values ​​is obtained by comparing the identification results of peak and trough elevation values ​​across multiple time windows or using various detection algorithms. In one implementation of this method, the current time window can be shifted forward and backward by several frames to form multiple partially overlapping time windows. Peak and trough elevation values ​​are identified in each overlapping window, and the consistency of the position and number of peak elevation values ​​in the identification results of each window is compared. The degree of consistency can be measured by calculating the standard deviation of the number of peak elevation values ​​identified in each window, or by calculating the mean deviation of the corresponding peak elevation value positions. The smaller the standard deviation or the smaller the deviation, the higher the identification consistency.

[0041] Signal-to-noise ratio (SNR) is obtained by power spectrum analysis of wavefront elevation sequence data. The wavefront elevation sequence data is transformed from the time domain to the frequency domain using a Fast Fourier Transform (FFT). In the frequency spectrum, the signal components with energy concentrated within the wave frequency band are separated from the noise components distributed throughout the band. The ratio of the signal power within the wave frequency band to the noise power in the remaining band is calculated; this is the SNR. A higher SNR indicates a larger proportion of useful signal components in the wavefront elevation signal, and thus better signal quality.

[0042] The occlusion ratio is obtained by performing target detection or motion segmentation on the wave surface region in the wave image. In this embodiment, the occlusion region, such as a ship, platform structural component, or splashing water mist, can be identified in the wave image using the difference information between adjacent frames or a pre-trained target detection model. The occlusion ratio is the percentage of pixel area occupied by the occlusion region in the wave surface region. A higher occlusion ratio indicates worse observation conditions and a smaller effective wave surface region for wave feature extraction.

[0043] Window stability is determined by calculating the degree of statistical drift in the wavefront elevation sequence data within a preset time window. Specifically, the preset time window is divided into several sub-windows, and the mean and variance of the wavefront elevation sequence data within each sub-window are calculated. Then, the degree of variation in the mean and variance among the sub-windows is calculated. The smaller the degree of variation, the more stable the statistical characteristics of the wavefront elevation data within the window, and the better the window stability.

[0044] After obtaining the specific values ​​of each quality indicator using the above method, a weighted sum is performed based on the pre-set weight coefficients for each quality indicator. These weight coefficients can be determined based on historical experience data or by tuning through a validation dataset. The weighted sum is then mapped using a Sigmoid function or truncated normalization to map the overall quality score to a preset range between 0 and 1, thus obtaining the feature quality score q. A feature quality score closer to 1 indicates higher quality and stronger reliability of the structured feature data set; a feature quality score closer to 0 indicates lower quality and weaker reliability of the structured feature data set.

[0045] Step 3: Use the structured template protocol to encapsulate and normalize the data, obtaining a normalized model input data sequence. After obtaining the structured feature data set and the feature quality score, a preset structured template protocol is invoked for encapsulation processing. The structured template protocol has predefined fields, which include at least a wave feature value field, a unit field, a context parameter field, a physical prior cue field, and an uncertainty field.

[0046] In this embodiment, the structured template protocol is a predefined data organization format, which can be implemented using JSON, key-value pairs, or other parsable structured formats. The functions and content of each field in the structured template protocol are as follows: The wave feature value field is used to carry the various values ​​contained in the structured feature data group. In this embodiment, the structured feature data group includes wave crest elevation value, wave trough elevation value, candidate wave height value, candidate period value, and FFT dominant frequency. All five values ​​are filled into the wave feature value field.

[0047] The unit field is used to carry the unit of measurement corresponding to each feature value in the wave feature value field. Specifically, the unit of measurement for wave crest elevation and wave trough elevation is meters, the unit of measurement for candidate wave height is meters, the unit of measurement for candidate period is seconds, and the unit of measurement for FFT master frequency is Hertz. This unit of measurement information can be directly determined from the physical definition of each feature value and stored corresponding to the position of each feature value in the wave feature value field.

[0048] The context parameter field is used to carry context parameters during the data acquisition process. Context parameters can be directly obtained from the wave video data acquisition configuration information. In this embodiment, context parameters may include at least one of the following: sampling duration, frame rate, observation distance, geometric parameters of the image acquisition device, and sea state label. Sampling duration refers to the continuous acquisition time of wave video data; frame rate refers to the number of wave image frames acquired per unit time; observation distance refers to the horizontal distance from the image acquisition device to the target sea surface; the geometric parameters of the image acquisition device may include installation height, pitch angle, and focal length; the sea state label can be a label indicating the current sea state level based on meteorological and oceanographic forecast data or manual judgment. Context parameters provide background information for data acquisition to the large model, assisting it in making more accurate judgments based on acquisition conditions during regression inference.

[0049] The physical prior information field is used to carry prior information related to the physical laws of waves. In this embodiment, the physical prior information can be obtained from external data sources or pre-input by operators based on the actual engineering situation. The physical prior information may include whether the water depth of the target sea area is known and its specific value, whether deep-water approximation conditions are applicable, and the wave type. The water depth refers to the seawater depth of the target sea area, which can be obtained from nautical charts or on-site measurement data; the deep-water approximation condition refers to the ability to simplify the dispersion relationship to a deep-water form when the ratio of water depth to wave wavelength is greater than a certain threshold; the wave type refers to whether the current wave belongs to wind waves, swells, or mixed waves, etc. The physical prior information provides a basic physical background reference for subsequent physical constraint correction.

[0050] The uncertainty field is used to carry the feature quality score. The feature quality score is then filled into the uncertainty field. The feature quality score characterizes the overall quality level and reliability of the structured feature data set.

[0051] After filling in each field, the values ​​filled into the wave feature value field are normalized. It should be noted that in this implementation, normalization only applies to the values ​​filled into the wave feature value field; the contents of the unit field, context parameter field, and physical prior cue field are not normalized. This is because the values ​​in the wave feature value field are extracted in real-time from wave surface height sequence data, and their ranges vary significantly under different sea states and acquisition conditions. Normalization is needed to eliminate the adverse effects of scale differences on subsequent large-scale model regression inference. The parameters in the context parameter field are acquisition configuration information with relatively fixed ranges, and the information in the physical prior cue field is mainly classification or labeling information. Preserving the contents of these fields in their original form makes it easier for the large model to directly understand the acquisition conditions and physical background.

[0052] The normalization process is as follows: For each value filled into the wave feature value field, a normalization mapping is performed based on the preset mean and preset standard deviation corresponding to each value to obtain the normalized value for each value. Taking candidate wave height and candidate period values ​​as examples, the preset mean and preset standard deviation can be obtained through statistical analysis of a large amount of historical wave observation data. For example, for candidate wave height, the mean and standard deviation of wave height values ​​in historical observation data can be used as the preset mean and preset standard deviation; for candidate period values, the mean and standard deviation of period values ​​in historical observation data can be used as the preset mean and preset standard deviation. The formula for normalization mapping is: the normalized value equals the original value minus the preset mean divided by the preset standard deviation. Wave crest elevation, wave trough elevation, and FFT dominant frequency also have their corresponding preset mean and preset standard deviation, and the normalization mapping method is the same as described above.

[0053] After normalization, the original values ​​of each item in the wave feature value field, the normalized values ​​corresponding to each item, the units of measurement corresponding to the unit field, the context parameters corresponding to the context parameter field, the physical prior information corresponding to the physical prior cue field, and the feature quality score corresponding to the uncertainty field are organized according to the field order and format defined in the structured template protocol to obtain the normalized model input data sequence.

[0054] The normalized model input data sequence is a structured data package that includes both the original and normalized values ​​of wave features, as well as metadata such as unit information, acquisition context information, physical prior information, and feature quality scores. This unified encapsulation of feature data and metadata makes the normalized model input data sequence self-descriptive and transferable, adaptable to wave feature data generated under different sea conditions and image acquisition devices, and provides a standardized input interface for the generalization inference of subsequent large-scale wave parameter regression models.

[0055] Step 4: Input the normalized model input data sequence into the wave parameter regression model to obtain the initial prediction results data. The obtained normalized model input data sequence is input into a pre-trained large-scale wave parameter regression model to perform regression inference. The large-scale wave parameter regression model is a neural network model trained with a large amount of wave observation data, which has the ability to infer and output wave parameters based on the input wave features and their contextual information.

[0056] In this embodiment, the large-scale wave parameter regression model can be an industry-specific large-scale model trained based on large-scale wave observation data, or a general-purpose large-scale model that has been fine-tuned. Pre-training means that the network structure and weight parameters of the large-scale model have been determined and stored in the computing device before this step. The training process of the large-scale wave parameter regression model can adopt a supervised learning approach: collect a large number of training samples containing wave videos or wave height sequences, each labeled with the actual wave height and period values; convert the training samples into a normalized model input data sequence according to the processing flow of steps one to three, input it into the large-scale model to be trained, and the large-scale model outputs the predicted wave height and period values; calculate the loss between the predicted values ​​and the labeled values, and update the model parameters through the backpropagation algorithm; repeat the above process until the model converges, obtaining the trained large-scale wave parameter regression model.

[0057] After the normalized model input data sequence is fed into the large model, regression inference is performed. Regression inference refers to the process by which the large model performs forward propagation calculations on the input data: the input data sequentially passes through the weight matrix multiplication and addition operations and nonlinear activation function transformations of each network layer of the large model, and finally obtains the inference result at the output layer.

[0058] It should be noted that the regression inference in this embodiment is a numerical regression task of wave parameters. That is, the large model directly outputs the specific values ​​of wave height and period based on the input feature data, rather than classifying wave types or determining sea state levels. This end-to-end numerical regression capability is a key feature that distinguishes the large wave parameter regression model from general large language models.

[0059] After the calculation is completed, the inference results are structured and parsed according to the preset output field format to obtain the initial prediction result data. The preset output field format refers to the predefined organization of the output data of the large model. This format can constrain the output content of the large model to include specific fields and data types, which facilitates automated processing in subsequent steps and ensures the stable operation of the system during engineering implementation.

[0060] In this embodiment, the preset output field format can be a fixed-pattern data structure definition. For example, the output must contain a set of key-value pairs with field names "wave height value", "cycle value", "confidence level", and "output basis", and the data types of each field are specified as floating-point number, floating-point number, floating-point number, and string, respectively. The structured parsing process is the process of extracting the corresponding values ​​of each field from the original output of the large model according to the preset format: performing field matching and type conversion on the output text or data structure to extract the values ​​of each field.

[0061] The initial prediction results data include at least the initial wave height, the initial period value, and the model confidence score. The initial wave height is a preliminary estimate of the wave height obtained by the large model based on the input structured feature data set, and the initial period value is a preliminary estimate of the period obtained by the large model based on the input structured feature data set.

[0062] Model confidence is a self-assessment value by which a large model evaluates the credibility of its inference result. In this implementation, model confidence can be obtained through several methods. In one approach, during training, the large model learns not only the output wave height and period value but also an output confidence estimate. This confidence estimate is trained by correlating it with the large model's prediction error, ensuring that a higher confidence level corresponds to a statistically smaller prediction error. In another approach, a dedicated confidence estimation branch network can be set after the large model's output layer. This branch network comprehensively evaluates the reliability of the inference based on the feature representations of the large model's intermediate layers and the numerical range of the final output, and outputs a confidence value. In yet another approach, an ensemble learning method can be used, determining the confidence level by observing the stability of the output after multiple inferences or applying small perturbations to the input data—the more stable the output, the higher the confidence level. Model confidence typically ranges from 0 to 1. A value closer to 1 indicates greater confidence in the inference result, while a value closer to 0 indicates greater uncertainty in the inference result.

[0063] It is important to note that model confidence and feature quality score are two different dimensions of reliability metrics. Feature quality score measures the quality of the input data, i.e., the reliability of the front-end feature extraction; model confidence measures the quality of the large model's inference process, i.e., the large model's confidence in its own output results. Both will work together in the subsequent confidence-weighted fusion step to comprehensively ensure the reliability of the final wave parameter data from both the input and output dimensions.

[0064] Preferably, the initial prediction results data may also include output basis information. This output basis information records the identifiers of key fields in the structured feature data set referenced in this regression inference and their degree of contribution. For example, the output basis information may indicate that the large model has the highest dependence weight on candidate wave height values, the second highest reference weight on the FFT dominant frequency, and a relatively low direct dependence on wave peak and trough elevation values ​​in this inference. The output basis information provides a degree of interpretability to the black-box inference process of the large model, facilitating result traceability and quality analysis in engineering applications.

[0065] Step 5: Load the physical constraint correction model and perform physical consistency correction operation. Although large wave parameter regression models possess strong nonlinear fitting and generalization capabilities, in complex sea conditions or with varying data distributions, the output of purely data-driven large models may exhibit physical inconsistencies. For example, the output wave height and period values ​​may not satisfy fundamental physical laws such as wave dispersion relations. To address this issue, this embodiment introduces a physical constraint correction model to perform physical consistency correction on the initial prediction results of the large model.

[0066] The physical constraint correction model incorporates wave dispersion relation constraints. The wave dispersion relation describes the physical relationship between wave angular frequency, wave number, and water depth. In marine engineering, for small-amplitude surface gravity waves, the linear dispersion relation is: the square of the angular frequency equals the gravitational acceleration multiplied by the wave number, then multiplied by the hyperbolic tangent function acting on the product of the wave number and water depth. In this embodiment, the wave dispersion relation constraint term r... disp Constructed based on the following physical residuals:

[0067] Where ω is the angular frequency, determined by the period value during the correction process, specifically ω = 2π / T, where T is the period value during the correction process. k is the wave number, determined by the wave wavelength, specifically k = 2π / λ. h is the water depth of the target sea area. g is the gravitational acceleration. tanh is the hyperbolic tangent function. The above physical residual r disp It measures the degree of deviation between the corrected wave height and period values ​​and the physical laws. The closer the residual is to zero, the more the correction result conforms to the physical laws.

[0068] In this embodiment, the physical constraint correction operation is an iterative optimization process. At the start of correction, the initial period value from the initial prediction data is used as the starting value for the period variable, and the initial wave height value is used as the starting value for the wave height variable. In each iteration, the correction optimizer updates the period and wave height variables based on the gradient information of the correction loss function. The updated period and wave height values ​​continue to participate in the loss function calculation for the next iteration. During this iteration, the current value of the period variable in each iteration is the "period value during the correction process." As the iteration progresses, the period value during the correction process is gradually adjusted from the initial period value. When the correction loss function converges or reaches the preset number of iterations, the iteration terminates, and the period value at this point is the corrected period value.

[0069] Therefore, the relationship between the "period value during the calibration process," the initial period value, and the calibrated period value can be summarized as follows: the initial period value is the preliminary result of the large model regression inference and the starting point for physical constraint calibration; the period value during the calibration process is an intermediate variable that is continuously updated during the optimization process, with each iteration generating a current period value; the calibrated period value is the final output result after optimization convergence. In the physical residual r... disp In the expression, T in angular frequency ω=2π / T refers to the period value in the correction process. In each iteration, the current period value of that iteration is used for calculation.

[0070] In practical applications, the wave wavelength λ can be obtained in various ways. In one implementation of this embodiment, the wave wavelength λ is determined based on the spatial distance between adjacent wave crests in the wavefront elevation sequence data. Specifically, in step two above, the elevation values ​​of each wave crest and their corresponding time and spatial location have been identified. The distance between the spatial locations corresponding to two adjacent wave crest elevation values ​​is one wave wavelength. By averaging the distances between multiple adjacent wave crests within a window, the estimated wave wavelength λ corresponding to that time window can be obtained. In another implementation of this embodiment, when the wave wavelength is difficult to accurately estimate from the wavefront elevation sequence data (e.g., the wavefront area is not large enough, or the number of identifiable wave crests is insufficient), the wave number k can be used as a latent variable and learned by the physical constraint correction model during training. Latent variables refer to variables that are not directly observed but participate in the calculation within the model. Through training on a large amount of data, the physical constraint correction model can learn a reasonable value pattern for the wave number k.

[0071] When performing physical consistency correction, the physical constraint correction model adaptively adjusts the weight ratio of data fidelity terms to physical residual terms based on the feature quality score.

[0072] In the calibration process, the physical constraint correction model adjusts the wave height and period values ​​by optimizing a correction loss function. This ensures that the corrected result neither deviates excessively from the initial predictions of the larger model nor fails to satisfy the physical constraints of the wave dispersion relation. The correction loss function is constructed as follows:

[0073] Among them, L data For data fidelity items, its expression is: . and These are the corrected wave height and the corrected period, respectively. and These represent the initial wave height and the initial period value, respectively. The data fidelity term measures the degree of deviation between the corrected result and the initial prediction result; the smaller the deviation, the closer the corrected result is to the inference result of the large model.

[0074] L phys The physical residual term is expressed as follows: That is, the wave dispersion relation constraint term r disp The square of . The physical residual term measures the degree of agreement between the correction result and the physical law.

[0075] L regλ is a regularization term used to constrain the range of corrected wave height and period values, ensuring the reasonableness of the output results. For example, it constrains wave height values ​​to be non-negative and period values ​​to be within a reasonable range. The specific form of the regularization term can be the L2 norm of the parameters or other regularization forms commonly used in this field. reg The weight coefficient for the regularization term is a preset positive constant.

[0076] λ data λ is the weighting coefficient for the data fidelity item, and its value is taken as the feature quality score q. phys The weighting coefficient for the physical residual term is 1 - q.

[0077] By setting the weighting coefficients as described above, the technical effect of adaptively adjusting the weight ratio of the data fidelity term and the physical residual term based on the feature quality score is achieved. The working mechanism under different feature quality score values ​​is explained in detail below.

[0078] When the feature quality score q is high (e.g., q is close to or equal to 1), it indicates that the structured feature data set is of good quality and has high reliability. In this case, the wave video data acquisition conditions are good, the wavefront region is clear and unobstructed, the identification of wave crest and trough elevation values ​​is accurate and reliable, the signal-to-noise ratio is high, and the window stability is good. data The value is close to 1 and λ phys When the value is close to 0, the data fidelity term dominates the correction loss function, and the correction result is more inclined to believe the initial prediction of the large model, making only minor adjustments to physical consistency. This mechanism ensures that the data-driven capability of the large model is fully utilized under high-quality data conditions, and the fitting accuracy of the large model is not lost due to excessive intervention of physical constraints.

[0079] When the feature quality score q is low (e.g., q is close to or equal to 0), it indicates poor quality and low reliability of the structured feature data set. In this case, wave video data may be affected by factors such as reflective saturation, water mist obstruction, platform vibration, or severe sea conditions, leading to inaccurate wave surface elevation measurements, significant errors in identifying wave crest and trough elevations, low signal-to-noise ratio, and poor window stability. In this situation, λ data The value is close to 0 and λ phys With a value close to 1, the physical residual term dominates the correction loss function, and the correction result tends to follow the constraints of physical laws, using physical laws to provide a fallback correction for potentially inaccurate initial predictions. This mechanism ensures that even with low-quality data, and even if the initial predictions of a large model are unreliable, the final wave parameter output still satisfies basic physical laws, and physically invalid values ​​will not appear.

[0080] When the feature quality score q is at an intermediate level (e.g., q is between 0.4 and 0.6), the weights of the data fidelity term and the physical residual term both account for a certain proportion, and the correction process achieves a balance between trusting the large model and following physical laws.

[0081] The training process of the physical constraint correction model can utilize a pre-collected wave observation dataset. Each sample in the training dataset contains a set of structured feature data extracted from wave video data, along with the corresponding true wave height and true period value. During the training phase, the structured feature data set of the sample is input into the wave parameter regression model to obtain initial prediction results. Then, the initial prediction results and the feature quality scores corresponding to the samples are input into the physical constraint correction model, and the parameters of the physical constraint correction model are optimized by minimizing the correction loss function. After training, the physical constraint correction model can perform physical consistency correction on new initial prediction results during the actual inference phase.

[0082] After processing by the physical constraint correction model, the corrected wave height and period values ​​are finally obtained. The corrected wave height and period values ​​retain the generalization advantage of large model data-driven regression, and have been verified and adjusted according to physical laws, thus having higher physical consistency and engineering credibility.

[0083] Step Six: Perform confidence-weighted fusion to generate the final wave parameter data. After obtaining the corrected wave height and corrected period values, the initial prediction data and the corrected wave height and corrected period values ​​are weighted and fused based on the feature quality score and model confidence to generate the final wave parameter data.

[0084] In this embodiment, the specific process of confidence-weighted fusion is as follows: First, based on the feature quality score q and the model confidence score ĉ LLM Calculate the fusion coefficient α. The fusion coefficient determines the relative weight of the initial prediction result and the physical correction result in the final output. The formula for calculating the fusion coefficient α is:

[0085] Here, η1 and η2 are preset weight coefficients. η1 and η2 can be determined through parameter tuning on the validation dataset. For example, they can be initially set to η1=0.6 and η2=0.4, or adjusted according to the relative importance of feature quality scores and model confidence in the actual application scenario. `clip` restricts the calculation results to the range of 0 to 1, ensuring that the fusion coefficients are always valid values.

[0086] As can be seen from the above formula, the fusion coefficient α comprehensively considers two dimensions of credibility information: the feature quality score q measures the quality of the feature data used for regression from the input end, and the model confidence score... LLM The output measures the large model's confidence in its own inference results. The weighted sum of both factors, after truncation, yields a comprehensive fusion coefficient.

[0087] Then, the initial prediction data, the corrected wave height value, and the corrected period value are weighted and summed using the fusion coefficient.

[0088] The formula for calculating the final wave height is: final =α· LLM +(1-α)· ; The formula for calculating the final period value is: final =α· LLM +(1-α)· ; Through the weighted summation described above, when both the feature quality score and model confidence are high, the fusion coefficient α is close to 1, and the final result is more biased towards the initial prediction. When the feature quality score or model confidence is low, the fusion coefficient α is smaller, and the final result is more biased towards the result after physical constraint correction. This fusion mechanism ensures that the final output wave parameters have high reliability and physical rationality under different data quality and model confidence levels.

[0089] The final wave parameter data includes at least the final wave height and the final period. These two parameters are core wave parameters that are directly relevant to marine engineering construction and can be used for subsequent determination of operational windows and risk assessment.

[0090] refer to Figure 2 As shown, this embodiment also provides a wave parameter regression system based on structured features and physical constraints. This system can be deployed in a computing device on a nearshore or offshore platform to execute the wave parameter regression method described above. The system includes the following functional units: The data preprocessing unit is used to preprocess the acquired wave video data of the target sea area and output wave surface height sequence data. The data preprocessing unit receives wave video data, performs physical space mapping and image quality enhancement processing according to the acquisition parameters, and converts the original image frames into wave surface height sequence data arranged in chronological order.

[0091] The feature extraction and quality assessment unit is used to identify peak and trough elevation values ​​from wavefront height sequence data, extract candidate wave height values ​​and candidate period values ​​to form a structured feature data set, and generate a feature quality score associated with the structured feature data set. The specific functions of the feature extraction and quality assessment unit include: analyzing wavefront height sequence data within a preset time window, identifying peak and trough elevation values ​​through extreme value detection, calculating candidate wave height values, obtaining the FFT dominant frequency and candidate period values ​​through Fast Fourier Transform, and generating a feature quality score by integrating multiple quality indicators.

[0092] The cue word encapsulation unit has a built-in preset structured template protocol. This protocol is used to fill in and normalize the structured feature data set and feature quality scores according to the template protocol, outputting a normalized model input data sequence. The cue word encapsulation unit maintains a preset structured template protocol that defines standardized field structures for wave feature value fields, unit fields, context parameter fields, physical prior cue fields, and uncertainty fields. This allows it to encapsulate wave feature data from different acquisition devices and sea conditions into a unified format model input data sequence.

[0093] The large model regression unit contains a pre-trained wave parameter regression model. It receives normalized input data sequences, performs regression inference, and outputs initial prediction results including initial wave height, initial period, and model confidence. The large model regression unit stores the network structure and weight parameters of the trained wave parameter regression model and has the ability to perform forward inference calculations on the input data.

[0094] The physical constraint correction unit, with its built-in wave dispersion relation constraint term, receives the initial prediction result data and feature quality score. Under the wave dispersion relation constraint term, it performs physical consistency correction on the initial prediction result data and adaptively adjusts the weight ratio of the data fidelity term and the physical residual term during the correction process based on the feature quality score. It then outputs the corrected wave height and corrected period values. The physical constraint correction unit is one of the core components of the system. Its embedded wave dispersion relation constraint term provides a hard constraint based on physical laws for the correction process, while the adaptive weight adjustment mechanism achieves a dynamic balance between data-driven and physical-driven paradigms.

[0095] The confidence fusion unit is used to weight and fuse the initial prediction data with the corrected wave height and period values ​​based on the feature quality score and model confidence, outputting the final wave parameter data. The confidence fusion unit integrates the input quality evaluation and the output confidence information, and calculates a fusion coefficient to weight and synthesize the two wave parameter estimation results, generating the final wave parameter estimation result.

[0096] The above units are connected sequentially according to the data processing flow: the output of the data preprocessing unit is connected to the input of the feature extraction and quality assessment unit; the output of the feature extraction and quality assessment unit is connected to the input of the prompt word encapsulation unit; the output of the prompt word encapsulation unit is connected to the input of the large model regression unit; the output of the large model regression unit is connected to the input of the physical constraint correction unit; the output of the feature extraction and quality assessment unit is also connected to the input of the physical constraint correction unit (for transmitting feature quality scores); the outputs of the large model regression unit and the physical constraint correction unit are respectively connected to the input of the confidence fusion unit; and the confidence fusion unit outputs the final wave parameter data.

[0097] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0098] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. Any of the claimed embodiments can be used in any combination.

Claims

1. A wave parameter regression method based on structured features and physical constraints, characterized in that, include: Acquire wave video data of the target sea area, preprocess the wave video data to obtain wave surface height sequence data; The wavefront elevation sequence data is used to identify peak and trough elevation values. Within a preset time window, candidate wave height values ​​and candidate period values ​​are extracted based on the identified peak and trough elevation values ​​to obtain a structured feature data set. A feature quality score associated with the structured feature data set is then generated. The structured feature data set includes at least the peak elevation value, the trough elevation value, the candidate wave height value, and the candidate period value. A preset structured template protocol is invoked. The structured template protocol has predefined fields, including at least a wave feature value field, a unit field, a context parameter field, a physical prior cue field, and an uncertainty field. The values ​​contained in the structured feature data group are filled into the wave feature value field, the feature quality score is filled into the uncertainty field, and the values ​​filled into the wave feature value field are normalized to obtain a normalized model input data sequence. The normalized model input data sequence is input into a pre-trained wave parameter regression model to perform regression inference. The inference results are then structured and parsed according to a preset output field format to obtain initial prediction result data. The initial prediction result data includes at least the initial wave height value, the initial period value, and the model confidence. A pre-built physical constraint correction model is loaded, and a physical consistency correction operation is performed based on the initial prediction result data and the feature quality score to obtain the corrected wave height value and the corrected period value. The physical constraint correction model has a built-in wave dispersion relation constraint term, and when performing the physical consistency correction operation, the weight ratio of the data fidelity term and the physical residual term is adaptively adjusted according to the feature quality score. Based on the feature quality score and the model confidence level, the initial prediction result data, the corrected wave height value, and the corrected period value are subjected to confidence-weighted fusion to generate the final wave parameter data.

2. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, The preprocessing of the wave video data to obtain wavefront height sequence data includes: The wave video data is analyzed frame by frame, and the pixels in each frame are physically mapped according to the acquisition parameters of the wave video data to generate wave surface height sequence data arranged in chronological order; the acquisition parameters include at least one of the installation height of the image acquisition device, pitch angle and focal length. Perform at least one of image stabilization and noise reduction operations on the wavefront high-order sequence data.

3. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, The structured feature data set includes: Within a preset time window, the wavefront elevation sequence data at the same spatial location or in the same cross-sectional direction are acquired. The maximum value of the wavefront elevation sequence data within the preset time window is determined as the wave crest elevation value, and the minimum value of the wavefront elevation sequence data within the preset time window is determined as the wave trough elevation value. The difference between the peak elevation value and the trough elevation value is determined as the candidate wave height value; Perform a Fast Fourier Transform (FFT) on the wavefront high-order sequence data to obtain the spectrum. Determine the frequency corresponding to the main peak in the spectrum as the FFT main frequency, and determine the reciprocal of the FFT main frequency as the candidate period value.

4. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, Generating feature quality scores associated with the structured feature data set specifically includes: The feature quality score is obtained by weighting and summing at least two of the following quality indicators: effective pixel ratio of wavefront, consistency index between wave crest elevation and wave trough elevation, signal-to-noise ratio, occlusion ratio, and window stability. The weighted sum is then subjected to Sigmoid function mapping or truncation normalization processing. The value range of the feature quality score is a preset interval.

5. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, The obtained normalized model input data sequence includes: For each value filled into the wave feature value field, a normalization mapping is performed according to the preset mean and preset standard deviation corresponding to each value to obtain the normalized value corresponding to each value. The original values ​​of each item in the wave feature value field, the normalized values ​​corresponding to each item, the unit of measurement corresponding to the unit field, the context parameters corresponding to the context parameter field, the physical prior information corresponding to the physical prior cue field, and the feature quality score corresponding to the uncertainty field are organized according to the structured template protocol to obtain the normalized model input data sequence.

6. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, The initial prediction result data also includes output basis information, which is used to record the key field identifiers in the structured feature data group referenced in this inference.

7. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, The wave dispersion relation constraint term Constructed based on the following physical residuals: Where ω is the angular frequency, determined by the period value during the correction process; k is the wave number, determined by the wave wavelength, k = 2π / λ; h is the water depth of the target sea area; g is the gravitational acceleration; tanh is the hyperbolic tangent function; The wave wavelength λ is determined based on the spatial distance between adjacent wave crests in the wavefront height sequence data, or the wave number k is obtained as a latent variable by the physical constraint correction model during the training process.

8. The wave parameter regression method based on structured features and physical constraints according to claim 7, characterized in that, The adaptive adjustment of the weight ratio between the data fidelity term and the physical residual term based on the feature quality score includes: Construct the corrected loss function: Among them, L data For data fidelity items, , and These are the corrected wave height value and the corrected period value, respectively. and The initial wave height and the initial period value are respectively, λ data The weighting coefficient for the data fidelity item is set to the feature quality score q. L phys For physical residuals, , λ phys The weighting coefficient for the physical residual term is 1 - q; L reg λ is the regularization term; reg The weight coefficients for the regularization term; When the feature quality score q is higher, the data fidelity item L... data The larger the weight, the lower the feature quality score q, and the lower the physical residual term L. phys The greater the weight, the better.

9. The wave parameter regression method based on structured features and physical constraints according to claim 1, characterized in that, The generation of the final wave parameter data includes: The fusion coefficient is calculated based on the feature quality score and the model confidence score, wherein the fusion coefficient α is specifically: in, LLM The confidence level of the model is defined by η1 and η2, which are preset weight coefficients, and clip indicates that the calculation results are restricted to the range of 0 to 1. The initial prediction data, the corrected wave height value, and the corrected period value are weighted and summed using the fusion coefficients to obtain the final wave height value. final =α· LLM +(1-α)· The final period value is final =α· LLM +(1-α)· ; The final wave parameter data includes at least the final wave height value and the final period value.

10. A wave parameter regression system based on structured features and physical constraints, characterized in that, include: The data preprocessing unit is used to preprocess the acquired wave video data of the target sea area and output wave surface height sequence data. The feature extraction and quality assessment unit is used to identify the peak elevation and trough elevation values ​​from the wavefront height sequence data, extract candidate wave height values ​​and candidate period values ​​to form a structured feature data set, and generate a feature quality score associated with the structured feature data set. The prompt word encapsulation unit has a built-in preset structured template protocol, which is used to fill and normalize the structured feature data group and the feature quality score according to the structured template protocol, and output a normalized model input data sequence. The large model regression unit has a built-in pre-trained wave parameter regression large model, which is used to receive the normalized model input data sequence, perform regression inference and output initial prediction result data including initial wave height value, initial period value and model confidence. The physical constraint correction unit has a built-in wave dispersion relation constraint term. It is used to receive the initial prediction result data and the feature quality score, perform physical consistency correction on the initial prediction result data under the wave dispersion relation constraint term, and adaptively adjust the weight ratio of the data fidelity term and the physical residual term during the correction process according to the feature quality score, and output the corrected wave height value and the corrected period value. The confidence fusion unit is used to perform weighted fusion of the initial prediction result data, the corrected wave height value, and the corrected period value based on the feature quality score and the model confidence, and output the final wave parameter data.

Images