Foundation bed surface flatness evaluation method, system and equipment and underwater leveling machine

Abstract

The invention provides a foundation bed surface flatness evaluation method, system and equipment and an underwater leveling machine, and relates to the technical field of underwater construction. The method comprises the following steps: acquiring multi-source sensing data of an operation sea area, and generating a multi-source fusion foundation bed surface model based on the multi-source sensing data; taking the multi-source fusion subgrade bed surface model as the input of a target subgrade bed surface prediction model, and outputting an elevation prediction map of the subgrade bed surface based on the target subgrade bed surface prediction model; and generating a flatness evaluation index of the operation sea area based on the elevation deviation distribution diagram. The method can improve the accuracy and reliability of flatness evaluation of the foundation bed in the operation sea area.

Description

Technical Field

[0001] This application relates to the field of underwater construction technology, specifically to a method, system, equipment, and underwater leveling machine for evaluating the flatness of a foundation surface. Background Technology

[0002] In subsea riprap construction and underwater engineering (such as gravity wharves and pipeline laying), subsea leveling is a crucial process that directly affects the stability and service life of the structure. Currently, underwater leveling machines or manual operation by divers are commonly used to repeatedly scrape the subsea surface until the elevation and flatness meet design requirements. Advanced underwater leveling equipment claims to achieve a leveling accuracy of approximately 5cm; however, ensuring a sufficiently flat subsea surface over a large area remains a challenge in actual construction. Uneven leveling may leave unevenly distributed shallow spots and other quality defects, threatening the safety of subsequent projects. Therefore, after construction, it is usually necessary to conduct acceptance measurements of the subsea surface to check indicators such as elevation, thickness, and flatness.

[0003] Traditional underwater measurement methods have shortcomings: single sensors (such as multibeam sonar) may be limited by resolution or noise interference, making it difficult to detect small-scale unevenness in a timely manner; manual measurement points are sparse and time-consuming, and cannot comprehensively reflect the overall flatness of the substrate. Therefore, existing technologies suffer from low accuracy in assessing substrate flatness. Summary of the Invention

[0004] This application provides a method, system, equipment, and underwater leveling machine for assessing the flatness of a subgrade surface, in order to overcome the shortcomings of existing measurement methods, realize automated and refined evaluation of subgrade flatness, and improve the accuracy and reliability of subgrade flatness assessment in the operating sea area by obtaining a comprehensive and objective assessment result of the subgrade flatness.

[0005] In a first aspect, embodiments of this application provide a method for assessing the flatness of a seabed surface, comprising: acquiring multi-source sensor data of the operational sea area, and generating a multi-source fusion seabed surface model based on the multi-source sensor data; using the multi-source fusion seabed surface model as input to a target seabed surface prediction model, and outputting an elevation prediction map of the seabed surface based on the target seabed surface prediction model; wherein the target seabed surface prediction model is obtained by training an initial surface prediction model, the initial surface prediction model is constructed by a convolutional neural network and a multi-head attention mechanism, the hyperparameters of the initial surface prediction model are optimized by an alpha evolution algorithm, and the loss function of the initial surface prediction model is a Huber function; the value of each grid point in the elevation prediction map represents the deviation of the seabed surface at that location from the ideal design plane or target elevation; and generating a flatness evaluation index for the operational sea area based on the elevation prediction map.

[0006] According to one embodiment of this application, the step of using the multi-source fusion substrate surface model as input to the target substrate surface prediction model, and outputting an elevation prediction map of the substrate surface based on the target substrate surface prediction model, includes: inputting the multi-source fusion substrate surface model into the target substrate surface prediction model, obtaining the elevation prediction map of the substrate surface from the target substrate surface prediction model through forward inference; calculating the elevation deviation value of each grid point based on the elevation prediction map and reference elevation data, and generating the elevation deviation distribution map; wherein, the reference elevation data is the measured substrate elevation or the ideal design plane elevation, the elevation deviation distribution map includes the raised or recessed areas of the substrate surface, and the magnitude of the deviation value in the elevation deviation distribution map characterizes the unevenness of the operating sea area.

[0007] According to one embodiment of this application, the smoothness evaluation index includes an overall smoothness index and a local smoothness index; wherein, the overall smoothness index includes root mean square error, used to quantify the overall elevation fluctuation range of the operating sea area; the local smoothness index includes the average slope change and local extreme value detection results of the operating sea area, and the local smoothness index is used to characterize the surface smoothness and local unevenness rate of the operating sea area.

[0008] According to one embodiment of this application, the step of optimizing the hyperparameters of the initial surface prediction model includes: generating multiple candidate solutions in a uniformly distributed manner in the hyperparameter space, each candidate solution representing a hyperparameter configuration vector of the initial surface prediction model; wherein the hyperparameters include the kernel size, learning rate, batch size, and network depth of the initial surface prediction model, and the value range of each hyperparameter in the hyperparameter space is determined according to the characteristics of the bed topography data; randomly sampling from the current candidate solutions to construct an evolution matrix, each row vector of the evolution matrix corresponding to a hyperparameter configuration; initializing the evolution matrix and calculating the evaluation function, the value of which is the loss function value calculated by the initial surface prediction model on the bed validation set; determining whether the current function evaluation count has reached the maximum evaluation count, and if not, calculating the evolution matrix, perturbation matrix, and decay factor; and selecting the base by generating random numbers. The vector calculation path is as follows: if the random number is less than a set threshold, the basis vector is calculated through the first path; otherwise, it is calculated through the second path. After obtaining the basis vector according to the selected path, the first individual with a fitness higher than a preset threshold and the second individual with a fitness lower than a preset threshold are selected from the current population, and control parameters are calculated. Using the basis vector, control parameters, decay factor, and perturbation matrix, the search operator is updated using the alpha operator. Boundary constraints are applied to the updated search operator, and a binary backoff strategy is used to adjust the solution components that exceed the boundary. The fitness of individuals before and after the update is compared using a greedy selection strategy to determine whether to replace the original individual. After updating all individuals in the current generation, it is determined whether the individual index has reached the population size; if not, the next individual is processed. After all individuals have been processed, it is determined again whether the function evaluation count has reached the maximum value; if so, the optimal hyperparameter combination is output.

[0009] According to one embodiment of this application, the method for calculating the basis vector through the first path includes: ; The methods for calculating the basis vectors through the second path include: ; in, Let be the basis vector. and The learning rate parameter, and These are the historical basis vectors based on sampling paths with replacement and those based on sampling paths without replacement, respectively; A is a D×D matrix obtained by sampling D times with replacement from the candidate solution matrix X, and B is a matrix obtained by sampling K times without replacement from the candidate solution matrix X. This is a fitness-based weight vector.

[0010] According to one embodiment of this application, generating the flatness evaluation index of the operating sea area based on the elevation prediction map includes: calculating the mean square error or root mean square error by statistically analyzing the deviation values ​​of each point, so as to quantify the magnitude of the overall elevation fluctuation of the operating sea area based on the mean square error or root mean square error; calculating the slope change of the subgrade surface, determining the local slope through the elevation difference between adjacent grid points, and evaluating the maximum rate of change of the operating sea area.

[0011] According to one embodiment of this application, the flatness evaluation index includes one or more of the following: root mean square error, average slope change, and local unevenness rate.

[0012] Secondly, embodiments of this application provide a substrate surface flatness assessment system, comprising: a sensing unit, including multi-source mapping sensors deployed in the operating sea area for collecting multi-source sensing data; a prediction unit, wherein the prediction unit is configured with a trained target substrate surface prediction model, the prediction unit being used to generate a multi-source fusion substrate surface model based on the multi-source sensing data, using the multi-source fusion substrate surface model as input to the target substrate surface prediction model, and outputting an elevation prediction map of the substrate surface based on the target substrate surface prediction model; wherein the target substrate surface prediction model is obtained by training an initial surface prediction model, the initial surface prediction model being constructed by a convolutional neural network and a multi-head attention mechanism, the hyperparameters of the initial surface prediction model being optimized by an alpha evolution algorithm, and the loss function of the initial surface prediction model being a Huber function; the value of each grid point in the elevation prediction map represents the deviation of the substrate surface at that location from the ideal design plane or target elevation; and an evaluation unit, used to generate a flatness evaluation index for the operating sea area based on the elevation prediction map.

[0013] Thirdly, embodiments of this application provide a computer device, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described above.

[0014] Fourthly, this application also provides an underwater leveling machine, on which the computer equipment described in the above embodiments is provided.

[0015] Compared with existing technologies, the beneficial effects of this application are as follows: By acquiring multi-source sensor data from the operational sea area to construct a multi-source fusion subgrade surface model, the accuracy and robustness of the basic data are effectively improved. Subsequently, the fusion model is processed using an optimized target subgrade surface prediction model. This model combines the feature extraction capabilities of convolutional neural networks and the advantages of multi-head attention mechanisms in capturing long-distance dependencies. Furthermore, the hyperparameters are optimized through evolutionary algorithms, resulting in a model with stronger adaptability and accuracy in predicting subgrade surface elevation. The adoption of the Huber loss function further ensures the training stability of the model in the presence of outliers. The final generated elevation prediction map intuitively reflects the deviation of each point on the subgrade from the design elevation, providing a direct basis for the generation of flatness evaluation indicators. This allows for a comprehensive and objective assessment of the subgrade flatness in the operational sea area, significantly improving the accuracy and reliability of subgrade flatness assessment in the operational sea area. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the steps of the method for evaluating the flatness of the substrate surface provided in the embodiments of this application.

[0017] Figure 2 A schematic diagram of the framework of the target substrate surface prediction model provided in the embodiments of this application.

[0018] Figure 3 This is a schematic diagram illustrating the specific process of the AE algorithm provided in the embodiments of this application. Detailed Implementation

[0019] The present application will now be described in further detail with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the subject matter of the present application to the following embodiments. All technologies implemented based on the content of the present application fall within the scope of protection of the present application.

[0020] Unless otherwise specified, the terms "upper," "lower," "left," "right," "center," "inner," "outer," and "side" used in the description of specific embodiments of this application to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationship in which the product / equipment / device is usually placed during use. These terms are merely for the purpose of facilitating the description of the solution in this application or simplifying the description in specific embodiments, so as to enable those skilled in the art to quickly understand the solution, and do not indicate or imply that a particular device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship. Therefore, they should not be construed as limitations on this application.

[0021] In the description of the embodiments of this application, technical terms such as "first" and "second" only distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary or secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] During the research process, the applicant discovered that a high-precision assessment of the flatness of the leveled seabed surface is essential. In recent years, deep learning technology has demonstrated enormous potential in the field of underwater topographic mapping—combining multi-source remote sensing / image / acoustic data can effectively improve the accuracy of seabed topographic reconstruction. For example, research has proposed a shallow sea topographic inversion model, CNN-NSTR, based on convolutional neural networks fusing multi-source multispectral image data. This model significantly improves the accuracy of water depth prediction, reducing the mean square error to 1.35m in the 0–30m water depth range, an improvement of more than 50% compared to traditional empirical models. Deep convolutional networks can model complex relationships between multi-source data and seabed elevation, capturing subtle topographic changes and ensuring that the predicted topography closely matches the actual elevation trend.

[0024] Based on this, the applicant, after research, proposed an intelligent method for evaluating the flatness of the subgrade surface. Please refer to [link / reference needed]. Figure 1 , Figure 1 This diagram illustrates the steps of the method for evaluating the flatness of a subgrade surface provided in this application embodiment. The steps of the method for evaluating the flatness of a subgrade surface may include: S1. Acquire multi-source sensor data of the operating sea area and generate a multi-source fusion substrate surface model based on the multi-source sensor data.

[0025] S2. Use the multi-source fusion substrate surface model as input to the target substrate surface prediction model, and output the elevation prediction map of the substrate surface based on the target substrate surface prediction model.

[0026] S3. Generate flatness evaluation indicators for the operating sea area based on the elevation prediction map.

[0027] In this embodiment of the application, the operating sea area refers to the area where the underwater leveling operation has been completed, and the multi-source sensing data refers to the multi-source sensing data (including three-dimensional morphology and texture data) of the substrate surface obtained from different physical dimensions (such as acoustics and optics) using different types of sensors.

[0028] For example, after the leveling construction is completed, detailed information about the subgrade surface can be obtained by multi-source mapping sensors deployed in the operating sea area, including (1) multibeam echo sounding: multibeam echo sounders can be deployed on the operating vessel to perform high-resolution scanning of the leveled area and obtain data such as the digital elevation model (DEM) and reflection intensity map of the subgrade surface; (2) lidar / structured light scanning: depending on the water depth and water clarity, underwater lidar, structured light projector and camera devices can be used to perform fine three-dimensional scanning of the local area and obtain high-precision point cloud data; (3) underwater camera acquisition: underwater high-definition cameras can be deployed to capture visible light image sequences of the subgrade surface for photometric three-dimensional reconstruction or as a texture auxiliary means.

[0029] By performing operations such as correction, noise filtering, and outlier removal on the acquired multi-source sensor data, and transforming all data into a unified coordinate system, the limitations of a single data source can be overcome (e.g., sonar is susceptible to noise interference, and optical equipment is affected by water clarity). Through coordinate unification, registration, and complementarity, a more complete and accurate multi-source fusion substrate surface model can be generated. Ultimately, a multi-source fusion substrate surface 3D model is generated. This multi-source fusion substrate surface model can be represented as a Digital Elevation Model (DEM) with texture information.

[0030] For example, data acquired by each sensor can be transmitted to a support platform via an underwater communication link or buoy relay. The support platform then generates a multi-source fusion bed surface model based on the multi-source sensor data. The support platform can be a data processing system on a work vessel, a shore-based or fixed platform control center, or a cloud computing service platform.

[0031] By performing sound velocity correction, noise filtering, and outlier removal on multibeam bathymetry data, a seabed topographic point cloud or raster elevation map covering the leveled area is generated. Distortion correction and coordinate transformation (considering the effects of water refraction) are performed on laser / structured light point cloud data, which is then registered and fused with sonar data. Color calibration and distortion correction are performed on underwater images, and sparse point clouds are extracted using photogrammetry algorithms, then combined with laser point clouds for densification. After these preprocessing steps, a multi-source fused seabed surface model and its image texture in a unified coordinate system are obtained. This data will then be used as input to a deep learning model.

[0032] Please refer to Figure 2 , Figure 2This is a schematic diagram of the framework of the target subgrade surface prediction model provided in this application embodiment. The target subgrade surface prediction model is an intelligent agent for evaluating the flatness of the subgrade surface. The initial surface prediction model is an end-to-end model constructed using Convolutional Neural Networks (CNN) and a multi-head attention mechanism. CNN excels at extracting local spatial features (such as minute undulations) from images or raster data, while the multi-head attention mechanism can capture the global dependencies between different data sources, weigh the importance of information from each source, and achieve more effective feature fusion. The target subgrade surface prediction model achieves accurate prediction of the subgrade surface elevation through an end-to-end model.

[0033] The model's input is a multi-source fused substrate surface model, specifically represented by three data extraction modules: raster elevation, depth, and texture features. The 3D model generated by the platform is converted into regular raster data, where each grid point contains not only elevation values ​​but also depth information derived from multibeam data and texture features extracted from underwater imagery.

[0034] These input data are fed into a convolutional neural network (CNN) module. The CNN layer shown in the diagram typically contains a series of sequential operations, including: convolution operations that use learnable convolutional kernels to perform sliding window calculations on the input data to capture local spatial features ranging from edges and shapes to more complex terrain structures; the ReLU activation function introduces a non-linear transformation into the network, enabling it to fit complex non-linear relationships; and max pooling operations, which reduce the spatial resolution of the feature maps while preserving salient features, increase the model's translation invariance and receptive field. This series of operations works together to transform the raw input into a high-level, abstract feature map.

[0035] The core of convolutional neural networks lies in convolution operators and feature fusion. By sliding and multiplying the convolutional kernel with the input feature map spatially and accumulating the results, weighted extraction of local patterns is achieved. The output of each convolutional layer undergoes nonlinear activation. (ReLU) mapping introduces sparsity and nonlinear discriminative power. The mathematical expression of the multi-head attention mechanism is: for the input feature sequence... After calculating the key, query, and value vectors, the self-attention output is performed. This mechanism can capture global dependencies. By combining local features extracted through convolution, the model can simultaneously obtain local details and global context, improving the accuracy of terrain prediction.

[0036] The hyperparameters of the initial surface prediction model are optimized using the Alpha Evolution (AE) algorithm, which achieves an efficient balance between global exploration and local exploitation through a core Alpha operator, thereby automatically finding an optimal set of hyperparameters.

[0037] After the CNN extracts spatial features, it enters a multi-head attention mechanism layer. This allows the model to autonomously focus on deeper dependencies between different features. For example, one set of heads might focus on the correlation between elevation and texture features, while another set might focus on the spatial relationships between distant grid points. In this way, the model can more intelligently fuse information from different data sources and understand the broader terrain context, rather than just the local neighborhood, thus making more accurate predictions.

[0038] For example, raster elevation, depth, and texture are fused using a multi-head attention mechanism and then input into the same convolutional network. During the encoding stage, multiple convolutional and pooling layers are stacked sequentially to extract high-level features. The convolutional layers use a certain kernel size (e.g., 3×3) to slide and extract local region features. Each convolutional layer is followed by a ReLU activation function to introduce non-linearity, combined with batch normalization (BN) to accelerate convergence and improve generalization performance. For example, four convolutional layers are used, each with a 3×3 kernel size, and the number of channels increases progressively (e.g., 32, 64, 128, 256), with ReLU and batch normalization applied after each layer. The convolutional computation satisfies: ,in Represents the input image tensor. For convolution kernel weights, It outputs the position index. The kernel's height and width range are defined. Local feature extraction is achieved through convolution operations, while pooling layers downsample the feature map using a 2×2 window, reducing resolution while extracting invariant features. After feature extraction via convolution, a fully connected layer merges spatial information to output a global flatness index for predicting continuous elevation or deviation values.

[0039] The high-level features, weighted and integrated by the multi-head attention mechanism, are flattened and fed into one or more fully connected layers. The fully connected layer acts like a powerful regressor, mapping the distributed features learned from all preceding steps to the final output space. In this task, the output layer is typically a linear unit, whose output value directly corresponds to the predicted elevation value for each grid point.

[0040] The elevation prediction map directly displays the elevation distribution of the leveled subgrade surface. Due to the use of a deep learning model, it can obtain more detailed terrain features than traditional interpolation methods. (Output elevation prediction map) The measured true elevation can be subtracted. (Or subtract the reference elevation surface) to obtain the deviation The deviation diagram visually displays which areas of the subgrade still have protrusions or depressions, and its numerical value reflects the unevenness. This output allows for convenient calculation of flatness evaluation indicators, such as root mean square error, for acceptance assessment.

[0041] In some embodiments, to train an initial surface prediction model, this application defines a loss function to measure the deviation between the predicted elevation and the true elevation. This application uses the Huber function as the loss function for the initial surface prediction model, measuring the difference between the model's predicted value and the true value. The Huber function combines the advantages of Mean Squared Error (MSE) and Mean Absolute Error (MAE), is insensitive to outliers in the training data, and improves the stability of model training.

[0042] The mathematical form of Huber loss is:

[0043] in, The difference between the predicted value and the actual value. The threshold hyperparameter controls the inflection point at which the loss transitions from quadratic to linear. This is achieved by selecting an appropriate threshold hyperparameter. The Huber loss is equivalent to the MSE when the error is small, and equivalent to the absolute error when the error is large, thus improving robustness to outliers. The model is trained to minimize the above loss by adjusting the network parameters through backpropagation, so that the predicted elevation approximates the true value.

[0044] Please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating the specific process of the AE algorithm provided in an embodiment of this application. In some embodiments, the step of optimizing the hyperparameters of the initial surface prediction model based on the AE algorithm may include: In the hyperparameter space, multiple candidate solutions are generated in a uniformly distributed manner. Each candidate solution represents a hyperparameter configuration vector of the initial surface prediction model. The hyperparameters include the kernel size, learning rate, batch size, and network depth of the initial surface prediction model. The value range of each hyperparameter in the hyperparameter space is determined based on the characteristics of the bed topography data. An evolution matrix is ​​constructed by randomly sampling from the current candidate solutions. Each row vector of the evolution matrix corresponds to a hyperparameter configuration. The evolution matrix is ​​initialized, and an evaluation function is calculated. The evaluation function is the loss function value calculated by the initial surface prediction model on the bed validation set. It is determined whether the current function evaluation count has reached the maximum number of evaluations. If not, the evolution matrix, perturbation matrix, and decay factor are calculated. The calculation path of the basis vector is selected by generating random numbers. When the random number is less than a set threshold... In the first case, the basis vector is calculated via the first path; otherwise, it is calculated via the second path. After obtaining the basis vector based on the selected path, the first individual with a fitness higher than a preset threshold and the second individual with a fitness lower than a preset threshold are selected from the current population, and control parameters are calculated. Using the basis vector, control parameters, decay factor, and perturbation matrix, the search operator is updated using the alpha operator. Boundary constraints are applied to the updated search operator, and a binary backoff strategy is used to adjust the solution components that exceed the boundary. The fitness of individuals before and after the update is compared using a greedy selection strategy to determine whether to replace the original individual. After updating all individuals in the current generation, it is determined whether the individual index has reached the population size; if not, the next individual is processed. After all individuals have been processed, it is determined again whether the number of function evaluations has reached the maximum value; if so, the optimal hyperparameter combination is output.

[0045] For example, in In the 3D hyperparameter space, the AE algorithm first generates the hyperparameters in a uniform distribution manner. There are several candidate solutions to form a candidate matrix:

[0046] in, Indicates the first One hyperparameter configuration vector, These are the lower and upper bounds of each dimension; This indicates the generation of a length of The vector, each component in Uniform sampling.

[0047] The Alpha operator, by simultaneously fusing three types of information—adaptive global starting point, global random perturbation, and local difference correction—takes into account both global exploration and local development in a single update, thereby efficiently mining and refining candidate solutions. Its mathematical expression is:

[0048] Among them, the evolution matrix By analyzing the candidate matrix A sample obtained by sampling with replacement Matrix. Specifically, in each generation, the algorithm selects candidate matrices... Randomly selected from candidate solutions (Repetition allowed) , its first OK That is the first A solution awaiting update; This represents the basis vector, which determines the starting position of the evolution; As a decay factor, it controls the exploration and development of algorithms; Indicates the first A random step size; To control the parameters, control the differential vector (adaptive step size); and That is, from The solution extracted from the middle satisfies .

[0049] For adaptive basis vectors, the evolutionary starting point is initially calculated in two ways:

[0050] in From the candidate matrix Sampling with replacement The square array obtained this time This indicates taking its diagonal; From Sampling without replacement The matrix obtained this time This is a fitness-based weight vector. To make... As historical information accumulates across generations, the AE algorithm introduces evolutionary paths for the two sampling methods.

[0051] The method for calculating the basis vector through the first path includes: ; The methods for calculating the basis vectors through the second path include: ; in, Let be the basis vector. and The learning rate parameter, and These are the historical basis vectors based on sampling paths with replacement and those based on sampling paths without replacement, respectively; A is a D×D matrix obtained by sampling D times with replacement from the candidate solution matrix X, and B is a matrix obtained by sampling K times without replacement from the candidate solution matrix X. This is a fitness-based weight vector.

[0052] Learning rate , This represents the current number of times the target function has been called. This represents the maximum number of calls allowed.

[0053] For random step size It provides a global search function. The attenuation factor is a non-linear decreasing value, related to the perturbation matrix. Closely related. Its decay process is as follows:

[0054] perturbation matrix The calculation is as follows:

[0055] in and Indicates by The generated random real matrices are used to generate perturbations. It is a collection of lines and A matrix of columns express The first in Row vectors.

[0056] To ensure the search remains within the feasible region, AE employs a "binary backoff" approach for variables that exceed the boundary:

[0057] After the update, a greedy strategy will be used to select and retain the best performers.

[0058] in, This is the value of the j-th dimension (i.e., a specific hyperparameter, such as the learning rate) of the i-th candidate solution after being updated by the alpha operator. Let j be the upper bound of the j-th hyperparameter in the search space. This is the lower bound of the j-th hyperparameter in the search space. Let be the hyperparameter configuration vector for the i-th candidate solution in the t-th generation of the population. The vector of new candidate solutions is generated after the i-th candidate solution is updated by the alpha operator and boundary constraints are applied.

[0059] During model training, supervised learning was employed, with label data derived from high-precision measured bed elevations (e.g., real elevations obtained from underwater laser point clouds or multibeam bathymetry). The measured elevation of a specific area after leveling construction was used as the ground truth. Multi-source observation data (sonar depth maps, image textures, etc.) of the corresponding region are used as model input. Construct training samples Yes. The loss function should use either Huber or MSE as mentioned above to measure the model output. and The difference is negligible. During training, the model parameters are optimized by minimizing the loss, making the predicted results approximate the true elevation distribution. The optimization algorithm uses the Adam adaptive gradient optimizer, with an initial learning rate set to 0.001, combined with a momentum factor. Adam optimization exhibits good convergence efficiency and robustness to hyperparameters in deep learning. To ensure training stability, embodiments of this application employ a learning rate decay strategy during training, such as multiplying the learning rate by 0.1 every few epochs to prevent later oscillations.

[0060] Regarding hyperparameter selection and validation, the batch size during training depends on the hardware memory and data size. For raster input, batch sizes such as 32 or 64 can be chosen; if training with small image patches, even larger batch sizes can be set (some studies have used 512 batches to train 9×9 small patches to fully utilize the data). Model training requires dividing the training and validation sets, typically randomly partitioning the data in a ratio of, for example, 8:2, while ensuring data coverage across different regions. To fully utilize limited data and evaluate model robustness, this application's embodiments introduce a K-fold cross-validation scheme. For example... The dataset is divided into five equal parts. One part is used for validation each time, and the remaining four parts are used for training. This process is repeated five times to obtain the average performance. This approach provides a more reliable evaluation of the model compared to a single partition, reducing bias caused by random partitions. A study compared the effects of traditional random partitioning with K-fold validation, finding that the latter improved test accuracy from 84.1% to 88.3% and reduced the spatial error standard deviation by 60%, indicating that cross-validation helps improve the model's generalization performance. Preventing overfitting: Due to the high cost of acquiring underwater terrain data and the limited number of training samples, models are prone to overfitting. To address this, several regularization strategies are employed: First, Dropout layers are used, randomly discarding neuron outputs with a certain probability (e.g., 30%) in fully connected layers or during the decoding stage to break feature dependencies. Dropout reduces over-reliance on certain local features during training, improving the model's adaptability to unseen data. Second, early stopping is applied, monitoring the validation set loss and stopping training when there is no improvement after several epochs to avoid overtraining. Third, data augmentation can be performed. If the original imagery and depth measurement data allow, random rotations, translational perturbations, or noise can be added to the training samples to increase data diversity and thus improve model robustness. Finally, introducing residual connections and regularization terms into the network structure is also helpful. Residual connections allow for training deeper networks without sacrificing stability; adding these to the loss function... This approach can suppress excessively large model parameters. By comprehensively utilizing the above strategies, the embodiments of this application can focus more on the performance of validation error while ensuring a reduction in training error, striving to obtain a flatness evaluation model with strong generalization ability and applicability to different environments.

[0061] Furthermore, to comprehensively evaluate model performance, embodiments of this application select multiple metrics for quantitative evaluation on the validation set, including RMSE, MAE, and These indicators measure the degree of agreement between predicted and actual elevations from different perspectives. Among them, the root mean square error (RMSE) measures the overall error magnitude, reflecting the standard deviation of the predicted value from the actual value. The mean absolute error (MAE) is the average of the absolute values ​​of the errors, and the formula is... The median (MAE) directly represents the magnitude of the mean deviation. Because the median is highly resistant to outliers, the MAE better reflects the typical error level of most points, ignoring a very small number of outliers. (Coefficient of determination) The measure of how well a model explains actual elevation changes is defined as follows: . The value ranges from 0 to 1. The closer it is to 1, the more terrain variance the model explains, meaning the better the prediction fits the true value.

[0062] In some embodiments, the smoothness evaluation index may include an overall smoothness index and a local smoothness index; wherein, the overall smoothness index includes root mean square error, used to quantify the overall elevation fluctuation of the operating sea area; the local smoothness index includes the average slope change and local extreme value detection results of the operating sea area, and the local smoothness index is used to characterize the surface smoothness and local unevenness rate of the operating sea area. Step S12 may specifically include: inputting the multi-source fusion substrate surface model into the target substrate surface prediction model, and obtaining the elevation prediction map of the substrate surface by forward inference from the target substrate surface prediction model; based on the elevation prediction map and reference elevation data, calculating the elevation deviation value of each grid point, and generating the elevation deviation distribution map; wherein, the reference elevation data is the measured substrate elevation or the ideal design plane elevation, the elevation deviation distribution map includes the convex or concave areas of the substrate surface, and the magnitude of the deviation value in the elevation deviation distribution map characterizes the unevenness of the operating sea area.

[0063] The trained target subgrade surface prediction model can output a corresponding elevation prediction map based on the input multi-source fused subgrade surface model. Each pixel or grid point in the map corresponds to a predicted absolute elevation value. By comparing this prediction map with the ideal design plane or target elevation (i.e., the perfectly flat surface required for construction), the deviation at each point can be obtained, thus forming an elevation deviation distribution map. The variation in the elevation deviation distribution map is the relative deviation value; a positive deviation value indicates a convexity, and a negative deviation value indicates a depression.

[0064] The steps for generating the flatness evaluation index of the operating sea area based on the elevation deviation distribution map may specifically include: calculating the mean square error or root mean square error of the deviation values ​​at each point, so as to quantify the magnitude of the overall elevation fluctuation of the operating sea area based on the mean square error or root mean square error; calculating the slope change of the subgrade surface, determining the local slope through the elevation difference between adjacent grid points, and evaluating the maximum rate of change of the operating sea area.

[0065] For example, after training, the model is deployed for flatness assessment of field data. Preprocessed and fused multi-source data is input into the trained CNN model, and forward inference is performed to obtain the predicted flatness of the subgrade surface. Next, flatness indices are calculated from the model output: the mean square error (MSE) or root mean square error (RMSE) is calculated based on the deviation values ​​at each point to quantify the overall elevation fluctuation; simultaneously, the slope variation of the subgrade surface is calculated, for example, by obtaining the local slope through the elevation difference between adjacent grid points and evaluating its maximum rate of change to capture steep slopes or uneven areas. These indices comprehensively reflect the flatness after leveling. Finally, a flatness assessment report is generated, and the construction quality is judged by comparing it with pre-set acceptance thresholds: if the mean square error and other indices are within the allowable range, the flatness of the subgrade surface is considered acceptable; otherwise, areas exceeding the tolerance are marked for rework and repair.

[0066] In this embodiment, a visualization method can be used to compare and analyze the elevation / deviation map output by the model with the actual measurement. On the one hand, a comparison map of the predicted elevation and the measured elevation (such as a profile comparison, a 3D surface map, etc.) is drawn to visually check whether the model has captured the key undulation features of the subgrade surface. For an ideally flat area, the predicted and measured curves should basically coincide; if there is a large deviation in the prediction at a certain point, it will be shown as a significant elevation or depression on the map. On the other hand, a heat map is used to display the distribution of deviations, and the magnitude of the deviation is represented by color, which can clearly locate the location and degree of residual unevenness. This scatter plot comparison map and deviation heat map can quantitatively and intuitively evaluate the leveling effect: dense point clouds that are close to the ideal line indicate good overall flatness, while outliers and areas with concentrated deviations indicate areas that need to be re-examined. Through the analysis of the above-mentioned multiple visualization methods, we not only verify the accuracy of the model prediction, but also can further guide the rework and optimization of construction based on the deviation map, ultimately achieving a comprehensive and objective evaluation of the leveling quality of the underwater subgrade.

[0067] In the aforementioned implementation process, a multi-source fusion model of the subgrade surface was constructed by acquiring multi-source sensor data from the operational sea area, effectively improving the accuracy and robustness of the basic data. Subsequently, the fusion model was processed using an optimized target subgrade surface prediction model. This model combines the feature extraction capabilities of convolutional neural networks and the advantages of multi-head attention mechanisms in capturing long-distance dependencies. Furthermore, evolutionary algorithms were used to optimize hyperparameters, resulting in a more adaptable and accurate model for predicting subgrade surface elevation. The adoption of the Huber loss function further ensured the model's training stability even in the presence of outliers. The final generated elevation prediction map visually reflects the deviation of each point on the subgrade from the design elevation, providing a direct basis for generating flatness evaluation indicators. This allows for a comprehensive and objective assessment of the subgrade flatness in the operational sea area, significantly improving the accuracy and reliability of subgrade flatness evaluation.

[0068] Based on the same concept, embodiments of this application also provide a substrate surface flatness assessment system, which may include: The sensing unit includes multi-source mapping sensors deployed in the operational sea area for collecting multi-source sensing data; a prediction unit configured with a trained target bed surface prediction model, which generates a multi-source fusion bed surface model based on the multi-source sensing data, uses the multi-source fusion bed surface model as input to the target bed surface prediction model, and outputs a predicted elevation map of the bed surface based on the target bed surface prediction model; wherein, the target bed surface prediction model is obtained by training an initial surface prediction model, which is constructed by a convolutional neural network and a multi-head attention mechanism, the hyperparameters of the initial surface prediction model are optimized by an alpha evolution algorithm, and the loss function of the initial surface prediction model is the Huber function; the value of each grid point in the elevation prediction map represents the deviation of the bed surface from the ideal design plane or target elevation; and an evaluation unit is used to generate a flatness evaluation index for the operational sea area based on the elevation prediction map.

[0069] It should be understood that when the various modules of the system provided in the above embodiments are working, the division of each functional module in the above description is only used as an example. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0070] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0071] Based on the same concept, embodiments of this application also provide a computer device, which may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described above.

[0072] Based on the same concept, this application also provides an underwater leveling machine, which is equipped with the computer device described above for executing the methods described above.

[0073] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for evaluating the flatness of a substrate surface, characterized in that, include: Acquire multi-source sensor data of the operating sea area, and generate a multi-source fusion substrate surface model based on the multi-source sensor data; The multi-source fusion bed surface model is used as input to the target bed surface prediction model to output an elevation prediction map of the bed surface based on the target bed surface prediction model. The target bed surface prediction model is obtained by training an initial surface prediction model, which is constructed using a convolutional neural network and a multi-head attention mechanism. The hyperparameters of the initial surface prediction model are optimized using an alpha evolution algorithm, and the loss function of the initial surface prediction model is the Huber function. The value of each grid point in the elevation prediction map represents the deviation of the bed surface at that location from the ideal design plane or target elevation. The flatness evaluation index of the operating sea area is generated based on the elevation prediction map.

2. The method according to claim 1, characterized in that, The step of using the multi-source fused substrate surface model as input to the target substrate surface prediction model, and outputting a predicted elevation map of the substrate surface based on the target substrate surface prediction model, includes: The multi-source fusion substrate surface model is input into the target substrate surface prediction model, and the target substrate surface prediction model obtains the elevation prediction map of the substrate surface through forward inference; Based on the elevation prediction map and reference elevation data, the elevation deviation value of each grid point is calculated to generate the elevation deviation distribution map; wherein, the reference elevation data is the measured subgrade elevation or the ideal design plane elevation, the elevation deviation distribution map includes the raised or recessed areas of the subgrade surface, and the magnitude of the deviation value in the elevation deviation distribution map characterizes the unevenness of the operating sea area.

3. The method according to claim 1, characterized in that, The smoothness evaluation index includes an overall smoothness index and a local smoothness index; wherein, the overall smoothness index includes the root mean square error, which is used to quantify the overall elevation fluctuation of the operating sea area; the local smoothness index includes the average slope change and local extreme value detection results of the operating sea area, and the local smoothness index is used to characterize the surface smoothness and local unevenness rate of the operating sea area.

4. The method according to claim 1, characterized in that, The hyperparameters of the initial surface prediction model are optimized by the following steps: In the hyperparameter space, multiple candidate solutions are generated in a uniformly distributed manner, each candidate solution representing a hyperparameter configuration vector of the initial surface prediction model; wherein, the hyperparameters include the kernel size, learning rate, batch size and network depth of the initial surface prediction model, and the value range of each hyperparameter in the hyperparameter space is determined according to the characteristics of the bed topography data; An evolution matrix is ​​constructed by randomly sampling from the current candidate solutions, and each row vector of the evolution matrix corresponds to a hyperparameter configuration; The evolution matrix is ​​initialized and the evaluation function is calculated. The evaluation function is the loss function value calculated by the initial surface prediction model on the bed validation set. Determine whether the current function evaluation count has reached the maximum number of evaluations. If not, calculate the evolution matrix, perturbation matrix, and decay factor. The calculation path of the basis vector is selected by generating a random number. If the random number is less than a set threshold, the basis vector is calculated through the first path; otherwise, the basis vector is calculated through the second path. After obtaining the basis vectors according to the selected path, the first individual with a fitness higher than a preset threshold and the second individual with a fitness lower than a preset threshold are selected from the current population, and the control parameters are calculated; using the basis vectors, control parameters, decay factors and perturbation matrix, the search operator is updated through the alpha operator; Boundary constraints are applied to the updated search operator, and a binary backoff strategy is used to adjust solution components that exceed the boundary. A greedy selection strategy is used to compare the fitness of individuals before and after the update to decide whether to replace the original individuals. After updating all individuals in the current generation, determine whether the individual index has reached the population size. If not, continue processing the next individual. After all individuals have been processed, check again whether the number of function evaluations has reached the maximum value. If it has, output the optimal combination of hyperparameters.

5. The method according to claim 4, characterized in that, in, The methods for calculating the basis vectors through the first path include: ； The methods for calculating the basis vectors through the second path include: ； in, Let be the basis vector. and The learning rate parameter, and These are the historical basis vectors based on sampling paths with replacement and those based on sampling paths without replacement, respectively; A is a D×D matrix obtained by sampling D times with replacement from the candidate solution matrix X, and B is a matrix obtained by sampling K times without replacement from the candidate solution matrix X. This is a fitness-based weight vector.

6. The method according to claim 1, characterized in that, The generation of the flatness evaluation index for the operating sea area based on the elevation prediction map includes: The mean square error or root mean square error is calculated by statistically analyzing the deviation values ​​at each point, and the magnitude of the overall elevation fluctuation of the operating sea area is quantified based on the mean square error or the root mean square error. The slope variation of the substrate surface is calculated, the local slope is determined by the elevation difference between adjacent grid points, and the maximum rate of change in the operating sea area is evaluated.

7. The method according to claim 1, characterized in that, The flatness evaluation index includes one or more of the following: root mean square error, average slope change, and local unevenness rate.

8. A system for evaluating the flatness of a substrate bed surface, characterized in that, include: The sensing unit includes multi-source mapping sensors deployed in the operating sea area to collect multi-source sensing data. The prediction unit is configured with a trained target bed surface prediction model. The prediction unit generates a multi-source fusion bed surface model based on the multi-source sensor data, uses this multi-source fusion bed surface model as input to the target bed surface prediction model, and outputs a predicted elevation map of the bed surface based on the target bed surface prediction model. The target bed surface prediction model is obtained by training an initial surface prediction model, which is constructed using a convolutional neural network and a multi-head attention mechanism. The hyperparameters of the initial surface prediction model are optimized using an alpha evolution algorithm, and the loss function of the initial surface prediction model is the Huber function. The value of each grid point in the elevation prediction map represents the deviation of the bed surface at that location from the ideal design plane or target elevation. An evaluation unit is used to generate a flatness evaluation index for the operating sea area based on the elevation prediction map.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1-7.

10. An underwater leveling machine, characterized in that, The underwater leveling machine is equipped with the computer equipment described in claim 9.

Images