Desilting strategy optimization method and system based on reinforcement learning

Abstract

The invention provides a reinforcement learning-based desilting strategy optimization method and system, and relates to the technical field of intelligent water area desilting, and the method comprises the steps: carrying out the multi-scale decomposition, adaptive histogram equalization, contrast optimization and edge sharpening processing of a collected image, and obtaining an optimized desilting environment image; then constructing a three-dimensional voxel grid model based on the optimized image, and establishing an environment interference source mathematical model and a dynamic evolution rule of sludge distribution; and finally, inputting the image features into a dredging strategy optimization network, through interaction with a simulation environment, updating network parameters based on a dual time sequence difference algorithm, generating a motion track planning scheme and an operation parameter configuration scheme of the dredging robot, and optimizing a value function on line according to an actual dredging effect. The desilting efficiency can be effectively improved, and the desilting cost is reduced.

Description

Technical Field

[0001] The present invention relates to intelligent water area dredging technology, and in particular to a dredging strategy optimization method and system based on reinforcement learning. Background Art

[0002] Traditional dredging methods mainly rely on mechanical removal and suction dredging technologies, which have problems such as low dredging efficiency, poor operation accuracy, and inability to cope with changes in complex water environments. At the same time, dredging operations usually rely on manual experience for decision-making, lacking the ability to be intelligent and adapt in real time, resulting in unstable dredging results and serious waste of resources.

[0003] In addition, environmental factors have an important impact on the effectiveness of dredging operations, and existing dredging technologies fail to effectively consider these factors. Although there are some water environment modeling methods based on physical models, these methods are computationally intensive, difficult to adjust in real time, and lack intelligent optimization mechanisms, and cannot fully cope with complex dynamic changes.

[0004] In recent years, deep learning and reinforcement learning technologies have been gradually applied to dredging operations in waters. Extracting image features through deep convolutional neural networks and optimizing operation strategies in combination with reinforcement learning algorithms have become the development trend of intelligent dredging technology. However, existing dredging methods based on reinforcement learning still face challenges such as multi-physics model integration, real-time environmental adaptation, and decision-making efficiency.

[0005] Therefore, proposing a new method that can simultaneously consider the dynamic changes of the water environment and optimize the dredging strategy in real time has important technical significance and application prospects. Summary of the invention

[0006] The embodiments of the present invention provide a dredging strategy optimization method and system based on reinforcement learning, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention, A dredging strategy optimization method based on reinforcement learning is provided, comprising: The desilting environment image of the silt is collected by a depth camera and a sonar sensor, and the desilting environment image is subjected to multi-scale decomposition based on orthogonal wavelet and curvelet transform to filter out noise to obtain a de-noised image, and the grayscale distribution probability density function is calculated by an adaptive histogram equalization algorithm for the de-noised image, and a piecewise linear grayscale stretching and gamma correction are performed according to the peak interval of the grayscale distribution probability density function to obtain a brightness equalized image, and the brightness equalized image is subjected to contrast optimization and edge sharpening processing by using a bilateral filtering algorithm and a fuzzy enhancement algorithm, and the enhancement effect is iteratively optimized by using a peak signal-to-noise ratio index and a structural similarity index to obtain an optimized desilting environment image; A three-dimensional voxel grid model is constructed based on the optimized desilting environment image, a state vector including a water velocity field, a turbulence intensity and a suspended solids concentration is established in each voxel grid unit of the three-dimensional voxel grid model, a Bayesian probability model is used to calculate the joint probability distribution of environmental interference factors in the voxel grid unit, a mathematical model of environmental interference sources is constructed according to the joint probability distribution, and the Navier-Stokes equations are used to model and solve the silt distribution state to obtain the dynamic evolution law of the silt distribution; The optimized dredging environment image is input into a deep convolutional neural network, and the multi-scale texture features, directional gradient features and regional semantic features of the optimized dredging environment image are extracted. The extracted features are fused using an attention weight mechanism to obtain a feature vector, and the feature vector is input into a classification model to determine the silt category. A three-dimensional distribution map of silt thickness is constructed by combining the mathematical model of the environmental interference source and the dynamic evolution law of silt distribution. The silt category and the three-dimensional distribution map of silt thickness are input into a dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward values ​​by interacting with the simulation environment, updates the strategy network parameters based on a dual temporal difference algorithm, generates a motion trajectory planning scheme and an operation parameter configuration scheme for the dredging robot, and performs online optimization of the value function of the dredging strategy optimization network according to the actual dredging effect.

[0008] In an optional embodiment, The brightness equalization image is subjected to contrast optimization and edge sharpening processing by using bilateral filtering algorithm and fuzzy enhancement algorithm, and the enhancement effect is iteratively optimized by using peak signal-to-noise ratio index and structural similarity index, and the optimized dredging environment image is obtained, including: Calculating pixel gradient values ​​and gradient directions for the brightness-equalized image, adaptively dividing the image area into blocks according to the gradient values, calculating the gradient variance and gradient direction consistency in each block, constructing a spatial weight function of an adaptive Gaussian kernel based on the gradient variance and the gradient direction consistency, performing bilateral filtering, and obtaining a first filtering result; Calculating a local contrast factor in an adaptive window for the first filtering result, wherein the size of the adaptive window is determined according to the gradient value, and performing secondary filtering by multiplying the local contrast factor with a grayscale weight function constructed based on the gradient variance to obtain a bilateral filtered image; Decomposing the bilateral filtered image at multiple scales, calculating local statistical features based on the gradient value at each scale layer, mapping the local statistical features into parameters of a fuzzy membership function, determining weight coefficients of each scale according to the gradient direction consistency, and weighted fusion of the multi-scale fuzzy membership values ​​to obtain fused fuzzy membership values; Calculating the edge degree based on the fused fuzzy membership value, constructing an adaptive enhancement coefficient in combination with the gradient value, multiplying the adaptive enhancement coefficient with the difference between the bilateral filtered image and the original brightness equalized image to obtain an edge enhancement amount, performing multi-scale decomposition on the edge enhancement amount and superimposing it on the original brightness equalized image to obtain an enhanced image; Calculating the peak signal-to-noise ratio, structural similarity, and edge preservation based on gradient features of the enhanced image relative to the original brightness-equalized image, constructing a multi-objective optimization function, and adaptively adjusting the weight coefficients of the optimization indicators according to the gradient variance distribution; The multi-objective optimization function is iteratively optimized using an adaptive learning rate based on a gradient value, bilateral filtering parameters and fuzzy enhancement parameters are updated, and a final optimized image is output when the relative change value of the optimization target after multiple consecutive iterations is less than a preset threshold.

[0009] In an optional embodiment, A three-dimensional voxel grid model is constructed based on the optimized desilting environment image, a state vector including a water velocity field, turbulence intensity and suspended solids concentration is established in each voxel grid unit of the three-dimensional voxel grid model, and a joint probability distribution of environmental interference factors in the voxel grid unit is calculated using a Bayesian probability model, including: Calculating the multi-scale gradient features of the optimized dredging environment image, constructing a structure tensor to obtain local eigenvalues, and performing adaptive meshing and smoothing processing on the dredging environment image based on the local eigenvalues ​​to obtain an optimized three-dimensional voxel grid; In each grid unit of the optimized three-dimensional voxel grid, the velocity field is discretized using a second-order central difference format, a trilinear interpolation basis function is constructed to calculate the grid node velocity component, a pressure Poisson equation is introduced to establish a divergence constraint, a pressure-velocity coupling algorithm is used to solve the intermediate velocity field, and a water flow velocity field that satisfies the divergence constraint of zero is obtained through multiple grid iterations; Performing time-average decomposition on the water flow velocity field to extract velocity pulsation components, calculating turbulent kinetic energy based on the velocity pulsation components, and calculating strain rate tensor based on the water flow velocity field to obtain turbulent dissipation rate, and combining the turbulent kinetic energy with the turbulent dissipation rate to obtain a turbulence intensity characteristic value; Establishing a mapping function between the grayscale value of the optimized desilting environment image and the suspended matter concentration, calibrating the coefficient of the mapping function using measured concentration data, and generating a suspended matter concentration field; The water flow velocity field, the turbulence intensity characteristic quantity and the suspended matter concentration field are combined into an environmental state vector, and the number of components of a Gaussian mixture model is determined based on the Bayesian information criterion, and the prior probability distribution of the environmental state vector is constructed using the Gaussian mixture model; Collect sensor observation data to construct an observation vector, establish an observation likelihood function based on the observation vector, use a covariance matrix to characterize the error correlation of the observation vector, use a Markov chain Monte Carlo method to sample the state vector based on the prior probability distribution and the observation likelihood function, and calculate the posterior probability distribution of the state vector in combination with an adaptive step size to control the sampling acceptance rate; The posterior probability distribution is converted into a joint probability distribution of environmental interference factors in voxel grid units using a kernel density estimation method, wherein a bandwidth parameter of the kernel density estimation is determined by a cross-validation method.

[0010] In an optional embodiment, The water velocity field, the turbulence intensity characteristic quantity and the suspended matter concentration field are combined into an environmental state vector, and the number of components of a Gaussian mixture model is determined based on the Bayesian information criterion. The prior probability distribution of the environmental state vector is constructed by using the Gaussian mixture model, including: Collecting velocity components of the water velocity field, turbulent kinetic energy and turbulent dissipation rate in the turbulence intensity characteristic quantity, and spatial distribution data of the suspended matter concentration field, forming the collected data into an environmental state vector sample set, and constructing a log-likelihood function of a Gaussian mixture model based on the environmental state vector sample set; Based on the dimension of the environmental state vector sample set, the total number of parameters corresponding to each Gaussian component is calculated, the total number of parameters includes the dimension of the mean vector, the number of independent elements of the covariance matrix and the degree of freedom of the mixing weight, the log-likelihood function is combined with the logarithmic term of the total number of parameters to construct the Bayesian information criterion, the number of parameters required for each additional Gaussian component is calculated, the number of parameters, the sample size and the log-likelihood function are substituted into the Bayesian information criterion, and the optimal number of components is obtained through grid search with different numbers of Gaussian components; Initialize the parameters of the Gaussian mixture model using the sample set of the environmental state vector and the optimal number of components, and calculate the posterior probability matrix of each Gaussian component corresponding to each sample based on the current parameters in the expected step; calculate the weighted statistics of the sample based on the posterior probability matrix, and update the mixing weights, mean vectors and covariance matrix corresponding to the optimal number of components through the weighted statistics; Calculating the ratio of the maximum eigenvalue to the minimum eigenvalue of the covariance matrix, and when the ratio exceeds a preset correction threshold, adaptively adjusting the regularization parameter and correcting the covariance matrix; The relative rate of change of the log-likelihood function and the parameter update amount between adjacent iterations are calculated, and the relative rate of change and the update amount are compared with a preset threshold to determine the iteration termination condition. When the iteration termination condition is met, the optimal number of components and their corresponding mixing weights, mean vector and modified covariance matrix are used to construct the prior probability distribution of the environment state vector.

[0011] In an optional embodiment, The mathematical model of environmental interference sources is constructed according to the joint probability distribution, and the Navier-Stokes equation is used to model and solve the silt distribution state to obtain the dynamic evolution law of silt distribution, including: Establishing a sludge motion equation based on the Stokes drag of spherical particles, wherein the sludge motion equation includes a gravity term, a buoyancy term, and a Brownian force term, establishing a Navier-Stokes equation group containing environmental interference sources based on the sludge motion equation, introducing the joint probability distribution of the environmental interference sources into the Navier-Stokes equation group, and constructing a momentum equation including a momentum source term and a mass equation including a mass source term; A sludge concentration transport equation is established according to the Navier-Stokes equations, wherein the sludge concentration transport equation includes a convection diffusion term and a concentration source term, wherein the convection diffusion term is obtained by solving the momentum equation, and the concentration source term is determined by the joint probability distribution of the environmental interference sources; Constructing a staggered grid system, spatially discretizing the momentum equation in the Navier-Stokes equations to obtain a momentum discrete equation, spatially discretizing the mass equation to obtain a mass discrete equation, and temporally discretizing the momentum discrete equation and the mass discrete equation to obtain a space-time discrete equation group; The turbulence kinetic energy and the turbulence dissipation rate are obtained according to the turbulence intensity, the momentum prediction equation in the space-time discrete equation group is introduced through the eddy viscosity coefficient, the intermediate velocity field is obtained by solving the momentum prediction equation, and the pressure Poisson equation is constructed based on the intermediate velocity field; The pressure Poisson equation is solved to obtain a pressure field, the intermediate velocity field is corrected using the pressure field to obtain a corrected velocity field, the corrected velocity field is substituted into the sludge concentration transport equation, the sludge concentration transport equation is solved in an explicit format to obtain the sludge concentration distribution at the current moment, the system boundary conditions for the next time step are updated according to the sludge concentration distribution at the current moment, and the steps from solving the momentum prediction equation to calculating the sludge concentration distribution are repeatedly executed based on the system boundary conditions until the simulation calculation of the specified time is completed to obtain the dynamic evolution law of the sludge distribution.

[0012] In an optional embodiment, Combining the mathematical model of environmental interference sources and the dynamic evolution law of silt distribution to construct a three-dimensional distribution map of silt thickness includes: A sludge deposition rate model is established based on the mathematical model of the environmental interference source and the dynamic evolution law of the sludge distribution, and the joint probability distribution of the environmental interference factors in the voxel grid unit obtained in advance is combined with the sludge concentration gradient to calculate the sludge deposition flux of each voxel grid unit, wherein when the sludge concentration is greater than a first preset threshold and the joint probability distribution of the environmental interference factors is less than a second preset threshold, the sludge deposition flux is determined to be a positive value, and when the joint probability distribution of the environmental interference factors is greater than the second preset threshold, the sludge deposition flux is determined to be a negative value; Integrating the silt deposition flux in the time domain to obtain the theoretical cumulative thickness of the silt in each voxel grid unit, and correcting the theoretical cumulative thickness of the silt according to the compaction correction coefficient to obtain the actual cumulative thickness of the silt in each voxel grid unit; Constructing a silt thickness spatial variation model, calculating the spatial correlation coefficient of silt thickness at different positions based on the silt thickness spatial variation model, and using a spatial interpolation algorithm to interpolate the actual cumulative thickness of the silt from each voxel grid unit to a preset spatial grid node to obtain the silt interpolation thickness of the preset spatial grid node; A spatial display grid is established in a three-dimensional coordinate system, the silt interpolation thickness is mapped onto the spatial display grid to obtain silt thickness display data, a plurality of silt thickness isosurfaces and a plurality of silt thickness profile layers are generated according to the silt thickness display data, and the plurality of silt thickness isosurfaces and the plurality of silt thickness profile layers are combined to form a three-dimensional distribution map of silt thickness.

[0013] In an optional embodiment, The silt category and the three-dimensional distribution map of silt thickness are input into the dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the double temporal difference algorithm, and generates the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot, including: Encoding the silt category information into a category feature vector, performing voxel processing on the three-dimensional distribution map of silt thickness to obtain a spatial feature tensor, inputting the category feature vector and the spatial feature tensor into a feature fusion network to generate a representation of the dredging environment state; Constructing a dredging simulation environment including a sludge resuspension model, a dredging robot kinematic model and an energy consumption model, and obtaining the dredging robot position information, sludge removal volume information and energy consumption information based on the dredging environment state representation and the dredging simulation environment; Establishing a comprehensive reward function according to the sludge removal volume information, the energy consumption information, and the distance information between the dredging robot and the obstacle, wherein the comprehensive reward function includes a dredging efficiency item positively correlated with the sludge removal volume information, an energy consumption item negatively correlated with the energy consumption information, and a safety item related to the distance information; Constructing a strategy network, wherein the strategy network includes an action generation network and a value evaluation network, wherein the action generation network outputs a motion trajectory sequence and an operation parameter sequence of the dredging robot based on the dredging environment state representation, and the value evaluation network predicts long-term benefits based on the dredging environment state representation, the motion trajectory sequence and the operation parameter sequence; The strategy network interacts with the dredging simulation environment to obtain state transition information and reward value information, calculates the value target using a temporal difference algorithm based on the state transition information and the reward value information, and updates the network parameters of the strategy network using a policy gradient method; Construct an experience replay buffer pool, store state transition information and reward value information in the experience replay buffer pool, and obtain training samples from the experience replay buffer pool using random sampling; Build a target network, migrate the network parameters of the policy network to the target network in a soft update manner, provide a learning target based on the target network, and calculate the loss function value of the policy network. When the change in the loss function value over multiple consecutive training cycles is less than the preset change threshold, the policy network training converges. Using the converged policy network, the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot are generated according to the input dredging environment state representation.

[0014] According to a second aspect of the embodiments of the present invention, A dredging strategy optimization system based on reinforcement learning is provided, comprising: The first unit is used to collect a dredging environment image of silt through a depth camera and a sonar sensor, perform multi-scale decomposition based on orthogonal wavelet and curvelet transform to filter out noise on the dredging environment image to obtain a denoised image, use an adaptive histogram equalization algorithm to calculate a grayscale distribution probability density function on the denoised image, perform piecewise linear grayscale stretching and gamma correction according to a peak interval of the grayscale distribution probability density function to obtain a brightness equalized image, use a bilateral filtering algorithm and a fuzzy enhancement algorithm to perform contrast optimization and edge sharpening on the brightness equalized image, iteratively optimize the enhancement effect through a peak signal-to-noise ratio index and a structural similarity index to obtain an optimized dredging environment image; The second unit is used to construct a three-dimensional voxel grid model based on the optimized dredging environment image, establish a state vector including water velocity field, turbulence intensity and suspended matter concentration in each voxel grid unit of the three-dimensional voxel grid model, calculate the joint probability distribution of environmental interference factors in the voxel grid unit using a Bayesian probability model, construct a mathematical model of environmental interference sources according to the joint probability distribution, model and solve the silt distribution state using the Navier-Stokes equations, and obtain the dynamic evolution law of silt distribution; The third unit is used to input the optimized dredging environment image into a deep convolutional neural network, extract the multi-scale texture features, directional gradient features and regional semantic features of the optimized dredging environment image, use the attention weight mechanism to fuse the extracted features to obtain a feature vector, input the feature vector into a classification model to determine the silt category, and construct a three-dimensional distribution map of silt thickness in combination with the mathematical model of the environmental interference source and the dynamic evolution law of silt distribution. The silt category and the three-dimensional distribution map of silt thickness are input into the dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the dual temporal difference algorithm, generates a motion trajectory planning scheme and an operation parameter configuration scheme for the dredging robot, and performs online optimization on the value function of the dredging strategy optimization network according to the actual dredging effect.

[0015] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0017] In this embodiment, image processing technologies such as multi-scale decomposition based on orthogonal wavelet and curvelet transform, adaptive histogram equalization, bilateral filtering and fuzzy enhancement are used to effectively filter out noise, equalize brightness, optimize contrast and sharpen edges, significantly improve the quality of dredging environment images, and lay the foundation for subsequent three-dimensional modeling and feature extraction. A three-dimensional voxel grid model is constructed based on the optimized image, and a mathematical model of environmental interference sources is established in combination with the Bayesian probability model and the Navier-Stokes equation, which can more accurately predict the dynamic evolution law of silt distribution and provide a reliable basis for the formulation of dredging strategies. Image features are extracted using a deep convolutional neural network, combined with silt category and thickness distribution information, and the dredging strategy is optimized through a reinforcement learning algorithm to generate a better robot motion trajectory and operation parameter configuration scheme, thereby improving dredging efficiency and reducing dredging costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a flow chart of a dredging strategy optimization method based on reinforcement learning according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a dredging strategy optimization system based on reinforcement learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG. 1 is a flow chart of a dredging strategy optimization method based on reinforcement learning according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Collect images of the dredging environment of the silt by using a depth camera and a sonar sensor, perform multi-scale decomposition based on orthogonal wavelet and curvelet transform to filter out noise on the dredging environment image, obtain a denoised image, calculate the grayscale distribution probability density function of the denoised image by using an adaptive histogram equalization algorithm, perform piecewise linear grayscale stretching and gamma correction according to the peak interval of the grayscale distribution probability density function to obtain a brightness equalized image, perform contrast optimization and edge sharpening on the brightness equalized image by using a bilateral filtering algorithm and a fuzzy enhancement algorithm, iteratively optimize the enhancement effect by using a peak signal-to-noise ratio index and a structural similarity index, and obtain an optimized dredging environment image; S102. constructing a three-dimensional voxel grid model based on the optimized desilting environment image, establishing a state vector including water velocity field, turbulence intensity and suspended matter concentration in each voxel grid unit of the three-dimensional voxel grid model, calculating the joint probability distribution of environmental interference factors in the voxel grid unit using a Bayesian probability model, constructing a mathematical model of environmental interference sources according to the joint probability distribution, modeling and solving the silt distribution state using the Navier-Stokes equations, and obtaining the dynamic evolution law of silt distribution; S103. Input the optimized dredging environment image into the deep convolutional neural network, extract the multi-scale texture features, directional gradient features and regional semantic features of the optimized dredging environment image, use the attention weight mechanism to fuse the extracted features to obtain a feature vector, input the feature vector into the classification model to determine the silt category, and construct a three-dimensional distribution map of silt thickness in combination with the mathematical model of the environmental interference source and the dynamic evolution law of silt distribution. Input the silt category and the three-dimensional distribution map of silt thickness into the dredging strategy optimization network. The dredging strategy optimization network obtains the dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the dual temporal difference algorithm, generates the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot, and optimizes the value function of the dredging strategy optimization network online according to the actual dredging effect.

[0022] In order to optimize the desilting environment image and plan the desilting strategy, we first collect the image data of the silt distribution, and then obtain the desilting environment image of the target area by combining the depth camera and sonar sensor. The depth camera generates three-dimensional depth information by measuring the reflection time of light, and the sonar sensor generates the contour image of underwater objects and silt distribution through sound wave detection. The data of the two are fused to obtain a more comprehensive and accurate desilting environment image.

[0023] After the image data is collected, in order to eliminate noise interference, multi-scale decomposition and noise reduction processing is performed on the image. Orthogonal wavelet and curvelet transform are used to decompose the image signal into components of different scales and directions. Through the threshold processing method, the components dominated by noise are removed and the image is reconstructed to obtain the de-noised dredging environment image.

[0024] The image after denoising may have uneven grayscale distribution, and the image quality needs to be further enhanced. Adaptive histogram equalization technology is used to estimate the probability density of the grayscale distribution of the image. Based on the peak interval of the probability density function, piecewise linear grayscale stretching is performed, and gamma correction technology is combined to improve the balance of brightness distribution. This process makes the grayscale contrast in the image more significant and enhances the distinction between the silt area and the background.

[0025] Subsequently, in order to optimize the contrast and detail of the image, the bilateral filtering algorithm is used to smooth the image while retaining the edge details. The edges of the image are sharpened by the fuzzy enhancement algorithm to further enhance the detail of the key areas. At this stage, the image enhancement effect is evaluated by calculating indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and the iterative optimization technology is used to continuously adjust the parameters to finally generate a high-quality optimized image.

[0026] In an optional implementation, the brightness equalization image is subjected to contrast optimization and edge sharpening processing using a bilateral filtering algorithm and a fuzzy enhancement algorithm, and the enhancement effect is iteratively optimized using a peak signal-to-noise ratio index and a structural similarity index, to obtain an optimized dredging environment image, including: Calculating pixel gradient values ​​and gradient directions for the brightness-equalized image, adaptively dividing the image area into blocks according to the gradient values, calculating the gradient variance and gradient direction consistency in each block, constructing a spatial weight function of an adaptive Gaussian kernel based on the gradient variance and the gradient direction consistency, performing bilateral filtering, and obtaining a first filtering result; Calculating a local contrast factor in an adaptive window for the first filtering result, wherein the size of the adaptive window is determined according to the gradient value, and performing secondary filtering by multiplying the local contrast factor with a grayscale weight function constructed based on the gradient variance to obtain a bilateral filtered image; Decomposing the bilateral filtered image at multiple scales, calculating local statistical features based on the gradient value at each scale layer, mapping the local statistical features into parameters of a fuzzy membership function, determining weight coefficients of each scale according to the gradient direction consistency, and weighted fusion of the multi-scale fuzzy membership values ​​to obtain fused fuzzy membership values; Calculating the edge degree based on the fused fuzzy membership value, constructing an adaptive enhancement coefficient in combination with the gradient value, multiplying the adaptive enhancement coefficient with the difference between the bilateral filtered image and the original brightness equalized image to obtain an edge enhancement amount, performing multi-scale decomposition on the edge enhancement amount and superimposing it on the original brightness equalized image to obtain an enhanced image; Calculating the peak signal-to-noise ratio, structural similarity, and edge preservation based on gradient features of the enhanced image relative to the original brightness-equalized image, constructing a multi-objective optimization function, and adaptively adjusting the weight coefficients of the optimization indicators according to the gradient variance distribution; The multi-objective optimization function is iteratively optimized using an adaptive learning rate based on a gradient value, bilateral filtering parameters and fuzzy enhancement parameters are updated, and a final optimized image is output when the relative change value of the optimization target after multiple consecutive iterations is less than a preset threshold.

[0027] Exemplarily, first, the input brightness equalization image is preprocessed. The preprocessing step includes converting the image into a grayscale image and removing noise from the image. Noise removal can use methods such as median filtering or Gaussian filtering to remove small noise points in the image and provide cleaner image data for subsequent processing. For example, using a median filter with a 3x3 window, the value of each pixel is replaced by the median of its neighboring pixel values.

[0028] Next, calculate the gradient value and gradient direction of each pixel in the image. The gradient value can be calculated using the Sobel operator, which approximates the gradient of the image by calculating the pixel differences in the horizontal and vertical directions. The gradient direction can be calculated using the inverse tangent function. For example, for a pixel (x, y), its horizontal gradient Gx=I(x+1, y)-I(x-1, y), vertical gradient Gy=I(x, y+1)-I(x, y-1), gradient amplitude G=sqrt(Gx^2+Gy^2), and gradient direction θ=arctan(Gy / Gx).

[0029] According to the calculated gradient value, the image is divided into several blocks of different sizes. Areas with larger gradient values ​​are divided into smaller blocks, and areas with smaller gradient values ​​are divided into larger blocks. This can better preserve image details and avoid over-smoothing. For example, a gradient threshold T can be set, and areas with gradient values ​​greater than T are divided into 5x5 blocks, and areas with gradient values ​​less than T are divided into 10x10 blocks.

[0030] In each block, the gradient variance and gradient direction consistency are calculated. The gradient variance reflects the texture complexity of the area, and the gradient direction consistency reflects the clarity of the edge of the area. For example, the gradient variance can be obtained by calculating the sum of the squares of the difference between the gradient value of each pixel and the average gradient value in the block. The gradient direction consistency can be obtained by calculating the average value of the cosine similarity of the gradient directions of all pixels in the block.

[0031] Based on the gradient variance and gradient direction consistency, a spatial weight function is constructed. This function is used to control the smoothness of bilateral filtering. The larger the gradient variance, the smaller the weight and the lower the smoothness; the higher the gradient direction consistency, the larger the weight and the higher the smoothness. The constructed spatial weight function is used to perform bilateral filtering to obtain the first filtering result. Bilateral filtering is a nonlinear filtering method that takes into account both the spatial distance and grayscale value difference between pixels, and can effectively remove noise and retain edges.

[0032] The first filtering result is subjected to secondary filtering. A local contrast factor is calculated in an adaptive window, and the window size is determined according to the gradient value. The local contrast factor is multiplied by a grayscale weight function constructed based on the gradient variance, and secondary filtering is performed to obtain a bilaterally filtered image. For example, the local contrast factor can be defined as the ratio of the current pixel grayscale value to the average grayscale value in the window.

[0033] Perform multi-scale decomposition on the bilateral filtered image, for example, using wavelet transform. At each scale level, calculate local statistical features, such as mean and variance, based on the gradient value. Map these features to parameters of the fuzzy membership function. Determine the weight coefficient of each scale based on the consistency of the gradient direction, and perform weighted fusion on the multi-scale fuzzy membership values ​​to obtain the final fuzzy membership value.

[0034] The edge degree is calculated based on the fused fuzzy membership value. The adaptive enhancement coefficient is constructed in combination with the gradient value, and the coefficient is multiplied by the difference between the bilateral filtered image and the original brightness equalized image to obtain the edge enhancement amount. After multi-scale decomposition of the edge enhancement amount, it is superimposed on the original brightness equalized image to obtain the enhanced image.

[0035] Finally, an iterative optimization method is used to adjust the bilateral filter parameters and blur enhancement parameters according to indicators such as peak signal-to-noise ratio, structural similarity, and edge preservation until the optimization target reaches a stable state and the final optimized image is output. For example, a gradient descent method can be used for iterative optimization.

[0036] In this embodiment, through adaptive blocking and Gaussian kernel bilateral filtering, noise is effectively suppressed while retaining details, enhancing the local contrast and edge clarity of the image. Multi-scale fuzzy membership fusion and gradient consistency analysis further optimize the detail performance of the image, making the edge more prominent. Through multi-objective optimization functions such as peak signal-to-noise ratio, structural similarity and edge retention, the optimization effect can be fully evaluated and adjusted to ensure that the image quality is optimally improved in various complex environments. Finally, through the adjustment of adaptive learning rate and gradient value, the optimization process is efficient and stable, and the final output image has higher visual quality and adaptability.

[0037] In an optional embodiment, a three-dimensional voxel grid model is constructed based on the optimized dredging environment image, a state vector including a water velocity field, a turbulence intensity and a suspended matter concentration is established in each voxel grid unit of the three-dimensional voxel grid model, and a joint probability distribution of environmental interference factors in the voxel grid unit is calculated using a Bayesian probability model, including: Calculating the multi-scale gradient features of the optimized dredging environment image, constructing a structure tensor to obtain local eigenvalues, and performing adaptive meshing and smoothing processing on the dredging environment image based on the local eigenvalues ​​to obtain an optimized three-dimensional voxel grid; In each grid unit of the optimized three-dimensional voxel grid, the velocity field is discretized using a second-order central difference format, a trilinear interpolation basis function is constructed to calculate the grid node velocity component, a pressure Poisson equation is introduced to establish a divergence constraint, a pressure-velocity coupling algorithm is used to solve the intermediate velocity field, and a water flow velocity field that satisfies the divergence constraint of zero is obtained through multiple grid iterations; Performing time-average decomposition on the water flow velocity field to extract velocity pulsation components, calculating turbulent kinetic energy based on the velocity pulsation components, and calculating strain rate tensor based on the water flow velocity field to obtain turbulent dissipation rate, and combining the turbulent kinetic energy with the turbulent dissipation rate to obtain a turbulence intensity characteristic value; Establishing a mapping function between the grayscale value of the optimized desilting environment image and the suspended matter concentration, calibrating the coefficient of the mapping function using measured concentration data, and generating a suspended matter concentration field; The water flow velocity field, the turbulence intensity characteristic quantity and the suspended matter concentration field are combined into an environmental state vector, and the number of components of a Gaussian mixture model is determined based on the Bayesian information criterion, and the prior probability distribution of the environmental state vector is constructed using the Gaussian mixture model; Collect sensor observation data to construct an observation vector, establish an observation likelihood function based on the observation vector, use a covariance matrix to characterize the error correlation of the observation vector, use a Markov chain Monte Carlo method to sample the state vector based on the prior probability distribution and the observation likelihood function, and calculate the posterior probability distribution of the state vector in combination with an adaptive step size to control the sampling acceptance rate; The posterior probability distribution is converted into a joint probability distribution of environmental interference factors in voxel grid units using a kernel density estimation method, wherein a bandwidth parameter of the kernel density estimation is determined by a cross-validation method.

[0038] Exemplarily, first, image data of the dredging environment is obtained and optimized. Gradient features of different scales in the image, such as edge and texture information, are calculated. A structure tensor is constructed using these gradient features, and its local eigenvalues ​​are calculated. According to the size of these eigenvalues, the image is adaptively divided to generate an initial three-dimensional voxel grid. In order to obtain a smoother grid, the generated grid is smoothed, such as Laplacian smoothing, and finally an optimized three-dimensional voxel grid is obtained. For example, the gradient features of the image can be divided into three scales: fine scale, medium scale and coarse scale, corresponding to the details, texture and contour information of the image, respectively.

[0039] Next, the water velocity field is calculated in each grid cell of the optimized three-dimensional voxel grid. The velocity field is discretized using the second-order central difference format, and the velocity components of the grid nodes are calculated using trilinear interpolation basis functions. In order to ensure the conservation of fluid mass, the pressure Poisson equation is introduced to establish the divergence constraint, and the pressure-velocity coupling algorithm is used to solve the intermediate velocity field. Finally, the water velocity field that satisfies the zero divergence constraint is obtained through the multi-grid iteration method. For example, in a 10x10x10 grid, the size of each grid cell is 0.1 meters and the time step is 0.01 seconds.

[0040] Then, the turbulence intensity is calculated. The water flow velocity field is decomposed into the time-averaged velocity and the velocity pulsation component. The turbulent kinetic energy is calculated based on the velocity pulsation component. At the same time, the strain rate tensor is calculated based on the water flow velocity field, and the turbulence dissipation rate is obtained. The turbulence kinetic energy and the turbulence dissipation rate are combined to obtain the turbulence intensity characteristic quantity. For example, the variance of the velocity pulsation component in each grid cell can be calculated as the turbulent kinetic energy, and the turbulence dissipation rate can be calculated based on the strain rate tensor.

[0041] Next, calculate the suspended matter concentration field. Establish a mapping function between the grayscale value of the optimized dredging environment image and the suspended matter concentration. Use the measured concentration data to calibrate the coefficients of the mapping function and finally generate the suspended matter concentration field. For example, it can be assumed that the grayscale value is linearly related to the suspended matter concentration, and the least squares method is used to fit the coefficients of the linear function. Assuming that the measured data shows that a grayscale value of 50 corresponds to a concentration of 10 mg / L, and a grayscale value of 100 corresponds to a concentration of 20 mg / L, a linear relationship can be fitted: concentration = 0.2*grayscale value.

[0042] The water velocity field, turbulence intensity characteristic and suspended solids concentration field are combined into an environmental state vector. The number of components of the Gaussian mixture model is determined using the Bayesian information criterion, and the model is used to construct the prior probability distribution of the environmental state vector. For example, if the Bayesian information criterion determines that the optimal number of components is 3, a mixture model of three Gaussian distributions is used to represent the prior probability distribution of the environmental state vector.

[0043] Collect sensor observation data and construct observation vectors. Build observation likelihood function based on observation vectors, and use covariance matrix to characterize error correlation of observation vectors. Use Markov chain Monte Carlo method to sample state vectors based on prior probability distribution and observation likelihood function. Combined with adaptive step size to control sampling acceptance rate, calculate posterior probability distribution of state vectors.

[0044] Finally, the kernel density estimation method is used to transform the posterior probability distribution into the joint probability distribution of environmental disturbance factors in the voxel grid unit. The bandwidth parameter of the kernel density estimation is determined by the cross-validation method. For example, a five-fold cross-validation can be used to select the optimal bandwidth parameter.

[0045] In this embodiment, by constructing a three-dimensional voxel grid model and considering key factors such as water velocity field, turbulence intensity and suspended matter concentration, the complex dynamic changes of the dredging environment can be simulated more accurately. Through the Bayesian probability model and Markov chain Monte Carlo method, the uncertainty of environmental interference factors can be quantified and their joint probability distribution can be predicted, providing a reliable basis for environmental risk assessment. Through accurate simulation and probabilistic prediction of the dredging environment, the dredging plan can be optimized, environmental risks can be reduced, and dredging efficiency can be improved.

[0046] In an optional embodiment, the water flow velocity field, the turbulence intensity characteristic quantity and the suspended matter concentration field are combined into an environmental state vector, and the number of components of a Gaussian mixture model is determined based on the Bayesian information criterion. The prior probability distribution of the environmental state vector constructed using the Gaussian mixture model includes: Collecting velocity components of the water velocity field, turbulent kinetic energy and turbulent dissipation rate in the turbulence intensity characteristic quantity, and spatial distribution data of the suspended matter concentration field, forming the collected data into an environmental state vector sample set, and constructing a log-likelihood function of a Gaussian mixture model based on the environmental state vector sample set; Based on the dimension of the environmental state vector sample set, the total number of parameters corresponding to each Gaussian component is calculated, the total number of parameters includes the dimension of the mean vector, the number of independent elements of the covariance matrix and the degree of freedom of the mixing weight, the log-likelihood function is combined with the logarithmic term of the total number of parameters to construct the Bayesian information criterion, the number of parameters required for each additional Gaussian component is calculated, the number of parameters, the sample size and the log-likelihood function are substituted into the Bayesian information criterion, and the optimal number of components is obtained through grid search with different numbers of Gaussian components; Initialize the parameters of the Gaussian mixture model using the sample set of the environmental state vector and the optimal number of components, and calculate the posterior probability matrix of each Gaussian component corresponding to each sample based on the current parameters in the expected step; calculate the weighted statistics of the sample based on the posterior probability matrix, and update the mixing weights, mean vectors and covariance matrix corresponding to the optimal number of components through the weighted statistics; Calculating the ratio of the maximum eigenvalue to the minimum eigenvalue of the covariance matrix, and when the ratio exceeds a preset correction threshold, adaptively adjusting the regularization parameter and correcting the covariance matrix; The relative rate of change of the log-likelihood function and the parameter update amount between adjacent iterations are calculated, and the relative rate of change and the update amount are compared with a preset threshold to determine the iteration termination condition. When the iteration termination condition is met, the optimal number of components and their corresponding mixing weights, mean vector and modified covariance matrix are used to construct the prior probability distribution of the environment state vector.

[0047] Exemplarily, first, collect environmental data. Use sensors carried by sensor networks or mobile platforms to collect velocity components of water flow velocity fields (x, y, and z components of three-dimensional velocity), turbulence intensity characteristics (turbulent kinetic energy and turbulence dissipation rate), and spatial distribution data of suspended matter concentration fields. For example, 10 measurement points are evenly arranged in a specific water area, and 100 sets of data are collected at each point to obtain 1,000 sample data in 6 dimensions, including 3 velocity components, 2 turbulence intensity characteristics, and 1 suspended matter concentration. These data are combined into an environmental state vector sample set.

[0048] Next, determine the number of components of the Gaussian mixture model. According to the dimension of the collected environmental state vector sample set (6 dimensions in this case), calculate the number of parameters that need to be estimated for each Gaussian component. For example, the mean vector has 6 parameters, the covariance matrix has 21 independent elements because it is a symmetric matrix, and the mixing weight has one degree of freedom (because the sum of all weights is 1), so each component has a total of 28 parameters. Then, construct the log-likelihood function of the Gaussian mixture model based on the environmental state vector sample set. Combine the log-likelihood function with the logarithm of the total number of parameters to construct the Bayesian information criterion. Assume that the number of test components ranges from 1 to 10, and calculate the Bayesian information criterion value corresponding to each number of components. By comparing the Bayesian information criterion values ​​corresponding to different numbers of components, select the number of components that maximizes the Bayesian information criterion value as the optimal number of components. Assume that in this example, the optimal number of components is determined to be 3.

[0049] Then, initialize the Gaussian mixture model parameters. Use the determined optimal number of components (for example, 3) and the environment state vector sample set to initialize the parameters of the Gaussian mixture model, including the mixture weight, mean vector, and covariance matrix of each component. The k-means algorithm can be used for initialization.

[0050] Next, an expectation maximization (EM) iterative optimization is performed. In the expectation step, the posterior probability of each sample belonging to each Gaussian component is calculated based on the current parameters. Then, in the maximization step, weighted statistics of the samples, such as weighted mean and weighted covariance, are calculated based on the posterior probabilities. These weighted statistics are used to update the mixing weights, mean vector, and covariance matrix of each component.

[0051] In order to avoid the covariance matrix from being singular during the iteration process, the covariance matrix needs to be corrected. The ratio of the maximum eigenvalue to the minimum eigenvalue of the covariance matrix is ​​calculated. If the ratio exceeds the preset correction threshold (for example, 100), the regularization parameter is adaptively adjusted and the regularization term is added to the diagonal of the covariance matrix for correction.

[0052] In order to determine the termination condition of the iteration, it is necessary to calculate the relative change rate of the log-likelihood function and the parameter update amount between adjacent iterations. If both the relative change rate and the parameter update amount are less than a preset threshold (for example, 0.001), the iteration is considered to have converged and the iteration is terminated.

[0053] Finally, the prior probability distribution of the environment state vector is constructed. The prior probability distribution of the environment state vector is constructed using the optimal number of components (e.g., 3) and their corresponding mixing weights, mean vectors, and modified covariance matrices.

[0054] In this embodiment, the probability distribution of multiple factors in a complex water environment can be accurately described, so that the mobile robot can perceive and understand the surrounding environment more accurately, thereby improving its adaptability and autonomous navigation capabilities in complex environments. By considering the probability distribution of environmental factors, the robot can more effectively predict environmental changes and plan safer and more reliable paths to avoid potential risks and improve the success rate of task completion. The Bayesian information criterion is used to automatically determine the optimal number of components of the Gaussian mixture model, and the model parameters are iteratively optimized through the expectation maximization algorithm, which reduces the workload of manual parameter adjustment and improves the practicality and efficiency of the model.

[0055] In an optional implementation, a mathematical model of environmental interference sources is constructed according to the joint probability distribution, and the sludge distribution state is modeled and solved using the Navier-Stokes equations to obtain the dynamic evolution law of sludge distribution, including: Establishing a sludge motion equation based on the Stokes drag of spherical particles, wherein the sludge motion equation includes a gravity term, a buoyancy term, and a Brownian force term, establishing a Navier-Stokes equation group containing environmental interference sources based on the sludge motion equation, introducing the joint probability distribution of the environmental interference sources into the Navier-Stokes equation group, and constructing a momentum equation including a momentum source term and a mass equation including a mass source term; A sludge concentration transport equation is established according to the Navier-Stokes equations, wherein the sludge concentration transport equation includes a convection diffusion term and a concentration source term, wherein the convection diffusion term is obtained by solving the momentum equation, and the concentration source term is determined by the joint probability distribution of the environmental interference sources; Constructing a staggered grid system, spatially discretizing the momentum equation in the Navier-Stokes equations to obtain a momentum discrete equation, spatially discretizing the mass equation to obtain a mass discrete equation, and temporally discretizing the momentum discrete equation and the mass discrete equation to obtain a space-time discrete equation group; The turbulence kinetic energy and the turbulence dissipation rate are obtained according to the turbulence intensity, the momentum prediction equation in the space-time discrete equation group is introduced through the eddy viscosity coefficient, the intermediate velocity field is obtained by solving the momentum prediction equation, and the pressure Poisson equation is constructed based on the intermediate velocity field; The pressure Poisson equation is solved to obtain a pressure field, the intermediate velocity field is corrected using the pressure field to obtain a corrected velocity field, the corrected velocity field is substituted into the sludge concentration transport equation, the sludge concentration transport equation is solved in an explicit format to obtain the sludge concentration distribution at the current moment, the system boundary conditions for the next time step are updated according to the sludge concentration distribution at the current moment, and the steps from solving the momentum prediction equation to calculating the sludge concentration distribution are repeatedly executed based on the system boundary conditions until the simulation calculation of the specified time is completed to obtain the dynamic evolution law of the sludge distribution.

[0056] Exemplarily, first, establish the motion equation of the sludge particles. Simplify the sludge particles into spheres and consider the various forces they are subject to in the water, including the downward force generated by the earth's gravity, the upward buoyancy generated by the water, and the Brownian force generated by the random collision of water molecules with the particles. These forces jointly determine the motion state of the sludge particles. For example, a sludge particle with a diameter of 0.1 mm and a density of 2650 kg per cubic meter is subject to a gravity of about 5.14x10^-10 Newtons and a buoyancy of about 4.08x10^-10 Newtons in water.

[0057] Next, we construct the Navier-Stokes equations that describe the motion of water flow. This equation group contains the momentum equation and the mass equation, which describe the velocity change and mass conservation of water flow, respectively. In order to consider the impact of environmental disturbances, the joint probability distribution of environmental disturbance sources is introduced. Assuming that the joint probability distribution of wind speed and water flow velocity is a normal distribution, it can be introduced into the equation group to simulate the impact of wind on water flow. This makes the momentum equation contain a momentum source term and the mass equation contain a mass source term.

[0058] Then, the silt concentration transport equation is established. This equation describes the change of silt concentration over time and space. The equation contains convection diffusion terms and concentration source terms. The convection diffusion term is obtained from the velocity field in the Navier-Stokes equations solved previously, and the concentration source term is determined by the joint probability distribution of environmental interference sources. For example, if rainfall causes an increase in silt concentration in a certain area, it can be reflected by the concentration source term.

[0059] In order to solve these equations, a staggered grid system is constructed to divide the calculation area into many small grids. The Navier-Stokes equations and sludge concentration transport equations established previously are discretized on these grids and converted into a set of algebraic equations that can be solved by computer. After spatial discretization, time discretization is performed to divide the continuous time into discrete time steps, and finally a set of space-time discrete equations is obtained.

[0060] According to the turbulence intensity of the water flow, the turbulent kinetic energy and turbulent dissipation rate are calculated and converted into eddy viscosity coefficients, which are introduced into the momentum prediction equation in the space-time discrete equation group. The momentum prediction equation is solved to obtain an intermediate velocity field.

[0061] Based on this intermediate velocity field, the pressure Poisson equation is constructed and solved to obtain the pressure field. The intermediate velocity field is corrected using the pressure field to obtain the final corrected velocity field.

[0062] Substitute the modified velocity field into the silt concentration transport equation and solve it in explicit format to obtain the silt concentration distribution at the current moment. For example, at a certain time step, the silt concentration of a certain grid is calculated to be 0.05 kg per cubic meter.

[0063] According to the silt concentration distribution at the current moment, the system boundary conditions for the next time step are updated. For example, if the silt concentration reaches a certain threshold, it is considered that siltation occurs at the boundary and the boundary conditions need to be adjusted.

[0064] Repeat the steps from solving the momentum prediction equation to calculating the silt concentration distribution until the simulation calculation of the specified time is completed, such as simulating the silt transport process for 10 hours, and finally obtain the dynamic evolution law of silt distribution.

[0065] In this embodiment, by considering the joint probability distribution of environmental interference sources and a more sophisticated silt particle motion model, the movement and distribution of silt in a complex environment can be simulated more accurately. This method can be applied to various environmental conditions, such as different water depths, different flow rates, different wind forces, etc., and has strong adaptability. By adopting an efficient numerical calculation method, the calculation efficiency can be improved and the simulation time can be shortened while ensuring the calculation accuracy.

[0066] In an optional implementation, combining the environmental interference source mathematical model and the dynamic evolution law of sludge distribution to construct a three-dimensional distribution map of sludge thickness includes: A sludge deposition rate model is established based on the mathematical model of the environmental interference source and the dynamic evolution law of the sludge distribution, and the joint probability distribution of the environmental interference factors in the voxel grid unit obtained in advance is combined with the sludge concentration gradient to calculate the sludge deposition flux of each voxel grid unit, wherein when the sludge concentration is greater than a first preset threshold and the joint probability distribution of the environmental interference factors is less than a second preset threshold, the sludge deposition flux is determined to be a positive value, and when the joint probability distribution of the environmental interference factors is greater than the second preset threshold, the sludge deposition flux is determined to be a negative value; Integrating the silt deposition flux in the time domain to obtain the theoretical cumulative thickness of the silt in each voxel grid unit, and correcting the theoretical cumulative thickness of the silt according to the compaction correction coefficient to obtain the actual cumulative thickness of the silt in each voxel grid unit; Constructing a silt thickness spatial variation model, calculating the spatial correlation coefficient of silt thickness at different positions based on the silt thickness spatial variation model, and using a spatial interpolation algorithm to interpolate the actual cumulative thickness of the silt from each voxel grid unit to a preset spatial grid node to obtain the silt interpolation thickness of the preset spatial grid node; A spatial display grid is established in a three-dimensional coordinate system, the silt interpolation thickness is mapped onto the spatial display grid to obtain silt thickness display data, a plurality of silt thickness isosurfaces and a plurality of silt thickness profile layers are generated according to the silt thickness display data, and the plurality of silt thickness isosurfaces and the plurality of silt thickness profile layers are combined to form a three-dimensional distribution map of silt thickness.

[0067] Exemplarily, obtaining environmental disturbance data and silt distribution data is the first step in constructing a three-dimensional distribution map of silt thickness. Environmental disturbance data include water flow velocity, wind speed, wave height, etc., which can be obtained through sensors, remote sensing images, etc. Silt distribution data include silt concentration, particle size distribution, etc., which can be obtained through field sampling, laboratory analysis, etc. For example, in a specific case, we collect water flow velocity, wind speed, wave height and other data every hour through a sensor network arranged in the water area, and collect water samples every other day through a water quality monitoring station for laboratory analysis to obtain silt concentration and particle size distribution data.

[0068] Next, a sludge deposition rate model is established. This model links environmental disturbance factors with sludge concentration gradients and is used to calculate the sludge deposition flux of each voxel grid unit. First, the acquired environmental disturbance data is subjected to probability distribution modeling, such as using Gaussian mixture model or kernel density estimation to obtain the joint probability distribution of environmental disturbance factors. Then, the sludge concentration gradient is calculated based on the sludge concentration data. Finally, the joint probability distribution of environmental disturbance factors is combined with the sludge concentration gradient to calculate the sludge deposition flux of each voxel grid unit. Two thresholds are set: the first preset threshold is used to determine whether the sludge concentration is high enough, and the second preset threshold is used to determine the degree of influence of environmental disturbance factors. When the sludge concentration is greater than the first preset threshold and the joint probability distribution of environmental disturbance factors is less than the second preset threshold, the sludge deposition flux is determined to be a positive value, indicating sludge deposition; when the joint probability distribution of environmental disturbance factors is greater than the second preset threshold, the sludge deposition flux is determined to be a negative value, indicating that the sludge is washed away. Assume that the first preset threshold is 100 mg / L and the second preset threshold is 0.5. In a voxel grid unit, if the silt concentration is 150 mg / L and the joint probability distribution of environmental disturbance factors is 0.3, the silt deposition flux is positive; if the joint probability distribution of environmental disturbance factors is 0.7, the silt deposition flux is negative.

[0069] Then, calculate the cumulative thickness of the silt. Integrate the silt deposition flux of each voxel grid unit in the time domain to obtain the theoretical cumulative thickness of the silt. Since compaction will occur after the silt is deposited, the theoretical cumulative thickness of the silt needs to be corrected according to the compaction correction coefficient to obtain the actual cumulative thickness of the silt. The compaction correction coefficient can be determined by laboratory tests or empirical formulas. Assuming that the silt deposition flux of a voxel grid unit is 1mm / day over a period of time, and the time period is 30 days, the theoretical cumulative thickness of the silt is 30mm. Assuming the compaction correction coefficient is 0.8, the actual cumulative thickness of the silt is 24mm.

[0070] In order to construct a complete three-dimensional distribution map, the calculated actual cumulative thickness of silt needs to be interpolated from the voxel grid unit to the preset spatial grid node. First, a spatial variation model of silt thickness is established, such as using a Kriging interpolation model or an inverse distance weighted interpolation model. Then, the spatial correlation coefficient of silt thickness at different locations is calculated based on the established spatial variation model. Finally, a spatial interpolation algorithm is used to interpolate the actual cumulative thickness of silt from each voxel grid unit to the preset spatial grid node to obtain the interpolated thickness of the silt at the preset spatial grid node.

[0071] Finally, the interpolated silt thickness is mapped to the spatial display grid in the three-dimensional coordinate system to obtain the silt thickness display data. According to the silt thickness display data, multiple silt thickness isosurfaces and multiple silt thickness profile layers are generated, and these isosurfaces and profile layers are combined to form a three-dimensional distribution map of silt thickness, which intuitively displays the spatial distribution of silt thickness.

[0072] In this embodiment, by combining environmental interference factors and the dynamic evolution law of silt distribution, the silt deposition process is simulated more accurately, thereby improving the accuracy of silt thickness prediction. The silt thickness data is mapped to three-dimensional space, and isosurfaces and profile layers are generated to achieve three-dimensional visualization of silt thickness, and the spatial distribution characteristics of silt thickness are more intuitively displayed. Accurate three-dimensional distribution maps of silt thickness can provide important basic data for projects such as silt dredging and waterway maintenance, helping to optimize engineering design and construction plans.

[0073] In an optional implementation, the silt category and the three-dimensional distribution map of silt thickness are input into a dredging strategy optimization network, the dredging strategy optimization network obtains dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the double temporal difference algorithm, and generates a motion trajectory planning scheme and an operation parameter configuration scheme for the dredging robot, including: Encoding the silt category information into a category feature vector, performing voxel processing on the three-dimensional distribution map of silt thickness to obtain a spatial feature tensor, inputting the category feature vector and the spatial feature tensor into a feature fusion network to generate a representation of the dredging environment state; Constructing a dredging simulation environment including a sludge resuspension model, a dredging robot kinematic model and an energy consumption model, and obtaining the dredging robot position information, sludge removal volume information and energy consumption information based on the dredging environment state representation and the dredging simulation environment; Establishing a comprehensive reward function according to the sludge removal volume information, the energy consumption information, and the distance information between the dredging robot and the obstacle, wherein the comprehensive reward function includes a dredging efficiency item positively correlated with the sludge removal volume information, an energy consumption item negatively correlated with the energy consumption information, and a safety item related to the distance information; Constructing a strategy network, wherein the strategy network includes an action generation network and a value evaluation network, wherein the action generation network outputs a motion trajectory sequence and an operation parameter sequence of the dredging robot based on the dredging environment state representation, and the value evaluation network predicts long-term benefits based on the dredging environment state representation, the motion trajectory sequence and the operation parameter sequence; The strategy network interacts with the dredging simulation environment to obtain state transition information and reward value information, calculates the value target using a temporal difference algorithm based on the state transition information and the reward value information, and updates the network parameters of the strategy network using a policy gradient method; Construct an experience replay buffer pool, store state transition information and reward value information in the experience replay buffer pool, and obtain training samples from the experience replay buffer pool using random sampling; Build a target network, migrate the network parameters of the policy network to the target network in a soft update manner, provide a learning target based on the target network, and calculate the loss function value of the policy network. When the change in the loss function value over multiple consecutive training cycles is less than the preset change threshold, the policy network training converges. Using the converged policy network, the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot are generated according to the input dredging environment state representation.

[0074] Exemplarily, first obtain the silt category and thickness information of the area to be desilted. Silt category information such as silt composition, particle size, etc. can be represented by different codes, for example, "silt A" is encoded as [1, 0, 0], "silt B" is encoded as [0, 1, 0], etc., to form a category feature vector. The three-dimensional distribution map of silt thickness can be obtained by measuring with sonar, multi-beam and other equipment, and voxelized, and each voxel records the silt thickness at that location to form a spatial feature tensor. For example, if the silt thickness of a voxel is 0.05m, the value of the corresponding position in the spatial feature tensor is 0.05.

[0075] The obtained category feature vector and spatial feature tensor are input into the feature fusion network. The network can be a convolutional neural network, a recurrent neural network or a combination thereof, which is used to extract and fuse the silt category and thickness information. For example, the spatial feature tensor can be processed by a convolutional neural network to extract the spatial features, and then the extracted spatial features are concatenated with the category feature vector and input into a fully connected network, and finally a fixed-length vector is output as a representation of the dredging environment state.

[0076] In order to train the control strategy, a dredging simulation environment is constructed. The environment includes a sludge resuspension model, a dredging robot kinematic model, and an energy consumption model. The sludge resuspension model simulates the diffusion and sedimentation of sludge during dredging, and the diffusion range of sludge can be calculated based on the water flow velocity and the size of sludge particles. The dredging robot kinematic model simulates the movement of the robot, and calculates the robot's posture change based on the robot's propeller power and direction. The energy consumption model simulates the energy consumption of the robot under different motion states and operating parameters, for example, the energy consumption is calculated based on the robot's propeller power and operating time. By inputting the dredging environment state representation into the simulation environment, the posture information, the volume of sludge removed, and the energy consumption information of the dredging robot can be obtained. At a certain moment, the robot is located at (x=2m, y=3m, z=1m), removes 0.01 cubic meters of sludge, and consumes 0.001kWh of energy.

[0077] According to the information returned by the simulation environment, a comprehensive reward function is established. The function contains the silt removal efficiency term, the energy consumption term and the safety term. The silt removal efficiency term is positively correlated with the volume of silt removed. For example, the larger the volume of silt removed, the higher the reward. The energy consumption term is negatively correlated with energy consumption. The lower the energy consumption, the higher the reward. The safety term is related to the distance between the robot and the obstacle. For example, the farther away from the obstacle, the higher the reward. Assuming that at a certain moment, the robot has removed 0.01 cubic meters of silt, consumed 0.001 kWh of energy, and is 1 meter away from the nearest obstacle, the total reward value can be calculated based on the pre-set weight coefficient.

[0078] Construct a strategy network, which includes an action generation network and a value evaluation network. The action generation network takes the desilting environment state representation as input and outputs the motion trajectory sequence and operation parameter sequence of the desilting robot, such as the robot's speed in the x, y, and z directions and the power of the dredge pump in the next 5 seconds. The value evaluation network takes the desilting environment state representation, motion trajectory sequence, and operation parameter sequence as input to predict long-term benefits.

[0079] The policy network is interacted with the dredging simulation environment for training. The actions output by the policy network control the robot to move and operate in the simulation environment, and the environment returns state transition information and reward value information. The value target is calculated using the temporal difference algorithm, and the policy network parameters are updated using the policy gradient method.

[0080] In order to improve the stability of training, an experience replay buffer pool is constructed, and the state transition information and reward value information are stored in the buffer pool, and training samples are obtained from it by random sampling. At the same time, a target network is constructed, and the parameters of the policy network are migrated to the target network in a soft update manner. The target network is used to provide learning objectives and calculate the loss function value of the policy network. When the change in the loss function value in multiple consecutive training cycles is less than the preset change threshold, the policy network training is determined to be converged.

[0081] After the training converges, the policy network can be used to control the actual dredging robot. By inputting the state information of the actual dredging environment into the trained policy network, the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot can be generated.

[0082] In this embodiment, by combining sludge category information with a three-dimensional distribution map to generate a representation of the dredging environment state, the strategy optimization network can efficiently plan the motion trajectory and operating parameters of the dredging robot. Through the interaction between the dredging simulation environment and the robot, the efficiency, energy consumption and safety of sludge removal are evaluated in real time, and these factors are comprehensively considered to optimize the dredging operation strategy. The training of the strategy network uses a dual temporal difference algorithm and an experience replay mechanism to ensure that the network gradually learns the optimal decision-making strategy. Ultimately, the generated motion trajectory planning scheme and operating parameter configuration scheme can maximize dredging efficiency, reduce energy consumption and ensure operational safety, significantly improving the automation and intelligence level of dredging operations.

[0083] Figure 2 FIG. 1 is a schematic diagram of the structure of a dredging strategy optimization system based on reinforcement learning according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect a dredging environment image of silt through a depth camera and a sonar sensor, perform multi-scale decomposition based on orthogonal wavelet and curvelet transform to filter out noise on the dredging environment image to obtain a denoised image, use an adaptive histogram equalization algorithm to calculate a grayscale distribution probability density function on the denoised image, perform piecewise linear grayscale stretching and gamma correction according to a peak interval of the grayscale distribution probability density function to obtain a brightness equalized image, use a bilateral filtering algorithm and a fuzzy enhancement algorithm to perform contrast optimization and edge sharpening on the brightness equalized image, iteratively optimize the enhancement effect through a peak signal-to-noise ratio index and a structural similarity index to obtain an optimized dredging environment image; The second unit is used to construct a three-dimensional voxel grid model based on the optimized dredging environment image, establish a state vector including water velocity field, turbulence intensity and suspended matter concentration in each voxel grid unit of the three-dimensional voxel grid model, calculate the joint probability distribution of environmental interference factors in the voxel grid unit using a Bayesian probability model, construct a mathematical model of environmental interference sources according to the joint probability distribution, model and solve the silt distribution state using the Navier-Stokes equations, and obtain the dynamic evolution law of silt distribution; The third unit is used to input the optimized dredging environment image into a deep convolutional neural network, extract the multi-scale texture features, directional gradient features and regional semantic features of the optimized dredging environment image, use the attention weight mechanism to fuse the extracted features to obtain a feature vector, input the feature vector into a classification model to determine the silt category, and construct a three-dimensional distribution map of silt thickness in combination with the mathematical model of the environmental interference source and the dynamic evolution law of silt distribution. The silt category and the three-dimensional distribution map of silt thickness are input into the dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the dual temporal difference algorithm, generates a motion trajectory planning scheme and an operation parameter configuration scheme for the dredging robot, and performs online optimization on the value function of the dredging strategy optimization network according to the actual dredging effect.

[0084] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0085] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0086] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dredging strategy optimization method based on reinforcement learning, characterized in that: include: The desilting environment image of the silt is collected by a depth camera and a sonar sensor, and the desilting environment image is subjected to multi-scale decomposition based on orthogonal wavelet and curvelet transform to filter out noise to obtain a de-noised image, and the grayscale distribution probability density function is calculated by an adaptive histogram equalization algorithm for the de-noised image, and a piecewise linear grayscale stretching and gamma correction are performed according to the peak interval of the grayscale distribution probability density function to obtain a brightness equalized image, and the brightness equalized image is subjected to contrast optimization and edge sharpening processing by using a bilateral filtering algorithm and a fuzzy enhancement algorithm, and the enhancement effect is iteratively optimized by using a peak signal-to-noise ratio index and a structural similarity index to obtain an optimized desilting environment image; A three-dimensional voxel grid model is constructed based on the optimized desilting environment image, a state vector including a water velocity field, a turbulence intensity and a suspended solids concentration is established in each voxel grid unit of the three-dimensional voxel grid model, a Bayesian probability model is used to calculate the joint probability distribution of environmental interference factors in the voxel grid unit, a mathematical model of environmental interference sources is constructed according to the joint probability distribution, and the Navier-Stokes equations are used to model and solve the silt distribution state to obtain the dynamic evolution law of the silt distribution; The optimized dredging environment image is input into a deep convolutional neural network, and the multi-scale texture features, directional gradient features and regional semantic features of the optimized dredging environment image are extracted. The extracted features are fused using an attention weight mechanism to obtain a feature vector, and the feature vector is input into a classification model to determine the silt category. A three-dimensional distribution map of silt thickness is constructed by combining the mathematical model of the environmental interference source and the dynamic evolution law of silt distribution. The silt category and the three-dimensional distribution map of silt thickness are input into a dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward values ​​by interacting with the simulation environment, updates the strategy network parameters based on a dual temporal difference algorithm, generates a motion trajectory planning scheme and an operation parameter configuration scheme for the dredging robot, and performs online optimization of the value function of the dredging strategy optimization network according to the actual dredging effect.

2. The method according to claim 1, characterized in that The brightness equalization image is subjected to contrast optimization and edge sharpening processing by using bilateral filtering algorithm and fuzzy enhancement algorithm, and the enhancement effect is iteratively optimized by using peak signal-to-noise ratio index and structural similarity index, and the optimized dredging environment image is obtained, including: Calculating pixel gradient values ​​and gradient directions for the brightness-equalized image, adaptively dividing the image area into blocks according to the gradient values, calculating the gradient variance and gradient direction consistency in each block, constructing a spatial weight function of an adaptive Gaussian kernel based on the gradient variance and the gradient direction consistency, performing bilateral filtering, and obtaining a first filtering result; Calculating a local contrast factor in an adaptive window for the first filtering result, wherein the size of the adaptive window is determined according to the gradient value, and performing secondary filtering by multiplying the local contrast factor with a grayscale weight function constructed based on the gradient variance to obtain a bilateral filtered image; Decomposing the bilateral filtered image at multiple scales, calculating local statistical features based on the gradient value at each scale layer, mapping the local statistical features into parameters of a fuzzy membership function, determining weight coefficients of each scale according to the gradient direction consistency, and weighted fusion of the multi-scale fuzzy membership values ​​to obtain fused fuzzy membership values; Calculating the edge degree based on the fused fuzzy membership value, constructing an adaptive enhancement coefficient in combination with the gradient value, multiplying the adaptive enhancement coefficient with the difference between the bilateral filtered image and the original brightness equalized image to obtain an edge enhancement amount, performing multi-scale decomposition on the edge enhancement amount and superimposing it on the original brightness equalized image to obtain an enhanced image; Calculating the peak signal-to-noise ratio, structural similarity, and edge preservation based on gradient features of the enhanced image relative to the original brightness-equalized image, constructing a multi-objective optimization function, and adaptively adjusting the weight coefficients of the optimization indicators according to the gradient variance distribution; The multi-objective optimization function is iteratively optimized using an adaptive learning rate based on a gradient value, bilateral filtering parameters and fuzzy enhancement parameters are updated, and a final optimized image is output when the relative change value of the optimization target after multiple consecutive iterations is less than a preset threshold.

3. The method according to claim 1, characterized in that A three-dimensional voxel grid model is constructed based on the optimized desilting environment image, a state vector including a water velocity field, turbulence intensity and suspended solids concentration is established in each voxel grid unit of the three-dimensional voxel grid model, and a joint probability distribution of environmental interference factors in the voxel grid unit is calculated using a Bayesian probability model, including: Calculating the multi-scale gradient features of the optimized dredging environment image, constructing a structure tensor to obtain local eigenvalues, and performing adaptive meshing and smoothing processing on the dredging environment image based on the local eigenvalues ​​to obtain an optimized three-dimensional voxel grid; In each grid unit of the optimized three-dimensional voxel grid, the velocity field is discretized using a second-order central difference format, a trilinear interpolation basis function is constructed to calculate the grid node velocity component, a pressure Poisson equation is introduced to establish a divergence constraint, a pressure-velocity coupling algorithm is used to solve the intermediate velocity field, and a water flow velocity field that satisfies the divergence constraint of zero is obtained through multiple grid iterations; Performing time-average decomposition on the water flow velocity field to extract velocity pulsation components, calculating turbulent kinetic energy based on the velocity pulsation components, and calculating strain rate tensor based on the water flow velocity field to obtain turbulent dissipation rate, and combining the turbulent kinetic energy with the turbulent dissipation rate to obtain a turbulence intensity characteristic value; Establishing a mapping function between the gray value of the optimized desilting environment image and the suspended matter concentration, using measured concentration data to calibrate the coefficient of the mapping function, and generating a suspended matter concentration field; The water flow velocity field, the turbulence intensity characteristic quantity and the suspended matter concentration field are combined into an environmental state vector, and the number of components of a Gaussian mixture model is determined based on the Bayesian information criterion, and the prior probability distribution of the environmental state vector is constructed using the Gaussian mixture model; Collect sensor observation data to construct an observation vector, establish an observation likelihood function based on the observation vector, use a covariance matrix to characterize the error correlation of the observation vector, use a Markov chain Monte Carlo method to sample the state vector based on the prior probability distribution and the observation likelihood function, and calculate the posterior probability distribution of the state vector in combination with an adaptive step size to control the sampling acceptance rate; The posterior probability distribution is converted into a joint probability distribution of environmental interference factors in voxel grid units using a kernel density estimation method, wherein a bandwidth parameter of the kernel density estimation is determined by a cross-validation method.

4. The method according to claim 3, characterized in that: The water velocity field, the turbulence intensity characteristic quantity and the suspended matter concentration field are combined into an environmental state vector, and the number of components of a Gaussian mixture model is determined based on the Bayesian information criterion. The prior probability distribution of the environmental state vector is constructed by using the Gaussian mixture model, including: Collecting velocity components of the water velocity field, turbulent kinetic energy and turbulent dissipation rate in the turbulence intensity characteristic quantity, and spatial distribution data of the suspended matter concentration field, forming the collected data into an environmental state vector sample set, and constructing a log-likelihood function of a Gaussian mixture model based on the environmental state vector sample set; Based on the dimension of the environmental state vector sample set, the total number of parameters corresponding to each Gaussian component is calculated, the total number of parameters includes the dimension of the mean vector, the number of independent elements of the covariance matrix and the degree of freedom of the mixing weight, the log-likelihood function is combined with the logarithmic term of the total number of parameters to construct the Bayesian information criterion, the number of parameters required for each additional Gaussian component is calculated, the number of parameters, the sample size and the log-likelihood function are substituted into the Bayesian information criterion, and the optimal number of components is obtained through grid search with different numbers of Gaussian components; Initialize the parameters of the Gaussian mixture model using the sample set of the environmental state vector and the optimal number of components, and calculate the posterior probability matrix of each Gaussian component corresponding to each sample based on the current parameters in the expected step; calculate the weighted statistics of the sample based on the posterior probability matrix, and update the mixing weights, mean vectors and covariance matrix corresponding to the optimal number of components through the weighted statistics; Calculating the ratio of the maximum eigenvalue to the minimum eigenvalue of the covariance matrix, and when the ratio exceeds a preset correction threshold, adaptively adjusting the regularization parameter and correcting the covariance matrix; The relative rate of change of the log-likelihood function and the parameter update amount between adjacent iterations are calculated, and the relative rate of change and the update amount are compared with a preset threshold to determine the iteration termination condition. When the iteration termination condition is met, the optimal number of components and their corresponding mixing weights, mean vector and modified covariance matrix are used to construct the prior probability distribution of the environment state vector.

5. The method according to claim 1, characterized in that The mathematical model of environmental interference sources is constructed according to the joint probability distribution, and the Navier-Stokes equation is used to model and solve the silt distribution state to obtain the dynamic evolution law of silt distribution, including: Establishing a sludge motion equation based on the Stokes drag of spherical particles, wherein the sludge motion equation includes a gravity term, a buoyancy term, and a Brownian force term, establishing a Navier-Stokes equation group containing environmental interference sources based on the sludge motion equation, introducing the joint probability distribution of the environmental interference sources into the Navier-Stokes equation group, and constructing a momentum equation including a momentum source term and a mass equation including a mass source term; A sludge concentration transport equation is established according to the Navier-Stokes equations, wherein the sludge concentration transport equation includes a convection diffusion term and a concentration source term, wherein the convection diffusion term is obtained by solving the momentum equation, and the concentration source term is determined by the joint probability distribution of the environmental interference sources; Constructing a staggered grid system, spatially discretizing the momentum equation in the Navier-Stokes equations to obtain a momentum discrete equation, spatially discretizing the mass equation to obtain a mass discrete equation, and temporally discretizing the momentum discrete equation and the mass discrete equation to obtain a space-time discrete equation group; The turbulence kinetic energy and the turbulence dissipation rate are obtained according to the turbulence intensity, the momentum prediction equation in the space-time discrete equation group is introduced through the eddy viscosity coefficient, the intermediate velocity field is obtained by solving the momentum prediction equation, and the pressure Poisson equation is constructed based on the intermediate velocity field; The pressure Poisson equation is solved to obtain a pressure field, the intermediate velocity field is corrected using the pressure field to obtain a corrected velocity field, the corrected velocity field is substituted into the sludge concentration transport equation, the sludge concentration transport equation is solved in an explicit format to obtain the sludge concentration distribution at the current moment, the system boundary conditions for the next time step are updated according to the sludge concentration distribution at the current moment, and the steps from solving the momentum prediction equation to calculating the sludge concentration distribution are repeatedly executed based on the system boundary conditions until the simulation calculation of the specified time is completed to obtain the dynamic evolution law of the sludge distribution.

6. The method according to claim 1, characterized in that Combining the mathematical model of environmental interference sources and the dynamic evolution law of silt distribution to construct a three-dimensional distribution map of silt thickness includes: A sludge deposition rate model is established based on the mathematical model of the environmental interference source and the dynamic evolution law of the sludge distribution, and the joint probability distribution of the environmental interference factors in the voxel grid unit obtained in advance is combined with the sludge concentration gradient to calculate the sludge deposition flux of each voxel grid unit, wherein when the sludge concentration is greater than a first preset threshold and the joint probability distribution of the environmental interference factors is less than a second preset threshold, the sludge deposition flux is determined to be a positive value, and when the joint probability distribution of the environmental interference factors is greater than the second preset threshold, the sludge deposition flux is determined to be a negative value; Integrating the silt deposition flux in the time domain to obtain the theoretical cumulative thickness of the silt in each voxel grid unit, and correcting the theoretical cumulative thickness of the silt according to the compaction correction coefficient to obtain the actual cumulative thickness of the silt in each voxel grid unit; Constructing a silt thickness spatial variation model, calculating the spatial correlation coefficient of silt thickness at different positions based on the silt thickness spatial variation model, and using a spatial interpolation algorithm to interpolate the actual cumulative thickness of the silt from each voxel grid unit to a preset spatial grid node to obtain the silt interpolation thickness of the preset spatial grid node; A spatial display grid is established in a three-dimensional coordinate system, the silt interpolation thickness is mapped onto the spatial display grid to obtain silt thickness display data, a plurality of silt thickness isosurfaces and a plurality of silt thickness profile layers are generated according to the silt thickness display data, and the plurality of silt thickness isosurfaces and the plurality of silt thickness profile layers are combined to form a three-dimensional distribution map of silt thickness.

7. The method according to claim 1, characterized in that The silt category and the three-dimensional distribution map of silt thickness are input into the dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the double temporal difference algorithm, and generates the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot, including: Encoding the silt category information into a category feature vector, performing voxel processing on the three-dimensional distribution map of silt thickness to obtain a spatial feature tensor, inputting the category feature vector and the spatial feature tensor into a feature fusion network to generate a representation of the dredging environment state; Constructing a dredging simulation environment including a sludge resuspension model, a dredging robot kinematic model and an energy consumption model, and obtaining the dredging robot position information, sludge removal volume information and energy consumption information based on the dredging environment state representation and the dredging simulation environment; Establishing a comprehensive reward function according to the sludge removal volume information, the energy consumption information, and the distance information between the dredging robot and the obstacle, wherein the comprehensive reward function includes a dredging efficiency item positively correlated with the sludge removal volume information, an energy consumption item negatively correlated with the energy consumption information, and a safety item related to the distance information; Constructing a strategy network, wherein the strategy network includes an action generation network and a value evaluation network, wherein the action generation network outputs a motion trajectory sequence and an operation parameter sequence of the dredging robot based on the dredging environment state representation, and the value evaluation network predicts long-term benefits based on the dredging environment state representation, the motion trajectory sequence and the operation parameter sequence; The strategy network interacts with the dredging simulation environment to obtain state transition information and reward value information, calculates the value target using a temporal difference algorithm based on the state transition information and the reward value information, and updates the network parameters of the strategy network using a policy gradient method; Construct an experience replay buffer pool, store state transition information and reward value information in the experience replay buffer pool, and obtain training samples from the experience replay buffer pool using random sampling; Build a target network, migrate the network parameters of the policy network to the target network in a soft update manner, provide a learning target based on the target network, and calculate the loss function value of the policy network. When the change in the loss function value over multiple consecutive training cycles is less than the preset change threshold, the policy network training converges. Using the converged policy network, the motion trajectory planning scheme and operation parameter configuration scheme of the dredging robot are generated according to the input dredging environment state representation.

8. A dredging strategy optimization system based on reinforcement learning, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect a dredging environment image of silt through a depth camera and a sonar sensor, perform multi-scale decomposition based on orthogonal wavelet and curvelet transform to filter out noise on the dredging environment image to obtain a denoised image, use an adaptive histogram equalization algorithm to calculate a grayscale distribution probability density function on the denoised image, perform piecewise linear grayscale stretching and gamma correction according to a peak interval of the grayscale distribution probability density function to obtain a brightness equalized image, use a bilateral filtering algorithm and a fuzzy enhancement algorithm to perform contrast optimization and edge sharpening on the brightness equalized image, iteratively optimize the enhancement effect through a peak signal-to-noise ratio index and a structural similarity index to obtain an optimized dredging environment image; The second unit is used to construct a three-dimensional voxel grid model based on the optimized dredging environment image, establish a state vector including water velocity field, turbulence intensity and suspended matter concentration in each voxel grid unit of the three-dimensional voxel grid model, calculate the joint probability distribution of environmental interference factors in the voxel grid unit using a Bayesian probability model, construct a mathematical model of environmental interference sources according to the joint probability distribution, model and solve the silt distribution state using the Navier-Stokes equations, and obtain the dynamic evolution law of silt distribution; The third unit is used to input the optimized dredging environment image into a deep convolutional neural network, extract the multi-scale texture features, directional gradient features and regional semantic features of the optimized dredging environment image, use the attention weight mechanism to fuse the extracted features to obtain a feature vector, input the feature vector into a classification model to determine the silt category, and construct a three-dimensional distribution map of silt thickness in combination with the mathematical model of the environmental interference source and the dynamic evolution law of silt distribution. The silt category and the three-dimensional distribution map of silt thickness are input into the dredging strategy optimization network. The dredging strategy optimization network obtains dredging operation status information and reward value by interacting with the simulation environment, updates the strategy network parameters based on the dual temporal difference algorithm, generates a motion trajectory planning scheme and an operation parameter configuration scheme for the dredging robot, and performs online optimization on the value function of the dredging strategy optimization network according to the actual dredging effect.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images