Offshore moon pool resonance suppression method based on multi-objective collaborative optimization

Abstract

The invention discloses an offshore moon pool resonance suppression method based on multi-objective collaborative optimization. The method comprises the following steps: obtaining an initial flow field model of a moon pool; determining a propagation time delay distribution matrix; extracting local eddy current intensity and direction characteristics caused by contour interference from the propagation time delay distribution matrix, and classifying contributions of different mechanism contours to flow field deformation through an improved random forest algorithm to obtain an interference characteristic vector; constructing a time delay interference joint tensor for the interference characteristic vector and the propagation time delay distribution matrix, and separating the time delay interference joint tensor to determine a dynamic mapping relation; if the time sequence deviation in the dynamic mapping relation exceeds a preset threshold value, adjusting the control instruction sequence through back propagation to compensate time delay mismatch so as to obtain a corrected instruction set; and obtaining an optimized execution scheme, and completing the resonance suppression of the offshore moon pool. According to the invention, the active identification of the moon pool resonance risk and the inhibition capability of key disturbance are improved.

Description

Technical Field

[0001] This invention relates to the field of marine lunar pool technology, and in particular to a method for suppressing resonance in marine lunar pools based on multi-objective collaborative optimization. Background Technology

[0002] With the continuous development of deep-sea operations and marine engineering technology, moon pool structures are increasingly widely used in various types of offshore platforms and operational equipment. The moon pool area provides access channels and operating windows for underwater robotic arms, operational thrusters, and measurement and control equipment, becoming an important support for high-precision operations in the deep-sea environment.

[0003] Existing methods for controlling flow field disturbances in moon pools mostly rely on fluid dynamics modeling and passive control of a single actuator. They primarily focus on simplified prediction and local suppression of flow field disturbances in a single actuator, making it difficult to address the complex disturbances and resonance problems caused by flow field coupling effects when multiple actuators operate in tandem.

[0004] Within the confined space of a lunar pool, traditional disturbance modeling methods struggle to accurately characterize the spatiotemporal interference and flow field response caused by the simultaneous operation of multiple actuators. Conventional flow field measurement and signal processing techniques are limited by their data fusion capabilities in high-noise environments, resulting in insufficient extraction of flow field disturbance features and significant errors in predicting flow field evolution paths. Furthermore, existing multi-actuator scheduling strategies often employ fixed or empirical parameters, lacking collaborative optimization based on real-time feedback. This makes it impossible to dynamically compensate for timing mismatches caused by fluid delays and structural differences between actuators, and also hinders the proactive avoidance and precise control of flow field effects within high-risk operating windows.

[0005] For the identification and suppression of resonance risk in the lunar pool region, existing algorithm models mostly ignore the dynamic coupling mechanism of disturbance propagation and local eddy current interference. Traditional machine learning methods have limitations in feature weighting, risk sensitivity and decision interpretability, and cannot achieve active sample weighting and multi-objective node optimization for resonance amplification risk. Summary of the Invention

[0006] One objective of this invention is to propose a method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization. This invention improves the ability to actively identify lunar pool resonance risks and suppress key disturbances.

[0007] The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to embodiments of the present invention includes:

[0008] The flow field disturbance data generated by the actuator and the water movement parameters in the key area of ​​the moon pool are collected by the sensor array. Real-time signal processing is used to filter out noise and obtain the initial flow field model of the moon pool.

[0009] The water density, temperature and salinity distribution along the propagation path from each actuator to the center of the moon pool are calculated based on the initial flow field model of the moon pool. Kalman filtering is used to fuse multi-source data to determine the propagation delay distribution matrix.

[0010] The intensity and direction features of local eddy currents caused by profile interference are extracted from the propagation delay distribution matrix. The contribution of different mechanism profiles to flow field deformation is classified by an improved random forest algorithm to obtain the interference characteristic vector.

[0011] A joint time-delay interference tensor is constructed for the interference characteristic vector and the propagation time delay distribution matrix, and the dynamic mapping relationship is determined by separating the joint time-delay interference tensor;

[0012] If the timing deviation in the dynamic mapping relationship exceeds the preset threshold, the control command sequence is adjusted through back propagation to compensate for the time delay mismatch and obtain the correction command set.

[0013] After obtaining the correction instruction set, the collaborative action logic is integrated, and the contour compensation factor is injected into the parallel scheduling of multiple actuators to determine the convergence time of the flow field effect, thereby obtaining an optimized execution scheme and completing the suppression of resonance in the lunar pool at sea.

[0014] Optionally, the step of using real-time signal processing to filter out noise and obtain the initial flow field model of the moon pool includes:

[0015] The instantaneous velocity and pressure signals of the actuator are synchronously acquired by a multi-channel sensor array to obtain raw multi-dimensional time-series data.

[0016] A finite impulse response filter is used to perform real-time bandpass filtering and notch filtering on the original multidimensional time series data to obtain denoised multidimensional time series data.

[0017] The instantaneous velocity vector and pressure gradient at each sensor location are calculated based on the denoised multidimensional time-series data to obtain the current flow field vector field.

[0018] The corrected real-time flow field distribution is obtained by fusing the current flow field vector field with the predicted flow field state from the previous moment using a Kalman filter.

[0019] The average vorticity and velocity fluctuation intensity in the key area of ​​the moon pool are extracted based on the corrected real-time flow field distribution to obtain the hydrodynamic parameters of the moon pool area.

[0020] If the hydrodynamic parameters in the moon pool area exceed the preset threshold range, the actuator control signal adjustment command is triggered to obtain the control quantity to suppress the disturbance;

[0021] The actuator driving parameters are directly updated based on the control quantity that suppresses disturbances, and the corrected real-time flow field distribution is used as the prior state of the Kalman filter in the next cycle to obtain the initial flow field model of the moon pool for continuous tracking.

[0022] Optionally, the step of using Kalman filtering to fuse multi-source data and determine the propagation delay distribution matrix includes:

[0023] By using the initial flow field model of the moon pool to obtain the original distribution data of water density, temperature and salinity along the propagation path from each actuator to the center of the moon pool, preliminary environmental parameter distribution results are obtained.

[0024] The initial environmental parameter distribution results are calibrated for consistency to determine the calibrated multi-source dataset.

[0025] From the calibrated multi-source dataset, feature values ​​related to water density, temperature and salinity are extracted. An environmental variable matrix along the propagation path is constructed through feature value analysis to obtain a comprehensive characterization of the environmental variables.

[0026] The Kalman filter method is applied to fuse the comprehensive representation of environmental variables, calculate the estimated time delay along the propagation path, and obtain the preliminary distribution of propagation time delay.

[0027] Based on the preliminary distribution of propagation delay, and combined with the positional relationship between the center of the moon pool and the actuator, the estimated delay value is spatially corrected to determine the deviation of the delay distribution. If the deviation exceeds the preset threshold, the delay value is iteratively adjusted to obtain the corrected delay distribution.

[0028] By constructing the propagation delay distribution matrix using the corrected delay distribution, and then smoothing outliers in the matrix, the final propagation delay distribution matrix is ​​obtained.

[0029] Optionally, the step of classifying the contribution of different mechanism profiles to flow field deformation by improving the random forest algorithm includes:

[0030] Each element in the propagation delay distribution matrix represents the perturbation propagation delay of a discrete path unit on the propagation path of an actuator to the center of the moon pool.

[0031] For each discrete path unit of each actuator, calculate the local propagation velocity increment;

[0032] The local eddy intensity index is defined based on the local propagation velocity increment and the path unit length;

[0033] From the initial flow field model of the moon pool, the three-dimensional fluid velocity vector corresponding to each propagation path unit is extracted, and the local vortex principal direction angle of the discrete path unit is calculated by using the ratio of the two orthogonal components of the three-dimensional fluid velocity vector in the horizontal plane.

[0034] The local eddy current intensity index, the local eddy current principal direction angle, the disturbance propagation delay, and the radially normalized distance are combined into a four-dimensional column vector to obtain the local interference feature vector.

[0035] The set of lunar pool resonance characteristic periods was obtained through experiments and numerical simulations, and the resonance sensitivity weight of the samples was calculated based on the perturbation propagation delay.

[0036] Based on the resonance sensitivity weight, a reference value for eddy current intensity and an eddy current amplification coefficient are introduced to construct a local comprehensive sample weight;

[0037] The local resonance risk index is obtained by multiplying the local comprehensive sample weight by the ratio of the local eddy current intensity reference value to the eddy current intensity reference value.

[0038] An improved random forest model is constructed, which uses the local interference feature vector, local resonance risk index and local comprehensive sample weight as the training input for each sample. The weighted Gini impurity and weighted resonance risk of the sample are combined into a partitioning objective function. By optimizing the partitioning objective function, the model takes into account both the deformation mode of the classified flow field and the priority separation of samples with high resonance risk, so as to realize node partitioning and tree structure generation based on multi-objective collaborative optimization.

[0039] For each decision tree, output the weighted probability contribution vector of each flow field deformation mode to which each sample belongs;

[0040] The interferometric characteristic vector is obtained by integrating the outputs of all decision trees based on the weighted probability contribution vector.

[0041] Optionally, constructing the time-delay interferometric joint tensor and determining the dynamic mapping relationship includes:

[0042] By mapping all actuator outline numbers, disturbance propagation path numbers, and interference feature dimensions one-to-one, a three-dimensional joint array is obtained;

[0043] Perform low-rank decomposition on a three-dimensional joint array, and the decomposition result includes three factor arrays;

[0044] Based on three factor arrays, a two-dimensional array of dynamic spatiotemporal mapping relationships is defined.

[0045] Normalization is performed on all elements in the two-dimensional array of dynamic spatiotemporal mapping relationships, and based on the normalization results, the average dynamic mapping difference between each pair of actuator contours under all disturbance propagation paths is calculated.

[0046] If the average dynamic mapping difference between any pair of actuator contours exceeds the preset contour coupling deviation threshold, the current dynamic mapping relationship is marked as a high deviation state, triggering the time delay compensation process, and the contour-path pairs in the dynamic mapping relationship that exceed the threshold are recorded as mapping indices to be corrected.

[0047] Optionally, the step of obtaining a correction instruction set by adjusting the control instruction sequence through backpropagation to compensate for time delay mismatch includes:

[0048] By analyzing the specific manifestations of time-series deviations through dynamic mapping relationships, the distribution of deviations can be obtained;

[0049] Based on the deviation distribution, abnormal time series points are identified for the portion exceeding the preset threshold.

[0050] If an abnormal timing point is detected, the specific range of time delay mismatch is determined by comparing the preset threshold with the actual deviation value.

[0051] For specific intervals of time delay mismatch, the back propagation method is used to adjust the parameters of the control command sequence and generate preliminary correction commands.

[0052] Key control points are extracted from the initial correction instructions, and their matching degree with the instruction sequence is analyzed to obtain the adjusted instruction set.

[0053] If the adjusted instruction set still has local time delay mismatches, the final corrected instruction set is determined by fine-tuning the local points again.

[0054] Optionally, the step of injecting a contour compensation factor to determine the convergence time of the flow field effect in the parallel scheduling of multiple actuators includes:

[0055] Based on the correction instruction set, and combined with the structural characteristics and control timing of multiple actuators, an initial motion coordination matrix is ​​constructed;

[0056] Based on the initial action coordination matrix and the two-dimensional array of dynamic spatiotemporal mapping relationship, the total disturbance superposition intensity at each moment is calculated to determine the convergence moment of the disturbance effect;

[0057] Based on the convergence moment candidate and target resonance avoidance window of the disturbance effect, a target offset evaluation function is constructed;

[0058] In the construction of the optimized execution scheme, a contour compensation factor is introduced, and the target offset evaluation function is used as the optimization objective. The control start time of each actuator is fine-tuned to minimize the target offset.

[0059] The final optimized execution plan is output, which requires that the target offset evaluation function not exceed the preset upper limit threshold, and that the actuator number, control start time and contour compensation factor correspond one-to-one.

[0060] Optionally, the final optimized execution scheme includes:

[0061] Actuator A: Control start time adjusted to T1, contour compensation factor set to C1;

[0062] Actuator B: Control start time adjusted to T2, contour compensation factor set to C2;

[0063] Actuator C: Control start time adjusted to T3, contour compensation factor set to C3;

[0064] Where T1, T2, and T3 are the control times that finally satisfy the convergence condition of the target offset evaluation function, and C1, C2, and C3 are the corresponding contour compensation factors.

[0065] The beneficial effects of this invention are:

[0066] This invention integrates sensor array flow field disturbance data with hydrodynamic parameters of key areas in the lunar pool in a high dimension. Through multi-source data Kalman filtering and path partitioning weighting, it achieves multi-dimensional modeling of environmental variables along the disturbance propagation path from the actuators to the central region of the lunar pool, constructs a dynamically updatable propagation delay distribution matrix, couples the interference characteristic vector and the delay distribution into a third-order joint tensor through low-rank tensor decomposition, and establishes a dynamic mapping relationship of multiple actuators, multiple paths, and multiple feature dimensions by extracting spatiotemporal mapping factors through decomposition. This enables precise quantification of the multi-path response of actuator disturbances to the lunar pool flow field, achieving global and dynamic control of complex flow field coupling effects at the system level.

[0067] This invention addresses the coupling risk between lunar pool flow field disturbances and dominant resonance modes by proposing an improved random forest algorithm based on resonance period-sensitive weighting and eddy current intensity weighting. The algorithm incorporates disturbance propagation delay, local eddy current characteristics, and radial distance into feature construction, and designs an exponentially decaying weighting function to measure the resonance amplification risk of disturbance samples. Node partitioning comprehensively considers weighted Gini impurity and weighted resonance risk, achieving multi-objective collaborative optimization. This not only improves the discriminative power of flow field deformation modes but also prioritizes the separation of high-risk samples in the decision tree structure, enhancing the proactive identification of lunar pool resonance risks and the suppression of key disturbances. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0069] Figure 1 This is a flowchart of a method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figure 1 As shown in Example 1: A method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization, comprising:

[0072] The flow field disturbance data generated by the actuator and the water movement parameters in the key area of ​​the moon pool are collected by the sensor array. Real-time signal processing is used to filter out noise and obtain the initial flow field model of the moon pool.

[0073] The water density, temperature and salinity distribution along the propagation path from each actuator to the center of the moon pool are calculated based on the initial flow field model of the moon pool. Kalman filtering is used to fuse multi-source data to determine the propagation delay distribution matrix.

[0074] The intensity and direction features of local eddy currents caused by profile interference are extracted from the propagation delay distribution matrix. The contribution of different mechanism profiles to flow field deformation is classified by an improved random forest algorithm to obtain the interference characteristic vector.

[0075] A joint time-delay interference tensor is constructed for the interference characteristic vector and the propagation time delay distribution matrix, and the dynamic mapping relationship is determined by separating the joint time-delay interference tensor;

[0076] If the timing deviation in the dynamic mapping relationship exceeds the preset threshold, the control command sequence is adjusted through back propagation to compensate for the time delay mismatch and obtain the correction command set.

[0077] After obtaining the correction instruction set, the collaborative action logic is integrated, and the contour compensation factor is injected into the parallel scheduling of multiple actuators to determine the convergence time of the flow field effect, thereby obtaining an optimized execution scheme and completing the suppression of resonance in the lunar pool at sea.

[0078] In this embodiment, real-time signal processing is used to filter out noise to obtain the initial flow field model of the moon pool, including:

[0079] The instantaneous velocity and pressure signals of the actuator are synchronously acquired by a multi-channel sensor array to obtain raw multi-dimensional time-series data.

[0080] A finite impulse response filter is used to perform real-time bandpass filtering and notch filtering on the original multidimensional time series data to obtain denoised multidimensional time series data.

[0081] The instantaneous velocity vector and pressure gradient at each sensor location are calculated based on the denoised multidimensional time-series data to obtain the current flow field vector field.

[0082] The corrected real-time flow field distribution is obtained by fusing the current flow field vector field with the predicted flow field state from the previous moment using a Kalman filter.

[0083] The average vorticity and velocity fluctuation intensity in the key area of ​​the moon pool are extracted based on the corrected real-time flow field distribution to obtain the hydrodynamic parameters of the moon pool area.

[0084] If the hydrodynamic parameters in the moon pool area exceed the preset threshold range, the actuator control signal adjustment command is triggered to obtain the control quantity to suppress the disturbance;

[0085] The actuator driving parameters are directly updated based on the control quantity that suppresses disturbances, and the corrected real-time flow field distribution is used as the prior state of the Kalman filter in the next cycle to obtain the initial flow field model of the moon pool for continuous tracking.

[0086] In this embodiment, Kalman filtering is used to fuse multi-source data to determine the propagation delay distribution matrix, including:

[0087] By using the initial flow field model of the moon pool to obtain the original distribution data of water density, temperature and salinity along the propagation path from each actuator to the center of the moon pool, preliminary environmental parameter distribution results are obtained.

[0088] The initial environmental parameter distribution results are calibrated for consistency to determine the calibrated multi-source dataset.

[0089] From the calibrated multi-source dataset, feature values ​​related to water density, temperature and salinity are extracted. An environmental variable matrix along the propagation path is constructed through feature value analysis to obtain a comprehensive characterization of the environmental variables.

[0090] In Example 1, statistical features of water density, temperature and salinity at each propagation path node are extracted from the calibrated multi-source dataset. These features include mean, standard deviation, maximum value, minimum value and rate of change. The five statistical features of each propagation path node are combined into a 15-dimensional feature vector to construct the environmental variable feature matrix of all nodes on the propagation path.

[0091] The Kalman filter method is applied to fuse the comprehensive representation of environmental variables, calculate the estimated time delay along the propagation path, and obtain the preliminary distribution of propagation time delay.

[0092] Based on the preliminary distribution of propagation delay, and combined with the positional relationship between the center of the moon pool and the actuator, the estimated delay value is spatially corrected to determine the deviation of the delay distribution. If the deviation exceeds the preset threshold, the delay value is iteratively adjusted to obtain the corrected delay distribution.

[0093] By constructing the propagation delay distribution matrix using the corrected delay distribution, and then smoothing outliers in the matrix, the final propagation delay distribution matrix is ​​obtained.

[0094] In Example 1, the corrected propagation path delay distribution data is reconstructed in two dimensions using a rasterization based on the spatial mapping relationship between the actuator and the center of the moon pool to generate an initial propagation delay distribution matrix. Anomalies in the initial propagation delay distribution matrix are identified based on the statistical threshold method, and local smoothing is performed on the anomalies to obtain the final propagation delay distribution matrix.

[0095] In this embodiment, the contribution of different mechanism profiles to flow field deformation is classified by improving the random forest algorithm, including:

[0096] Each element in the propagation delay distribution matrix represents the perturbation propagation delay of a discrete path unit on the propagation path of an actuator to the center of the moon pool.

[0097] For each discrete path unit of each actuator, calculate the local propagation velocity increment;

[0098] In Example 1, the physical length of the discrete path unit is divided by the corresponding disturbance propagation delay to obtain the local equivalent propagation velocity of the discrete path unit. The difference between the local equivalent propagation velocities of adjacent path units is used to obtain the local propagation velocity increment. The local propagation velocity increment represents the local acceleration or deceleration phenomenon when the disturbance propagates in the moon pool channel.

[0099] The local eddy intensity index is defined based on the local propagation velocity increment and the path unit length;

[0100] In Example 1, the absolute value of the local propagation velocity increment is divided by the path element length and multiplied by the eddy current intensity scaling factor to obtain the local eddy current intensity index of the discrete path element. The local eddy current intensity index is used to reflect the intensity change of the local rotating flow induced by contour interference in the propagation channel.

[0101] From the initial flow field model of the moon pool, the three-dimensional fluid velocity vector corresponding to each propagation path unit is extracted, and the local vortex principal direction angle of the discrete path unit is calculated by using the ratio of the two orthogonal components of the three-dimensional fluid velocity vector in the horizontal plane.

[0102] The three-dimensional fluid velocity vector includes velocity components along three orthogonal directions in the lunar pool coordinate system. The principal direction angle of the local vortex is used to characterize the principal rotation direction of the vortex in the horizontal profile below the free surface of the lunar pool.

[0103] The local eddy current intensity index, the local eddy current principal direction angle, the disturbance propagation delay, and the radially normalized distance are combined into a four-dimensional column vector to obtain the local interference feature vector.

[0104] In Example 1, for each path unit, the radial distance from the center of the discrete path unit to the geometric center of the moon pool is obtained, and normalized by the maximum radial distance from the moon pool section to the center to obtain the radially normalized distance. The local eddy current intensity index, the local eddy current principal direction angle, the disturbance propagation delay and the radially normalized distance are combined into a local interference feature vector. The local interference feature vector represents the influence of the actuator contour interference on the local morphology of the ocean moon pool flow field.

[0105] The set of lunar pool resonance characteristic periods was obtained through experiments and numerical simulations, and the resonance sensitivity weight of the samples was calculated based on the perturbation propagation delay.

[0106] In Example 1, based on the geometry, boundary conditions, and fluid properties of the lunar pool at sea, a three-dimensional numerical simulation method was used to perform modal analysis of the fluid oscillations in the lunar pool under different working conditions. By solving the eigenvalue problem of fluid-structure interaction, the eigencycles of several dominant resonance modes were obtained. At the same time, combined with physical experimental methods, excitation was applied on the lunar pool model experimental platform, and the response signals of the free surface and internal flow field of the lunar pool were collected. Time-frequency analysis and spectral analysis methods were used to verify and supplement the dominant resonance cycles obtained by numerical simulation. All dominant resonance cycles confirmed by simulation and experiment were organized into a set of lunar pool resonance characteristic cycles. Based on the set of lunar pool resonance characteristic cycles, the absolute difference between each disturbance propagation delay sample and all dominant resonance cycles was calculated. An exponential decay function was applied to the difference, and the maximum value was taken as the resonance sensitivity weight of the disturbance propagation delay sample.

[0107] ;

[0108] in, This represents the characteristic period of the m-th dominant resonance mode in the marine lunar pool. In the basic random forest, the first... The decision tree is for the first The resonance-sensitive weights initially assigned to each contour region under conventional features are used to characterize the discriminative contribution of flow field deformation under the standard algorithm. This is the resonance time deviation attenuation coefficient, used to control the degree of weight amplification when the disturbance propagation delay approaches the characteristic period of lunar pool resonance. This is the propagation delay of the disturbance.

[0109] The physical meaning of resonance sensitivity weight is that when the propagation delay of a disturbance approaches the characteristic period of any dominant resonant mode in the lunar pool, its impact on the risk of resonance amplification will be amplified exponentially. The higher the resonance sensitivity weight, the stronger the sensitivity of the disturbance in the resonance risk zone.

[0110] Based on the resonance sensitivity weight, a reference value for eddy current intensity and an eddy current amplification coefficient are introduced to construct a local comprehensive sample weight;

[0111] In Example 1, a reference eddy current intensity index under the lunar pool condition at sea is selected from experimental statistics or simulation data as the reference value of eddy current intensity. The ratio of the eddy current intensity of each local path unit to the reference eddy current intensity is multiplied by the eddy current amplification factor. The result reflects the amplification or weakening effect of local disturbance relative to intensity. The local comprehensive sample weight indicates that the closer the disturbance propagation delay is to the dominant resonance period of the lunar pool, the greater the local eddy current intensity, and the closer the location is to the center of the lunar pool, the higher the physical influence of the disturbance sample on the lunar pool resonance risk. All local path units obtain their corresponding local comprehensive sample weights in this way.

[0112] ;

[0113] in, This indicates that in the improved random forest algorithm, the th... The decision tree is for the first... Local comprehensive sample weights assigned to each contour region feature branch. This is the vorticity gain coefficient. This indicates that along the propagation path of the flow field disturbance, the first... The decision tree sampling point corresponds to the first... The reference value for local eddy current intensity in each contour region indicates that the corresponding region is more strongly affected by flow field disturbances and contributes more significantly to the risk of resonance. This serves as a reference value for eddy current intensity, used to standardize eddy current data of different regions and intensities. It is selected as the historical maximum safe eddy current for key areas of the lunar pool, ensuring that the characteristics of each region are weighted and adjusted under the same discrimination scale. Indicates the first The disturbance risk coefficient of each contour region This represents the maximum disturbance risk coefficient across all contour regions.

[0114] The local resonance risk index is obtained by multiplying the local comprehensive sample weight by the ratio of the local eddy current intensity reference value to the eddy current intensity reference value.

[0115] The local resonance risk index reflects the potential impact of this flow field disturbance during the resonance amplification process in the ocean moon pool.

[0116] An improved random forest model is constructed, which uses the local interference feature vector, local resonance risk index and local comprehensive sample weight as the training input for each sample. The weighted Gini impurity and weighted resonance risk of the sample are combined into a partitioning objective function. By optimizing the partitioning objective function, the model takes into account both the deformation mode of the classified flow field and the priority separation of samples with high resonance risk, so as to realize node partitioning and tree structure generation based on multi-objective collaborative optimization.

[0117] In Example 1, when constructing the improved random forest model, for each training sample corresponding to a local interference feature vector, the local resonance risk index associated with the sample and the local comprehensive sample weight are used as sample attributes in the node partitioning calculation. During the growth of each decision tree, when evaluating a candidate partitioning node, the class distribution of samples in each child node before and after partitioning is statistically analyzed, and the samples of each class are weighted according to the local comprehensive sample weight to obtain the weighted class proportion result. Based on the weighted class proportion result, the weighted Gini impurity corresponding to the current node and its child nodes is calculated to measure the degree of influence of the partitioning on the classification purity of the flow field deformation mode. At the same time, the local resonance risk index of all samples contained in the current node is weighted and averaged within the same node range to obtain the weighted Gini impurity corresponding to the node. The resonance risk assessment value represents the potential risk level of the node sample as a whole to the amplification of lunar pool resonance at sea. The weighted Gini impurity and weighted resonance risk assessment value are weighted and combined according to the preset impurity-risk synergistic weight ratio to form a comprehensive evaluation index for node division judgment, and the comprehensive evaluation index is used as the evaluation result of the current candidate division scheme. By comparing the comprehensive evaluation index of all candidate division schemes under the same node, the division scheme with the best comprehensive evaluation index is selected as the final division method of the corresponding node. Thus, while ensuring the classification ability of flow field deformation mode, high resonance risk samples are preferentially classified into independent branches in the tree structure, realizing node division and decision tree structure generation based on multi-objective collaborative optimization. The improved random forest model can accurately distinguish the flow field deformation feature regions that are most critical to the suppression of lunar pool resonance.

[0118] For each decision tree, output the weighted probability contribution vector of each flow field deformation mode to which each sample belongs;

[0119] In Example 1, each sample is divided into a unique leaf node along the decision tree path. In the leaf node, the weight normalization ratio of all weighted training samples belonging to the node is counted in different flow field deformation mode categories to obtain the probability distribution vector of the sample in each flow field deformation mode category under the decision tree. For the same actuator contour, the probability distribution vectors of all its local path units are weighted and averaged according to their respective comprehensive sample weights to obtain the weighted probability contribution vector of the contour in the corresponding decision tree corresponding to each flow field deformation mode.

[0120] Each actuator profile is mapped to a multidimensional array under all decision trees according to the weighted probability contribution vector. The arrays are arranged according to the actuator number, path unit index, and flow field deformation mode category, and then summarized to form a three-dimensional component. The first dimension is the actuator number, the second dimension is the path unit number, and the third dimension is the probability dimension of the flow field deformation mode category. Each element corresponds to the weighted probability contribution value of a certain actuator profile under a certain dominant path unit and a certain flow field deformation mode.

[0121] The interferometric characteristic vector is obtained by integrating the outputs of all decision trees based on the weighted probability contribution vector.

[0122] ;

[0123] in, Let be the interference characteristic vector of the t-th decision tree on the profile of the k-th actuator. The weighted probability contribution vector. This is the set of local interference path unit indices corresponding to the k-th actuator contour.

[0124] In this embodiment, constructing the time-delay interferometry joint tensor and determining the dynamic mapping relationship includes:

[0125] By mapping all actuator outline numbers, disturbance propagation path numbers, and interference feature dimensions one-to-one, a three-dimensional joint array is obtained;

[0126] In Example 1, a set of actuator contour numbers is set, and the interference characteristic vector of each actuator contour is represented as a fixed-length vector. Each component of the interference characteristic vector represents the contribution of the actuator contour to the local flow field disturbance morphology under different interference feature dimensions.

[0127] The propagation delay distribution corresponding to each actuator profile is represented as a two-dimensional array. Each element of the array represents the propagation delay of the actuator profile from a discrete path unit on a specific disturbance propagation path to the center of the moon pool.

[0128] Each element of the three-dimensional joint array is obtained by weighted fusion of the specified component in the interference characteristic vector with the average propagation delay of the corresponding actuator contour on the corresponding disturbance propagation path. The three-dimensional joint array is used to integrate and describe the time delay interference coupling relationship under different actuator contours, disturbance propagation paths and interference characteristic dimensions.

[0129] Perform low-rank decomposition on a three-dimensional joint array, and the decomposition result includes three factor arrays;

[0130] The three factor arrays are: a contour factor array, representing the coupling response characteristics of each actuator contour; a path factor array, representing the spatial interference mode of each disturbance propagation path; and a feature factor array, representing the contribution ratio of each interference feature dimension in resonant coupling. Each element in the three-dimensional joint array can be approximated by the weighted product of these three factor arrays.

[0131] Based on three factor arrays, a two-dimensional array of dynamic spatiotemporal mapping relationships is defined.

[0132] In Example 1, for each actuator contour and each disturbance propagation path, the contour factor array of the actuator contour on all rank components and the path factor array of the corresponding disturbance propagation path on the same rank component are multiplied one by one, and the product results of all rank components are accumulated. The accumulated result is used as the dynamic mapping intensity of the actuator contour on the disturbance propagation path, forming a two-dimensional array of dynamic spatiotemporal mapping relationship with the actuator contour number as the row index and the disturbance propagation path number as the column index.

[0133] The dynamic mapping intensity is used to characterize the degree of comprehensive influence of the flow field interference characteristics generated by the actuator profile on the corresponding disturbance propagation path, which is transmitted to the central region of the moon pool and participates in the resonant coupling.

[0134] Normalization is performed on all elements in the two-dimensional array of dynamic spatiotemporal mapping relationships, and based on the normalization results, the average dynamic mapping difference between each pair of actuator contours under all disturbance propagation paths is calculated.

[0135] In Example 1, the dynamic mapping intensities corresponding to all actuator contours and all disturbance propagation paths in the two-dimensional array of dynamic spatiotemporal mapping relationships are used as statistical objects. The maximum and minimum dynamic mapping intensity values ​​are obtained. For each dynamic mapping intensity in the two-dimensional array of dynamic spatiotemporal mapping relationships, normalization is performed by subtracting the minimum dynamic mapping intensity and dividing by the difference between the maximum and minimum dynamic mapping intensities. All dynamic mapping intensities are uniformly mapped to the same numerical range, resulting in a normalized two-dimensional array of dynamic spatiotemporal mapping relationships. After obtaining the normalized two-dimensional array of dynamic spatiotemporal mapping relationships, the average dynamic mapping difference between each pair of actuator contours under all disturbance propagation paths is calculated. The average dynamic mapping difference is obtained as follows:

[0136] For any two different actuator profiles, the normalized dynamic mapping intensity of each actuator profile on each disturbance propagation path is compared one by one. The absolute value of the difference in dynamic mapping intensity on the corresponding path is calculated. The absolute differences in dynamic mapping intensity of the two actuator profiles on all disturbance propagation paths are accumulated, and the accumulated result is divided by the total number of disturbance propagation paths to obtain the average dynamic mapping difference between the two actuator profiles. This is used to quantify the overall coupling deviation of different actuator profiles in the disturbance propagation response.

[0137] If the average dynamic mapping difference between any pair of actuator contours exceeds the preset contour coupling deviation threshold, the current dynamic mapping relationship is marked as a high deviation state, triggering the time delay compensation process, and the contour-path pairs in the dynamic mapping relationship that exceed the threshold are recorded as mapping indices to be corrected.

[0138] In this embodiment, a correction instruction set is obtained by adjusting the control command sequence through backpropagation to compensate for time delay mismatch, including:

[0139] By analyzing the specific manifestations of time-series deviations through dynamic mapping relationships, the distribution of deviations can be obtained;

[0140] Based on the deviation distribution, abnormal time series points are identified for the portion exceeding the preset threshold.

[0141] In Example 1, the timing deviation data reflected by the dynamic mapping relationship is statistically analyzed. According to the preset time delay tolerance threshold, the actual disturbance response time of all control command sequences is compared with the target response time one by one, and the time delay deviation at each time is calculated. For all times when the time delay deviation exceeds the tolerance threshold, the corresponding control command time point or interval is marked as an abnormal timing point, forming an abnormal timing point set.

[0142] If an abnormal timing point is detected, the specific range of time delay mismatch is determined by comparing the preset threshold with the actual deviation value.

[0143] For specific intervals of time delay mismatch, the back propagation method is used to adjust the parameters of the control command sequence and generate preliminary correction commands.

[0144] In Example 1, for the control command interval where a time delay mismatch is detected, the difference between the target output and the actual disturbance propagation delay is set as the optimization target. By analyzing the influence path of each control parameter on the disturbance propagation delay within this interval, a mapping relationship between the control command and the time delay response is established. Based on the mapping relationship, each control parameter is gradually adjusted. In each adjustment process, the correction direction and magnitude of each control parameter are adaptively updated according to the optimization target, and the optimization is iteratively optimized in real time by combining historical data feedback results, gradually reducing the time delay deviation between the target output and the actual response until the optimization target reaches the preset convergence standard or the number of optimizations reaches the set threshold. Finally, a new set of control commands that can compensate for the time delay mismatch and match the dynamic mapping relationship is obtained as the initial correction command set.

[0145] Key control points are extracted from the initial correction instructions, and their matching degree with the instruction sequence is analyzed to obtain the adjusted instruction set.

[0146] If the adjusted instruction set still has local time delay mismatches, the final corrected instruction set is determined by fine-tuning the local points again.

[0147] In this embodiment, injecting a contour compensation factor to determine the convergence time of the flow field effect in the parallel scheduling of multiple actuators includes:

[0148] Based on the correction instruction set, and combined with the structural characteristics and control timing of multiple actuators, an initial motion coordination matrix is ​​constructed;

[0149] In Example 1, for each actuator's control command, the control start time, duration, and actuator number are extracted from the calibration command set. A two-dimensional motion coordination matrix is ​​constructed with the actuator number as the index and the control timing as the content, which is used to describe the collaborative relationship and time coverage interval when multiple actuators operate in parallel.

[0150] Based on the initial action coordination matrix and the two-dimensional array of dynamic spatiotemporal mapping relationship, the total disturbance superposition intensity at each moment is calculated to determine the convergence moment of the disturbance effect;

[0151] In Example 1, based on the initial action coordination matrix and the two-dimensional array of dynamic spatiotemporal mapping relationship, all time steps of the total scheduling time interval are traversed. The product of the action activation state of all actuators through each disturbance propagation path and the corresponding dynamic mapping intensity is calculated at each moment. The total disturbance superposition intensity at each moment is accumulated to obtain the sequence of the total disturbance superposition intensity over time. By finding the maximum point of the sequence, the moment when the disturbance effect converges is determined, that is, the time point when the flow field disturbance of each actuator is strongest in the key area of ​​the moon pool.

[0152] Based on the convergence moment candidate and target resonance avoidance window of the disturbance effect, a target offset evaluation function is constructed;

[0153] In Example 1, a resonance avoidance time window for the lunar pool system is preset. Based on the time distance between the convergence time of the disturbance effect and the target resonance avoidance window, a target offset evaluation function is constructed to quantify the degree of overlap between the convergence time of the disturbance and the resonance avoidance window. The smaller the target offset, the easier it is for the disturbance to induce lunar pool resonance.

[0154] In the construction of the optimized execution scheme, a contour compensation factor is introduced, and the target offset evaluation function is used as the optimization objective. The control start time of each actuator is fine-tuned to minimize the target offset.

[0155] In Example 1, the contour compensation factor is a time adjustment parameter determined by the actuator contour shape, disturbance propagation characteristics, and dynamic mapping relationship. The adjustment step size of the compensation factor is set to Δt. The gradient descent strategy is used with the target offset evaluation function as the loss function to iteratively update the control start time. In each round of adjustment, the disturbance superposition intensity curve and the corresponding target offset are recalculated. When the output value of the target offset evaluation function meets the preset tolerance, the optimization process is terminated, forming a control command scheduling sequence that meets the requirements of disturbance peak shifting and resonance avoidance. The contour compensation factor is used to adjust the local response rhythm in the disturbance propagation path, thereby avoiding the concentrated outbreak of flow field resonance effects in the moon pool region.

[0156] The final optimized execution plan is output, which requires that the target offset evaluation function not exceed the preset upper limit threshold, and that the actuator number, control start time and contour compensation factor correspond one-to-one.

[0157] In this embodiment, the final optimized execution plan includes:

[0158] Actuator A: Control start time adjusted to T1, contour compensation factor set to C1;

[0159] Actuator B: Control start time adjusted to T2, contour compensation factor set to C2;

[0160] Actuator C: Control start time adjusted to T3, contour compensation factor set to C3;

[0161] Where T1, T2, and T3 are the control times that finally satisfy the convergence condition of the target offset evaluation function, and C1, C2, and C3 are the corresponding contour compensation factors.

[0162] Example 2: During a maritime operation exercise, after the platform flow field monitoring system was activated, it collected raw flow field disturbance data of five actuators (numbered M1-M5) during their 0-60 second operation period via a sensor array. The instantaneous velocity signal, pressure gradient, and triaxial acceleration data of each actuator were uploaded in real time at a sampling rate of 1000Hz. The system initially calculated that the average flow field velocity of M1 was 0.46 m / s, and the pressure fluctuation variance was 0.13 Pa². The values ​​for M2, M3, M4, and M5 were 0.44 m / s, 0.48 m / s, 0.47 m / s, and 0.45 m / s, respectively, and the pressure variances of all of them did not exceed 0.18 Pa².

[0163] The real-time signal processing module first performs FIR filtering and notch filtering on the raw data of each actuator over 100 seconds. After removing frequency band noise, the signal-to-noise ratio of the denoised flow field vector signal is improved to over 28dB, which is 5-7dB higher than that of traditional FFT processing. Subsequently, by fusing the current and historical vector field states through a Kalman filter, the platform obtains a continuous corrected distribution of the moon pool flow field. The average vorticity peak values ​​of M2 and M3 during the 20-28 second period are detected to be 0.67 and 0.71, respectively, which are significantly higher than the static mean of 0.35.

[0164] In the propagation delay matrix calculation, the system selected the main propagation path from M1 to M5 to the center of the moon pool. Each path was divided into 16 units with a unit size of 0.5 meters. Water density, temperature, and salinity were collected for each unit. Taking unit S8 of path M3 as an example, the density was 1024.5 kg / m³, the temperature was 16.8℃, and the salinity was 35.2 ppt. Through multi-source data fusion and consistency calibration, a 5×16 propagation delay distribution matrix was finally constructed. The propagation delay of path M3 in unit S8 was 0.034 seconds. Compared with the traditional single-sensor method, the delay deviation was reduced from 0.013 seconds to 0.006 seconds, and the standard deviation of the delay data was reduced by 45%.

[0165] The system further processes the propagation delay distribution matrix in blocks, calculating the local equivalent propagation velocity for each actuator and each path unit, and obtaining the velocity increment accordingly. In Example 2, the local propagation velocities of M4 in units S11 and S12 are 0.51 m / s and 0.57 m / s, respectively, with a velocity increment of 0.06 m / s. Combining this with the path unit length (0.5 m), the system calculates the local eddy current intensity of M4 in S12 to be 0.144 using a formula.

[0166] The three-dimensional fluid velocity vector is decomposed, and the system obtains the principal direction angle of each element. In Example 2, the horizontal component ratio of M2 in element S9 is 1.15, and the principal direction angle is 48°. The local eddy current intensity, principal direction angle, time delay, and radial distance of all elements are combined to form the feature vector. Taking element S6 of M1 as an example, the feature vector is [0.091, 42°, 0.027, 0.42].

[0167] Based on simulation and experimental results, the system obtains a resonant period set of ${2.8, 4.9, 7.5}$ seconds. The perturbation propagation delay of M3 at S8 is 0.034 seconds, and the corresponding resonant sensitivity weight is calculated to be 0.88 according to the formula.

[0168] The sample weights are calculated by combining the weights, eddy current intensity reference values, and radial distance. The local combined weight of the S12 unit in M4 is 0.62. All sample weights are combined with the feature vectors and used as input to the improved random forest. Traditional random forests only use physical features without weighting.

[0169] On the training set, the system sampled 3,500 feature samples. After improving the random forest training, the system identified 93% of high-risk perturbations (true positive rate), compared to 77% for the traditional method, and the AUC value was improved by 0.14.

[0170] After constructing the three-dimensional joint array, low-rank decomposition yields the actuator factor array, path factor array, and feature factor array. In Example 2, the mapping intensities of M2 and M4 at path S12 are 0.74 and 0.69, respectively. After normalization, the difference between the two is 0.05, which does not exceed the threshold. The difference between M1 and M3 at S8 is 0.21, indicating a high deviation state detected by the system.

[0171] For the deviation region, the system backpropagates the optimized instructions. The original start time of scheduling M3 was 15.6 seconds, which was reduced to 17.2 seconds after compensation. The local delay mismatch was reduced from 0.12 seconds to 0.04 seconds. The final average delay deviation at all compensation points was 0.028 seconds, which is better than the 0.082 seconds achieved by traditional instruction adjustment.

[0172] During the parallel scheduling phase of the actuators, the system generates an initial action coordination matrix. M2 and M3, which were originally scheduled to start synchronously at 16.0 seconds and 15.6 seconds respectively, have their start times adjusted to 17.5 seconds and 17.2 seconds after contour compensation. The compensation factors are 1.08 and 0.96, successfully avoiding the resonance avoidance window (19.5-21.5 seconds). The peak value of the disturbance superposition intensity curve decreased from 9.8 to 3.2, and the overall offset was optimized from 1.06 seconds to 0.14 seconds, significantly lower than the preset threshold.

[0173] After the final operation was completed, the maximum disturbance amplitude of the moon pool flow field was 0.037m, far lower than the 0.112m under traditional scheduling. The average positioning accuracy of the robotic arm was 2.1mm (5.8mm under the traditional method), the success rate of the task was 100% (compared to 70% success rate with 3 failures in the control group), and the operation time was reduced by 16%. The system's high-risk warning lead time was increased from 1.2 seconds under the traditional method to 5.3 seconds.

[0174] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization, characterized in that, include: The flow field disturbance data generated by the actuator and the water movement parameters in the key area of ​​the moon pool are collected by the sensor array. Real-time signal processing is used to filter out noise and obtain the initial flow field model of the moon pool. The water density, temperature and salinity distribution along the propagation path from each actuator to the center of the moon pool are calculated based on the initial flow field model of the moon pool. Kalman filtering is used to fuse multi-source data to determine the propagation delay distribution matrix. The intensity and direction features of local eddy currents caused by profile interference are extracted from the propagation delay distribution matrix. The contribution of different mechanism profiles to flow field deformation is classified by an improved random forest algorithm to obtain the interference characteristic vector. A joint time-delay interference tensor is constructed for the interference characteristic vector and the propagation time delay distribution matrix, and the dynamic mapping relationship is determined by separating the joint time-delay interference tensor; If the timing deviation in the dynamic mapping relationship exceeds the preset threshold, the control command sequence is adjusted through back propagation to compensate for the time delay mismatch and obtain the correction command set. After obtaining the correction instruction set, the collaborative action logic is integrated, and the contour compensation factor is injected into the parallel scheduling of multiple actuators to determine the convergence time of the flow field effect, thereby obtaining an optimized execution scheme and completing the suppression of resonance in the lunar pool at sea.

2. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The process of using real-time signal processing to filter out noise and obtain the initial flow field model of the moon pool includes: The instantaneous velocity and pressure signals of the actuator are synchronously acquired by a multi-channel sensor array to obtain raw multi-dimensional time-series data. A finite impulse response filter is used to perform real-time bandpass filtering and notch filtering on the original multidimensional time series data to obtain denoised multidimensional time series data. The instantaneous velocity vector and pressure gradient at each sensor location are calculated based on the denoised multidimensional time-series data to obtain the current flow field vector field. The corrected real-time flow field distribution is obtained by fusing the current flow field vector field with the predicted flow field state from the previous moment using a Kalman filter. The average vorticity and velocity fluctuation intensity in the key area of ​​the moon pool are extracted based on the corrected real-time flow field distribution to obtain the hydrodynamic parameters of the moon pool area. If the hydrodynamic parameters in the moon pool area exceed the preset threshold range, the actuator control signal adjustment command is triggered to obtain the control quantity to suppress the disturbance; The actuator driving parameters are directly updated based on the control quantity that suppresses disturbances, and the corrected real-time flow field distribution is used as the prior state of the Kalman filter in the next cycle to obtain the initial flow field model of the moon pool for continuous tracking.

3. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The step of using Kalman filtering to fuse multi-source data and determine the propagation delay distribution matrix includes: By using the initial flow field model of the moon pool to obtain the original distribution data of water density, temperature and salinity along the propagation path from each actuator to the center of the moon pool, preliminary environmental parameter distribution results are obtained. The initial environmental parameter distribution results are calibrated for consistency to determine the calibrated multi-source dataset. From the calibrated multi-source dataset, feature values ​​related to water density, temperature and salinity are extracted. An environmental variable matrix along the propagation path is constructed through feature value analysis to obtain a comprehensive characterization of the environmental variables. The Kalman filter method is applied to fuse the comprehensive representation of environmental variables, calculate the estimated time delay along the propagation path, and obtain the preliminary distribution of propagation time delay. Based on the preliminary distribution of propagation delay, and combined with the positional relationship between the center of the moon pool and the actuator, the estimated delay value is spatially corrected to determine the deviation of the delay distribution. If the deviation exceeds the preset threshold, the delay value is iteratively adjusted to obtain the corrected delay distribution. By constructing the propagation delay distribution matrix using the corrected delay distribution, and then smoothing outliers in the matrix, the final propagation delay distribution matrix is ​​obtained.

4. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The improvement of the random forest algorithm to classify the contribution of different mechanism profiles to flow field deformation includes: Each element in the propagation delay distribution matrix represents the perturbation propagation delay of a discrete path unit on the propagation path of an actuator to the center of the moon pool. For each discrete path unit of each actuator, calculate the local propagation velocity increment; The local eddy intensity index is defined based on the local propagation velocity increment and the path unit length; From the initial flow field model of the moon pool, the three-dimensional fluid velocity vector corresponding to each propagation path unit is extracted, and the local vortex principal direction angle of the discrete path unit is calculated by using the ratio of the two orthogonal components of the three-dimensional fluid velocity vector in the horizontal plane. The local eddy current intensity index, the local eddy current principal direction angle, the disturbance propagation delay, and the radially normalized distance are combined into a four-dimensional column vector to obtain the local interference feature vector. The set of lunar pool resonance characteristic periods was obtained through experiments and numerical simulations, and the resonance sensitivity weight of the samples was calculated based on the perturbation propagation delay. Based on the resonance sensitivity weight, a reference value for eddy current intensity and an eddy current amplification coefficient are introduced to construct a local comprehensive sample weight; The local resonance risk index is obtained by multiplying the local comprehensive sample weight by the ratio of the local eddy current intensity reference value to the eddy current intensity reference value. An improved random forest model is constructed, which uses the local interference feature vector, local resonance risk index and local comprehensive sample weight as the training input for each sample. The weighted Gini impurity and weighted resonance risk of the sample are combined into a partitioning objective function. By optimizing the partitioning objective function, the model takes into account both the deformation mode of the classified flow field and the priority separation of samples with high resonance risk, so as to realize node partitioning and tree structure generation based on multi-objective collaborative optimization. For each decision tree, output the weighted probability contribution vector of each flow field deformation mode to which each sample belongs; The interferometric characteristic vector is obtained by integrating the outputs of all decision trees based on the weighted probability contribution vector.

5. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The construction of the time-delay interferometric joint tensor and the determination of the dynamic mapping relationship include: By mapping all actuator outline numbers, disturbance propagation path numbers, and interference feature dimensions one-to-one, a three-dimensional joint array is obtained; Perform low-rank decomposition on a three-dimensional joint array, and the decomposition result includes three factor arrays; Based on three factor arrays, a two-dimensional array of dynamic spatiotemporal mapping relationships is defined. Normalization is performed on all elements in the two-dimensional array of dynamic spatiotemporal mapping relationships, and based on the normalization results, the average dynamic mapping difference between each pair of actuator contours under all disturbance propagation paths is calculated. If the average dynamic mapping difference between any pair of actuator contours exceeds the preset contour coupling deviation threshold, the current dynamic mapping relationship is marked as a high deviation state, triggering the time delay compensation process, and the contour-path pairs in the dynamic mapping relationship that exceed the threshold are recorded as mapping indices to be corrected.

6. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The method of obtaining a correction instruction set by adjusting the control instruction sequence through backpropagation to compensate for time delay mismatch includes: By analyzing the specific manifestations of time-series deviations through dynamic mapping relationships, the distribution of deviations can be obtained; Based on the deviation distribution, abnormal time series points are identified for the portion exceeding the preset threshold. If an abnormal timing point is detected, the specific range of time delay mismatch is determined by comparing the preset threshold with the actual deviation value. For specific intervals of time delay mismatch, the back propagation method is used to adjust the parameters of the control command sequence and generate preliminary correction commands. Key control points are extracted from the initial correction instructions, and their matching degree with the instruction sequence is analyzed to obtain the adjusted instruction set. If the adjusted instruction set still has local time delay mismatches, the final corrected instruction set is determined by fine-tuning the local points again.

7. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The step of injecting a contour compensation factor in the parallel scheduling of multiple actuators to determine the convergence time of the flow field effect includes: Based on the correction instruction set, and combined with the structural characteristics and control timing of multiple actuators, an initial motion coordination matrix is ​​constructed; Based on the initial action coordination matrix and the two-dimensional array of dynamic spatiotemporal mapping relationship, the total disturbance superposition intensity at each moment is calculated to determine the convergence moment of the disturbance effect; Based on the convergence moment candidate and target resonance avoidance window of the disturbance effect, a target offset evaluation function is constructed; In the construction of the optimized execution scheme, a contour compensation factor is introduced, and the target offset evaluation function is used as the optimization objective. The control start time of each actuator is fine-tuned to minimize the target offset. The final optimized execution plan is output, which requires that the target offset evaluation function not exceed the preset upper limit threshold, and that the actuator number, control start time and contour compensation factor correspond one-to-one.

8. The method for suppressing lunar pool resonance at sea based on multi-objective collaborative optimization according to claim 1, characterized in that, The final optimized execution plan includes: Actuator A: Control start time adjusted to T1, contour compensation factor set to C1; Actuator B: Control start time adjusted to T2, contour compensation factor set to C2; Actuator C: Control start time adjusted to T3, contour compensation factor set to C3; Where T1, T2, and T3 are the control times that finally satisfy the convergence condition of the target offset evaluation function, and C1, C2, and C3 are the corresponding contour compensation factors.

Images