A method and system for monitoring underwater noise of marine wind power

Abstract

The present invention discloses a method and system for monitoring underwater noise of marine wind power plants, which belongs to the field of marine environmental monitoring, artificial intelligence and autonomous system technology. The method includes constructing a digital twin model that integrates multimodal perception; deducing the future state of the system and environment based on the digital twin model to generate a predictive state sequence; solving and generating a collaborative control strategy around the joint optimization goal of data quality and energy consumption; and analyzing the collaborative control strategy to synchronously regulate the physical system. The present invention adopts a cognitive decision-making and autonomous control architecture driven by digital twins. By constructing a high-fidelity digital twin, it conducts forward-looking, uncertainty deduction of the physical system and the marine environment, and makes closed-loop decisions based on the principle of risk-benefit optimization. This enables the monitoring system to have long-term, autonomous, and online self-improvement and risk avoidance capabilities, significantly improving the reliability of monitoring data in complex marine environments and the system's full life cycle autonomous operation level.

Description

Technical Field

[0001] The present invention relates to the technical fields of marine environment monitoring, artificial intelligence and autonomous systems, and in particular to a method and system for monitoring underwater noise of marine wind power plants. Background Art

[0002] With the rapid development of the offshore wind power industry, long-term monitoring of underwater noise generated during wind farm construction and operation is crucial for assessing environmental impacts and protecting marine ecosystems. Existing technologies typically use acoustic sensors, such as hydrophones, deployed in target waters via fixed or moored platforms to collect underwater sound field data, providing fundamental data support for environmental regulation and scientific research.

[0003] However, existing monitoring systems face challenges in maintaining data continuity, stability, and reliability when dealing with harsh marine environments. Furthermore, most systems still rely on frequent manual intervention for energy supply and equipment maintenance, making it difficult to meet the urgent need for long-term, uninterrupted, and highly autonomous monitoring of specific sea areas. This limits the ability to conduct detailed analysis and forward-looking assessments of dynamic changes in the marine environment. Summary of the Invention

[0004] To solve the above problems, the present invention provides a method and system for monitoring underwater noise of marine wind power plants, which adopts a cognitive decision-making and autonomous control architecture driven by digital twins. By constructing a high-fidelity digital twin, it conducts forward-looking and uncertainty deduction of the physical system and the marine environment, and makes closed-loop decisions based on the principle of optimal risk-return. This enables the monitoring system to have long-term, autonomous, and online self-improvement and risk avoidance capabilities, significantly improving the reliability of monitoring data in complex marine environments and the system's autonomous operation level throughout its life cycle.

[0005] The above objectives can be achieved through the following solutions:

[0006] A method for monitoring underwater noise in marine wind power plants comprises real-time acquisition and fusion of underwater acoustic signals, equipment attitude data, and environmental energy data to construct a digital twin model; based on the digital twin model, deducing equipment energy consumption, spatiotemporal distribution of environmental noise, and hydrodynamic disturbances at a preset time scale to generate a predictive state sequence; taking monitoring data quality and long-term operating energy consumption as joint optimization goals, solving and generating a collaborative control strategy based on the predictive state sequence; parsing the collaborative control strategy, synchronously regulating attitude adjustment and energy management, comparing monitoring data with the predictive state sequence in real time, and quantifying deviations; and triggering a safety plan to online correct the collaborative control strategy when the quantified deviation exceeds a preset deviation threshold.

[0007] Optionally, the construction of the digital twin model includes: abstracting the physical components of the system equipment and the external environmental elements as nodes, defining the physical coupling relationship and information interaction relationship of the physical components as edges, and constructing a dynamic system topology map; associating the underwater acoustic signals, equipment posture data and environmental energy data with the nodes in the dynamic system topology map to generate a historical time series attribute sequence training set; taking the historical time series attribute sequence as input and the time series attribute sequence at the next moment as output, and using the historical time series attribute sequence training set to establish and train a graph neural network to obtain a digital twin model.

[0008] Optionally, the construction of a dynamic system topology map includes: generating different node types using the physical components and the external environment, generating meta-paths using the paths between the node types, and fusing the node types and the meta-paths into a heterogeneous graph pattern; analyzing the node status based on the heterogeneous graph pattern and calculating an attention weight matrix; assigning the attention weight matrix to the corresponding edges in the heterogeneous graph pattern to generate a dynamic system topology map.

[0009] Optionally, generating a predictive state sequence includes: obtaining current real-time state data, and using the current real-time state data to initialize the state of the digital twin model to obtain a digital twin model in an initial state; using the state of the digital twin model at the previous moment as input, using the physical evolution law of the digital twin model in the initial state to perform forward deduction, and recursively calculating to generate a multidimensional state matrix; performing task decoding on the multidimensional state matrix, separating and extracting the evolution trajectory of the equipment energy consumption, the spatiotemporal distribution of the environmental noise, and the hydrodynamic disturbance, to obtain a predictive state sequence.

[0010] Optionally, the recursive calculation to generate a multidimensional state matrix includes: parameterizing the physical evolution law into a stochastic differential equation, and using the digital twin model of the initial state to fit the drift term and the diffusion term of the stochastic differential equation to obtain a stochastic differential equation model; performing multi-path forward integration based on the stochastic differential equation model to calculate the state matrix probability distribution; performing statistical moment calculation on the state matrix probability distribution, extracting the mathematical expectation of the state matrix probability distribution, and generating a multidimensional state matrix.

[0011] Optionally, the method also includes: quantifying the uncertainty of the predictive state sequence based on the state matrix probability distribution, and tracing it back to the weakly correlated edges in the dynamic system topology map to obtain a model uncertainty attribution report; calculating the model parameter correction amount for the weakly correlated edges according to the model uncertainty attribution report; and using the model parameter correction amount to perform online parameter updates on the digital twin model.

[0012] Optionally, the solution to generate a collaborative control strategy includes: constructing the monitoring data quality and long-term operating energy consumption into a joint optimization utility function, and discretizing the control instructions to obtain a finite control action space; guided by the joint optimization utility function, performing a forward search in the finite control action space, and evaluating in combination with the predictive state sequence to obtain a first action sequence; mapping and converting the first action sequence into a timing control instruction to generate a collaborative control strategy.

[0013] Optionally, the method also includes: taking the collaborative control strategy as input, performing forward-looking deduction on the state matrix probability distribution to obtain an execution effect distribution; based on the execution effect distribution, calculating the risk value of the collaborative control strategy, and quantifying the strategy execution risk; when the strategy execution risk is higher than a preset risk tolerance, generating a risk boundary condition, and using the risk boundary condition to re-solve the collaborative control strategy.

[0014] Optionally, the triggering of the safety plan to online correct the collaborative control strategy includes: determining the risk level in real time based on the quantitative deviation, and calling a safety plan that matches the risk level in a preset safety library; converting the safety plan into a safety control instruction, inserting it into the execution flow of the collaborative control strategy, and implementing instruction-level correction on the collaborative control strategy; after the execution of the safety control instruction is completed, re-evaluating the deviation threshold based on the latest monitoring data, and resuming the normal cyclic execution of the collaborative control strategy.

[0015] Based on the same inventive concept, the present invention also provides an underwater noise monitoring system for marine wind power, which includes: a multi-dimensional data fusion and modeling module, which is used to collect and fuse underwater acoustic signals, equipment posture data and environmental energy data in real time to construct a digital twin model; a state deduction and prediction module, which is used to deduce the equipment energy consumption, spatiotemporal distribution of environmental noise and hydrodynamic disturbances at a preset time scale based on the digital twin model to generate a predictive state sequence; a collaborative control strategy solution module, which is used to solve and generate a collaborative control strategy based on the predictive state sequence with monitoring data quality and long-term operation energy consumption as joint optimization goals; a strategy analysis and synchronous regulation module, which is used to analyze the collaborative control strategy, synchronously regulate posture adjustment and energy management, compare monitoring data with the predictive state sequence in real time and quantify the deviation. When the quantified deviation exceeds the preset deviation threshold, the safety plan is triggered to correct the collaborative control strategy online.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. By building a digital twin model of the device and inferring its interaction with the environment, this invention elevates the traditional passive protection approach, which relies on the material's inherent corrosion and impact resistance, to an intelligent adaptive approach that actively senses posture changes and makes real-time adjustments. This closed-loop control capability fundamentally solves the problems of monitoring position offset and data distortion caused by dynamic environmental disturbances such as wind, waves, and currents.

[0018] 2. This invention doesn't view monitoring tasks and energy management in isolation. Instead, it uses "monitoring data quality" and "long-term operating energy consumption" as joint optimization objectives. By predicting future states, it dynamically calculates the optimal coordinated control strategy. This approach intelligently balances and switches between "acquiring high-quality data" and "extremely energy-efficient operation," significantly enhancing the device's long-term autonomous survivability in volatile marine environments.

[0019] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 The present invention is a flow chart of a method for monitoring underwater noise of marine wind power plants.

[0022] Figure 2 It is a dynamic system topology map and attention weight map of an embodiment of the present invention.

[0023] Figure 3 It is a probabilistic evolution diagram of the predictive device energy consumption according to an embodiment of the present invention.

[0024] Figure 4 Schematic diagram of the Pareto optimal frontier of the collaborative control strategy according to an embodiment of the present invention.

[0025] Figure 5 It is a structural schematic diagram of an ocean wind power underwater noise monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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 making creative efforts shall fall within the scope of protection of the present invention.

[0027] Reference Figure 1 One embodiment of the present invention proposes a method for monitoring underwater noise of marine wind power plants. It adopts a cognitive decision-making and autonomous control architecture driven by digital twins. By constructing a high-fidelity digital twin, it conducts forward-looking and uncertainty deduction of the physical system and the marine environment, and makes closed-loop decisions based on the principle of risk-return optimization. This enables the monitoring system to have long-term, autonomous, and online self-improvement and risk avoidance capabilities, significantly improving the reliability of monitoring data in complex marine environments and the system's full life cycle autonomous operation level.

[0028] The method of this embodiment specifically includes:

[0029] Real-time collection and integration of underwater acoustic signals, equipment attitude data, and environmental energy data to build a digital twin model;

[0030] Based on the digital twin model, the equipment energy consumption, spatiotemporal distribution of environmental noise, and hydrodynamic disturbances at a preset time scale are simulated to generate a predictive state sequence;

[0031] Taking monitoring data quality and long-term operating energy consumption as joint optimization goals, a collaborative control strategy is calculated and generated according to the predictive state sequence;

[0032] The collaborative control strategy is analyzed, attitude adjustment and energy management are synchronously regulated, monitoring data is compared with the predictive state sequence in real time, and deviations are quantified. When the quantified deviation exceeds a preset deviation threshold, a safety plan is triggered to correct the collaborative control strategy online.

[0033] By adopting a cognitive decision-making and autonomous control architecture driven by digital twins, and by constructing high-fidelity digital twins to conduct forward-looking, uncertainty deductions of physical systems and marine environments, and making closed-loop decisions based on the principle of risk-return optimization, the monitoring system can have long-term, autonomous, and online self-improvement and risk avoidance capabilities, significantly improving the reliability of monitoring data in complex marine environments and the system's full life cycle autonomous operation level.

[0034] Optionally, building a digital twin model includes:

[0035] The physical components of the system equipment and the external environment elements are abstracted as nodes, and the physical coupling relationship and information interaction relationship of the physical components are defined as edges to construct a dynamic system topology map;

[0036] Specifically, this step aims to mathematically express complex physical systems using a structured, machine-readable graph data structure. A graph construction process first abstracts the key physical components of the monitoring equipment itself, such as hydrophones, attitude sensors, batteries, solar panels, wave energy harvesting devices, and the external environmental factors in which they are located, into nodes in the graph. Subsequently, the coupling relationships or information interaction relationships that have clear physical meanings between these nodes are defined as edges connecting the nodes. The collection of all these nodes and edges together constitutes a dynamic system topology graph that can characterize the topological structure of the entire system. The underwater acoustic signals, equipment attitude data, and environmental energy data are associated with the nodes in the dynamic system topology graph to generate a historical time series attribute sequence training set;

[0037] Specifically, this step aims to assign raw, multimodal sensor data to the graph structure constructed in the previous step. A data association process uses the real-time underwater acoustic signals, attitude data from the six-axis attitude sensor, and energy data from solar panels, wave energy harvesters, and batteries as dynamic attributes of their corresponding nodes. By continuously and chronologically recording this data, one or more time-series attribute sequences are generated for each node, representing the changes in its state over time. The collection of such attribute sequences for all nodes at all historical moments constitutes the historical time-series attribute sequence training set used for subsequent model training.

[0038] Taking the historical time series attribute sequence as input and the time series attribute sequence at the next moment as output, a graph neural network is established and trained using the historical time series attribute sequence training set to obtain a digital twin model.

[0039] Specifically, a model training process employs a graph neural network (GNN) architecture. This network takes a training set of historical time-series attribute sequences as input, and its training goal is to learn a mapping function that accurately predicts the graph state at the next moment based on the current state. This process not only learns statistical correlations in the data but also incorporates known physical laws into the neural network's loss function through an innovative physical constraint regularization term. By minimizing this hybrid loss function, which combines data-driven loss with physical constraints, a converged, physically constrained graph neural network (GNN) is ultimately obtained, whose internal evolutionary laws are highly consistent with the real world. This network is known as the digital twin model.

[0040] Optionally, constructing a dynamic system topology map includes:

[0041] Generating different node types using the physical components and the external environment, generating meta-paths using the paths between the node types, and fusing the node types and the meta-paths into a heterogeneous graph pattern;

[0042] Specifically, this step aims to construct a heterogeneous graph that can contain different types of information and characterize their complex relationships. First, entities are defined as different node types, such as "hydrophone node", "attitude sensor node", "energy management node" and "marine environment node". Subsequently, a set of meta-paths are defined based on the potential associations between these different types of nodes that have physical or information transfer significance. A meta-path describes a composite association path between different types of nodes. For example, the meta-path of "environment node → attitude sensor node → hydrophone node" represents the physical process of the marine environment affecting the device attitude and ultimately affecting the underwater acoustic signal acquisition. By fusing all predefined node types and meta-paths, a heterogeneous graph model that can describe multi-dimensional, heterogeneous relationships is constructed and generated.

[0043] Analyze the node status based on the heterogeneous graph pattern and calculate the attention weight matrix;

[0044] Specifically, this step aims to quantify the mutual influence between different nodes at the current moment. A Graph Attention Network (GAT) is used for this process. The network takes the time-series attribute sequence associated with each node as input and uses a self-attention mechanism to learn and calculate the attention weight between two nodes on each edge. This attention weight indicates how much attention should be given to the information of its neighboring nodes when predicting the future state of a node. The calculation process of an attention weight can be expressed by the formula: ,

[0045] in, It is a slave node To Node The attention weight of and Node and The eigenvector of is a shared weight matrix; is the weight vector of a single-layer feedforward neural network; Represents vector concatenation operation; is a node The set of all neighbor nodes of . The set of attention weights between all pairs of nodes constitutes the attention weight matrix.

[0046] The attention weight matrix is ​​assigned to the corresponding edges in the heterogeneous graph pattern to generate a dynamic system topology map.

[0047] Specifically, this step is the last step in completing the construction of the graph. An assignment process will use the attention weight matrix calculated in the previous step, which changes dynamically over time, as the weight value of the corresponding edge in the heterogeneous graph pattern. Through this step, a static graph pattern that only describes the connection relationship is converted into a dynamic weighted directed graph whose edge weights can reflect the importance and influence between nodes in real time. This weighted directed graph is the final generated dynamic system topology map, which provides highly concentrated and structured input for the subsequent training of higher-level digital twin models, such as Figure 2 As shown in the figure, the dynamic influence relationship between the physical components of the monitoring equipment and the external environmental factors is displayed in the form of a network diagram, where the thickness and grayscale of the edges represent the real-time influence weights between nodes under specific working conditions calculated by the graph attention network.

[0048] Optionally, generating a predictive state sequence includes:

[0049] The current real-time state data is obtained and used to initialize the state of the digital twin model to obtain a digital twin model in its initial state. Specifically, this step aims to align a universal, offline-trained digital twin model with the instantaneous real working conditions on site. A state initialization process collects the current, unprocessed real-time underwater acoustic signals, equipment posture data, and environmental energy data. Subsequently, this set of data is used as input to perform a one-time update or "calibration" of the network state of the obtained digital twin model, thereby obtaining a digital twin model in its initial state whose internal latent state is highly consistent with the current physical world, providing a reliable starting point for subsequent forward-looking deductions.

[0050] Taking the state of the digital twin model at the previous moment as input, forward deduction is performed using the physical evolution law of the digital twin model in the initial state, and a multi-dimensional state matrix is ​​generated by recursive calculation;

[0051] Specifically, the core of this step is to iteratively and step-by-step predict the state at multiple future time steps. A forward deduction process takes the complete digital twin model state at the previous moment as input, which can be represented as a high-dimensional tensor containing all node attributes and connection relationships. A forward propagation is performed using the learned physical evolution laws built into the graph neural network model to calculate the state at the next moment. By using this calculated new state as the input for the next iteration, and recursively performing this calculation, a multidimensional state matrix covering the entire prediction time scale and containing the states of all nodes at all future moments can be generated.

[0052] The multi-dimensional state matrix is ​​task-decoded to separate and extract the equipment energy consumption, the spatiotemporal distribution of the environmental noise, and the evolution trajectory of the hydrodynamic disturbance to obtain a predictive state sequence.

[0053] Specifically, this step aims to extract a sequence of understandable physical quantities that are directly related to the decision-making objectives of the present invention from a high-dimensional state matrix containing massive amounts of information. A task decoding process will first identify the nodes in the multidimensional state matrix that correspond to specific physical components or external environmental elements. Then, through a decoding function or a pre-trained output layer, the temporal attribute sequences of these nodes in the high-dimensional feature space are reversely mapped back to their corresponding evolutionary trajectories with clear physical meanings. For example, a certain attribute sequence of the "battery" node is decoded into the future "device energy consumption" curve, and the attribute sequences of the "hydrophone" and "water body" nodes are jointly decoded into the future "temporal and spatial distribution of environmental noise". The collection of all these separated and extracted trajectories constitutes the final predictive state sequence, such as Figure 3 As shown, the prediction results of the digital twin model of the present invention on future equipment energy consumption are displayed in the form of confidence bands. The figure not only gives the most likely evolution trend, but also quantifies the "uncertainty" of the prediction at different times through the width of the gray area.

[0054] Optionally, the recursive calculation to generate a multidimensional state matrix includes:

[0055] Parameterizing the physical evolution law into a stochastic differential equation, and fitting the drift term and the diffusion term of the stochastic differential equation using the digital twin model of the initial state to obtain a stochastic differential equation model;

[0056] Specifically, this step aims to explicitly and parameterize the complex physical evolution laws implicitly learned by the digital twin model constructed using a graph neural network (GNN) using a stochastic differential equation (SDE) with a well-defined mathematical form. A SDE can simultaneously describe the deterministic trends and stochastic fluctuations in the time evolution of a state. A model fitting process uses the initial state GNN model to fit the drift and diffusion terms in the SDE by applying a large number of small, known perturbations and observing the responses. This process distills the dynamic knowledge embedded in the "black-box" GNN model into a "white-box" SDE model that is mathematically easier to analyze and reason about. This model captures not only the predictable evolution of the state but also the inherent randomness caused by internal noise and external unmodeled factors, which is crucial for long-term, high-precision state reasoning.

[0057] Perform multipath forward integration based on the stochastic differential equation model to calculate the state matrix probability distribution;

[0058] Specifically, this step aims to utilize the stochastic differential equation model obtained in the previous step to simulate and predict all possible future state trajectories through numerical integration. Because stochastic differential equations contain random terms, each integration produces a unique future state evolution path that incorporates random perturbations. A forward integration process employs a high-order numerical solution algorithm, such as the Runge-Kutta method or the more suitable Euler-Maruyama method, to perform thousands of independent forward integrations starting from the current initial state. Each integration is like conducting an independent "future experiment" in the digital twin world, generating a vast collection of possible future state evolution paths. This collection collectively forms a state matrix probability distribution that comprehensively describes all future possibilities, not just the most likely trends, providing an extremely rich data foundation for subsequent risk assessment and robust decision-making.

[0059] Statistical moment calculation is performed on the state matrix probability distribution, the mathematical expectation of the state matrix probability distribution is extracted, and a multi-dimensional state matrix is ​​generated.

[0060] Specifically, this step aims to extract the most valuable key statistics for decision-making from the probability distribution obtained in the previous step, which contains a vast amount of possible information. A statistical moment calculation process analyzes the probability distribution of the state matrix at each future time step. It first calculates the first-order statistical moment of the probability distribution, known as the mathematical expectation (mean). This mean represents the most likely value at that moment and constitutes the core prediction for the future. To quantify the reliability of this prediction, the process also calculates the second-order statistical moment, known as the variance or covariance matrix. This variance precisely quantifies the uncertainty of the prediction at that moment. By extracting the mathematical expectation values ​​at all future time steps and arranging them in chronological order, the final multidimensional state matrix is ​​obtained. As the core of the predictive state sequence, each element of this matrix not only represents a predicted value but also contains an assessment of the confidence level of that predicted value.

[0061] Optionally, the method further includes:

[0062] Based on the state matrix probability distribution, the uncertainty of the predictive state sequence is quantified and traced back to the weakly associated edges in the dynamic system topology map to obtain a model uncertainty attribution report;

[0063] Specifically, this step aims not only to understand how large the "uncertainty" of the prediction is, but also to explore which part of the model "causes" this uncertainty. An uncertainty tracing process first extracts the covariance part from the obtained state matrix probability distribution as an overall quantification of the uncertainty of the predictive state sequence. Subsequently, an attribution algorithm based on sensitivity analysis or gradient backpropagation is used. This algorithm calculates the partial derivative of the overall uncertainty with respect to the weight parameter of each edge in the constructed dynamic system topology map. A high partial derivative value indicates that the edge is a "weakly correlated edge" and a small change in its parameter will have a huge impact on the stability of the final prediction. The results of all these sensitivity analyses are summarized to generate a model uncertainty attribution report that indicates the main sources of model uncertainty.

[0064] Calculating a model parameter correction amount for the weakly correlated edge according to the model uncertainty attribution report;

[0065] Specifically, this step aims to convert passive analysis results into active, executable optimization instructions. A correction calculation process takes the model uncertainty attribution report generated in the previous step as input. The report contains the identification of all weakly correlated edges and their sensitivity ranking. For each identified weakly correlated edge, a correction calculation function will calculate a specific parameter adjustment value for it. The calculation goal of this adjustment value is to make the digital twin model more "robust" in the direction of the weakly correlated edge, that is, to reduce its sensitivity to the uncertainty of the final prediction. An exemplary correction calculation can be expressed by the formula:

[0066] ,

[0067] in, It is for The parameter correction amount calculated from the model parameters of the weakly correlated edges; Is a learning rate hyperparameter that controls the step size of the correction; is a loss function that quantifies the uncertainty of the prediction; The loss function is parameterized The gradient of , whose value is provided by the model uncertainty attribution report.

[0068] The model parameter correction amount is used to perform online parameter update on the digital twin model.

[0069] Specifically, this step completes the online self-optimization loop of the model by implementing online parameter updates for the digital twin model using model parameter corrections. The online update process applies the model parameter corrections calculated for all weakly connected edges in the previous step to the resulting digital twin model. These corrections are then applied to the corresponding weight parameters of the graph neural network using an optimization algorithm, such as stochastic gradient descent (SGD). This online parameter update allows the digital twin model to dynamically and targetedly "reinforce" weak links in its internal structure during use, thereby demonstrating greater stability and lower prediction uncertainty in subsequent forecasting and deduction.

[0070] Optionally, the solving and generating the collaborative control strategy includes:

[0071] The monitoring data quality and long-term operating energy consumption are constructed as a joint optimization utility function, and the control instructions are discretized to obtain a finite control action space;

[0072] Specifically, this step aims to express the two conflicting top-level optimization goals, namely "obtaining high-quality data" and "maximum energy saving", with a unified mathematical function. A utility function construction process first discretizes all possible control instructions, such as "speed gear of attitude adjustment mechanism" and "power supply mode of energy system", into a finite control action space containing a finite number of optional actions. Subsequently, a joint optimization utility function is constructed, which is used to evaluate the comprehensive "utility" or "reward" that can be obtained after executing any control action. An exemplary joint optimization utility function can be expressed by the formula:

[0073] ,

[0074] in, Is performing an action and transfer to the next state The joint optimization utility value obtained after ; It is a function that quantifies the quality of monitoring data, and its value is related to the predicted environmental noise level and the device attitude stability; It is a function that quantifies the long-term energy consumption, and its value is determined by the predicted device energy consumption and battery health status; is a dynamic trade-off factor, with a value range of [0, 1], used to dynamically balance data quality and energy consumption. For example, when the battery is high and the environment is noisy, the trade-off factor is automatically increased, favoring actions that improve data quality; vice versa.

[0075] Guided by the joint optimization utility function, a forward search is performed in the finite control action space, and an evaluation is performed in combination with the predictive state sequence to obtain a first action sequence;

[0076] Specifically, this step aims to find an optimal action sequence that maximizes overall utility over a period of time in the future, within a limited, discretized action space. A forward search process uses a decision tree-based search algorithm, such as Monte Carlo Tree Search (MCTS). This algorithm uses the current state as the root node and constructs a search tree by performing a large number of fast, forward-looking simulations in a limited control action space. In each simulation, the predictive state sequence generated by the digital twin model is used to evaluate the possible future states that may be achieved after executing a certain action sequence, and the joint optimization utility function constructed in the previous step is used to score the path. By continuously exploring high-scoring paths at a deeper level, the algorithm can eventually find an action sequence starting from the current state with the highest expected cumulative utility value. This sequence is the first action sequence.

[0077] The first action sequence is mapped and converted into a timing control instruction to generate a collaborative control strategy.

[0078] Specifically, this step is to convert the abstract, optimal action sequence obtained in the previous step into a control instruction with a precise timestamp that can be directly executed by the physical hardware. An instruction conversion process will parse each "atomic action" in the first action sequence. Subsequently, the process generates one or more specific timing control instructions that comply with the underlying hardware communication protocol for each atomic action. All these instructions are organized and packaged according to their order in the action sequence, and their collection constitutes the final collaborative control strategy. This strategy will be distributed to the attitude adjustment system and the energy management system to guide their collaborative operation in the next time window, such as Figure 4 As shown, the present invention shows a set of optimal trade-off solutions found by a multi-objective optimization algorithm between the two conflicting goals of "monitoring data quality" and "long-term operating energy consumption", and illustrates how the system dynamically selects the most appropriate collaborative control strategy according to real-time operating conditions.

[0079] Optionally, the method further includes:

[0080] Taking the collaborative control strategy as input, performing forward-looking deduction on the state matrix probability distribution to obtain execution effect distribution;

[0081] Specifically, this step aims to conduct a comprehensive, probabilistic "sandbox simulation" of the generated, preliminary collaborative control strategy. A forward-looking simulation process will apply the collaborative control strategy to the generated state matrix probability distribution. Specifically, for the tens of thousands of independent future state trajectory samples contained in the probability distribution, a simulator will simulate the execution effect that each trajectory sample can ultimately achieve under the guidance of the collaborative control strategy. The execution effect here is a multidimensional vector that includes multiple key performance indicators such as "predicted monitoring data signal-to-noise ratio" and "predicted total equipment energy consumption". Since the input future state itself is probabilistic, the output execution effect is also a probability distribution containing multiple possible results, namely the execution effect distribution. This distribution not only gives the most likely effect after executing the strategy, but more importantly, it reveals the potential risks faced by the strategy under future uncertainty through its variance or tail shape.

[0082] Based on the execution effect distribution, the risk value of the collaborative control strategy is calculated to quantify the strategy execution risk;

[0083] Specifically, this step aims to extract a single, decision-making risk metric from the complex probability distribution of execution effects obtained in the previous step. This process uses a calculation method similar to the Value at Risk (VaR) in financial risk management. A risk quantification process analyzes the tail characteristics of the execution effect distribution. For example, it calculates the probability that the quality of monitoring data will fall below a critical threshold under this distribution, or the probability that long-term operating energy consumption will exceed a certain budget red line. This calculated value, which represents the probability of an adverse consequence occurring under the current strategy, or the value obtained by weighting this probability with the severity of the adverse consequence, is the final quantified strategy execution risk. This risk value evaluates the robustness of the current collaborative control strategy in an intuitive and comparable manner.

[0084] When the strategy execution risk is higher than a preset risk tolerance, a risk boundary condition is generated, and the collaborative control strategy is recalculated using the risk boundary condition.

[0085] Specifically, this step is a closed-loop link for "reinforcement" and "robustness enhancement" of the execution strategy. A decision-making process will compare the strategy execution risk quantified in the previous step with a risk tolerance preset based on equipment safety requirements or mission importance. If the strategy execution risk is higher than the risk tolerance, it indicates that the currently preliminarily generated collaborative control strategy is too "aggressive" and its performance under future uncertainties is unacceptable. At this point, a strategy recalculation process will be triggered. This process will convert potential adverse consequences that lead to high risks into a new set of hard constraints that must be avoided, namely risk boundary conditions. Subsequently, an optimizer will add this new set of risk boundary conditions to the joint optimization utility function and resolve the original optimization problem to obtain a new version of the collaborative control strategy that is more conservative, more robust, and optimized with risk constraints.

[0086] Optionally, triggering the safety plan to online correct the collaborative control strategy includes:

[0087] Determine the risk level in real time based on the quantitative deviation, and call a safety plan that matches the risk level from a preset safety library;

[0088] Specifically, this step aims to conduct a real-time, multi-dimensional risk assessment of the inconsistencies between the digital twin model and physical reality. A risk level determination process not only considers the quantitative deviation size at the current moment, but also creatively introduces the time evolution characteristics of the deviation. The process calculates the first-order derivative of the deviation value and the integral value within a specific time window. Through a multi-dimensional risk assessment function, the instantaneous amplitude, change rate and continuous accumulation of the deviation are nonlinearly fused to calculate a comprehensive real-time risk level. An exemplary risk level calculation function can be expressed by the formula:

[0089] ,

[0090] in, yes Real-time risk level at each moment; is the quantized deviation vector at that moment; 、 、 are the preset weights of deviation amplitude, change rate and accumulation; 、 、 A nonlinear function is used to map physical quantities to risk scores. After calculating the risk level, a plan call process calls a safety plan from a pre-defined safety plan library that precisely matches the risk level. The safety plan is a structured data object that defines a set of principled response measures to be taken at that risk level.

[0091] Converting the safety plan into safety control instructions, inserting the instructions into the execution flow of the collaborative control strategy, and performing instruction-level correction on the collaborative control strategy;

[0092] Specifically, this step aims to convert the relatively macro-level safety plan invoked in the previous step into precise instructions that can be immediately executed by the underlying hardware. An instruction conversion and correction process parses the principled measures contained in the safety plan, such as "reducing the response sensitivity of attitude adjustment" or "switching to the most reliable backup energy source." This process converts these principled measures into a set of specific safety control instructions, such as reducing the proportional-integral-derivative gain parameters of the attitude controller by 20%, or sending an instruction to the energy management to force a switch to lithium battery power. Subsequently, an instruction stream correction process pauses the currently executing, generated collaborative control strategy and dynamically inserts this newly generated, higher-priority safety control instruction set into the execution flow of the collaborative control strategy, thereby achieving instruction-level online correction of the original strategy.

[0093] After the safety control instruction is executed, the deviation threshold is re-evaluated based on the latest monitoring data, and the normal cyclic execution of the collaborative control strategy is resumed.

[0094] Specifically, this step is a key link in ensuring that learning can be done from a safety intervention and that the system can be restored to its normal optimized operating state safely and smoothly. After all safety control instructions are confirmed to have been executed, a post-processing and recovery process is initiated. This process will first re-evaluate the deviation threshold based on the monitoring data of the recent period. For example, if the deviation is caused by a severe but short-term environmental disturbance, the deviation threshold may be appropriately relaxed in the short term to avoid unnecessary and overly frequent plan triggering due to environmental aftermath. After completing the adaptive adjustment of the deviation threshold, the process will send a recovery instruction to all sub-systems to terminate the execution of the safety plan and resume the normal cycle execution of the generated collaborative control strategy. This process ensures that both decisive intervention can be made in times of danger and a smooth and intelligent return to normalcy can be achieved after the danger has passed. Example 1

[0095] To verify the feasibility of this invention, it was applied to a set of underwater noise long-term monitoring buoys deployed near a key waterway at a wind farm. The waters in which these buoys are located are known for complex hydrological conditions, frequent typhoons, and severe salt spray corrosion, posing significant challenges to the monitoring equipment's long-term autonomous operation, data collection quality, and energy self-sustainability.

[0096] To validate the beneficial effects of the present invention, a six-month continuous monitoring period was selected. The control group employed a conventional monitoring buoy equipped with a high-performance battery pack and a fixed hydrophone, whose control logic was based on pre-set threshold rules. The experimental group fully deployed the present invention, whose core features are an intelligent monitoring device equipped with a hydrophone, a six-axis attitude sensor, a micro-vector thruster, a flexible solar panel, and a built-in magnetorheological fluid wave energy harvester.

[0097] In this embodiment, a central processor integrates acoustic signals collected by hydrophones, attitude sensor data, and real-time data from various energy system components. Using a physically constrained graph neural network, it constructs a digital twin model that represents the complex coupling relationship between the device itself and the marine environment. This model not only includes physical component nodes such as hydrophones, batteries, and various sensors, but also innovatively introduces virtual nodes representing external environmental factors such as hydrodynamics, salinity, and temperature. Through a heterogeneous graph network based on an attention mechanism, the influence weights between these internal and external nodes are dynamically calculated, generating a dynamic system topology map.

[0098] Subsequently, based on this digital twin model, a deduction engine began to make forward-looking predictions about the state over the next 24 hours. By parameterizing the model's physical evolution laws as a set of stochastic differential equations and performing multipath forward integration, not only was the most likely future state trajectory of the device derived, but more importantly, a state matrix probability distribution that incorporated the uncertainty of the predictions was obtained. For example, in one deduction, it was predicted that 12 hours into the future, the device's attitude would be significantly deflected by a strong ocean current, resulting in an 85% probability of the hydrophone's signal-to-noise ratio dropping by more than 30%. Furthermore, if the current attitude adjustment force was maintained, the device's energy consumption would increase by 50%, resulting in a 70% probability of the battery being depleted within 20 hours.

[0099] Next, in response to this prediction, a collaborative control strategy solution module was activated. It considered "maintaining data quality" and "ensuring long-term energy consumption" as two conflicting optimization objectives. Using a multi-objective evolutionary algorithm, it calculated a "Pareto front" encompassing all optimal trade-offs. A dynamic decision function determined that the current battery charge level was still healthy and, based on this frontier, selected a collaborative control strategy that prioritized data quality at the expense of energy consumption. Before executing this strategy, robustness testing was performed. After simulating a virtual disturbance event in which ocean current intensity exceeded expectations by 20%, the strategy was fine-tuned and reinforced.

[0100] Ultimately, this reinforced collaborative control strategy was parsed into two parallel sub-strategies: an attitude control sub-strategy instructs the vector thrusters in the attitude control system to proactively adjust the buoy's attitude in an energy-optimal manner, minimizing its surface area. Simultaneously, an energy management sub-strategy instructs the energy system to temporarily switch to "wave energy priority" mode and appropriately reduce the frequency of data uploads from non-core sensors. The two sub-systems achieve highly precise synchronized control by exchanging predictive states in real time.

[0101] After a six-month comparative test, the present invention has demonstrated significant technical advantages in monitoring data quality, energy management efficiency, and autonomous operation capabilities. For specific data, please refer to Tables 1, 2, and 3.

[0102] Table 1 Comparison of monitoring data validity and quality

[0103] Test Group Data efficiency Average attitude deviation angle control group 76.40% 15.8 Experimental group 98.10% 1.2 Performance improvements +28.4% -92.40% ;

[0104] Table 2 Comparison of energy system autonomous operation efficiency

[0105] Test Group Energy self-sufficiency rate Battery exhaustion times Maximum continuous trouble-free operation time control group 45.30% 7 18 Experimental group 99.20% 0 >180 Performance improvements +119% -100% >900% ;

[0106] Table 3 Verification of survivability and decision-making intelligence under complex sea conditions

[0107] Test Group Number of simulated typhoon events Equipment survival rate Control group (traditional equipment) 2 0% Experimental group (present invention) 2 100% Improved robustness - N / A ;

[0108] Tables 1 to 3 above record comparative data of the method of the present invention in practical applications of underwater noise monitoring for marine wind power plants, demonstrating in detail the superior performance of the present invention in terms of improving data quality, energy self-sustaining capability, and intelligent adaptability to complex environments.

[0109] Table 1 shows the core differences between the two methods in ensuring monitoring data quality. The data shows that during the six-month test period, due to its fixed monitoring posture, the control group's data efficiency was only 76.4% when encountering strong ocean currents, and the average posture deviation angle was large. However, through active posture adjustment, the experimental group's data efficiency reached 98.1%, and the average posture deviation was controlled within a very small range. This clearly demonstrates the adaptive anti-interference capability of the present invention, which can fundamentally guarantee the accuracy and reliability of monitoring data.

[0110] Table 2 quantifies the level of intelligent energy management. The data shows that the control group's battery was completely depleted seven times during the entire test period, resulting in long data interruptions. However, the experimental group, through intelligent multi-energy collaboration and forward-looking predictions of future energy consumption, achieved an energy self-sufficiency rate of 99.2%. Operation was never interrupted due to battery depletion throughout the test period, and its longest continuous trouble-free operation time was over ten times that of traditional methods. This demonstrates that the collaborative control strategy of this invention can significantly enhance the long-term autonomous survivability of equipment in remote waters.

[0111] Table 3 examines the decision-making capabilities of the system in response to extreme weather events. During two simulated typhoon events, the control group's equipment terminated monitoring prematurely due to attitude loss or energy management failure. However, the experimental group was able to anticipate the significant hydrodynamic disturbances and zero solar input brought about by the typhoon and independently generate a collaborative control strategy prioritized for survival. Data shows that the experimental group successfully survived both typhoon events, achieving a 100% equipment survival rate, fully demonstrating the advanced nature and practical value of the combined optimization and dynamic decision-making mechanism embodied in this invention.

[0112] Based on the same inventive concept, the present invention also provides an underwater noise monitoring system for marine wind power generation, such as Figure 5 As shown, the system includes:

[0113] Multi-dimensional data fusion and modeling module, used to collect and fuse underwater acoustic signals, equipment posture data, and environmental energy data in real time to build a digital twin model;

[0114] A state deduction and prediction module is used to deduce the equipment energy consumption, spatiotemporal distribution of environmental noise, and hydrodynamic disturbances at a preset time scale based on the digital twin model to generate a predictive state sequence;

[0115] A collaborative control strategy solving module is used to solve and generate a collaborative control strategy based on the predictive state sequence, taking the monitoring data quality and long-term operation energy consumption as joint optimization objectives;

[0116] The strategy analysis and synchronous control module is used to analyze the collaborative control strategy, synchronously control posture adjustment and energy management, compare the monitoring data with the predictive state sequence in real time and quantify the deviation. When the quantified deviation exceeds the preset deviation threshold, the safety plan is triggered to correct the collaborative control strategy online.

[0117] It should be noted that the functional division and information interaction between the aforementioned modules are logical. Physically, they can be integrated into the same software platform or deployed in a distributed manner. The connections between them represent data and control flows, designed to collaboratively achieve the dynamic optimization of building energy consumption of the present invention. The foregoing description is merely an exemplary embodiment of the present invention and is not intended to limit its scope.

Claims

1. A method for monitoring underwater noise of marine wind power, characterized in that: The method comprises: Real-time collection and integration of underwater acoustic signals, equipment attitude data, and environmental energy data to build a digital twin model; Based on the digital twin model, the equipment energy consumption, spatiotemporal distribution of environmental noise, and hydrodynamic disturbances at a preset time scale are simulated to generate a predictive state sequence: Taking monitoring data quality and long-term operating energy consumption as joint optimization goals, a collaborative control strategy is calculated and generated according to the predictive state sequence; Analyze the collaborative control strategy, synchronously regulate attitude adjustment and energy management, compare monitoring data with the predictive state sequence in real time and quantify the deviation. When the quantified deviation exceeds a preset deviation threshold, trigger the safety plan to correct the collaborative control strategy online; Generating a predictive state sequence includes: Acquire current real-time state data, and use the current real-time state data to initialize the state of the digital twin model to obtain a digital twin model in an initial state; Taking the state of the digital twin model at the previous moment as input, forward deduction is performed using the physical evolution law of the digital twin model in the initial state, and a multi-dimensional state matrix is ​​generated by recursive calculation; Performing task decoding on the multidimensional state matrix, separating and extracting the equipment energy consumption, the spatiotemporal distribution of the environmental noise, and the evolution trajectory of the hydrodynamic disturbance, and obtaining a predictive state sequence; The recursive calculation to generate a multi-dimensional state matrix includes: Parameterizing the physical evolution law into a stochastic differential equation, and fitting the drift term and the diffusion term of the stochastic differential equation using the digital twin model of the initial state to obtain a stochastic differential equation model; Perform multipath forward integration based on the stochastic differential equation model to calculate the state matrix probability distribution; Performing statistical moment calculation on the state matrix probability distribution, extracting the mathematical expectation of the state matrix probability distribution, and generating a multidimensional state matrix; The method further comprises: Based on the probability distribution of the state matrix, the uncertainty of the predictive state sequence is quantified and traced back to the weakly associated edges in the dynamic system topology map to obtain a model uncertainty attribution report; Calculating a model parameter correction amount for the weakly correlated edge according to the model uncertainty attribution report; The model parameter correction amount is used to perform online parameter update on the digital twin model.

2. The method for monitoring underwater noise of marine wind power according to claim 1, characterized in that: The construction of the digital twin model includes: The physical components of the system equipment and the external environment elements are abstracted as nodes, and the physical coupling relationship and information interaction relationship of the physical components are defined as edges to construct a dynamic system topology map; Associating the underwater acoustic signal, device posture data, and environmental energy data with nodes in the dynamic system topology map to generate a historical time series attribute sequence training set; Taking the historical time series attribute sequence as input and the time series attribute sequence at the next moment as output, a graph neural network is established and trained using the historical time series attribute sequence training set to obtain a digital twin model.

3. The method for monitoring underwater noise of marine wind power according to claim 2, characterized in that: The construction of the dynamic system topology map includes: Generating different node types using the physical components and the external environment, generating meta-paths using the paths between the node types, and fusing the node types and the meta-paths into a heterogeneous graph pattern; Analyze the node status based on the heterogeneous graph pattern and calculate the attention weight matrix; The attention weight matrix is ​​assigned to the corresponding edges in the heterogeneous graph pattern to generate a dynamic system topology map.

4. The method for monitoring underwater noise of marine wind power according to claim 1, characterized in that: The solution to generate the collaborative control strategy includes: The monitoring data quality and long-term operating energy consumption are constructed as a joint optimization utility function, and the control instructions are discretized to obtain a finite control action space; Guided by the joint optimization utility function, a forward search is performed in the finite control action space, and an evaluation is performed in combination with the predictive state sequence to obtain a first action sequence; The first action sequence is mapped and converted into a timing control instruction to generate a collaborative control strategy.

5. The method for monitoring underwater noise of marine wind power according to claim 4, characterized in that: The method further comprises: Taking the collaborative control strategy as input, performing forward-looking deduction on the state matrix probability distribution to obtain execution effect distribution; Based on the execution effect distribution, the risk value of the collaborative control strategy is calculated to quantify the strategy execution risk; When the strategy execution risk is higher than a preset risk tolerance, a risk boundary condition is generated, and the collaborative control strategy is recalculated using the risk boundary condition.

6. The method for monitoring underwater noise of marine wind power according to claim 1, characterized in that: The triggering of the safety plan to online correct the collaborative control strategy includes: Determine the risk level in real time based on the quantitative deviation, and call a safety plan that matches the risk level from a preset safety library; Converting the safety plan into safety control instructions, inserting the instructions into the execution flow of the collaborative control strategy, and performing instruction-level correction on the collaborative control strategy; After the safety control instruction is executed, the deviation threshold is re-evaluated based on the latest monitoring data, and the normal cyclic execution of the collaborative control strategy is resumed.

7. A marine wind power underwater noise monitoring system, applied to a marine wind power underwater noise monitoring method according to any one of claims 1 to 6, characterized in that: The system comprises: Multi-dimensional data fusion and modeling module, used to collect and fuse underwater acoustic signals, equipment posture data, and environmental energy data in real time to build a digital twin model; A state deduction and prediction module is used to deduce the equipment energy consumption, spatiotemporal distribution of environmental noise, and hydrodynamic disturbances at a preset time scale based on the digital twin model to generate a predictive state sequence; A collaborative control strategy solving module is used to solve and generate a collaborative control strategy based on the predictive state sequence, taking the monitoring data quality and long-term operation energy consumption as joint optimization objectives; The strategy analysis and synchronous control module is used to analyze the collaborative control strategy, synchronously control posture adjustment and energy management, compare the monitoring data with the predictive state sequence in real time and quantify the deviation. When the quantified deviation exceeds the preset deviation threshold, the safety plan is triggered to correct the collaborative control strategy online.

Images