Adaptive Observation Dynamic Optimization Method and System for Extreme Sea States

Abstract

This application relates to the field of marine environmental monitoring and intelligent equipment technology, and provides a dynamic optimization method and system for adaptive observation in extreme sea states. The method includes: acquiring multi-source observation data; fusing features from the multi-source observation data to generate environmental state information; inputting the environmental state information into a reinforcement learning model to generate an observation control strategy; dynamically adjusting the operating parameters of the observation system based on the observation control strategy; and updating the reinforcement learning model based on the execution results after adjusting the operating parameters. This application achieves closed-loop linkage between perception, decision-making, and execution, effectively solving the problems of poor data reliability and insufficient system adaptability in extreme environments, and achieving a dynamic balance between observation quality, system safety, and energy consumption.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring and intelligent equipment technology, specifically to a dynamic optimization method and system for adaptive observation in extreme sea conditions. Background Technology

[0002] With the increasing demands for marine resource development and marine environmental protection, higher requirements are being placed on marine environmental monitoring technologies. Marine environmental monitoring under extreme sea conditions has always faced significant technical challenges. Current marine monitoring technologies mainly rely on platforms such as satellite remote sensing, buoy sensors, and unmanned surface vessels for data acquisition. For example, some existing technologies integrate multi-source data for real-time 3D modeling, but these have significant shortcomings in terms of data reliability and system adaptability under extreme sea conditions. Some existing technologies optimize data collection through deep reinforcement learning, but do not fully consider the system stability and observation quality assurance issues under extreme sea conditions. Other technologies focus on data processing and forecasting algorithms, with insufficient research on the adaptive capabilities of the observation system itself. Existing solutions suffer from problems such as traditional buoy systems relying on preset thresholds to trigger observations, which cannot adapt to sudden nonlinear changes in wind, waves, and turbulence; single-sensor architecture leading to spatiotemporal calibration errors; insufficient multimodal data fusion; and a lack of adaptive capabilities in the observation system, making it unable to dynamically adjust observation strategies according to sea state changes. This results in incomplete data acquisition, poor system adaptability, and low energy efficiency under extreme conditions. Summary of the Invention

[0003] To address the problems of poor adaptability and incomplete data acquisition in existing technologies for observation systems under extreme sea conditions, this application proposes an adaptive observation dynamic optimization method and system for extreme sea conditions to achieve a dynamic balance between observation quality, system safety, and energy consumption.

[0004] To achieve the above objectives, this application adopts the following technical solution: An adaptive observation dynamic optimization method for extreme sea states includes: acquiring multi-source observation data; performing feature fusion on the multi-source observation data to generate environmental state information; inputting the environmental state information into a reinforcement learning model to generate an observation control strategy; dynamically adjusting the operating parameters of the observation system based on the observation control strategy; and updating the reinforcement learning model according to the execution results after the operating parameters are adjusted.

[0005] The above scheme uses the environmental state information generated by fusing multi-source observation data as input to the reinforcement learning model, and dynamically adjusts the system operating parameters based on the observation control strategy output by the model. Then, it updates the model based on the execution results, realizing the closed-loop linkage of perception, decision-making and execution, and solving the problems of poor data reliability and insufficient system adaptability in extreme environments.

[0006] Optionally, feature fusion is performed on the multi-source observation data to generate environmental state information, including: using an attention mechanism to perform feature weighted fusion on the multi-source observation data, dynamically evaluating the confidence weight of each data source, and generating the environmental state information.

[0007] The above scheme introduces an attention mechanism to dynamically evaluate the confidence weight of each data source, which can effectively cope with the lack of single sensor data or noise interference under extreme sea conditions, and improve the accuracy and robustness of the fused environmental state information.

[0008] Optionally, the feature-weighted fusion of the multi-source observation data using the attention mechanism includes: for optical wave sensor data, calculating weights based on image sharpness and jitter level; for microwave surface flow sensor data, calculating weights based on signal-to-noise ratio and spectral quality; and for intelligent temperature, salinity, and depth meter data, calculating weights based on self-correction historical error and data stability.

[0009] The above scheme achieves refined confidence assessment of heterogeneous data sources in extreme environments by setting specific weight calculation criteria for different types of sensors, further improving the accuracy of multimodal data fusion.

[0010] Optionally, updating the reinforcement learning model based on the execution result after adjusting the running parameters includes: calculating a reward value based on the execution result using a multi-objective reward function; and updating the parameters of the reinforcement learning model based on the reward value.

[0011] The above scheme comprehensively evaluates the execution results through a multi-objective reward function, which can guide the reinforcement learning model to find the optimal solution among multiple dimensions such as observation quality, energy consumption and security, thereby improving the scientific nature of model evolution.

[0012] Optionally, the multi-objective reward function includes a data quality reward, an energy consumption penalty, a security reward, and an exploration reward; the data quality reward is calculated based on data integrity and signal-to-noise ratio; the energy consumption penalty is calculated based on total system power consumption; the security reward is calculated based on device status and environmental risk; and the exploration reward is used to provide positive incentives when exploring unknown sea conditions within a safe range.

[0013] The above scheme refines the reward function into four specific dimensions, enabling the model to comprehensively consider data acquisition quality, system endurance, device security, and the ability to explore unknown environments during the update process.

[0014] Optionally, the reinforcement learning model uses a proximal policy optimization algorithm for policy updates.

[0015] The above scheme, by employing a near-end policy optimization algorithm, ensures the stability of policy network training in non-stationary extreme sea state environments and avoids system loss of control due to excessively large policy update steps.

[0016] Optionally, the dynamic adjustment of the operating parameters of the observation system includes: when the wave height exceeds a preset threshold, reducing the sampling frequency of the optical wave sensor and increasing the sampling frequency of the microwave surface flow sensor; when the remaining energy of the system is lower than a preset threshold, suspending the auxiliary observation task to ensure the operation of the core sensors; and when a sudden ocean phenomenon is identified, increasing the sampling frequency of the relevant sensors and adjusting the course of the unmanned observation platform for tracking and observation.

[0017] The above solution enables the system to flexibly switch working modes according to real-time sea conditions by setting specific control strategies for different extreme scenarios, thereby extending the system's endurance while ensuring the collection of core data and improving the response capability to emergencies.

[0018] Furthermore, this application also provides an adaptive observation system for extreme sea states, comprising: a cross-modal fusion module configured to acquire multi-source observation data, perform feature fusion on the multi-source observation data, and generate environmental state information; a reinforcement learning decision engine connected to the cross-modal fusion module, configured to receive the environmental state information and generate an observation control strategy; an execution module connected to the reinforcement learning decision engine, configured to dynamically adjust the operating parameters of the observation system based on the observation control strategy; and an update module connected to the execution module, configured to update the reinforcement learning decision engine according to the execution result after the operating parameters are adjusted.

[0019] The above scheme achieves adaptive observation and continuous optimization of the system under extreme sea conditions through data flow and closed-loop interaction mechanisms between modules.

[0020] Optionally, the system further includes a multimodal sensor array, which includes: an optical wave sensor configured to measure wave height, wave direction, and wave period; a microwave surface flow sensor configured to retrieve the sea surface flow field distribution; an intelligent temperature, salinity, and depth meter configured to collect temperature, salinity, and depth data; an underwater acoustic communication module configured to realize data transmission between the underwater sensors and the surface platform; and a meteorological parameter acquisition unit configured to measure meteorological elements such as wind speed, wind direction, air temperature, humidity, and air pressure.

[0021] The above scheme constructs a comprehensive marine environmental raw data acquisition network by configuring multiple types of sensors, providing a rich data foundation for subsequent fusion and decision-making.

[0022] In addition, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described above.

[0023] Beneficial effects: This application generates environmental state information by fusing features from multi-source observation data. This environmental state information is then used as input to a reinforcement learning model. Based on the reinforcement learning model, an observation control strategy is output to dynamically adjust the operating parameters of the observation system. Finally, the model is updated based on the execution results, forming a closed loop. This mechanism, through the deep linkage between fused perception and reinforcement learning decision-making, enables the system to adaptively adjust its observation strategy in real time according to the nonlinear changes in extreme sea states, overcoming the problem of poor adaptability caused by traditional reliance on preset thresholds. Simultaneously, by combining a multi-objective reward mechanism with a near-end strategy optimization algorithm, the system effectively reduces energy consumption while ensuring data integrity and system security, significantly improving the endurance and data acquisition reliability of the observation system under extreme conditions. Specifically, through multimodal sensor data fusion and adaptive observation strategies, data integrity is improved by more than 30% under extreme sea states; the reinforcement learning-based dynamic optimization method enables the system to maintain normal operation even when wave height exceeds 4 meters; intelligent scheduling and resource allocation extend the system's endurance by more than 40%; and the edge computing and cloud-based collaborative processing architecture reduces the system response time from hours to minutes, significantly improving the early warning capability for sudden marine phenomena. This application achieves a dynamic balance between observation quality, system security, and energy consumption. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a basic method for adaptive observation and dynamic optimization in extreme sea states, according to an embodiment of this application. Figure 2 This is a schematic diagram of the extreme sea state adaptive observation system according to an embodiment of this application; Figure 3 This is a flowchart of the multimodal data fusion processing according to an embodiment of this application; Figure 4 This is a framework diagram of the reinforcement learning dynamic optimization algorithm according to an embodiment of this application; Figure 5 This is a flowchart illustrating the adaptive observation process under extreme sea conditions, as described in this application embodiment. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In the description of this application, it should be understood that the terms “first,” “second,” etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated.

[0027] Example 1: This embodiment provides a basic method for adaptive observation and dynamic optimization in extreme sea states. This method is applied to applications such as... Figure 2 The overall architecture shown in the diagram depicts the four core layers of the system and their interactions when operating in extreme sea conditions: The perception layer consists of a multimodal sensor array deployed on the main control buoy platform, auxiliary unmanned surface vessels, and diving buoys. It is responsible for collecting raw multimodal data on the marine environment, including waves, currents, temperature, salinity, depth, and meteorology.

[0028] Edge Intelligence Layer: This is the core intelligent unit of the system. Data from the perception layer is gathered here via underwater acoustics or short-range wireless communication. Edge computing nodes utilize GPU computing power for real-time preprocessing and feature extraction; the cross-modal fusion module employs an attention-based algorithm for feature fusion; the reinforcement learning decision engine (using a PPO-RMSprop hybrid algorithm) makes real-time decisions based on the fused environmental state information, generating adaptive observation strategies or control commands (such as adjusting sensor modes or platform attitude).

[0029] Network transport layer: Employs a multi-mode communication architecture (satellite, 4G / 5G, microwave, underwater acoustic) to ensure reliable communication in various scenarios under extreme sea conditions. It is responsible for uplinking fused data and status reports from the edge layer to the cloud platform, and transmitting optimization strategies and control commands issued by the cloud platform to the edge layer and execution units.

[0030] The cloud platform application layer, centered on an autonomous observation and management platform, provides data storage, global status visualization, early warning information dissemination, and remote control interfaces. It receives information from multiple edge nodes, enabling users to perform global monitoring. Simultaneously, it possesses powerful computing capabilities, allowing for large-scale simulation training and model optimization, and synchronizes the updated policy model downlink to the edge-side reinforcement learning decision engine, achieving continuous system evolution.

[0031] The diagram clearly shows the data flow (uplink) and control flow (downlink), forming a closed loop from perception to decision-making to execution, and achieving continuous system optimization through cloud-edge collaboration.

[0032] like Figure 1 As shown, the method includes the following steps.

[0033] Step S110: Acquire multi-source observation data, perform feature fusion on the multi-source observation data, and generate environmental status information.

[0034] Specifically, multi-source observation data refers to raw, heterogeneous data collected by different types of sensors under extreme sea conditions, such as optical image sequences, microwave radar point clouds, and temperature, salinity, and depth time series. Because a single sensor is easily affected by environmental noise or data loss under extreme sea conditions, directly using raw data for decision-making can lead to system misjudgments. By fusing features from multi-source observation data, a comprehensive feature vector representing the true intensity of the current sea state and the system's own state can be extracted, i.e., environmental state information. In this embodiment, environmental state information serves as the core bridge between perception and decision-making, transforming multi-dimensional physical quantities into standard state inputs that can be processed by reinforcement learning models, shielding the heterogeneity of the underlying sensors, and enabling the decision engine to focus on the evolution of the environmental situation. It should be understood that multi-source observation data is not limited to the types listed above, but can also include meteorological parameters, underwater acoustic communication status data, etc. Any data that can reflect the marine environment or system operating status is within the scope of this application.

[0035] Step S120: Input the environmental state information into the reinforcement learning model to generate an observation and control strategy.

[0036] Specifically, reinforcement learning models learn optimal strategies through continuous interaction with the environment. In this embodiment, after environmental state information is input into the reinforcement learning model, the model evaluates the potential benefits of each action based on the current state and outputs an observation control strategy. It is crucial to distinguish that the observation control strategy refers to a set of instructions used to adjust the operating parameters or working mode of the observation system hardware, such as adjusting sensor sampling frequency, changing platform heading, or reallocating energy; while observation data refers to the physical quantities of the marine environment collected by sensors. The observation control strategy is a means for the system to actively adapt to the environment, while observation data is the result of the system's passive perception. The reinforcement learning model generates the observation control strategy based on environmental state information, enabling the system to make proactive adaptive adjustments to nonlinear changes in extreme sea states, rather than simply relying on preset thresholds for passive triggering.

[0037] Step S130: Based on the observation control strategy, dynamically adjust the operating parameters of the observation system.

[0038] Specifically, the operating parameters of the observation system include, but are not limited to, sensor sampling frequency, transmission power, observation platform attitude control parameters, and communication link priority. Once the observation control strategy is generated, the system execution module parses it into specific hardware control instructions. For example, when sea conditions deteriorate, the observation control strategy might instruct the reduction of the sampling frequency of optical sensors susceptible to wave spray, while simultaneously increasing the sampling frequency of microwave sensors with stronger penetrating power. This dynamic adjustment mechanism allows the system to concentrate limited computational and energy resources on the most effective observation methods under extreme conditions, ensuring the continuity of core data.

[0039] Step S140: Update the reinforcement learning model based on the execution results after adjusting the running parameters.

[0040] Specifically, the execution results include the quality of new observation data collected by the adjusted system, changes in system energy consumption, and equipment safety status. The system feeds these results back to the reinforcement learning model, which calculates a reward value to evaluate the effectiveness of the previously adopted observation and control strategy and updates its internal network parameters accordingly. This step forms a closed-loop linkage mechanism between perception, decision-making, and execution. Through this closed-loop feedback, the reinforcement learning model can continuously learn from its experience with extreme sea conditions, optimizing its strategy output. This allows the system to make better parameter adjustment decisions when encountering similar sea conditions again, thereby achieving a dynamic balance between observation quality, system safety, and energy consumption.

[0041] Example 2: This embodiment, based on Embodiment 1, further refines the feature fusion process for multi-source observation data. Specifically, it involves fusing features from multi-source observation data to generate environmental state information, including: using an attention mechanism to perform weighted feature fusion of multi-source observation data, dynamically evaluating the confidence weights of each data source, and generating environmental state information.

[0042] The system inputs multimodal raw data, including heterogeneous data from multiple sources such as optical, microwave, hydrological, and meteorological data. The data then enters a parallel preprocessing and standardization module for denoising, calibration, and format unification, laying the foundation for subsequent fusion.

[0043] Combination Figure 3As shown, after the multi-source observation data enters the fusion module, it is not simply averaged or spliced. Instead, an attention mechanism network scores the reliability of each data source at the current moment. Because the degree of interference to different sensors varies dynamically under extreme sea conditions—for example, the data quality of optical sensors drops sharply when blocked by high waves and spray, while microwave sensors, with their stronger penetration, are less affected—the attention mechanism can capture these quality fluctuations in real time. It assigns higher confidence weights to high-quality data sources and lower weights to interfered data sources, thus effectively suppressing noise contamination during the fusion stage and generating highly robust environmental state information.

[0044] Furthermore, an attention mechanism is used to perform feature-weighted fusion of multi-source observation data, including specific weight calculation logic for different types of sensors. For optical wave sensor data, weights are calculated based on image sharpness and jitter; for microwave surface flow sensor data, weights are calculated based on signal-to-noise ratio and spectral quality; and for intelligent temperature, salinity, and depth meter data, weights are calculated based on self-correcting historical errors and data stability.

[0045] Specifically, optical wave sensors primarily rely on visual images to measure wave height and direction. When strong winds cause significant sea spray to obscure the lens, or when the buoy platform oscillates violently, image clarity decreases significantly and image jitter intensifies. The system quantifies clarity by calculating the energy of high-frequency components or edge sharpness of the image, and quantifies jitter by calculating the variance of optical flow changes between adjacent frames. Lower clarity and higher jitter are assigned lower weights. Microwave surface flow sensors invert the flow field by transmitting and receiving radar waves. Under extreme sea conditions, they are susceptible to sea clutter interference. The system assesses data reliability by analyzing the signal-to-noise ratio (SNR) of the echo signal and the concentration of the main lobe (i.e., spectral quality). Higher SNR and more concentrated main lobes indicate less clutter interference, and are assigned higher weights. Intelligent temperature, salinity, and depth meters are typically deployed underwater, where they are less directly affected by surface wind and waves. However, they may be affected by water turbulence or biological attachment. The system uses its built-in self-calibration mechanism to output historical error correction values ​​and the stability of data variance within a certain time window as evaluation criteria. The smaller the self-calibration historical error and the more stable the data fluctuations, the higher the weight assigned.

[0046] Furthermore, before feature fusion, multi-source observation data needs to undergo feature extraction processing. Specifically, for optical wave sensor data, a convolutional neural network encoder is used to extract visual features, including in-depth characterization of wave morphology parameters such as wave height, wave direction, and wave period; for microwave surface flow sensor data, point cloud neural networks or fast Fourier transforms are used to extract flow field features, including frequency domain features of flow velocity and direction; for intelligent temperature, salinity, and depth instrument data, long short-term memory networks or statistical feature extraction methods are used to extract hydrological features, including the temporal variation patterns of temperature, salinity, and depth time series; for meteorological data, statistical feature extraction methods are used to extract features of meteorological elements such as wind speed, wind direction, air temperature, humidity, and air pressure. The extracted features are then fed into an attention-based feature weighting and fusion module to generate a fused feature vector.

[0047] Furthermore, the fused feature vectors need to undergo quality assessment and confidence calculation before being output.

[0048] Specifically, the system sets a confidence threshold θ to evaluate the quality of the fused feature vectors. When the confidence level is higher than the threshold θ, the fused data is deemed reliable, a high-quality fused data product is generated, and output to the reinforcement learning decision engine and the data storage / visualization platform. When the confidence level is lower than the threshold θ or a conflict is detected between data sources, the fused data is deemed unreliable, a feedback optimization mechanism is initiated, and the reinforcement learning decision engine is triggered to readjust the sensor configuration and re-acquire and fuse data, forming a closed-loop optimization. The specific value of the confidence threshold θ can be configured according to the task accuracy requirements and historical data statistics; for example, it can be set to any value between 0.8 and 0.95. This mechanism ensures that the environmental state information input to the reinforcement learning decision engine always meets the quality requirements, avoiding decision bias caused by low-quality data.

[0049] Example 3: This embodiment further refines the update mechanism of the reinforcement learning model based on Embodiment 1. Specifically, the reinforcement learning model is updated according to the execution result after adjusting the running parameters, including: calculating the reward value based on the execution result using a multi-objective reward function; and updating the parameters of the reinforcement learning model based on the reward value.

[0050] Combination Figure 4As shown, after executing the observation and control strategy, the reinforcement learning decision engine collects the execution results after adjusting the system operating parameters and maps them to a reward signal that the reinforcement learning model can recognize. The reward value is a comprehensive evaluation of the merits of the observation and control strategy just adopted. The model guides the update direction of network parameters by maximizing the long-term cumulative reward value. Because the system faces multiple mutually constraining objectives under extreme sea conditions—for example, increasing the sampling frequency can acquire more data but increases energy consumption—a single-objective reward function can cause the model to get stuck in local optima, such as excessively shutting down sensors to save power or exhausting energy to pursue data quality. Therefore, this embodiment introduces a multi-objective reward function, unifying conflicting optimization objectives within the same mathematical framework, enabling the model to find the global optimum under complex constraints.

[0051] Furthermore, the multi-objective reward function includes a data quality reward, an energy consumption penalty, a security reward, and an exploration reward. The data quality reward is calculated based on data integrity and signal-to-noise ratio. The energy consumption penalty is calculated based on the total power consumption of the system. The security reward is calculated based on equipment status and environmental risk. The exploration reward is used to provide positive incentives when exploring unknown sea conditions within a safe range.

[0052] Specifically, multi-objective reward function The calculation formula can be expressed as: .in, For data quality awards, As an energy consumption penalty item, For safety rewards, To explore the rewards, , , , These are the weighting coefficients for each reward item. These four dimensions do not exist independently, but rather achieve a synergistic balance between observation quality, energy consumption, and safety through dynamic adjustment of the weighting coefficients.

[0053] The data quality reward is calculated based on data integrity and signal-to-noise ratio (SNR). Data integrity can be quantified by the ratio of the actual number of valid data packets received to the theoretical number of data packets that should be sent; the SNR is directly derived from the underlying signal processing module of the sensor. When the system adjusts the sensor sampling frequency under extreme sea conditions, a positive reward is given if data integrity and SNR remain at a high level, and a negative penalty is given otherwise. This dimension encourages the model to acquire as much high-quality environmental information as possible while ensuring the system's survival.

[0054] The energy consumption penalty is calculated based on the total system power consumption. Total system power consumption includes the sum of the instantaneous power of all sensors, communication modules, computing units, and platform thrusters. When the observation and control strategy output by the model causes the total system power consumption to exceed a preset power baseline, the energy consumption penalty generates a negative reward. This dimension interacts with the data quality reward, forcing the model to learn to reduce energy consumption by shutting down redundant modules or reducing the frequency of non-critical sensors, while still meeting basic data acquisition requirements, thereby extending the system's endurance under extreme conditions.

[0055] The safety reward is calculated based on equipment status and environmental risk. Equipment status includes indicators such as platform attitude angle, cabin watertightness, and battery temperature; environmental risk includes the degree to which sea state parameters such as real-time wave height and current velocity exceed the equipment's safety threshold. When abnormal equipment attitude or environmental parameters approaching their limits are detected, the safety reward generates a strong negative reward, which can even directly override the weights of other reward items, forcing the model to output a conservative control strategy. This dimension constitutes the bottom-line constraint for system survival, ensuring that the model does not sacrifice platform safety in pursuit of data or exploration.

[0056] The exploration reward term provides a positive incentive for exploring unknown sea states within a safe range. In reinforcement learning, models are prone to relying solely on known optimal strategies, leading to sluggish responses to unfamiliar extreme sea state evolution patterns. The exploration reward term encourages the model to attempt to adjust sensor combinations or change platform course to capture rare ocean phenomena (such as sudden internal waves) within boundary conditions where the safety reward term does not trigger an alarm. This design allows the system to not only passively adapt to known environments but also actively accumulate experience in coping with new environments.

[0057] Through the synergistic effect of the above four dimensions, the multi-objective reward function can comprehensively evaluate the execution results. (Weight coefficients) , , , The weighting can be configured based on task priority. In a preferred embodiment, the weighting coefficients are configured as follows: data quality reward weight 0.5, energy consumption penalty weight 0.3, security reward weight 0.2, and exploration reward weight dynamically allocated within the range of 0 to 0.1 according to task requirements. For example, in the core area of ​​a typhoon, the weighting can be appropriately increased. The proportion is emphasized to highlight survival; when the typhoon is on the periphery and energy is abundant, it can increase and The proportion of data is increased to obtain more information. Based on the calculated comprehensive reward value, the model updates the parameters of the policy network and value network through the backpropagation algorithm, so that when encountering similar environmental conditions in the future, it can output observation and control strategies that obtain higher expected rewards, thereby realizing the continuous evolution and dynamic optimization of the system under extreme sea conditions.

[0058] Example 4: This embodiment further refines the policy update mechanism of the reinforcement learning model based on Embodiment 1. Specifically, the reinforcement learning model uses a proximal policy optimization algorithm for policy update.

[0059] Combination Figure 4 As shown, after collecting empirical data, the reinforcement learning decision engine uses the Proximal Policy Optimization (PPO) algorithm to update the policy network. In non-stationary extreme sea state environments, environmental conditions such as waves and wind fields change drastically and unpredictably. Traditional reinforcement learning algorithms (such as DQN or unrestricted policy gradient methods) are prone to policy collapse when updating network parameters due to excessively large single step sizes. This can lead to the observation system outputting incorrect observation and control policies, resulting in equipment damage or data loss. The PPO algorithm ensures the stability of the training process by limiting the differences between the old and new policies.

[0060] Specifically, the PPO algorithm uses a truncation mechanism for its objective function to update the policy. Its standard objective function formula is as follows: ; in, Let the objective function of the policy network be... These are the parameters to be optimized for the policy network; Indicates at time step The expectation operator on; The policy probability ratio represents the probability of the current policy in state . Take action below The ratio of the probability of taking the same action to the probability of taking the same action under the old strategy; The advantage function is used to evaluate the superiority of the current action relative to the average level. The advantage function is calculated using the generalized advantage estimation method. The variance of the advantage estimation is reduced by exponentially weighting the time-series difference residuals, thereby improving the accuracy of the policy gradient estimation. This is a hyperparameter used to limit the magnitude of policy updates; As the truncation function, the policy probability ratio Limited to Within the interval. Through this objective function, when the dominant function... When the current action is positive (i.e., better than the average), the policy probability ratio is limited to [value missing]. Within this range, to prevent policy abrupt changes due to overuse of the action; when the advantage function When it is negative, the policy probability ratio is restricted to Within a safe and reliable range, the system prevents excessive deviation from its strategy. This mechanism ensures that, even in extreme sea conditions, each strategy update is performed within a safe and reliable range, achieving a smooth transition from "automation" to "intelligence."

[0061] Furthermore, in solving the gradient of the objective function and performing backpropagation of parameters, this embodiment preferably uses the RMSprop optimizer for adaptive learning rate adjustment. Under extreme sea conditions, the environmental state exhibits strong non-stationary characteristics, and the feature distributions of data from different sensors vary greatly. If a fixed learning rate is used, convergence may be slow in gradient oscillation regions or the step size may be too small in flat regions. The RMSprop optimizer dynamically adjusts the learning rate of each parameter by using an exponentially weighted moving average of the squared gradient. A smaller learning rate is assigned to frequently updated parameters, and a larger learning rate is assigned to sparsely updated parameters. This adaptive mechanism enables the model to quickly decay abnormal gradients and maintain robustness of parameter updates when faced with sudden, severe gradient fluctuations caused by sudden sea conditions. It should be understood that this application is not limited to using the RMSprop optimizer; in other embodiments, equivalent alternatives with adaptive learning rate adjustment functions, such as the Adam optimizer, can also be used. However, the reason why this application prefers RMSprop is that the Adam optimizer introduces a momentum term, which may cause the policy update direction to lag behind the environmental change due to the accumulation of historical momentum in an extremely non-stationary environment. In contrast, RMSprop only relies on the second moment estimation of the current gradient, which is more agile in response to non-stationary environments and can adapt to the instantaneous deterioration of sea state more quickly. Thus, it can achieve rapid convergence and optimization of the observation policy while ensuring the survival of the system.

[0062] Example 5: This embodiment, based on Embodiment 1, further refines the specific actions for dynamically adjusting the operating parameters of the observation system. Specifically, dynamically adjusting the operating parameters of the observation system includes: when the detected wave height exceeds a preset threshold, reducing the sampling frequency of the optical wave sensor and increasing the sampling frequency of the microwave surface flow sensor; when the remaining energy of the system is lower than a preset threshold, suspending auxiliary observation tasks to ensure the operation of the core sensors; and when a sudden ocean phenomenon is identified, increasing the sampling frequency of the relevant sensors and adjusting the course of the unmanned observation platform for tracking and observation.

[0063] Combination Figure 5 As shown, the system executes the following stages sequentially: Initial response (milliseconds to seconds): Based on a pre-defined rule base, protective actions are immediately triggered, such as adjusting precision sensors to protection mode, adjusting the attitude of buoys or unmanned surface vessels to enhance stability, and activating backup power. The goal of this stage is to ensure system survival and buy time for subsequent intelligent decision-making.

[0064] Environmental perception and assessment: The multimodal fusion module begins operation, comprehensively assessing the true intensity of the current sea state (such as significant wave height, wave period, and current velocity) and the system's own state (such as equipment safety thresholds and energy levels). This provides accurate state input for intelligent decision-making.

[0065] Intelligent decision generation: The reinforcement learning decision engine calculates a dynamic optimization strategy in real time based on the current state. This strategy is no longer a simple "switch", but a fine-grained control instruction, such as: dynamically adjusting the working mode of the sensor (e.g., temporarily shutting down the optical sensor and increasing the frequency of the microwave sensor), reconfiguring the priority of the observation task (ensuring only the monitoring of core elements), optimizing energy allocation (reducing the power consumption of non-core functions), and adjusting the communication strategy (reducing the transmission frequency and sending in standby mode).

[0066] Strategy Execution and Continuous Monitoring: The system executes the above-mentioned refined strategies and continuously monitors environmental changes and the effectiveness of strategy execution, forming a rapid "perception-decision-execution" closed loop.

[0067] Learning and Optimization Phase: Once sea conditions calm down, the system enters the learning phase. Instead of resetting, it evaluates the performance and data value of the strategies employed throughout the extreme event, storing these experiences in an experience replay buffer and fine-tuning the policy network parameters online. This allows the system to perform better the next time it encounters a similar situation.

[0068] Recovery and Readiness: The system eventually recovers to high-performance observation mode and generates a complete, high-quality dataset about this extreme event, while waiting for the next event.

[0069] The flowchart clearly shows that when facing extreme sea conditions, the system of this invention does not simply exhibit "shutdown protection," but rather a resilient, gradually degraded, and intelligently recoverable "adaptive" capability, truly achieving the design goal of "being able to survive, measure accurately, and learn" in extreme environments.

[0070] In the system's adaptive observation workflow under extreme sea conditions, the observation control strategy generated by the reinforcement learning decision engine includes specific adjustment actions for different extreme scenarios. By binding these adjustment actions to specific threshold parameters, the system can flexibly switch operating modes according to real-time sea conditions, extending system endurance while ensuring core data acquisition and improving response capabilities to emergencies.

[0071] Scenario 1: When the wave height exceeds a preset threshold, the sampling frequency of the optical wave sensor is reduced, while the sampling frequency of the microwave surface flow sensor is increased. Specifically, the optical wave sensor relies on visual images for measurement. When strong winds cause significant wave spray to obscure the lens, image clarity decreases considerably. The system identifies when the current wave height exceeds the preset threshold (e.g., 4 meters) based on environmental status information, triggering a control action. This involves reducing the sampling frequency of the optical wave sensor (e.g., from 10Hz to 1Hz or 0.5Hz) to minimize invalid data collection and energy consumption. Simultaneously, the sampling frequency of the microwave surface flow sensor, which has stronger penetrating power, is increased (e.g., from 2Hz to 5Hz) to compensate for the data loss caused by the optical sensor's frequency reduction, ensuring the continuity of flow field distribution data. It should be understood that the preset wave height threshold is not limited to 4 meters and can be configured according to the observation platform's wave resistance level and mission requirements. For example, it can be set to any value between 3 and 6 meters. When the wave height reaches 3 meters, primary frequency reduction protection is triggered; when the wave height reaches 6 meters, the optical sensor is completely shut down. The control logic in this scenario is based on matching the physical characteristics of the sensors with the degree of sea state interference, which enables the precise transfer of observation resources under adverse sea conditions.

[0072] Scenario 2: When the system's remaining energy falls below a preset threshold, auxiliary observation tasks are suspended to ensure the operation of core sensors. Specifically, in extreme sea conditions, the system may face interruptions in solar power supply or limitations in generator operation, making energy consumption a key factor restricting the system's endurance. When the system detects through its energy management module that the remaining energy is below a preset threshold, such as 20%, an energy allocation optimization strategy is triggered, suspending auxiliary observation tasks. For example, non-core temperature and humidity sensors in the meteorological parameter acquisition unit may be shut down, or the transmission frequency of the underwater acoustic communication module may be reduced. Simultaneously, the operation of core sensors, such as the intelligent temperature, salinity, and depth analyzer and the microwave surface current sensor, is ensured to guarantee the continuous acquisition of critical marine environmental data. Similarly, the preset threshold for remaining energy is not limited to 20% and can be set to any value between 15% and 25%, depending on the system's battery capacity and task priorities. The control logic in this scenario is driven by an energy consumption penalty term in the multi-objective reward function, forcing the model to learn to degrade tasks when energy is scarce, thereby extending the system's survival time under extreme conditions.

[0073] Scenario 3: When a sudden ocean phenomenon is detected, the sampling frequency of relevant sensors is increased, and the course of the unmanned observation platform is adjusted for tracking observation. Specifically, sudden ocean phenomena, such as internal ocean waves, are characterized by short duration, small spatial scale, and strong destructive power, making them easy to miss with conventional low-frequency observations. The system identifies signs of sudden ocean phenomena through multimodal data fusion, such as dramatic stratification or abrupt changes in current velocity in temperature, salinity, and depth data, triggering tracking observation actions. This involves increasing the sampling frequency of relevant sensors, for example, increasing the sampling frequency of the intelligent temperature, salinity, and depth meter from 1Hz to 5Hz to capture high-frequency change characteristics. Simultaneously, the course of the unmanned observation platform is adjusted for tracking observation, for example, controlling the unmanned surface vessel to sail against the current at a specific heading angle or drift with the waves to remain located in the core area of ​​the phenomenon. The control logic in this scenario is driven by the exploration reward term in the multi-objective reward function, encouraging the model to actively capture rare phenomena within a safe range and improving the system's ability to explore unknown environments. Scenario 4: When the system detects abnormal attitude of the observation platform or a decrease in platform stability due to sea conditions, the thruster power and attitude of the observation platform are adjusted. Specifically, the reinforcement learning decision engine 20 sends power adjustment commands to the platform's thrusters through the execution module 30. For example, it increases the thruster power of the unmanned surface vessel from 50% to 80% of its rated power, while adjusting the heading angle to the wave-facing direction to reduce the roll angle and enhance the platform's stability in large waves. The control logic in this scenario is driven by the safety reward term in the multi-objective reward function, ensuring the platform's physical safety under extreme sea conditions. Scenario 5: When the system detects a decline in communication link quality or bandwidth limitations, it adjusts the priority and transmission frequency of the data communication links. Specifically, under extreme sea conditions, satellite communication may experience signal attenuation due to atmospheric disturbances, and 4G / 5G mobile communication may be interrupted due to limited base station coverage. The system evaluates the signal-to-noise ratio and available bandwidth of each communication link through the communication quality monitoring module, dynamically adjusting data transmission strategies. For example, it reduces the transmission frequency of non-core data, prioritizes the uplink transmission of environmental status information and early warning information, or switches the communication link from 4G / 5G to satellite communication to ensure uninterrupted data link operation. The control logic in this scenario ensures the reliability of communication between the system and the cloud platform under extreme conditions. Through the synergistic effect of the above five scenarios, the system can achieve a dynamic balance between observation quality, system energy consumption, and safety under extreme sea conditions.

[0074] Example 6: This embodiment provides an adaptive observation system for extreme sea states. Combined with... Figure 1As shown, the system includes: a cross-modal fusion module configured to acquire multi-source observation data, perform feature fusion on the multi-source observation data, and generate environmental state information; a reinforcement learning decision engine connected to the cross-modal fusion module, configured to receive environmental state information and generate observation control strategies; and an execution module connected to the reinforcement learning decision engine, configured to dynamically adjust the operating parameters of the observation system based on the observation control strategies, such as... Figure 2 As shown, the execution module can send the fused data and control commands to the corresponding observation system through the network transport layer; the update module is connected to the execution module and configured to update the reinforcement learning decision engine based on the execution results adjusted according to the running parameters.

[0075] Specifically, the cross-modal fusion module, reinforcement learning decision engine, execution module, and update module together constitute a system-level closed-loop architecture for perception, decision-making, and execution. The cross-modal fusion module corresponds to step S110 in Embodiment 1. It receives raw heterogeneous data from the underlying multimodal sensor array, performs feature weighting and spatiotemporal registration on the data through an internally deployed attention mechanism network, and outputs highly robust environmental state information. Physically, this module can be deployed on the local edge computing nodes of the observation platform to reduce data transmission latency and ensure real-time perception.

[0076] The reinforcement learning decision engine corresponds to step S120 in Embodiment 1, which establishes a high-speed data connection with the cross-modal fusion module through an internal bus or network interface. The reinforcement learning decision engine 20 receives the fused environmental state information and uses its built-in near-end policy optimization algorithm to evaluate the current sea state and output specific observation and control strategies. At the hardware level, this engine can be implemented using an embedded computing board equipped with a graphics processing unit (GPU) or a neural network computing acceleration chip to meet the computational power requirements of complex matrix operations.

[0077] The execution module corresponds to step S130 in Embodiment 1. It is connected to the reinforcement learning decision engine, receives the observation control strategy, and parses it into hardware-recognizable level signals or control commands. The execution module includes various drive circuits and relay arrays for directly controlling the sensor's sampling frequency, adjusting the platform thruster power, or switching communication links. For example, when the observation control strategy instructs a reduction in the optical sensor's sampling frequency, the execution module achieves the frequency reduction operation by sending a frequency division command to the sensor's clock control pin.

[0078] The update module corresponds to step S140 in Embodiment 1. It is connected to the execution module and collects the execution results after the operating parameters are adjusted, including the signal-to-noise ratio of the newly acquired data, the instantaneous power consumption of the system, and the device attitude feedback. The update module converts these execution results into reward signals and transmits them back to the reinforcement learning decision engine to trigger online fine-tuning of the model parameters. Through this data flow and closed-loop interaction mechanism between the cross-modal fusion module, the reinforcement learning decision engine, the execution module, and the update module, the system can not only passively adapt to extreme sea conditions, but also continuously optimize the decision network based on the execution results, achieving continuous evolution and dynamic equilibrium of the system in extreme environments.

[0079] Example 7: Based on Embodiment 6, this embodiment further includes a multimodal sensor array, which comprises: an optical wave sensor configured to measure wave height, wave direction, and wave period; a microwave surface flow sensor configured to retrieve the sea surface flow field distribution; an intelligent temperature, salinity, and depth meter configured to collect temperature, salinity, and depth data; an underwater acoustic communication module configured to enable data transmission between the underwater sensors and the surface platform; and a meteorological parameter acquisition unit configured to measure meteorological elements such as wind speed, wind direction, air temperature, humidity, and air pressure.

[0080] Combination Figure 2 As shown, the multimodal sensor array constitutes the system's sensing layer and is the physical basis for acquiring multi-source observation data in extreme sea states. By configuring multiple types of sensors, a comprehensive raw marine environmental data acquisition network is constructed, providing a rich data foundation for subsequent fusion and decision-making. It should be understood that the multimodal sensor array is not limited to the specific models or parameters listed below; any sensor capable of performing the corresponding measurement function is within the scope of protection of this application.

[0081] For optical wave sensors, measurements are achieved based on high-precision camera equipment and computer vision algorithms. In a preferred embodiment, the wave height measurement range of the optical wave sensor is 0 to 25 meters, with an accuracy of ±0.1 meters; the wave direction measurement range is 0 to 360 degrees, with an accuracy of ±2 degrees; the wave period measurement accuracy is within ±0.1 seconds, and the sampling frequency can reach 10 Hz. Due to the extremely low visibility and large amount of spray on the sea surface during extreme sea states such as typhoons, optical wave sensors are usually equipped with optical protective covers with automatic water droplet removal functions and installed at the highest point of the observation platform to reduce direct obstruction by waves. This high sampling rate configuration can capture high-frequency changes in wave morphology in the early stages of sea state deterioration, providing fine-grained visual feature input for the cross-modal fusion module.

[0082] For microwave surface flow sensors, frequency-modulated continuous wave radar technology is used to invert the sea surface flow field distribution. Specifically, the microwave surface flow sensor measures flow velocity from 0 to 5 m / s with an accuracy of ±0.05 m / s; flow direction measures flow direction from 0 to 360 degrees with an accuracy of ±3 degrees, and the sampling frequency is 2 Hz. Compared to optical sensors, microwave radar has stronger penetration capabilities through rain and fog. When optical wave sensors are degraded due to droplet obstruction, microwave surface flow sensors can continuously provide flow field data as the core sensing method. Its installation location is typically on the side of the observation platform, with the radar beam illuminating the sea surface at a specific grazing angle. Surface flow velocity and direction are calculated using the echo Doppler frequency shift.

[0083] The intelligent CTD (Conductivity, Temperature, and Depth) meter integrates temperature, salinity, and depth sensing functions and employs a self-calibration mechanism to ensure data accuracy. In terms of hardware configuration, the temperature measurement range is -5 to 40 degrees Celsius with an accuracy of ±0.01 degrees Celsius; the salinity measurement range is 0 to 40 PSU with an accuracy of ±0.01 PSU; and the depth measurement range is 0 to 6000 meters with an accuracy of ±0.05% of full scale, with a sampling frequency of 1 Hz. The intelligent CTD meter is typically deployed at a certain depth underwater and connected to the main control platform via cable. Its built-in self-calibration mechanism can automatically compensate for sensor drift based on historical profile data. Even in extreme sea conditions causing severe water mixing, it can still output stable CTD profile data, providing a reliable hydrological characteristic benchmark for multimodal data fusion.

[0084] The underwater acoustic communication module is configured to enable data transmission between underwater sensors and the surface platform. In extreme sea conditions, surface cable connections are prone to breakage due to severe tension; underwater acoustic communication, as a wireless transmission method, ensures uninterrupted data link between underwater nodes and the surface control platform. The underwater acoustic communication module employs a spread-spectrum communication system resistant to multipath interference, guaranteeing data transmission integrity in noisy marine environments. It should be understood that the underwater acoustic communication module can not only transmit data from the intelligent CTD-DDE instrument but also serve as a convergence channel for underwater node data within the entire multimodal sensor array.

[0085] The meteorological parameter acquisition unit is configured to measure meteorological elements such as wind speed, wind direction, air temperature, humidity, and air pressure. During a typhoon, the rapid changes in air pressure and wind speed are important indicators of the severity of sea conditions. The anemometers in the meteorological parameter acquisition unit typically employ ultrasonic anemometers designed to withstand strong winds, avoiding the jamming or damage of traditional mechanical anemometers under extremely high wind speeds. The measured meteorological and hydrological data are simultaneously input into the cross-modal fusion module, enabling the generated environmental state information to comprehensively reflect the air-sea coupling effect, providing a more comprehensive state perception input for the reinforcement learning decision engine. Through the above hardware configuration, the multimodal sensor array achieves full-element coverage from underwater to the surface and from meteorological to hydrological data at the physical level. Furthermore, each sensor has performance redundancy reserved to cope with extreme environments, thus ensuring the system's survivability and data acquisition reliability under extreme sea conditions at the hardware level.

[0086] Example 8: This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can perform the steps of the method as described in any one of Embodiments 1 to 5.

[0087] Specifically, the computer-readable storage medium can be any physical medium capable of storing program instructions and being read by a computing device. In an extreme sea state adaptive observation system, this storage medium is typically deployed within an edge computing node local to the observation platform to persistently store the underlying driver code of the reinforcement learning decision engine, the multimodal data fusion algorithm model, and system operation logs. When the system powers on or is restarted due to extreme sea states, the processor reads and executes the computer program from the medium, thereby completing the fusion generation of environmental state information, the output of the observation control strategy, the dynamic adjustment of system operating parameters, and the closed-loop update of the reinforcement learning model, in the order of steps S110 to S140 in Embodiment 1.

[0088] It should be understood that the physical form of computer-readable storage media is not limited to solid-state drives or hard disk drives. For example, on platforms with limited space and harsh environments, such as ocean buoys, the media can employ highly reliable read-only memory (ROM) or flash memory to ensure that program instructions are not lost under strong vibration and high salt spray conditions. In scenarios requiring a large amount of temporary data interaction, the media can also include random access memory (RAM) or electrically erasable programmable read-only memory (EEPROM). Furthermore, the media can also be portable storage devices such as magnetic disks and optical disks, facilitating direct upgrades and replacements of algorithm models by personnel when the platform is docked for maintenance. By providing this technical solution of computer-readable storage media, this application achieves comprehensive protection of the adaptive observation dynamic optimization method for extreme sea states at the software carrier level, ensuring that the core closed-loop decision logic can be stably deployed and run on different hardware platforms.

[0089] Example 9: The following is combined with Figure 5 Taking a specific application scenario of a passing typhoon as an example, this study demonstrates the complete closed-loop process of an adaptive observation system for extreme sea states. In this scenario, the typhoon center is set to gradually approach the observation area, and environmental parameters deteriorate rapidly, with real-time monitored wave heights exceeding 4 meters and wind speeds exceeding 15 meters per second.

[0090] Upon arrival at the typhoon's leading edge, the system first enters the initial response phase. Within a timescale of milliseconds to seconds, the system immediately triggers protective actions based on a pre-set rule base. The execution module sends emergency commands to the hardware layer, controlling precision sensors to enter protection mode, such as closing the lens cap of the optical wave sensor or adjusting it to a safe attitude; simultaneously, it adjusts the attitude of the observation platform to enhance stability and activates the emergency power supply. The core objective of this phase is to ensure system survival and buy time for subsequent intelligent decision-making.

[0091] As the main body of the typhoon covered the area, the system entered the multimodal sensing and fusion assessment phase. The multimodal sensor array continuously collected data, and the cross-modal fusion module acquired multi-source observation data, performing feature fusion to generate environmental state information. Specifically, due to the presence of a large amount of spray and strong winds on the sea surface, the image clarity of the optical wave sensor decreased sharply, while the microwave surface flow sensor, with its penetrating power, acquired relatively reliable flow field data. The cross-modal fusion module used an attention mechanism to dynamically evaluate the confidence weight of each data source, significantly reducing the weight of optical data and increasing the weight of microwave data, thereby generating environmental state information that accurately reflects the true intensity of the current severe sea conditions.

[0092] Subsequently, the system enters the reinforcement learning stage to generate refined strategies. The reinforcement learning decision engine receives environmental state information and generates observation and control strategies. Guided by a multi-objective reward function, the reinforcement learning decision engine calculates a set of dynamically optimized strategies: shutting down the vulnerable optical wave sensor, increasing the sampling frequency of the microwave surface flow sensor from 2Hz to 5Hz to compensate for data loss, adjusting the unmanned observation platform's heading towards the waves to reduce the roll angle, and reducing the transmission frequency of the underwater acoustic communication module to save energy. Based on this observation and control strategy, the execution module dynamically adjusts the operating parameters of the observation system.

[0093] During the passage of a typhoon, the system is in a continuous monitoring cycle. The system repeatedly goes through the processes of sensing, fusing, decision-making, and execution, forming a rapid closed loop. If a sudden ocean phenomenon, such as internal ocean waves, is detected during this period, the system will increase the sampling frequency of the relevant sensors and adjust the course of the unmanned observation platform to track and observe it, in order to acquire high-frequency data of rare phenomena within a safe range.

[0094] As the typhoon gradually dissipates and sea conditions ease, the system enters a learning and optimization phase. The update module updates the reinforcement learning decision engine based on the adjusted execution results and operating parameters. The system evaluates the performance and data value of the strategies adopted throughout the extreme event, storing these experiences in an experience replay buffer and fine-tuning the policy network parameters online. In this way, the system learns from its response to this typhoon, enabling it to perform better in similar sea conditions the next time.

[0095] Ultimately, the system resumes its high-performance observation mode, generating a complete and high-quality dataset of this typhoon event, and awaits the next event. As can be seen from the above application scenarios, the system of this application demonstrates a resilient, progressively degraded, and intelligently recoverable adaptive capability when facing extreme sea conditions, truly achieving the design goals of being able to withstand, accurately measure, and learn in extreme environments.

[0096] Example 10: This embodiment, based on Embodiments 6 and 7, further explains the network transmission layer, cloud platform application layer, system hardware configuration, and software implementation of the extreme sea state adaptive observation system.

[0097] Combination Figure 2As shown, the system adopts a four-level collaborative architecture, including a perception layer, an edge intelligence layer, a network transmission layer, and a cloud platform application layer. The network transmission layer employs a multi-mode communication architecture, including four communication modes: satellite communication, 4G / 5G mobile communication, microwave relay, and underwater acoustic communication. Under extreme sea conditions, the availability of different communication modes varies dynamically. For example, satellite communication may experience signal attenuation due to atmospheric disturbances during typhoons, 4G / 5G mobile communication may be interrupted due to limited coverage of coastal base stations, and microwave relay may experience beam deviation due to antenna swaying. The system uses a communication quality monitoring module to evaluate the signal-to-noise ratio and available bandwidth of each communication link in real time, automatically selecting the optimal communication link for data transmission, or performing load balancing when multiple links are available. The network transmission layer is responsible for uplinking the fused data and status reports from the edge intelligence layer to the cloud platform, and downlinking the optimization strategies and control commands issued by the cloud platform to the edge layer and execution units, forming a bidirectional interaction between data flow and control flow.

[0098] The cloud platform application layer, centered on an autonomous observation and management platform, provides functions such as data storage and management, global status visualization, early warning information dissemination, and remote control interfaces. Specifically, the cloud platform receives fused data and status reports from multiple edge nodes, centrally stores them, and displays the global situation for users to monitor globally. Simultaneously, the cloud platform possesses powerful computing capabilities, enabling large-scale simulation training and model optimization. It then downlinks the updated policy model to the edge-side reinforcement learning decision engine, achieving continuous system evolution through cloud-edge collaboration. In extreme sea conditions, when the edge-side reinforcement learning decision engine experiences policy performance degradation due to limited computing resources or sudden environmental changes, the cloud platform can perform offline training based on global historical data, generating a superior policy model and deploying it downlink, thus compensating for the limitations of online learning at the edge.

[0099] In terms of system hardware configuration, the system hardware platform adopts a modular design. The main control buoy platform is a large buoy with a diameter of 3 meters, made of stainless steel, equipped with solar panels and a backup wind turbine, providing a continuous power output of no less than 2000W. The buoy integrates data acquisition devices, communication equipment, and edge computing nodes. Auxiliary observation nodes include 3 to 5 small unmanned surface vessels (USVs) and several diving buoys, forming a distributed observation network. The USVs are approximately 2 meters long, adopt a catamaran design to enhance stability, and are equipped with electric propulsion, achieving a speed of up to 5 knots. The above hardware configuration provides physical assurance for the system's continuous operation in extreme sea conditions. The combined power supply scheme of solar panels and wind turbines ensures energy redundancy, and the catamaran USV design improves the platform's anti-capsulation capability in large waves. In terms of software implementation, the system software adopts a microservice architecture, mainly including the following modules: a data acquisition service, responsible for real-time acquisition and caching of multimodal sensor data; a preprocessing service, implementing data quality control, outlier detection, and missing data imputation; a fusion computing service, executing cross-modal data fusion algorithms to generate comprehensive observation products; a reinforcement learning decision service, running the PPO-RMSprop hybrid algorithm to generate optimal observation strategies; and a visualization service, providing a web interface to display real-time observation data and analysis results. The microservice architecture allows each functional module to be deployed and upgraded independently. For example, when the fusion algorithm is updated, only the fusion computing service needs to be restarted without affecting the operation of other modules, improving the system's maintainability and robustness.

[0100] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic optimization method for adaptive observation in extreme sea states, characterized in that, include: Acquire multi-source observation data, perform feature fusion on the multi-source observation data, and generate environmental state information; The environmental state information is input into a reinforcement learning model to generate an observation and control strategy. Based on the aforementioned observation and control strategy, the operating parameters of the observation system are dynamically adjusted. The reinforcement learning model is updated based on the execution results after the operating parameters are adjusted.

2. The method according to claim 1, characterized in that, The multi-source observation data is fused to generate environmental state information, including: The multi-source observation data is fused using an attention mechanism with feature weighting to dynamically evaluate the confidence weight of each data source and generate the environmental state information.

3. The method according to claim 2, characterized in that, The feature-weighted fusion of the multi-source observation data using an attention mechanism includes: For optical wave sensor data, weights are calculated based on image sharpness and jitter level; For microwave surface flow sensor data, weights are calculated based on signal-to-noise ratio and spectral quality. For data from intelligent temperature, salinity, and depth instruments, weights are calculated based on self-correcting historical errors and data stability.

4. The method according to claim 1, characterized in that, Based on the execution results after adjusting the aforementioned operating parameters, the reinforcement learning model is updated, including: Based on the multi-objective reward function, the reward value is calculated according to the execution result; The parameters of the reinforcement learning model are updated based on the reward value.

5. The method according to claim 4, characterized in that, The multi-objective reward function includes a data quality reward, an energy consumption penalty, a security reward, and an exploration reward. The data quality reward is calculated based on data integrity and signal-to-noise ratio. The energy consumption penalty is calculated based on the total power consumption of the system. The safety reward items are calculated based on equipment status and environmental risk. The exploration reward is used to provide positive incentives when exploring unknown sea conditions within a safe range.

6. The method according to claim 1, characterized in that, The reinforcement learning model uses a proximal policy optimization algorithm for policy updates.

7. The method according to claim 1, characterized in that, The dynamic adjustment of the operating parameters of the observation system includes: When the detected wave height exceeds the preset threshold, the sampling frequency of the optical wave sensor is reduced and the sampling frequency of the microwave surface flow sensor is increased. When the remaining energy of the system is lower than a preset threshold, the auxiliary observation task is suspended to ensure the operation of the core sensors; When a sudden ocean phenomenon is detected, the sampling frequency of the relevant sensors is increased, and the course of the unmanned observation platform is adjusted for tracking and observation.

8. An adaptive observation system for extreme sea states, characterized in that, include: The cross-modal fusion module is configured to acquire multi-source observation data, perform feature fusion on the multi-source observation data, and generate environmental state information; A reinforcement learning decision engine is connected to the cross-modal fusion module and configured to receive the environmental state information and generate an observation and control strategy. The execution module is connected to the reinforcement learning decision engine and is configured to dynamically adjust the operating parameters of the observation system based on the observation control strategy. An update module, connected to the execution module, is configured to update the reinforcement learning decision engine based on the execution results adjusted according to the running parameters.

9. The system according to claim 8, characterized in that, The system further includes a multimodal sensing array, which comprises: An optical wave sensor configured to measure wave height, wave direction, and wave period; A microwave surface flow sensor configured to retrieve sea surface flow field distribution; The intelligent temperature, salinity, and depth meter is configured to collect temperature, salinity, and depth data. The underwater acoustic communication module is configured to enable data transmission between underwater sensors and the surface platform. The meteorological parameter acquisition unit is configured to measure meteorological elements such as wind speed, wind direction, temperature, humidity, and air pressure.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images