Anti-sway predictive control method for offshore cranes

Abstract

The application discloses an offshore crane anti-swing predictive control method, and belongs to the technical field of offshore engineering equipment control. The method aims to solve the problem of poor load swing suppression effect of the offshore crane caused by control lag and inaccurate model in the prior art. The scheme points are as follows: a state sensor group and a multi-source environment perception system are used to acquire the crane state and the predicted information of future waves in real time; a hybrid prediction model is used to predict the crane system state sequence under different control instructions in the future time domain through rolling prediction; based on the predicted sequence, a reference control trajectory is solved in the upper optimization, the reference control trajectory aims to suppress swing and reduce structural fatigue, an instant control instruction is solved in the lower fast optimization, and the instant control instruction meets the actuator constraint and is output; meanwhile, the system health management is independently executed, and the control mode is dynamically adjusted according to the evaluation. The method is mainly used for precise anti-swing control of the offshore crane under complex sea conditions, can effectively suppress the load swing in advance, and improves the operation accuracy and equipment safety.

Description

Technical Field

[0001] This invention belongs to the field of marine engineering equipment control technology, specifically relating to a method for predictive control of anti-swaying of offshore cranes. Background Technology

[0002] In the field of marine engineering, the operational stability of offshore floating cranes has always been a major concern. These devices face significant technical challenges in wave environments, with the core issues stemming from the uncertainty of the marine environment and the complexity of the system itself.

[0003] Periodic disturbances caused by ocean waves to operating vessels are transmitted to the crane system through the hull structure, causing boom vibration and load sway. Existing control methods mostly employ feedback control strategies, which implement control compensation after detecting load sway. This method has an inherent response delay, with the control system always lagging behind the disturbance. When encountering continuously changing wave sequences, the controller struggles to achieve precise phase compensation, resulting in limited suppression effectiveness and the residual sway amplitude of the load often exceeding the ideal range.

[0004] From a system modeling perspective, accurately describing the dynamic characteristics of cranes is quite challenging. Crane systems involve complex characteristics such as multibody dynamics coupling, structural flexible deformation, and hydraulic drive nonlinearity. Establishing a high-precision physical model requires balancing computational efficiency and real-time performance, which is difficult to achieve simultaneously in practical engineering. A common approach is to simplify the model as necessary, but this simplification loses some of the system's dynamic characteristics, causing discrepancies between model predictions and actual responses. Controllers designed based on incomplete models face performance bottlenecks.

[0005] Another noteworthy challenge lies in the effective utilization of environmental information. Wave disturbances have a temporal continuity; theoretically, obtaining future disturbance information in advance could significantly improve control performance. However, achieving this goal faces numerous difficulties: wave information acquired by different sensors varies in timing and accuracy; the morphological evolution during wave propagation increases prediction uncertainty; and the fusion of multi-source asynchronous data requires addressing issues such as time registration and confidence assessment. Previous studies, relying solely on historical ship motion data for extrapolation prediction, have yielded prediction times and accuracy insufficient for high-precision control requirements.

[0006] These problems limit the operational efficiency of offshore cranes in harsh sea conditions, especially in lifting operations requiring precise positioning. The continuous swaying of the load not only affects operational efficiency but may also pose safety hazards. Overcoming the limitations of existing control methods has become a key technical challenge that needs to be addressed in this field. Summary of the Invention

[0007] This invention provides a method for predictive control of anti-swaying of offshore cranes, which can effectively suppress load swaying in advance based on accurate prediction and feedforward control of future waves; improve the accuracy of system state prediction through a hybrid model that combines physical mechanisms and data-driven approaches; and adopt a two-layer optimization structure to ensure tracking performance while taking into account structural fatigue damage and actuator constraints, ultimately achieving high-precision and robust anti-swaying control under complex sea conditions.

[0008] To achieve these objectives and other advantages of the present invention, the present invention provides a method for predictive control of anti-swaying of an offshore crane, comprising the following steps: S1. The current state of the crane is acquired in real time through a group of state sensors installed on the crane system. The current state of the crane includes the spatial swing angle of the load, the pitch angle and rotation angle of the boom, and the real-time length of the wire rope; and the prediction information of future waves is acquired through a multi-source environmental sensing system. S2. Input the current state of the crane and the predicted information of the future waves into a pre-trained hybrid prediction model. The hybrid prediction model is constructed based on the parameterized physical dynamics model of the crane system and supplemented by a data-driven residual correction network to predict the state sequence of the crane system under different control commands in the future time domain. S3. Taking the current state of the crane as the initial condition, based on the predicted state sequence output by the hybrid prediction model, solve the first optimization problem within the first control cycle to obtain a reference control trajectory in the future time domain. The objective function of the first optimization problem includes the weighted sum of load swing kinetic energy and structural fatigue damage. S4. In a second control cycle that is shorter than the first control cycle, based on the reference control trajectory, solve a second optimization problem to obtain the real-time control command at the current moment; the objective function of the second optimization problem is to minimize the error of tracking the reference control trajectory and satisfy the operation constraints of the electro-hydraulic servo system. S5. Convert the real-time control command into a drive signal and output it to the electro-hydraulic servo actuator of the crane; S6. Simultaneously, independently execute the system health management process, evaluate the health status of the multi-source environmental perception system, the hybrid prediction model, and each sensor in real time, and dynamically adjust the control modes from S1 to S5 based on the evaluation results.

[0009] Preferably, in step S1, the multi-source environmental sensing system includes: Shipborne coherent lidar acquires the sea surface wave height sequence and wave propagation direction in real time based on the principle of coherent detection; X-band marine radar extracts wave spectrum features, including significant wave height, average wave period, and wave direction, through a sea surface echo inversion algorithm. A ship motion prospective observer based on ship inertial measurement unit data predicts ship motion over a future time domain by integrating current and historical ship motion attitude data based on ship motion equations. The multi-source environmental perception system uses a data fusion algorithm to fuse environmental and motion prediction information from different sources. The data fusion algorithm employs extended Kalman filtering or unscented Kalman filtering, uses the wave surface height sequence measured by lidar as the observation update, uses the wave spectrum features extracted by marine radar as the prior distribution, and combines the output of the ship motion prospective observer to generate prediction information of future waves through a multi-source asynchronous data fusion framework. The prediction information of future waves includes wave excitation force sequence, spatiotemporal distribution of wave surface disturbance, and ship motion response.

[0010] Preferably, in step S2, the construction and training method of the hybrid prediction model specifically includes: S210. Establish a parameterized physical dynamics model of the crane system as a basic predictor; S220. Collect historical operating data, which includes crane status sequence, control command sequence and corresponding environmental data; S230. Using the historical operating data, a residual correction network is trained through supervised learning. The training objective of the residual correction network is to predict the residual between the output of the parameterized physical dynamics model and the actual crane state. S240. The output of the parameterized physical dynamics model is added to the prediction output of the residual correction network to form the final prediction result of the hybrid prediction model.

[0011] Preferably, in step S3, the first optimization problem is solved using a simplified constraint set, which includes only the kinematic constraints of the boom joint space and the stress amplitude constraints of key parts.

[0012] Preferably, in step S4, the second optimization problem is solved by a model predictive controller or an exact feedback linearization controller, whose internal model is a simplified model characterizing the core nonlinearity of the electro-hydraulic servo system, and the operating constraints include valve core displacement limit, system pressure limit and actuator flow limit.

[0013] Preferably, in step S6, the system health management process specifically includes: S610: Real-time monitoring of the data quality and reliability of each sensing unit; S620. Evaluate the error between the prediction output of the hybrid prediction model and the measured state sequence. When the cumulative error exceeds a first preset threshold, determine that the health of the hybrid prediction model has decreased. S630. Based on the health status assessment results, dynamically and seamlessly switch between multiple control modes, wherein the control modes include at least: Full-function mode, all perception and prediction functions are enabled normally; Degraded sensing mode, which relies on the remaining effective sensing sources when either lidar or marine radar fails; In the safety fallback mode, when the entire environmental perception system fails or the health of the hybrid prediction model deteriorates significantly, the system switches to a robust feedback control method based solely on the data from the state sensor group. This robust feedback control method is either sliding mode variable structure control or fuzzy PID control.

[0014] Preferably, it also includes an online model update loop, which periodically uses actual operating data within a preset time window to incrementally learn and update the residual correction network in the hybrid prediction model, and the update process fixes the parameters of the parameterized physical dynamics model and only optimizes the weights of the residual correction network.

[0015] Preferably, in step S3, the estimation process of structural fatigue damage in the objective function of the first optimization problem specifically includes: S310. Based on the predicted state sequence of the crane system in the future time domain output by the hybrid prediction model, the dynamic equivalent stress time sequence of one or more key parts of the boom in the future time domain is calculated through a parameterized boom structure stress estimation model. S320. The stress estimation model of the boom structure is a parametric model pre-calibrated based on finite element analysis. Its inputs include at least the real-time pitch angle, slewing angle, wire rope tension, and load swing inertial force derived from the predicted state. The output is the equivalent stress value of key parts. S330. The dynamic equivalent stress time series is processed by rainflow counting method to identify complete stress cycles; S340. For each identified stress cycle, calculate the fatigue damage caused by that stress cycle based on its stress amplitude and the fatigue life model established based on the material SN curve. S350. Using Miner's linear cumulative damage rule, the fatigue damage caused by all stress cycles in the future prediction time domain is accumulated to estimate the structural fatigue damage forecast in the future time domain.

[0016] Preferably, in step S4, during the solution of the second optimization problem, an adaptive weight adjustment mechanism is introduced. This adaptive weight adjustment mechanism dynamically adjusts the weight coefficients of the tracking error term in the objective function based on real-time load weight, wire rope length, and environmental disturbance intensity. Specifically, the adaptive weight adjustment mechanism includes: S410. Real-time acquisition of load weight, wire rope length, and environmental disturbance intensity; the load weight is obtained through the weighing sensor of the crane system or based on the wire rope tension estimation model; the wire rope length is measured by an encoder or laser ranging sensor; the environmental disturbance intensity is extracted from the wave prediction information output by the multi-source environmental sensing system, including significant wave height, average wave period, and wave propagation direction, and is comprehensively calculated into a dimensionless disturbance intensity index; S 420. Input the load weight, wire rope length, and environmental disturbance intensity into a preset weight calculation function. This weight calculation function is a nonlinear mapping function obtained based on multi-parameter fitting or machine learning regression, and its output is the weight coefficient of the tracking error term. The form of the weight calculation function is as follows: , in, Here are the adjusted weighting coefficients, where m is the load weight, l is the wire rope length, d is the environmental disturbance intensity index, f1, f2, and f3 are the normalized influence functions for each parameter, and k1, k2, and k3 are adjustable gain coefficients. The benchmark weighting coefficient; It is an increasing function of the load weight; It is a decreasing function of the wire rope length; It is a non-monotonic function of the intensity of environmental disturbance; S430. The adaptive weight adjustment mechanism further includes an online update module for online optimization of the gain coefficients k1, k2, and k3. The optimization triggering condition is any of the following: a) reaching a preset fixed time interval; b) the accumulated number of control cycles reaches a preset threshold; c) the system performance evaluation index deteriorates to below a preset standard. The performance evaluation index is the weighted sum of the average kinetic energy of the load swing and the tracking error over a past period.

[0017] Preferably, it also includes a fail-safe loop that monitors the pressure and flow signals of the electro-hydraulic servo actuator in real time. When an actuator jamming or leakage fault is detected, it immediately switches to a redundant actuator or triggers an emergency braking procedure, and at the same time sends a fault alarm signal to the system health management process.

[0018] The present invention has at least the following beneficial effects: First, by integrating feedforward (wave prediction) and feedback (current state), combining long-cycle optimization with short-cycle command tracking, and introducing system health management, a multi-level, adaptive high-order control system was constructed. Wave prediction information enabled control that anticipates disturbances, reducing system response lag. Incorporating structural fatigue damage into the optimization objective achieved a balance between sway suppression and structural protection, facilitating long-term equipment operation. The health management module ensures the system maintains usable control capabilities even when some functions degrade, significantly improving the robustness and reliability of the entire system. By fusing sensing data from three different sources—LiDAR, marine radar, and ship motion observers—and comprehensively utilizing the high-precision wavefront sequences from LiDAR, the statistical wave spectrum characteristics from marine radar, and the motion trend predictions from ship motion observers, a data fusion algorithm was used to leverage the strengths of each, generating more comprehensive, accurate, and reliable predictions of future wave disturbances. By constructing a hybrid prediction model that combines a physical model and a data-driven model, the parameterized physical dynamics model ensures that the prediction results conform to basic physical laws, providing interpretability and extrapolation capabilities. Meanwhile, the data-driven residual correction network can learn and compensate for unmodeled dynamic and nonlinear characteristics neglected by the simplified physical model. The combination of these two approaches, compared to a single model, improves the prediction accuracy of future system state sequences, thus providing a more reliable basis for model-based predictive control.

[0019] Secondly, by employing a simplified constraint set in the first optimization problem, the computational complexity of the optimization problem is significantly reduced. While ensuring key kinematic safety and structural strength within limits, secondary constraints are discarded, enabling the optimizer to reliably solve for a feasible reference control trajectory within a finite computation time (i.e., within the first control cycle). This ensures the real-time performance of the upper-level planning layer and avoids control interruptions or performance degradation due to computation timeouts. By configuring a solver and simplified model specifically for the electro-hydraulic servo system for the second optimization problem, a precise and smooth transition from high-level instructions to low-level execution is achieved. This design fully considers the dynamic response characteristics and physical operational limits of the electro-hydraulic servo system, ensuring that the generated real-time control instructions are not only optimal in performance but also practically executable. This avoids actuator saturation and system pressure surges caused by overly aggressive instructions, guaranteeing control stability and system safety.

[0020] Third, by implementing a system health management process, the control system is endowed with the ability to self-aware of its state and reconfigure its modes. By continuously monitoring sensor data and model accuracy, the system can proactively identify performance degradation or faults and seamlessly switch to a control mode that matches its capabilities. This capability avoids the risk of a single point of failure causing the entire system to fail, and enhances its survivability and mission completion capabilities during long-term operations in complex and uncertain marine environments.

[0021] Fourth, by introducing an online model update loop, the hybrid predictive model acquires the ability to continuously learn and adapt to the time-varying characteristics of the system. By periodically updating the residual correction network with the latest operational data, model deviations caused by factors such as equipment aging and environmental changes can be continuously corrected, enabling the predictive model to always track the dynamics of the actual system. This ensures the long-term performance stability of the control system and reduces the reliance on periodic manual recalibration.

[0022] Fifth, by integrating fatigue damage estimation based on rainflow counting and cumulative damage theory into the optimization objective, the optimization of the control strategy is elevated from a simple "instantaneous performance" level to a "long-term lifespan" level. When making decisions, the controller proactively avoids control actions that, while quickly suppressing sway, would cause significant fatigue damage to the boom structure. This makes the control strategy more intelligent and forward-looking, helping to extend the service life of critical structural components of the crane and reduce total life-cycle maintenance costs.

[0023] Sixth, by introducing an adaptive weight adjustment mechanism, the control system can automatically adjust its control parameters according to real-time operating conditions (load, rope length, wind and waves). This allows the system to maintain excellent control performance under different operating scenarios: for example, it prioritizes stability under heavy loads or long rope conditions, and balances tracking accuracy and system robustness under strong disturbances. This adaptive capability improves the adaptability to complex and changing operating environments and the consistency of control effects.

[0024] Seventh, by setting up an independent fail-safe loop, a fast and reliable protection mechanism is provided for the actuator, a critical end-point component. Real-time monitoring of signals such as pressure and flow can sensitively detect typical faults such as jamming and leakage, and immediately trigger predetermined safety strategies (switching or braking). This sets up a final safety barrier for the entire control system, effectively preventing serious consequences such as load runaway caused by actuator failure, and greatly improving the overall safety of the system.

[0025] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the anti-sway prediction and control method for offshore cranes according to the present invention. Detailed Implementation

[0027] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0028] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0029] like Figure 1 As shown, this embodiment of the invention provides a method for predictive control of anti-sway of an offshore crane, comprising the following steps: S1. The current state of the crane is acquired in real time through a group of state sensors installed on the crane system. The current state of the crane includes the spatial swing angle of the load, the pitch angle and rotation angle of the boom, and the real-time length of the wire rope; and the prediction information of future waves is acquired through a multi-source environmental sensing system. S2. Input the current state of the crane and the predicted information of the future waves into a pre-trained hybrid prediction model. The hybrid prediction model is constructed based on the parameterized physical dynamics model of the crane system and supplemented by a data-driven residual correction network to predict the state sequence of the crane system under different control commands in the future time domain. S3. Taking the current state of the crane as the initial condition, based on the predicted state sequence output by the hybrid prediction model, solve the first optimization problem within the first control cycle to obtain a reference control trajectory in the future time domain. The objective function of the first optimization problem includes the weighted sum of load swing kinetic energy and structural fatigue damage. S4. In a second control cycle that is shorter than the first control cycle, based on the reference control trajectory, solve a second optimization problem to obtain the real-time control command at the current moment; the objective function of the second optimization problem is to minimize the error of tracking the reference control trajectory and satisfy the operation constraints of the electro-hydraulic servo system. S5. Convert the real-time control command into a drive signal and output it to the electro-hydraulic servo actuator of the crane; S6. Simultaneously, independently execute the system health management process, evaluate the health status of the multi-source environmental perception system, the hybrid prediction model, and each sensor in real time, and dynamically adjust the control modes from S1 to S5 based on the evaluation results.

[0030] In the above embodiments, the method integrates multi-source environmental perception, hybrid model prediction, two-layer optimization decision-making and system health management to achieve active suppression and safe control of the load sway of the offshore crane.

[0031] In the initial steps, the current operating status of the crane is acquired in real time through a set of status sensors installed on the crane system. This set of sensors includes tilt sensors or a visual recognition system for measuring the swing angle of the load space, encoders or inertial measurement units for detecting the boom pitch and slewing angles, and real-time wire rope length information acquired through encoders or laser rangefinders. Simultaneously, a multi-source environmental sensing system begins operation. This system may include devices such as lidar, marine radar, and ship motion observers. These devices work together to generate predictions of future waves over a given period by analyzing sea surface conditions and ship motion trends.

[0032] After acquiring the current state of the crane and the predicted future wave information, these are input into a pre-trained hybrid prediction model. This hybrid prediction model combines a parametric physical dynamics model with a data-driven residual correction network. The parametric physical dynamics model, based on Newtonian mechanics or Lagrange equations, describes the basic dynamic behavior of the crane system under ideal conditions. The residual correction network, trained using historical operating data and employing data-driven methods such as neural networks, is specifically designed to predict the deviation between the output of the parametric physical dynamics model and the actual system response. This residual correction network can be implemented using a fully connected neural network or a long short-term memory network with 1-2 hidden layers. It learns the systematic errors of the physical model through offline training and performs only forward propagation during online operation, keeping the computational load controllable. Working together, the parametric physical dynamics model provides basic predictions consistent with physical laws, while the residual correction network compensates for unmodeled dynamic and nonlinear characteristics. Together, they predict the crane system state sequence under different control commands in the future time domain, providing a reliable basis for subsequent optimization decisions.

[0033] Based on the predicted state sequence output by the hybrid prediction model, the control system enters the upper-level optimization stage. This stage uses the crane's current state as the initial condition and solves the first optimization problem within a relatively long control cycle. The objective function of this first optimization problem is specially designed, not only including the traditional load swing kinetic energy term but also innovatively introducing the predicted value of structural fatigue damage. The two are weighted and summed to form a comprehensive optimization objective. During the solution process, to improve real-time performance, a simplified constraint set can be used, for example, considering only the kinematic constraints of the boom joint space and the stress amplitude constraints of key components. By solving this optimization problem, the system obtains a reference control trajectory in the future time domain, which achieves a good balance between suppressing swing and protecting structural lifespan.

[0034] Within a shorter second control cycle, the system enters the next optimization phase. This phase solves the second optimization problem based on the reference control trajectory generated in the previous phase. The objective function of this second optimization problem aims to minimize the tracking error between the actual control command and the reference trajectory, while strictly satisfying the operational constraints of the electro-hydraulic servo system. These constraints include valve spool displacement limits, system pressure limits, and actuator flow limits. This second optimization problem can be solved by a model predictive controller or a precise feedback linearized controller, which internally employs a simplified model characterizing the core nonlinearities of the electro-hydraulic servo system to ensure that the generated real-time control commands meet performance requirements while remaining within the physical capabilities of the actual actuator.

[0035] Upon receiving real-time control commands, the system converts them into specific drive signals and outputs them to the crane's electro-hydraulic servo actuators. This conversion process involves signal amplification, format conversion, and interface adaptation to ensure that the control commands can accurately drive actuators such as hydraulic cylinders or hydraulic motors, ultimately achieving precise control of the crane's movement.

[0036] Running parallel to the aforementioned control process is the system health management process. This process continuously monitors the operational status of the multi-source environmental sensing system, the hybrid prediction model, and various sensors. By evaluating indicators such as data quality and prediction error, it determines the health of each component. When a performance degradation of certain components is detected, the system can dynamically adjust the control mode, such as switching from full-function mode to degraded sensing mode or safety fallback mode, ensuring that the system can still maintain basic control capabilities and guarantee operational safety even when some functions fail.

[0037] This technical solution, through the coordinated efforts of the aforementioned multiple stages, achieves a significant improvement in anti-sway performance compared to traditional feedback control methods. Its feedforward control, utilizing wave prediction information, effectively overcomes the hysteresis problem of simple feedback control, significantly reducing the residual sway amplitude of the load. Simultaneously, by incorporating structural fatigue damage into the optimization objective, the control system considers the long-term service life of the equipment during sway suppression—a feature not addressed in traditional methods. The dual-layer optimization architecture ensures both optimal long-term performance and short-term feasibility and safety. Furthermore, the introduction of system health management greatly enhances the adaptability and reliability of the control system in complex marine environments, enabling it to maintain usable control capabilities even in the face of component failures or performance degradation.

[0038] In one specific embodiment, in step S1, the multi-source environmental sensing system includes: Shipborne coherent lidar acquires the sea surface wave height sequence and wave propagation direction in real time based on the principle of coherent detection; X-band marine radar extracts wave spectrum features, including significant wave height, average wave period, and wave direction, through a sea surface echo inversion algorithm. A ship motion prospective observer based on ship inertial measurement unit data predicts ship motion over a future time domain by integrating current and historical ship motion attitude data based on ship motion equations. The multi-source environmental perception system uses a data fusion algorithm to fuse environmental and motion prediction information from different sources. The data fusion algorithm employs extended Kalman filtering or unscented Kalman filtering, uses the wave surface height sequence measured by lidar as the observation update, uses the wave spectrum features extracted by marine radar as the prior distribution, and combines the output of the ship motion prospective observer to generate prediction information of future waves through a multi-source asynchronous data fusion framework. The prediction information of future waves includes wave excitation force sequence, spatiotemporal distribution of wave surface disturbance, and ship motion response.

[0039] In the above embodiments, shipborne coherent lidar is a high-precision sensor based on the principle of coherent detection. It acquires the wave surface height sequence and wave propagation direction in real time by emitting a laser beam and analyzing the phase changes of the reflected echoes from the sea surface. This type of shipborne coherent lidar typically has high measurement accuracy; for example, the wave surface height measurement error can be controlled at the centimeter level, and the wave direction measurement accuracy is within a few degrees, providing the system with detailed spatiotemporal wave surface data. X-band marine radar uses X-band electromagnetic waves to illuminate the sea surface and extracts wave spectrum features, including significant wave height, average wave period, and wave direction, through a sea surface echo inversion algorithm. The significant wave height measurement range is between 0.5 meters and 5 meters, and the average wave period is between 3 seconds and 15 seconds. These statistical characteristics serve as a macroscopic description of the wave environment, supplementing the detailed information of the shipborne coherent lidar. A ship motion look-ahead observer based on ship inertial measurement unit (IMU) data predicts the ship's motion over a future time domain by fusing current and historical ship motion attitude data provided by IMUs (such as accelerometers and gyroscopes) and combining them with ship motion equations (such as rigid body dynamics models). The prediction time domain can be set according to actual needs, such as 5 to 30 seconds, depending on the ship size and sea conditions. These three components capture environmental information from different dimensions: the shipborne coherent lidar provides high-resolution wavefront dynamics, the X-band ocean radar provides wave statistical characteristics, and the hull motion observer predicts the ship's own response. Together, they form a multi-layered sensing network to ensure that the system can fully cover the direct and indirect effects of wave disturbances.

[0040] The multi-source environmental perception system employs either Extended Kalman Filtering (EPF) or Unscented Kalman Filtering (UKF) as the core data fusion algorithm. Both methods effectively handle the uncertainties of nonlinear systems. EPF estimates the state through linearization approximation, while UKF uses sampling points to more accurately capture nonlinear distributions, making it suitable for the complex dynamics of the marine environment. During the fusion process, the wavefront height sequence measured in real-time by the shipborne coherent lidar serves as the observation update input. Due to its high frequency and high precision characteristics (e.g., update rates exceeding 10 Hz), it can promptly correct prediction biases. Wave spectrum features extracted by X-band marine radar (such as significant wave height and wave direction) serve as the prior distribution, providing the statistical basis for the wave environment. This data can be updated every 1 to 5 seconds, ensuring the stability of the fusion results. The output of the ship motion look-ahead observer supplements the system state prediction, generating the ship's future trajectory by integrating ship motion equations (such as a six-degree-of-freedom model considering hydrodynamics). The multi-source asynchronous data fusion framework is responsible for coordinating the temporal differences between different sensors. For example, it handles millisecond to second-level data delays that may exist from shipborne coherent lidar, X-band marine radar, and inertial measurement units through timestamp alignment and buffer management. In actual operation, the fusion algorithm is executed periodically, for example, updating the prediction every 50 to 200 milliseconds. Wave prediction results are continuously optimized through iterative correction, ensuring that the generated information is both real-time and reliable. This data fusion framework addresses the asynchronous problem by assigning a unified timestamp to all sensor data and using interpolation algorithms to align them to the same series of fusion time points. Simultaneously, it comprehensively evaluates the real-time confidence of each data source by analyzing the quality of each sensor signal in real time, cross-referencing the consistency of data from different sources, and verifying the deviation between observed and predicted values. Based on this confidence, it dynamically adjusts their weights in the fusion result, thus outputting robust environmental prediction information even when some sensor data is unreliable.

[0041] The predicted information includes wave excitation force sequences, spatiotemporal distribution of wave surface disturbances, and ship motion response. The wave excitation force sequence describes the changes in the force exerted by waves on the ship in the future time domain, which can be represented as a force vector with a frequency range between 0.05 Hz and 0.5 Hz, corresponding to typical ocean wave cycles. The spatiotemporal distribution of wave surface disturbances represents the evolution of wave height in space and time using gridded data; for example, the spatial resolution can be set to meters, and the time step to seconds, helping the system understand wave propagation patterns. The ship motion response predicts the ship's roll, pitch, and other attitude angles, with amplitudes ranging from a few degrees to tens of degrees, depending on the sea state intensity. This predicted information is input in real time into a hybrid prediction model to predict the state sequence of the crane system in the future time domain. During implementation, a multi-source environmental perception system operates continuously: the shipborne coherent lidar generates point cloud data by scanning the sea surface, and outputs a wave surface sequence after filtering and interpolation; the X-band ocean radar inverts wave spectrum parameters through a signal processing chain; and the ship motion observer solves the motion equations using numerical integration methods. The predictive information is updated once per control cycle, for example, every 0.1 to 1 second, to ensure that the control system can respond to disturbances in advance and achieve feedforward compensation.

[0042] This implementation method, through the integrated design of a multi-source environmental sensing system, can significantly improve the accuracy and robustness of wave prediction. The technical benefits are as follows: by deeply fusing data from multiple sensors, the system effectively reduces prediction errors caused by environmental uncertainties, enabling anti-sway control to respond to wave disturbances earlier, thereby significantly reducing load sway amplitude. Simultaneously, the system's adaptive characteristics enhance reliability in complex sea conditions; even if the performance of some sensors degrades, usable predictive capabilities can still be maintained through data fusion, thus improving the overall stability and safety of crane operations.

[0043] In one specific implementation, the construction and training method of the hybrid prediction model in step S2 specifically includes: S210. Establish a parameterized physical dynamics model of the crane system as a basic predictor; S220. Collect historical operating data, which includes crane status sequence, control command sequence and corresponding environmental data; S230. Using the historical operating data, a residual correction network is trained through supervised learning. The training objective of the residual correction network is to predict the residual between the output of the parameterized physical dynamics model and the actual crane state. S240. The output of the parameterized physical dynamics model is added to the prediction output of the residual correction network to form the final prediction result of the hybrid prediction model.

[0044] In the above implementation, the construction of the hybrid prediction model first requires establishing a parametric physical-dynamic model of the crane system as the basic predictor. This parametric physical-dynamic model, based on physical principles such as Newtonian mechanics or Lagrange's equations, describes the dynamic behavior of the crane system under ideal conditions, including the pitch and rotation motion of the boom, the swing dynamics of the load, and the extension and contraction effects of the wire rope. Parametric design means that key parameters in the model, such as the mass, length, moment of inertia of the boom, and the stiffness of the wire rope, can be calibrated and adjusted according to the actual crane's design specifications. For example, the boom length parameter can be selected within the range of 20 meters to 100 meters, depending on the crane model and operational requirements; the load mass range can be from several tons to hundreds of tons to cover different lifting scenarios. During implementation, engineers obtain these parameter values ​​by measuring equipment geometry, consulting technical manuals, or conducting on-site tests, and then substitute them into the dynamic equations to construct a mathematical model capable of simulating the basic motion of the crane. This model calculates the system state sequence over a future period based on current state inputs (such as boom angle and load position) and control commands (such as hydraulic cylinder displacement commands) using numerical integration methods (such as the Runge-Kutta method). Examples include the load swing angle and boom attitude over the next 5 to 30 seconds. Because the parametric physical dynamics model strictly follows the laws of mechanics, its predictions have good interpretability and extrapolation capabilities. However, it may contain systematic errors due to simplifying assumptions (such as neglecting structural flexibility or friction effects).

[0045] Next, the system trains a residual correction network by collecting historical operational data to compensate for prediction biases in the parametric physical dynamics model. Historical operational data includes crane state sequences, control command sequences, and corresponding environmental data, typically obtained from sensor networks and control system logs. State sequences can include the load's spatial swing angle, boom pitch and slewing angles, etc., with sampling frequencies ranging from 10 Hz to 100 Hz to ensure temporal resolution. Control command sequences record past signals sent to the electro-hydraulic servo actuators, such as valve opening or motor speed commands. Environmental data comes from a multi-source environmental sensing system, including wave height, period, and propagation direction. During the data acquisition phase, the system continuously records operational data for several days or weeks, forming a large-scale dataset covering different sea states and load conditions, potentially reaching thousands of hours. This data is then used to train the residual correction network using supervised learning. This network typically employs deep learning structures, such as multilayer perceptrons or long short-term memory networks. Its inputs include the physical model's predicted output, the current state, and control commands; the output is the residual between the physical model's prediction and the actual state. During training, the data is randomly divided into training and validation sets. Optimization algorithms such as stochastic gradient descent are used to minimize the mean squared error of the residual predictions. The training period can last from several hours to several days, depending on the data size and network complexity. By learning historical error patterns, the residual correction network can capture nonlinear characteristics not covered by the physical model, external disturbances (such as the effects of wind and waves), or time-varying behavior of the system (such as changes in hydraulic oil temperature), thereby significantly improving prediction accuracy.

[0046] Finally, the output of the parametric physical dynamics model is added to the prediction output of the residual correction network to form the final prediction result of the hybrid prediction model. During operation, for each control cycle, the parametric physical dynamics model first generates a basic prediction based on the current state and the assumed control command sequence, such as the load swing angle sequence with a time step of 0.1 seconds over the next 10 seconds. Simultaneously, the residual correction network calculates the corresponding residual sequence based on the same input. The two are then fused through element-wise addition to form the corrected state prediction. For example, if the parametric physical dynamics model predicts a load swing angle of 12 degrees at a certain moment, and the residual correction network outputs a residual of -2 degrees, the final prediction value is 10 degrees. This combination method maintains the mechanistic nature of the parametric physical dynamics model while introducing data-driven adaptability, enabling the prediction result to reflect the basic dynamics of the system while correcting deviations in actual operation. In implementation, the addition operation is computationally lightweight, typically completed within milliseconds, ensuring that the hybrid model can continuously update predictions within real-time control cycles (e.g., every 0.1 to 1 second). Hybrid predictive models were then used to optimize control decisions by continuously re-predicting based on the latest state, dynamically adapting to environmental changes and system disturbances.

[0047] This implementation significantly improves the accuracy and robustness of crane system state prediction by constructing a hybrid prediction model that organically combines a parametric physical dynamics model with a data-driven residual correction network. The parametric physical dynamics model provides stable and interpretable basic predictions, ensuring reasonable extrapolation capabilities under unseen operating conditions. Meanwhile, the residual correction network effectively reduces prediction bias caused by model simplification or unmodeled dynamics by learning and compensating for historical errors, particularly excelling in handling nonlinear responses and external disturbances. This hybrid architecture makes the prediction results closer to the actual system behavior, providing more reliable input for subsequent optimized control, thereby enhancing the overall performance of the anti-sway control system, significantly reducing load sway amplitude, and improving operational stability and safety. Furthermore, the modular design of the model facilitates maintenance and updates; the residual network can be continuously retrained as the system operates to adapt to equipment aging or environmental changes during long-term use.

[0048] In one specific implementation, in step S3, the first optimization problem is solved using a simplified constraint set, which includes only the kinematic constraints of the boom joint space and the stress amplitude constraints of key parts.

[0049] In the above implementation, the simplified constraint set means that only the constraints most critical to the safety and stability of the system are considered during the optimization process, while secondary or non-urgent constraints are ignored, thereby significantly reducing the complexity of the optimization problem. Specifically, the constraint set only includes two types: kinematic constraints of the boom joint space and stress amplitude constraints of key parts. The principle is based on the strict requirements of computation time in real-time control: the first control cycle is usually set in the range of several hundred milliseconds to several seconds (e.g., 0.5 seconds to 2 seconds). Introducing too many constraints will lead to excessively long solution time, affecting the real-time performance of control. Therefore, by carefully selecting these two types of constraints, the basic safety of the crane operation is ensured, and the optimizer can reliably converge within a finite time to generate a feasible reference control trajectory. In actual implementation, the optimization algorithm (such as gradient descent or interior point method) is executed periodically. At the beginning of each first control cycle, with the current crane state as the initial condition, the objective function (the weighted sum of load swing kinetic energy and structural fatigue damage) is minimized under the simplified constraint set based on the state sequence output by the hybrid prediction model, thereby outputting the reference control trajectory in the future time domain.

[0050] Kinematic constraints in the boom joint space primarily define safe operating ranges for motion parameters such as the boom's pitch and slewing angles to prevent mechanical interference or excessive movement. For example, the pitch angle can be limited to -10 degrees to +80 degrees, and the slewing angle to -180 degrees to +180 degrees; these values ​​can be adjusted based on the specific crane model and operating environment. The principle is based on the crane's mechanical structure and kinematic model, ensuring that the boom does not collide with the ship's structure or other equipment during movement, while also avoiding entering non-working areas. In optimization problems, these constraints exist in the form of inequalities, such as requiring the pitch angle to always remain between minimum and maximum limits. During implementation, the optimizer checks the boom attitude at each predicted step; if it exceeds the limits, the control commands are adjusted to bring the trajectory back to the feasible region. This use of constraints not only ensures equipment integrity but also indirectly reduces the possibility of load sway by limiting the range of motion.

[0051] Stress amplitude constraints at critical locations focus on the dynamic stress levels at key points of the boom structure (such as hinge points, welds, or high-stress areas) to ensure they do not exceed the material's allowable fatigue limit. For example, the stress amplitude can be limited to between 100 MPa and 200 MPa. This limit is determined based on the fatigue strength of the high-strength steel materials typically used in boom structures (e.g., marine high-strength steel with a yield strength of not less than 690 MPa, such as Q690D or AH36). The value is calculated using the material's SN curve and the allowable stress amplitude under the target design life, aiming to control the alternating stress at critical locations below the fatigue limit to ensure long-term structural safety. The principle is based on structural mechanics and fatigue analysis, preventing structural damage by estimating the equivalent stress (such as Von Mises stress) at critical locations in real time. In optimization, stress constraints are integrated through a parametric stress estimation model: this model takes predicted states such as boom attitude and load swing inertia as input, outputs stress values ​​at critical locations, and requires these values ​​to be below a preset threshold. During implementation, the optimizer evaluates the impact of each control command on the stress sequence. If the predicted stress exceeds the limit, the trajectory is adjusted first to reduce the load. This use of constraints incorporates structural protection into the control decision, avoiding long-term fatigue damage that may result from short-term performance optimization.

[0052] These technical features work together to achieve efficient and safe optimization solutions. Kinematic constraints ensure the feasibility and safety of the boom's movement, while stress constraints protect structural integrity; together, they form a concise yet comprehensive set of constraints. In this process, the optimizer first generates an initial trajectory under kinematic constraints, then refines it using stress constraints, ultimately outputting a reference control trajectory that suppresses swaying while also considering structural lifespan. This simplification strategy significantly reduces the computational burden, making reliable solutions possible within a finite first control cycle (e.g., within 1 second), thus supporting the real-time performance of the upper planning layers.

[0053] In one specific implementation, in step S4, the second optimization problem is solved by a model predictive controller or an exact feedback linearization controller, whose internal model is a simplified model characterizing the core nonlinearity of the electro-hydraulic servo system, and the operating constraints include valve core displacement limit, system pressure limit and actuator flow limit.

[0054] In the above implementation, in step S4, the solution process for the second optimization problem involves selecting an appropriate controller type, wherein a model predictive controller or an exact feedback linearization controller is employed. A model predictive controller is a control strategy based on rolling optimization and feedback correction. It uses an internal model to predict the system's behavior in the future finite time domain and generates control commands by solving a constrained optimization problem. Its principle lies in compensating for system dynamic changes and uncertainties through iterative online optimization. An exact feedback linearization controller, on the other hand, is a nonlinear control method that precisely transforms the original nonlinear system into a linear system through coordinate transformation and state feedback, thereby applying linear control theory for design. Its core principle is to eliminate the nonlinear characteristics of the system, simplifying controller design. In this step, these controllers are used to meet the real-time control requirements of electro-hydraulic servo systems. Within a short second control cycle (e.g., 20 to 100 milliseconds, depending on system hardware performance and real-time requirements), they rapidly solve for the immediate control command at the current moment to ensure tracking of the reference control trajectory generated by the upper layer. The second control cycle (20 to 100 milliseconds) is primarily set to match the rapid dynamic response characteristics of the electro-hydraulic servo actuator and the natural frequency of load oscillation (typically 0.1 Hz to 0.5 Hz). This cycle is much faster than the mechanical dynamics of the controlled object, sufficient to achieve accurate trajectory tracking, while this timescale is also within the real-time computing capabilities of modern embedded processors. In implementation, the model predictive controller periodically performs optimization calculations, for example, calling the solver every 50 milliseconds to minimize the tracking error based on the current system state and the predictive model output; while the precise feedback linearization controller directly calculates the control method output through real-time state measurement and transformation. The choice between these two controllers allows the system to switch flexibly according to actual operating conditions. For example, when the system is changing rapidly, the model predictive controller can be used first to handle constraints, while when computational resources are limited, the exact feedback linearization controller can be used to improve response speed. They work together to achieve efficient and stable instruction tracking.

[0055] The controller's internal model is a simplified model specifically designed to characterize the core nonlinearities of the electro-hydraulic servo system. This simplified model focuses on the system's main dynamic characteristics, such as the flow-pressure relationship of the valve-controlled hydraulic cylinder, the inertial effect of the actuator, and frictional nonlinearity, while ignoring some secondary factors to reduce computational complexity. The principle is that by preserving key nonlinear behaviors (e.g., the flow nonlinearity of the valve orifice or the force-velocity characteristics of the hydraulic cylinder), the simplified model accurately reflects the system's essence and is suitable for real-time optimization calculations. This simplified model is used as the basis for the controller to predict the response of the electro-hydraulic servo system under control commands, or for transformation calculations in feedback linearization. During implementation, the simplified model can be constructed based on physical equations (such as fluid dynamics and mechanics equations) and its parameters calibrated using experimental data. For example, the valve core displacement-flow relationship can be simplified to a piecewise linear function, or the system pressure dynamics can be approximated as a first-order inertial element. The simplification level of the simplified model can be referenced from the actual system. For example, in a typical offshore crane, the simplified model can consider only the nonlinearity of the main hydraulic actuators, while simplifying the pipeline dynamics to lumped parameters, with the update frequency synchronized with the control cycle (e.g., updating the state prediction every 50 milliseconds). This simplified model works closely with the controller, enabling rapid evaluation of the impact of different control commands when solving optimization problems, thus meeting real-time requirements while ensuring accuracy. For example, in model predictive controllers, it is used to generate prediction sequences; in precise feedback linearization, it is used to derive the linearization transformation.

[0056] Operational constraints include valve spool displacement limits, system pressure limits, and actuator flow limits. These constraints represent the physical operational boundaries of the electro-hydraulic servo system, ensuring that control commands are feasible and safe in actual equipment. Valve spool displacement limits refer to the mechanical range of valve spool movement, preventing excessive displacement that could cause hardware damage or performance degradation. System pressure limits relate to the maximum permissible pressure of the hydraulic circuit, avoiding overpressure that could lead to leaks or component failures. Actuator flow limits restrict the flow of hydraulic oil, ensuring that the actuator speed remains within a safe range. The principle is that by incorporating these physical constraints into the constraints of the optimization problem, the solver is forced to generate control commands within the equipment's capabilities. These operational constraints are used as hard or soft constraints in the secondary optimization problem. For example, in model predictive controllers, they can be defined in the form of inequalities, such as valve spool displacement not exceeding ±10 mm (the specific value varies depending on the equipment model; for example, some systems may set it to ±8 mm to ±12 mm), system pressure maintained between 0 and 25 MPa (refer to the actual hydraulic system design; pressure limits can be adjusted according to a safety factor), and actuator flow limited to within 100 liters per minute (flow limits can be selected based on pump capacity and pipe size, for example, within the range of 80 liters / minute to 120 liters / minute). During implementation, these constraints are checked and applied in each control cycle, for example, by using online optimization algorithms (such as the interior-point method or the active set method) to ensure that the solution satisfies all constraints. If a constraint conflict is detected, the system can adjust control commands or trigger a degradation strategy. These constraints work in conjunction with the controller and the internal model to ensure control smoothness and system reliability, preventing actuator saturation or overload.

[0057] In one specific implementation, step S6 specifically includes the following system health management process: S610: Real-time monitoring of the data quality and reliability of each sensing unit; S620. Evaluate the error between the prediction output of the hybrid prediction model and the measured state sequence. When the cumulative error exceeds a first preset threshold, determine that the health of the hybrid prediction model has decreased. S630. Based on the health status assessment results, dynamically and seamlessly switch between multiple control modes, wherein the control modes include at least: Full-function mode, all perception and prediction functions are enabled normally; Degraded sensing mode, which relies on the remaining effective sensing sources when either lidar or marine radar fails; In the safety fallback mode, when the entire environmental perception system fails or the health of the hybrid prediction model deteriorates significantly, the system switches to a robust feedback control method based solely on the data from the state sensor group. This robust feedback control method is either sliding mode variable structure control or fuzzy PID control.

[0058] In the above implementation, the system health management process first involves real-time monitoring of the data quality and reliability of each sensing unit. These sensing units include a group of status sensors installed on the crane system, such as tilt sensors or visual recognition systems to measure the load space swing angle, encoders or inertial measurement units to detect the boom pitch and slewing angles, and encoders or laser rangefinders to acquire the real-time length of the wire rope. Data quality monitoring involves checking the integrity, consistency, and stability of sensor outputs, for example, by analyzing the packet loss rate, signal noise level, or frequency of outliers. The noise level can be set within a range of 5% to 10% as a reference threshold, with the specific value adjusted according to the actual sensor performance. Reliability assessment is based on the physical rationality and historical behavior patterns of the sensor data, for example, by comparing readings from multiple redundant sensors or using statistical methods (such as Kalman filtering) to detect deviations. If the reliability score of a sensor's data is lower than a preset standard (e.g., the confidence interval is lower than the reference range of 80% to 90%), the system will mark the sensor as suspicious. The principle of this monitoring mechanism is to identify sensor failures or performance degradation early, preventing erroneous data from affecting subsequent control decisions. During implementation, the monitoring module operates at a fixed frequency, such as performing a comprehensive check every 100 to 500 milliseconds. The specific frequency depends on the system's real-time requirements, and timing differences are managed through a data buffer to ensure timely assessment. Through continuous monitoring, the system can promptly detect problems and trigger corresponding handling strategies, providing reliable input for subsequent health assessments.

[0059] Next, the system evaluates the error between the predicted output of the hybrid predictive model and the measured state sequence to determine the model's health. The hybrid predictive model combines a parametric physical dynamics model and a data-driven residual correction network to predict the future state sequence of the crane system, such as load swing angle and boom attitude. The evaluation process compares the model's predicted values ​​with the actual sensor measurements. Error calculation can use indicators such as root mean square error or mean absolute error, accumulating error values ​​over a period of time, such as the sum of errors over the past 10 to 30 seconds. The specific time window can be dynamically adjusted according to the control cycle. When the accumulated error exceeds a first preset threshold, the system determines that the health of the hybrid predictive model has declined. The first preset threshold can be set based on historical performance data, such as the average error under normal operation plus two standard deviations. The error range for the corresponding load swing angle can be between 2 and 5 degrees as a reference, but the actual value needs to be calibrated according to the crane model and operating environment. The judgment principle lies in the direct correlation between model prediction accuracy and control system performance: if the model error remains consistently high, it indicates that the hybrid predictive model cannot accurately capture system dynamics, which may lead to control command failure. During implementation, the error assessment module runs periodically, updating the cumulative error every 1 to 5 seconds and comparing it with a threshold. It also records error trends to distinguish between temporary fluctuations and persistent degradation. If the health status declines, the system triggers an alarm and prepares to switch control modes, ensuring the controller does not rely on inaccurate predictions, thereby maintaining overall control stability.

[0060] Finally, based on the health status assessment results, the system dynamically and seamlessly switches between multiple control modes. These control modes include full-function mode, degraded sensing mode, and safety fallback mode. Full-function mode is activated when all sensing and prediction functions are normal, fully utilizing the multi-source environmental sensing system and hybrid predictive model for advanced control to achieve optimal anti-sway performance. Degraded sensing mode is activated when some sensors fail, such as when one of the lidar or ocean radar fails. The system relies on the remaining effective sensing sources (e.g., using only ocean radar data or ship motion observer output) to generate wave prediction information, ensuring that the control still has some feedforward capability. In this mode, the data fusion algorithm adaptively adjusts weights, prioritizing the use of available sources. Safety fallback mode is triggered when the entire environmental sensing system fails or the health of the hybrid predictive model deteriorates significantly. At this time, the system switches to a robust feedback control method based solely on state sensor group data, such as sliding mode variable structure control or fuzzy PID control. These methods do not rely on predictive models but directly calculate control commands based on the current state feedback, exhibiting strong robustness to adapt to uncertainty. The switching process is designed to be seamless, meaning that the control output transitions smoothly during mode changes, avoiding abrupt changes that could lead to system instability. During implementation, the health management module monitors sensor and model status in real time. When a fault or performance degradation is detected, it immediately calculates the switching signal and applies the new mode within the next control cycle (e.g., within 100 milliseconds), while interpolation or buffering mechanisms ensure command continuity. This dynamic switching mechanism ensures that the control system maintains basic functionality under various fault conditions. Through collaborative work, various technical features achieve a smooth degradation from advanced predictive control to basic feedback control, guaranteeing operational safety and system resilience.

[0061] In one specific embodiment, an online model update loop is also included. This online model update loop periodically uses actual operating data within a preset time window to incrementally learn and update the residual correction network in the hybrid prediction model. The update process fixes the parameters of the parameterized physical dynamics model and only optimizes the weights of the residual correction network.

[0062] In the above implementation, the online model update loop is a crucial component of the offshore crane anti-sway predictive control method. It dynamically adjusts the hybrid predictive model periodically using actual operating data to ensure its continuous adaptation to system changes and environmental disturbances. The online model update loop operates on the principle of incremental learning, allowing the system to fine-tune the model using the latest operational data without interrupting the control process, thus maintaining predictive accuracy. During implementation, the online model update loop triggers updates at preset time intervals, such as every 24 hours or every 1000 control cycles. The specific time interval can be adjusted between 8 and 48 hours depending on the application requirements. This periodic update mechanism ensures that the hybrid predictive model can promptly capture slow, dynamic changes in the system, such as equipment aging, changes in hydraulic oil characteristics, or long-term evolution of environmental conditions. During operation, the update loop works in parallel with the control system. When a preset time or event trigger condition is met, the system automatically collects the actual operating data from the most recent period and initiates the hybrid predictive model update process without affecting the current anti-sway control operation, thereby achieving seamless integration and continuous optimization.

[0063] In the online model update loop, the incremental learning update process relies on actual operational data within a preset time window. This data includes crane state sequences, control command sequences, and corresponding environmental measurements, such as load swing angle, boom angle, and wave prediction information. The length of the time window can be selected based on the speed of system dynamic changes and the amount of data; for example, it can be set to the most recent 7 days or the most recent 10,000 data points. The window size can be adjusted from 1 to 30 days to adapt to different learning needs. In practice, the system continuously records operational data and, with each update, extracts data from the preset window as a training set from storage. This data is then used to train the residual correction network in a supervised learning manner, with the training objective being to minimize the difference between the network's predicted residuals and the actual observation errors. Incremental learning means that each update does not start training from scratch but rather further optimizes the network based on its current weights. This reduces computational burden and allows for rapid adaptation. For example, the update process may use stochastic gradient descent or its variants, completing within minutes to tens of minutes, depending on the data scale and hardware performance, thus ensuring that real-time performance is not affected.

[0064] During the hybrid prediction model update process, the system fixes the parameters of the parametric physical dynamics model and optimizes only the weights of the residual correction network. This design is based on a balance between the stability of the physical model and the adaptability of the data-driven network. The parametric physical dynamics model is constructed based on the mechanical principles of the crane system. Its parameters, such as mass, inertia, and geometry, are usually determined during system design and are relatively stable and do not easily change over time. Therefore, fixing these parameters can avoid unnecessary disturbances and overfitting risks. For example, parameters such as boom length and load mass may remain at their initial calibration values ​​and not change during the update process. Conversely, the residual correction network is responsible for compensating for nonlinear dynamics and external disturbances not covered by the physical model. These components may drift with changes in operating conditions and therefore need to be adjusted periodically. The principle is to maintain the basic predictive power of the physical model while capturing residual changes through the data-driven network, thereby achieving long-term model accuracy. In practice, the update process only calculates the weight gradient of the residual network and uses optimization algorithms such as Adam or RMSprop to update the weights, while the parameters of the physical model are treated as constants. This division of labor and cooperation enables the hybrid prediction model to adaptively correct prediction biases while maintaining physical interpretability. For example, when a crane operates under specific sea conditions for a long time, the residual correction network will learn the error patterns specific to that environment and adjust the output accordingly, thereby improving the overall prediction reliability.

[0065] In one specific embodiment, in step S3, the estimation process of structural fatigue damage in the objective function of the first optimization problem specifically includes: S310. Based on the predicted state sequence of the crane system in the future time domain output by the hybrid prediction model, the dynamic equivalent stress time sequence of one or more key parts of the boom in the future time domain is calculated through a parameterized boom structure stress estimation model. S320. The stress estimation model of the boom structure is a parametric model pre-calibrated based on finite element analysis. Its inputs include at least the real-time pitch angle, slewing angle, wire rope tension, and load swing inertial force derived from the predicted state. The output is the equivalent stress value of key parts. S330. The dynamic equivalent stress time series is processed by rainflow counting method to identify complete stress cycles; S340. For each identified stress cycle, calculate the fatigue damage caused by that stress cycle based on its stress amplitude and the fatigue life model established based on the material SN curve. S350. Using Miner's linear cumulative damage rule, the fatigue damage caused by all stress cycles in the future prediction time domain is accumulated to estimate the structural fatigue damage forecast in the future time domain.

[0066] In the above implementation, in step S3, the objective function of the first optimization problem includes a predicted value of structural fatigue damage. This estimation is achieved through a multi-step process, aiming to quantify the cumulative damage that the boom structure may suffer from dynamic loads in the future time domain. First, based on the predicted state sequence of the crane system in the future time domain output by the hybrid prediction model, the system uses a parameterized boom structure stress estimation model to calculate the dynamic equivalent stress time series of one or more key parts of the boom over a future period. This stress estimation model is pre-calibrated through finite element analysis, and its input parameters include the real-time pitch angle, slewing angle, wire rope tension, and load swing inertial force derived from the predicted state. These parameters collectively reflect the stress state of the boom during operation. For example, the pitch angle may be limited to the range of -10 degrees to +80 degrees in actual operation, the slewing angle may be set between -180 degrees and +180 degrees, and the wire rope tension varies dynamically according to the load weight, which may range from several tons to hundreds of tons, depending on the crane model and operational requirements. The principle of this model is based on structural mechanics and material properties. By simplifying the finite element analysis results, it constructs a parameterized relationship for rapid calculation, thereby efficiently outputting the equivalent stress values ​​of key components, such as Von Mises stress, in real-time control. The output range may be between tens of megapascals and hundreds of megapascals, serving as a reference. During implementation, the model runs periodically, for example, once every first control cycle (e.g., 0.5 to 2 seconds). It generates a stress time series based on the latest predicted state sequence, thus providing basic data for subsequent fatigue analysis. The setting of the first control cycle (0.5 to 2 seconds) mainly matches the low-frequency characteristics of ocean wave disturbances (the dominant frequency is usually below 0.2 Hz) and provides sufficient computational time window for solving multi-objective optimization problems involving load oscillation and structural fatigue, ensuring the feasibility and foresight of the upper-level planning strategy.

[0067] Next, the system processes the calculated dynamic equivalent stress time series using the rainflow counting method to identify complete stress cycles. Rainflow counting is an algorithm based on the principle of material fatigue. It decomposes the complex random load history into a series of complete stress cycles by analyzing the peaks and troughs in the stress time series. Each cycle includes a stress amplitude and a mean. For example, in typical offshore operations, the stress amplitude may range from several megapascals to tens of megapascals, depending on sea conditions and load conditions, while the number of cycles may reach dozens in the future time domain (e.g., 5 to 30 seconds). The principle of this method is to simulate the fatigue behavior of materials under actual loads and quantify damage accumulation by identifying cycles. In implementation, the system inputs the stress sequence into the rainflow counting module, which may be implemented as a software algorithm, operating at a frequency synchronized with the first control cycle, for example, processing the sequence data once per optimization cycle. Then, for each identified stress cycle, the system combines a fatigue life model based on the material's SN curve to calculate the fatigue damage caused by that stress cycle. The SN curve describes the number of cycles required for a material to fail at a specific stress amplitude. For example, for commonly used structural steel, a stress amplitude of 100 MPa may correspond to millions of cycles, but the actual value needs to be calibrated based on material certificates and experimental data. Damage degree calculation is achieved by mapping the stress amplitude to the equivalent life loss on the SN curve, for example, using the linear damage assumption in the Palmgren-Miner rule. During operation, the system calculates the damage degree in real time for each identified cycle. These calculations are typically completed within milliseconds, ensuring that real-time control is not affected. These technical features, through the combination of rainflow counting and the SN model, transform complex stress histories into quantifiable damage indicators, providing fatigue-related inputs for optimization decisions.

[0068] Finally, the system employs Miner's linear cumulative damage rule to sum the fatigue damage caused by all stress cycles within the future prediction time domain, thereby estimating the structural fatigue damage forecast for the future time domain. Miner's rule is based on the principle of linear superposition of fatigue damage, assuming that the damage from each stress cycle is independent and additive, and the total damage equals the sum of the damage from each cycle. For example, within the prediction time domain of the next 10 seconds, the identified stress cycles may produce a total damage between 0.001 and 0.1, with the specific value depending on the load intensity and the number of cycles, serving as a possible reference range. This principle allows the system to comprehensively consider short-term performance and long-term structural health in control optimization, avoiding excessive pursuit of sway suppression that could accelerate fatigue. During implementation, the cumulative damage calculation module executes within each first control cycle, performing a rapid summation based on the damage values ​​of each cycle output from the previous step, and inputting the result as part of the objective function into the optimizer. For example, when solving the first optimization problem, this forecast is weighted and combined with the load sway kinetic energy to guide the generation of a reference control trajectory. Through a coherent process from stress estimation to damage accumulation, each technical feature enables dynamic prediction and integration of structural fatigue, ensuring that the control strategy not only suppresses sway but also proactively reduces the negative impact on boom life.

[0069] To ensure the real-time performance of fatigue damage estimation, this system employs the following simplification strategies: First, the parameterized boom structure stress estimation model is essentially an empirical formula (such as a polynomial or response surface model) pre-calibrated through extensive offline finite element analysis. Its online calculations consist only of algebraic operations, consuming extremely little time (microseconds). Second, the rainflow counting method has been modified for real-time processing. Instead of processing the complete historical stress time series, it only performs simplified iterative identification (e.g., extracting only the main peak and valley values) on the predicted stress series output by the hybrid prediction model within a finite future time domain. This data volume is small, the period is known, and the computational load is controllable. Finally, fatigue damage accumulation (Miner's rule) is a simple arithmetic summation. The entire estimation process is designed to be completed within the first control cycle (0.5~2 seconds), and its computational load is far lower than that of trajectory optimization problems in the same period, thus not becoming a limitation to the system's real-time performance.

[0070] In one specific implementation, in step S4, during the solution of the second optimization problem, an adaptive weight adjustment mechanism is introduced. This adaptive weight adjustment mechanism dynamically adjusts the weight coefficients of the tracking error term in the objective function based on real-time load weight, wire rope length, and environmental disturbance intensity. Specifically, the adaptive weight adjustment mechanism includes: S410. Real-time acquisition of load weight, wire rope length, and environmental disturbance intensity; the load weight is obtained through the weighing sensor of the crane system or based on the wire rope tension estimation model; the wire rope length is measured by an encoder or laser ranging sensor; the environmental disturbance intensity is extracted from the wave prediction information output by the multi-source environmental sensing system, including significant wave height, average wave period, and wave propagation direction, and is comprehensively calculated into a dimensionless disturbance intensity index; S 420. Input the load weight, wire rope length, and environmental disturbance intensity into a preset weight calculation function. This weight calculation function is a nonlinear mapping function obtained based on multi-parameter fitting or machine learning regression, and its output is the weight coefficient of the tracking error term. The form of the weight calculation function is as follows: , in, Here are the adjusted weighting coefficients, where m is the load weight, l is the wire rope length, d is the environmental disturbance intensity index, f1, f2, and f3 are the normalized influence functions for each parameter, and k1, k2, and k3 are adjustable gain coefficients. The benchmark weighting coefficient; It is an increasing function of the load weight; It is a decreasing function of the wire rope length; It is a non-monotonic function of the intensity of environmental disturbance; S430. The adaptive weight adjustment mechanism further includes an online update module for online optimization of the gain coefficients k1, k2, and k3. The optimization triggering condition is any of the following: a) reaching a preset fixed time interval; b) the accumulated number of control cycles reaches a preset threshold; c) the system performance evaluation index deteriorates to below a preset standard. The performance evaluation index is the weighted sum of the average kinetic energy of the load swing and the tracking error over a past period.

[0071] In the above implementation, the adaptive weight adjustment mechanism first requires real-time acquisition of key operating parameters, including load weight, wire rope length, and environmental disturbance intensity. Load weight is typically measured directly by a weighing sensor integrated into the crane system, or indirectly derived based on a wire rope tension estimation model. For example, in typical applications, the load weight can range from several tons to hundreds of tons, depending on the crane's design capacity and actual operational requirements, serving as a reference range. Wire rope length is precisely measured using an encoder or laser rangefinder. These sensors can track rope length changes with millimeter-level accuracy. In offshore cranes, the rope length can be adjusted within tens of meters based on the operating height, for example, selected between 10 meters and 100 meters. The environmental disturbance intensity is extracted from wave prediction information output by a multi-source environmental sensing system, including factors such as significant wave height, mean wave period, and wave propagation direction. This is then comprehensively calculated and transformed into a dimensionless disturbance intensity index. For example, the significant wave height can be selected between 0.5 meters and 5 meters, referencing actual sea conditions, and the mean wave period between 3 seconds and 15 seconds. The disturbance intensity index quantifies the severity of the environmental impact by weightedly combining these parameters, with its value set between 0 and 1 as a normalization reference. The principle is that these parameters collectively reflect the dynamic load conditions, geometric configuration, and external disturbance level of the crane system, thus providing an objective basis for weight adjustment. During implementation, these parameters are sampled at a fixed frequency, such as every 100 to 500 milliseconds, with the specific frequency depending on the real-time requirements of the control system. A data preprocessing module ensures the integrity and consistency of the data, laying the foundation for subsequent weight calculations. The various technical features work collaboratively through real-time data acquisition and preprocessing, ensuring the accuracy and timeliness of parameter acquisition and providing reliable input for adaptive adjustment.

[0072] After acquiring real-time parameters, the system inputs them into a preset weight calculation function. This function is a nonlinear mapping function based on multi-parameter fitting or machine learning regression, used to output the weight coefficients of the tracking error term. The weight calculation function involves several components: load weight affects the weight through an increasing function, meaning that as the load increases, the system tends to increase the weight of the tracking error to enhance control accuracy; wire rope length affects the weight through a decreasing function, because increasing rope length usually amplifies load sway, so the system can appropriately reduce the weight to avoid over-response; environmental disturbance intensity is handled through a non-monotonic function, reflecting that the weight can be increased under moderate disturbances to cope with uncertainty, while the weight can be decreased under extreme disturbances to maintain stability. The function also includes adjustable gain coefficients and a baseline weight coefficient. The initial values ​​of these coefficients can be determined based on historical data or simulation. For example, gain coefficients k1, k2, and k3 can be selected in the range of 0.1 to 1.0, and the baseline weight coefficient β0 can be set to a value between 0.5 and 1.5. The specific values ​​need to be optimized according to the actual control system. The principle is to adaptively adjust the weights so that the controller automatically balances tracking performance and system robustness under different operating conditions. During operation, the weight calculation function is executed within each control cycle, for example, calculating new weights every 50 milliseconds, and directly applying them to the objective function of the second optimization problem, ensuring that the control commands can adapt to changing conditions in real time. Various technical features work collaboratively through function mapping to transform multi-source parameters into a unified adjustment coefficient, thereby dynamically optimizing the control strategy. This allows the system to prioritize accuracy under heavy loads, stability over long ropes, and flexibly adjust its response intensity under environmental disturbances.

[0073] It should be noted that the function Designed as a non-monotonic system, it is based on the differentiated strategies that the control system should adopt under different environmental disturbance intensities. Its design principle is as follows: At low to moderate disturbance intensities, the amplitude of wave excitation is finite and relatively regular. In this case, the control system has the capability and should enhance the tracking weight (i.e., increase...) The system actively compensates for disturbances by adjusting the tracking weight (i.e., by reducing the tracking weight) to achieve higher control accuracy and sway suppression performance. However, when the disturbance intensity exceeds a certain critical value and enters the high-intensity range, wave excitation becomes more severe and its randomness significantly increases, leading to greater uncertainty in the prediction model. If high tracking weight is maintained at this point in pursuit of tracking accuracy, the control command may become overly aggressive, causing frequent saturation of the actuator, severe fluctuations in system pressure, and even inducing system instability. Therefore, under strong disturbances, the control strategy should shift from prioritizing "precise tracking" to prioritizing "robust stability," by appropriately reducing the tracking weight (i.e., decreasing the tracking weight). The value of the variable (increases first, then decreases) sacrifices some performance for system stability and safety. This non-monotonic characteristic allows the controller to adaptively strike an optimal balance between performance and robustness, representing an intelligent strategy for coping with complex and ever-changing marine environments. In practical design, Specifically, it can be implemented as a piecewise function based on expert experience or fitting of massive simulation data (e.g., linearly increasing when d is small, and turning to linear or exponential decay after d exceeds a threshold), or directly mapped by a lightweight neural network to map this complex nonlinear relationship.

[0074] The adaptive weight adjustment mechanism also includes an online update module for periodically optimizing the gain coefficients k1, k2, and k3 to further enhance the system's adaptability. The optimization trigger conditions for the online update module include various scenarios: for example, reaching a preset fixed time interval, such as every 24 hours or every 1000 control cycles; or the cumulative number of control cycles reaching a preset threshold, such as 10,000 operations; or the system performance evaluation index falling below a preset standard. This index is typically calculated based on the weighted sum of the average kinetic energy of load fluctuations and tracking errors over a past period, such as performance data from the past 10 minutes, where the performance threshold can be set as a percentage of the historical average as a reference. The principle is to continuously monitor system performance and automatically adjust the gain coefficients to compensate for long-term changes or unmodeled dynamics. During implementation, when the trigger conditions are met, the system collects recent operating data and recalculates the gain coefficients using optimization algorithms such as gradient descent or evolutionary strategies, completing the update within minutes without interrupting normal control. The updated gain coefficients are immediately applied to the weight calculation function, forming a closed-loop optimization. Through the coordination of triggering conditions, performance evaluation, and optimization algorithms, the various technical features ensure that the weight adjustment mechanism can self-improve, thereby maintaining the control effect at its best and adapting to long-term changes caused by equipment aging or environmental evolution.

[0075] In one specific embodiment, a fail-safe loop is also included. This fail-safe loop monitors the pressure and flow signals of the electro-hydraulic servo actuator in real time. When an actuator jamming or leakage fault is detected, it immediately switches to a redundant actuator or triggers an emergency braking procedure, and at the same time sends a fault alarm signal to the system health management process.

[0076] In the above implementation, the fail-safe loop, as a key component of the offshore crane anti-sway predictive control method, is primarily responsible for real-time monitoring and protection of the electro-hydraulic servo actuator. This loop achieves continuous monitoring by acquiring pressure and flow signals from the actuator at high frequency. The pressure signal reflects the internal working state of the hydraulic system, such as changes in system load and resistance, while the flow signal indicates the flow of hydraulic oil, used to assess the actuator's movement rate and efficiency. During implementation, pressure and flow sensors typically acquire data at a sampling frequency of 100 to 500 times per second. This frequency range can be adjusted according to the system's real-time requirements; for example, it can be increased to 1000 Hz in some high-dynamic scenarios, but 200 Hz is generally used as a reference to ensure a balance between data timeliness and processing load. The monitoring principle is based on hydraulic system dynamics: an abnormal rise in pressure suggests actuator jamming or a sudden increase in external load, while an abnormal drop in flow indicates internal leakage or oil circuit blockage. In the claims, this real-time monitoring is used to provide a basic data stream for fault diagnosis. By continuously comparing the signal with a preset normal range (e.g., the pressure operating range can be set between 5 MPa and 30 MPa, and the flow rate range is between 20 liters / minute and 120 liters / minute, with the specific values ​​fluctuating slightly depending on the crane model and hydraulic design), the system can quickly capture deviations, thereby supporting subsequent fault judgment and response.

[0077] Based on real-time monitored signals, the fail-safe loop employs a dedicated algorithm to detect actuator jamming or leakage faults. Jamming faults typically manifest as a pressure signal consistently exceeding a threshold accompanied by flow fluctuations, while leakage faults are characterized by a slow decrease in pressure and an unexpected increase in flow. The detection principle relies on signal pattern recognition and threshold comparison. For example, anomalies are identified by calculating the rate of pressure change or flow deviation. The pressure change rate threshold can be set to an increase of more than 5 MPa per second, or a flow deviation exceeding the normal value by 15% to 25%. These thresholds are based on historical operating data and system calibration, but in practical applications, they can be flexibly adjusted within a range of 10% to 30% depending on the operating environment. In implementation, the detection algorithm runs on an embedded controller, performing an analysis every control cycle (e.g., 10 to 50 milliseconds). Data is smoothed using sliding window statistics or moving average methods to reduce noise interference. This detection mechanism is used to proactively identify potential faults, preventing control performance degradation or safety accidents caused by actuator failure. The specific operation involves real-time calculation of signal characteristic values ​​and comparison with dynamic thresholds. Once the limit is exceeded for multiple consecutive cycles, a fault flag is triggered, ensuring the accuracy and timeliness of detection.

[0078] When a jamming or leakage fault is detected, the fail-safe loop immediately initiates response measures, including switching to a redundant actuator or triggering an emergency braking procedure, while simultaneously sending a fault alarm signal to the system health management process. The redundant actuator is a pre-configured backup hydraulic component, such as an additional hydraulic cylinder or backup valve assembly. The switching process is managed by the electronic control unit and is completed within a very short time after fault confirmation, for example, activating the redundant component within 20 to 100 milliseconds. The specific switching time depends on the hardware response speed, but is typically targeted at completion within 50 milliseconds for a seamless transition. The emergency braking procedure involves rapidly stopping the crane's movement, for example, by shutting down the main hydraulic pump, releasing the safety valve, or applying mechanical braking. The braking process can be designed to be gradual to avoid shock, for example, gradually reducing the system pressure to a safe level over 1 to 3 seconds. Simultaneously, a fault alarm signal is sent to the system health management process in real time. This signal contains information such as the fault type, time of occurrence, and severity, so that the health management module can record the event and trigger control mode adjustments. In the claims, these response measures are used to achieve rapid fault isolation and system protection. The various technical features work together to form a closed loop: monitoring provides the data foundation, detects and identifies risks, and the response measures execute recovery actions and are integrated with health management processes to ensure the adaptability of the overall control strategy. For example, during operation, the fail-safe loop will prioritize attempting redundancy switching to maintain operation, initiating emergency braking only if switching fails or a severe fault occurs, thereby maximizing system availability while ensuring safety.

[0079] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0080] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for predictive control of sway resistance in offshore cranes, characterized in that, Includes the following steps: S1. The current state of the crane is acquired in real time through a group of state sensors installed on the crane system. The current state of the crane includes the spatial swing angle of the load, the pitch angle and rotation angle of the boom, and the real-time length of the wire rope; and the prediction information of future waves is acquired through a multi-source environmental sensing system. S2. Input the current state of the crane and the predicted information of the future waves into a pre-trained hybrid prediction model. The hybrid prediction model is constructed based on the parameterized physical dynamics model of the crane system and supplemented by a data-driven residual correction network to predict the state sequence of the crane system under different control commands in the future time domain. S3. Taking the current state of the crane as the initial condition, based on the predicted state sequence output by the hybrid prediction model, solve the first optimization problem within the first control cycle to obtain a reference control trajectory in the future time domain. The objective function of the first optimization problem includes the weighted sum of load swing kinetic energy and structural fatigue damage. S4. In a second control cycle that is shorter than the first control cycle, based on the reference control trajectory, solve the second optimization problem to obtain the real-time control command at the current moment. The objective function of the second optimization problem is to minimize the error in tracking the reference control trajectory while satisfying the operational constraints of the electro-hydraulic servo system; S5. Convert the real-time control command into a drive signal and output it to the electro-hydraulic servo actuator of the crane; S6. Simultaneously, independently execute the system health management process, evaluate the health status of the multi-source environmental perception system, the hybrid prediction model, and each sensor in real time, and dynamically adjust the control modes from S1 to S5 based on the evaluation results.

2. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S1, the multi-source environmental sensing system includes: Shipborne coherent lidar acquires the sea surface wave height sequence and wave propagation direction in real time based on the principle of coherent detection; X-band marine radar extracts wave spectrum features, including significant wave height, average wave period, and wave direction, through a sea surface echo inversion algorithm. A ship motion prospective observer based on ship inertial measurement unit data predicts ship motion over a future time domain by integrating current and historical ship motion attitude data based on ship motion equations. The multi-source environmental perception system uses a data fusion algorithm to fuse environmental and motion prediction information from different sources. The data fusion algorithm employs extended Kalman filtering or unscented Kalman filtering, uses the wave surface height sequence measured by lidar as the observation update, uses the wave spectrum features extracted by marine radar as the prior distribution, and combines the output of the ship motion prospective observer to generate prediction information of future waves through a multi-source asynchronous data fusion framework. The prediction information of future waves includes wave excitation force sequence, spatiotemporal distribution of wave surface disturbance, and ship motion response.

3. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S2, the construction and training method of the hybrid prediction model specifically includes: S210. Establish a parameterized physical dynamics model of the crane system as a basic predictor; S220. Collect historical operating data, which includes crane status sequence, control command sequence and corresponding environmental data; S230. Using the historical operating data, a residual correction network is trained through supervised learning. The training objective of the residual correction network is to predict the residual between the output of the parameterized physical dynamics model and the actual crane state. S240. The output of the parameterized physical dynamics model is added to the prediction output of the residual correction network to form the final prediction result of the hybrid prediction model.

4. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S3, the first optimization problem is solved using a simplified constraint set, which includes only the kinematic constraints of the boom joint space and the stress amplitude constraints of key parts.

5. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S4, the second optimization problem is solved by a model predictive controller or an exact feedback linearization controller. Its internal model is a simplified model that characterizes the core nonlinearity of the electro-hydraulic servo system. The operational constraints include valve core displacement limit, system pressure limit, and actuator flow limit.

6. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S6, the system health management process specifically includes: S610: Real-time monitoring of the data quality and reliability of each sensing unit; S620. Evaluate the error between the prediction output of the hybrid prediction model and the measured state sequence. When the cumulative error exceeds a first preset threshold, determine that the health of the hybrid prediction model has decreased. S630. Based on the health status assessment results, dynamically and seamlessly switch between multiple control modes, wherein the control modes include at least: Full-function mode, all perception and prediction functions are enabled normally; Degraded sensing mode, which relies on the remaining effective sensing sources when either lidar or marine radar fails; In the safety fallback mode, when the entire environmental perception system fails or the health of the hybrid prediction model deteriorates significantly, the system switches to a robust feedback control method based solely on the data from the state sensor group. This robust feedback control method is either sliding mode variable structure control or fuzzy PID control.

7. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, It also includes an online model update loop, which periodically uses actual operating data within a preset time window to incrementally learn and update the residual correction network in the hybrid prediction model. The update process keeps the parameters of the parameterized physical dynamics model fixed and only optimizes the weights of the residual correction network.

8. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S3, the estimation process for structural fatigue damage in the objective function of the first optimization problem specifically includes: S310. Based on the predicted state sequence of the crane system in the future time domain output by the hybrid prediction model, the dynamic equivalent stress time sequence of one or more key parts of the boom in the future time domain is calculated through a parameterized boom structure stress estimation model. S320. The stress estimation model of the boom structure is a parametric model pre-calibrated based on finite element analysis. Its inputs include at least the real-time pitch angle, slewing angle, wire rope tension, and load swing inertial force derived from the predicted state. The output is the equivalent stress value of key parts. S330. The dynamic equivalent stress time series is processed by rainflow counting method to identify complete stress cycles; S340. For each identified stress cycle, calculate the fatigue damage caused by that stress cycle based on its stress amplitude and the fatigue life model established based on the material SN curve. S350. Using Miner's linear cumulative damage rule, the fatigue damage caused by all stress cycles in the future prediction time domain is accumulated to estimate the structural fatigue damage forecast in the future time domain.

9. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, In step S4, during the solution of the second optimization problem, an adaptive weight adjustment mechanism is introduced. This mechanism dynamically adjusts the weight coefficients of the tracking error term in the objective function based on real-time load weight, wire rope length, and environmental disturbance intensity. Specifically, the adaptive weight adjustment mechanism includes: S410. Real-time acquisition of load weight, wire rope length, and environmental disturbance intensity; the load weight is obtained through the weighing sensor of the crane system or based on the wire rope tension estimation model; the wire rope length is measured by an encoder or laser ranging sensor; the environmental disturbance intensity is extracted from the wave prediction information output by the multi-source environmental sensing system, including significant wave height, average wave period, and wave propagation direction, and is comprehensively calculated into a dimensionless disturbance intensity index; S 420. Input the load weight, wire rope length, and environmental disturbance intensity into a preset weight calculation function. This weight calculation function is a nonlinear mapping function obtained based on multi-parameter fitting or machine learning regression, and its output is the weight coefficient of the tracking error term. The form of the weight calculation function is as follows: ， in, Here are the adjusted weighting coefficients, where m is the load weight, l is the wire rope length, d is the environmental disturbance intensity index, f1, f2, and f3 are the normalized influence functions for each parameter, and k1, k2, and k3 are adjustable gain coefficients. The benchmark weighting coefficient; It is an increasing function of the load weight; It is a decreasing function of the wire rope length; It is a non-monotonic function of the intensity of environmental disturbance; S430. The adaptive weight adjustment mechanism further includes an online update module for online optimization of the gain coefficients k1, k2, and k3. The optimization trigger condition is any of the following: a) reaching a preset fixed time interval; b) the accumulated number of control cycles reaches a preset threshold; c) the system performance evaluation index deteriorates to below a preset standard. The performance evaluation index is the weighted sum of the average kinetic energy of the load swing and the tracking error over a period of time.

10. The anti-sway prediction and control method for offshore cranes as described in claim 1, characterized in that, It also includes a fail-safe loop that monitors the pressure and flow signals of the electro-hydraulic servo actuator in real time. When an actuator jamming or leakage fault is detected, it immediately switches to a redundant actuator or triggers an emergency braking procedure, and at the same time sends a fault alarm signal to the system health management process.

Images