Ship intelligent management method and system based on digital twinning

Abstract

This invention relates to the fields of marine engineering and communication technology, and discloses a ship intelligent management method and system based on digital twins, including: a shipborne sensing and modeling device and a computing network resource orchestration controller. The device collects ship motion and radar data to reconstruct a forward-looking wave digital twin; the controller determines the data coverage scale based on the forward-looking prediction window and grid resolution, and determines the effective sampling update frequency by combining the ship's speed and relative wave encounter angle; based on the product of these two factors, it calculates a navigation condition-related load factor characterizing the nonlinear growth of service demand with speed; this factor is used to determine the minimum service throughput requirement, and it is used as a feedforward parameter input to a multi-dimensional resource adaptation model including latency, energy consumption, and cost; finally, a swarm intelligence algorithm is used to solve and distribute network slices and edge computing node configurations. This invention can effectively cope with the dual dynamic changes of navigation conditions and network environment, ensuring real-time spatiotemporal synchronization of the digital twin.

Description

Technical Field

[0001] This invention relates to the fields of marine engineering and communication technology, and more specifically, to a method and system for intelligent ship management based on digital twins. Background Technology

[0002] The development of intelligent ships and maritime IoT technologies is rapid. To achieve real-time monitoring of ship navigation status, seakeeping analysis, and assisted navigation, digital twin technology is widely used to construct virtual mappings of ships. Such systems typically require the ship to collect massive amounts of sensor data (such as radar, attitude, and operational conditions) and transmit it back to shore or the cloud in real time via wireless networks for high-precision simulation and decision-making. In actual maritime communication scenarios, to overcome the limitations of single-link coverage and stability, a hybrid heterogeneous network architecture consisting of 5G / 6G shore-based networks, ship-to-ship multi-hop ad hoc networks, and satellite communication is usually adopted. Simultaneously, to reduce end-to-end latency, edge computing technology has been introduced, offloading some computing tasks to edge nodes closer to the ship and using network slicing technology to isolate and protect service traffic. However, existing ship digital twin resource scheduling schemes have significant limitations when dealing with highly dynamic sea conditions, mainly in the following two aspects: First, the maritime wireless propagation environment is highly unstable and prone to discrete jumps. While evaporative waveguides in the ocean atmosphere can be beneficial for beyond-line-of-sight transmission under certain conditions, they are greatly affected by weather conditions, causing communication links to not degrade smoothly but rather exhibit abrupt changes in connectivity. This phenomenon often forces communication systems to frequently switch topology between direct connections, multi-hop relays, or satellite links. Existing resource management strategies are mostly based on feedback adjustments using current channel state information. When the network topology changes drastically, they often cannot respond in time, leading to data transmission interruptions or a sharp increase in latency.

[0003] Secondly, existing scheduling strategies fail to fully consider the nonlinear dynamic changes in the business needs of ship digital twins. Traditional views tend to treat twin data streams as constant bit streams or simple periodic reporting. However, in high-precision intelligent ship management, to achieve accurate prediction and control of the ship's future motion, it is necessary to reconstruct and transmit wave field information in front of the hull in real time. In actual navigation, the frequency at which a ship encounters waves (wave encounter frequency) is directly affected by its speed and heading. When the ship accelerates or navigates head-on, to maintain the spatiotemporal synchronization accuracy between the twin and the physical ship, the system's demand for the update frequency and data scale of the forward-looking wave field information exhibits a sharp, nonlinear increase. Existing technical solutions lack the ability to predict this demand surge mechanism caused by the coupling of navigation conditions and the natural environment. In summary, when a ship is in high-speed or complex wave conditions (leading to a surge in business demand) and simultaneously encounters changes in evaporating waveguides or network topology switching (leading to a sharp drop in network supply), existing feedback-based scheduling mechanisms are prone to failure, unable to meet millisecond-level spatiotemporal synchronization requirements, thus affecting the real-time performance and security of ship control commands. Summary of the Invention

[0004] This invention provides a digital twin-based intelligent ship management method and system, which solves the technical problems mentioned in the background art.

[0005] The first aspect is a ship intelligent management system based on digital twins, including: The shipborne sensing and modeling device is configured to collect radar images, speed, heading and attitude data of the ship, and reconstruct a forward-looking wave digital twin based on the radar images. The network resource orchestration controller is configured to perform resource orchestration within a scheduling cycle, wherein the resource orchestration includes: The coverage scale of the twin data is determined based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin. Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined; Based on the product relationship between the twin data coverage scale and the effective sampling update frequency, calculate the navigation condition-related load factor that characterizes the non-linear growth of twin service demand with the speed. The minimum service throughput requirement to meet the spatiotemporal synchronization requirement is determined by using the associated load factor of the navigation condition, and the minimum service throughput requirement is used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption and cost. The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions.

[0006] Secondly, a digital twin-based intelligent ship management method, realizing a digital twin-based intelligent ship management system as described in any one of the claims, including: Collect radar images, speed, heading, and attitude data of the ship, and reconstruct a forward-looking wave digital twin based on the radar images; Resource orchestration is performed within a scheduling period, and the resource orchestration includes: The coverage scale of the twin data is determined based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin. Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined; Based on the product relationship between the twin data coverage scale and the effective sampling update frequency, calculate the navigation condition-related load factor that characterizes the non-linear growth of twin service demand with the speed. The minimum service throughput requirement to meet the spatiotemporal synchronization requirement is determined by using the associated load factor of the navigation condition, and the minimum service throughput requirement is used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption and cost. The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions.

[0007] The beneficial effects of this invention include: by constructing a navigation condition-related load factor, this invention characterizes the nonlinear burst pattern of ship digital twin service demand as speed and sea state change, and transforms the nonlinear burst pattern into a deterministic feedforward driving quantity introduced into the resource adaptation model; compared with traditional feedback scheduling, this invention can proactively sense the surge in service throughput demand caused by changes in navigation state before network topology mutations or evaporative waveguide effects cause fluctuations in communication supply, thereby completing the joint orchestration of network slicing and edge computing power in advance; this feedforward mechanism effectively avoids the risk of resource mismatch under highly dynamic sea states, significantly improves the effectiveness of end-to-end transmission while ensuring millisecond-level spatiotemporal synchronization accuracy, and achieves globally optimal configuration of energy consumption and cost. Attached Figure Description

[0008] Figure 1 This is a flowchart of the ship intelligent management system based on digital twins of the present invention; Figure 2 This is a schematic diagram of a specific implementation scenario of the present invention. Detailed Implementation

[0009] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0010] Example 1: As Figure 1 As shown, the ship intelligent management system based on digital twins includes: The shipborne sensing and modeling device is configured to collect radar images, speed, heading and attitude data of the ship, and reconstruct a forward-looking wave digital twin based on the radar images. The network resource orchestration controller is configured to perform resource orchestration within a scheduling cycle, wherein the resource orchestration includes: The coverage scale of the twin data is determined based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin. Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined; Based on the product relationship between the twin data coverage scale and the effective sampling update frequency, calculate the navigation condition-related load factor that characterizes the non-linear growth of twin service demand with the speed. The minimum service throughput requirement to meet the spatiotemporal synchronization requirement is determined by using the associated load factor of the navigation condition, and the minimum service throughput requirement is used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption and cost. The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions.

[0011] Preferably, the shipborne sensing and modeling device acquires radar images, speed, heading, and attitude data of the ship, and reconstructs a forward-looking wave digital twin based on the radar images, specifically including: The shipborne sensing and modeling device performs linear interpolation alignment of the acquired radar images of non-same-source frequencies with the attitude data using a unified time reference; it then uses the attitude data to construct a rotation matrix, transforms the radar images from the radar local coordinate system and projects them onto the ground-fixed local horizontal coordinate system, thereby obtaining a motion-compensated radar grid image sequence. A three-dimensional fast Fourier transform is performed on the radar grid image sequence to obtain the frequency domain wavenumber spectrum. A filter mask is then constructed using the deep-water gravity wave dispersion relation to extract the effective wave signal components. The calculation formula for the filter mask is as follows: ; in, The filter mask, and These are the wave numbers in the two horizontal directions, Angular frequency, It is the acceleration due to gravity. For wavenumber modulus, Let be the dispersion tolerance constant; The wave surrogate quantity is obtained through three-dimensional fast Fourier inverse transform, and the phase-resolved wavefront elevation field is calculated using a linear calibration formula: ; in, The phase-resolved wavefront elevation field, The wave surrogate quantity is obtained after the three-dimensional inverse fast Fourier transform. and The linear coefficients and bias constants were determined through offline comparison calibration; The historical frame sequence of the phase-resolved wavefront elevation field is input into a pre-trained and fixed Fourier neural operator prediction model. Through inference calculation, a three-dimensional forward-looking wave field within a future time window is generated as a digital twin of the forward-looking wave. The inference formula of the Fourier neural operator prediction model is as follows: ; in, This is the output tensor that contains wave field data within the future prediction time window. The input tensor contains the phase-resolved wavefront elevation field within the historical time window. For parameters The Fourier neural operator network has been solidified.

[0012] Radar imagery consists of data on the intensity of electromagnetic waves reflected from the sea surface, captured during ship navigation. This data can be acquired using a shipborne X-band sea surface radar, operating at approximately 9.41 GHz, with a range resolution of about 8.5 meters and a frame rate of 2 Hz. Speed ​​is the distance traveled by the ship per unit time, obtained through the fusion output of a Global Navigation Satellite System (GNSS) and a log. Heading is the direction of the ship's navigation, obtained through a gyro and GNSS heading data, measured in radians, ranging from -3.14 to 3.14. Attitude data includes the ship's spatial attitude and rate of attitude change, including roll, pitch, yaw, and corresponding angular velocities, acquired through an inertial measurement unit (IMU) or a motion reference unit (RPU). The rotation matrix describes the attitude transformation relationship between the radar local coordinate system and the ground-based local horizontal coordinate system, constructed based on the roll, pitch, and yaw angles from the attitude data. The radar local coordinate system is a coordinate system fixed at the phase center of the radar antenna. Preferably, the X-axis points towards the bow, the Y-axis points to the starboard side, and the Z-axis points downwards, conforming to the general design specifications for ship navigation equipment coordinate systems. The ground-based local horizontal coordinate system is a coordinate system fixed to a local area of ​​the Earth's surface. Preferably, it is an east-facing, north-facing, or celestial coordinate system to unify the measurement benchmarks of different devices and eliminate the influence of ship motion on the data. The radar grid image sequence is a continuous collection of radar images organized in a uniform grid after motion compensation. The dispersion tolerance constant is a parameter controlling the looseness of the filter mask, preferably 0.15 times the square of the angular frequency, to balance the accuracy of gravity wave signal extraction and noise immunity. The wave surrogate quantity is obtained through a three-dimensional inverse fast Fourier transform and is intermediate data that indirectly reflects wave characteristics. The linearity coefficient is a proportionality coefficient used to map the wave surrogate quantity to the actual wave surface elevation. Its preferred value is determined through offline least squares fitting, with the fitting benchmark being the measured data from the shipborne wave sensor to ensure the accuracy of the wave surface elevation field calculation. The bias constant is a parameter used to correct the deviation in the wave surrogate mapping. It is determined through offline least squares fitting, with the fitting benchmark being the measured data from the shipborne wave sensor, to eliminate the influence of systematic errors on the wave surface elevation calculation. The phase-resolved wave surface elevation field is spatial field data that reflects the three-dimensional shape and phase information of the sea surface, and is the core data for describing wave morphology. The historical frame sequence consists of continuously acquired multi-frame phase-resolved wave surface elevation field data, used to provide training and inference basis for the prediction model. The Fourier neural operator prediction model is a deep learning model based on Fourier transform, used for short-term wave field prediction. It preferably uses a network structure fixed after offline training, with 20 input channels, 10 output channels, and a grid size of 128x128, to balance prediction accuracy and real-time inference speed. The network parameters are the internal weights and biases of the Fourier neural operator prediction model, preferably fixed parameters obtained through training with synthetic training data, to ensure the model's predictive ability for the wave field.The future time window is the time range within which the model predicts the wave field, preferably 1 second, to accommodate a 0.1-second scheduling cycle and provide timely demand input for resource orchestration. The input tensor is a multidimensional data structure that organizes historical frame sequences according to model requirements, used for model inference calculations. The output tensor is a multidimensional data structure obtained from model inference, containing wave field data within the future time window. The three-dimensional forward-looking wave field is a digital twin reflecting the three-dimensional morphology of sea surface waves ahead of the ship over a future period, serving as core reference data for intelligent ship management.

[0013] Linear interpolation alignment of non-homogeneous frequency data with a unified time base includes: firstly, selecting a precise time protocol as the system's global clock, prioritizing the use of the Global Navigation Satellite System (GNSS) for time synchronization to ensure time synchronization of all devices. A fixed scheduling period of 0.1 seconds is defined. For non-homogeneous data with different sampling frequencies, such as radar images and attitude data, the original sampling points before and after each scheduling moment are found, and interpolation coefficients are calculated. These coefficients are the time difference between the scheduling moment and the previous sampling moment divided by the time interval between the two sampling moments. The data for the scheduling moment is then calculated using a linear interpolation formula. For example, if the inertial measurement unit sampling rate is 100 Hz and the airspeed sampling rate is 10 Hz, both are interpolated and aligned according to the scheduling period. When the radar frame rate is 2 Hz, the nearest frame is used in adjacent scheduling periods in conjunction with motion compensation.

[0014] Radar image motion compensation includes: first, constructing a rotation matrix based on roll, pitch, and yaw angles from the attitude data. This matrix is ​​generated using the Euler angle transformation method and describes the attitude transformation from the radar local coordinate system to the ground-based local horizontal coordinate system. Next, the polar coordinate pixels of the radar image are converted to Cartesian coordinates in the radar local coordinate system, and then transformed to the ground-based local horizontal coordinate system using the rotation matrix. The horizontal component is extracted as the coordinates of the resampled target. Finally, a bilinear interpolation method is used to remap the original radar image onto a uniform ground-based coordinate system grid, resulting in a radar grid image sequence that eliminates ship motion distortion.

[0015] Wave signal reconstruction includes: performing three-dimensional spatiotemporal spectral analysis on the radar grid image sequence, i.e., simultaneously performing Fast Fourier Transform on the two-dimensional spatial grid and the time dimension to obtain the frequency domain wavenumber spectrum. Based on the deep-water gravity wave dispersion relation, i.e., the square of the angular frequency equals the product of gravitational acceleration and wavenumber, a filter mask is constructed. This mask is in the form of an exponential function, with the exponent being negative. The square of the difference between the square of the angular frequency and the gravitational acceleration multiplied by the wavenumber is divided by twice the square of the dispersion tolerance constant. The frequency domain wavenumber spectrum is multiplied by the filter mask to retain the signal components that conform to the propagation law of gravity waves and filter out noise and non-gravity wave interference.

[0016] Wavefront elevation field mapping includes: performing a three-dimensional inverse fast Fourier transform on the filtered frequency domain wavenumber spectrum to obtain the wave surrogate quantity. Linear coefficients and bias constants are determined through offline calibration. The calibration method involves using measured wavefront elevation data from a shipborne wave sensor and performing a least-squares fit with the corresponding wave surrogate quantity, minimizing the sum of squares of their differences to obtain the optimal linear coefficients and bias constants. Finally, the wave surrogate quantity is mapped to the actual phase-resolved wavefront elevation field using a linear formula.

[0017] The Fourier neural operator prediction model includes: model training data composed of synthetic data generated by nonlinear wavefield numerical simulation combined with numerical radar imaging, with offline training. The training objective is to minimize the error between the predicted wavefront elevation field and the actual wavefront elevation field. During model deployment, only inference calculations are performed. The input is a historical frame sequence consisting of the most recent 20 frames of phase-resolved wavefront elevation fields, and the output is the prediction results for the next 10 steps, with each step corresponding to 0.1 seconds, adapting to scheduling cycle requirements. An edge graphics processor is used as the inference device to ensure real-time performance.

[0018] The radar acquisition parameters include: the operating frequency band is selected as 9.41 GHz, which has low propagation loss in the marine environment and is suitable for wave measurement; the range resolution is 8.5 meters and the frame rate is 2 Hz. This combination of parameters can control the amount of data while ensuring measurement accuracy and avoiding overloading edge computing resources.

[0019] Unified time reference and synchronization accuracy include: the time reference adopts a precise time protocol, and when the global navigation satellite system is available, the time is calibrated once every second to ensure that the system clock error is less than 1 microsecond; the synchronization accuracy of the linear interpolation timestamp is required to be controlled within 10 microseconds to meet the high-precision time synchronization requirements of phase-resolved wavefield reconstruction.

[0020] The rotation matrix construction includes: constructing the rotation matrix using the Euler angle transformation method in the order of Z, Y, X. First, rotate the yaw angle around the Z-axis of the radar local coordinate system, then rotate the pitch angle around the Y-axis, and finally rotate the roll angle around the X-axis. Each rotation angle corresponds to a rotation sub-matrix. The three sub-matrices are multiplied to obtain the final rotation matrix. For example, the yaw angle is 0.1 radians, the pitch angle is 0.05 radians, and the roll angle is 0.08 radians. The corresponding rotation matrix is ​​obtained by multiplying them in the above order.

[0021] Three-dimensional spatiotemporal spectrum analysis includes: a time window length of 2 seconds, corresponding to 20 frames of radar grid images; the number of spatial two-dimensional fast Fourier transform points is consistent with the radar grid image size, for example, 128 by 128; the number of temporal dimension fast Fourier transform points is 20; the overlap rate is 50%, and a Hanning window is used. This parameter combination can effectively extract the spatiotemporal frequency characteristics of waves while suppressing spectral leakage.

[0022] The specific structure of the Fourier neural operator model includes: an input layer with 20 channels, corresponding to 20 frames of historical wavefront elevation fields; an intermediate layer containing 3 Fourier neural operator blocks, each with a feature dimension of 64 and employing 16 frequency components; an output layer with 10 channels, corresponding to 10 steps of future prediction results; the activation function uses a modified linear unit, and the optimizer uses adaptive moment estimation. This structure can balance prediction accuracy and inference speed.

[0023] The calibration benchmarks for linear coefficients and offset constants include: a shipborne laser rangefinder is used as the calibration wave sensor, with a measurement accuracy of ±1 mm and a sampling rate of 10 Hz. The calibration process is conducted under three typical sea conditions: calm sea state, moderate wave sea state, and large wave sea state. Data is collected for 10 minutes in each scenario, and the linear coefficients and offset constants for each scenario are fitted. In practical applications, the corresponding parameters are selected according to the current sea state. For example, in calm sea state, the linear coefficient is set to 0.85, and the offset constant is set to 0.02 meters.

[0024] The future time window is preferably 1 second, corresponding to 10 prediction steps, each step being 0.1 seconds, consistent with the scheduling cycle. Therefore, a prediction window that is too short cannot provide effective prediction for resource scheduling, while a prediction window that is too long will lead to a decrease in prediction accuracy. A 1-second prediction window can achieve a balance between accuracy and practicality.

[0025] Preferably, the computing network resource orchestration controller determines the twin data coverage scale based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin, using the following specific calculation formula: ; ; in, To predict the distance for forward-looking waves, For the speed mentioned, The prediction time window; The coverage scale of the twin data represents the total number of spatial grid points contained in the forward-looking wave digital twin; The effective coverage width; This refers to the vertical resolution value within the grid resolution. This refers to the horizontal resolution value within the grid resolution. This indicates the rounding up operation.

[0026] The prediction time window is a fixed time length for the forward-looking wave digital twin to predict the future wave field, preferably 600 seconds. This is to reference the publicly available research context of ship wave prediction, balance the prediction range with computational feasibility, and meet the advance prediction needs of navigation decisions. The effective coverage width is the lateral width of the sea surface that the radar can effectively observe and use for wave field reconstruction, preferably 2000 meters. This is to combine the field of view of the shipborne X-band radar, avoiding data redundancy due to an excessively large observation range and affecting the integrity of the wave field if the observation range is too small. The grid resolution is the unit size when the forward-looking wave field is discretized into a spatial grid, including both longitudinal and lateral resolutions. A preferred value of 10 meters is used for both to match the common measurement resolution of X-band radar, balancing wave field detail representation with computational efficiency. The forward-looking wave prediction distance is the farthest distance of the wave field that can be predicted ahead of the ship. It is obtained by multiplying the ship's speed by the prediction time window and directly reflects the spatial extension range of the forward-looking wave field. The longitudinal grid point count is the number of grids in the forward-looking wave field along the ship's navigation direction. It is obtained by dividing the forward-looking wave prediction distance by the longitudinal resolution and rounding up to ensure a prediction distance with complete coverage. The number of horizontal grid points is the number of grid points in the forward-looking wave field perpendicular to the navigation direction. It is obtained by dividing the effective coverage width by the horizontal resolution and rounding up to ensure the effective observation width with complete coverage. The twin data coverage scale is the total number of spatial grid points contained in the forward-looking wave digital twin. It is obtained by multiplying the number of vertical grid points by the number of horizontal grid points and is the core indicator for quantifying the twin data volume.

[0027] The quantification of forward wave prediction distance includes: When a ship is sailing, the spatial range of the forward wave field changes with the speed. The faster the speed, the farther the distance of the sea surface in front needs to be covered in the same amount of time. Therefore, the method of multiplying the speed by the prediction time window is used to directly correlate the ship's motion state with the spatial range of the wave field. For example, if the speed is 15 meters per second and the prediction time window is 600 seconds, the forward wave prediction distance is 9000 meters. This logic fits the structural characteristics of the ship's forward twin and ensures that the prediction range matches the sailing state.

[0028] The number of grid points is rounded up, including: since the predicted distance of the forward wave and the effective coverage width are not necessarily integer multiples of the grid resolution, if conventional division is used for calculation, wave field data in some areas will be missing. Therefore, the number of grid points in both the vertical and horizontal directions is rounded up.

[0029] The coverage scale of twin data includes: The core of the forward wave digital twin is three-dimensional wave field data, and its data volume directly depends on the total number of spatial grid points. Therefore, the total number of grid points is defined as the coverage scale of twin data. This standard is directly related to the subsequent information generation rate and provides a clear basis for the quantification of minimum business throughput requirements.

[0030] The prediction time window is fixed at 600 seconds, or 10 minutes, which provides sufficient lead time for ship navigation control and does not reduce prediction accuracy due to excessive time, which is in line with the actual engineering application scenario.

[0031] Effective coverage width includes: a preferred value of 2000 meters, which corresponds to the effective sector width of the shipborne X-band radar. In engineering, this can be achieved by setting the radar's observation area of ​​interest (ROI) to ensure that the data collected by the radar can completely cover this width and avoid invalid or missing data due to exceeding the range.

[0032] The grid resolution, including both longitudinal and lateral resolution, is fixed at 10 meters. This scale is on the same order of magnitude as the range resolution of X-band radar, which is about 8.5 meters. This can accurately depict the three-dimensional shape of waves without causing a surge in the number of grid points due to excessively high resolution, thus avoiding increased computational and transmission pressure.

[0033] Preferably, the network resource orchestration controller determines the effective sampling update frequency for the forward-looking wave digital twin based on the flight speed, the relative wave encounter angle calculated from the course and the radar image, and the upper limit of the wave frequency band. The specific calculation formula is as follows: ; ; ; in, This is the upper limit of the wave frequency band. The maximum eigenfrequency, Pi is a constant. For the speed mentioned, The relative wave angle, It is the acceleration due to gravity. These are the deep-water wave terms derived from the deep-water gravity wave dispersion relation; The maximum wave encounter frequency represents the maximum wave excitation frequency after considering the Doppler effect; The effective sampling update frequency is determined based on the Nyquist sampling theorem.

[0034] The two-dimensional spectral analysis of the wavefront elevation field yields frequency domain energy distribution data obtained after performing a two-dimensional Fourier transform on the phase-resolved wavefront elevation field. This data is used to extract the spatial frequency characteristics of waves. The main wave propagation direction, determined by the energy peak position in the two-dimensional spectral analysis, is the direction in which wave energy is most concentrated and is a core parameter describing wave propagation characteristics. The relative encounter angle is the angle between the ship's heading and the main wave propagation direction, reflecting the relative motion between the ship and the wave and directly affecting the excitation effect of the wave on the ship. The time series of the wavefront elevation field is a set of wavefront elevation data collected at continuous time points, used to analyze the temporal variation of waves. The power spectral density analysis results are frequency-energy distribution data obtained after performing power spectral analysis on the wavefront elevation field time series, used to identify the main frequency components of waves. The cumulative energy quantile rule is a statistical rule for determining the upper limit of the wave frequency band, preferably 0.95, to ensure coverage of 95% of wave energy while eliminating high-frequency noise interference. The upper limit of the wave frequency band is the highest frequency containing the main wave energy and is the critical frequency for distinguishing effective wave signals from noise. The maximum intrinsic angular frequency (MIR) is the angular frequency corresponding to the upper limit of the wave frequency band, obtained by multiplying the upper limit of the wave frequency band by twice pi, reflecting the inherent time characteristics of waves. The deep-water wave number is a parameter describing the spatial periodicity of deep-water waves, obtained by dividing the square of the MIR by gravitational acceleration, conforming to the propagation law of deep-water gravity waves. The Doppler shift term is a correction term considering the influence of ship motion on the wave encounter frequency, obtained by multiplying the deep-water wave number, ship speed, and the cosine of the relative wave encounter angle, reflecting the change in wave observation frequency caused by ship motion. The maximum wave encounter angular frequency is the maximum wave excitation angular frequency actually encountered by the ship after considering the Doppler effect, obtained by taking the absolute value of the difference between the MIR and the Doppler shift term. The effective sampling update frequency is the minimum sampling frequency required to capture the dynamic characteristics of waves, obtained by dividing the maximum wave encounter angular frequency by pi, ensuring that the Nyquist sampling theorem is satisfied. Gravitational acceleration is the acceleration caused by Earth's gravity, with a value of 9.8 meters per second squared, a fundamental physical constant describing the propagation law of gravity waves. Pi is a fundamental constant in mathematics, with a value of 3.1416, and is used for conversion calculations between angular frequency and frequency.

[0035] Two-dimensional spectral analysis extracts the main propagation direction of waves, including: performing a two-dimensional fast Fourier transform on the phase-resolved wavefront elevation field to obtain the wavenumber spectrum in the spatial frequency domain, where the horizontal and vertical axes of the spectrum correspond to the wavenumbers in two horizontal directions, respectively.

[0036] The wavenumber spectrum is traversed to find the wavenumber point with the highest energy. The wavenumber direction corresponding to this point is the main propagation direction of the wave. For example, if the wavenumber vector corresponding to the energy peak in the wavenumber spectrum points 30 degrees east of north, then the main propagation direction of the wave is 30 degrees east of north. A Hanning window is used to suppress spectral leakage during the analysis, and the number of points in the two-dimensional Fourier transform is consistent with the grid size of the wavefront elevation field to ensure analysis accuracy.

[0037] Power spectral density analysis combined with quantiles determines the upper limit of the wave band. This involves: first, preprocessing the time series of the wavefront elevation field to remove the mean and linear trend; then, using the Welch method to estimate the power spectral density with a 2-second window, 50% overlap, and 256 Fourier transform points; finally, calculating the cumulative energy of the power spectrum and finding the frequency corresponding to 95% of the total energy, which is the upper limit of the wave band.

[0038] The deep-water wave number is derived based on the dispersion relation of deep-water gravity waves. This involves the following: the dispersion relation of deep-water gravity waves is that the square of the angular frequency equals the product of gravitational acceleration and the wave number. Based on this physical law, a mathematical transformation is performed, resulting in the wave number being equal to the square of the angular frequency divided by gravitational acceleration. Substituting the maximum eigenfrequency into this formula yields the deep-water wave number.

[0039] The calculation of the maximum wave encounter frequency (MWF) considering the Doppler effect includes the following: During ship movement, the observed wave frequency shifts due to the Doppler effect; the frequency increases when facing the wave and decreases when with the wave. Combining the definitions of deep-water wave number and the Doppler shift term, the MWF is equal to the absolute value of the difference between the maximum intrinsic angular frequency and the Doppler shift term.

[0040] The effective sampling update frequency is derived based on the Nyquist sampling theorem, including the following: The Nyquist sampling theorem requires that the sampling frequency be no less than twice the highest frequency of the signal to avoid signal aliasing. Since the maximum encounter angle frequency is the highest angular frequency of the wave, and the corresponding frequency is the maximum encounter angle frequency divided by twice pi, the effective sampling update frequency must be no less than twice this frequency, i.e., the maximum encounter angle frequency divided by pi.

[0041] Two-dimensional spectral analysis includes: employing a two-dimensional fast Fourier transform algorithm, with the number of transform points matching the grid size of the wavefront elevation field; for example, if the wavefront elevation field grid is 128x128, then the number of transform points is 128x128; using the Hanning window function to suppress spectral leakage; and zero-padding the wavefront elevation field data before analysis, resulting in a size of 256x256 to improve spectral resolution and ensure accurate extraction of the main wave propagation direction.

[0042] The specific quantile values ​​for the cumulative energy quantile rule include: preferably 0.95. This value is a commonly used standard for extracting the main wave components in marine engineering, which can cover the vast majority of effective wave energy, while avoiding the inclusion of high-frequency noise in the frequency band and ensuring the rationality of the upper limit of the wave frequency band.

[0043] The minimum length requirement for wavefront elevation field time series includes: a minimum length of 2 seconds, which ensures that the frequency resolution of power spectral density analysis is not less than 0.5 Hz, which is sufficient to identify the main frequency components of ocean waves. For example, the main frequency of waves in medium wave states is between 0.5 and 1.5 Hz, and a 2-second time series can accurately capture the frequency characteristics in this range.

[0044] The normalization process for the relative encounter angle includes: normalization using modular arithmetic to map the calculated relative encounter angle to the range of -3.1416 to 3.1416. Specifically, when the angle is greater than 3.1416, subtract twice the value of pi; when the angle is less than -3.1416, add twice the value of pi. For example, if the calculated angle is 4.0 radians, after normalization, it is 4.0 minus 6.2832, resulting in -2.2832 radians, ensuring a consistent angle range.

[0045] Preferably, the computing network resource orchestration controller calculates a flight condition-related load factor that characterizes the non-linear growth of twin service demands with flight speed based on the product relationship between the twin data coverage scale and the effective sampling update frequency. The specific calculation formula is as follows: ; in, The load factor associated with the navigation condition represents the quadratic coefficient of the minimum information generation rate of the twin service with respect to the speed. The pre-defined single-grid-point encoding load represents the number of data bits required for a single grid point update in the forward-looking wave digital twin; The prediction time window, The effective coverage width, This refers to the vertical resolution value within the grid resolution. This refers to the horizontal resolution value within the grid resolution. Pi is a constant, and gravitational acceleration is π / 2. The relative wave angle, The maximum eigenfrequency, This represents the nonlinear contribution of wave frequency band energy distribution to the refresh frequency requirement.

[0046] The single-grid point coding load is the number of bits of data required for a single update of a spatial grid point in the forward-looking wave digital twin, preferably 48 bits, to cover necessary fields such as sea surface elevation, phase derivative, confidence level, checksum, and timestamp increment, balancing data integrity and transmission efficiency. The numerator is the part used to calculate the basic structure constant, obtained by multiplying the single-grid point coding load, prediction time window, and effective coverage width, reflecting the impact of the twin's spatial characteristics on data requirements. The denominator is the part used to calculate the basic structure constant, obtained by multiplying the constant of pi, gravitational acceleration, longitudinal resolution, and lateral resolution, correcting for the constraints of physical laws and grid settings on data requirements. The basic structure constant is an intermediate parameter reflecting the inherent structural characteristics of the twin, obtained by dividing the numerator by the denominator, providing a basic calculation benchmark for the navigation condition-related load factor. The absolute value of the cosine of the relative wave encounter angle is the non-negative value after cosine operation of the relative wave encounter angle, eliminating the directional influence of the wave encounter direction on the calculation results and retaining only the magnitude of the influence. The square of the maximum intrinsic angular frequency is the result of multiplying the maximum intrinsic angular frequency by itself, reflecting the nonlinear contribution of wave frequency band energy distribution to the data update frequency requirement. The navigation condition-related load factor is a core coefficient characterizing the minimum information generation rate of twin operations with respect to the second growth law of speed. It integrates spatial characteristics, physical laws, and sea state parameters, directly linking operational requirements with navigation status.

[0047] The single-grid point encoded payload fields include: The field comprises four core components: sea surface elevation stored at 16-bit specific points, covering a range of ±8 meters with a resolution of 0.25 millimeters, meeting the accuracy requirements for wave elevation measurement; phase derivative stored at 16-bit specific points to capture the wave's time-varying rate of change; confidence level using 8 bits to quantify data reliability; and checksum and timestamp increment using 8 bits to ensure data transmission integrity and time synchronization. These four components total 48 bits. This design ensures the integrity of necessary information while avoiding redundant fields that increase data volume. For example, a grid point with a sea surface elevation of 2.35 meters, a phase derivative of 0.12 meters per second, a confidence level of 200, and a timestamp increment of 10 corresponds to 48 bits of encoded data.

[0048] The construction of the numerator and denominator includes: In the numerator, the single-grid point coding load determines the data volume of a single grid, the prediction time window and effective coverage width determine the spatial range, and the product of these three directly relates to the spatial scale of the twin and the data volume of a single grid; In the denominator, the constant pi is used to unify the calculation dimension, gravitational acceleration reflects the physical laws of waves, and vertical and horizontal resolutions reflect the degree of grid discretization, and the product of these four corrects the impact of physical constraints and discretization processing on data requirements. This logic integrates spatial characteristics and physical laws into computable fundamental constants through the ratio of the numerator and denominator.

[0049] The navigation condition-related load factors include: the basic structure constant integrating the inherent structural characteristics of the twin; the absolute value of the cosine of the relative wave encounter angle reflecting the relative attitude influence between the ship and the waves; and the square of the maximum intrinsic angular frequency reflecting the energy contribution of the wave frequency band. The coefficient obtained by multiplying the three is essentially the quadratic coefficient of the twin business information generation rate with respect to the speed.

[0050] The specific bit allocation for the single-grid point coding load includes: 16 bits for sea surface elevation, using Q-format encoding, with 1 sign bit, 4 integer bits, and 11 decimal bits, covering a range of ±8 meters and a resolution of 0.25 millimeters; 16 bits for phase derivative, with the same encoding format as sea surface elevation, covering a range of ±8 meters per second; 8 bits for confidence, using an unsigned integer ranging from 0 to 255, where 255 indicates that the data is completely reliable; and 8 bits for checksum and timestamp increment, with 4 bits used for checksum and 4 bits used for timestamp increment. This allocation scheme ensures that the precision of each field matches the actual requirements.

[0051] The offline calibration and verification of the basic structural constants includes: After the basic structural constants are calculated, they need to be verified through offline calibration. The calibration method is to select three typical sea states: calm sea state, medium wave sea state, and large wave sea state. Data is collected for 10 minutes under each sea state, the actual information generation rate at different speeds is calculated, and compared with the theoretical value derived based on the basic structural constants. Parameters such as resolution are adjusted to ensure that the error between the theoretical value and the actual value is less than 5%, thus ensuring the accuracy of the constants.

[0052] The numerical range and typical values ​​of the navigation condition-related load factor include: the value of this factor varies under different sea states and navigation attitudes. In calm sea states, the relative cosine of the angle of encounter is 0.3, the square of the maximum intrinsic angular frequency is 20, and the typical value is about 5000. In moderate sea states, the relative cosine of the angle of encounter is 0.6, the square of the maximum intrinsic angular frequency is 30, and the typical value is about 12000. In rough sea states, the relative cosine of the angle of encounter is 0.9, the square of the maximum intrinsic angular frequency is 50, and the typical value is about 22500. This range provides a reference for resource orchestration and facilitates the assessment of the intensity of demand growth.

[0053] Preferably, the multi-dimensional resource adaptation model, which includes end-to-end latency, energy consumption, and cost, specifically includes: The end-to-end delay model is calculated using the following formula: ; ; ; in, This refers to the end-to-end delay; Baseline latency of the basic protocol; Allocate computing power to edge computing nodes. Configure network slice bandwidth; For twin computing task workloads, The scale of the twin data coverage, The cost constant for wave field processing per unit grid is [missing information]. To solve for the cost constant in a coupled manner, To couple the solution of the state dimension, To determine the number of iterations for coupled solution; This represents the minimum data volume per cycle. This is the minimum service throughput requirement. The scheduling period; The effective transmission efficiency coefficient; The energy consumption model is calculated using the following formula: ; in, The energy consumption is as described; The energy consumption coefficient of the edge node. Energy consumption coefficient per unit bandwidth; The cost model and its calculation formula are as follows: ; in, The cost; This refers to the unit price for slice rental. This is the unit price for computing power billing.

[0054] End-to-end latency is the total time taken for a twin service from data acquisition to command issuance. It is calculated by summing the baseline latency of the basic protocol, the computational processing latency, and the network transmission latency, and is a core indicator for measuring system real-time performance. The baseline latency of the basic protocol is the inherent latency of the communication protocol and fixed processing flow, preferably 10 milliseconds, to reference the typical protocol overhead of 5G and edge computing, ensuring that the value closely matches engineering realities. The computational processing latency is the time taken for edge nodes to process twin computing tasks, calculated by dividing the twin computing task load by the allocated computing power of the edge computing nodes, reflecting the impact of computing power configuration on processing speed. The network transmission latency is the time taken for twin data to be transmitted in the network, calculated by dividing the minimum data volume per cycle by the product of the effective transmission efficiency coefficient and the network slice bandwidth configuration, reflecting the constraint of bandwidth on transmission speed. The twin computing task load is the computational load required to process the forward-looking wave digital twin, calculated by summing the wave field processing computational load and the coupled solution computational load, and is linearly proportional to the twin data coverage scale. The unit grid wavefield processing computational cost constant represents the computational overhead of wavefield reconstruction and prediction at a single grid point. Its preferred value is calibrated through offline benchmark testing, based on the hardware performance of the target edge nodes to ensure accurate quantification of computational load. The coupled solution computational cost constant represents the computational overhead of a single seakeeping-maneuvering coupled solution. Its preferred value is calibrated through offline benchmark testing, based on the state dimension and iteration count of the coupled solution to match actual computational requirements. The coupled solution state dimension is the number of variables in the seakeeping-maneuvering coupled solution, preferably 12, to include six degrees of freedom motion and additional states, covering core variables of ship dynamics. The coupled solution iteration count is the number of iteration steps in the coupled solution algorithm, preferably 5, to balance accuracy and computational efficiency, ensuring reliable results and controllable time consumption. The edge computing node allocated computing power is the computing capacity allocated to the edge nodes for the twin task, measured in cycles per second, reflecting the intensity of processing resources available from the nodes. The minimum service throughput requirement is the minimum data transmission rate required to maintain spatiotemporal synchronization of the twin, obtained by summing linear and nonlinear demand components. The scheduling cycle is a fixed time interval for resource orchestration, preferably 0.1 seconds, to match the spatiotemporal synchronization error requirements of the twin and ensure timely dynamic resource adaptation. The minimum data volume per cycle is the amount of twin data that must be transmitted within each scheduling cycle, obtained by multiplying the minimum service throughput requirement by the scheduling cycle, quantifying the transmission pressure per cycle. The effective transmission efficiency coefficient is the ratio of the actual transmission rate to the theoretical bandwidth, which can be measured using a 1-second sliding window by statistically analyzing the ratio of the number of bytes successfully delivered by the application layer to the theoretical number of bytes transmitted. Network slice bandwidth configuration is the dedicated network bandwidth allocated for twin services and is a key resource parameter for ensuring transmission rate. Energy consumption is the energy consumed during system operation, obtained by adding computational and communication energy consumption, and is an important optimization target for resource adaptation. Computational energy consumption is the energy consumed by edge nodes processing twin tasks, proportional to the square of the allocated computing power of the edge computing node and the scheduling cycle, conforming to the energy consumption characteristics of dynamic voltage and frequency adjustment.Communication energy consumption refers to the energy consumed in transmitting twin data over a network. It is directly proportional to the network slice bandwidth configuration and scheduling cycle, aligning with the energy consumption patterns of wireless transmission. The edge node energy consumption coefficient is an energy consumption characteristic parameter of edge nodes. Preferred values ​​are obtained through offline fitting, based on measured power consumption at different computing power levels to ensure accurate energy consumption calculations. The unit bandwidth energy consumption coefficient is a transmission energy consumption parameter per unit bandwidth. Preferred values ​​are obtained through offline measurement, based on transmission power consumption under different bandwidth configurations, matching actual network energy consumption characteristics. Cost is the economic overhead of system operation, obtained by adding slice resource costs and computing power resource costs. It is an economic constraint indicator for resource adaptation. Slice resource cost is the fee for using network slices, obtained by multiplying the slice rental unit price, network slice bandwidth configuration, and scheduling cycle, reflecting the economic cost of slice resources. Computing power resource cost is the fee for using edge computing power, obtained by multiplying the computing power billing unit price, edge computing node allocation computing power, and scheduling cycle, reflecting the economic overhead of computing power resources. The slice rental price is the cost per second of bandwidth used in a slice. Ideally, it should be a fixed price agreed upon in the contract, based on the operator's slice service pricing standards to ensure cost calculation compliance. The computing power billing price is the cost per second of computing power used. Ideally, it should be based on internal billing standards or cloud service provider pricing, balancing economic efficiency and rationality between the hardware and operating costs of computing power resources.

[0055] The twin computing workload consists of two parts: the wavefield processing computation is linearly related to the twin data coverage scale and is obtained by multiplying the unit grid wavefield processing computation cost constant by the total number of grid points. For example, if the total number of grid points is 100,000 and the unit cost constant is 1,000 cycles, then the wavefield processing computation is 1e8 cycles. The coupled solution computation is obtained by multiplying the coupled solution computation cost constant, the state dimension, and the number of iterations.

[0056] The superposition relationship of the end-to-end latency model includes: the baseline latency of the basic protocol is a fixed overhead, for example, the initialization processing of the 5G control plane protocol and the edge node takes 10 milliseconds; the computation processing latency is inversely proportional to the computing power, for example, if the computation load is 1e8 cycles and the computing power is 1e9 cycles per second, then the processing latency is 0.1 seconds; the network transmission latency is inversely proportional to the bandwidth.

[0057] The energy consumption calculation model is a quadratic proportional model, which shows that the energy consumption of edge nodes is a quadratic function of computing power. This model conforms to the energy consumption characteristics of dynamic voltage and frequency adjustment technology. For example, the energy consumption is 20 watts when the computing power is 5e8 cycles per second and 60 watts when the computing power is 1e9 cycles per second, rather than increasing linearly. This model fits the hardware energy consumption law and ensures the authenticity of energy consumption calculation.

[0058] The linear proportional model of communication energy consumption includes: network transmission energy consumption is approximately linearly related to bandwidth. The higher the bandwidth, the greater the energy consumption for transmitting the same amount of data. For example, the transmission energy consumption is 5 watts when the bandwidth is 5 megabits per second, and 9.5 watts when the bandwidth is 10 megabits per second. This model balances accuracy and computational complexity and is suitable for engineering applications.

[0059] The cost model is billed independently, including: the cost of slice resources is billed based on bandwidth, which is in line with the charging model of network slices of operators; the cost of computing power resources is billed based on computing power, which matches the computing power pricing standards of cloud service providers. The two are calculated independently and then added together to fully reflect the economic expenses of system operation.

[0060] The baseline latency of the basic protocol includes: a fixed value of 10 milliseconds. This value includes 3 milliseconds of 5G user plane protocol overhead, 4 milliseconds of edge node data reception and parsing, and 3 milliseconds of instruction encapsulation and delivery. The time consumption of each step is based on actual engineering scenarios to ensure that the value is reasonable.

[0061] The effective transmission efficiency coefficient is measured by using a 1-second sliding window, counting the number of bytes successfully delivered by the application layer within the window, multiplying by 8 to convert to bits, and dividing by the product of the network slice bandwidth configuration and the window duration to obtain the effective transmission efficiency coefficient.

[0062] The calibration of the unit grid wave field processing computation cost constant includes: selecting wave field data of 10,000 grid points on the target edge node, running a complete wave field reconstruction and Fourier neural operator prediction, recording the time taken as 100 milliseconds, and the node physical computing power as 1e9 cycles per second. Then the unit cost constant is (1e9×0.1)÷1e4=1e4 cycles. After calibration, it is fixed and written into the system.

[0063] The calibration of the cost constant for coupled solution calculation includes: on the target edge node, with a fixed state dimension of 12 and 5 iterations, run one Nebula-manipulation coupled solution and record the time taken as 50 milliseconds. Then the unit cost constant is (1e9×0.05)÷(12×5)≈8.3e5 cycles, ensuring that it matches the actual computational overhead.

[0064] The fitting of the edge node energy consumption coefficient includes: setting five different computing power levels on the edge node, 5e8, 6e8, 8e8, 1e9, and 1.2e9 cycles per second, measuring the power consumption of each level, and using a quadratic function to fit the energy consumption coefficient.

[0065] The measurement of energy consumption coefficient per unit bandwidth includes: measuring the transmission power consumption of different bandwidth configurations under three typical topologies: 5G direct connection, ship-to-ship multi-hop, and satellite backup, and taking the average value as the energy consumption coefficient per unit bandwidth.

[0066] The pricing of slice leasing unit price and computing power billing unit price includes: the slice leasing unit price refers to the public quotation of 5G industry slices of operators, and the computing power billing unit price refers to the edge computing power pricing of mainstream cloud service providers. Combined with the real-time needs of the business, medium and high priority pricing is selected to ensure that the cost calculation is in line with the actual market.

[0067] Preferably, the network resource orchestration controller uses the navigation condition-related load factor to determine the minimum service throughput requirement to meet the spatiotemporal synchronization requirements, and uses the minimum service throughput requirement as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption, and cost. The specific calculation formula is as follows: Calculate the minimum traffic throughput requirement: ; in, This refers to the minimum service throughput requirement; The speed is denoted as ; the load factor associated with the navigation condition is denoted as . The coefficient for the first-order term is only related to the coverage scale of the twin data and the upper limit of the wave frequency band; Generate dynamic weights and construct a comprehensive adaptation objective function: ; ; ; in, For delay weighting, As energy consumption weight, Cost weighting; As a sensitivity adjustment constant, It is a natural exponential function; The comprehensive adaptation objective function; , , These are the normalized values ​​for the end-to-end latency model, the energy consumption model, and the cost model, respectively; the weight configuration ensures that when the speed causes the load factor associated with the navigation condition to increase, the optimization direction of the multi-dimensional resource adaptation model automatically tilts towards reducing latency.

[0068] The linear term coefficient is a fundamental parameter for calculating the linear demand component. It is related to the single-grid point coding load and the forward-looking spatial domain size. Preferably, it is calculated by multiplying the single-grid point coding load by the forward-looking spatial domain size correlation coefficient to match the linear contribution of the upper limit of the wave frequency band to the demand, ensuring accurate calculation of the linear component. The linear demand component is the portion of the minimum service throughput demand that increases linearly with speed. It is obtained by multiplying the linear term coefficient by the speed, reflecting the fundamental impact of speed on service demand. The nonlinear demand component is the portion of the minimum service throughput demand that increases with the square of the speed. It is obtained by multiplying the navigation condition-related load factor by the square of the speed, reflecting the amplification effect of the square of the speed on service demand. The minimum service throughput demand is the minimum data transmission rate required to maintain twin spatiotemporal synchronization. It is obtained by adding the linear and nonlinear demand components and is the core demand benchmark for resource adaptation. The sensitivity adjustment constant is a parameter that controls the sensitivity of delay weight changes. A value of 0.1 is preferred to balance the smoothness of weight adjustment with the timeliness of condition response, avoiding abrupt weight changes. The latency priority factor is an intermediate variable that dynamically adjusts the latency weights. It is obtained by multiplying the sensitivity adjustment constant and the load factor associated with the flight conditions using a natural exponential function, and increases with the load factor. The latency weight is a weighting coefficient for latency in the comprehensive adaptation objective function, and its value increases with the load factor associated with the flight conditions, ensuring latency priority in high-demand scenarios. The energy consumption weight is a weighting coefficient for energy consumption in the comprehensive adaptation objective function, and its value decreases with the load factor associated with the flight conditions, balancing the optimization priorities under different scenarios. The cost weight is a weighting coefficient for cost in the comprehensive adaptation objective function, and its value decreases with the load factor associated with the flight conditions, consistent with the trend of the energy consumption weight. The normalized latency value is latency data that eliminates the influence of dimensions, obtained by dividing the end-to-end latency by a preset latency benchmark, ensuring dimensional uniformity in multi-objective optimization. The normalized energy consumption value is energy consumption data that eliminates the influence of dimensions, obtained by dividing the system energy consumption by the energy consumption reference scale, adapting to multi-objective weighted calculations. The normalized cost value is the cost data free from the influence of dimensions. It is obtained by dividing the system cost by the cost reference scale, ensuring the rationality of multi-objective optimization. The energy consumption reference scale is the benchmark value for energy consumption normalization, preferably the statistical value of system energy consumption under typical operating conditions to cover common sea states and navigation conditions, ensuring the stability of the normalization effect. The cost reference scale is the benchmark value for cost normalization, preferably the statistical value of system cost under typical operating conditions. Its value is determined based on the same criteria as the energy consumption reference scale, maintaining the consistency of the normalization logic. The comprehensive adaptation objective function is the optimization objective of resource adaptation. It is obtained by weighted summation of the time delay weight, energy consumption weight, and cost weight with their corresponding normalized values, guiding resource allocation decisions.

[0069] The minimum throughput requirements include: the linear demand component, which is derived from the linear increase in the size of the forward-looking space domain with the speed of the aircraft. For example, if the speed of the aircraft is 10 meters per second and the coefficient of the first term is 500, then the linear component is 5000 bits per second; the nonlinear demand component is derived from the coupling effect between the wave encounter frequency and the speed of the aircraft.

[0070] The generation of dynamic weights includes: constructing a latency priority factor based on the natural exponential function. For example, if the load factor is 1000 and the sensitivity constant is 0.1, then the priority factor is a natural exponent of 100, the latency weight is 100 divided by 102, approximately 0.98, and the energy consumption and cost weights are each 1 divided by 102, approximately 0.01. In this case, the optimization focus is on latency. If the load factor is 100 and the priority factor is a natural exponent of 10, the latency weight is approximately 0.91, and the energy consumption and cost weights are approximately 0.04. The weights are dynamically adjusted to adapt to changes in operating conditions.

[0071] The construction of the comprehensive adaptation objective function includes: first, normalizing the latency, energy consumption, and cost to eliminate dimensional differences. For example, if the actual latency is 0.08 seconds and the baseline is 0.1 seconds, the normalized value is 0.8 seconds; then, the function is weighted and summed according to dynamic weights.

[0072] The first-order coefficients include: the first-order coefficients are equal to the single-grid point coding load multiplied by the upper limit of the wave frequency band, and then divided by the relevant structural parameters.

[0073] Sensitivity adjustment constants include: a preferred value of 0.1. At this value, when the load factor increases from 100 to 1000, the delay weight smoothly increases from 0.91 to 0.98 without abrupt changes, which responds to changes in operating conditions and avoids drastic fluctuations in the optimization direction.

[0074] The determination of energy consumption and cost reference scales includes: selecting three sea states: calm, medium waves, and large waves; selecting five typical sailing speeds under each sea state; collecting 10 minutes of system energy consumption and cost data; and taking the average value as the reference scale. For example, the energy consumption reference scale is 50 watts, and the cost reference scale is 0.01 yuan per second, to ensure that normalization covers the main application scenarios.

[0075] Preferably, the network resource orchestration controller uses a swarm intelligence optimization algorithm to solve the multi-dimensional resource adaptation model, generates network slice bandwidth configuration instructions and edge computing node allocation instructions, and issues them out. The specific calculation formula and execution steps are as follows: Define the particle position vector: ; in, The particle position vector; Configure network slice bandwidth; Index continuous variables for edge computing nodes; Allocate computing power to candidate edge computing nodes; Constructing the fitness function and node mapping: ; ; ; in, Calculate node identifiers for the mapped discrete target edges. For floor operations, To correct for trace amounts; Let be the penalty function. The penalty coefficient is... This is the calculated value of the end-to-end delay model at the current particle position. The preset spatiotemporal synchronization threshold; For the fitness function, The comprehensive adaptation objective function; The global optimal particle position is obtained by iteratively solving using the particle swarm optimization algorithm. And execute the following instruction mapping and distribution: ; ; in, To ensure bit rate, The maximum bit rate is used to configure network slicing; Allocate execution cycle quotas for container processors. The total physical computing power of the target edge computing node is used to configure edge computing container resources.

[0076] The particle position vector is a multi-dimensional data structure in the particle swarm optimization algorithm that describes candidate solutions. It includes network slice bandwidth configuration, continuous variables for edge computing node indices, and allocated computing power for candidate edge computing nodes, comprehensively covering the core decision variables for resource adaptation. The continuous variables for edge computing node indices are continuous values ​​used to map discrete node identifiers, ranging from 1 to the total number of candidate nodes plus 1, providing a branchless deterministic mapping basis for node selection. The candidate edge computing node set is the total number of edge nodes that can participate in resource allocation, preferably 5 nodes, to balance the diversity of resource selection with algorithm solution efficiency and avoid computational complexity caused by too many nodes. The allocated computing power of candidate edge computing nodes is the computing power that each candidate node can provide to the twin task, measured in cycles per second, reflecting the node's processing resource potential. The correction increment is a tiny value to avoid ambiguity in the node index mapping, preferably 10 to the power of -9, to ensure that the continuous variable maps to a unique discrete node, eliminating boundary judgment errors. The target edge computing node identifier is the finally selected edge node number, obtained by rounding down the corrected continuous variables for edge computing node indices, clearly defining the target object for resource allocation. The penalty function is a penalty term that strengthens the delay constraint. It outputs a penalty value when the end-to-end delay exceeds the spatiotemporal synchronization threshold, otherwise outputting zero, ensuring that the delay constraint is strictly satisfied. The penalty coefficient is a parameter controlling the strength of the penalty function, preferably 10 to the power of 6, to significantly increase the fitness of candidate solutions with excessive delay, thus eliminating them from the algorithm. The spatiotemporal synchronization threshold is the maximum allowable end-to-end delay, preferably 0.1 seconds, to match the spatiotemporal synchronization error requirements of twins and ensure system real-time performance. The fitness function is the evaluation criterion of the particle swarm optimization algorithm, obtained by adding the comprehensive fitness objective function and the penalty function, guiding the algorithm to find the optimal solution that satisfies the constraints. The particle swarm optimization algorithm parameters are configuration items for the algorithm's operation, including the number of particles, number of iterations, inertia weight, learning factor, and random seed. Preferred values ​​are 16, 8, 0.6, 1.2, 1.2, and 20260121, respectively, to balance solution accuracy and real-time performance, adapting to a 0.1-second scheduling cycle. The globally optimal particle position is the optimal candidate solution obtained through iterative processing using the particle swarm optimization algorithm. It includes the optimal network slice bandwidth configuration, the target edge node identifier, and the node's allocated computing power. The guaranteed bit rate is the minimum transmission rate provided by the network slice for twin services, equal to the optimal network slice bandwidth configuration, ensuring the minimum requirements for service transmission. The maximum bit rate is the highest allowed transmission rate of the network slice, equal to 1.2 times the guaranteed bit rate, to prevent network congestion caused by sudden service surges. The total physical computing power of the target edge computing node is the maximum hardware computing capability of the target node, obtainable by querying node hardware parameters, including CPU clock speed, number of cores, and utilization limit, providing a benchmark for computing power allocation. The container processor's execution cycle quota is the proportion of CPU resources allocated to the twin task containers, obtained by dividing the target edge computing node's allocated computing power by the node's total physical computing power.

[0077] The construction of particle position vectors includes: the vector simultaneously contains network slice bandwidth configuration, continuous variable node index, and node allocated computing power. These three correspond to the three core decision dimensions of network resources, node selection, and computing resources, respectively. For example, the vectors are 10 megabits per second, 3.2, 5e8 cycles per second, 2e8 cycles per second, and 3e8 cycles per second. This not only covers all the elements of resource adaptation but also provides the algorithm with a continuous and optimizable decision space, adapting to the needs of joint scheduling of computing networks.

[0078] The deterministic mapping of node indices includes: the continuous variable of the edge computing node index takes values ​​from 1 to the total number of candidate nodes plus 1. For example, if there are 5 candidate nodes, the variable takes values ​​from 1 to 6. After subtracting the correction amount, the values ​​are rounded down to obtain discrete node identifiers from 1 to 5.

[0079] The penalty function and fitness function are as follows: The penalty function adopts a quadratic penalty form. The more the delay exceeds the limit, the larger the penalty value. For example, if the delay is 0.12 seconds, the threshold is 0.1 seconds, and the penalty coefficient is 1e6, then the penalty value is 1e6×(0.12-0.1)²=400. The fitness function is the comprehensive adaptation objective function value plus 400, so that the fitness of the out-of-limit solution is significantly higher than that of the compliant solution, and it is quickly eliminated by the algorithm to ensure the rigidity of the delay constraint.

[0080] The instruction mapping between network slicing and computing power includes: network slice bandwidth configuration mapping to a guaranteed bit rate and a maximum bit rate, with a ratio of 1.2 to meet sudden business demands while avoiding network resource waste, for example, a guaranteed bit rate of 10 megabits per second and a maximum bit rate of 12 megabits per second; computing power allocation mapping to container CPU execution cycle quota, for example, if the total physical computing power of a node is 1e9 cycles per second and the allocated computing power is 5e8 cycles per second, then the quota is 0.5, corresponding to 50% of the CPU resources. This mapping directly connects to the network slice management and MEC resource scheduling interface to ensure that instructions can be implemented.

[0081] The hot-start mechanism of the particle swarm optimization algorithm includes: using the global optimal particle position of the previous cycle as the particle initialization center of the current cycle, and adding a random perturbation of a fixed amplitude. For example, if the optimal bandwidth of the previous cycle is 10 megabits per second and the perturbation amplitude is 0.5 megabits per second, then the particle initialization bandwidth of the current cycle is between 9.5 and 10.5 megabits per second, which shortens the convergence time of the algorithm and adapts to the real-time requirements of a 0.1-second scheduling cycle.

[0082] The size and selection criteria of the candidate edge computing node set include: prioritizing 5 nodes, with the selection criteria being the 5 edge nodes closest to the ship's current location, including shore-based edge nodes and shipborne edge nodes, to ensure controllable network transmission latency and node coverage of different topology scenarios.

[0083] The specific values ​​and selection of micro-values ​​are corrected, including: prioritizing values ​​of 10 to the power of -9, which are much smaller than the minimum interval of continuous variables, such as the node index interval of 1, so as not to affect the rounding result, and to avoid boundary ambiguity when rounding continuous variables.

[0084] The penalty coefficient includes a preferred value of 1e6. When the time delay exceeds the limit by 0.01 seconds, the penalty value is 100, which significantly increases the fitness function value. The algorithm will prioritize the solution with compliant time delay and will not get stuck in a local optimum due to an excessively large penalty coefficient.

[0085] The detailed configuration of the particle swarm optimization algorithm parameters includes: 16 particles, 8 iterations (this configuration takes about 20 milliseconds to run on an edge CPU, meeting the 0.1-second scheduling cycle); inertia weight of 0.6, balancing the algorithm's exploration and exploitation capabilities; and a learning factor of 1.2, guiding particles toward both individual and global optima.

[0086] The configuration of container CPU execution cycle quota includes: it is implemented using Linux container group (CGROUPS) technology. The quota corresponds to the cpu.max parameter. For example, if the quota is 0.5, then cpu.max is set to 50000, which means 50% of the CPU time. The configuration is issued through the MEC orchestrator interface and takes effect in real time without restarting the container.

[0087] The network slice configuration command distribution interface includes: interfacing with the network slice management interface defined by 3GPP, adopting the RESTful (representational state transfer) protocol, and the command includes 5QI, guaranteed bit rate, maximum bit rate, ARP, DSCP and other parameters. The core network slice management unit completes the configuration within 10 milliseconds after receiving the command to ensure the real-time nature of the command distribution.

[0088] The interaction of edge computing node allocation instructions includes: the MEC orchestrator communicates with the target edge node through the Mm5 reference point interface. The instructions include container identifiers and execution cycle quotas. After receiving the instructions, the node resource management unit adjusts the container resources through CGROUPS. The entire process takes less than 10 milliseconds, which is suitable for scheduling cycle requirements.

[0089] Example 2: A digital twin-based intelligent ship management method, implementing the digital twin-based intelligent ship management system as described in any one of the examples, including: Collect radar images, speed, heading, and attitude data of the ship, and reconstruct a forward-looking wave digital twin based on the radar images; Resource orchestration is performed within a scheduling period, and the resource orchestration includes: The coverage scale of the twin data is determined based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin. Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined; Based on the product relationship between the twin data coverage scale and the effective sampling update frequency, calculate the navigation condition-related load factor that characterizes the non-linear growth of twin service demand with the speed. The minimum service throughput requirement to meet the spatiotemporal synchronization requirement is determined by using the associated load factor of the navigation condition, and the minimum service throughput requirement is used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption and cost. The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions.

[0090] like Figure 2 As shown, Figure 2 The architecture of a ship intelligent management system based on digital twins was demonstrated: the main vessel is equipped with shipborne radar, shipborne edge computing equipment and orchestration controller. The shipborne radar collects data to generate a forward-looking wave twin, and at the same time obtains speed, heading and attitude information. The main vessel connects to a 5G base station through wireless communication with the help of a multi-hop relay ship, and then connects to shore-based edge nodes and core network slice management units to realize the coordination of shipborne edge computing, multi-hop communication, shore-based resources and core network slices to support intelligent management related to ship digital twins.

[0091] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A ship intelligent management system based on digital twins, characterized in that, include: The shipborne sensing and modeling device is configured to collect radar images, speed, heading and attitude data of the ship, and reconstruct a forward-looking wave digital twin based on the radar images. The network resource orchestration controller is configured to perform resource orchestration within a scheduling cycle, wherein the resource orchestration includes: The coverage scale of the twin data is determined based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin. Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined; Based on the product relationship between the twin data coverage scale and the effective sampling update frequency, calculate the navigation condition-related load factor that characterizes the non-linear growth of twin service demand with the speed. The minimum service throughput requirement to meet the spatiotemporal synchronization requirement is determined by using the associated load factor of the navigation condition, and the minimum service throughput requirement is used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption and cost. The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions.

2. The ship intelligent management system based on digital twins according to claim 1, characterized in that, Acquiring radar images, speed, heading, and attitude data of the ship, and reconstructing a forward-looking wave digital twin based on the radar images, including: The shipborne sensing and modeling device performs linear interpolation alignment of the acquired radar images of non-same-source frequencies with the attitude data using a unified time reference; it then uses the attitude data to construct a rotation matrix, transforms the radar images from the radar local coordinate system and projects them onto the ground-fixed local horizontal coordinate system, thereby obtaining a motion-compensated radar grid image sequence. Three-dimensional spatiotemporal spectrum analysis is performed on the radar grid image sequence. A filter mask is constructed using the deep-water gravity wave dispersion relation to extract effective wave signal components that conform to the gravity wave propagation law. The effective wave signal components are mapped into a phase-resolved wavefront elevation field through inverse transformation and linear calibration. The historical frame sequence of the phase-resolved wavefront elevation field is input into a pre-trained Fourier neural operator prediction model, and a three-dimensional forward wave field within a future time window is generated through inference calculation, which serves as the digital twin of the forward wave.

3. The ship intelligent management system based on digital twins according to claim 1, characterized in that, Based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin, the twin data coverage scale is determined, including: The computing network resource orchestration controller multiplies the prediction time window with the speed to obtain the forward wave prediction distance; Divide the forward wave prediction distance by the vertical resolution value in the grid resolution and round up to obtain the vertical grid point count; divide the effective coverage width by the horizontal resolution value in the grid resolution and round up to obtain the horizontal grid point count; The number of vertical grid points is multiplied by the number of horizontal grid points to obtain the total number of spatial grid points contained in the forward-looking wave digital twin, and the total number of spatial grid points is determined as the coverage scale of the twin data.

4. The ship intelligent management system based on digital twins according to claim 1, characterized in that, Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined, including: The computing network resource orchestration controller performs two-dimensional spectral analysis on the wavefront elevation field of the forward-looking wave digital twin to extract the main propagation direction of the wave, and determines the difference between the heading and the main propagation direction of the wave as the relative wave encounter angle; it performs power spectral density analysis on the time series of the wavefront elevation field, determines the upper limit of the wave frequency band based on the cumulative energy quantile rule, and converts the upper limit of the wave frequency band into the maximum intrinsic angular frequency; Using the deep-water gravity wave dispersion relation, the deep-water wave number is obtained by dividing the square of the maximum intrinsic angular frequency by the gravitational acceleration; the Doppler frequency shift term is obtained by multiplying the deep-water wave number, the speed, and the cosine of the relative wave encounter angle. Calculate the absolute value of the difference between the maximum intrinsic angular frequency and the Doppler frequency shift term to obtain the maximum encounter angular frequency; divide the maximum encounter angular frequency by the constant pi to obtain the effective sampling update frequency that satisfies the Nyquist sampling theorem.

5. The ship intelligent management system based on digital twins according to claim 4, characterized in that, Based on the product of the twin data coverage scale and the effective sampling update frequency, a flight condition-related load factor characterizing the non-linear growth of twin service requirements with the flight speed is calculated, including: The computing network resource orchestration controller presets the single-grid point coding load; it then multiplies the single-grid point coding load, the prediction time window, and the effective coverage width together to obtain the numerator. The constant of pi, the acceleration due to gravity, the vertical resolution value and the horizontal resolution value in the grid resolution are multiplied together and used as the denominator. Calculate the absolute value of the cosine of the relative encounter angle and the square of the maximum intrinsic angular frequency; The fundamental structure constant is obtained by dividing the numerator by the denominator. The fundamental structure constant is then multiplied by the absolute value of the cosine of the relative wave encounter angle and the square of the maximum intrinsic angular frequency to obtain the navigation condition associated load factor.

6. The ship intelligent management system based on digital twins according to claim 1, characterized in that, A multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption, and cost, including: The end-to-end latency model consists of the base protocol baseline latency, computation processing latency, and network transmission latency. The computation processing latency is calculated by dividing the twin computing task load by the computing power allocated to the edge computing nodes. The network transmission latency is calculated by dividing the product of the minimum service throughput requirement and the scheduling period by the product of the effective transmission efficiency coefficient and the network slice bandwidth configuration. The twin computing task load is calculated linearly based on the twin data coverage scale. The energy consumption model consists of the sum of computing energy consumption and communication energy consumption; wherein, the computing energy consumption is proportional to the product of the square of the computing power allocated to the edge computing node and the scheduling period; and the communication energy consumption is proportional to the product of the network slice bandwidth configuration and the scheduling period. The cost model consists of the sum of slice resource cost and computing power resource cost; wherein, the slice resource cost is calculated by multiplying the slice rental unit price, the network slice bandwidth configuration, and the scheduling period; the computing power resource cost is calculated by multiplying the computing power billing unit price, the computing power allocated to the edge computing node, and the scheduling period.

7. The ship intelligent management system based on digital twins according to claim 6, characterized in that, The minimum service throughput requirement to meet spatiotemporal synchronization is determined using the associated load factor of the navigation conditions. This minimum service throughput requirement is then used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption, and cost. The network resource orchestration controller calculates the linear term coefficient based on the preset single-grid point coding load and forward-looking spatial domain size; multiplies the linear term coefficient with the flight speed to obtain the linear demand component; multiplies the flight condition associated load factor with the square of the flight speed to obtain the nonlinear demand component; and adds the linear demand component and the nonlinear demand component to obtain the minimum service throughput requirement. Set a sensitivity adjustment constant, calculate the product of the sensitivity adjustment constant and the load factor associated with the navigation condition, and calculate the natural exponential function value of the product as the time delay priority factor; Based on the aforementioned delay priority factor, a delay weight whose value increases with the increase of the flight condition associated load factor is generated, and an energy consumption weight and a cost weight whose value decreases with the increase of the flight condition associated load factor. Using the latency weight, energy consumption weight, and cost weight, the normalized end-to-end latency model, energy consumption model, and cost model are weighted and summed to construct a comprehensive adaptation objective function for solving.

8. The ship intelligent management system based on digital twins according to claim 6, characterized in that, The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions, including: The computing network resource orchestration controller is constructed including network slice bandwidth configuration, continuous variables of edge computing node index, and particle position vector of edge computing node computing power allocation. A penalty function is constructed based on the end-to-end delay model. When the calculated end-to-end delay exceeds a preset spatiotemporal synchronization threshold, the penalty function outputs a penalty value; otherwise, it outputs a zero value. The penalty function is added to the comprehensive adaptation objective function to form the fitness function. The particle position vector is iteratively updated using a particle swarm optimization algorithm, and the value of the fitness function is calculated in each iteration to find the globally optimal particle position. Map the continuous variable of the edge computing node index in the global optimal particle position to a node identifier to determine the target edge computing node; The network slice bandwidth configuration instruction is generated based on the globally optimal particle position, and the network slice bandwidth configuration is mapped to the guaranteed bit rate and maximum bit rate in the quality of service parameters and sent to the core network slice management unit. The edge computing node allocation instruction is generated based on the globally optimal particle position, the computing power allocated to the edge computing node is mapped to the execution cycle quota of the container processor, and then sent to the resource management unit of the target edge computing node.

9. A ship intelligent management method based on digital twins, characterized in that, Implementing a ship intelligent management system based on digital twins as described in any one of claims 1-8, comprising: Collect radar images, speed, heading, and attitude data of the ship, and reconstruct a forward-looking wave digital twin based on the radar images; Resource orchestration is performed within a scheduling period, and the resource orchestration includes: The coverage scale of the twin data is determined based on the prediction time window, effective coverage width, and grid resolution of the forward-looking wave digital twin. Based on the speed, and the relative wave encounter angle and wave frequency band upper limit calculated from the heading and the radar image, the effective sampling update frequency for the forward-looking wave digital twin is determined; Based on the product relationship between the twin data coverage scale and the effective sampling update frequency, calculate the navigation condition-related load factor that characterizes the non-linear growth of twin service demand with the speed. The minimum service throughput requirement to meet the spatiotemporal synchronization requirement is determined by using the associated load factor of the navigation condition, and the minimum service throughput requirement is used as a feedforward load parameter input to a multi-dimensional resource adaptation model that includes end-to-end latency, energy consumption and cost. The multidimensional resource adaptation model is solved using a swarm intelligence optimization algorithm, generating and issuing network slice bandwidth configuration instructions and edge computing node allocation instructions.

Images