Multi-link Cooperative Scheduling Method for Marine Emergency Communications

By inverting wave height through multipath channels to obtain sea state levels, and combining adaptive windowing and counterfactual reward mechanisms to optimize DRL training, the problems of real-time sea state acquisition and emergency dispatch strategy convergence in marine communication are solved, and efficient marine emergency communication is achieved.

CN123123441APending Publication Date: 2026-11-17CHINA TRANSPORT TELECOMM & INFORMATION CENT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611300729.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In marine communications, existing technologies struggle to acquire real-time information on ship positions and sea conditions, making it difficult to adapt to non-stationary marine channels. This results in the inability of emergency dispatch strategies to converge, hindering the ability of agents to learn the optimal dispatch strategy for emergency services. Furthermore, existing observation windows cannot simultaneously achieve rapid tracking and noise suppression.

Method used

Real-time sea state levels are obtained by inverting wave height through multipath channels. An adaptive window is used to display stationary and non-stationary ocean channels. The standard DRL training module for decision scheduling is optimized. Sea state is inverted using VDES signal multipath fingerprints. The observation window is dynamically adjusted by combining CUSUM or Page-Hinkley detection adaptive window mechanism. The counterfactual reward mechanism is combined to accelerate DRL convergence.

Benefits of technology

It enables the acquisition of real-time sea conditions without the need for additional wave height sensors, significantly reducing system deployment costs, improving link quality assessment accuracy, rapidly tracking channel changes, significantly accelerating DRL convergence speed, and improving the quality of emergency dispatch strategies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123441A_ABST
    Figure CN123123441A_ABST
Patent Text Reader

Abstract

This invention provides a multi-link collaborative scheduling method for marine emergency communications. It obtains sea state levels by acquiring real-time sea state parameters from the VDES (Vacuum Demand System), then imports these parameters into the quality assessment of each link to calculate quality assessment values. These quality assessment values ​​are then used to adjust the calculation results for each link, ultimately outputting the optimal link among four links to the user. Furthermore, it provides an adaptive observation window to display stable and non-stationary marine channels. Finally, it offers a deep reinforcement learning scheduling training method for a standard DRL training module based on counterfactual link assessment, providing more accurate support for decision-making. This invention has good engineering deployability. Each module can be implemented on existing shipborne communication equipment and edge computing nodes without large-scale modifications to antennas, radio frequency components, and other hardware. It improves the reliability of marine emergency communications. Through differentiated service priority scheduling and high-quality link selection, the latency and reliability of critical services such as search and rescue positioning are significantly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of communications, and in particular to a multi-link collaborative scheduling method for marine emergency communications. Background Technology

[0002] The ocean covers approximately 71% of the Earth's surface, but global ocean communication coverage is less than 15%. The marine emergency communication environment is inherently heterogeneous. Common methods for acquiring marine information include: space-based links (LEO / MEO satellites, Ka / Ku bands), air-based links (UAVs / balloon platforms), ground-based links (shore-based 5G / microwave), and sea-based links (ship VDES / BeiDou short message / 5G NR NTN). These four network layers differ greatly in terms of frequency band, bandwidth, latency, coverage, and mobility. Furthermore, the network requirements also vary for differentiated services such as search and rescue positioning, situational awareness transmission, and command and dispatch video.

[0003] The existing technology has the following problems:

[0004] I. Marine communication requires real-time information on the location of the communicating vessel and its surrounding sea area, including key parameters such as real-time wave height and sea state. However, the distance between shore-based weather stations / buoys is typically over 100 kilometers, and shipborne wave height sensors are expensive and have a low installation rate, and are prone to failure in extreme weather conditions. Using recent buoy data or historical statistical data for interpolation results in severely insufficient spatial resolution and real-time performance, leading to inaccurate quality assessments of various links.

[0005] Second, the acquired real-time sea state information needs to be displayed in the observation window, but the current observation window can only display stationary ocean channels and cannot adapt to non-stationary ocean channels. Because ocean channels are affected by multiple factors such as waves, rainfall, and elevation angle, their statistical characteristics change continuously on time scales ranging from seconds to minutes. The existing observation window length cannot simultaneously take into account fast tracking and noise suppression. Incomplete state information affects the quality of decision-making. For example, if the window is too long, it will mask channel abrupt changes, while if the window is too short, it will lead to excessive estimation variance.

[0006] Third, the sparse rewards lead to the failure of emergency dispatch strategies to converge. Marine emergency events (distress calls, SOS alarms, rescue video transmissions) occur very infrequently in actual operation. Standard DRL (deep reinforcement learning) training modules spend most of their time experiencing normal communication scenarios and are not optimized for the non-stationarity of marine channels. This makes it difficult to obtain differentiated feedback signals under emergency scenarios, resulting in the agent's inability to learn the optimal dispatch strategy for emergency services. Convergence is extremely slow or even non-convergent, making standard DRL strategies prone to "catastrophic forgetting" or insufficient generalization ability. Summary of the Invention

[0007] This invention provides a multi-link collaborative scheduling method for marine emergency communication.

[0008] One of the objectives of this invention is to provide a method for obtaining real-time sea state levels through multipath channels.

[0009] The second objective of the invention is to provide a method for displaying stable and unbalanced ocean channels through an adaptive window.

[0010] The third objective of the invention is to provide a standard DRL training module that can optimize decision scheduling.

[0011] To achieve the above objectives, embodiments of this application provide a flow control method, comprising: acquiring the output of a pressure controller, the valve position value of a first flow valve, and the set value of a second flow controller; constructing a fuzzy controller; using the valve position value of the first flow valve and the set value of the second flow controller as inputs to the fuzzy controller, and outputting interference factor calculation coefficients; determining an updated set value of the second flow controller based on the range of the second flow controller, the interference factor calculation coefficients, and the output of the pressure controller according to a first preset rule; and adjusting the second flow rate based on the updated set value of the second flow controller.

[0012] In one embodiment, the calculation process for the inverted wave height in step 200 is as follows:

[0013] Assume the VDES operating frequency is f = 160MHz, the wavelength is λ = c / f = 1.875m, and the antenna height is h. _ant = 10m, antenna elevation angle θ = 10°, the power attenuation of the sea surface reflection path relative to the direct path is expressed as: L _ref = L _geo ·ρ², where L _geo The geometric path loss difference is given by ρ, where ρ is the sea surface reflection coefficient, and ρ is related to the sea surface roughness parameter ε=h. _w / λ is related; when ε < 0.1, it is a calm sea surface, and when ρ is close to 1, it is a specular reflection; when ε > 1, it is a rough sea surface, and ρ is diffuse scattering.

[0014] Derivation of the inversion formula:

[0015] Step a, measure the Rice factor K=P from the impulse response time h(τ). _LoS / P _scatter Given P _scatter Proportional to ρ²·(1-ρ²)^{-1}, K has an analytical relationship with ρ: K=(1-ρ²) / ρ²;

[0016] Step b, calculate ρ from K: ρ = 1 / sqrt(1+K);

[0017] Step c: Calculate the wave height h from the relationship between ρ and sea surface roughness._w According to the Miller-Brown model, the effective reflectance ρ of the sea surface _eff =ρ _0 ·exp(-2(2π·h _w ·sin(θ) / λ)²), where ρ _0 The reflection coefficient of a calm sea surface can be solved by taking the logarithm:

[0018] h _w =(λ / (2π·sin(θ)))·sqrt(-ln(ρ _eff / ρ _0 ) / 2);

[0019] Substitute the VDES parameters: h _w =(1.875 / (2π·sin(10°)))·sqrt(-ln(ρ _eff / 0.9) / 2)≈1.72·sqrt(-ln(ρ _eff / 0.9) / 2);

[0020] This yields the wave height estimate for a single sea surface reflection path. Then, the wave height estimate is obtained by calibrating and correcting the wave height using actual sea measurement data. Repeating the above process yields the wave height estimate for each reflection path in the multipath channel.

[0021] The wave height after multi-path wave height aggregation is obtained by the following formula:

[0022] h _w_final =Σ(w _i ×h _w,i ) / Σ(w _i );

[0023] Where the weight w _i =SNR _i / σ² _i SNR _i Let σ² be the signal-to-noise ratio of the i-th path. _i The inversion variance of the path is used. When the signal-to-noise ratio of a path is lower than the threshold, it is directly removed from the aggregation.

[0024] In one embodiment, the specific process of CUSUM or Page-Hinkley detection in step 300 is as follows:

[0025] Let the historical mean of the quality index of each link within the current adaptive window be μ, the standard deviation be σ, and each new sampled value be x. _t Upon arrival:

[0026] Step a: Calculate the normalized deviation value: z _t =(x _t -μ) / σ;

[0027] Step b: Recursive cumulative statistic: S _t =max(0,S_ {t-1} +z _t -k);

[0028] Where k is the sensitivity parameter;

[0029] Step c: Decision: If S _t > H, determine if a change point has occurred, reset S _t =0, update the baseline mean μ to the mean of the data within the new adaptive window.

[0030] In one embodiment, the impact of calculating sea state parameters in CUSUM or Page-Hinkley during step 300 is as follows:

[0031] The sea state level is divided into 1 to 9, and the sea state parameter serves as the basis for adaptive adjustment of the sensitivity parameter k.

[0032] Step a: When the sea state is 1 to 3, the quality fluctuations of each link are caused by noise, and k takes a value of 0.3σ. When the sea state is 7 to 9, the quality fluctuations of each link include real channel mutations, and k takes a value of 0.8σ.

[0033] Step b: When the sea state level changes, the CUSUM or Page-Hinkley baseline reset is triggered directly, without waiting for the CUSUM or Page-Hinkley statistics to accumulate to the threshold;

[0034] Step c: Reference window W when sea state is 1-3 _base =100, reference window W when sea state is 4-6 _base =50; sea state 7-9 reference W _base =20.

[0035] In one embodiment, the automatic adjustment process of the baseline window and the variable point window in step 300 is as follows:

[0036] Window size W(t) = W base * exp(-α * D(t));

[0037] Among them W base Let α be the reference window size, α be the sensitivity coefficient, and D(t) be the sampling point statistic. When D(t) is close to 0, W(t) is close to W. base As D(t) increases, W(t) changes according to W(t) = W _base The contraction is performed using ×exp(-α ×D(t)).

[0038] In one embodiment, in step 400, the training module constructs / updates the channel prediction model using the estimation results displayed in the adaptive window, and updates the adaptive window display, as follows:

[0039] Step a: Construct the channel prediction model:

[0040] An autoregressive model AR(p) is used, where p is the model order. For each link j, an independent set of AR coefficients {a_1,j, a_2,j, ..., a_p,j} is maintained.

[0041] Prediction formula: Q _pred (j|t+1) = a _1 ,j×Q _j (t) + a _2 ,j×Q _j (t-1) + ... + a_p,j×Q _j (t-p+1);

[0042] Step b: Channel prediction model initialization:

[0043] Collection link j is the closest to W _base A quality assessment value {Q} _j (t), Q _j (t-1), ..., Q _j (tW _base +1)};

[0044] Construct a training sample matrix X and a label vector Y, where each row of X is [Q]. _j (ti), Q _j (ti-1),...,Q _j [(ti-p+1)], where the corresponding element of Y is Q. _j (t-i+1), where i ranges from 0 to W _base -p-1;

[0045] The AR coefficients are solved using the least squares method: A = (X^T × X)^{-1} × X^T × Y;

[0046] Calculate the residual variance σ² _res = mean((YX×A)²), which serves as the benchmark for predicting uncertainty;

[0047] Step c: Online update of the channel prediction model:

[0048] Each time a new quality assessment value Q is obtained _j After (t+1), the AR coefficients are updated online using the Recursive Least Squares (RLS) algorithm;

[0049] Calculate the prediction error e(t+1)=Q _j (t+1)-Q _pred (j|t+1);

[0050] Update the gain vector K(t+1) = R(t)^{-1}×X(t+1) / (1+X(t+1)^T×R(t)^{-1}×X(t+1)), where R(t) is the input autocorrelation matrix;

[0051] Update the AR coefficients A(t+1) = A(t) + K(t+1) × e(t+1);

[0052] Update the autocorrelation matrix R(t+1) = R(t) + X(t+1) × X(t+1)^T;

[0053] Update residual variance σ² _res (t+1)=(1-γ)×σ² _res (t)+γ×e(t+1)², where γ is the forgetting factor;

[0054] Step d: Channel prediction model and adaptive window are updated in tandem.

[0055] When CUSUM or Page-Hinkley detects a change point, historical data before the change point is cleared, and the AR model is reinitialized using only the new data after the change point; if there are fewer than p new data points, it degenerates into moving average prediction; the initial stage prediction variance σ² after the change point is... _pred The confidence weight w of the counterfactual reward is automatically increased. _conf The corresponding reduction.

[0056] In one embodiment, the process of predicting the future quality of each unselected link in step 500 is as follows:

[0057] Step a: Obtain the most recent p=10 historical quality assessment values ​​for the unselected links: [Q _sat (t),Q _sat (t-1),...,Q _sat (t-9)];

[0058] Step b: Obtain the AR coefficients of the unselected links [a] _1 ,sat,a _2 ,sat,...,a _10 [sat] and residual variance σ² _res ,sat;

[0059] Step c: Calculate the predicted value: Q _pred (sat|t)=a _1 ,sat×Q _sat (t)+a _2,sat×Q_sat(t-1) +...+a _10 ,sat×Q _sat (t-9);

[0060] Step d: Calculate the prediction variance: σ² _pred (sat)=σ² _res ,sat×(1+[Q _sat (t),...,Q _sat (t-9)]×R _sat ^{-1}×[Q _sat (t),...,Q _sat (t-9)]^T), where R _sat This is the autocorrelation matrix of the historical quality of the unselected links;

[0061] Step e: Repeat the above process for each unselected link to obtain the predicted quality value and prediction variance for all unselected links.

[0062] In one embodiment, the calculation process for the counterfactual reward in step 500 is as follows:

[0063] Counterfactual reward r cf (j)=Q pred (j|s t )-Q actual (s t ,a t );

[0064] Among them, Q actual (st,at): The actual quality assessment value of the selected link at the current moment, Q. pred (j|st): If link j were chosen at that time, the quality value of link j estimated by the prediction model. This formula represents the value if link j were chosen instead of a. t The expected performance differences exist between the two.

[0065] In one embodiment, the training module in step 500 further includes a dense training signal generation procedure:

[0066] Step 1: At the end of each scheduling cycle, calculate:

[0067] Total training signal r total =rt+λ×Σ _j (wconf(j)×rcf(j));

[0068] Where rt is the actual reward, rcf(j) is the counterfactual reward for each unselected link, wconf(j) is the confidence weight, and λ is the counterfactual reward weight coefficient (typical value 0.3).

[0069] Step 2: Transfer the empirical tuple (s) _t ,a _t ,r total ,s _{t+1} Store it in the experience replay buffer;

[0070] Step 3: During training, a small batch is randomly sampled from the experience replay buffer, using the total training signal r. total Update the Actor-Critic network as a target reward signal:

[0071] - Critic network update: Minimize the temporal difference error δ=r total +γ×V(s _{t+1} )-V(s _t ), where V(·) is the state value output by the Critic network, and γ is the discount factor (0.99).

[0072] - Actor network update: using policy gradient ∇θJ=E[∇θlogπ(a _t |s _t )×δ], where δ is an estimate of the dominance function;

[0073] Step 4: The Actor-Critic network parameters are updated via backpropagation.

[0074] In one embodiment, the training module in step 500 further includes an emergency event generation program:

[0075] Step a: Virtual emergency event triggering mechanism:

[0076] The training module has a virtual injection scheduler. During normal communication, the scheduler randomly triggers a virtual emergency injection every N scheduling cycles.

[0077] Step a: Virtual emergency event generation:

[0078] Upon triggering, the training module automatically generates a virtual emergency event descriptor, which includes the following fields:

[0079] 1. Event type; 2. Priority level; 3. Amount of randomly generated data; 4. Latency requirements.

[0080] Step c: Virtual emergency event execution process:

[0081] Inject virtual emergency event descriptors into the state space of the training module, including event type and priority information;

[0082] After receiving a status containing virtual emergency information, the policy network of the training module outputs a decision on which link to select;

[0083] The training module does not actually execute the link selection, but instead calculates the reward signal for the decision through a counterfactual evaluation mechanism;

[0084] Generate training experience and store it in the experience replay buffer.

[0085] This invention solves the problem of directly acquiring sea state parameters. Through VDES signal multipath fingerprint inversion, real-time sea state information of a vessel's location can be obtained without the need for additional wave height sensors, significantly reducing system deployment costs and complexity. An adaptive observation window mechanism dynamically adjusts according to the rate of channel state change, balancing fast tracking and low variance estimation, improving link quality assessment accuracy and overcoming the limitation of existing fixed windows in handling non-stationary channels. A counterfactual link evaluation mechanism provides dense training signals for the agent even during normal communication, significantly accelerating convergence speed and improving the quality of emergency scheduling strategies, solving the sparse reward problem of existing channel module DRLs in emergency communication scenarios.

[0086] In this scheme, sea state awareness provides environmental parameters for channel assessment, channel assessment provides a prediction basis for the counterfactual training module DRL, the counterfactual training module DRL outputs scheduling decisions, and SLA feedback forms a closed loop. The overall system performance is better than that of each module used independently.

[0087] This solution boasts excellent engineering deployability. Each module can be implemented on existing shipborne communication equipment and edge computing nodes without requiring large-scale modifications to antennas, radio frequency components, or other hardware. It enhances the reliability of marine emergency communications. Through differentiated service priority scheduling and high-quality link selection, the latency and reliability of critical services such as search and rescue positioning are significantly improved. Attached Figure Description

[0088] Figure 1 Flowchart of the method steps of this invention;

[0089] Figure 2 A flowchart of the present invention. Detailed Implementation

[0090] The structure and working process of this solution will be further explained below through specific implementation methods and accompanying drawings.

[0091] This invention addresses four links for collecting sea state information: space-based links (satellites), air-based links (UAVs), ground-based links (shore-based 5G / microwave), and sea-based links (VDES / BeiDou / 5G NTN). The sampling of each link is affected by sea state conditions, such as:

[0092] Space-based links (satellites): Sea state affects the multipath scattering environment of the sea surface, which in turn affects the reflection interference of satellite signals and the prediction of rainfall attenuation (sea state level is related to the probability of rainfall).

[0093] Airborne links (UAVs): Sea conditions affect the height of the evaporation waveguide on the sea surface, which in turn affects the propagation environment of the UAV relay link;

[0094] Ground-based links (shore-based 5G / microwave): Sea conditions affect multipath reflection from the sea surface, which in turn affects the propagation quality of shore-based signals;

[0095] Sea-based links (VDES / BeiDou / 5G NTN): Sea conditions directly affect multipath delay spread and fading depth.

[0096] Therefore, this scheme obtains the sea state level by acquiring real-time sea state parameters from VDES, then imports it into the quality assessment of each link to calculate the quality assessment value, and then uses the quality assessment value to adjust the calculation results of each link, finally outputting the optimal link among the four links to the user; in addition, it provides an adaptive observation window to display the stationary and non-stationary ocean channels; finally, it provides a deep reinforcement learning scheduling training method for the standard DRL training module based on counterfactual link assessment, providing more accurate support for decision-making.

[0097] The specific process is as follows: This invention discloses a multi-link cooperative scheduling method for marine emergency communication, including the following steps:

[0098] Step 100: Obtain information for each link, perform sliding correlation between each multipath channel of the VDES signal in the sea-based link and the known training sequence to obtain the impulse response of each multipath channel;

[0099] The link information in this step is obtained by measuring each link layer or by external sensors (such as network sniffers, channel detectors, and weather stations) to obtain quality indicators such as received signal strength, signal-to-noise ratio, bit error rate, and latency of each link.

[0100] VDES (VHF Data Exchange System) is a dedicated system for maritime data exchange between ships and between ships and shore, operating in the VHF band (approximately 160 MHz). During maritime communication, VDES signals, in addition to the direct path (LoS), also undergo multipath reflection from the sea surface. The degree of sea surface undulation (wave height) directly affects the time delay difference, phase difference, and fading intensity of the multipath path. Therefore, the multipath characteristics of the VDES received signal naturally carry sea state information and can serve as a "fingerprint" for indirectly sensing sea conditions.

[0101] This step obtains the multipath channel impulse response h(τ) by performing a sliding correlation between the VDES received signal r(t) and the known training sequence. Each peak in h(τ) corresponds to one of the multipath paths, and the time τ when the peak appears is the propagation delay difference of that path relative to the direct path.

[0102] The specific calculation process is as follows: Let the height of the VDES antenna be h. _ant The direct path delay is τ _0 = d / c, where d is the transmission and reception distance, c is the speed of light, and the path delay due to reflection from the sea surface is τ. _1 =sqrt(d²+(2h _ant )²) / c. Then the time delay difference between the reflected path and the direct path is Δτ=τ. _1 -τ _0 When the sea has waves as high as h _w When the height of the reflection point is in [-h] _w ,+h _w The wave height information of the corresponding multipath channel can be indirectly obtained by measuring the statistical characteristics (mean and variance) of Δτ within the range of Δτ.

[0103] For example, when the VDES antenna is 10 meters high and the transmission distance is 5 kilometers, the Δτ on a calm sea surface (without waves) is approximately 0.13 microseconds; when the wave height is 2 meters, Δτ fluctuates between 0.10 and 0.16 microseconds, and the fluctuation range is positively correlated with the wave height.

[0104] This method of retrieving sea state parameters from VDES communication signals does not require the additional deployment of wave height sensors or buoys.

[0105] Step 200: Obtain the characteristic parameters of each multipath channel from the impulse response, then establish the mapping relationship between the characteristic parameters and the wave height to obtain the inverted wave height of each multipath channel, use CFOSAT data to correct the inverted wave height through adaptive filtering, and then combine the wave height period statistical relationship to obtain the sea state level at the current communication vessel.

[0106] The feature parameter extraction process for each multipath channel in this step is explained as follows:

[0107] (1) Multipath delay spread δ _τ Identify all peak values ​​from the impulse response h(τ), calculate the difference between the delay of each reflection path and the delay of the direct path, and take the root mean square value.

[0108] Specific formula: δ _τ =sqrt(Σ(τ _i -τ _mean )²·|h _i |² / Σ|h _i |²);

[0109] Where τ _i Let |h| be the delay of the i-th path. _i |For its magnitude, τ _mean This is the power-weighted average time delay.

[0110] (2) Rice factor K_factor Defined as the ratio of the direct path power to the scattered path power, the power of the first peak (direct path) in the impulse response h(τ) is P. _LoS =|h(0)|², the sum of all other peak powers is P _scatter =Σ _{i>0} |h _i |², then K=P _LoS / P _scatter .

[0111] (3) Phase difference δ _φ The direct path phase φ is extracted from the complex channel impulse response h(τ) =|h(τ)|·exp(jφ(τ)). _0 =arg(h(0)), reflection path phase φ _1 =arg(h(τ _1 )), then δ _φ =φ _1 -φ _0 .

[0112] (4) Antenna elevation angle θ: It is directly provided by the ship's GPS / BeiDou positioning and attitude sensor (gyroscope) and does not need to be extracted from the signal.

[0113] The specific calculation process for inverting wave height mapping in this step is as follows (taking a single sea surface reflection path as an example):

[0114] Assume the VDES operating frequency is f = 160MHz, the wavelength is λ = c / f = 1.875m, and the antenna height is h. _ant =10m, antenna elevation angle θ=10°.

[0115] According to electromagnetic scattering theory (ITM-R P.680 sea surface scattering model), the power attenuation of the sea surface reflection path relative to the direct path can be expressed as: L _ref =L _geo ·ρ², where L _geo The geometric path loss difference is given by ρ, where ρ is the sea surface reflection coefficient, and ρ is related to the sea surface roughness parameter ε=h. _w / λ correlation: When ε < 0.1 (calm sea surface), ρ approaches 1 (specular reflection); when ε > 1 (rough sea surface), ρ decreases sharply (diffuse scattering).

[0116] (1) The Rice factor K = P_LoS / P_scatter is measured from the impulse response h(τ). It is known that P_scatter is proportional to ρ²·(1-ρ²)^{-1}, so there is an analytical relationship between K and ρ: K = (1-ρ²) / ρ².

[0117] (2) Find ρ from K: ρ = 1 / sqrt(1+K).

[0118] (3) Calculate the wave height h by inversely using the relationship between ρ and sea surface roughness. _w According to the Miller-Brown model, the effective reflectance ρ of the sea surface _eff =ρ _0 ·exp(-2(2π·h _w ·sin(θ) / λ)²), where ρ _0 The reflectance coefficient of a calm sea surface (approximately 0.9) can be obtained by taking the logarithm:

[0119] h _w =(λ / (2π·sin(θ)))·sqrt(-ln(ρ _eff / ρ _0 ) / 2);

[0120] Substitute the VDES parameters: h _w =(1.875 / (2π·sin(10°)))·sqrt(-ln(ρ _eff / 0.9) / 2)≈1.72·sqrt(-ln(ρ _eff / 0.9) / 2).

[0121] Therefore, starting from the measured K-factor, the wave height estimate can be obtained through three steps: "K → ρ → h_w". This mapping relationship can be calibrated and corrected using marine measurement data.

[0122] The sea state classification in this step is divided into Beaufort states 1 to 9. The following are examples of calculations for three sea state classifications:

[0123] Basic parameters: VDES frequency f=160MHz, λ=1.875m, antenna height h _ant =10m, antenna elevation angle θ=10°. Inversion formula: h _w =1.72×sqrt(-ln(ρ _eff / 0.9) / 2), where ρ _eff = 1 / sqrt(1+K).

[0124] Example 1: Sea state 1 (Beaufort 1, wave height 0-0.1m, sea surface as smooth as a mirror)

[0125] Measured channel characteristics: extremely strong direct path, weak reflected path. The peak amplitude of the direct path in the impulse response h(τ) is much larger than that of the reflected path. Measured Rice factor K≈100 (i.e., 20dB).

[0126] Calculation process: ρ _eff =1 / sqrt(1+100)=0.0995; ρ _eff / ρ _0=0.0995 / 0.9=0.1106;-ln(0.1106)=2.201;h _w =1.72×sqrt(2.201 / 2)=1.72×1.049=1.80m.

[0127] In this embodiment, the result does not match the actual wave height (0~0.1m) in sea state 1, indicating that the inversion formula has a saturation effect when K is very large, requiring segmented calibration. In practice, when K>30dB, the wave height is already below the lower limit of detection sensitivity, and h is directly output. _w <0.1m, classified as sea state 1.

[0128] Example 2: Sea state 5 (Beaufort 5, wave height 1.5~2.5m, moderate waves)

[0129] Measured channel characteristics: The power of the direct path and the reflected path are close, and a significant reflection peak appears in the impulse response h(τ). The measured Rice factor K≈3.16 (i.e. 5dB).

[0130] Calculation process: ρ _eff =1 / sqrt(1+3.16)=1 / sqrt(4.16)=0.490;ρ _eff / ρ _0 = 0.490 / 0.9=0.544;-ln(0.544)=0.609;h _w =1.72×sqrt(0.609 / 2)=1.72× 0.552=0.95m.

[0131] After correction (multiplied by a calibration factor of 2.1), h _w The value is approximately 2.0m, which corresponds to sea state 5 (wave height 1.5~2.5m), so the result is reasonable.

[0132] Example 3: Sea state 9 (Beaufort 9, wave height 7-9m, very rough seas)

[0133] Measured channel characteristics: The direct path is almost completely submerged by scattering, and there is no obvious direct peak in the impulse response h(τ), exhibiting diffuse scattering characteristics. Measured K≈1 (i.e., 0dB).

[0134] Calculation process: ρ _eff =1 / sqrt(1+1)=0.707; ρ _eff / ρ _0 =0.707 / 0.9=0.786;-ln(0.786)=0.241;h _w =1.72×sqrt(0.241 / 2)=1.72×0.347=0.60m.

[0135] After correction (multiplied by a calibration factor of 13.3), h _w ≈8.0m, corresponding to sea state 9 (wave height 7~9m), the result is reasonable.

[0136] The calibration coefficients in the above embodiments were obtained by fitting sea trial data, and vary slightly depending on the sea area and season. The essence of calibration is to correct the deviation between the theoretical model and the actual sea surface scattering characteristics. In practical engineering, a piecewise linear fitting method is used to establish a lookup table from the K-factor to wave height, which is more accurate than a single analytical formula.

[0137] The corrections in the above embodiments are dynamic calibrations, which use large-scale wave height data provided by satellite scatterometers (such as CFOSAT) as a low-frequency calibration benchmark, and use adaptive filtering to correct the long-term drift of the inversion model, thereby ensuring the continuous stability of the inversion accuracy.

[0138] The following is the calculation process for the inversion wave height aggregation of each individual multipath channel:

[0139] (1) In actual sea areas, the impulse response h(τ) of the multipath channel of the VDES received signal may contain multiple reflection paths (first reflection from the sea surface, second reflection, reflection from islands and reefs, etc.). The wave height estimate h can be independently calculated for each reflection path i. _w The final wave height is obtained by weighted averaging of the estimates from each path:

[0140] h _w_final =Σ(w _i ×h _w ,i) / Σ(w _i );

[0141] Where the weight w _i =SNR _i / σ² _i SNR _i Let σ² be the signal-to-noise ratio of the i-th path (the stronger the path, the more reliable the estimate). _i This represents the inversion variance of the path (the smaller the variance, the greater the weight). When the signal-to-noise ratio of a path is below the threshold, it is directly removed from the aggregation.

[0142] (2) Perform dynamic calibration on the aforementioned calculation results:

[0143] Calibration benchmark: Large-scale wave height data (h) with a resolution of 25 km provided by the CFOSAT satellite scatterometer. _CFOSAT As a low-frequency reference (updated approximately every 3 days).

[0144] Calibration procedure: Within the calibration window (e.g., 1 hour), accumulate the VDES-derived wave height h. _w_VDES The mean h _VDES_avgObtain the wave height h at the corresponding time and location from CFOSAT. _CFOSAT ; Calculate the deviation Δ=h _CFOSAT -h _VDES_avg The calibration coefficients β of the inversion model are updated using an adaptive filter (LMS or RLS): β _new =β _old +μ×Δ×X, where μ is the step size and X is the current feature vector; subsequent inversions use the updated calibration coefficients.

[0145] (3) Differentiated calibration for different sea state levels:

[0146] The main sources of error in the inversion model differ under different sea state levels, thus requiring a differentiated calibration strategy:

[0147] Example 1: Low sea state (1-3, wave height <1m): At this time, the K factor is very large (>15dB), and the inversion is extremely sensitive to the measurement error of K (a small change in K corresponds to a large change in wave height). Calibration focus: systematic bias correction of the K factor. Calibration method: Under calm sea state, confirm that the wave height is close to 0 using CFOSAT. The K factor measured at this time is used as the reference K. _0 Subsequent K value minus (K _measured -K _0 The bias of ).

[0148] Example 2: Medium sea state (4-6, wave height 1-4m): At this time, the K factor is in the range of 3-10dB, and the time delay spread δ _τ Both the K factor and the K factor carry valid information. Calibration focus: Simultaneously calibrate K→h. _w and δ _τ →h _w The slopes of the two mapping curves. Calibration method: Using CFOSAT wave height as the label, perform linear regression on the two mapping curves respectively, and update the slope and intercept.

[0149] Example 3: High sea state (7-9, wave height > 5m): At this time, the K-factor is close to 0dB, the direct path is almost indistinguishable, the distinguishability of the K-factor is lost, and the calibration focus shifts to relying on the time delay spread δ. _τ As the primary inversion index, the calibration method is as follows: Establish δ _τ The piecewise linear relationship with wave height was established, and the piecewise slope was calibrated using CFOSAT data. Simultaneously, the calibration frequency was increased (from 3 days to 1 day) under high sea states because the model drifts faster in extreme sea states.

[0150] Step 300: Using the mean and variance of each link's quality index as a benchmark, continuously calculate the input samples and current sea state level of each link using CUSUM or Page-Hinkley, and display them in an adaptive window. Calculate the cumulative offset of each link. When the cumulative offset of a link reaches a threshold, the benchmark window for that link will be automatically changed to a change point window that displays the change point status. When the change point ends, the benchmark window for that link will be automatically restored.

[0151] The change-point detection mechanism in this step is an adaptive observation window mechanism. The mechanism is as follows: it detects whether there is a sudden change in the statistical distribution of link quality online. When a change point is detected, it immediately shrinks the baseline window to quickly capture the new state. When no change point is detected, it gradually expands the baseline window to reduce the estimated variance.

[0152] Link determination method: The system continuously scans the physical layer status of all available links during operation. For each link, the communication hardware (satellite modem, VDES transceiver, BeiDou receiver, 5G modem) automatically outputs the quality indicators of the current link, such as Received Signal Strength (RSSI), Signal-to-Noise Ratio (SNR), and Bit Error Rate (BER).

[0153] The adaptive observation window operates independently for each available link. That is, if there are currently 3 available links (such as satellite, VDES, and BeiDou), the system maintains 3 independent adaptive windows simultaneously, each tracking the quality index changes of its corresponding link. CUSUM change point detection monitors the quality index sequence for each link.

[0154] When a change point is detected in the CUSUM of a link, only the reference window of that link shrinks, without affecting the window state of other links. The reference window size may differ for each link, depending on the degree of non-stationarity of its respective channel.

[0155] The following is a specific example of the adaptive window change process:

[0156] Initial parameter settings: Baseline window W _base =100 sampling points (sampling interval of 1 second, i.e., the reference window covers 100 seconds of data), minimum reference window W _min =5 sampling points, sensitivity coefficient α=0.3, CUSUM threshold H=10, sensitivity parameter k=0.5σ.

[0157] Phase 1: Steady-state operation (t=0~50s);

[0158] The sea conditions were stable, and the link SNR fluctuated around the mean μ=20dB with a standard deviation σ=1dB. The CUSUM statistic D(t) fluctuated slightly near 0 (D(t)<2) and did not exceed the threshold H=10. The adaptive window size W(t) =100×exp(-0.3×D(t))≈100×exp(-0.3×1)=74~100, basically remaining at 100. At this point, using all 100 sampling points to calculate the average SNR minimized the estimated variance.

[0159] Phase 2: Change point occurrence and detection (t=50~54s);

[0160] At t=50s, a gust of wind caused the sea state to change abruptly from level 2 to level 5, and the SNR dropped from 20dB to 14dB. The CUSUM statistics began to accumulate: D(50)=0, D(51)=3, D(52)=7, D(53)=12; at t=53s, D(53)=12>H=10, indicating that a change point had occurred.

[0161] Phase 3: The baseline window rapidly shrinks to form a change point window (t=53~55s);

[0162] After detecting the change point, W(53) = 100 × exp(-0.3 × 12) = 100 × 0.027 = 2.7, and the lower limit W is taken. _min =5. At this point, only the most recent 5 sampling points are used to calculate the average signal-to-noise ratio (SNR). Although the variance increases, it can quickly track the change in SNR from 20dB to 14dB.

[0163] Phase 4: Establishment of a new steady state and recovery of the adaptive window (t=55~75s);

[0164] After the new baseline is established (new mean μ=14dB), D(t) gradually decays: D(55)=8, D(58)=4, D(62)=1, D(65)=0. The baseline window is restored accordingly.

[0165] W(55)=100×exp(-0.3×8)=9, W(58)=100×exp(-0.3×4)=30, W(62)=100×exp(-0.3×1)=74, W(65)=100×exp(0)=100. The baseline window is completely restored to 100, and the system re-enters the low variance estimation state.

[0166] 1. In this step, taking CUSUM as an example, the specific detection process is as follows:

[0167] Let μ be the historical mean and σ be the standard deviation of the link quality metric (SNR) within the current window, and let x be the new sample value. _t Upon arrival:

[0168] Step 1: Calculate the normalized deviation value: z _t =(x _t -μ) / σ;

[0169] Step 2: Recursively calculate cumulative statistics: S _t =max(0,S _{t-1} +z _t -k), where k is the sensitivity parameter (k=0.5, meaning that accumulation only begins when the deviation exceeds 0.5 times the standard deviation);

[0170] Step 3: Judgment: If S _t >H, determine if a change point has occurred, reset S _t =0, update the baseline mean μ to the mean of the data in the new window.

[0171] 2. Criteria for determining the preset threshold H:

[0172] The threshold H is determined based on the Average Run Length (ARL) theory. _0 ARL is the average false alarm interval when there are no change points. _1 This represents the average detection delay when there are changing points.

[0173] In engineering, ARL is usually required _0 ≥500 (meaning there is only one false alarm on average every 500 sampling points when there are no variable points). Based on the Page approximation formula: ARL _0 ≈exp(2×(H / σ-1)×k / σ). Take k=0.5σ, ARL _0 When = 500, we can solve for H ≈ 5σ.

[0174] Therefore, the preset threshold H=5σ (based on the standard deviation of the current adaptive window signal-to-noise ratio SNR) is set. When σ=1dB, H=5dB; when σ=2dB, H=10dB. The threshold is adaptive to the noise level of the current channel.

[0175] 3. The role of sea state data in change point detection:

[0176] (1) The sea state parameter serves as the basis for adaptive adjustment of the CUSUM sensitivity parameter k: When the sea state level is low (level 1 to 3), the link quality fluctuation is mainly caused by noise, so k is taken as a small value (k=0.3σ) to reduce sensitivity and reduce false alarms; when the sea state level is high (level 7 to 9), the link quality fluctuation may contain real channel mutations, so k is taken as a large value (k=0.8σ) to increase sensitivity and detect quickly.

[0177] (2) Sea state changes themselves trigger baseline reset: When a change in sea state level is detected (e.g., from level 3 to level 5), the CUSUM baseline reset is directly triggered (S _t =0, μ is recalculated), without waiting for the CUSUM statistic to accumulate to the threshold. This allows change point detection to not only rely on indirect changes in link quality statistics but also directly utilize physical information about sea state changes.

[0178] (3) Sea state level determines the adaptive window reference W _base When the sea is calm (level 1-3), W _base =100 (slow channel change, reference window available); W in moderate sea state (4-6) _base =50; W in severe sea states (7-9) _base =20 (The channel changes rapidly, requiring a shorter variable point window).

[0179] Step 400: The fast-changing factor and the slow-changing factor in the input sample are estimated using the change point window and the baseline window, respectively. The confidence of the estimation results are weighted and fused to obtain the quality assessment value of each link. Subsequently, the quality assessment value is used to replace the input sample and substitute it into CUSUM or Page-Hinkley for change point detection. At the same time, the updated estimation results of CUSUM or Page-Hinkley are used to construct a channel prediction model for change point processing.

[0180] In this step, the rapidly changing factors (multipath, beam sway) are derived from real-time physical layer measurements of the current link. Multipath delay spread is extracted from the channel's impulse response h(τ); beam sway is obtained from the beam tracking error output of the satellite modem. These parameters change rapidly (on the order of seconds) and are smoothed using a variable-point window.

[0181] The slow-varying factors include rainfall attenuation and satellite elevation angle: rainfall attenuation is calculated using the ITU-RP.838 model combined with local meteorological data (input is rainfall rate R mm / h, output is attenuation value dB); satellite elevation angle is calculated directly from satellite ephemeris and ship GPS position. These parameters change slowly (on the order of minutes) and are smoothed using a reference window.

[0182] The following provides an example of the variable point window and the baseline window, along with the fusion calculation process:

[0183] Let short window W _short =10 sampling points, long window W _long =100 sampling points. Taking a satellite link as an example:

[0184] Variable point window estimation: Take the most recent 10 signal-to-noise ratio (SNR) samples and calculate the quality score Q of the rapidly changing factor. _fast =mean(SNR _1~SNR _10 Q = 15.2 dB, normalized to Q _fast_norm =(15.2-5) / (25-5)=0.51.

[0185] Baseline window estimation: Take the most recent 100 rainfall attenuation and elevation angle data, and calculate the quality score Q of the slow variation factor. _slow =mean(rain _attenuation_1 The link margin corresponding to ~100) is normalized to Q. _slow_norm = 0.78.

[0186] Fusion weight: w _fast =1 / σ² _fast (σ) _fast =2dB, w _fast =0.25); w _slow =1 / σ² _slow (σ) _slow =0.5dB, w _slow =4.0).

[0187] Fusion result: Q _total =(0.25×0.51+4.0×0.78) / (0.25+4.0)=(0.1275+3.12) / 4.25=0.764.

[0188] In this step, the quality assessment value serves as the input signal for CUSUM testing: CUSUM versus Q _total Perform change point monitoring, rather than monitoring the original signal-to-noise ratio (SNR), because Q _total By integrating multi-factor information, it can better reflect the true quality changes of the link.

[0189] Step 500: Import the information of each link, sea state level and confidence assessment value into the training module. Select the priority link according to the type and priority of the service to be sent and display it in the window. At the same time, perform future quality prediction and counterfactual reward for each unselected link to determine its difference relative to the priority link. Then import the real-time status information into the training module and finally output the decision window for each service type.

[0190] This step introduces a counterfactual link evaluation mechanism, which, after each actual scheduling decision is made, in addition to obtaining the actual reward r... t In addition, a channel prediction model is established to estimate "what the link quality would have been if the link had been selected" for each unselected alternative link, and a counterfactual reward is constructed accordingly. In this way, even when no real emergency event occurs, the training module can obtain differentiated scheduling quality feedback, thereby generating intensive training signals.

[0191] The training module here refers to the DRL scheduling decision module Agent, which is a software module that automatically selects communication links based on the current environment state. Its functions are as follows:

[0192] 1. At time t, the training module adjusts the current state s. t Select link a t And execute, to obtain the actual reward r t and the next state s t+1 ;

[0193] Current state s _t It is a multidimensional vector containing the following information:

[0194] (1) Fusion quality assessment value Q of each available link _total ;

[0195] (2) Current sea state level;

[0196] (3) The type and priority of the service to be sent (P1 / P2 / P3);

[0197] (4) Historical quality trends of each link (rising / stable / falling);

[0198] (5) Satellite elevation angle, for example: s _t =[Q _sat =0.72,Q _vdes =0.85,Q _beidou =0.60,sea _state =4,biz _type =P1, elev=35°].

[0199] Next state s _{t+1} This refers to the new state vector observed by the system after a scheduling cycle (e.g., 1 second) following the execution of link selection. The state value will change due to channel changes and service changes.

[0200] Selected link a _t This refers to the policy network π(a|s) of the training module. _t According to the current state s _t The output shows the selection probability of each link. The training module samples and selects one link based on the probability, choosing the optimal link from all currently available links (space-based / air-based / ground-based / sea-based). For example, when s _t When the sea state is level 4 and the service is P1 search and rescue positioning, the strategy may be to choose the VDES link (low latency and good sea state adaptability).

[0201] The channel prediction model construction and update process in this step is as follows:

[0202] The channel prediction model consists of four sub-modules:

[0203] (1) Submodule A "Data Input Interface": Receives the historical quality assessment value sequence {Q} of each link from the adaptive window module. _j (t), Q _j (t-1), ...}. Each time the adaptive window updates, the new Q is automatically updated. _total The value is appended to the history sequence;

[0204] (2) Submodule B "AR Coefficient Library": Maintains a set of AR coefficients {a} for each available link. _1 ,j,..., a _p ,j} and residual variance σ² _res j is initially calculated offline using the least squares method and updated online during runtime using the RLS algorithm;

[0205] (3) Submodule C "Prediction Calculator": Given the historical sequence and AR coefficients of link j, calculate the predicted value Q. _pred (j|t+1)=Σ(a _i ,j×Q _j (t-i+1)), and simultaneously calculate the prediction variance σ² _pred (j)=σ² _res ,j×(1+X^T×R^{-1}×X);

[0206] (4) Submodule D "Change Point Responder": Monitors the CUSUM change point detection results. When a change point notification is received:

[0207] a. Clear the historical data prior to the change point of this link;

[0208] b. Reset the AR coefficients to a zero vector (degenerate to mean prediction);

[0209] c. Temporarily increase the prediction variance (lower the confidence level);

[0210] d. Wait for new data to accumulate to p before reinitializing the AR model.

[0211] Inter-module data flow: Input interface (→ AR coefficient library (→ Prediction calculator → Output [Q]) _pred , σ ²_pred The change-point responder listens to the CUSUM output and intervenes in both the AR coefficient library and the prediction calculator simultaneously when a change occurs.

[0212] Training prediction model output [Q] _pred (j), σ² _pred (j)] is used in the training module for counterfactual reward calculation and confidence weighting.

[0213] How to construct a channel prediction model:

[0214] 1. Model Structure:

[0215] An autoregressive model AR(p) is used, where p is the model order (typically p=10). An independent set of AR coefficients {a} is maintained for each link j. _1 ,j,a _2 ,j,...,a _p ,j};

[0216] Prediction formula: Q _pred (j|t+1)=a_1,j×Q _j (t)+a _2 ,j×Q _j (t-1)+...+ a _p ,j×Q _j (t-p+1).

[0217] The reasons for choosing the AR model over LSTM are: the AR model has low computational cost, strong interpretability, and is suitable for embedded deployment; the statistical characteristics of the ocean channel are close to stationary within the adaptive window (time without change points), which the AR model can adequately characterize.

[0218] 2. Channel prediction model construction process:

[0219] (1) Collect the nearest W link j _base A quality assessment value {Q} _j (t),Q _j (t-1),...,Q _j (tW _base +1)};

[0220] (2) Construct the training sample matrix X and the label vector Y, where each row of X is [Q _j (ti),Q _j (ti-1),...,Q _j [(ti-p+1)], where the corresponding element of Y is Q. _j (t-i+1), where i ranges from 0 to W _base -p-1;

[0221] (3) Solve for the AR coefficients using the least squares method: A = (X^T × X)^{-1} × X^T × Y;

[0222] Step d: Calculate the residual variance σ² _res =mean((YX×A)²), which serves as the benchmark for predicting uncertainty.

[0223] 3. Online update process of the channel prediction model:

[0224] Each time a new quality assessment value Q is obtained _j After (t+1), the AR coefficients are updated online using the Recursive Least Squares (RLS) algorithm:

[0225] (1) Calculate the prediction error e(t+1)=Q _j (t+1)-Q _pred (j|t+1);

[0226] (2) Update the gain vector K(t+1)=R(t)^{-1}×X(t+1) / (1+X(t+1)^T×R(t)^{-1}×X(t+1)), where R(t) is the input autocorrelation matrix;

[0227] (3) Update the AR coefficients A(t+1) = A(t) + K(t+1) × e(t+1);

[0228] (4) Update the autocorrelation matrix R(t+1) = R(t) + X(t+1) × X(t+1)^T;

[0229] (5) Update the residual variance σ² _res (t+1)=(1-γ)×σ² _res (t)+γ×e(t+1)², where γ is the forgetting factor (typical value 0.01).

[0230] 4. Linkage between channel prediction model and adaptive window:

[0231] When CUSUM detects a change point, the channel prediction model needs to be specially handled:

[0232] (1) Clear the historical data before the change point and reinitialize the AR model using only the new data after the change point;

[0233] (2) When there are fewer than p new data points, the prediction method degenerates into a moving average forecast.

[0234] (3) Initial stage prediction variance σ² after the change point _pred The confidence weight w of the counterfactual reward is automatically increased. _conf The corresponding reduction is to prevent inaccurate predictions from misleading training.

[0235] The counterfactual reward in this step is calculated for each unselected link j:

[0236] r cf (j) = Q pred (j|s t ) - Q actual (s t , a t );

[0237] Among them, Qactual (s t , a t ): The link a actually selected at the current moment t The true quality assessment value. This is a measurable, real result that has already occurred.

[0238] Q pred (j|s t ): If link j were selected at that time, the quality value of link j estimated by the prediction model. This is a result "hypothetically" derived by the prediction model that did not occur.

[0239] rcf(j) = Q pred (j|s t ) - Q actual (s t , a t ): The difference between the two indicates "if j was chosen instead of a". t "How much better or worse will the quality be?"

[0240] Example: Current state s t Below, the training module actually selected the VDES link (a t =VDES), and the mass Q of VDES was measured. actual =0.85. Meanwhile, the channel prediction model estimates that if a satellite link is selected, the quality Q... pred If (sat) = 0.72, then the counterfactual reward rcf(sat) = 0.72 - 0.85 = -0.13. This means that "choosing satellite is 0.13 worse than choosing VDES", therefore choosing VDES is correct, and the training module should reinforce this choice.

[0241] Conversely, if Q pred If (sat)=0.92, then rcf(sat)=0.92-0.85=+0.07, which means "choosing satellite would be better". The training module should consider choosing satellite link in similar situations in the future.

[0242] In this way, even in the absence of real emergency events, the identification module can obtain feedback on "what would happen if other links were selected" after each scheduling, thereby continuously learning the optimal scheduling strategy and no longer relying on rare real emergency events to obtain training signals.

[0243] Step 600 involves monitoring the data of each business online and calculating the deviation between the real-time data and the displayed data. The deviation is then used to adjust the weights in step 400 and to adjust the counterfactual rewards in step 500.

[0244] In this step, the prediction variance or confidence level w of the channel prediction model is used. conf(j) Weighting of counterfactual rewards. For links with large prediction variance, the contribution of counterfactual rewards is automatically suppressed to avoid misleading training with erroneous signals.

[0245] 1. Methods for obtaining prediction variance:

[0246] The prediction variance is directly given by the statistical properties of the AR model. For the AR(p) model, the prediction variance σ² _pred (j)=σ² _res ,j×(1+X _j ^T×R _j ^{-1}×X _j );

[0247] Where: σ² _res ,j: The residual variance of the AR model for link j (calculated during training and updated online), reflecting the uncertainty of the channel prediction model itself.

[0248] X _j : Current input vector [Q _j (t),Q _j (t-1),...,Q _j [(t-p+1)];

[0249] R _j : Autocorrelation matrix of historical input data of link j (normalized form of X^T×X);

[0250] (1+X _j ^T×R _j ^{-1}×X _j This term reflects the distance between the current input and the distribution of the training data. If the current input is far from the center of the training data (i.e., the current channel state differs greatly from historical states), this term increases, and the prediction variance increases.

[0251] 2. Source of confidence weights wconf:

[0252] Confidence weight wconf(j) = 1 / (1 + σ²) _pred (j) / σ² _ref );

[0253] Where σ² _ref The reference variance is the mean of the variances of all link residuals.

[0254] When σ² _pred When (j) is very small (accurate prediction), wconf(j) is close to 1, and counterfactual rewards are almost entirely involved;

[0255] When σ² _predWhen (j) is large (inaccurate prediction), wconf(j) approaches 0, and counterfactual rewards hardly participate in training.

[0256] 3. Implementation method of "automatic suppression":

[0257] In the formula for the total training signal:

[0258] r total =rt+λ×Σ _j (wconf(j)×rcf(j));

[0259] When the prediction variance of a link j is large, wconf(j) automatically approaches 0, and the counterfactual reward rcf(j) for that link increases relative to r. total The contribution is automatically multiplied by a weight close to 0, which is equivalent to being "suppressed".

[0260] The suppression process is smooth and continuous (rather than a 0 / 1 switch), avoiding abrupt changes in the training signal due to hard thresholds.

[0261] In one embodiment of the present invention, the training module further includes a dense training signal generation program:

[0262] Step 1: At the end of each scheduling cycle, calculate:

[0263] Total training signal r total =rt+λ×Σ _j (wconf(j)×rcf(j));

[0264] Where rt is the actual reward (from the SLA feedback unit), rcf(j) is the counterfactual reward for each unselected link, wconf(j) is the confidence weight, and λ is the counterfactual reward weight coefficient (typical value 0.3).

[0265] Step 2: Transfer the empirical tuple (s) _t ,a _t ,r total ,s _{t+1} Experiences are stored in the experience replay buffer. This is a circular queue data structure, typically with a capacity of 100,000 experiences. When the buffer is full, new experiences overwrite the oldest ones.

[0266] Step 3: During training, randomly sample a small batch (e.g., 64 records) from the experience replay buffer, and use r total Update the Actor-Critic network as a target reward signal:

[0267] - Critic network update: Minimize the temporal difference error δ=r total +γ×V(s_{t+1} )-V(s _t ), where V(·) is the state value output by the Critic network, and γ is the discount factor (0.99).

[0268] - Actor network update: using policy gradient ∇θJ=E[∇θlogπ(a _t |s _t )×δ], where δ is an estimate of the dominance function.

[0269] Step 4: The Actor-Critic network parameters are updated via backpropagation (Adam optimizer, learning rate 3×10^{-4}).

[0270] Key difference: In the standard training module DRL, the reward stored in the replay buffer is rt (only the actual reward; in emergency scenarios, rt is almost constant, and the training signal is sparse). This invention replaces rt with r. total (Including counterfactual rewards), such that even during normal communication, the total training signal r total It also exhibits differentiation due to the presence of counterfactual rewards, thus providing dense training signals.

[0271] In one embodiment of the present invention, the training module further includes virtual injection of emergency scenarios. During normal communication, emergency service requirements (such as sudden high-priority data packets) are randomly injected virtually, and emergency dispatch training samples are generated using counterfactual evaluation, enabling the training module to learn emergency strategies without the need for real emergency events.

[0272] The following is a specific example of how virtual injection can be implemented in emergency scenarios:

[0273] 1. Triggering mechanism:

[0274] The training module internally uses a virtual injection scheduler (timer). During normal communication, every N scheduling cycles (e.g., N=1000, i.e., every 1000 seconds), the scheduler randomly triggers a virtual emergency injection. The trigger probability can also be controlled by a Poisson process to simulate the randomness of real emergency events.

[0275] 2. Virtual emergency event generation:

[0276] Upon triggering, the system automatically generates a virtual emergency event descriptor, which includes the following fields:

[0277] - Event type: Randomly select one from [SOS alarm, distress call, rescue video transmission, medical data transmission];

[0278] - Priority level: Automatically assigned based on event type (SOS→P1, distress call→P1, rescue video→P2, medical data→P2);

[0279] - Data volume: Randomly generated (e.g., P1 service 10KB~100KB, P2 service 1MB~10MB).

[0280] - Latency requirements: P1 service requires <500ms, P2 service requires <2s.

[0281] 3. Execution process:

[0282] Step a: Inject the virtual emergency event descriptor into the state space of the training module. Specifically, this is done by adding a virtual emergency event descriptor to the state vector s. _t Add a "Virtual Emergency Event" field to the system, and fill in the event type and priority information mentioned above.

[0283] Step b: The policy network of the training module receives the state s containing the virtual emergency information. _t_prime Then, the output link selection decision a _t_prime (e.g., choose to transmit rescue videos via satellite link).

[0284] Step c: The training module does not actually execute the link selection (does not actually route the virtual emergency data), but instead calculates the reward signal for the decision through a counterfactual evaluation mechanism (steps 2-5).

[0285] Step d: Generate training experience (s) _t_prime , a _t_prime , r total_prime , s _{t+1}_prime ), and store it in the experience replay buffer.

[0286] The virtual injection module in this scheme is an independent thread that runs in parallel with the normal scheduling thread. It generates virtual state and sends it to the policy network for inference, but does not interfere with the execution path of the normal scheduling. The generated training experience is passed to the training thread through a thread-safe queue.

[0287] In one embodiment of the present invention, the training module further includes a policy network update step, which updates the total training signal r. total The training signal, used as the objective function of reinforcement learning, updates the policy and value parameters of the Actor-Critic network. The specific process is as follows:

[0288] 1. The parameter updates of the Actor-Critic network are automatically implemented through the standard backpropagation algorithm, requiring no manual intervention. The specific steps for each training iteration are as follows:

[0289] (1) Randomly sample 64 experiences {(s) from the experience replay buffer_i ,a _i ,r total_i ,s _{i+1} )};

[0290] (2) The Critic network calculates the state value V(s) of each experience. _i ) and V(s _{i+1} );

[0291] (3) Calculate the timing difference error δ _i =r total_i +γ×V(s _{i+1} )-V(s _i ).

[0292] (4) Critic loss function L _critic =mean(δ _i ²), the Adam optimizer is used to optimize the Critic network parameters θ _critic Perform gradient descent update: θ _critic ←θ _critic -lr×∇θ _critic L _criti c);

[0293] (5) The Actor network calculates the policy gradient ∇θ _actor J=mean(∇θ _actor logπ(a _i |s _i )×δ _i ), for the Actor network parameters θ _actor Perform gradient ascent update: θ _actor ←θ _actor +lr×∇θ _actor J;

[0294] (6) The above updates are executed automatically on GPU / CPU, and are completed by the back propagation mechanism of the PyTorch / TensorFlow framework.

[0295] (7) The counterfactual reward calculation module runs after each scheduling cycle to calculate the total training signal r. total Total training signal r total The data, along with the state and actions, is packaged into an experience tuple and written to the experience replay buffer (a data structure in memory). The training thread samples data from the buffer and updates the network weights through backpropagation. After the network weights are updated, the training module uses the updated policy network to make decisions during the next online scheduling.

[0296] The entire process requires no code modification or recompilation; network weight updates are automatically performed at runtime via gradient descent. The counterfactual evaluation module, as a parallel computation unit, runs after each scheduling, calculating the counterfactual reward and replacing the standard reward rt with r. total .

[0297] This invention solves the problem of directly acquiring sea state parameters. Through VDES signal multipath fingerprint inversion, real-time sea state information of a vessel's location can be obtained without the need for additional wave height sensors, significantly reducing system deployment costs and complexity. An adaptive observation window mechanism dynamically adjusts according to the rate of channel state change, balancing fast tracking and low variance estimation, improving link quality assessment accuracy and overcoming the limitation of existing fixed windows in handling non-stationary channels. A counterfactual link evaluation mechanism provides dense training signals for the agent even during normal communication, significantly accelerating convergence speed and improving the quality of emergency scheduling strategies, solving the sparse reward problem of existing channel module DRLs in emergency communication scenarios.

[0298] In this scheme, sea state awareness provides environmental parameters for channel assessment, channel assessment provides a prediction basis for the counterfactual training module DRL, the counterfactual training module DRL outputs scheduling decisions, and SLA feedback forms a closed loop. The overall system performance is better than that of each module used independently.

[0299] This solution boasts excellent engineering deployability. Each module can be implemented on existing shipborne communication equipment and edge computing nodes without requiring large-scale modifications to antennas, radio frequency components, or other hardware. It enhances the reliability of marine emergency communications. Through differentiated service priority scheduling and high-quality link selection, the latency and reliability of critical services such as search and rescue positioning are significantly improved.

[0300] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A multi-link cooperative scheduling method for marine emergency communication, characterized in that, Includes the following steps: Step 100: Obtain information for each link, perform sliding correlation between each multipath channel of the VDES signal in the sea-based link and the known training sequence to obtain the impulse response of each multipath channel; Step 200: Obtain the characteristic parameters of each multipath channel from the impulse response, then establish the mapping relationship between the characteristic parameters and the wave height to obtain the inverted wave height of each multipath channel, use CFOSAT data to correct the inverted wave height through adaptive filtering, aggregate the inverted wave height of each multipath channel, and then combine the wave height period statistical relationship to obtain the sea state level at the current communication vessel. Step 300: Using the mean and variance of each link's quality index as a benchmark, continuously calculate the input samples and current sea state level of each link using CUSUM or Page-Hinkley, and display them in an adaptive window. Calculate the cumulative offset of each link. When the cumulative offset of a link reaches a threshold, the benchmark window for that link will be automatically changed to a change point window that displays the change point status. When the change point ends, the benchmark window for that link will be automatically restored. Step 400: The fast-changing factor and the slow-changing factor in the input sample are estimated using the change point window and the baseline window, respectively. The confidence of the estimation results are weighted and fused to obtain the quality assessment value of each link. Subsequently, the quality assessment value is used to replace the input sample and substitute it into CUSUM or Page-Hinkley for change point detection. At the same time, the updated estimation results of CUSUM or Page-Hinkley are used to construct a channel prediction model for change point processing. Step 500: Import the information of each link, sea state level and confidence assessment value into the training module. Select the priority link according to the type and priority of the service to be sent and display it in the adaptive window. At the same time, perform future quality prediction and counterfactual reward for each unselected link to determine its difference relative to the priority link. Then import the real-time status information into the training module and finally output the decision window for each service type. Step 600 involves monitoring the data of each business online and calculating the deviation between the real-time data and the displayed data. The deviation is then used to adjust the weights in step 400 and to adjust the counterfactual rewards in step 500.

2. The multi-link cooperative scheduling method for marine emergency communication according to claim 1, characterized in that, The calculation process for the inverted wave height in step 200 is as follows: Assume the VDES operating frequency is f = 160MHz, the wavelength is λ = c / f = 1.875m, and the antenna height is h. _ant = 10m, antenna elevation angle θ = 10°, the power attenuation of the sea surface reflection path relative to the direct path is expressed as: L _ref = L _geo ·ρ², where L _geo The geometric path loss difference is given by ρ, where ρ is the sea surface reflection coefficient, and ρ is related to the sea surface roughness parameter ε=h. _w / λ is related; when ε < 0.1, it is a calm sea surface, and when ρ is close to 1, it is a specular reflection; when ε > 1, it is a rough sea surface, and ρ is diffuse scattering. Derivation of the inversion formula: Step a, measure the Rice factor K=P from the impulse response time h(τ). _LoS / P _scatter Given P _scatter Proportional to ρ²·(1-ρ²)^{-1}, K has an analytical relationship with ρ: K=(1-ρ²) / ρ²; Step b, calculate ρ from K: ρ = 1 / sqrt(1+K); Step c: Calculate the wave height h from the relationship between ρ and sea surface roughness. _w According to the Miller-Brown model, the effective reflectance ρ of the sea surface _eff =ρ _0 ·exp(-2(2π·h _w ·sin(θ) / λ)²), where ρ _0 The reflection coefficient of a calm sea surface can be solved by taking the logarithm: h _w =(λ / (2π·sin(θ)))·sqrt(-ln(ρ _eff / r _0 ) / 2); Substitute the VDES parameters: h _w =(1.875 / (2π·sin(10°)))·sqrt(-ln(ρ _eff / 0.9) / 2)≈1.72·sqrt(-ln(ρ _eff / 0.9) / 2); This yields the wave height estimate for a single sea surface reflection path. Then, the wave height estimate is obtained by calibrating and correcting the wave height using actual sea measurement data. Repeating the above process yields the wave height estimate for each reflection path in the multipath channel. The wave height after multi-path wave height aggregation is obtained by the following formula: h _w_final =Σ(w _i ×h _w,i ) / Σ(w _i ); Where the weight w _i =SNR _i / σ² _i SNR _i Let σ² be the signal-to-noise ratio of the i-th path. _i The inversion variance of the path is used. When the signal-to-noise ratio of a path is lower than the threshold, it is directly removed from the aggregation.

3. The multi-link cooperative scheduling method for marine emergency communication according to claim 1, characterized in that, In step 300, the specific process of CUSUM or Page-Hinkley detection is as follows: Let the historical mean of the quality index of each link within the current adaptive window be μ, the standard deviation be σ, and each new sampled value be x. _t Upon arrival: Step a: Calculate the normalized deviation value: z _t =(x _t -μ) / σ; Step b: Recursive cumulative statistic: S _t =max(0,S_ {t-1} +z _t -k); Where k is the sensitivity parameter; Step c: Decision: If S _t > H, determine if a change point has occurred, reset S _t =0, update the baseline mean μ to the mean of the data within the new adaptive window.

4. The multi-link cooperative scheduling method for marine emergency communication according to claim 3, characterized in that, In step 300, the impact of calculating sea state parameters in CUSUM or Page-Hinkley is as follows: The sea state level is divided into 1 to 9, and the sea state parameter serves as the basis for adaptive adjustment of the sensitivity parameter k. Step a: When the sea state is 1 to 3, the quality fluctuations of each link are caused by noise, and k takes a value of 0.3σ. When the sea state is 7 to 9, the quality fluctuations of each link include real channel mutations, and k takes a value of 0.8σ. Step b: When the sea state level changes, the CUSUM or Page-Hinkley baseline reset is triggered directly, without waiting for the CUSUM or Page-Hinkley statistics to accumulate to the threshold; Step c: Reference window W when sea state is 1-3 _base =100, reference window W when sea state is 4-6 _base =50; sea state 7-9 reference W _base =20.

5. The multi-link cooperative scheduling method for marine emergency communication according to claim 3, characterized in that, The automatic adjustment process of the reference window and the variable point window in step 300 is as follows: Window size W(t) = W base * exp(-α * D(t)); Among them W base Let α be the reference window size, α be the sensitivity coefficient, and D(t) be the sampling point statistic. When D(t) is close to 0, W(t) is close to W. base As D(t) increases, W(t) changes according to W(t) = W _base The contraction is performed using ×exp(-α ×D(t)).

6. The multi-link cooperative scheduling method for marine emergency communication according to claim 1, characterized in that, In step 400, the training module constructs / updates the channel prediction model using the estimation results displayed in the adaptive window, and updates the adaptive window display, as follows: Step a: Construct the channel prediction model: An autoregressive model AR(p) is used, where p is the model order. For each link j, an independent set of AR coefficients {a_1,j, a_2,j, ..., a_p,j} is maintained. Prediction formula: Q _pred (j|t+1) = a _1 ,j×Q _j (t) + a _2 ,j×Q _j (t-1) + ... + a_p,j×Q _j (t-p+1); Step b: Channel prediction model initialization: Collection link j is the closest to W _base A quality assessment value {Q} _j (t), Q _j (t-1), ..., Q _j (tW _base +1)}; Construct a training sample matrix X and a label vector Y, where each row of X is [Q]. _j (ti), Q _j (ti-1),...,Q _j [(ti-p+1)], where the corresponding element of Y is Q. _j (t-i+1), where i ranges from 0 to W _base -p-1; The AR coefficients are solved using the least squares method: A = (X^T × X)^{-1} × X^T × Y; Calculate the residual variance σ² _res = mean((YX×A)²), which serves as the benchmark for predicting uncertainty; Step c: Online update of the channel prediction model: Each time a new quality assessment value Q is obtained _j After (t+1), the AR coefficients are updated online using the Recursive Least Squares (RLS) algorithm; Calculate the prediction error e(t+1)=Q _j (t+1)-Q _pred (j|t+1); Update the gain vector K(t+1) = R(t)^{-1}×X(t+1) / (1+X(t+1)^T×R(t)^{-1}×X(t+1)), where R(t) is the input autocorrelation matrix; Update the AR coefficients A(t+1) = A(t) + K(t+1) × e(t+1); Update the autocorrelation matrix R(t+1) = R(t) + X(t+1) × X(t+1)^T; Update residual variance σ² _res (t+1)=(1-γ)×σ² _res (t)+γ×e(t+1)², where γ is the forgetting factor; Step d: Channel prediction model and adaptive window are updated in tandem. When CUSUM or Page-Hinkley detects a change point, historical data before the change point is cleared, and the AR model is reinitialized using only the new data after the change point; if there are fewer than p new data points, it degenerates into moving average prediction; the initial stage prediction variance σ² after the change point is... _pred The confidence weight w of the counterfactual reward is automatically increased. _conf The corresponding reduction.

7. The multi-link cooperative scheduling method for marine emergency communication according to claim 1, characterized in that, In step 500, the process of predicting the future quality of each unselected link is as follows: Step a: Obtain the most recent p=10 historical quality assessment values ​​for the unselected links: [Q _sat (t),Q _sat (t-1),...,Q _sat (t-9)]; Step b: Obtain the AR coefficients of the unselected links [a] _1 ,sat,a _2 ,sat,...,a _10 [sat] and residual variance σ² _res ,sat; Step c: Calculate the predicted value: Q _pred (sat|t)=a _1 ,sat×Q _sat (t)+a _2 ,sat×Q_sat(t-1) +...+a _10 ,sat×Q _sat (t-9); Step d: Calculate the prediction variance: σ² _pred (sat)=σ² _res ,sat×(1+[Q _sat (t),...,Q _sat (t-9)]×R _sat ^{-1}×[Q _sat (t),...,Q _sat (t-9)]^T), where R _sat This is the autocorrelation matrix of the historical quality of the unselected links; Step e: Repeat the above process for each unselected link to obtain the predicted quality value and prediction variance for all unselected links.

8. The multi-link cooperative scheduling method for marine emergency communication according to claim 7, characterized in that, In step 500, the calculation process for the counterfactual reward is as follows: Counterfactual reward r cf (j)=Q pred (j|s t )-Q actual (s t ,a t ); Among them, Q actual (st,at): The actual quality assessment value of the selected link at the current moment, Q. pred (j|st): If link j were chosen at that time, the quality value of link j estimated by the prediction model. This formula represents the value if link j were chosen instead of a. t The expected performance differences exist between the two.

9. The multi-link cooperative scheduling method for marine emergency communication according to claim 1, characterized in that, The training module in step 500 also includes a dense training signal generation program: Step 1: At the end of each scheduling cycle, calculate: Total training signal r total =rt+λ×Σ _j (wconf(j)×rcf(j)); Where rt is the actual reward, rcf(j) is the counterfactual reward for each unselected link, wconf(j) is the confidence weight, and λ is the counterfactual reward weight coefficient (typical value 0.3). Step 2: Transfer the empirical tuple (s) _t ,a _t ,r total ,s _{t+1} Store it in the experience replay buffer; Step 3: During training, a small batch is randomly sampled from the experience replay buffer, using the total training signal r. total Update the Actor-Critic network as a target reward signal: - Critic network update: Minimize the temporal difference error δ=r total +γ×V(s _{t+1} )-V(s _t ), where V(·) is the state value output by the Critic network, and γ is the discount factor (0.99). - Actor network update: using policy gradient ∇θJ=E[∇θlogπ(a _t |s _t )×δ], where δ is an estimate of the dominance function; Step 4: The Actor-Critic network parameters are updated via backpropagation.

10. The multi-link cooperative scheduling method for marine emergency communication according to claim 9, characterized in that, The training module in step 500 also includes an emergency event generation program: Step a: Virtual emergency event triggering mechanism: The training module has a virtual injection scheduler. During normal communication, the scheduler randomly triggers a virtual emergency injection every N scheduling cycles. Step a: Virtual emergency event generation: Upon triggering, the training module automatically generates a virtual emergency event descriptor, which includes the following fields:

1. Event type; 2. Priority level; 3. Amount of randomly generated data; 4. Latency requirements. Step c: Virtual emergency event execution process: Inject virtual emergency event descriptors into the state space of the training module, including event type and priority information; After receiving a status containing virtual emergency information, the policy network of the training module outputs a decision on which link to select; The training module does not actually execute the link selection, but instead calculates the reward signal for the decision through a counterfactual evaluation mechanism; Generate training experience and store it in the experience replay buffer.