A three-stage sea state unmanned surface vehicle six-degree-of-freedom multi-scale prediction feedforward compensation method

Abstract

The application discloses a kind of six-degree-of-freedom multiscale prediction feedforward compensation methods of three-level sea state unmanned ship, it is related to unmanned system cooperative control and ship motion prediction field.This method obtains unmanned ship motion data, control input data and sea state parameters, and combines unmanned aerial vehicle flight state, reference trajectory and landing stage, to motion data is carried out abnormal identification, filtering, normalization and state fusion;Based on six-degree-of-freedom dynamics model estimation sea wave disturbance, obtain disturbance mode through time-varying sparse block decoupling matrix;According to the average period of sea wave decomposition low-frequency and high-frequency disturbance, use common coding, time series attention and double-branch decoding model for long short-term prediction;The prediction result is converted into deck landing point motion and is weighted fusion according to landing stage, generates feedforward compensation instruction, and feedback control instruction is fused to obtain landing control instruction, and according to prediction residual updates model.The application can improve the stability and adaptability of unmanned aerial vehicle shipborne landing under three-level sea state.

Description

Technical Field

[0001] This invention relates to the fields of cooperative control of unmanned systems and prediction of ship motion, and particularly to a six-degree-of-freedom multi-scale prediction feedforward compensation method for unmanned surface vessels in sea state level 3. Background Technology

[0002] With the development of collaborative operation technology between unmanned surface vessels (USVs) and unmanned aerial vehicles (UAVs), using USVs as maritime take-off and landing platforms for UAVs can be applied to scenarios such as maritime inspection, emergency rescue, environmental monitoring, and maritime reconnaissance. When UAVs autonomously land on the deck of an USV, they need to continuously adjust their flight trajectory according to the deck's position, attitude, and motion trend. However, in sea state 3, USVs are affected by periodic waves, random fluctuations, and the actions of propellers and servos, resulting in six degrees of freedom motion: sway, roll, pitch, roll, and bow. Significant inertial coupling, hydrodynamic coupling, and attitude transition coupling exist between these degrees of freedom, causing the deck landing point to exhibit non-stationary, time-varying, and multi-frequency mixed motion characteristics.

[0003] Existing methods for predicting ship motion typically utilize motion data collected by inertial measurement units (IMUs) or satellite positioning devices. However, maritime data collection may contain issues such as pulse anomalies, random noise, inconsistent sampling times from different sensors, and inconsistencies in coordinate systems, which can easily affect the six-degree-of-freedom state estimation results. Some methods only use normalization to unify the dimensions of different data, making it difficult to simultaneously perform anomaly identification, noise suppression, and multi-source motion state fusion, resulting in the input data of subsequent prediction models still potentially containing significant errors.

[0004] In processing six-degree-of-freedom motion, existing methods typically employ fixed-parameter dynamic models, fixed decoupling matrices, or predict each degree of freedom independently. Since the unmanned surface vessel's speed, significant wave height, average wave period, and control input all change over time, fixed decoupling parameters struggle to adapt to dynamic changes in sea state and navigation status. Furthermore, treating each degree of freedom completely independently fails to reflect the coupling relationships between sway, heave, and pitch, as well as between sway, roll, and bow. Moreover, if the self-control actions generated by the thrusters and servos are not deducted from the six-degree-of-freedom dynamic response, the resulting disturbance sequence will simultaneously include wave disturbances and the unmanned surface vessel's own maneuvering actions, hindering the accurate extraction of motion changes caused by sea state.

[0005] Unmanned surface vessels (USVs) in sea state 3 exhibit significant multi-scale motion characteristics. Low-frequency components are primarily characterized by long-duration, large-amplitude overall hull motion, while high-frequency components are mainly characterized by short-duration, rapidly changing local deck motion. Existing single-time-window or single-branch prediction models struggle to simultaneously account for both long-term low-frequency trends and high-frequency transient changes. The disclosure document explicitly addresses the coexistence of long-period large-amplitude swaying and short-period deck micro-shocks, proposing the use of shared coding, temporal attention, and bi-branch prediction to handle motion information at different time scales.

[0006] Furthermore, existing shipborne UAV landing control typically relies primarily on feedback control based on the current relative position and attitude error between the UAV and the deck. Feedback control requires correction only after motion errors have occurred, which can easily lead to compensation lag when there is rapid deck movement or communication / computing delays. While some predictive control methods can predict ship motion, they do not fully consider the spatial difference between the UAV's reference point and the actual landing point on the deck, nor do they adjust the proportion of low-frequency and high-frequency prediction results according to different landing stages such as long-distance approach, close-range hovering, and contact landing. This makes it difficult to meet the different compensation requirements during the phased landing process of the UAV.

[0007] To address this, a six-degree-of-freedom multi-scale prediction feedforward compensation method for unmanned surface vessels in sea state level 3 is proposed. Summary of the Invention

[0008] The main objective of this invention is to provide a six-degree-of-freedom multi-scale prediction feedforward compensation method for unmanned surface vessels in sea state level 3, which can effectively solve the problems in the background technology.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A six-DOF multi-scale prediction feedforward compensation method for a Class 3 sea state unmanned surface vessel includes the following steps:

[0011] S1. Acquire raw motion data, control input data, and sea state parameters of the unmanned surface vessel (USV) under sea state 3, as well as the target USV's flight status, reference landing trajectory, and landing phase information. The raw motion data includes inertial measurement data and satellite positioning data. The sea state parameters include at least the significant wave height and average wave period. The landing phase includes a long-range approach phase, a close-range hovering phase, and a contact landing phase. Perform anomaly identification, filtering, and normalization on the raw motion data. Then, perform time synchronization, coordinate transformation, and state fusion on the processed inertial measurement data and satellite positioning data to obtain the USV's six-degree-of-freedom pose vector. Six-degree-of-freedom velocity vector and the six-degree-of-freedom acceleration vector ;

[0012] S2. Determine the estimated vectors of control force and control torque corresponding to the unmanned surface vessel's self-control action based on the control input data. And calculate the wave disturbance estimation vector based on the six-degree-of-freedom dynamic model of the unmanned surface vessel:

[0013]

[0014] in, The inertia matrix, The matrix of Coriolis force and centripetal force. For hydrodynamic damping matrix, This represents the vector of the restoring force and restoring moment in still water.

[0015] S3. Dimensionlessly transform the estimated wave disturbance vector according to the reference quantities corresponding to each of the six degrees of freedom to obtain a standardized wave disturbance vector; based on the sea state parameters and the current six-degree-of-freedom state of the unmanned surface vessel (USV), locally linearize the USV's six-degree-of-freedom dynamic model, extract the coupling coefficients between the six degrees of freedom, and construct a sixth-order local coupling response matrix. With the optimization objective of reducing the off-diagonal coupling after matrix transformation and the matrix changes between adjacent time steps, an invertible and normalized time-varying decoupling matrix is ​​solved within a pre-defined sparse block matrix set. And determine the perturbation mode vector:

[0016]

[0017] in, The standardized wave disturbance vector;

[0018] S4. Determine the frequency band division parameters based on the average wave period, and decompose the disturbance mode vector into a low-frequency disturbance mode sequence and a high-frequency disturbance mode sequence; input the low-frequency disturbance mode sequence, the high-frequency disturbance mode sequence, and the corresponding six-degree-of-freedom state sequence into a multi-scale temporal prediction model to obtain the low-frequency six-degree-of-freedom motion prediction sequence in the first prediction time domain, the high-frequency six-degree-of-freedom motion prediction sequence in the second prediction time domain, and the corresponding prediction confidence; wherein, the first prediction time domain is larger than the second prediction time domain; wherein, the multi-scale temporal prediction model includes a common coding layer, a temporal attention layer, a low-frequency long temporal decoding branch, and a high-frequency short temporal decoding branch, and the first prediction time domain is larger than the second prediction time domain;

[0019] S5. Based on the fixed spatial relationship between the unmanned surface vessel reference point and the deck landing point, the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence are converted into a low-frequency deck landing point motion prediction sequence and a high-frequency deck landing point motion prediction sequence, respectively; based on the landing phase information, the low-frequency prediction weight and the high-frequency prediction weight are determined, and the low-frequency deck landing point motion prediction sequence and the high-frequency deck landing point motion prediction sequence are weighted and fused to obtain the deck landing point fusion prediction result;

[0020] S6. Based on the deck landing point fusion prediction result, the flight status of the target UAV, and the reference landing trajectory, determine the predicted relative motion error, generate a feedforward compensation command based on the predicted relative motion error, and fuse the feedforward compensation command with the feedback control command of the target UAV to obtain the landing control command:

[0021]

[0022] in, For landing control commands, For feedback control commands, The feedforward enable coefficient, This is the feedforward compensation gain matrix corresponding to the current landing phase. To predict relative motion error, The compensation time corresponding to the current landing phase, To compensate for the amplitude limiting function;

[0023] S7. Save the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence and their prediction target time. After reaching the prediction target time, perform time matching between the actual six-degree-of-freedom motion data and the corresponding prediction sequence, calculate the prediction residual, and update the adaptive parameters of the multi-scale time series prediction model when the prediction residual meets the preset update conditions.

[0024] Furthermore, in step S1, for each degree of freedom data in the original motion acquisition data, the median and median absolute deviation are determined respectively based on the historical sliding window excluding the current data;

[0025] When the Current data for each degree of freedom satisfy:

[0026]

[0027] When the current data is identified as pulse abnormal data, the pulse abnormal data is replaced with the middle value of the historical sliding window, the interpolation result of adjacent valid data, or the state prediction value.

[0028] in, The median of the historical sliding window. The median absolute deviation of the historical sliding window. The anomaly detection coefficient;

[0029] After replacing the outlier data, perform state filtering on the data and normalize it to the maximum and minimum values ​​according to the following formula:

[0030]

[0031] in, The state-filtered first One degree of freedom data, and These are the valid maximum and minimum values ​​within the reference window, respectively. It is a preset positive number.

[0032] Furthermore, in step S1, the inertial measurement data and satellite positioning data are synchronized in time and transformed in coordinates, and then fused through a state estimation model to determine the position, attitude, linear velocity, angular velocity and linear acceleration of the unmanned surface vessel.

[0033] In step S2, the control input vector is determined based on the thruster operating data and the servo motor operating data. The control force and control torque estimation vectors are determined according to the following formula:

[0034]

[0035] in, This is the control allocation matrix determined based on the calibration data of the thrusters and servos.

[0036] Furthermore, in step S3, sway, heave, and pitch are divided into longitudinal coupling groups, and sway, roll, and bow roll are divided into transverse coupling groups.

[0037] The time-varying decoupling matrix includes a longitudinal motion decoupling sub-matrix corresponding to the longitudinal coupling group and a transverse motion decoupling matrix corresponding to the transverse coupling group. Off-diagonal elements are allowed within the same decoupling sub-matrix, and matrix elements between different decoupling sub-matrixes are set to zero.

[0038] The time-varying decoupling matrix is ​​updated based on the significant wave height, average wave period, unmanned surface vessel speed, and historical decoupling residuals.

[0039] Furthermore, in step S3, each row of the time-varying decoupling matrix is ​​normalized, and the condition number of the time-varying decoupling matrix is ​​limited to within a preset condition number threshold.

[0040] The time-varying decoupling matrix is ​​determined according to the following formula:

[0041]

[0042] in, For a predefined set of sparse block matrices, This indicates the extraction of off-diagonal elements from the matrix. These are the matrix transformation constraint coefficients. This represents the Frobenius norm.

[0043] Furthermore, in step S4, based on the average wave period... Determine the cutoff frequency:

[0044]

[0045] in, The cutoff frequency, The preset frequency adjustment coefficient;

[0046] A low-pass processing method with the cutoff frequency as the frequency band boundary is used to extract the low-frequency disturbance mode vector from the disturbance mode vector, and the difference between the disturbance mode vector and the low-frequency disturbance mode vector is determined as the high-frequency disturbance mode vector.

[0047] A low-frequency long-time-series input window is constructed based on low-frequency disturbance mode vectors at multiple consecutive time points, and a high-frequency short-time-series input window is constructed based on high-frequency disturbance mode vectors at multiple consecutive time points. The duration of the low-frequency long-time-series input window is greater than the duration of the high-frequency short-time-series input window.

[0048] Furthermore, in step S4, the low-frequency perturbation mode sequence, the high-frequency perturbation mode sequence, and the corresponding six-degree-of-freedom state sequence are time-aligned and combined into a multi-scale input sequence;

[0049] The common coding layer uses a shared long short-term memory network to extract the shared temporal features of the multi-scale input sequence, and the temporal attention layer is used to determine the temporal weights based on the shared temporal features at different times.

[0050] The low-frequency long-time-series decoding branch outputs a low-frequency six-degree-of-freedom motion prediction sequence for the next 2 seconds based on the shared timing features and timing weights, while the high-frequency short-time-series decoding branch outputs a high-frequency six-degree-of-freedom motion prediction sequence for the next 0.5 seconds based on the shared timing features and timing weights.

[0051] Furthermore, in step S5, for each prediction moment in the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence, the position, attitude, velocity, and angular velocity of the unmanned surface vessel (USV) reference point corresponding to that prediction moment are substituted into the deck landing point motion transformation relationship to determine the predicted position and predicted velocity of the deck landing point.

[0052]

[0053]

[0054] in, The speed of the unmanned surface vessel as a reference point. For the angular velocity of the unmanned surface vessel, This refers to any prediction moment in the low-frequency six-degree-of-freedom motion prediction sequence or the high-frequency six-degree-of-freedom motion prediction sequence;

[0055] The sum of the low-frequency prediction weight and the high-frequency prediction weight is 1; the low-frequency prediction weight is increased during the long-distance approach phase, and the high-frequency prediction weight is increased during the close-range hovering phase and the contact landing phase.

[0056] In step S6, the feedforward enable coefficient is determined based on the prediction confidence, the reliability of the original motion acquisition data, and the communication status between the unmanned surface vessel and the target UAV. When the prediction confidence is lower than a preset confidence threshold, the reliability of the original motion acquisition data is lower than a preset data threshold, or the communication delay exceeds a preset delay threshold, the feedforward enable coefficient is set to zero, and the amplitude and rate of change of the allowed feedforward compensation command are limited.

[0057] A six-DOF multi-scale prediction feedforward compensation system for a Class 3 sea state unmanned surface vessel (USV) includes a motion data acquisition interface, a control input interface, a sea state data acquisition interface, a USV data interface, a communication unit, a memory, and a processor. The memory stores a computer program executable by the processor, and the processor is configured with:

[0058] The data processing module is used to perform abnormal data identification, filtering, normalization and state fusion on the raw motion acquisition data obtained through the motion data acquisition interface to obtain the six degrees of freedom state of the unmanned surface vessel.

[0059] The disturbance estimation module is used to determine the wave disturbance estimation vector based on the six degrees of freedom state of the unmanned surface vessel and the control input data obtained through the control input interface;

[0060] The time-varying decoupling module is used to construct a sixth-order local coupling response matrix based on the sea state parameters obtained through the sea state data acquisition interface and the current six-degree-of-freedom state of the unmanned surface vessel, determine the time-varying decoupling matrix through optimization, and determine the disturbance mode vector based on the time-varying decoupling matrix.

[0061] The multi-scale prediction module is used to decompose the perturbation mode vector into a low-frequency perturbation mode sequence and a high-frequency perturbation mode sequence, and obtain the low-frequency six-degree-of-freedom motion prediction sequence, the high-frequency six-degree-of-freedom motion prediction sequence, and the corresponding prediction confidence through the multi-scale time series prediction model;

[0062] The landing point prediction module is used to convert the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence into a deck landing point motion prediction sequence based on the fixed spatial positional relationship between the unmanned surface vessel reference point and the deck landing point.

[0063] The feedforward compensation module is used to generate feedforward compensation instructions based on the deck landing point motion prediction sequence and the target UAV flight status, reference landing trajectory and landing phase information obtained through the UAV data interface, and send the feedforward compensation instructions to the flight controller of the target UAV through the communication unit.

[0064] The online correction module is used to update the adaptive parameters of the multi-scale time series prediction model based on the prediction residual between the actual six-degree-of-freedom motion data and the corresponding prediction sequence.

[0065] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. This invention sequentially performs anomaly identification, state filtering, normalization, time synchronization, coordinate transformation, and state fusion on inertial measurement data and satellite positioning data. This can reduce the impact of pulse anomalies, random noise, and asynchronous multi-source data on six-degree-of-freedom state estimation, and uniformly obtain the six-degree-of-freedom pose, velocity, and acceleration of the unmanned surface vessel, providing consistent state data for dynamic disturbance estimation and subsequent time series prediction.

[0068] 2. This invention determines the control force and control torque corresponding to the thruster and servo motor based on the control input data of the unmanned surface vessel (USV), and uses a six-degree-of-freedom dynamic model to subtract its own control action from the overall dynamic response of the USV to obtain the estimated wave disturbance vector. Furthermore, by locally linearizing the dynamic model to construct a sixth-order local coupling response matrix, and optimizing the solution of the time-varying decoupling matrix within a preset sparse block matrix set, this invention can reduce the coupling interference between different degrees of freedom of motion while preserving the main physical correlations within the longitudinal and lateral coupling groups, thereby improving the adaptability of the disturbance mode extraction results to dynamic sea conditions.

[0069] 3. This invention decomposes the disturbance mode into low-frequency disturbance mode sequences and high-frequency disturbance mode sequences based on the average wave period. It then uses a multi-scale temporal prediction model, including a common coding layer, a temporal attention layer, a low-frequency long-time-series decoding branch, and a high-frequency short-time-series decoding branch, to predict low-frequency motion trends in a longer prediction time domain and high-frequency transient changes in a shorter prediction time domain, respectively. The common coding layer reduces redundant feature extraction between different prediction branches, while the temporal attention layer highlights temporal features that have a significant impact on the current motion prediction, thus balancing the prediction needs of both long-period large-amplitude motions and short-period rapid motions.

[0070] 4. This invention converts the predicted six-degree-of-freedom motion of the unmanned surface vessel (USV) into the predicted position and velocity of the deck landing point based on the fixed spatial relationship between the USV reference point and the deck landing point. It dynamically adjusts the low-frequency and high-frequency prediction weights according to the long-distance approach, close-range hovering, and contact landing phases. Furthermore, it generates feedforward compensation commands based on the predicted relative motion error between the fused deck landing point prediction and the USV's reference landing trajectory, enabling the USV to pre-correct deck motion errors before they actually materialize. Simultaneously, by incorporating prediction confidence, data reliability, communication status, compensation limiting, and an online prediction residual update mechanism, this invention improves the stability, reliability, and long-term sea state adaptability of the feedforward compensation process. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0072] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] Example 1

[0075] like Figure 1-2 As shown, a six-degree-of-freedom multi-scale prediction feedforward compensation method for unmanned surface vessels (USVs) in sea state 3 is used to predict the future motion state of the USV's deck landing point when the USV is subjected to sea waves in sea state 3, and to generate feedforward compensation commands for the shipborne landing process of the target USV.

[0076] In this embodiment, the unmanned surface vessel (USV) is equipped with an inertial measurement unit (IMU), a satellite positioning device, a sea state sensing device, a thruster controller, a servo controller, and an onboard computing device. It communicates with the flight controller of the target unmanned aerial vehicle (UAV) via a wireless communication link. The IMU collects the USV's three-axis acceleration and three-axis angular velocity; the satellite positioning device collects the USV's position and velocity; and the sea state sensing device acquires the significant wave height, average wave period, and selectable wave direction. The thruster controller and servo controller provide control input data such as thruster speed, thrust command, and rudder angle, respectively.

[0077] The target UAV's flight controller provides the onboard computing device with the UAV's position, speed, attitude, reference landing trajectory, and current landing phase information. The landing phase includes a long-range approach phase, a close-range hovering phase, and a contact landing phase.

[0078] This embodiment retains the technical concepts proposed in the disclosure document, such as six-degree-of-freedom dynamic modeling, sea state adaptive decoupling, common coding layer, temporal attention layer, low-frequency long-time-series prediction branch, high-frequency short-time-series prediction branch, and online residual correction.

[0079] Coordinate system and six-degree-of-freedom state definition

[0080] Establish a geographic coordinate system and an unmanned surface vessel (USV) hull coordinate system. The six-degree-of-freedom pose vector of the USV is represented as:

[0081]

[0082] in, , and These represent the longitudinal, lateral, and vertical positions of the unmanned surface vessel, respectively. , and These represent the roll angle, pitch angle, and bow angle, respectively.

[0083] The six-degree-of-freedom velocity vector of an unmanned surface vessel is represented as:

[0084]

[0085] in, , and These represent the longitudinal velocity, lateral velocity, and vertical velocity in the ship's coordinate system, respectively. , and These represent the roll rate, pitch rate, and yaw rate, respectively.

[0086] The six-degree-of-freedom acceleration vector is:

[0087]

[0088] Step S1: Motion data preprocessing and state fusion

[0089] Acquire raw motion data, control input data, and sea state parameters of the unmanned surface vessel in sea state 3, as well as the flight status, reference landing trajectory, and landing phase information of the target UAV.

[0090] The raw motion acquisition data includes inertial measurement data and satellite positioning data. Inertial measurement data can be acquired at a frequency of not less than 100Hz, while satellite positioning data can be acquired at a frequency of 5Hz to 20Hz. The onboard computing unit performs time alignment based on the timestamps carried by each data point and maps the low-frequency satellite positioning data to the corresponding time series of the inertial measurement data through interpolation or state prediction.

[0091] Pulse anomaly data identification

[0092] Regarding the first Current data for each degree of freedom Create a historical sliding window that does not include current data:

[0093]

[0094] in, This is the length of the sliding window. In one embodiment, Take 20 to 100 sampling points.

[0095] Calculate the median of the historical sliding window:

[0096]

[0097] Calculate the absolute deviation of the median:

[0098]

[0099] When the following conditions are met:

[0100]

[0101] At that time, the current data is determined to be pulse abnormal data. Among them, This is the anomaly detection coefficient, which can be set according to different sensor accuracies and data fluctuation ranges with different degrees of freedom.

[0102] For data identified as pulse anomalies, the median of the historical sliding window, the interpolation results of adjacent valid data, or the predicted value output by the state estimation model can be used for replacement.

[0103] State filtering

[0104] State filtering is performed on the inertial measurement data and satellite positioning data after replacing the abnormal data. This embodiment can use an unscented Kalman filter, an extended Kalman filter, or other filters suitable for six-degree-of-freedom motion state estimation.

[0105] The filter's state variables include at least position, attitude, linear velocity, angular velocity, and the zero-bias parameter of the inertial measurement unit. Inertial measurement data is used to perform high-frequency state recursion, while satellite positioning data is used to correct the position and velocity states.

[0106] The linear acceleration output by the inertial measurement unit is subjected to gravity component elimination and installation error compensation, and coordinate transformation is completed through a rotation matrix between the ship coordinate system and the geographic coordinate system.

[0107] Maximum-Minimum Normalization

[0108] After state filtering, the first Data of degrees of freedom Perform maximum-minimum normalization:

[0109]

[0110] in, and Each of the following is the first in the reference window The effective maximum and effective minimum values ​​of each degree of freedom To prevent the default positive number with a denominator of zero.

[0111] It should be noted that the maximum-minimum normalization in this embodiment is mainly used to unify the numerical scale of data with different degrees of freedom; the identification of pulse anomaly data is achieved by the aforementioned median absolute deviation method, and random noise suppression is achieved by the state filtering process.

[0112] After time synchronization, coordinate transformation, and state fusion, the six-DOF pose vector of the unmanned surface vessel is obtained. Six-degree-of-freedom velocity vector and the six-degree-of-freedom acceleration vector .

[0113] Step S2: Wave Disturbance Estimation

[0114] Construct control input vectors based on thruster and servo motor operation data:

[0115]

[0116] in, to This indicates the rotational speed, thrust command, or equivalent control quantity of each thruster. Represents the rudder angle.

[0117] Determine the control allocation matrix based on the calibration data of the thrusters and servos. Then, the estimated vectors of control force and control torque corresponding to the unmanned surface vessel's own control action are calculated:

[0118]

[0119] The six-degree-of-freedom dynamic model of the unmanned surface vessel is expressed as follows:

[0120]

[0121] in:

[0122] The inertia matrix includes the rigid body mass and the added mass; The matrix represents the Coriolis force and centripetal force. The hydrodynamic damping matrix; This represents the vector of the restoring force and restoring moment in still water. The control force and control torque generated by the unmanned surface vessel's own control action; The generalized disturbance force and disturbance moment generated by the action of ocean waves.

[0123] Therefore, the estimated wave disturbance vector is calculated:

[0124]

[0125] By subtracting the control actions corresponding to the thrusters and servos, the impact of the unmanned surface vessel's own maneuvering actions on the wave disturbance estimation results is reduced.

[0126] Step S3: Standardization and Time-varying Sparse Block Decoupling

[0127] Since the first three translational degrees of freedom correspond to stress, and the last three rotational degrees of freedom correspond to stress torque, their physical dimensions and numerical ranges are different. Therefore, the wave disturbance estimation vector is made dimensionless according to the reference quantities corresponding to each degree of freedom.

[0128] Construct the reference matrix:

[0129]

[0130] in, , and The reference forces corresponding to the three translational degrees of freedom are: , and These are the reference torques corresponding to the three rotational degrees of freedom.

[0131] The standardized wave disturbance vector is:

[0132]

[0133] Construction of Local Coupling Response Matrix

[0134] The dynamics model of the unmanned surface vessel is represented as a local state mapping:

[0135]

[0136] in, This represents the sea state parameters, which consist of the significant wave height, the average wave period, and the optional direction of incoming waves.

[0137] At the current six-degree-of-freedom state, the dynamic model is locally linearized, the local coupling coefficients between each degree of freedom are extracted, and a sixth-order local coupling response matrix is ​​constructed:

[0138]

[0139] In other implementations, the local coupled response matrix can also be identified online using the recursive least squares method based on the six-degree-of-freedom state changes and the standardized wave disturbance changes at adjacent time points.

[0140] Sparse block structure

[0141] Sway, heave, and pitch are classified into a longitudinal coupling group, while sway, roll, and bow roll are classified into a transverse coupling group.

[0142] The time-varying decoupling matrix includes a longitudinal motion decoupling submatrix and a lateral motion decoupling submatrix, which can be represented as:

[0143]

[0144] in, For longitudinal motion decoupling submatrix, For lateral motion decoupling submatrix, This is the permutation matrix determined based on the order of the six degrees of freedom.

[0145] Off-diagonal elements are allowed within the same decoupling submatrix to reduce mutual influence between different degrees of freedom within the same coupling group; matrix elements between two decoupling submatrices are set to zero to reduce unnecessary calculations between degrees of freedom that do not have a major physical coupling relationship.

[0146] Optimization solution of time-varying decoupling matrix

[0147] In the preset sparse block matrix set Internal solution for the time-varying decoupling matrix:

[0148]

[0149] The first term reduces the off-diagonal coupling after the matrix transformation, and the second term limits the change in the decoupling matrix between adjacent time steps. These are the matrix transformation constraint coefficients.

[0150] During the solution process, each row of the time-varying decoupling matrix is ​​normalized, and its condition number is limited to a preset condition number threshold. When the condition number of the solution exceeds the threshold, the matrix change constraint coefficient is increased or the effective decoupling matrix from the previous time step is used to avoid ill-conditioned decoupling matrix caused by sudden changes in sea state.

[0151] The local coupling response matrix and optimization parameters are dynamically updated based on the significant wave height, average wave period, unmanned surface vessel speed, and historical decoupling residuals, thereby enabling the time-varying decoupling matrix to adapt to changes in sea state and vessel motion.

[0152] The final perturbation mode vector is obtained as follows:

[0153]

[0154] Step S4: High and low frequency disturbance decomposition and multi-scale prediction

[0155] Frequency band division

[0156] Based on the average wave period Determine the cutoff frequency:

[0157]

[0158] in, This is the preset frequency adjustment coefficient.

[0159] Adopted For causal low-pass processing at the cutoff frequency, the low-frequency perturbation mode vector is extracted from the perturbation mode vector:

[0160]

[0161] in, This indicates the low-pass processing operator.

[0162] The high-frequency disturbance mode vector is:

[0163]

[0164] A long-duration low-frequency input window is constructed based on low-frequency disturbance mode vectors from multiple consecutive time points, and a short-duration high-frequency input window is constructed based on high-frequency disturbance mode vectors from multiple consecutive time points. The duration of the long-duration low-frequency input window is longer than that of the short-duration high-frequency input window.

[0165] For example, a low-frequency long-time-series input window can include data from the past 4 to 10 seconds, while a high-frequency short-time-series input window can include data from the past 0.5 to 2 seconds.

[0166] Multi-scale input sequence construction

[0167] The low-frequency perturbation mode sequence, the high-frequency perturbation mode sequence, and the corresponding six-degree-of-freedom state sequence are aligned according to timestamps, and a multi-scale input sequence is constructed.

[0168] For sampling time Multi-scale input features can be represented as:

[0169]

[0170] in, The effective wave height.

[0171] Common coding layer

[0172] The common coding layer uses a shared long short-term memory network to extract shared temporal features of multi-scale input sequences.

[0173] For the Input features at each time step Long Short-Term Memory (LSTM) units are calculated in the following manner:

[0174]

[0175]

[0176]

[0177]

[0178]

[0179]

[0180] in, To share the hidden state, For the unit state, function, This indicates that corresponding elements are multiplied.

[0181] Temporal attention layer

[0182] The temporal attention layer determines the temporal weights based on the shared hidden states at each time step:

[0183]

[0184]

[0185]

[0186] in, For the first The temporal weight at each moment, The shared temporal features are weighted by attention.

[0187] Two-branch prediction

[0188] The attention-weighted shared temporal features are input into the low-frequency long-time-series decoding branch and the high-frequency short-time-series decoding branch, respectively.

[0189] The low-frequency long-time decoding branch outputs a low-frequency six-degree-of-freedom motion prediction sequence for the first prediction time domain. In this embodiment, the first prediction time domain is 2 seconds.

[0190] The high-frequency short-time decoding branch outputs a high-frequency six-degree-of-freedom motion prediction sequence for the future second prediction time domain. In this embodiment, the second prediction time domain is 0.5 seconds.

[0191] If the model input sampling period is Then the low-frequency prediction sequence is represented as:

[0192]

[0193] in:

[0194]

[0195] The high-frequency prediction sequence is represented as:

[0196]

[0197] in:

[0198]

[0199] Model training

[0200] During the offline training phase of the model, historical six-degree-of-freedom motion data, control input data, and sea state parameters of the unmanned surface vessel under sea state level 3 are collected, and low-frequency disturbance mode sequences, high-frequency disturbance mode sequences, and corresponding future actual motion sequences are generated according to the aforementioned steps.

[0201] The actual low-frequency six-degree-of-freedom motion sequence following the current input window is used as the training label for the low-frequency branch, and the actual high-frequency six-degree-of-freedom motion sequence following the current input window is used as the training label for the high-frequency branch.

[0202] The model's total loss function can be expressed as:

[0203]

[0204] in, For low-frequency prediction loss, For high-frequency prediction loss, For model parameter regularization loss, , and The corresponding loss weights.

[0205] Low-frequency prediction loss and high-frequency prediction loss can be calculated using mean squared error, mean absolute error, or a weighted combination of the two.

[0206] Prediction confidence

[0207] The prediction confidence level is determined based on the prediction error of the prediction model within the historical residual window. For example, the prediction confidence level of the b-th prediction branch is expressed as:

[0208]

[0209] in, For the first The average prediction error of each prediction branch within the historical residual window This is the error scale parameter corresponding to this branch.

[0210] Step S5: Deck landing point motion prediction and multi-scale fusion

[0211] The predicted motion of an unmanned surface vessel (USV) in six degrees of freedom is usually based on the USV's center of mass or a preset hull reference point, while the actual landing position of the USV is at the deck landing point. Because there is a fixed lever between the deck landing point and the USV's reference point, the rotation of the hull will cause additional linear displacement and additional linear velocity at the deck landing point.

[0212] For each prediction moment in the low-frequency and high-frequency six-degree-of-freedom motion prediction sequences, the predicted position and velocity of the deck landing point are determined based on the predicted position, attitude, velocity, and angular velocity of the unmanned surface vessel reference point.

[0213] For any prediction time The predicted location of the landing point on the deck is:

[0214]

[0215] The predicted speed at the deck landing point is:

[0216]

[0217] in:

[0218] The position of the unmanned surface vessel reference point at the predicted time; The velocity of the unmanned surface vessel at the reference point at the predicted time; The attitude of the unmanned surface vessel at the predicted moment; Let be the angular velocity of the unmanned surface vessel at the predicted time. This is a fixed position vector from the unmanned surface vessel's reference point to its landing point on the deck. The rotation matrix is ​​determined based on the attitude of the unmanned surface vessel; This represents the time interval between the predicted time and the current time in the predicted sequence.

[0219] The above transformations are performed on the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence respectively to obtain the low-frequency deck landing point motion prediction sequence and the high-frequency deck landing point motion prediction sequence.

[0220] Determine the low-frequency prediction weights based on the landing phase of the target UAV. and high-frequency prediction weights And satisfy:

[0221]

[0222] During the long-distance approach phase, the low-frequency prediction weight is greater than the high-frequency prediction weight, so that the UAV mainly follows the long-term motion trend of the unmanned surface vessel deck.

[0223] During the close-range hovering phase, the weight of high-frequency predictions is gradually increased to enhance the response capability to short-term rapid changes in the deck.

[0224] During the contact and landing phase, the weight of high-frequency prediction is further increased and the compensation lead time is shortened to reduce the impact of rapid local deck movement on the landing stability of the UAV at the moment of contact.

[0225] The deck landing point fusion prediction results are as follows:

[0226]

[0227] in, This is the prediction result for low-frequency deck landing point motion. This is the result of the high-frequency deck landing point motion prediction.

[0228] For cases where the low-frequency branch and the high-frequency branch have different sampling intervals or prediction time domains, time interpolation can be used to map the prediction results of the two branches to the current compensation advance time. Weighted fusion is then performed at the same prediction time.

[0229] Step S6: Feedforward Compensation Command Generation and Control Fusion

[0230] The prediction relative motion error is determined based on the deck landing point fusion prediction results, the flight status of the target UAV, and the reference landing trajectory.

[0231] The predicted relative motion error may include the predicted relative position error, the predicted relative velocity error, and the predicted relative attitude error:

[0232]

[0233] in:

[0234]

[0235]

[0236] The location corresponding to the reference landing trajectory of the target drone. The speed corresponding to the reference landing trajectory of the target drone.

[0237] Generate feedforward compensation instructions based on the predicted relative motion error:

[0238]

[0239] in, This is the feedforward compensation gain matrix corresponding to the current landing phase.

[0240] The feedforward enable coefficient is determined based on the prediction confidence level, the reliability of the raw motion acquisition data, and the communication status between the unmanned surface vessel and the target unmanned aerial vehicle. .

[0241] In one embodiment:

[0242]

[0243] in, To predict confidence levels, To ensure the reliability of the raw motion data, This is due to communication delay.

[0244] When the prediction confidence is lower than the preset confidence threshold, the confidence of the original motion acquisition data is lower than the preset data threshold, or the communication delay exceeds the preset delay threshold, the feedforward enable coefficient is set to zero, so that the target UAV only uses the original feedback control commands.

[0245] When outputting feedforward compensation commands is allowed, the amplitude and rate of change of the feedforward compensation commands are limited to prevent short-term sudden changes in the prediction results from causing control command jumps.

[0246] The final landing control command is:

[0247]

[0248] in, The feedback control commands generated by the target UAV based on the current actual flight error.

[0249] In one distributed implementation, the onboard computing device is responsible for generating feedforward compensation commands and sending them to the flight controller of the target UAV via a communication unit. The flight controller of the target UAV is responsible for fusing the received feedforward compensation commands with its internally generated feedback control commands to obtain the final landing control commands. This distributed approach corresponds to the feedforward compensation command generation and transmission structure defined in claim 9.

[0250] Step S7: Predicted Residual Calculation and Online Update

[0251] Save the following each time you output the low-frequency six-DOF motion prediction sequence and the high-frequency six-DOF motion prediction sequence:

[0252] Prediction generation time; target time for each prediction; low-frequency prediction results; high-frequency prediction results; corresponding sea state parameters; current model version identifier.

[0253] When the time reaches a certain predicted target time, the actual six-degree-of-freedom motion data at that time is obtained and matched with the previously saved prediction results according to the timestamp.

[0254] The low-frequency prediction residual is expressed as:

[0255]

[0256] The high-frequency prediction residual is expressed as:

[0257]

[0258] Calculate the residual loss within the sliding residual window:

[0259]

[0260] When the online residual loss exceeds the preset update threshold, and the current data credibility and communication status meet the preset update conditions, the adaptive parameters of the multi-scale time series prediction model are updated.

[0261] To avoid disrupting common temporal features during online updates, it is preferable to keep the basic parameters of the common coding layer unchanged and only update the output layers of the temporal attention layer, the low-frequency long temporal decoding branch, the high-frequency short temporal decoding branch, or the additional adaptive correction layer.

[0262] Model parameter updates can be represented as:

[0263]

[0264] in, For the adaptive parameters to be updated, For online learning rate, Indicates gradient limiting, This is the gradient limiting threshold.

[0265] The updated model is first tested for error within a preset verification window. If the error of the updated model is lower than that of the model before the update, the updated model parameters replace the current online model parameters; otherwise, the model parameters before the update continue to be used.

[0266] By using delay matching and conditional online updates between prediction results and actual motion data, the performance degradation caused by unmanned surface vessel load changes, sensor drift, and slow changes in sea state can be suppressed.

[0267] System operation mode

[0268] The system in this embodiment can be deployed in a centralized or distributed manner.

[0269] When a centralized deployment is adopted, motion data preprocessing, wave disturbance estimation, time-varying decoupling, multi-scale prediction, deck landing point conversion, and feedforward compensation command generation are all completed by the onboard computing device, which sends landing control-related data to the UAV flight controller.

[0270] When a distributed deployment is adopted, the onboard computing device completes the six-degree-of-freedom state estimation, wave disturbance estimation, time-varying decoupling, deck motion prediction, and feedforward compensation command generation of the unmanned surface vessel. The UAV flight controller completes the generation of feedback control commands and the fusion of feedforward compensation commands and feedback control commands.

[0271] When the prediction confidence, data credibility, or communication status does not meet the preset conditions, the system stops using prediction feedforward compensation, and the target UAV continues to use its original feedback control method to complete hovering, go-around, or landing operations.

[0272] Through the above technical solution, the present invention can estimate wave disturbances from six-degree-of-freedom motion data including the unmanned surface vessel's own maneuvering in sea state 3, reduce the coupling effect between the main degrees of freedom by using a time-varying sparse block decoupling matrix, and take into account both low-frequency large-amplitude motion and high-frequency fast motion by using a dual-branch time series prediction model with different prediction time domains, thereby compensating for the UAV landing control in advance based on the future motion state of the deck landing point.

[0273] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A six-degree-of-freedom multi-scale prediction feedforward compensation method for a Class 3 sea state unmanned surface vessel, characterized in that, Includes the following steps: S1. Acquire raw motion data, control input data, and sea state parameters of the unmanned surface vessel (USV) under sea state 3, as well as the target USV's flight status, reference landing trajectory, and landing phase information. The raw motion data includes inertial measurement data and satellite positioning data. The sea state parameters include at least the significant wave height and average wave period. The landing phase includes a long-range approach phase, a close-range hovering phase, and a contact landing phase. Perform anomaly identification, filtering, and normalization on the raw motion data. Then, perform time synchronization, coordinate transformation, and state fusion on the processed inertial measurement data and satellite positioning data to obtain the USV's six-degree-of-freedom pose vector. Six-degree-of-freedom velocity vector and the six-degree-of-freedom acceleration vector ; S2. Determine the estimated vectors of control force and control torque corresponding to the unmanned surface vessel's self-control action based on the control input data. And calculate the wave disturbance estimation vector based on the six-degree-of-freedom dynamic model of the unmanned surface vessel: in, The inertia matrix, The matrix of Coriolis force and centripetal force. For hydrodynamic damping matrix, This represents the vector of the restoring force and restoring moment in still water. S3. Dimensionlessly transform the estimated wave disturbance vector according to the reference quantities corresponding to each of the six degrees of freedom to obtain a standardized wave disturbance vector; based on the sea state parameters and the current six-degree-of-freedom state of the unmanned surface vessel (USV), locally linearize the USV's six-degree-of-freedom dynamic model, extract the coupling coefficients between the six degrees of freedom, and construct a sixth-order local coupling response matrix. With the optimization objective of reducing the off-diagonal coupling after matrix transformation and the matrix change between adjacent time steps, an invertible and normalized time-varying decoupling matrix is ​​solved within a pre-defined sparse block matrix set. And determine the perturbation mode vector: in, The standardized wave disturbance vector; S4. Determine the frequency band division parameters based on the average wave period, and decompose the disturbance mode vector into a low-frequency disturbance mode sequence and a high-frequency disturbance mode sequence; input the low-frequency disturbance mode sequence, the high-frequency disturbance mode sequence, and the corresponding six-degree-of-freedom state sequence into a multi-scale temporal prediction model to obtain the low-frequency six-degree-of-freedom motion prediction sequence in the first prediction time domain, the high-frequency six-degree-of-freedom motion prediction sequence in the second prediction time domain, and the corresponding prediction confidence; wherein, the first prediction time domain is larger than the second prediction time domain; wherein, the multi-scale temporal prediction model includes a common coding layer, a temporal attention layer, a low-frequency long temporal decoding branch, and a high-frequency short temporal decoding branch, and the first prediction time domain is larger than the second prediction time domain; S5. Based on the fixed spatial relationship between the unmanned surface vessel reference point and the deck landing point, the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence are converted into a low-frequency deck landing point motion prediction sequence and a high-frequency deck landing point motion prediction sequence, respectively; based on the landing phase information, the low-frequency prediction weight and the high-frequency prediction weight are determined, and the low-frequency deck landing point motion prediction sequence and the high-frequency deck landing point motion prediction sequence are weighted and fused to obtain the deck landing point fusion prediction result; S6. Based on the deck landing point fusion prediction result, the flight status of the target UAV, and the reference landing trajectory, determine the predicted relative motion error, generate a feedforward compensation command based on the predicted relative motion error, and fuse the feedforward compensation command with the feedback control command of the target UAV to obtain the landing control command: in, For landing control commands, For feedback control commands, The feedforward enable coefficient, This is the feedforward compensation gain matrix corresponding to the current landing phase. To predict relative motion error, The compensation time corresponding to the current landing phase, To compensate for the amplitude limiting function; S7. Save the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence and their prediction target time. After reaching the prediction target time, perform time matching between the actual six-degree-of-freedom motion data and the corresponding prediction sequence, calculate the prediction residual, and update the adaptive parameters of the multi-scale time series prediction model when the prediction residual meets the preset update conditions.

2. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a Class III sea state unmanned surface vessel according to claim 1, characterized in that, In step S1, for each degree of freedom data in the original motion acquisition data, the median and median absolute deviation are determined respectively based on the historical sliding window excluding the current data; When the Current data for each degree of freedom satisfy: When the current data is identified as pulse abnormal data, the pulse abnormal data is replaced with the middle value of the historical sliding window, the interpolation result of adjacent valid data, or the state prediction value. in, The median of the historical sliding window. The median absolute deviation of the historical sliding window. The anomaly detection coefficient; After replacing the outlier data, perform state filtering on the data and normalize it to the maximum and minimum values ​​according to the following formula: in, The state-filtered first One degree of freedom data, and These are the valid maximum and minimum values ​​within the reference window, respectively. It is a preset positive number.

3. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a Class III sea state unmanned surface vessel according to claim 1, characterized in that, In step S1, the inertial measurement data and satellite positioning data are synchronized in time and transformed in coordinates, and then fused through a state estimation model to determine the position, attitude, linear velocity, angular velocity and linear acceleration of the unmanned surface vessel. In step S2, the control input vector is determined based on the thruster operating data and the servo motor operating data. The control force and control torque estimation vectors are determined according to the following formula: in, This is the control allocation matrix determined based on the calibration data of the thrusters and servos.

4. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a Class III sea state unmanned surface vessel according to claim 1, characterized in that, In step S3, sway, heave, and pitch are divided into longitudinal coupling groups, and sway, roll, and bow roll are divided into transverse coupling groups. The time-varying decoupling matrix includes a longitudinal motion decoupling sub-matrix corresponding to the longitudinal coupling group and a transverse motion decoupling matrix corresponding to the transverse coupling group. Off-diagonal elements are allowed within the same decoupling sub-matrix, and matrix elements between different decoupling sub-matrixes are set to zero. The time-varying decoupling matrix is ​​updated based on the significant wave height, average wave period, unmanned surface vessel speed, and historical decoupling residuals.

5. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a Class III sea state unmanned surface vessel according to claim 4, characterized in that, In step S3, each row of the time-varying decoupling matrix is ​​normalized, and the condition number of the time-varying decoupling matrix is ​​limited to within a preset condition number threshold. The time-varying decoupling matrix is ​​determined according to the following formula: in, For a predefined set of sparse block matrices, This indicates the extraction of off-diagonal elements from the matrix. These are the matrix transformation constraint coefficients. This represents the Frobenius norm.

6. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a three-level sea state unmanned surface vessel according to claim 1, characterized in that, In step S4, based on the average wave period Determine the cutoff frequency: in, The cutoff frequency, The preset frequency adjustment coefficient; A low-pass processing method with the cutoff frequency as the frequency band boundary is used to extract the low-frequency disturbance mode vector from the disturbance mode vector, and the difference between the disturbance mode vector and the low-frequency disturbance mode vector is determined as the high-frequency disturbance mode vector. A low-frequency long-time-series input window is constructed based on low-frequency disturbance mode vectors at multiple consecutive time points, and a high-frequency short-time-series input window is constructed based on high-frequency disturbance mode vectors at multiple consecutive time points. The duration of the low-frequency long-time-series input window is greater than the duration of the high-frequency short-time-series input window.

7. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a Class III sea state unmanned surface vessel according to claim 1, characterized in that, In step S4, the low-frequency perturbation mode sequence, the high-frequency perturbation mode sequence, and the corresponding six-degree-of-freedom state sequence are time-aligned and combined into a multi-scale input sequence; The common coding layer uses a shared long short-term memory network to extract the shared temporal features of the multi-scale input sequence, and the temporal attention layer is used to determine the temporal weights based on the shared temporal features at different times. The low-frequency long-time-series decoding branch outputs a low-frequency six-degree-of-freedom motion prediction sequence for the next 2 seconds based on the shared timing features and timing weights, while the high-frequency short-time-series decoding branch outputs a high-frequency six-degree-of-freedom motion prediction sequence for the next 0.5 seconds based on the shared timing features and timing weights.

8. The method for six-degree-of-freedom multi-scale prediction and feedforward compensation of a Class III sea state unmanned surface vessel according to claim 1, characterized in that, In step S5, for each prediction moment in the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence, the position, attitude, velocity, and angular velocity of the UAV reference point corresponding to that prediction moment are substituted into the deck landing point motion transformation relationship to determine the predicted position and predicted velocity of the deck landing point. in, The speed of the unmanned surface vessel as a reference point. For the angular velocity of the unmanned surface vessel, This refers to any prediction moment in the low-frequency six-degree-of-freedom motion prediction sequence or the high-frequency six-degree-of-freedom motion prediction sequence; The sum of the low-frequency prediction weight and the high-frequency prediction weight is 1; the low-frequency prediction weight is increased during the long-distance approach phase, and the high-frequency prediction weight is increased during the close-range hovering phase and the contact landing phase. In step S6, the feedforward enable coefficient is determined based on the prediction confidence, the reliability of the original motion acquisition data, and the communication status between the unmanned surface vessel and the target UAV. When the prediction confidence is lower than a preset confidence threshold, the reliability of the original motion acquisition data is lower than a preset data threshold, or the communication delay exceeds a preset delay threshold, the feedforward enable coefficient is set to zero, and the amplitude and rate of change of the allowed feedforward compensation command are limited.

9. A six-degree-of-freedom multi-scale prediction feedforward compensation system for a Class III sea state unmanned surface vessel, characterized in that, It includes a motion data acquisition interface, a control input interface, a sea state data acquisition interface, a UAV data interface, a communication unit, a memory, and a processor. The memory stores a computer program that can be executed by the processor, and the processor is configured with: The data processing module is used to perform abnormal data identification, filtering, normalization and state fusion on the raw motion acquisition data obtained through the motion data acquisition interface to obtain the six degrees of freedom state of the unmanned surface vessel. The disturbance estimation module is used to determine the wave disturbance estimation vector based on the six degrees of freedom state of the unmanned surface vessel and the control input data obtained through the control input interface; The time-varying decoupling module is used to construct a sixth-order local coupling response matrix based on the sea state parameters obtained through the sea state data acquisition interface and the current six-degree-of-freedom state of the unmanned surface vessel, determine the time-varying decoupling matrix through optimization, and determine the disturbance mode vector based on the time-varying decoupling matrix. The multi-scale prediction module is used to decompose the perturbation mode vector into a low-frequency perturbation mode sequence and a high-frequency perturbation mode sequence, and obtain the low-frequency six-degree-of-freedom motion prediction sequence, the high-frequency six-degree-of-freedom motion prediction sequence, and the corresponding prediction confidence through the multi-scale time series prediction model; The landing point prediction module is used to convert the low-frequency six-degree-of-freedom motion prediction sequence and the high-frequency six-degree-of-freedom motion prediction sequence into a deck landing point motion prediction sequence based on the fixed spatial positional relationship between the unmanned surface vessel reference point and the deck landing point. The feedforward compensation module is used to generate feedforward compensation instructions based on the deck landing point motion prediction sequence and the target UAV flight status, reference landing trajectory and landing phase information obtained through the UAV data interface, and send the feedforward compensation instructions to the flight controller of the target UAV through the communication unit. The online correction module is used to update the adaptive parameters of the multi-scale time series prediction model based on the prediction residual between the actual six-degree-of-freedom motion data and the corresponding prediction sequence.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 8.

Images