Joint modeling float measurement compensation method and system based on digital twinning

Abstract

This application provides a joint modeling buoy measurement compensation method and system based on digital twins, relating to the field of marine environmental monitoring technology. The method includes constructing a digital twin, comprising a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer; supervising the training of a time-series compensation network based on multi-condition time-series simulation samples generated from the scene twin layer; acquiring original observation time-series, attitude time-series, and drift time-series sequences, performing time-series synchronization alignment and channel splicing to obtain a multi-channel joint input time-series sequence, which is then input into the pre-trained time-series compensation network, outputting a disturbance term; and completing real-time compensation for buoy sensor measurements. This application, by constructing a four-layer architecture digital twin of the buoy measurement system, fully recreates the formation mechanism of measurement errors, generates simulation samples covering all operating conditions with ground truth annotations, and ensures the model's generalization ability and long-term operational stability under all sea states and operating conditions.

Description

Technical Field

[0001] This application relates to the field of marine environmental monitoring technology, and more specifically, to a method and system for buoy measurement compensation based on digital twin-based joint modeling. Background Technology

[0002] Ocean buoys, shipborne platforms, and nearshore observation platforms typically require long-term acquisition of continuous time-series information such as wave undulations, current velocity, and direction. These measurements are directly related to tasks such as marine environmental monitoring, sea state forecasting, equipment safety assessment, and navigation assistance. However, sensors are usually mounted on carriers whose attitude and position constantly change with sea state, so the sensor coordinate system is not always consistent with the coordinate system of the target physical quantity.

[0003] When the carrier undergoes roll, pitch, or heading changes, the wave elevation or flow velocity components will experience projection distortion and rotational mixing. When the carrier drifts and translates, the platform velocity will couple into the flow velocity observation. In addition, long-term sensor operation will also superimpose random biases, noise, and occasional spikes. Therefore, the raw observation values ​​often contain both the true signal and the spurious signals introduced by platform disturbances, leading to a decrease in measurement accuracy.

[0004] Currently, compensation methods for measurement errors of marine buoy sensors are mainly divided into two categories: The first common approach is to directly apply low-pass filtering, moving averages, or empirical corrections to the original observations. These methods are simple to implement, but they assume the error is primarily high-frequency noise and cannot explicitly address attitude rotation projection errors and platform velocity coupling errors. For targets like waves and current velocities, which are inherently dynamic, simple filtering may also weaken the true changes.

[0005] The second type of method builds a regression model based solely on the observations themselves, or considers only attitude or a single factor in drift. While this type of method is more flexible than pure filtering, it often fails to accurately distinguish between real environmental changes and spurious changes caused by platform disturbances due to the lack of complete information on the source of the disturbance, resulting in unstable compensation performance under complex operating conditions.

[0006] Therefore, the key problem to be solved in this application is: how to establish a sensor measurement compensation method that can be uniformly modeled, sequentially compensated, and stably deployed under the conditions of attitude disturbance, position drift, platform velocity coupling, bias, noise and spike interference, so that the output results are closer to the real physical quantities. Summary of the Invention

[0007] To address the aforementioned problems, this application employs a joint modeling buoy measurement compensation method based on digital twins, comprising the following steps: A digital twin of the buoy measurement system is constructed, which includes a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer that are sequentially associated from top to bottom. Based on the multi-condition time-series simulation samples generated by the scene twin layer, the time-series compensation network of the compensation decision layer is trained under supervision to obtain the pre-trained time-series compensation network. The original observation time series of the target physical quantity sensor of the buoy platform, the attitude time series output by the attitude measurement unit, and the drift time series output by the navigation and positioning unit are obtained. The three sets of sequences are then processed by time synchronization alignment and channel splicing to obtain a multi-channel joint input time series. The multi-channel joint input time series is input into the pre-trained time series compensation network, and the perturbation term is output; the compensated target physical quantity time series is calculated to complete the real-time compensation of the buoy sensor measurement.

[0008] Optionally, the target physical quantity sensors, attitude measurement units and navigation and positioning units of the buoy platform are mapped at the physical entity layer to synchronously acquire and output real observation time series data; In the mechanism twin layer, a virtual mechanism model is constructed to represent the true value of the target physical quantity, the carrier attitude rotation, the platform drift motion, and the sensor bias and noise. The mapping relationship between the sensor observation value and the true value of the target physical quantity and the disturbance factor is established. A simulation scenario library covering different attitude levels, drift levels and abnormal disturbance conditions is constructed in the scenario twin layer. Multi-condition time-series simulation samples with real value annotations are generated based on the mechanism twin layer for the training and verification of the compensation model. The compensation decision layer is used to deploy the time-series compensation network and outputs disturbance estimates and final compensation results based on the input time-series data.

[0009] Optionally, in the mechanistic twin layer: The true value vector x(t) of the target physical quantity is represented as: x(t) = [z(t), u(t), v(t)]; In the formula, t represents the time variable, z(t) represents the true value of the vertical wave elevation at time t, and u(t) and v(t) represent the true values ​​of the flow velocities in the two orthogonal horizontal directions at time t, respectively. The original observation vector is represented as: y(t)=[y z (t),y u (t),y v (t)]; In the formula, y(t) represents the vector of original sensor observations at time t. z (t) represents the observed vertical wave elevation at time t, y u (t), y v(t) represents the observed velocity values ​​in two orthogonal horizontal directions at time t.

[0010] Alternatively, the vertical wave elevation observation model is as follows: y z (t)=z(t)·cos(roll(t))·cos(pitch(t))+b z (t)+n z (t)+s z (t); The horizontal velocity observation model is as follows: [u r (t),v r (t),0] T =R(yaw,pitch,roll)·[u(t),v(t),0] T ; y u (t)=u r (t)+α·v x (t)+b u (t)+n u (t)+s u (t); y v (t)=v r (t)+α·v y (t)+b v (t)+n v (t)+s v (t); In the formula, roll(t) and pitch(t) are the roll and pitch angles of the buoy platform at time t, respectively; b z (t), b u (t), b v (t) represents the slowly varying random bias corresponding to the z, u, and v channels, respectively, and n z (t), n u (t), n v (t) represents the random observation noise corresponding to the z, u, and v channels, respectively, and s z (t), s u (t), s v (t) represent the abnormal spike disturbances corresponding to the z, u, and v channels, respectively. r (t), v r (t) represent the velocity components in channels u and v after rotation, respectively; u(t) and v(t) represent the actual velocities in channels u and v in the geographic coordinate system, respectively; R(yaw,pitch,roll) represents the coordinate rotation matrix constructed based on the attitude angle; α is the platform velocity coupling coefficient; v x (t), vy (t) represents the drift velocity of the buoy platform in two orthogonal horizontal directions at time t.

[0011] Optionally, the attitude timing sequence att(t) is a 6-channel timing sequence, represented as: att(t)=[roll(t),pitch(t),yaw(t),roll rate (t),pitch rate (t),yaw rate (t)]; In the formula, roll(t), pitch(t), and yaw(t) are the roll angle, pitch angle, and yaw angle of the buoy platform at time t, respectively. rate (t), pitch rate (t), yaw rate (t) represents the roll rate, pitch rate, and heading rate of the buoy platform at time t, respectively.

[0012] Optionally, the drift time series sequence drift(t) is a 4-channel time series sequence, represented as: ; In the formula, , Let be the coordinates of the buoy platform's drift position in two orthogonal horizontal directions as measured by GPS at time t. denoted as t, representing the drift velocities of the buoy platform in two orthogonal horizontal directions at time t.

[0013] Optionally, the temporal compensation network adopts a multi-scale temporal convolutional network (TCN), which consists of four concatenated temporal convolutional blocks. Each temporal convolutional block contains two one-dimensional convolutional layers, a batch normalization layer, a nonlinear activation layer, a dropout layer, and a residual connection structure. The dilation coefficients of the four temporal convolutional blocks are 1, 2, 4, and 8, respectively.

[0014] Optionally, the acquisition of the multi-channel joint input time series is performed using a sliding time window method; The multi-channel joint input time series is a 13-channel time series, obtained by concatenating the 3-channel original observation time series y(t), the 6-channel attitude time series att(t), and the 4-channel drift time series drift(t) along the channel dimension. The input tensor dimension of the time series compensation network is (13,T), where T is the sliding time window length. The output tensor is the 3-channel perturbation term Δ(t), and the compensation result is output as follows: (t): (t)=y(t)-Δ(t).

[0015] Optionally, when supervising the training of the time-compensation network, the loss function used is a weighted sum of the mean squared error loss and the smoothing regularization term, expressed as: L=MSE( ,x)+λ·mean(∣ (t) (t 1)∣); In the formula, MSE( (x) represents the sequence of target physical quantities after compensation. The mean square error between x(t) and the true value sequence x(t); λ is the weighting coefficient of the smoothing regularization term; mean(| (t)- (t-1)|) represents the mean of the absolute values ​​of the differences between adjacent time steps of the target physical quantity sequence after compensation.

[0016] This application also provides a digital twin-based co-modeling buoy measurement compensation system, characterized in that it includes methods for performing any of the aforementioned digital twin-based co-modeling buoy measurement compensation methods, comprising: The digital twin construction module is used to build a digital twin of the buoy measurement system. The digital twin includes a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer. The network training module is used to supervise the training of the time-compensation network based on the multi-condition time-series simulation samples generated by the scene twin layer, so as to obtain the pre-trained time-compensation network. The data acquisition and preprocessing module is used to acquire the original observation time series of the target physical quantity sensor of the buoy platform, the attitude time series output by the attitude measurement unit, and the drift time series output by the navigation and positioning unit, and to perform time synchronization alignment and channel splicing on the three sets of sequences to obtain a multi-channel joint input time series. The real-time compensation module is used to input the multi-channel joint input time series into the pre-trained time series compensation network, output the disturbance term and calculate the compensated target physical quantity time series, so as to realize the real-time compensation of buoy sensor measurement.

[0017] The beneficial effects of the joint modeling buoy measurement compensation method and system based on digital twins provided in this application are as follows: (1) This application constructs a four-layer architecture digital twin of the buoy measurement system to fully restore the formation mechanism of measurement error and generate simulation samples with true value annotation covering all working conditions. This provides a unified virtual support platform for the training, verification, parameter calibration and mechanism analysis of the compensation model. It fundamentally solves the problems of insufficient training samples, incomplete coverage of working conditions and inability to verify extreme working conditions, and ensures the generalization ability and long-term operational stability of the model under all sea conditions and working conditions.

[0018] (2) Experimental results show that the complete joint modeling scheme of this application can reduce the average root mean square error (RMSE) of the three channels z / u / v of the test set from 0.1330 in the original observation to 0.0573, with an overall reduction of 56.9%; under the most difficult working condition of L4-D3 with strong attitude and high drift, the average RMSE is reduced from 0.2235 to 0.0782, with a reduction of 65.0%, and the compensation effect is far superior to the existing technical scheme.

[0019] (3) This application adopts the output method of first predicting the disturbance term and then back-calculating the compensation value, which takes into account both error elimination and effective information retention. It reduces the complexity of the model's learning objective, effectively eliminates the errors introduced by various disturbances, and can completely retain the real dynamic changes of the target physical quantity in the original observation, avoiding the defects of loss of effective information and over-smoothing in traditional filtering methods. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0021] Figure 1 The average RMSE curves of the complete joint modeling scheme provided in this application embodiment at different pose levels; Figure 2 The average RMSE curves of the complete joint modeling scheme provided in this application embodiment under different drift levels; Figure 3 This is a timing comparison of the z-channel under the most difficult operating condition (L4-D3) provided in the embodiments of this application; Figure 4 The average RMSE curves of the attitude removal input ablation scheme provided in this application embodiment at different attitude levels; Figure 5 The average RMSE curves of the attitude removal input ablation scheme provided in this application embodiment under different drift levels; Figure 6 This is a timing comparison of the z-channel of the attitude removal input ablation scheme provided in the embodiments of this application under the most difficult working condition (L4-D3); Figure 7 These are the average RMSE curves of the drift-reducing input ablation scheme provided in this application embodiment at different pose levels; Figure 8 These are the average RMSE curves of the drift-removal input ablation scheme provided in this application embodiment under different drift levels; Figure 9 This is a timing comparison of the z-channel of the drift-removal input ablation scheme provided in the embodiments of this application under the most difficult operating conditions (L4-D3). Detailed Implementation

[0022] To make the technical problems, technical solutions, and beneficial effects to be solved by 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 are not intended to limit the scope of this application.

[0023] Example 1 A joint modeling buoy measurement compensation method based on digital twins includes the following steps: A digital twin of the buoy measurement system is constructed, which includes a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer that are sequentially associated from top to bottom. Based on the multi-condition time-series simulation samples generated by the scene twin layer, the time-series compensation network of the compensation decision layer is trained under supervision to obtain the pre-trained time-series compensation network. The original observation time series of the target physical quantity sensor of the buoy platform, the attitude time series output by the attitude measurement unit, and the drift time series output by the navigation and positioning unit are obtained. The three sets of sequences are then processed by time synchronization alignment and channel splicing to obtain a multi-channel joint input time series. The multi-channel joint input time series is input into the pre-trained time series compensation network, and the perturbation term is output; the compensated target physical quantity time series is calculated to complete the real-time compensation of the buoy sensor measurement.

[0024] The solution provided in this application recovers real physical quantities from complex platform perturbations. The difficulty of this operation lies in: 1. The sources of error are physically coupled. The observation bias is simultaneously affected by attitude projection, coordinate rotation, platform translation velocity coupling, offset drift, and random noise, and is not a single noise term.

[0025] 2. Error overlaps with the frequency band of the real signal. Attitude changes and drift often have low- to mid-frequency components, which are not completely separate from the dynamic components of the real wave / flow velocity, so traditional filtering cannot be simply relied upon.

[0026] 3. The goal is dynamic compensation, not static calibration. This task requires outputting continuous timing data, which must reduce errors while preserving the true dynamic response. Excessive smoothing should not be used to sacrifice useful information.

[0027] 4. Disturbance intensity varies significantly with operating conditions. Attitude and drift levels will change significantly under different sea states and platform motion states, and the model needs to remain stable under both weak and strong disturbance conditions.

[0028] 5. The method needs to be feasible for engineering deployment. Practical applications typically require the method to perform continuous inference over a sliding time window; therefore, in addition to accuracy, the model's real-time performance, stability, and implementation complexity also need to be considered.

[0029] To address the specific characteristics of the task, this solution avoids both simple analytical compensation and a completely black-box end-to-end truth regression. Instead, it employs a method where digital twins provide mechanistic and scenario support, joint inputs provide perturbation information, and data-driven temporal networks complete the compensation. Relying solely on analytical formulas for compensation would require precise knowledge of installation parameters, attitude changes, platform motion states, bias evolution patterns, and anomalous noise modes—factors that are difficult to model accurately. Conversely, relying entirely on black-box models risks overlooking the explicit physical meaning of attitude and drift, resulting in models that, while fitting the data, are difficult to interpret and generalize stably.

[0030] Therefore, this application adopts a joint modeling strategy, which explicitly takes attitude and drift information directly related to the error formation mechanism as input, while using digital twins to provide controllable scenarios, mechanism mapping and parameter verification capabilities, and then learning perturbation terms through a time series network, thereby taking into account physical relevance, data adaptability and engineering feasibility.

[0031] The digital twin provided in this application offers a virtual entity capable of mapping the relationship between "real environmental quantities, carrier motion state, and sensor observation results" for ocean observation compensation tasks. In real sea conditions, it is often difficult to simultaneously obtain long-term continuous measured data, strictly synchronized real reference values, and samples covering extreme attitudes and drift conditions. Therefore, without the support of a digital twin, model training, error attribution, and extreme condition verification would be greatly limited.

[0032] The initial implementation of a digital twin mainly involves: establishing a virtual mechanical model around the actual wave / flow velocity quantities, platform attitude, drift state, sensor observation links, bias noise, and anomalous disturbances; and generating multi-condition time-series samples under this mechanical model for training and validating the compensation network. Digital twins are not only used for simulation but also for constructing interpretable mapping relationships between physical and virtual systems.

[0033] At the physical entity layer, target physical quantity sensors, attitude measurement units, and navigation and positioning units of the buoy platform are mapped to synchronously acquire and output real observation time series data. In the mechanism twin layer, a virtual mechanism model is constructed to represent the true value of the target physical quantity, the carrier attitude rotation, the platform drift motion, and the sensor bias and noise. The mapping relationship between the sensor observation value and the true value of the target physical quantity and the disturbance factor is established. A simulation scenario library covering different attitude levels, drift levels and abnormal disturbance conditions is constructed in the scenario twin layer. Multi-condition time-series simulation samples with real value annotations are generated based on the mechanism twin layer for the training and verification of the compensation model. The compensation decision layer is used to deploy the time-series compensation network and outputs disturbance estimates and final compensation results based on the input time-series data.

[0034] The correspondence and associated variables between the physical entity layer, the mechanism twin layer, the scene twin layer, and the compensation decision layer are shown in Table 1. Table 1. Correspondence and Related Variables of Each Functional Layer in the Digital Twin

[0035] This application establishes a joint model around the source of the disturbance: the observation sequence, attitude sequence, and drift sequence are fed together into a time-series network to first predict the disturbance term, and then the compensation result is recovered by subtracting the disturbance term from the observation value. This makes the model's learning objective clearer and its physical meaning more explicit.

[0036] In the mechanistic twin layer: The true value vector x(t) of the target physical quantity is represented as: x(t) = [z(t), u(t), v(t)]; In the formula, t represents the time variable, z(t) represents the true value of the vertical wave elevation at time t, and u(t) and v(t) represent the true values ​​of the flow velocities in the two orthogonal horizontal directions at time t, respectively; the true value vector is the target signal to be recovered.

[0037] The original observation vector is represented as: y(t)=[y z (t),y u (t),y v (t)]; In the formula, y(t) represents the vector of original sensor observations at time t. z (t) represents the observed vertical wave elevation at time t, y u (t), y v (t) represents the observed velocity values ​​in two orthogonal horizontal directions at time t.

[0038] The vertical wave elevation observation model is as follows: y z (t)=z(t)·cos(roll(t))·cos(pitch(t))+b z (t)+n z (t)+s z (t); The horizontal velocity observation model is as follows: [u r (t),vr (t),0] T =R(yaw,pitch,roll)·[u(t),v(t),0] T ; y u (t)=u r (t)+α·v x (t)+b u (t)+n u (t)+s u (t); y v (t)=v r (t)+α·v y (t)+b v (t)+n v (t)+s v (t); In the formula, roll(t) and pitch(t) are the roll and pitch angles of the buoy platform at time t, respectively; b z (t), b u (t), b v (t) represents the slowly varying random bias corresponding to the z, u, and v channels, respectively, and n z (t), n u (t), n v (t) represents the random observation noise corresponding to the z, u, and v channels, respectively, and s z (t), s u (t), s v (t) represent the abnormal spike disturbances corresponding to the z, u, and v channels, respectively. r (t), v r (t) represent the velocity components in channels u and v after rotation, respectively; u(t) and v(t) represent the actual velocities in channels u and v in the geographic coordinate system, respectively; R(yaw,pitch,roll) represents the coordinate rotation matrix constructed based on the attitude angle, characterizing the coordinate rotation relationship caused by the attitude angle; α is the platform velocity coupling coefficient; v x (t), v y (t) represents the drift velocity of the buoy platform in two orthogonal horizontal directions at time t. α·v x (t), α·v y (t) characterizes the coupling effect of platform drift velocity on the velocity channel.

[0039] The attitude timing sequence att(t) is a 6-channel timing sequence, represented as follows: att(t)=[roll(t),pitch(t),yaw(t),roll rate (t),pitch rate(t),yaw rate (t)]; In the formula, roll(t), pitch(t), and yaw(t) are the roll angle, pitch angle, and yaw angle of the buoy platform at time t, respectively. rate (t), pitch rate (t), yaw rate (t) represents the roll, pitch, and yaw angular velocities of the buoy platform at time t, respectively. The attitude angles reflect the current directional relationship, while the angular velocities reflect the rate of attitude change. By using the combined input of attitude angles and angular velocities, the ability to characterize the effects of transient disturbances and dynamic rotation is enhanced.

[0040] The drift time series sequence drift(t) is a 4-channel time series sequence, represented as: ; In the formula, , Let be the coordinates of the buoy platform's drift position in two orthogonal horizontal directions as measured by GPS at time t. These represent the drift velocities of the buoy platform in two orthogonal horizontal directions at time t. Position reflects the drift trajectory state, while velocity more directly corresponds to the platform coupling term. By inputting both drift position and drift velocity, a more complete reflection of the drift process and its impact on observations can be obtained.

[0041] The multi-channel joint input time series is a 13-channel time series, which is obtained by splicing the 3-channel original observation time series y(t), the 6-channel attitude time series att(t), and the 4-channel drift time series drift(t) along the channel dimension. The perturbation term Δ(t) is predicted by the time series compensation network.

[0042] The input tensor of the time-series compensation network has dimensions (13, T), where T is the length of the sliding time window. The output tensor is a 3-channel perturbation term Δ(t), and the compensation result is output as follows: (t): (t)=y(t)-Δ(t).

[0043] Compared to directly regressing to the true value, this method of first estimating the perturbation and then inversely compensating is more conducive to preserving the effective dynamic information in the original observations.

[0044] The temporal compensation network employs a multi-scale temporal convolutional network (TCN), which consists of four concatenated temporal convolutional blocks. Each concatenated block contains two one-dimensional convolutional layers, a batch normalization layer, a non-linear activation layer, a dropout layer, and a residual connection structure. The dilation coefficients of the four blocks are 1, 2, 4, and 8, respectively. This allows for a larger temporal receptive field with fewer layers, enabling the simultaneous perception of short-term perturbations and trend changes within longer time windows. This improves training stability and inference efficiency while maintaining long-term dependency modeling capabilities.

[0045] The acquisition of multi-channel joint input time series is performed using a sliding time window method; When performing supervised training on a time-compensation network, the loss function used is a weighted sum of the mean squared error loss and the smoothing regularization term, expressed as: L=MSE( ,x)+λ·mean(∣ (t) (t 1)∣); In the formula, MSE( (x) represents the sequence of target physical quantities after compensation. The mean square error between x(t) and the true value sequence x(t); λ is the weighting coefficient of the smoothing regularization term; mean(| (t)- (t-1)|) represents the mean of the absolute values ​​of the differences between adjacent time steps in the compensated target physical quantity sequence. During training, supervised mean square error is used to constrain the compensation result to be close to the true value. At the same time, a smoothing regularization term is added, which constrains the difference amplitude of the compensation output between adjacent time steps, making the result more continuous, avoiding sudden jumps that do not conform to physical laws, and improving the physical rationality and usability of the sequence output.

[0046] Verification was performed using multi-condition time-series samples generated by digital twins. The sampling frequency was 10Hz, with each segment lasting 10 minutes, totaling 192 segments, divided into 128 / 32 / 32 segments for training / verification / testing. Attitude levels were set to L1-L4, corresponding to attitude RMS values ​​of 1°, 3°, 6°, and 10°, respectively; drift levels were set to D0-D3, corresponding to drift velocities of 0, 0.1, 0.3, and 0.5 m / s, respectively.

[0047] The simulation further considered GPS position noise, observation noise, random walk bias, and spike disturbances, and introduced a platform velocity coupling term in the current velocity channel to simulate the sources of measurement deviation of actual buoy sensors in complex sea conditions as closely as possible. Specifically, the platform velocity coupling coefficient was set to α=0.8, the standard deviation of vertical observation noise was 0.02, the standard deviation of horizontal current velocity observation noise was 0.03, and the standard deviation of GPS noise was 2.0m.

[0048] The network uses a 4-layer TCN with 64 / 64 / 64 / 64 channels, a dilation factor of 1 / 2 / 4 / 8, a kernel size of 3, dropout of 0.1, a window size of 256, an Adam optimizer, and a learning rate of 1e. -3 The loss function is the supervised mean squared error plus 0.05 times the smoothing constraint.

[0049] Detailed results are shown in Tables 2-4: Table 2. Comparison of RMSE accuracy of different compensation schemes under test set and extreme conditions.

[0050] Table 3. Comparison of RMSE accuracy in ablation experiments based on attitude and drift inputs.

[0051] Table 4 Comparison of RMSE Accuracy Before and After Compensation for the Z / U / V Three Measurement Channels

[0052] As can be seen from Table 2, using the average RMSE of the three channels z / u / v of the test set as a comprehensive index, the complete joint modeling scheme can reduce the average RMSE from 0.1330 to 0.0573, an overall reduction of 56.9%; under the most difficult L4-D3 condition, the average RMSE is reduced from 0.2235 to 0.0782, a reduction of 65.0%.

[0053] Compared to traditional low-pass filtering, this scheme not only has lower errors but also does not significantly weaken the true dynamic changes like traditional low-pass filtering. Under the most challenging L4-D3 condition, the average RMSE of the test set using traditional low-pass filtering actually increased to 0.4551, indicating that for time-series observations affected by attitude and platform motion coupling, frequency domain smoothing alone is insufficient to effectively recover the true quantities.

[0054] The ablation results further illustrate the necessity of joint modeling. As shown in Table 3, after removing the attitude input, the average RMSE on the test set increased to 0.0985; after removing the drift input, it increased to 0.1071. Compared with the complete model, the errors of these two ablation versions on the test set increased by 71.9% and 86.9%, respectively, and by 114.9% and 99.6%, respectively, under the most difficult condition (L4-D3 condition), indicating that both attitude and drift information are necessary inputs for achieving high-precision compensation.

[0055] From the perspective of channel dimensions, this scheme provides the most significant improvement to the horizontal velocity component u / v, which is more significantly affected by attitude rotation and platform drift: the RMSE on the test set decreased by 62.0% and 65.5%, respectively. In the current simulation settings, the original error of the z-channel is relatively small, so the complete model mainly reflects a significant improvement in the u / v channels while maintaining the same order of magnitude of error.

[0056] Figure 1 , Figure 2 , Figure 4 , Figure 5 , Figure 7 , Figure 8 In the legend, "original" refers to the original observation data, "filtered" refers to the traditional low-pass filter, and "this application" refers to the buoy measurement compensation method based on digital twin provided in this application. Figure 3 . Figure 6 , Figure 9 In the legend, "True value" represents the true value of the vertical wave elevation, "Original observation" represents the vertical elevation value of the original observation, "Low-pass filter" represents the vertical elevation value after traditional low-pass filtering, and "This application" represents the vertical elevation estimate after compensation by the digital twin-based joint modeling buoy measurement compensation method provided in this application.

[0057] Figures 1-3 The complete model test results diagram demonstrates the overall performance of the complete digital twin + attitude / drift joint modeling scheme, illustrating the complete model's compensation capabilities under attitude changes, drift changes, and the most challenging conditions. Figure 1 This is used to demonstrate the stability compensation effect of the complete model at different attitude levels; the overall RMSE reduction relative to the original observations for each attitude level is as follows: L1 - 55.79%; L2 - 62.76%; L3 - 52.05%; L4 - 56.81%.

[0058] Figure 2 This is used to demonstrate the stability compensation effect of the complete model under different drift levels; under the D3 condition, the overall RMSE reduction is 61.81%, and the reduction for each drift level is as follows: D0 - 55.92%; D1 - 48.76%; D2 - 55.11%; D3 - 61.81%.

[0059] Figure 3 This figure shows the time series comparison of the z-channel under the most challenging operating conditions (L4-D3). It demonstrates the time series tracking capability of the complete model under these conditions, showing that the compensated output is closer to the true curve than the original observations and traditional filtering.

[0060] Figures 4-6The image shows the ablation test results after removing the attitude input. The results clearly demonstrate that without attitude angle and angular velocity information, the model's ability to compensate for rotational projection errors and dynamic attitude disturbances significantly decreases. Figure 4 Used for attitude information ablation analysis; the average RMSE of the test set after attitude removal was 0.0985, which is 71.9% higher than that of the complete model.

[0061] Figure 5 The figure shows the average RMSE curves of the attitude-input-free ablation scheme at different drift levels. This figure illustrates that without attitude input, the model struggles to adequately compensate for errors introduced by attitude rotation and projection; in most cases, the overall curve is higher than that of the complete model.

[0062] Figure 6 The figure shows a time-series comparison of the z-channel performance of the attitude-input-removed ablation scheme under the most challenging load case (L4-D3). This illustrates that under the most challenging load case, removing the attitude input significantly weakens the model's ability to track the true curve compared to the fully functional model.

[0063] Figures 7-9 The image shows the ablation test results after removing the drift input. The results clearly demonstrate that without drift position and velocity information, the model's ability to compensate for platform velocity coupling errors and drift-related disturbances significantly decreases. Figure 7 The figure shows the average RMSE curves of the drift-removal input ablation scheme at different pose levels. This figure is used for drift information ablation analysis; the average RMSE of the test set after drift removal is 0.1071, which is an increase of 86.9% compared to the complete model.

[0064] Figure 8 The figure shows the average RMSE curves of the drift-free input ablation scheme at different drift levels. This illustrates that without drift input, the model's ability to compensate for platform velocity coupling and drift-related disturbances decreases significantly, especially at high drift levels.

[0065] Figure 9 The figure shows a time-series comparison of the z-channel performance of the drift-removed input ablation scheme under the most challenging condition (L4-D3). This illustrates that under the most challenging condition, removing the drift input significantly weakens the model's recovery of key dynamic processes compared to the full model.

[0066] This application also provides a method and system for buoy measurement compensation based on digital twin co-modeling, characterized in that, for performing any of the aforementioned methods for buoy measurement compensation based on digital twin co-modeling, it includes: The digital twin construction module is used to build a digital twin of the buoy measurement system. The digital twin includes a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer. The network training module is used to supervise the training of the time-compensation network based on the multi-condition time-series simulation samples generated by the scene twin layer, so as to obtain the pre-trained time-compensation network. The data acquisition and preprocessing module is used to acquire the original observation time series of the target physical quantity sensor of the buoy platform, the attitude time series output by the attitude measurement unit, and the drift time series output by the navigation and positioning unit, and to perform time synchronization alignment and channel splicing on the three sets of sequences to obtain a multi-channel joint input time series. The real-time compensation module is used to input the multi-channel joint input time series into the pre-trained time series compensation network, output the disturbance term and calculate the compensated target physical quantity time series, so as to realize the real-time compensation of buoy sensor measurement.

[0067] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A joint modeling buoy measurement compensation method based on digital twinning, characterized in that, Includes the following steps: A digital twin of a buoy measurement system is constructed, comprising a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer, which are sequentially associated from top to bottom. Based on the multi-condition time-series simulation samples generated by the scenario twin layer, the time-series compensation network of the compensation decision layer is trained under supervision to obtain the pre-trained time-series compensation network. The original observation time series of the target physical quantity sensor of the buoy platform, the attitude time series output by the attitude measurement unit, and the drift time series output by the navigation and positioning unit are obtained. The three sets of sequences are then processed by time synchronization alignment and channel splicing to obtain a multi-channel joint input time series. The multi-channel joint input time series is input into a pre-trained time compensation network, and a perturbation term is output. The time sequence of the target physical quantity after compensation is calculated, and real-time compensation of the buoy sensor measurement is completed.

2. The method for joint modeling buoy measurement compensation based on digital twins according to claim 1, characterized in that: The target physical quantity sensors, attitude measurement units and navigation and positioning units of the buoy platform are mapped in the physical entity layer to synchronously acquire and output real observation time series data. A virtual mechanism model is constructed in the mechanism twin layer to represent the true value of the target physical quantity, the carrier attitude rotation, the platform drift motion, the sensor bias and noise, and to establish the mapping relationship between the sensor observation value and the true value of the target physical quantity and the disturbance factor. A simulation scenario library covering different attitude levels, drift levels and abnormal disturbance conditions is constructed in the scenario twin layer. Based on the mechanism twin layer, multi-condition time-series simulation samples with real value annotations are generated for the training and verification of the compensation model. The compensation decision layer is used to deploy a time-series compensation network and outputs disturbance estimates and final compensation results based on the input time-series data.

3. The joint modeling buoy measurement compensation method based on digital twinning according to claim 2, characterized in that: In the aforementioned mechanistic twin layer: The true value vector x(t) of the target physical quantity is represented as: x(t) = [z(t), u(t), v(t)]; In the formula, t represents the time variable, z(t) represents the true value of the vertical wave elevation at time t, and u(t) and v(t) represent the true values ​​of the flow velocities in the two orthogonal horizontal directions at time t, respectively. The original observation vector is represented as: y(t)=[y z (t),y u (t),y v (t)]; In the formula, y(t) represents the vector of original sensor observations at time t. z (t) represents the observed vertical wave elevation at time t, y u (t), y v (t) represents the observed velocity values ​​in two orthogonal horizontal directions at time t.

4. The buoy measurement compensation method based on digital twins according to claim 3, characterized in that: The vertical wave elevation observation model is as follows: y z (t)=z(t)·cos(roll(t))·cos(pitch(t))+b z (t)+n z (t)+s z (t); The horizontal velocity observation model is as follows: [u r (t),v r (t),0] T =R(yaw,pitch,roll)·[u(t),v(t),0] T ; y u (t)=u r (t)+α·v x (t)+b u (t)+n u (t)+s u (t) y v (t)=v r (t)+α·v y (t)+b v (t)+n v (t)+s v (t) In the formula, roll(t) and pitch(t) are the roll and pitch angles of the buoy platform at time t, respectively; b z (t), b u (t), b v (t) represents the slowly varying random bias corresponding to the z, u, and v channels, respectively, and n z (t), n u (t), n v (t) represents the random observation noise corresponding to the z, u, and v channels, respectively, and s z (t), s u (t), s v (t) represent the abnormal spike disturbances corresponding to the z, u, and v channels, respectively. r (t), v r (t) represent the velocity components in channels u and v after rotation, respectively; u(t) and v(t) represent the actual velocities in channels u and v in the geographic coordinate system, respectively; R(yaw,pitch,roll) represents the coordinate rotation matrix constructed based on the attitude angle; α is the platform velocity coupling coefficient; v x (t), v y (t) represents the drift velocity of the buoy platform in two orthogonal horizontal directions at time t.

5. The joint modeling buoy measurement compensation method based on digital twinning according to claim 1, wherein: The attitude timing sequence att(t) is a 6-channel timing sequence, represented as follows: att(t)=[roll(t),pitch(t),yaw(t),roll rate (t),pitch rate (t),yaw rate (t)]; In the formula, roll(t), pitch(t), and yaw(t) are the roll angle, pitch angle, and yaw angle of the buoy platform at time t, respectively. rate (t), pitch rate (t), yaw rate (t) represents the roll rate, pitch rate, and heading rate of the buoy platform at time t, respectively.

6. The joint modeling buoy measurement compensation method based on digital twinning of claim 1, wherein: The drift time series sequence drift(t) is a 4-channel time series sequence, represented as: ； wherein, , are the coordinates of the drift position of the buoy platform in two orthogonal horizontal directions measured by GPS at time t, are the drift velocities of the buoy platform in two orthogonal horizontal directions at time t.

7. The joint modeling buoy measurement compensation method based on digital twinning according to claim 5, wherein: The temporal compensation network adopts a multi-scale temporal convolutional network (TCN), which consists of four concatenated temporal convolutional blocks. Each temporal convolutional block contains two one-dimensional convolutional layers, a batch normalization layer, a nonlinear activation layer, a dropout layer, and a residual connection structure. The dilation coefficients of the four temporal convolutional blocks are 1, 2, 4, and 8, respectively.

8. The buoy measurement compensation method based on digital twins according to claim 1, characterized in that: The acquisition of the multi-channel joint input time sequence is performed using a sliding time window method; The multi-channel joint input time series is a 13-channel time series, obtained by concatenating the 3-channel original observation time series y(t), the 6-channel attitude time series att(t), and the 4-channel drift time series drift(t) along the channel dimension. The input tensor dimension of the time series compensation network is (13,T), where T is the sliding time window length. The output tensor is the 3-channel perturbation term Δ(t), and the compensation result is output as follows: (t): (t) = y(t) - Δ(t).

9. The method for joint modeling buoy measurement compensation based on digital twins according to claim 2, characterized in that: When performing supervised training on a time-compensation network, the loss function used is a weighted sum of the mean squared error loss and the smoothing regularization term, expressed as: L=MSE( ,x)+λ·mean(∣ (t) (t 1)∣); In the formula, MSE( (x) represents the sequence of target physical quantities after compensation. The mean square error between x(t) and the true value sequence x(t); λ is the weighting coefficient of the smoothing regularization term; mean(| (t)- (t-1)|) represents the mean of the absolute values ​​of the differences between adjacent time steps of the target physical quantity sequence after compensation.

10. A joint modeling buoy measurement compensation system based on digital twins, characterized in that, To perform the digital twin-based joint modeling buoy measurement compensation method as described in any one of claims 1-9, comprising: A digital twin construction module is used to construct a digital twin of a buoy measurement system. The digital twin includes a physical entity layer, a mechanism twin layer, a scene twin layer, and a compensation decision layer. The network training module is used to supervise the training of the timing compensation network based on the multi-condition timing simulation samples generated by the scenario twin layer, so as to obtain the pre-trained timing compensation network. The data acquisition and preprocessing module is used to acquire the original observation time series of the target physical quantity sensor of the buoy platform, the attitude time series output by the attitude measurement unit, and the drift time series output by the navigation and positioning unit, and to perform time synchronization alignment and channel splicing on the three sets of sequences to obtain a multi-channel joint input time series. The real-time compensation module is used to input the multi-channel joint input time series into the pre-trained time series compensation network, output the disturbance term and calculate the compensated target physical quantity time series, so as to realize the real-time compensation of the buoy sensor measurement.

Images