A GNSS / INS integrated navigation method for unmanned surface vessels that considers control information

Abstract

This invention discloses a GNSS / INS integrated navigation method for unmanned surface vessels (USVs) that considers control information, belonging to the field of USV navigation. The method includes the following steps: constructing a CNN-LSTM-AM model, taking IMU data and control information as input, and outputting forward and lateral velocities; when GNSS is normal, the IMU obtains position, velocity, and attitude through mechanical orchestration, subtracts these from the GNSS position, inputs them to a filter for fusion, outputs the error, corrects the mechanical orchestration results, and calculates the velocity; simultaneously, the model is trained using IMU data, control information, and velocity; when lock is lost, real-time IMU data and control information are input to the model to predict forward and lateral velocities, then subtracts these from the velocity calculated by mechanical orchestration, and inputs them to a filter for state updates, obtaining error data to correct the mechanical orchestration results, achieving continuous state estimation. This invention can suppress INS error divergence and improve positioning accuracy in GNSS denied environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned surface vessel (USV) navigation, and more specifically to a GNSS / INS integrated navigation method for USVs that takes control information into account. Background Technology

[0002] With the rapid development of the marine economy, unmanned surface vessels (USVs) have shown great promise in marine environmental monitoring, near-shore patrol and search and rescue, marine resource exploration, and maritime operations due to their core advantages such as low cost, remote autonomous control, and no risk of personnel casualties. The high-precision positioning and environmental awareness capabilities of USVs are the core foundation for their stable and reliable completion of various complex maritime tasks. Global Navigation Satellite Systems (GNSS) offer outstanding advantages in all-weather, all-day, and continuous high-precision positioning; however, their signals are highly susceptible to interference from complex external environments, such as bridge obstructions, dense port structures, large ships, and vegetation obstructions, all of which can cause GNSS positioning to deviate, experience a sharp drop in accuracy, or even fail completely. Inertial Navigation Systems (INS) offer fully autonomous navigation, fast data refresh rates, and excellent short-term navigation accuracy; however, due to inherent errors in their inertial devices and limitations imposed by the integration calculation principle, positioning errors accumulate and rapidly diverge over time. Currently, GNSS and INS, with their complementary performance characteristics, have become the mainstream combined navigation solution in the USV field. However, when unmanned surface vessels navigate in complex and restricted environments such as bridges, waterways, ports, drilling platforms, sea caves, and densely built-up areas near the coast, they are prone to losing GNSS signal lock and may even be in a GNSS denied environment for a long time. In this case, relying solely on INS for independent navigation will cause serious positioning drift due to the continuous accumulation of errors in inertial devices, which will greatly reduce the overall accuracy of the navigation system and the reliability of maritime operations.

[0003] To address the issue of rapid divergence in INS positioning errors in areas where GNSS signal lock-off or denial occurs, some current solutions involve introducing external auxiliary sensors to expand the integrated navigation architecture. These sensors, such as magnetometers, barometers, lidar, and visual cameras, fuse multi-source sensing information to improve the overall robustness and positioning accuracy of the GNSS / INS integrated navigation system. While adding external auxiliary sensors can effectively improve the positioning accuracy of the integrated navigation system in GNSS lock-off or denial situations, this approach directly increases the hardware cost and overall architectural complexity of the navigation system. Furthermore, sensors such as lidar and visual cameras need to be installed on the exterior of the unmanned surface vessel (USV), which is exposed to the marine environment for extended periods, making them susceptible to seawater splash and salt spray corrosion. This not only reduces positioning accuracy but also shortens their service life. Additionally, some sensors (such as odometers) are difficult to install due to the physical characteristics of the USV.

[0004] Other solutions are based on constructing motion constraint models according to the inherent motion laws of the vehicle. The core of this approach is to convert the vehicle's prior motion information into equivalent measurements and incorporate them into the filter measurement update process, thereby suppressing the accumulation of INS errors. Commonly used constraint algorithms include elevation constraints, velocity constraints, and attitude constraints. Constructing motion constraint models does not require additional sensors and can improve the positioning accuracy and operational reliability of the integrated navigation system to a certain extent. However, during unmanned surface vessels (USVs), they are subject to severe disturbances from the marine environment, such as wind, waves, and currents, making it difficult to meet the applicable prerequisites of traditional motion constraint algorithms, such as non-holonomic constraints, resulting in a significant limitation on the constraint effect. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a GNSS / INS integrated navigation method for unmanned surface vessels that considers control information. This method combines IMU data (specific force, angular velocity), control information (thrust, rudder force), and neural networks to mitigate INS error divergence in GNSS lock-off or denial environments, thereby improving the positioning accuracy of the integrated navigation.

[0006] The technical solution adopted in this invention is: A method for GNSS / INS integrated navigation of unmanned surface vessels that takes control information into account includes the following steps: Step a: Construct a CNN-LSTM-AM model, using IMU data and control information as input data to the CNN-LSTM-AM model, and outputting forward velocity and lateral velocity; Step b: An error state Kalman filter is used as the fusion algorithm for GNSS and INS. When the GNSS signal is normal, the inertial measurement unit outputs IMU data, which is mechanically arranged and calculated to obtain position, velocity, and attitude information. The GNSS receiver calculates the position information, and the difference between the position calculated by the mechanical arrangement and the position information is input into the filter for fusion calculation. The output position error, velocity error, and attitude error data are used to correct the position, velocity, and attitude information obtained by the mechanical arrangement, and the corrected position, velocity, and attitude results are obtained. Based on this, the forward velocity and lateral velocity of the unmanned surface vessel are calculated. Meanwhile, the constructed CNN-LSTM-AM model is trained using the IMU data, control information, and calculated forward and lateral velocities of the unmanned surface vessel during this phase. Step c: When the unmanned surface vessel enters a GNSS lockout or signal denial environment, the real-time IMU data and control information are input into the trained CNN-LSTM-AM model to predict the forward and lateral velocities of the unmanned surface vessel at that moment. Simultaneously, the IMU data output by the inertial measurement unit is mechanically programmed and calculated to obtain position, velocity, and attitude information; The predicted forward and lateral velocities of the unmanned surface vessel are subtracted from the velocities calculated by the mechanical orchestration, and the difference is input into a filter for state update to obtain position error, velocity error, and attitude error data. These data are then used to correct the position, velocity, and attitude information obtained by the mechanical orchestration, resulting in corrected position, velocity, and attitude results, which serve as the final integrated navigation output, enabling continuous state estimation in GNSS lock-out or denial environments.

[0007] The beneficial technical effects of the present invention are as follows: (1) This invention introduces a new data source by considering control information (thrust, rudder force), and does not require additional sensors. It can achieve high-precision navigation in GNSS denied environments based on the control information of the unmanned vessel itself, thus alleviating the problem of reduced positioning accuracy of combined navigation caused by INS error divergence in GNSS denied environments.

[0008] (2) The CNN-LSTM-AM model constructed in this invention adopts a dual-branch structure, namely the first CNN-LSTM branch and the second CNN-LSTM branch, which respectively process IMU data and thrust information. Combined with subsequent gating attention mechanisms, it can achieve high-precision prediction of forward and lateral speed of unmanned surface vessels.

[0009] (3) This invention constructs a CNN-LSTM-AM model and combines IMU data (specific force, angular velocity) and control information (thrust, rudder force) with the CNN-LSTM-AM model. The IMU data and control information are used as model inputs, and the forward and lateral velocities of the unmanned surface vessel are used as outputs. This allows the forward and lateral velocities of the unmanned surface vessel to be predicted in GNSS lockout or denial environments, so as to correct the position, velocity and attitude information obtained by mechanical arrangement calculations and realize continuous state estimation in GNSS lockout or denial environments. This invention can effectively alleviate INS error divergence in GNSS lockout or denial environments and improve the positioning accuracy of integrated navigation during GNSS denial. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an unmanned surface vessel GNSS / INS integrated navigation method that takes control information into account, according to the present invention. Figure 2 This is a schematic diagram of the integrated navigation principle of the method of the present invention when GNSS signal is good; Figure 3 This is a schematic diagram of the integrated navigation principle of the method of the present invention when GNSS signals are denied; Figure 4 This is a schematic diagram of the architecture of the CNN-LSTM-AM model constructed by the method of the present invention; Figure 5 This is a schematic diagram of the motion control system for an unmanned surface vessel. Figure 6 This is a graph showing the change in forward velocity as a function of the resultant thrust force. Figure 7 This is a graph showing the variation of lateral velocity with the resultant force of the thruster. Figure 8 A local coordinate system trajectory diagram of the reservoir scene; Figure 9 A local coordinate system trajectory diagram of the bay scene; Figure 10 Positioning error map for the reservoir scene (north / east); Figure 11 This is a map showing the positioning error in the north / east direction for a bay scene. Detailed Implementation

[0011] Existing methods mitigate and resolve the problem of rapid divergence of INS positioning errors in GNSS-denied environments by adding external sensors, but this increases additional costs and algorithm complexity, and may also damage the sensors. In addition, due to the influence of wind, waves and currents at sea, traditional speed, attitude and position constraint methods are not suitable for unmanned surface vessels.

[0012] Based on this, this invention creatively incorporates control information as an additional input into the integrated navigation system by exploring the relationship between the control information and motion state of the unmanned surface vessel (USV). This input, combined with the constructed CNN-LSTM-AM model, improves positioning accuracy in GNSS-denied environments. Specifically, this invention combines IMU data (specific force, angular velocity), control information (thrust, rudder force), and the CNN-LSTM-AM model to effectively mitigate INS error divergence in GNSS lock-out or denied environments, thereby improving the positioning accuracy of integrated navigation during GNSS denial.

[0013] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0014] like Figure 1 As shown, an unmanned surface vessel (USV) GNSS / INS integrated navigation method considering control information includes the following steps: Step a: Construct a CNN-LSTM-AM model, using IMU data and control information as inputs to the CNN-LSTM-AM model, and outputting forward velocity and lateral velocity.

[0015] Step b, as follows Figure 2As shown, an error-state Kalman filter is used as the fusion algorithm for GNSS and INS. When the GNSS signal is normal, the inertial measurement unit outputs IMU data, which is mechanically choreographed to obtain position, velocity, and attitude information. The GNSS receiver calculates the position information, and the difference is calculated with the position obtained through mechanical choreographing. The difference is then input into the filter for fusion calculation, outputting position error, velocity error, and attitude error data, which are fed back to the mechanically choreographed position, velocity, and attitude information to obtain high-precision position, velocity, and attitude results. Based on the velocity and heading information calculated from these results, the forward and lateral velocities of the unmanned surface vessel (USV) are calculated. Simultaneously, the constructed CNN-LSTM-AM model is trained using the IMU data, control information, and forward and lateral velocities of the USV at this stage.

[0016] Step c, as Figure 3 As shown, when an unmanned surface vessel (USV) enters a GNSS lock-out or signal denial environment, real-time IMU data and control information are input into a trained CNN-LSTM-AM model to predict the USV's forward and lateral velocities. Simultaneously, the IMU data output from the inertial measurement unit (IMU) is mechanically programmed to obtain position, velocity, and attitude information. The predicted forward and lateral velocities are subtracted from the mechanically programmed velocities, and the difference is input into a filter for state updates, yielding position, velocity, and attitude errors. These errors are then used to correct the mechanically programmed position, velocity, and attitude information, resulting in corrected position, velocity, and attitude values, thus achieving continuous state estimation in GNSS lock-out or signal denial environments.

[0017] In step a: the CNN-LSTM-AM model includes a first CNN-LSTM branch, a second CNN-LSTM branch, and a gating attention mechanism module, i.e., the AM module. For example... Figure 4As shown, the first CNN-LSTM branch includes a first convolutional neural network and a first long short-term memory network. The specific force and angular velocity signals output by the inertial measurement unit (IMU) enter the first convolutional neural network, are processed to obtain the first convolutional output features, are then mapped to the non-negative interval by the first ReLU nonlinear activation function, and subsequently input into the first long short-term memory network. These features are then globally integrated along the time dimension and processed by the first fully connected layer to obtain the output features of the first CNN-LSTM branch. The second CNN-LSTM branch includes a second convolutional neural network and a second long short-term memory network. The thrust and rudder force signals output by the control system enter the second convolutional neural network, are processed to obtain the second convolutional output features, are then mapped to the non-negative interval by the second ReLU nonlinear activation function, and subsequently input into the second long short-term memory network. These features are then globally integrated along the time dimension and processed by the second fully connected layer to obtain the output features of the second CNN-LSTM branch. The output features of the first and second CNN-LSTM branches are jointly input into the gating attention mechanism module for adaptive weighted fusion to obtain the attention fusion features. The attention fusion features are processed by the third ReLU nonlinear activation function, and then passed through two independent fully connected output heads, namely the third fully connected layer and the fourth fully connected layer, to obtain the forward velocity and the lateral velocity, respectively.

[0018] Among them, attention fusion features are the output features of the first CNN-LSTM branch (i.e., the IMU branch). The output features of the second CNN-LSTM branch (i.e., the control information branch) Projected onto the same dimension respectively: ; ; in, Output features for the first CNN-LSTM branch (i.e., the IMU branch) The weight matrix, Its bias vector; both are parameters automatically learned during the training process. This is the feature vector projected onto the first CNN-LSTM branch (i.e., the IMU branch). , , Similarly, the output features of the second CNN-LSTM branch (i.e., the control information branch) are also calculated. The weight matrix, bias vector, and projected feature vector are also parameters that are automatically learned during the training process.

[0019] Then the two projection features , The weights are concatenated, and a fully connected layer (in the gated attention layer) with a sigmoid function is used to generate the dimension-wise gated weights of the first CNN-LSTM branch (i.e., the IMU branch). .

[0020] ; in, This represents the sigmoid function. for and In concatenating vectors The weight matrix, for and In concatenating vectors The bias vector, It is a length of (here The vector is equal to the fusion layer dimension 32 of the gated attention mechanism, with each element between 0 and 1, so each fused feature dimension has an adaptive weight. The second CNN-LSTM branch (i.e., the control information branch) uses complementary weights. .

[0021] Ultimately, attention fusion features for: ; in, It represents element-wise multiplication; its adaptability is reflected in the weights. Instead of fixed hyperparameters, they are dynamically calculated by inputting the features output from the IMU branch of the current sample and the features output from the control information branch into the gating attention module, and are automatically learned during training through backpropagation of regression loss.

[0022] Both the first and second convolutional neural networks mentioned above are one-dimensional convolutional neural networks. The IMU data mentioned above includes specific force and angular velocity, and the control information includes the thrust and rudder force output by the control system.

[0023] The following is a more detailed explanation of the underlying principles.

[0024] There is a clear relationship between the control information and the motion state of an unmanned surface vessel (USV): when the USV control system calculates the actions the USV should take based on mission requirements—that is, calculates the specific thrust and rudder force information—it sends this information to the actuator. The actuator maps this information to the two thrusters according to certain principles. Then, the drive module, according to the USV's differential control logic, converts this into independent control signals for the left and right thrusters. After pulse width modulation (PWM) signal conversion, the thrusters are driven to operate in coordination, ultimately enabling the USV to move forward, backward, and turn, thus completing the predetermined mission. The schematic diagram of the USV motion control system is shown below. Figure 5 As shown.

[0025] Depend on Figure 5 It is evident that there is a clear relationship between the control information of the unmanned surface vessel (USV) and its motion state. To visualize this relationship, trend graphs of thrust information and the forward and lateral velocities of the USV are presented here, as shown in the figure. Figure 6 , Figure 7 As shown in the figure, it is clear that there is a significant correlation between the thrust information output by the unmanned surface vessel (USV) control system and the forward and lateral velocities of the USV. However, this relationship is non-linear, and it is difficult to accurately express the changes in motion state using conventional fitting functions. Therefore, this method introduces a Convolutional Neural Network-Long Short-Term Memory Neural Network-Gated Attention Mechanism (CNN-LSTM-AM) model to express the non-linear relationship between the thrust and the forward and lateral velocities. Since a six-axis IMU can accurately sense the motion changes of the vehicle, this method also uses it as one of the information sources for predicting the forward and lateral velocities of the USV.

[0026] Because the raw IMU observation data and the control information such as thrust and rudder force output by the control system differ significantly in physical meaning, data distribution, noise characteristics, and temporal variation patterns, direct mixing and splicing can easily lead to feature interference and information aliasing. Therefore, it is necessary to use independent CNN-LSTM branches for feature modeling of IMU data and control information respectively. The dual-branch structure allows for the targeted extraction of local temporal features and long-term dependency features specific to each modality, avoiding mutual contamination between the two types of signals. Simultaneously, it maintains alignment of the feature dimensions and representation levels of the two branches, facilitating subsequent adaptive weighted fusion of multimodal features using attention mechanisms. This enables the model to dynamically weigh the contribution ratio of inertial observation information and control input information according to different navigation conditions of the unmanned surface vessel, improving velocity estimation accuracy and model robustness.

[0027] The force and angular velocity signals output by the IMU exhibit strong temporal correlation within a short time window. Affected by sensor errors and environmental disturbances such as wind, waves, and currents at sea, the raw observation data contains a large amount of high-frequency noise. The raw IMU data is high-dimensional temporal information, containing six dimensions of observations: three-axis acceleration and three-axis angular velocity. If the sequence length is K, the data dimension can reach K×6. Directly inputting this high-dimensional raw sequence into subsequent models can easily lead to a surge in the number of parameters and computational load, and can also cause overfitting. As for the thrust and rudder force information output by the control system, their effect on the unmanned surface vessel's motion state is continuous and temporally correlated, not an instantaneous and isolated command action. Furthermore, the calculated thrust and rudder force are prone to numerical spikes and anomalous disturbances. This invention employs a one-dimensional convolutional neural network (1D CNN), which can effectively capture local temporal variation features and achieve weighted local mean smoothing through convolution operations. This suppresses high-frequency noise while preserving effective low-frequency trend components of the signal. Furthermore, 1D CNN, through convolution and pooling operations, can refine and reduce the dimensionality of high-dimensional raw data, significantly reducing the computational overhead and overfitting risk of subsequent networks while retaining key representational information. Therefore, 1D CNN is also used to extract local features and reduce noise from IMU data and the thrust and rudder force data output by the control system. Since there are strong nonlinear relationships between the specific force and angular velocity output by the IMU and the forward and lateral velocities of the carrier, and between the control information and the forward and lateral velocities of the carrier, a ReLU nonlinear activation function is used after 1D CNN processing to introduce nonlinear relationships. This maps the features of the convolution output to a non-negative interval, unifying the feature distribution and making it easier for subsequent models to learn effective relationships.

[0028] Limited by the local sliding window characteristic of convolutional kernels, 1D CNNs can only capture local, fragmented features within a short time domain, failing to establish long-term dependencies across time steps. However, the motion of an unmanned surface vessel (USV) is a continuous dynamic process. Its current velocity is not only related to recent acceleration and rudder angle commands but also influenced by control inputs and inertial states from several seconds prior. Therefore, a Long Short-Term Memory (LSTM) network is introduced. Its built-in gating mechanism (forget gate, input gate, output gate) selectively retains key historical information and suppresses irrelevant noise interference, effectively modeling long-term dependencies in temporal data. This allows for the global integration of local features extracted by the CNN along the time dimension, enabling the understanding and modeling of the complete motion process of the USV. To adapt to subsequent multimodal fusion requirements, fully connected layers are added after each branch of the LSTM, mapping the temporal features of different modalities to a unified dimensional space and completing feature condensation, achieving dimensional alignment and enhanced representation of the features from both branches.

[0029] Under different operating conditions, the contributions of the IMU branch and the control information branch to the velocity estimation results differ significantly: in still water environments with minimal wind, wave, and current disturbances (such as lakes and reservoirs), the motion of the unmanned surface vessel (USV) is mainly controlled by the control system; however, in complex marine environments such as nearshore areas, in addition to control inputs, external disturbances such as wind, waves, and currents also significantly affect the USV's navigation status. To address this, the model introduces a gated attention mechanism (AM) to adaptively weight and fuse the features output by the IMU and control information branches. This dynamically allocates the contribution weights of the two types of information according to the actual navigation conditions of the USV, achieving complementary advantages of different modal features while effectively suppressing the adverse effects of noise and disturbances. Furthermore, considering the significant differences in the physical characteristics, dominant influencing factors, and temporal variation patterns of forward and lateral velocities, two independent fully connected output heads are set after the attention-fused features to establish mapping relationships from the fused features to the forward and lateral velocities, respectively. This avoids mutual interference in the prediction processes of the two velocity components, thereby improving the model's estimation accuracy and robustness for different velocity components.

[0030] In summary, the CNN-LSTM-AM model constructed in this invention employs a multi-branch fusion architecture with two paths outputting forward and lateral velocities respectively. First, the two 1D CNNs extract and denoise local features from the proportional and angular velocity data output by the IMU and the temporal data of the control system, respectively. This preserves key local patterns while compressing redundant information, providing a more robust feature representation for subsequent temporal modeling. Second, the Long Short-Term Memory (LSTM) network, relying on its gating structure, effectively captures long-term temporal dependencies during motion, transforming the local features extracted by the CNN into a global temporal understanding of the inertial state and control commands. Finally, a gating attention mechanism enables adaptive fusion of features from the two modalities, dynamically allocating weights for the IMU and control signals according to different operating conditions, highlighting key information and suppressing noise interference, thereby achieving accurate estimation of forward and lateral velocities.

[0031] Meanwhile, when training the CNN-LSTM-AM model, a sliding window is used, with each window containing IMU data for 20 time steps and corresponding control information.

[0032] The input to the CNN-LSTM-AM model is and ,in, This is the input to the first CNN-LSTM branch (i.e., the IMU branch). This is the input to the second CNN-LSTM branch (i.e., the control information branch).

[0033] Includes comparison information within a window period and heading angular velocity information , and The letter 'b' in the equation represents the projection onto the sensor coordinate system, i.e. Tie, In Represents the sensor coordinate system ( (system) relative to the inertial coordinate system ( Transformation relationship of (system); and letters in For the first in the sliding window Group data. Among them, the sensor coordinate system ( The origin of the coordinate system is the measurement center of the IMU. The three coordinate axes point to the positive directions of the sensitive axes of the accelerometer and gyroscope, respectively. According to the actual installation method, in this invention, the three coordinate axes are positively pointed to "front, right and down". The inertial coordinate system is a basic coordinate system in the spatial stability mechanics arrangement system. The coordinate axes remain unchanged relative to the star and do not rotate with the earth, which can provide a reference benchmark for the inertial sensor. The formula is as follows: ; Contains all thrust information within a single window. With steering force information , of which letters For the first in the sliding window The data set is shown in the following formula: ; The output of the CNN-LSTM-AM model is , Includes the last moment within a window period, i.e. forward speed of unmanned surface vessel Lateral (rightward) velocity ,in The carrier coordinate system represents the coordinate system that is fixed to the carrier and moves with it. In this paper, the form "Front Right Lower (FRD)" is selected.

[0034] .

[0035] Combined with the predicted forward speed of the unmanned surface vessel With lateral velocity Then it is represented as: ; The aforementioned This refers to the measured velocities (forward and lateral velocities) in the carrier coordinate system, used in the filtering update process, and their values ​​are related to... same.

[0036] Corresponding observation vector This is expressed as the difference between the speed of INS solution and the model prediction: ; Velocity observation equation or observation vector It can also be expressed in the following form: ; in, For the observation matrix: ; In the formula: The projection of the velocity calculated by INS mechanical orchestration into the navigation coordinate system. Let be the rotation matrix that rotates the carrier coordinate system to the sensor coordinate system, where the superscript is denoted as . Represents the carrier coordinate system, subscript Represents the sensor coordinate system. For the sensor coordinate system ( Rotate to the navigation coordinate system. The rotation matrix of the system. Where the subscripts are... The navigation coordinate system is a local geographic coordinate system whose origin is located at the measurement center of the navigation system or the centroid of the vehicle. Depending on the selection order and orientation of the coordinate axes, it is divided into the Northeast-Upper-High (ENU) coordinate system and the Northeast-Earth (NED) coordinate system. This paper uses the Northeast-Earth (NED) navigation coordinate system as the navigation coordinate system. Represents an antisymmetric matrix. The link between the carrier coordinate system and the sensor coordinate system is located in the sensor coordinate system. Let be the antisymmetric matrix of the lever arm. It is the identity matrix. The state vector in the error Kalman filter includes position error, velocity error, attitude error, accelerometer bias error, gyroscope bias error, accelerometer scaling factor error, and gyroscope scaling factor error. This is the measurement noise vector.

[0037] Based on the above formula or measurement equation, the forward and lateral velocities of the unmanned surface vessel (USV) predicted by the CNN-LSTM-AM model are used as observations. In the case of GNSS denial, an error state Kalman filter update stage is introduced to correct the position, velocity and attitude of the USV.

[0038] In summary, this invention proposes a GNSS / INS integrated navigation method for unmanned surface vessels (USVs) that considers control information. When GNSS signals are normal, the IMU outputs specific force and angular velocity, which are then mechanically choreographed to obtain position, velocity, and attitude information. The GNSS receiver receives satellite signals and calculates the GNSS position information. The difference between the GNSS position and the mechanically choreographed position is then calculated, and this difference is used in filter fusion calculations to output position error, velocity error, and attitude error information. These errors are fed back to the mechanically choreographed position, velocity, and attitude results, yielding high-precision position, velocity, and attitude data. Based on this, the forward and lateral velocities of the USV are calculated using the velocity and heading angle. Simultaneously, a CNN-LSTM-AM model is trained offline using the raw IMU data, control information, and high-precision result data (forward and lateral velocities of the USV) to output forward and lateral velocities when GNSS signals are denied, providing constraint information for the INS.

[0039] When an unmanned surface vessel (USV) enters a GNSS lock-out or signal denial environment, this CNN-LSTM-AM model, combined with real-time IMU data and thrust information, can be used to predict the forward and lateral velocities of the USV. Subsequently, the difference between the predicted velocity and the velocity calculated by the mechanical orchestration is calculated and then input into a filter for state updates. The output includes position error, velocity error, and attitude error. These errors are then fed back to the position, velocity, and attitude results obtained from the mechanical orchestration, ultimately yielding corrected position, velocity, and attitude results. This enables high-precision continuous state estimation in GNSS lock-out or signal denial environments.

[0040] To verify the effectiveness of this invention, relevant experiments were conducted, specifically as follows: GNSS interruption simulation tests were performed based on two sets of measured unmanned surface vessel GNSS / INS integrated navigation data from a reservoir and a bay. Four sets of comparative experiments were used to verify the effectiveness of the method of this invention. The local coordinate system trajectory diagram of the reservoir scene is shown below. Figure 8 As shown, the trajectory diagram of the local coordinate system of the bay scene is as follows: Figure 9 As shown, in both scenarios, the thrust is not constant but continuously changing. In the reservoir scenario, each GNSS simulation interruption lasted 40 seconds, with an interval of approximately 150 seconds between interruptions, for a total of 13 interruptions, including 8 straight-line driving segments and 5 turning segments. To further verify the performance of the method of this invention, in the bay scenario, each GNSS interruption lasted 50 seconds, with an interval of approximately 130 seconds between interruptions, for a total of 11 interruptions, including 5 straight-line driving segments and 6 turning segments. The experimental setup for each method is shown in Table 1, and the parameters of the MO model are shown in Table 2. Both sets of experiments used the root mean square error of position (RMSE) during the GNSS simulation interruption period as the metric.

[0041] Table 1

[0042] The parameters of the MO model are shown in Table 2.

[0043] Table 2

[0044] During mission execution, unmanned surface vessels (USVs) may perform repeated tasks, leading to overlapping trajectories. Therefore, when visualizing the trajectory, the X and Y axes are defined as eastward and northward planar coordinates in meters (m), and the Z axis is set to the corresponding GPST time in seconds (s). This time dimension allows for effective differentiation of multiple overlapping trajectories. For example... Figure 8 , Figure 9 As shown, the blue trajectory represents the navigation trajectory during the GNSS signal availability phase, while the red trajectory corresponds to the navigation trajectory under the GNSS simulation denial environment.

[0045] The wind and wave conditions differ significantly between the reservoir and bay scenarios, and these environments are representative of common mission environments for unmanned surface vessels (USVs). The northward and eastward errors in the two scenarios are as follows: Figure 10 , Figure 11 As shown, Figure 10 This is a map showing the positioning error in the north / east direction for a reservoir scene. Figure 11 The positioning error map for the bay scene is shown in the north / east direction; the error statistics are shown in Tables 3 and 4.

[0046] Table 3 is a statistical table of errors in the reservoir scenario.

[0047] Table 3

[0048] Table 4 is a statistical table of errors in the bay scene.

[0049] Table 4

[0050] In the reservoir scenario, the MO model proposed in this invention achieved optimal positioning performance. In 13 GNSS simulation rejection tests with a single interruption duration of 40 seconds, the MO model's plane RMSE was 1.2073m. Compared to the pure inertial GNSS interruption condition without any external constraints and the constant velocity constraint scheme, the MO model reduced positioning errors by 80.17% and 67.78%, respectively. This is because pure inertial calculations lead to the continuous accumulation and divergence of inertial device errors over time, while constant velocity constraints cannot adapt to the dynamic navigation conditions of reservoir water flow disturbances and UAV speed and direction changes, exhibiting poor constraint adaptability. The MO model, however, uses dual-path CNN-LSTM to mine IMU inertial temporal features and thrust force dynamic control features respectively, outputting real-time dynamic velocity as a virtual observation, continuously correcting INS position and attitude errors, and effectively suppressing inertial calculation drift. Notably, the MO model further achieves complementary correction of IMU and thrust force information through a dual-branch structure, compensating for the shortcomings of single-source IMU input feature dimensions and insufficient representation capabilities. Therefore, compared to the CL-I scheme which relies on only a single information source, the positioning accuracy is improved by 29.50%.

[0051] In the Gulf scenario, the MO model also demonstrated superior positioning performance. In 11 GNSS simulation rejection tests with a single outage duration of 50 seconds, the MO model achieved a planar RMSE of 3.6463m, ranking first among all comparison scenarios. Compared to the pure inertial GNSS outage scenario (Outage, 2D RMSE = 15.8245m) without any external constraints and the constant velocity constraint scenario (ConVel, 2D RMSE = 5.0652m), the MO model reduced positioning errors by 76.96% and 28.01%, respectively. In the bay scenario, the water flow disturbances are stronger and the maneuvering conditions are more complex. Under pure inertial calculation, the INS error accumulates and diverges rapidly over time, and constant velocity constraints cannot adapt to the dynamic velocity changes in actual navigation, thus limiting the ability to constrain the INS. In contrast, the MO model uses dual-path CNN-LSTM to mine IMU inertial time-series features and thrust force dynamic control features respectively, outputting the dynamic velocity at the corresponding moment as the ESKF virtual observation value, which can continuously correct INS position and attitude errors, significantly improving positioning stability in denied environments. At the same time, the MO model achieves complementary correction of IMU and thrust force information through a dual-branch structure, giving full play to the information redundancy advantage of multi-source features. Therefore, compared with the CL-I scheme that relies on only a single information source, the positioning accuracy is improved by 13.45%, achieving the best positioning performance among all compared schemes.

[0052] This invention combines the constructed CNN-LSTM-AM model, introduces thrust information to construct multi-source feature input, uses dual-path CNN-LSTM for temporal feature extraction, and combines gated attention to achieve dynamic fusion. It achieves significant improvement in positioning performance in two typical scenarios, providing effective technical support for reliable navigation of unmanned surface vessels in GNSS-denied environments, and has good prospects for engineering applications.

[0053] For any parts not mentioned above, existing technologies can be adopted or referenced.

[0054] It should be noted that the above description is only a preferred embodiment of the present invention. For those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. All components not explicitly stated in this embodiment can be implemented using existing technology.

Claims

1. A GNSS / INS integrated navigation method for unmanned surface vessels that considers control information, characterized in that... Includes the following steps: Step a: Construct a CNN-LSTM-AM model, using IMU data and control information as input data to the CNN-LSTM-AM model, and outputting forward velocity and lateral velocity; Step b: The error state Kalman filter is used as the fusion algorithm for GNSS and INS. When the GNSS signal is normal, the inertial measurement unit outputs IMU data, which is then mechanically arranged and calculated to obtain position, velocity and attitude information. The GNSS receiver calculates the position information, subtracts it from the position calculated by mechanical arrangement, and inputs it into the filter for fusion calculation. The outputs position error, velocity error and attitude error data, corrects the position, velocity and attitude information obtained by mechanical arrangement, and obtains the corrected position, velocity and attitude results. Based on this, the forward velocity and lateral velocity of the unmanned surface vessel are calculated. Meanwhile, the constructed CNN-LSTM-AM model is trained using the IMU data, control information, and calculated forward and lateral velocities of the unmanned surface vessel during this phase. Step c: When the unmanned surface vessel enters a GNSS lockout or signal denial environment, the real-time IMU data and control information are input into the trained CNN-LSTM-AM model to predict the forward and lateral velocities of the unmanned surface vessel at that moment. Simultaneously, the IMU data output by the inertial measurement unit is mechanically programmed and calculated to obtain position, velocity, and attitude information; The predicted forward and lateral velocities of the unmanned surface vessel are subtracted from the velocities calculated by the mechanical orchestration, and the difference is input into a filter for state update to obtain position error, velocity error, and attitude error data. These data are then used to correct the position, velocity, and attitude information obtained by the mechanical orchestration, resulting in corrected position, velocity, and attitude results, which serve as the final integrated navigation output, enabling continuous state estimation in GNSS lock-out or denial environments.

2. The GNSS / INS integrated navigation method for unmanned surface vessels considering control information according to claim 1, characterized in that, In step a: the IMU data includes specific force and angular velocity, and the control information includes the thrust and rudder force output by the control system.

3. The GNSS / INS integrated navigation method for unmanned surface vessels considering control information according to claim 2, characterized in that, In step a: the CNN-LSTM-AM model includes a first CNN-LSTM branch, a second CNN-LSTM branch, and a gating attention mechanism module, namely the AM module; The first CNN-LSTM branch includes a first convolutional neural network and a first long short-term memory network. The specific force and angular velocity signals output by the inertial measurement unit enter the first convolutional neural network, are processed to obtain the first convolutional output features, are then mapped to the non-negative interval by the first ReLU nonlinear activation function, and are then input into the first long short-term memory network. The network is then globally integrated according to the time dimension, and processed by the first fully connected layer to obtain the output features of the first CNN-LSTM branch. The second CNN-LSTM branch includes a second convolutional neural network and a second long short-term memory network. The thrust and rudder force signals output by the control system are fed into the second convolutional neural network. After processing, the output features of the second convolutional network are obtained. Then, they are mapped to the non-negative interval by the second ReLU nonlinear activation function and then input into the second long short-term memory network. They are globally integrated according to the time dimension and then processed by the second fully connected layer to obtain the output features of the second CNN-LSTM branch. The output features of the first CNN-LSTM branch and the output features of the second CNN-LSTM branch are input together into the gating attention mechanism module and adaptively weighted fusion is performed to obtain the attention fusion features; The attention fusion features are processed by the third ReLU nonlinear activation function, and then passed through two independent fully connected output heads, namely the third fully connected layer and the fourth fully connected layer, to obtain the forward velocity and the lateral velocity, respectively.

4. The GNSS / INS integrated navigation method for unmanned surface vessels considering control information according to claim 3, characterized in that: Both the first and second convolutional neural networks are one-dimensional convolutional neural networks.

5. The GNSS / INS integrated navigation method for unmanned surface vessels considering control information according to claim 4, characterized in that: When training the CNN-LSTM-AM model, a sliding window is used, with each window containing IMU data for multiple time steps and corresponding control information. The input to the CNN-LSTM-AM model is and ,in, This is the input to the first CNN-LSTM branch. This is the input to the second CNN-LSTM branch; Includes comparison information within a window period and angular velocity information Comparison Information With angular velocity information The letter 'b' in the equation represents the projection onto the sensor coordinate system, i.e., the 'b' system; angular velocity information In Represents the sensor coordinate system, i.e. The system is relative to the inertial coordinate system, i.e. Transformation relationships of the system; comparative force information With angular velocity information Letters in subscript For the first in the sliding window Group data; Each window is configured to contain IMU data for 20 time steps and corresponding control information, as shown in the following formula: ； Contains all thrust information within a single window. With steering force information , of which letters For the first in the sliding window The data set is shown in the following formula: ； The output of the CNN-LSTM-AM model is , Includes the last moment within a window period, i.e. forward speed of unmanned surface vessel With lateral velocity superscript Represents the carrier coordinate system; 。 6. The GNSS / INS integrated navigation method for unmanned surface vessels considering control information according to claim 5, characterized in that: Combined with the predicted forward speed of the unmanned surface vessel With lateral velocity Then it is represented as: ； The measured velocity in the carrier coordinate system, its value is the same as... same; Corresponding observation vector This is expressed as the difference between the speed of INS solution and the speed of model prediction: ； Observation vector It can also be expressed in the following form: ； in, For the observation matrix: ； In the formula: The projection of the velocity calculated by INS mechanical orchestration into the navigation coordinate system. Let be the rotation matrix that rotates the carrier coordinate system to the sensor coordinate system, where the superscript is denoted as . Represents the carrier coordinate system, subscript Represents the sensor coordinate system; Let be the rotation matrix that rotates the sensor coordinate system to the navigation coordinate system, where the subscripts are... Represents the navigation coordinate system; Represents an antisymmetric matrix. The link between the carrier coordinate system and the sensor coordinate system is located in the sensor coordinate system. Let be the antisymmetric matrix of the lever arm, and diag be the identity matrix. This represents the state vector in the error Kalman filter. For measuring noise vectors; Based on the above formula, the forward and lateral velocities of the unmanned surface vessel predicted by the CNN-LSTM-AM model are used as additional observations. An error Kalman filter update stage is introduced under GNSS rejection conditions to obtain position error, velocity error and attitude error data. The position, velocity and attitude information obtained by mechanical arrangement calculation are corrected to obtain the corrected position, velocity and attitude results.

7. A GNSS / INS integrated navigation method for unmanned surface vessels considering control information according to claim 6, characterized in that, The steps of the adaptive weighted fusion are as follows: Output features of the first CNN-LSTM branch Features output by the second CNN-LSTM branch Projected onto the same dimension respectively: ； ； in, Output features for the first CNN-LSTM branch The weight matrix, This is the bias vector for the first CNN-LSTM branch. The feature vector after projection of the first CNN-LSTM branch; Output features for the second CNN-LSTM branch The weight matrix, This is the bias vector for the second CNN-LSTM branch. The feature vector after projection of the second CNN-LSTM branch; Then the feature vector after projecting the first CNN-LSTM branch is... Feature vectors projected from the second CNN-LSTM branch The weights are concatenated, and the sigmoid function is added to the fifth fully connected layer to generate the dimensionally gated weights of the first CNN-LSTM branch. ; ； in, This represents the sigmoid function. for and In concatenating vectors The weight matrix, for and In concatenating vectors The bias vector, It is a length of The vector; the second CNN-LSTM branch uses complementary weights. ; Attention fusion features for: ； in, This indicates element-wise multiplication.

Images