A ship trajectory prediction method based on attention-n-beats

Abstract

This invention provides a ship trajectory prediction method based on Attention-N-beats, relating to the field of intelligent shipping technology. It acquires discrete dynamic signals to construct an observation sequence, and generates an interpolation confidence mask by combining time residuals. The attention weight matrix is ​​calculated and the input tensor is reconstructed to achieve multi-dimensional physical state topological binding and natural low-pass filtering. A time-series refinement module with decreasing configuration of reconstructed feature inputs is used for dimensionality reduction, and physical constraint truncation is performed using the extreme value mask tensor. A customized physical constraint loss function is constructed by fusing high-order dynamic integral derivation and Fraser topological measure, and the parameters are updated. This invention breaks the assumption of pure numerical decoupling, eliminates the tearing of kinematic features and trajectory accumulation drift of large-inertia ships, and forces the predicted trajectory to converge within a zero-fault-tolerant physical safety envelope boundary.

Description

Technical Field

[0001] This invention relates to the field of intelligent shipping technology, and in particular to a ship trajectory prediction method based on Attention-N-beats. Background Technology

[0002] As the modern shipping industry transitions to new energy systems, large ships powered by lithium and hydrogen fuel cells are becoming increasingly common. Because these ships are high-energy, high-risk vessels, a yaw collision during navigation could easily trigger catastrophic consequences such as thermal runaway or a chain reaction of explosions. Therefore, modern maritime regulations impose stringent physical limits of "zero tolerance" on the navigational safety control of these ships, particularly regarding the absolute distance error (ADE / FDE) of multi-step trajectory prediction. High-precision ship trajectory prediction technology has become a core prerequisite for achieving the aforementioned microscopic safety envelope control.

[0003] Currently, this technical field typically employs deep learning time series models, such as the related time series prediction architecture published in CN114298411A, to perform multi-step trajectory prediction. These conventional solutions usually convert the ship's Automatic Identification System (AIS) data (such as longitudinal position, lateral position, longitudinal velocity, lateral velocity, turning angular velocity, and sailing angle) into sequence tensors, directly inputting them into a multi-layer fully connected network, aiming to fit the future evolution trend of the trajectory by progressively stripping residuals.

[0004] However, the aforementioned patents exhibit limitations in practical engineering applications, such as rapid decline in prediction accuracy and severe divergence in long-term predictions. The drawbacks are: First, the lack of implicit kinematic topological constraints across dimensions easily leads to physical feature fragmentation. Existing fully connected network layers rely excessively on isolated linear weighted view fitting of historical inputs in each dimension. This mechanism forcibly presupposes the static independence of each physical feature channel, completely ignoring the insurmountable nonlinear differential coupling relationship between the complex multidimensional kinematic states of a ship (such as the relationship between instantaneous changes in heading angle and lateral velocity and position displacement). For ships with significant inertia and maneuvering lag, this purely numerical decoupling process lacking physical priors inevitably leads to inconsistencies between the output coordinate differences and velocity vectors at physical boundaries, resulting in severe kinematic feature fragmentation.

[0005] Second, the temporal residual mechanism lacks a natural low-pass filtering capability for high-frequency white noise, making it highly susceptible to trajectory accumulation drift. The AIS data actually transmitted from ships inevitably contains a large number of high-frequency abrupt jumps (white noise) caused by sensor errors. Existing technologies, when performing residual stripping and downward propagation, lack a long-range physical response mechanism with global context, failing to eliminate disordered features at the network's inception. This not only forces deep networks to easily overfit to local random noise but also inevitably leads to huge systematic coordinate accumulation drift in multi-step temporal prediction due to the progressive accumulation of historical errors.

[0006] In summary, existing technologies lack effective coupling mechanisms for multi-dimensional kinematic features in multi-step time-series prediction. This isolated feature processing approach not only fails to dynamically filter out high-frequency data noise in the input signal, but also leads to severe nonlinear accumulation of errors as the prediction time domain extends when the system processes ship trajectories with large inertial characteristics. Ultimately, this results in the predicted trajectory being prone to divergence and drift, failing to accurately converge and be constrained within a safe and permissible physical envelope boundary. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a ship trajectory prediction method based on Attention-N-beats, which can dynamically filter high-frequency white noise and cancel systematic cumulative errors during the time-domain refinement process of multi-step prediction, thereby forcibly converging the motion trajectory of ships with large inertial characteristics and locking it within an extremely stringent microscopic safety envelope physical boundary.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides a ship trajectory prediction method based on Attention-N-beats, comprising the following steps: S1. Obtain discrete multidimensional dynamic signals to construct the original observation vector sequence. Calculate the time interval residual Construct a decay map to generate an interpolation confidence mask. And the interpolation confidence mask With the original observation vector sequence The concatenation mapping is used as the initial input tensor. ; S2, for the initial input tensor and its latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. Combined with interpolation confidence mask For the global covariance matrix Normalization is performed to calculate the attention weight matrix. And using the attention weight matrix For the initial input tensor Perform weighted recombination to generate normalized feature stream variables ; S3, Normalize the feature stream variables The input is fed into the temporal refinement module, which compresses the high-dimensional feature space into a funnel-shaped latent vector by progressively decreasing the neuron node parameters. Using the projection matrix to analyze the latent vector of the funnel Perform an up-dimensional mapping and connect it to a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Perform constraint truncation to generate deep prediction components. Inter-level residual tensor ; S4, Aggregate all deep prediction components Generate aggregated predicted trajectory tensors Using higher-order dynamic numerical integration operators to aggregate the predicted trajectory tensor Generate the physical trajectory tensor through time-domain integral derivation. A spatial topological difference measure mechanism is introduced to compare and aggregate predicted trajectory tensors. With integral derivation of physical trajectory tensor Generate spatial topological difference penalty term And combined with inter-level residual tensors Constructing a customized physical constraint loss function Based on customized physical constraint loss function Perform high-dimensional parameter updates.

[0009] In the preferred scheme, discrete multidimensional dynamic signals are obtained to construct the original observation vector sequence. ,include: Acquire historical navigation automatic identification system data from receiver terminals deployed on large lithium and hydrogen fuel cell ships; Based on historical Automatic Identification System (AIS) data, a six-dimensional set of physical quantities, including longitudinal position coordinates, lateral position coordinates, longitudinal velocity, lateral velocity, turning angular velocity, and navigation angle, is extracted. Under conditions of VHF channel congestion and base station packet loss, this six-dimensional set of physical quantities is used to construct an unaligned original observation vector sequence. ; Using a zero-order hold or a basic linear interpolation algorithm, the original observation vector sequence is interpolated according to a set reference time step. Mechanical fill-in is performed at missing moments to generate a time-equidistant basic interpolation observation vector sequence. .

[0010] In the preferred embodiment, in step S2, based on the global covariance matrix... Normalization is performed to calculate the attention weight matrix. Generate normalized feature stream variables A multi-level time series refining module is constructed, from coarse to fine, and the normalized feature stream variables are... Input to the timing refinement module includes: Extracting latent space feature dimensions The square root of the value is used to construct a scaling penalty factor, which is then applied to the global covariance matrix. Perform element-wise numerical smoothing to generate a scaled covariance matrix. ; Scaling the covariance matrix using the exponentially normalized activation function Perform feature mapping processing and calculate the attention weight matrix. This is to cut off the computing memory path that transmits high-frequency noise to deep networks through a dynamic spatiotemporal low-pass filter. A layer normalization mechanism is introduced to perform feature alignment processing on the feature dimensions, generating normalized feature stream variables. ; The time series refining module is determined to be the basic module of N-Beats, which will normalize the feature stream variables. Input to In the N-Beats basic module composed of stack modules, the cascaded structure within each stack module is utilized. Each Block submodule, and the 4-layer fully connected network FC Stack contained within each Block submodule, performs feature parsing processing to achieve decoupling between macroscopic physical baseline extraction and microscopic cumulative drift cancellation in the network topology. By utilizing the progressively decreasing neuron node parameters in the temporal refinement module, computational dimensionality compression is performed on the high-dimensional feature space, reducing the latent space feature dimension. Forced compression into funnel-shaped hidden vectors .

[0011] In the preferred scheme, a customized physical constraint loss function is used. Perform high-dimensional parameter updates, including: In the initial stage of model training, the basic loss function is constructed using the standard mean square error (MSE) and accelerated warm-up processing is performed on the network parameters. The Adaptive Moment Estimator (Adam) is introduced, utilizing a customized physical constraint loss function. The feedback gradient is used to perform the update calculation of model parameters, and the optimal combination of key hyperparameters is determined through cross-validation throughout the entire cycle.

[0012] In the preferred scheme, based on the time interval residual Process the attenuation mapping to generate an interpolation confidence mask. and the initial input tensor and initial input tensor The corresponding latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. ,include: Determine the physical attenuation penalty coefficient Based on physical attenuation penalty coefficient With time interval residual Perform exponential decay mapping and calculate the interpolation confidence mask. ; For the initial input tensor and the latent space transpose matrix Perform global matrix inner product operations to construct high-order cross-product terms involving heading angle, velocity, and displacement, generating a global covariance matrix. .

[0013] In the preferred scheme, the physical extremum mask tensor, which includes multidimensional kinematic extremum thresholds, is used. Perform constraint truncation to generate inter-level residual tensors. ,include: Determine the physical extremum mask tensor This includes the absolute maximum acceleration threshold and the maximum steering angular velocity threshold for ships under extreme operating conditions; Introducing the Hadamard product operator, which multiplies corresponding elements of representation matrices. Based on the Hadamard product operator Hidden vectors of the funnel The output after dimensionality-upgrading mapping and the physical extremum mask tensor Perform feature truncation mapping processing; The inter-stage network input parameters of the temporal refinement module during the feature transition process are obtained. With backward reconstruction parameters Input parameters for inter-level networks With backward reconstruction parameters Perform residual stripping calculations to generate inter-level residual tensors. .

[0014] In the preferred scheme, higher-order dynamic numerical integration operators are used to aggregate the predicted trajectory tensor. Perform time-domain integral derivation to generate the physical trajectory tensor for integral derivation. A spatial topological difference measure mechanism is introduced to compare and aggregate predicted trajectory tensors. With integral derivation of physical trajectory tensor Generate spatial topological difference penalty term And constructing a customized physical constraint loss function ,include: The higher-order dynamics numerical integration operator was determined to be the fourth-order Runge-Kutta numerical integration operator RK4. The fourth-order Runge-Kutta numerical integration operator RK4 was used in conjunction with the rigid body kinematic differential equations to converge and predict the trajectory tensor. Perform a second consecutive derivation process in the physical domain to generate an integral derivation physical trajectory tensor. ; The spatial topological difference measure mechanism is determined to be the Fraser distance algorithm, which is used to aggregate the predicted trajectory tensor. With integral derivation of physical trajectory tensor Perform dynamic search and comparison processing to calculate the aggregated predicted trajectory tensor. With integral derivation of physical trajectory tensor The maximum physical offset in the worst case generates a spatial topology difference penalty term. ; For all inter-level residual tensors Perform the extraction operation of the Frobenius norm square, combined with the set progressive weights. The results of the extraction operation are subjected to a weighted summation to generate an inter-level residual dissipation penalty term. To conduct rigorous physical energy dissipation assessment and monitoring of the disordered drift characteristic energy that has failed to be absorbed by networks at all levels; Obtain the fundamental coordinate mean square error characterizing the error of the conventional model. Based on the mean square error of the basic coordinates Spatial topological difference penalty item The corresponding first constant weighting coefficient and inter-level residual dissipation penalty The corresponding second constant weighting coefficient Mean square error of basic coordinates Spatial topological difference penalty item Inter-level residual dissipation penalty term Perform full mathematical aggregation calculations to construct a customized physical constraint loss function. This generates a combined penalized gradient and locks the descent solution space of the inverse gradient.

[0015] In a preferred embodiment, the present invention also provides a device for a ship trajectory prediction method based on Attention-N-beats, comprising: The initial input building block is used to acquire discrete multidimensional dynamic signals to construct the original observation vector sequence. Based on time interval residuals Process the attenuation mapping to generate an interpolation confidence mask. To characterize the original observation vector sequence Physical distortion characteristics; The initial input building block is also used to construct the interpolation confidence mask. With the original observation vector sequence Perform concatenation mapping to generate the initial input tensor. To provide a high-dimensional data base for pre-embedded penalty targets for downstream models; The normalized feature reconstruction module is used to normalize the initial input tensor. and initial input tensor The corresponding latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. To construct higher-order cross-product terms across physical quantities; The normalized feature recombination module is also used to combine interpolation confidence masks. For the global covariance matrix Perform normalization and calculate the attention weight matrix. This is to achieve low-pass filtering of high-frequency disordered features; The normalized feature reorganization module is also used to utilize the attention weight matrix. For the initial input tensor Perform weighted reorganization to generate normalized feature stream variables. In order to complete the feature fusion of physical priors and spatiotemporal context; The deep feature mapping module is used to normalize feature stream variables. The input is fed into the temporal refinement module, which utilizes the neuron node parameters configured in a progressively decreasing manner to refine the normalized feature stream variables. Compressive mapping is performed on the high-dimensional feature space to generate funnel-shaped latent vectors. This is to eliminate high-frequency nonlinear mutation features by reducing the computational power required for feature representation; The deep feature mapping module is also used to apply the projection matrix to the funnel-shaped latent vectors. Perform up-dimensional mapping processing and base it on a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Constraint truncation is performed on the output of the dimensionality-upgrading mapping to generate deep prediction components. Inter-level residual tensor This is to forcibly remove abnormal data that violates the physical laws of ship rigidity; The parameter prediction update module is used to update all deep prediction components. Perform aggregation processing to generate an aggregated predicted trajectory tensor. ; The parameter prediction update module is also used to aggregate the predicted trajectory tensor using higher-order dynamic numerical integration operators. Perform time-domain integral derivation to generate the physical trajectory tensor for integral derivation. To generate a theoretical continuous trajectory reference that obeys the integral equations of rigid body dynamics; The parameter prediction update module is also used to perform aggregated prediction trajectory tensor analysis based on a spatial topological difference metric mechanism. With integral derivation of physical trajectory tensor Perform continuous parameterized alignment processing to generate a spatial topological difference penalty term. To quantify the geometric distortions during the global evolution process; The parameter prediction and update module is also used to extract the residual tensors of all levels. The physical energy dissipation characteristics are compared with the spatial topological difference penalty term. Perform aggregation processing to construct a customized physical constraint loss function. ; The parameter prediction and update module is also used for loss functions based on customized physical constraints. The high-dimensional parameter update of the execution time-series refining module is performed to aggregate the predicted trajectory tensor. It converges and is constrained within the defined physical safety envelope boundary.

[0016] In a preferred embodiment, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the Attention-N-beats-based ship trajectory prediction method described above.

[0017] In a preferred embodiment, the present invention provides a computer non-transitory readable storage medium storing computer instructions thereon, characterized in that, when the computer instructions are executed by a processor, they implement the steps of the ship trajectory prediction method based on Attention-N-beats described above.

[0018] In a preferred embodiment, the present invention further provides a computer program product, including computer instructions, characterized in that, when the computer instructions are executed by one or more processors, they implement the steps of the Attention-N-beats-based ship trajectory prediction method described above.

[0019] This invention provides a ship trajectory prediction method based on Attention-N-beats. Through the coordination of the aforementioned structures, it offers the following advantages compared to existing methods: First, by constructing a soft penalty mapping for discrete signals with objective distortion characteristics and coordinating with the natural low-pass filtering mechanism in the feature recombination stage, it is possible to dynamically identify and eliminate high-frequency white noise and abrupt jumps caused by sensor errors or channel congestion from the source of computing power. This effectively overcomes the inherent problem of deep networks being prone to deep overfitting to local random noise and completely solves the problem of systematic trajectory accumulation and drift that is very easy to occur in multi-step long time domain prediction of large ships. Secondly, by forcibly establishing a high-order topological strong binding relationship covering multi-dimensional physical states in the latent space, and superimposing the computational dimension reduction and physical extreme value boundary hard truncation mechanism in the feature refinement process, the traditional model's linear fitting assumption of isolated decoupling of each physical dimension is broken through. This forces the prediction network to output only feature components that conform to the evolution law of large inertia and maneuvering lag of large ships, fundamentally eliminating the "kinematic feature tearing" phenomenon where the predicted coordinates and velocity vectors cannot be self-consistent on the physical boundary, and eliminating distorted data output that violates the rigid body law. Third, by abandoning the conventional pure mathematical numerical fitting path, a novel measure of the difference between high-order dynamic continuous deduction and spatial geometric topology is introduced in the backpropagation stage. This constructs a comprehensive penalty gradient flow full of multi-dimensional physical laws. This mechanism directly locks the descent solution space of the model weights, forcing the predicted trajectory in the long time domain to converge and be firmly constrained within the "zero-tolerance" microscopic safety physical envelope boundary required by maritime supervision. This significantly improves the absolute safety of ships in autonomous driving and high-risk navigation control under complex and extreme conditions. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a main view diagram of the process structure of this invention; Figure 2 This is a simulation diagram comparing the trajectory drift convergence and the physical boundary of the microscopic safety envelope for long-time domain prediction of large inertial ships according to the present invention; Figure 3 This is a block diagram of the device according to Embodiment 3 of the present invention; Figure 4 A schematic diagram of the structure of the computer device of the present invention. Detailed Implementation

[0021] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The exemplary embodiments disclosed in this application will be described below with reference to the accompanying drawings, which include specific technical details disclosed in this embodiment to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures (such as standard image processing algorithms and common communication protocols) are omitted in the following description.

[0022] With the development of intelligent ships and autonomous navigation technologies, real-time perception of the multi-dimensional kinematic state of ships can be achieved based on Automatic Identification Systems (AIS) and onboard Inertial Measurement Units (IMUs). Collaborating with maritime supervision and intelligent navigation systems can effectively improve shipping efficiency and safety. However, the navigation safety of large lithium- and hydrogen-fueled ships heavily relies on these perception networks. In the real physical ocean environment, very high frequency (VHF) communication channels are prone to congestion, and base stations frequently experience packet loss. Furthermore, sensor signals may also experience high-frequency abrupt changes (white noise) due to interference from severe sea conditions, resulting in highly non-equidistant and physically distorted multi-dimensional dynamic signals input to the system.

[0023] In the current technological context, multi-step trajectory prediction for ships is typically based on conventional deep learning time series models (such as standard N-BEATS or purely fully connected network architectures). However, these methods have serious limitations when dealing with large ships with high inertia and maneuvering lag: on the one hand, existing technologies often perform isolated linear fitting of historical inputs in various dimensions, ignoring the nonlinear differential coupling relationship between the transient change in heading angle and lateral velocity and displacement, resulting in "kinematic tearing" of the output trajectory; on the other hand, existing time-series residual mechanisms lack long-range physical response mechanisms to suppress disordered high-frequency noise, causing deep networks to easily fall into local overfitting, leading to severe "trajectory accumulation drift" as the prediction time domain extends. Therefore, current technologies cannot accurately converge the predicted trajectory and constrain it within the "zero-tolerance" microscopic safety physical envelope boundary required by maritime regulations.

[0024] Example 1 like Figure 1 As shown, this embodiment proposes a ship trajectory prediction method based on Attention-N-beats. Specifically, this step involves establishing the underlying physical object processed by the algorithm, and, while preserving the imperfect physical reality of the original data, providing a legitimate tensor basis for downstream attention filtering and residual networks through soft mapping of the trust factor and high-dimensional latent space embedding.

[0025] S11: Acquisition of Discrete Non-Isometric Multidimensional Dynamic Signals First, discrete, non-uniformly spaced, multi-dimensional dynamic signals are acquired. Specifically, high-frequency discrete data from historical Automatic Identification System (AIS) data are continuously collected using receiving terminals deployed on large lithium- and hydrogen-fueled vessels. In practice, due to objective physical limitations such as VHF channel congestion and base station packet loss, the signals acquired by the receiving end exhibit highly non-uniformly spaced discrete distribution characteristics on the time axis.

[0026] Secondly, for each effectively received frame of data, a set of six-dimensional physical quantities containing differentially constrained relationships is extracted and constructed, denoted as the original observation vector. ; ; in, For the first The original observation vector at each actual sampling time; and The first The longitudinal and lateral position coordinates of the ship at a real sampling moment; and These correspond to the longitudinal and lateral velocities. The angular velocity of the bow turn; The sailing angle.

[0027] Next, the multiple raw observation vectors collected consecutively will be... Arrange the observation vectors in one dimension according to their chronological order to construct the original, unaligned sequence of observations. Therefore, this original observation vector sequence This forms the foundational physical data source for all subsequent analyses in this system.

[0028] S12: Basic Physical Placement Interpolation and Timing Alignment Then, basic physical placeholder interpolation and timing alignment are performed. Specifically, to meet the strict time-step alignment requirements of subsequent tensor matrix operations, the original observation vector sequence to be processed is... Perform basic time step regularization.

[0029] In one feasible approach, a zero-order hold or a basic linear interpolation algorithm is used to sample the original observation vector sequence at a fixed frequency. Mechanically fills in the missing moments due to channel congestion, thereby generating a time-equidistant basic interpolation observation vector. .

[0030] In the preferred scheme, the set reference time step .

[0031] In practice, this step deliberately avoids high-order dynamic filtering.

[0032] Preferably, the higher-order dynamic filtering is an extended Kalman filter (EKF).

[0033] Specifically, the technical intent is to expose the packet loss errors and distortion characteristics of the original signal to the downstream algorithm intact, thereby preserving the objective physical authenticity of the time-series data. Furthermore, it involves generating all continuously linked basic interpolation observation vectors... Reorganized along the time axis, a strictly equidistant basic interpolated observation vector sequence is obtained. .

[0034] S13: Trust Factor Soft Mapping and High-Dimensional Tensor Expansion Next, a trust factor soft mapping and high-dimensional tensor augmentation are performed. Specifically, to compensate for data artifacts introduced by low-order interpolation in the previous step and to provide identifiable penalty targets for the downstream self-attention mechanism, the basic interpolation observation vector sequence is... Each basic interpolation observation vector in Force physical expansion processing.

[0035] First, calculate the time interval residual between the current interpolation time and the most recent actual AIS signal reception time. Based on this, an exponential decay mapping is constructed to generate an interpolation confidence mask. ; ; in, For the first Interpolation confidence mask for each time step; This is a custom physical attenuation penalty coefficient.

[0036] In the preferred embodiment, the physical attenuation penalty coefficient The range of values ​​is . in, This refers to the time interval residual. Specifically, when the data is actually received... At this point, the mask value is 1. Furthermore, when the data is interpolated after a long period of packet loss, this mask value approaches 0 exponentially.

[0037] Secondly, the time interval residual Interpolation confidence mask As an increment in the physical penalty dimension, it is related to the basic interpolation observation vector. Channel concatenation is performed to generate an expanded baseline vector containing eight dimensions. .

[0038] Then, after processing them sequentially, all the expanded dimension reference vectors are... Sequential combination generates an extended-dimensional benchmark sequence. .

[0039] S14: Historical Backtracking Sliding Window Segmentation Next, the history backtracking sliding window is segmented. In one feasible approach, a set length is used. Historical backtracking sliding window for extended dimension benchmark sequences Perform a cut operation with a fixed step size.

[0040] In practice, at each prediction time point, the current moment and historical continuous data are extracted. Extended dimensionality reference vector at each time step , will Stacking one-dimensional vectors along the time dimension generates a vector with dimension . Local observation matrix Therefore, this local observation matrix This constitutes the complete historical input view upon which the model relies for single-step trajectory prediction.

[0041] S15: High-Dimensional Latent Space Mapping and Residual Bridging Initialization Then, high-dimensional latent space mapping and residual bridging initialization are performed. Specifically, since an 8-dimensional physical space is too narrow, directly importing this physical space downstream for global covariance dot product calculation can easily lead to a computational bottleneck in feature representation. Therefore, the local observation matrix is ​​forcibly... The input is fed into the Linear Embedding Layer for dimensionality upscaling of the feature projection space; ; in, This is the initial input tensor of the mapped output. Specifically, the matrix dimension of this initial input tensor is... .

[0042] in, This is the local observation matrix output from the previous process.

[0043] in, Let be the learnable weight matrix for the linear embedding layer. Specifically, the dimension of this learnable weight matrix is ​​. ;in, The implicit space feature dimension set for the system.

[0044] In the preferred scheme, the latent space feature dimension The value is 512.

[0045] in, This is a bias vector used to prevent overfitting of the data.

[0046] In practice, the above mapping completes a legitimate transformation from the original low-dimensional physical space to the algorithm's high-dimensional latent space. The generated initial input tensor... It will serve as the primary data flow foundation for the entire underlying architecture, directly feeding into the next process of the subsequent cascaded network to perform spatiotemporal covariance filtering and physical topology binding.

[0047] In this embodiment, a cross-dimensional implicit kinematic topological binding and natural low-pass filtering method is proposed. In specific implementation, this process serves as the first component of the core computational architecture of this case, namely the Attention-N-Beats model, specifically manifested as an attention mechanism, namely the Attention module, that is forcibly introduced before the subsequent residual base module.

[0048] Specifically, this step is based on the initial input tensor. The initial input tensor is captured in depth using a scaling dot product self-attention mechanism. The output weights of each module are dynamically adjusted by combining the context information of historical trajectory fragments in order to effectively capture long-range dependencies in the time series.

[0049] In addition, its technical purpose is to break the traditional fully connected network's assumption of isolation of the physical dimension. Before the input features enter the deep N-Beats basic module, it forcibly constructs physical topology constraints that conform to the characteristics of the ship's large inertia, and works with the confidence penalty mechanism pre-embedded in the previous process to complete the soft strangulation of high-frequency disorder noise.

[0050] S21: Calculate the global spatiotemporal covariance matrix First, the calculation process for the global spatiotemporal covariance matrix is ​​executed. Specifically, the initial input tensor generated in the previous step S15 is obtained. Specifically, this initial input tensor The matrix dimension is Then, for this initial input tensor... Perform a matrix transpose operation to generate the corresponding latent space transpose matrix. Specifically, the latent space transpose matrix The dimension is .

[0051] Next, the initial input tensor With the implicit space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. ; ; in, This is the global covariance matrix. Specifically, this global covariance matrix... The dimension is .

[0052] In one feasible approach, this matrix inner product operation implements the globally correlated dot product of historical trajectory sequences in the time dimension at a purely mathematical level. Specifically, at the physical mapping level, due to the initial input tensor... In the previous process of dimension expansion and mapping, the static representation of the ship's position and the dynamic differential representation of its velocity and angular velocity have been integrated simultaneously. As a result, this inner product operation inevitably generates higher-order cross-product terms across physical quantities.

[0053] Specifically, this cross-product term profoundly encompasses the implicit nonlinear coupling relationship between the heading angle, lateral velocity, and positional displacement. In this embodiment, this cross-dimensional implicit kinematic topology binding mechanism ensures that the subsequently generated feature matrix naturally carries the physical priors of the dynamic differential equations, thus depriving the subsequent fully connected network of the computational power to output fracture trajectories that violate the laws of rigid body physics from the input source.

[0054] S22: Perform feature dimension scaling and smoothing processing Secondly, feature dimension scaling and smoothing are performed. In practice, due to the latent space feature dimension... The values ​​are on the order of magnitude. In the preferred scheme, the latent space feature dimension... The value is 512. Specifically, the unscaled global covariance matrix. The variance of internal elements increases dramatically with the increase of the dimension of the latent space, making subsequent exponential activation operations prone to falling into the saturation blind zone of gradient vanishing.

[0055] Then, the latent space feature dimension is introduced. The square root of the value is used as a scaling penalty factor for the global covariance matrix. Perform element-wise numerical smoothing to generate a scaled covariance matrix. ; ; in, To scale the covariance matrix; The implicit space feature dimension set for the system.

[0056] S23: Activating the synergistic effect of trust factors and natural low-pass filtering Then, the trust factor synergy is activated with the natural low-pass filtering. In one feasible approach, the scaling covariance matrix is... Each time step vector is subjected to exponential normalization, and the corresponding attention weight matrix is ​​calculated. ; ; in, This is the attention weight matrix; This is the exponentially normalized activation function.

[0057] In this embodiment, an extremely strong nonlinear cooperative gain occurs during this step. Specifically, the interpolation confidence mask is forcibly introduced in the previous step S13. With time interval residual As a physical penalty dimension, it is already rigidly embedded in the high-dimensional latent space features. Specifically, when the matrix performs global inner product and exponential normalization calculations, low-confidence interpolation jumps caused by VHF communication packet loss or base station sensor delays are addressed by their corresponding interpolation confidence masks. Approaching 0, this low-confidence data will inevitably be significantly diluted in the global dot product summation equation. Therefore, this low-confidence data will be significantly diluted in the attention weight matrix. It automatically obtains extremely small convergence weights.

[0058] Furthermore, this addresses the occasional high-frequency white noise generated by sensor measurement errors in the trajectory data of the automatic identification system. In practice, because this high-frequency white noise lacks a continuous physical response to the global context on the time axis, it hinders the calculation of the global covariance matrix. This inevitably results in an extremely low inner product projection value. After exponential normalization... The exponential decay suppression will force the elements of these weight matrices, which represent disordered characteristics, to zero.

[0059] Therefore, this step constructs a powerful dynamic spatiotemporal low-pass filter using underlying mathematical computing power without changing the original neural network topology. This removes disordered features at the input end and completely cuts off the computing memory path for high-frequency noise to be transmitted to the deep fully connected network.

[0060] S24: Fusing context to generate spatiotemporal co-feature tensors Next, a spatiotemporal collaborative feature tensor is generated by fusing the context. In specific implementation, the attention weight matrix generated in this process is used. For the initial input tensor Perform global weighted summation and context feature reorganization; ; in, This is a spatiotemporal collaborative feature tensor. Specifically, this spatiotemporal collaborative feature tensor... The matrix dimension is strictly maintained as .

[0061] In one feasible approach, through this matrix multiplication operation, the features of each specific time step are deeply integrated with the global contextual information of the entire historical trajectory sequence.

[0062] In the preferred scheme, low-frequency core features that conform to the dynamic laws of large inertia of large ships are given high weights and significantly enhanced, while high-frequency jump point features that violate the laws of physical continuity are deeply suppressed.

[0063] S25: Residual Normalization Bridging and Feature Flow Output Next, residual normalization bridging and feature transfer output are performed. In specific implementation, if the filtered spatiotemporal co-operational feature tensor... Directly replacing the original physical features with input to the deep residual network can easily sever the gradient backpropagation path of the original physical baseline features, thereby inducing gradient vanishing and semantic discontinuity during the model's backpropagation training phase.

[0064] Specifically, in order to completely bridge the gap in the spatiotemporal co-operational feature tensor To address the logical gap between this implementation and the subsequent fully connected stack, this embodiment forcibly introduces a residual connection mechanism. Additionally, a layer normalization mechanism is introduced synchronously.

[0065] Then, the initial input tensor Spatiotemporal co-feature tensor Element-wise summation of the matrices is performed, and the resulting tensor is then fed into a layer normalization function for rigorous alignment of the high-dimensional mean and variance, generating normalized feature stream variables. ; ; in, For normalized feature flow variables; This is the layer normalization function.

[0066] In this embodiment, the normalized feature stream variable It perfectly integrates the original physical prior features with the contextual features after global low-pass filtering.

[0067] Thus, the generated normalized feature stream variables As a secure and reliable underlying data bridge, it is directly fed into the next step of this technical solution.

[0068] In practice, the next step is a time-domain cascaded residual refining step based on an asymmetric physical funnel architecture.

[0069] This embodiment proposes a temporal cascaded residual refinement method based on an asymmetric physical funnel architecture. Specifically, this process is mapped to the N-Beats module, the foundational module in the Attention-N-Beats model. Specifically, this N-Beats module consists of… It consists of a stack of cascaded modules. Then, based on the normalized feature stream variables output in step S2... By constructing a multi-level temporal refinement mechanism from coarse to fine, the decoupling of macroscopic physical baseline extraction and microscopic cumulative drift cancellation is achieved in the network topology.

[0070] S31: Latitude Baseline Extraction of First-Level Network Features First, the tolerance baseline of the first-level network features is extracted. Specifically, the normalized feature stream variables generated in the previous step S2 are obtained. .

[0071] In one feasible approach, to satisfy the standardized tensor interface definition for cascading transfer of each Stack module in the N-Beats module, the normalized feature stream variable is passed before entering the network stack. Perform a matrix assignment mapping to establish the input starting point of the first-level Stack module at the beginning, and generate the first-level network input tensor. ; ; in, Input tensors to the first-level network; To normalize the feature stream variables.

[0072] In practice, each Stack module consists of... The system consists of several block sub-modules connected in series. Furthermore, each block sub-module contains a 4-layer fully connected network (FC) stack. In a preferred embodiment, the first-level stack module at the beginning is constructed using a tolerance-based fully connected layer matrix.

[0073] In the preferred embodiment, the number of neurons in the fully connected layer matrix with this tolerance is equal to the latent space feature dimension set by the system. .

[0074] Among them, the latent space feature dimension The preferred value is 512.

[0075] Then, input tensors to the first-level network The input is fed into a fully connected layer matrix of this tolerance for feature parsing to extract the first-level primary latent vector representing the macroscopic low-frequency motion of the ship. .

[0076] Next, using the preset first-level forward projection matrix With the first-level backward projection matrix For the first-level primary hidden vector Perform dimensional transformation to generate first-level prediction components. Reconstructing tensor with first-level baseline .

[0077] In addition, the input tensor of the first-level network is calculated. Reconstructing tensor with first-level baseline The matrix differences are used to generate the initial inter-level residual tensor. ; ; in, This is the initial inter-level residual tensor. Specifically, this initial inter-level residual tensor... It includes micro-temporal evolution features that are not fully covered by the macro baseline, and is then cascaded downstream.

[0078] S32: Self-attention bridging and normalization of deep feature propagation Secondly, self-attention bridging and normalization are performed for deep feature propagation. Specifically, in deep residual networks, if the unprocessed attention output is directly replaced with the original feature input to subsequent fully connected layers, the gradient propagation of the original physical features will be directly broken, leading to gradient vanishing during model training. Therefore, this step embeds a standard combination of bridging variables before entering each level of the deep network.

[0079] In one feasible approach, for the first The deep residual module obtains the first level of the previous level's output. Inter-level residual tensor .

[0080] Then, the first Inter-level residual tensor The inner product projection is performed on its corresponding transpose matrix, and the deep attention weight matrix is ​​calculated using the same scaled dot product self-attention operator as in the previous process. And the deep attention weight matrix Acting on the first Inter-level residual tensor Generate deep collaborative feature tensors .

[0081] Next, a residual connection mechanism and a layer normalization mechanism are introduced to... Inter-level residual tensor With deep collaborative feature tensor Element-wise summation is performed, and the mean and variance are aligned using the layer normalization function to generate the first... Level network input tensor ; ; in, For the first Level network input tensor; For layer normalization function; For the first Inter-level residual tensor; It is a deep collaborative feature tensor.

[0082] In this embodiment, the formula completely bridges the logical gap between the attention filter output and the fully connected stack. Specifically, this mechanism ensures that the physical baseline data, which is not distorted by the exponential function, can be safely integrated with the context features and gradients can be continuously propagated backward.

[0083] S33: Physical Feature Compression and Funnel Dimensionality Reduction of Deep Residual Networks Then, physical feature compression and funnel dimensionality reduction of the deep residual network are performed. Specifically, the dimensionality reduction of the deep residual network is... Level network input tensor It is then fed into the corresponding deep residual network module.

[0084] In this embodiment, a funnel-shaped bottleneck architecture is implanted inside the deep residual network module. Specifically, in the layer-by-layer propagation operation of the multi-layer fully connected network, the high-dimensional space of the input features is physically reduced in computational power by setting progressively decreasing neuron node parameters.

[0085] In the preferred scheme, the high-dimensional space is compressed layer by layer from 512 dimensions to 64 dimensions, ultimately outputting a low-dimensional funnel-shaped latent vector. ; ; in, The hidden vector of the funnel; This is a mapping function for the fully connected layer in a funnel-shaped bottleneck architecture.

[0086] Therefore, the funnel-shaped hidden vector At the mathematical level, the computational power to represent the complex characteristics of high-frequency nonlinear mutations is lost, forcing the deep residual network module to output only the system cumulative drift compensation value that conforms to the large inertia evolution law.

[0087] S34: Projective Dimensional Upgrading and Cascade Transmission under Physical Extremum Mask Constraints Next, projection dimensionality increase and cascaded propagation under the physical extremum mask constraint are performed. In specific implementation, due to the funnel-shaped hidden vector... Directly outputting a purely 64-dimensional scalar set would result in black-box discontinuities in the matrix transformation rules. Therefore, this embodiment introduces a deep forward projection matrix. With deep back projection matrix .

[0088] In one feasible approach, this deep forward projection matrix is ​​utilized. With deep back projection matrix Hidden vectors of the funnel To increase the dimensionality of mathematical computation, the deep initial prediction components were calculated separately. With deep initial reconstruction tensor ; ; ; in, For deep initial prediction components; For deep initial reconstruction tensors; This is the deep forward projection matrix; This is the deep back projection matrix.

[0089] Specifically, in order to establish the physical causality between the above-mentioned pure mathematical dimensionality increase operation and the actual rigid body motion, it is necessary to introduce entirely new constraint variables to eliminate dirty data that may be generated during matrix dimensionality increase that violates the laws of physics.

[0090] In this embodiment, a physical extremum mask tensor is constructed. Among them, the physical extremum mask tensor It includes the absolute maximum acceleration threshold and maximum turning angular velocity threshold of this type of lithium and hydrogen-powered ships under extreme operating conditions.

[0091] Then, the deep initial prediction component and deep initial reconstruction tensor Respectively with the physical extremum mask tensor Perform the Hadamard product operation to generate the final deep prediction components. With deep reconstruction tensor ; ; ; in, For deep prediction components; For deep reconstruction of tensors; For physical extremum mask tensors; This is the Hadamard product operator, which characterizes the element-wise multiplication of matrices.

[0092] In addition, calculate the first Level network input tensor With this deep reconstruction tensor The matrix difference is used to generate the updated first... Inter-level residual tensor In specific implementation, the first... Inter-level residual tensor The computation continues to be passed down through the cascaded layers until all layers have completed their calculations.

[0093] Therefore, all predicted components output from all network levels are collected and transferred to the state physics dimensionality upgrade and differential constraint loss network.

[0094] In this embodiment, a state physics dimensionality upgrade and differential constraint loss network method is proposed.

[0095] In practice, this process inherits the multi-level predicted features and inter-level feature flows output from the previous process. By completely abandoning the pure mathematical fitting path of a single basic mean square error, it implants rigid body dynamics differential equations, spatial topology metric mechanisms, and residual energy penalty terms into the bottom-level backpropagation gradient flow. This fundamentally locks the physical convergence boundary of the trajectory prediction of high-inertia, high-risk ships and ensures a tight closed loop for the feature flow of the entire network link.

[0096] S41: Linear aggregation of multi-level prediction components and inter-level residual collection First, perform linear aggregation of multi-level prediction components and collection of inter-level residuals.

[0097] Specifically, the first-level prediction components collected in the previous process are obtained. and all deep prediction components .

[0098] In addition, the first generation generated by the cascading transmission of all network layers in the previous process is extracted synchronously. Inter-level residual tensor .

[0099] In one feasible approach, the first-level prediction component With all deep prediction components Linear summation is performed across the time dimension and feature channels to generate an aggregated predicted trajectory tensor. ; ; in, To aggregate the predicted trajectory tensor; This is the first-level prediction component; For deep prediction components; This represents the total number of layers in the residual network module.

[0100] In practice, due to the dimensionality reduction and masking constraints achieved through the preceding physical funnel, the aggregated predicted trajectory tensor... During the forward propagation phase, high-frequency jump points and dirty data that violate continuity have been removed, forming the baseline discrete output values ​​for this forward prediction. Simultaneously, all collected data are fully retained. Inter-level residual tensor This serves as the core energy target for subsequently determining the dissipation rate of network physical characteristics at each level.

[0101] S42: Derivation of Physical Reference Trajectory Based on Higher-Order Dynamic Numerical Integrals Secondly, physical baseline trajectory derivation based on high-order dynamic numerical integration is performed. In practice, traditional deep learning networks only pursue the minimization of coordinate errors in a purely mathematical sense.

[0102] Specifically, the mechanism cannot perceive the continuous calculus relationship between discrete coordinate transformation and velocity vector, which inevitably leads to severe kinematic tearing of the output trajectory.

[0103] In this embodiment, a higher-order dynamic numerical integral operator is introduced to perform a secondary derivation of the physical domain of the prediction results.

[0104] Preferably, the higher-order dynamic numerical integration operator is a fourth-order Runge-Kutta numerical integration operator.

[0105] Then, from the aggregated predicted trajectory tensor Extract the initial position state matrix of the initial time step. With the dynamic differential prediction vector in the entire time domain .

[0106] Specifically, this dynamic differential prediction vector It includes the longitudinal speed, lateral speed, and turning angular velocity of the actual ship.

[0107] Next, the initial position state matrix With the dynamic differential prediction vector The data is synchronously input into the fourth-order Runge-Kutta numerical integral operator, and a four-step weighted cumulative integration is performed in the time domain strictly according to the rigid body kinematic differential equations to generate the physical trajectory tensor for integral derivation. .

[0108] Therefore, this integral deduces the physical trajectory tensor. It represents the purely theoretical continuous trajectory that the physical ship will inevitably travel from its initial state, under the boundary of strictly obeying the a priori laws of large inertial dynamics.

[0109] S43: Customized physical constraint loss function calculation integrating spatial topology metrics and residual dissipation assessment Then, a customized physical constraint loss function incorporating spatial topology metrics and residual dissipation assessment is calculated. In one feasible approach, the true labeled trajectory tensor is obtained. .

[0110] Specifically, the aggregated predicted trajectory tensor is calculated. With the true label trajectory tensor Mean square error of the basic coordinates between .

[0111] In practical implementation, the mean square error of the base coordinates It can only measure Euclidean distance at discrete time points and cannot capture the continuous geometric distortion of the entire trajectory sequence during its global evolution. Therefore, the Fraser distance algorithm is introduced as a limit penalty calculation rule for spatial topological differences.

[0112] Next, the Fraser distance algorithm is used to aggregate the predicted trajectory tensors. The network directly outputs coordinate sequences and integrally extrapolates physical trajectory tensors. The pure physical deduction coordinate sequence is used for continuous curve parameterization comparison.

[0113] Specifically, the maximum physical offset of the two sets of trajectories under the worst-case extreme condition is dynamically searched and calculated on the time axis to generate the Fraser distance penalty term. .

[0114] Furthermore, in order to completely close the multi-level residual network flow path in the previous process and completely eliminate isolated features, all collected residual network paths are processed. Inter-level residual tensor Conduct a basic physical energy dissipation assessment.

[0115] In practical implementation, calculate all the first... Inter-level residual tensor The matrix norm is determined, and a weighted summation is performed using progressive weights to generate an inter-level residual dissipation penalty term. ; ; in, This is a penalty term for inter-level residual dissipation. For the corresponding number The residual decay weight coefficients for each network layer; For the first Inter-level residual tensor The square of the Frobenius norm.

[0116] Specifically, the inter-level residual dissipation penalty term The energy of disordered drift characteristics that were not completely absorbed by the residual networks at each level was rigorously quantified in the physical domain.

[0117] In the preferred scheme, the residual decay weight coefficient increases with the depth of network hierarchy variables. This results in an exponential increase. Consequently, the model is forced to thoroughly absorb and dissipate all kinematic anomalies in a step-by-step manner during the feature transfer process from coarse to fine.

[0118] Next, the mean square error of the basic coordinates is... Fraser distance penalty item and inter-level residual dissipation penalty Perform full mathematical aggregation to construct a customized physical constraint loss function. ; ; in, For customized physical constraint loss functions; The physical topology penalty weight coefficient set for the system; The residual dissipation penalty weight coefficient set for the system.

[0119] In the preferred scheme, the physical topology penalty weighting coefficient The value is set to 10.0, which is the residual dissipation penalty weighting coefficient. The value is set to 5.0.

[0120] S44: Zero-tolerance backpropagation and convergence lock at the boundary of the microscopic collision envelope Next, zero-tolerance backpropagation and convergence lock-up are performed at the microscopic collision envelope boundary. Specifically, the calculated customized physical constraint loss function is used... As the sole global gradient feedback source during the model's backpropagation phase, it performs high-dimensional parameter updates on the full learnable projection matrix of the underlying architecture and the network weight view.

[0121] In this embodiment, once the predicted trajectory tensor is aggregated... A tiny physical tearing phenomenon, violating the rigid body differential equations, occurs in the process. This tiny error will be drastically amplified in the higher-order numerical integration derivation of the previous process, and will directly affect the Fraser distance penalty term calculated in this process. The values ​​increase exponentially. Furthermore, if the deep funnel network fails to fully utilize the residuals, leading to unused feature overflow, the residual dissipation penalty term between levels... This will generate a very large resistance gradient simultaneously.

[0122] The resulting massive comprehensive penalty gradient will forcibly correct the feature calculation direction of the deep network architecture. Specifically, this customized physical constraint loss function directly locks the descent solution space of the model's backpropagation gradient, forcing the weight parameters of the entire underlying computational architecture to approach only the legitimate physical convergence domain that perfectly conforms to the real ship dynamics integral equations and has zero residual overflow.

[0123] Ultimately, by forcibly suppressing and rigidly constraining the multi-step time-series aggregation output within an extremely narrow microscopic safety envelope physical boundary, the algorithm fundamentally achieves the absolute zero-tolerance requirement for predicted trajectory errors in high-risk navigation control.

[0124] S45: Model Training Optimization Based on Cross-Validation and Adaptive Momentum Finally, model training optimization based on cross-validation and adaptive momentum is performed. In specific implementation, this is done to ensure the aforementioned customized physical constraint loss function... It can converge stably and extract massive amounts of historical real trajectory data as a training set to iteratively train the entire Attention-N-Beats model.

[0125] In one feasible approach, the initial phase of model training uses the Standard Mean Squared Error (MSE) as the base loss function to accelerate network warm-up. Subsequently, the customized physical constraint loss function constructed in this embodiment is dynamically switched and adopted. Perform in-depth rigid body physics parameter tuning.

[0126] Next, the adaptive moment estimator optimizer, namely the Adam optimizer, is used to perform the update calculation of the model parameters. Specifically, the Adam optimizer is used based on the customized physical constraint loss function. The calculated massive comprehensive penalty gradient adaptively adjusts the learning rate of the network weights of each Stack module and Block sub-module, ensuring smooth convergence of the microscopic physical boundary constraints.

[0127] In specific implementation, step S3, during the model's forward propagation phase, cuts off the computational pathway for deep networks to memorize high-frequency disordered features at the structural level through feature dimension compression and physical pruning, and disposes of undigested, discarded features through the... Inter-level residual tensor Perform explicit isolation.

[0128] Furthermore, in step S4, during the model backpropagation stage, not only is the physical convergence lower bound of feature prediction locked by fusing high-order numerical integrals and spatial topological metrics, but the energy decay supervision of the isolated residual waste in S3 is also performed using the inter-level residual dissipation penalty term.

[0129] Example 2 In practice, the specific implementation of this case is compared with the comparative example CN114298411A in the background technology. The actual measured R&D data table of the deduced case exposes the fundamental logical shortcomings and physical limitations of the comparative example in dealing with the trajectory prediction scenario of large and high-risk ships.

[0130] First, the analysis begins with the underlying data characteristics and input processing mechanism. Specifically, the GDP data processed in this comparative model possesses inherent low frequency and smoothness, requiring only basic normalization to meet the input requirements of the subsequent fully connected network. However, in the navigation control scenario of large ships, the high-frequency discrete dynamic signals collected by the Automatic Identification System (AIS) are constantly constrained by the physical environment of VHF channel congestion. In practice, directly applying the basic data processing logic of this comparative model would completely fail to quantify the feature distortion caused by data latency. Therefore, this embodiment introduces an interpolation confidence mask. A soft mapping mechanism for trust factors based on imperfect physical reality data was constructed, successfully preventing low-confidence interpolation jump points from intruding into deep computing networks.

[0131] Secondly, a comparison is made regarding the topological coupling network mechanism for cross-dimensional features. In one feasible approach, this comparison directly processes multidimensional features using a stack of general blocks, forcibly presupposing a static assumption that each data channel is independent. Specifically, there are extremely tight nonlinear differential coupling relationships between the six-dimensional physical quantities of a physical ship. Therefore, this comparison lacks a global spatiotemporal covariance matrix. The strong binding between attention filtering mechanism and the construction of cross-dimensional implicit kinematic topology inevitably leads to severe coordinate differences and physical feature tearing of velocity vectors when dealing with large inertial entities.

[0132] Then, a final blind spot examination is conducted from the feature flow and physical boundary convergence of long-term prediction. In specific implementation, the architecture of this comparison model lacks an asymmetric physical funnel compression mechanism for feature dimensions. As a result, its deep network retains excessive high-dimensional computing power, making it extremely prone to deep overfitting to local high-frequency noise. Furthermore, this comparison model only pursues the minimization of a single mean square error in a purely mathematical sense, completely lacking high-order dynamic numerical integral derivation and spatial topological metric penalty terms. As a result, the gradient solution space generated by its backpropagation is not rigidly supervised by any physical laws and cannot aggregate the predicted trajectory tensor at all. Forced convergence and rigidly locked within the physical safety boundary of absolute zero fault tolerance.

[0133] To intuitively quantify the aforementioned physical disparity, actual measurement comparison data from a real R&D environment are introduced. In the preferred scheme, the test environment is set as a complex waterway navigation condition for a large ship with a 30% signal packet loss rate and interference from ocean waves, and the prediction time domain is extended to 120 seconds.

[0134]

[0135] Specifically, as can be seen from the above-mentioned measured R&D data table.

[0136] like Figure 2As shown, in one feasible approach, this comparative example, when processing ship trajectories with large inertial characteristics, exhibits a final distance error of up to 92.45 meters after 120 seconds as the prediction time domain extends, a kinematic feature tearing rate of 24.33%, and the most critical extreme yaw boundary overflow rate of 18.20%. This objectively demonstrates that this purely mathematical time-series mining tool has serious performance limitations when facing zero-fault-tolerant physical scenarios. Furthermore, this embodiment introduces a customized physical constraint loss function. and the global spatiotemporal covariance matrix The deeply coupled architecture successfully converged the overflow rate of the extreme yaw boundary to 1.22%, achieving absolute locking of the microscopic physical boundary based on measured data.

[0137] Example 3 Figure 3 This document presents a schematic diagram of the underlying hardware and architecture of a ship trajectory prediction device based on Attention-N-beats, as provided in an embodiment of this application. It should be understood that large ships (such as lithium- or hydrogen-fueled vessels) possess significant inertia and maneuvering lag, requiring a "zero-tolerance" physical safety boundary for trajectory prediction. This solution abandons conventional filtering methods that forcibly smooth data, instead preserving the objective distortion characteristics of the underlying data through soft mapping of the trust factor. Before features enter the deep residual network, a self-attention mechanism module is forcibly introduced to construct a cross-dimensional implicit kinematic topology strong binding, achieving natural low-pass filtering of high-frequency disordered features. Simultaneously, the traditional residual module undergoes funnel dimensionality reduction and physical extreme value mask constraints. Finally, a single basic mean square error is completely abandoned, and a high-order numerical integral derivation and spatial topology metric mechanism are forcibly implanted into the underlying backpropagation gradient flow. This application can fundamentally lock the physical convergence boundary of the predicted trajectory of large-inertia ships under extremely harsh communication and environmental interference conditions, completely solving the problems of long-term divergence and physical feature tearing, and significantly improving the absolute safety of high-risk navigation control.

[0138] This intelligent large vessel may include a sensing system, a computing platform, a communication system, and an execution control system. The sensing system may include one or more sensors that sense the surrounding marine environment, weather conditions, and the vessel's own kinematic state. For example, the sensing system may include one or more of the following: an Automatic Identification System (AIS) receiver, a Global Navigation Satellite System (GNSS / RTK), an Inertial Measurement Unit (IMU), shipborne navigation radar, lidar, and a panoramic camera device.

[0139] Some or all of the functions of intelligent large ships can be controlled by a computing platform. As the "central brain" of the ship, the computing platform is responsible for acquiring discrete multidimensional dynamic signals from the sensing system, running the ship trajectory prediction model provided in the embodiments of this application, and outputting accurate future time-series trajectories and safety envelopes.

[0140] In some possible implementations, intelligent large vessels may include Advanced Maritime Navigation Assistance Systems (ADAS). These systems utilize one or more sensors and communication systems within a perception system to acquire information about complex waters surrounding the vessel, and then analyze and process this information. Based on the high-precision, zero-fault-tolerance trajectory prediction algorithm provided in this application, the system can achieve functions such as dynamic obstacle trajectory tracking, microscopic collision envelope warning, autonomous trajectory planning, and automatic berthing and unberthing assistance, thereby improving the safety and automation of large, high-risk vessels navigating in complex, narrow waterways or the open ocean.

[0141] Under different levels of autonomous surface navigation (MASS, typically classified as L1 to L4), this computing platform can achieve different levels of autonomous navigation assistance based on artificial intelligence algorithms and multi-sensor fusion. L1 is a system with automatic system assistance; L2 is a remotely controlled vessel with a crew; L3 is a remotely controlled vessel without a crew; and L4 is a fully autonomous vessel. It should be understood that the higher the level of intelligent navigation, the more complex the corresponding collision avoidance mode, and the higher the accuracy requirements for the trajectory prediction model described in this application in terms of physical convergence and noise drift resistance in the long-term domain.

[0142] The ship trajectory prediction device can be implemented entirely within the computing platform of the aforementioned intelligent ship as an algorithm module; alternatively, it can be partially implemented within large intelligent ships (e.g., for edge data extraction and mask generation) and partially implemented in shore-based maritime cloud servers. The ship trajectory prediction device includes: an initial input construction module, a normalized feature reconstruction module, a deep feature mapping module, a parameter prediction and update module, and underlying physical sensing interfaces and heterogeneous computing hardware supporting the operation of the aforementioned logic modules. Specifically, this includes, but is not limited to, application processors (APs), graphics processing units (GPUs), neural network processors (NPUs), digital signal processors (DSPs), independent safety microcontrollers (MCUs, i.e., safety islands), and shipborne high-speed communication buses.

[0143] Data flow source and sensing interface: To ensure the fidelity of the underlying physical data, the device first handles data throughput through an onboard communication and sensing interface module. This interface module is directly connected in hardware to the ship's VHF communication baseband RF chip to continuously acquire historical Automatic Identification System (AIS) signals that are not evenly distributed due to channel congestion or base station packet loss. Simultaneously, this interface module, in conjunction with the onboard CAN-FD bus and inertial measurement unit (IMU), acquires the ship's longitudinal / lateral coordinates, velocity, and bow turning angular velocity, and reports these high-frequency signals with real physical distortions to the domain controller's main memory in real time.

[0144] The initial input construction module is used to acquire the above discrete multidimensional dynamic signals to construct the original observation vector sequence, calculate the time interval residual to construct the decay map to generate the interpolation confidence mask, and concatenate the interpolation confidence mask with the original observation vector sequence to map it into the initial input tensor.

[0145] At the underlying hardware execution level, the logic of this module is mainly executed by the central processing unit (CPU) or application processor (AP) in the domain controller. This is because calculating the time interval residual and aligning the time steps for non-equidistant data gaps involves a large number of timing logic judgments, out-of-order instruction scheduling, and branch control. The CPU's control flow architecture can most efficiently handle the alignment and mask soft mapping of imperfect physical real data, thus pre-embedding reliable penalty targets for downstream processes.

[0146] The normalized feature reorganization module is used to perform matrix inner product operation on the initial input tensor and its latent space transpose to generate a global covariance matrix. It then uses the interpolation confidence mask to normalize the global covariance matrix, calculates the attention weight matrix, and uses the attention weight matrix to perform weighted reorganization on the initial input tensor to generate normalized feature stream variables.

[0147] Since the core objective of this step is to break the isolated linear assumptions between dimensions and establish implicit kinematic topological bindings through high-order cross-product terms across physical quantities, it involves extremely large-scale high-dimensional tensor concurrent operations (MAC). Therefore, the computational task of this module is directly bypassed by the underlying operating system and hard-distributed to the graphics processing unit (GPU) or a dedicated neural network processor (NPU / Tensor Core) through system middleware. The large-scale systolic array inside the NPU can concurrently perform the dot product and exponential normalization of the attention mechanism with extremely high memory bandwidth, instantly eliminating high-frequency white noise and abrupt jumps with weights close to zero from the source of computing power.

[0148] The deep feature mapping module is used to input the normalized feature stream variable into the temporal refinement module. By progressively decreasing the neuron node parameters, the high-dimensional feature space is compressed into a funnel latent vector. The projection matrix is ​​used to perform dimensionality-up mapping on the funnel latent vector, and constraint truncation processing is performed with the physical extreme value mask tensor containing multi-dimensional kinematic extreme value thresholds to generate deep prediction components and inter-level residual tensors.

[0149] In terms of hardware mapping, the funnel dimensionality reduction (computational cost reduction) of the deep N-BEATS residual network is still accelerated by the aforementioned NPU. However, in order to completely eliminate the output of predicted coordinates that violate rigid body physical boundaries (i.e., to solve the "kinematic tearing" problem), the core task of "performing constraint truncation processing with the physical extreme value mask tensor" in this module is directed to a physically isolated independent microcontroller (MCU, or Functional Safety Island) within the domain controller. The absolute maximum acceleration and maximum turning angular velocity thresholds of the ship are hard-coded in the read-only memory (ROM) inside the MCU. As a low-level hardware watchdog, the MCU forcibly truncates the high-dimensional tensor output by the NPU by performing extremely low-latency Hadamard product operations; once dirty data exceeding the physical laws of ship inertia is detected, the MCU will directly clear it to zero at the register level to ensure the "zero-fault-tolerance" safety envelope for maritime supervision.

[0150] The parameter prediction and update module is used to aggregate all deep prediction components to generate an aggregated prediction trajectory tensor. It then uses a high-order dynamic numerical integration operator to perform time-domain integration on the aggregated prediction trajectory tensor to generate an integrally derived physical trajectory tensor. A spatial topological difference measure mechanism is introduced to compare the aggregated prediction trajectory tensor with the integrally derived physical trajectory tensor to generate a spatial topological difference penalty term. Finally, a customized physical constraint loss function is constructed by combining the inter-level residual tensor, and high-dimensional parameter updates are performed based on this customized physical constraint loss function.

[0151] In the heterogeneous computing power allocation, the fourth-order Runge-Kutta (RK4) numerical integration and Fréchet distance penalty term calculations performed by this module highly rely on rigorous continuous physical model measures rather than simple tensor concurrency. Therefore, these high-precision continuous calculus calculations are precisely scheduled to be executed in a digital signal processor (DSP) equipped with an ultra-high precision floating-point unit (FPU). After the DSP derives the theoretical continuous trajectory that strictly obeys the rigid body dynamics integral equations, it shares the customized physical constraint loss results with the GPU through a unified memory architecture (UMA). Finally, based on this comprehensive penalty gradient that includes evaluations of spatial topological distortion and residual dissipation, the GPU locks the descent solution space and performs backpropagation of the entire network weights and high-dimensional parameter updates.

[0152] Memory hardware support: Data flow between the aforementioned processors is highly dependent on a high-speed storage system. The memory in this device includes Universal Flash Memory (UFS) for persistent storage of large model weights and the operating system kernel, and Static Random Access Memory (SRAM / Cache) with extremely high bandwidth directly packaged inside the heterogeneous SoC. During the RK4 derivation and attention weight calculation stages, SRAM effectively eliminates the memory wall bottleneck that hinders the flow of high-dimensional latent vectors between multi-level residual networks, ensuring extremely low latency in the prediction response.

[0153] It is understood that the underlying architecture illustrated in the embodiments of this application does not constitute an absolute limitation on the virtual device of the present invention. In other embodiments of this application, the device can be implemented by application-specific integrated circuits (ASICs) or field-programmable gate arrays (FPGAs) with higher integration through pure hardware wiring to realize the above-mentioned logic of physical binding, funnel dimensionality reduction and integral loss calculation, thereby further ensuring the determinism and physical consistency of the system under extreme ocean conditions.

[0154] Furthermore, when the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] As mentioned above, with the evolution of large ships (such as lithium-ion and hydrogen-powered vessels) towards higher levels of intelligent navigation, collaborative maritime supervision and autonomous driving systems heavily rely on shipborne communication networks (such as VHF / AIS channels) and vehicle / shipborne sensors (such as IMUs and GNSS). However, in real-world ocean or busy port conditions, these communication networks and sensors are highly susceptible to anomalies due to channel congestion, adverse sea conditions, or signal blind spots. This results in highly non-uniformly distributed dynamic signals, and even the inclusion of numerous high-frequency abrupt jumps (white noise). Under the current technological context, relying solely on centralized, purely data-driven machine learning algorithms (such as conventional fully connected networks) can easily induce overfitting in deep networks when dealing with physically distorted data. This leads to divergence and drift in predicted trajectories, failing to meet the "zero-fault-tolerance" requirements of maritime supervision.

[0156] Therefore, the ship trajectory prediction device is not an ordinary computer in terms of physical entity, but a highly reliable computing platform that is deeply integrated or directly deployed in the ship's underlying control network. In order to accurately converge the predicted trajectory and constrain it within a safe physical envelope boundary without relying on perfect physical real data, the ship trajectory prediction device of this application can be specifically embodied as a shipborne autonomous driving domain controller, a shipborne edge computing node, or a heterogeneous central processing unit (SoC) of an advanced maritime navigation system, depending on the ship's intelligence level.

[0157] In some more specific implementation scenarios, the physical carrier of the ship trajectory prediction device described in this application can be specifically instantiated as a shipboard autonomous driving domain controller, a shipboard edge computing node, or a highly integrated heterogeneous central processing unit (SoC) according to the evolution level of the ship's electronic and electrical architecture (E / E Architecture).

[0158] In one feasible approach, when an onboard autonomous driving domain controller is deployed on a large vessel employing a centralized electronic and electrical architecture, it is specifically manifested as an onboard autonomous driving domain controller.

[0159] This domain controller serves as the central hub for the ship's overall planning and control, encompassing system layers such as perception analysis, state observation (anomaly detection), and planning and control.

[0160] In this architecture, the domain controller utilizes its internal graphics processing unit (GPU) or neural network processing unit (NPU) to run the "normalized feature reconstruction module" and the "deep feature mapping module." By performing inner product operations on the global covariance matrix of the input tensor, higher-order cross-product terms across physical quantities are artificially created to overcome the technical bias of conventional time-series models that isolate the processing of each dimension. More importantly, to ensure the absolute safety of planning control, this domain controller architecture mandates a monitoring and interception mechanism executed by an independent safety microcontroller (MCU, i.e., a functional safety island). When the NPU outputs deep prediction components, the MCU invokes the hard-coded maximum ship acceleration and maximum steering angular velocity thresholds to perform constraint truncation processing using physical extremum masks on the prediction results. In the execution logic of the planning control module: In one example, when the MCU detects that the predicted parameters have not exceeded the extreme value mask threshold, the planning control module normally sends the rudder angle and speed control values ​​to the actuator according to the predicted trajectory. In another example, when the predicted trajectory suffers severe "kinematic tearing" due to a network attack or sudden sensor failure, causing the MCU to truncate a large amount of abnormal data that violates the laws of physics, the planning and control module will immediately trigger a degradation protection mechanism based on this interception state (e.g., controlling the ship's intelligent driving level to degrade from L4 high automation to L2 partial automation, or directly triggering a safe berthing protocol), thereby safeguarding the physical safety boundary of the large inertia ship at the architectural level.

[0161] In practice, the shipboard autopilot domain controller serves as the computing hub for the entire ship, employing a high-bandwidth PCIe bus and multi-gigabit automotive Ethernet as its internal data backbone. Physically, the domain controller utilizes a liquid-cooled heat dissipation substrate to support ultra-high computing power consumption of hundreds of TOPS (trillions of operations per second).

[0162] In the preferred embodiment, the domain controller internally operates on Hypervisor-based hardware virtualization technology, strictly isolating the underlying computing resources into a "high real-time security domain" and a "high-performance computing domain." Specifically, tasks involving the absolute safety boundaries of rigid body dynamics, such as the aforementioned "fourth-order Runge-Kutta (RK4) numerical integration" and "physical extremum mask tensor truncation," are deployed in the security domain running a strong real-time operating system (RTOS) like QNX or VxWorks. Meanwhile, high-performance AI network computations, such as "parallel inner product of global covariance matrices" and "funnel latent vector feature compression," are deployed in the high-performance computing domain running a customized Linux operating system. Thus, the domain controller achieves absolute decoupling and coordination between "complex AI deduction" and "rigid body physical safety" in ship trajectory prediction at the system physical architecture level.

[0163] In this context, the domain controller can serve as a central nervous system and command center. The domain controller generates operational control signals based on instruction opcodes and timing signals to control instruction fetching and execution.

[0164] In another feasible approach, for ultra-large lithium- or hydrogen-fueled ships with highly dispersed radar and Automatic Identification System (AIS) sensors, the "initial input building block" of the shipborne edge computing node can be submerged and specifically instantiated as the shipborne edge computing node.

[0165] For ultra-large ships with highly dispersed sensing devices, the above system architecture can adopt a collaborative physical deployment of "cloud-edge-device". Among them, the "initial input building module" can be separately deployed and materialized as a shipborne edge computing node.

[0166] For example, this shipboard edge computing node is physically deployed close to the edge of the VHF antenna and sensors. Since it directly deals with massive amounts of raw, noisy, discrete high-frequency data, the edge computing node is primarily implemented using a Field-Programmable Gate Array (FPGA). Upon receiving a raw observation vector sequence containing packet loss and delay, the edge computing node utilizes the parallel pipeline within the FPGA to directly calculate the time interval residual at the hardware level and generate an "interpolation confidence mask" based on an exponential decay mapping. Thus, the edge computing node intercepts and quantifies signal anomalies at the forefront of data transmission, sending only the initial input tensor with confidence penalty targets to the main control system via the shipboard Ethernet. This architecture effectively avoids congestion of the shipboard core communication network caused by unprocessed high-frequency white noise and redundant data, improving the data interaction security of the entire shipboard system.

[0167] In practice, the shipborne edge computing node is physically located very close to the ship's VHF receiving antenna and sensing front end, and its internal structure is typically built using low-power field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs).

[0168] Specifically, due to the massive concurrent high-frequency discrete signals caused by VHF channel congestion, directly transmitting all the raw data back to the central processing unit would trigger severe bus storms and delays. In this embodiment, the edge computing node utilizes the abundant parallel logic units within the FPGA to directly execute the "extraction of discrete non-equidistant multidimensional dynamic signals" and "construction of attenuation mapping using time interval residuals" operations in step S1 at the very beginning of data generation (i.e., the edge side). Thus, the edge computing node directly purifies the raw signal carrying a large amount of white noise and applies an "interpolation confidence mask" within a nanosecond-level hardware pipeline. Subsequently, only the high-dimensional state tensor, after initial purification and mask mapping, is reported via the bus. This edge implementation significantly reduces the communication load on the shipboard backbone network, ensuring extremely low latency for multi-step trajectory prediction from a physical link perspective.

[0169] In a preferred embodiment, in order to achieve the ultimate computing power density and physical volume control, all the logic modules of the above-mentioned ship trajectory prediction device are fully miniaturized and solidified into a heterogeneous central processing unit (SoC) chip of an advanced maritime auxiliary navigation system.

[0170] In practice, the heterogeneous central processing unit (SoC) uses a network-on-chip (NoC) as a micro data exchange matrix to integrate application processing cores (ARM Cortex-A series), high-speed neural network engine (NPU), digital signal processing core (DSP), and lockstep security core (Lockstep Cortex-R series) on a single silicon chip with high density.

[0171] Specifically, within the micro-execution cycle of this heterogeneous central processing unit (SoC), data flows through the system with zero copies, following the Unified Memory Addressing (UMA) protocol. The attention weight matrix in step S2 mentioned above... The computational task is precisely mapped to a large-scale matrix multiply-accumulate (MAC) array within the SoC's NPU for Single Instruction Multiple Data (SIMD) acceleration; the spatial topology difference measure and loss function in step S4... The high-precision floating-point operations are scheduled to be processed in the DSP's dedicated Very Long Instruction Word (VLIW) execution unit. Furthermore, the lockstep security core within the SoC acts as the highest-level "supervisor," exclusively reading the physical extremum mask tensor to perform clock-cycle-level microscopic Hadamard product truncation verification on the deep prediction components output by the NPU. Thus, this heterogeneous SoC chip perfectly replicates the core technical loop of this solution—"high-computing-power topology binding" and "zero-fault-tolerant physical convergence"—at the nanometer-level microscopic physical silicon wafer level.

[0172] In another possible implementation, to meet the requirements of extremely low latency microcontrolling, all modules of the above device are fully integrated into a single heterogeneous central processing unit (SoC). This heterogeneous SoC tightly couples the application processor (AP), NPU, digital signal processor (DSP), and lockstep core through a network on-chip (NoC).

[0173] For example, this heterogeneous central processing unit (SoC) completely abandons the traditional path of calculating the mean square error using a single processor. It independently performs fourth-order Runge-Kutta (RK4) numerical integration using the floating-point unit inside the DSP to deduce the baseline trajectory under pure physical evolution; at the same time, it dynamically compares the Frescher distance (spatial topological difference) between the network output coordinates and the physical baseline trajectory. This underlying architecture, which heterogeneously parallelizes the "deep learning computing power (NPU)" and the "rigorous physical calculus computing power (DSP)" within a single chip, perfectly supports the calculation of the customized physical constraint loss function (i.e., the fusion of basic error, spatial topological penalty, and residual dissipation penalty) and the high-dimensional parameter update at the hardware micro level, ensuring that the backpropagation gradient flow always approaches the legal convergence domain that conforms to the rigid body dynamics integral equation.

[0174] It is understood that the embodiments of this application do not constitute a specific limitation on the ship trajectory prediction device based on Attention-N-beats. In other embodiments of this application, the device may include more or fewer components than illustrated (e.g., adding a display panel for human-computer interaction, including LCD, OLED, AMOLED, MiniOLED, etc.), or combining certain components, splitting certain components, or arranging different components. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0175] Example 4 Further explanation in conjunction with Example 1, such as Figure 4 The structure shown. Figure 4 A schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes: Processor, memory, communication bus, and computer programs stored in memory that can run on the processor.

[0176] The processor can call a computer program in memory, and when executing the program, implement the ship trajectory prediction method based on Attention-N-beats provided in the above embodiments. The method includes: S1, acquiring discrete multidimensional dynamic signals to construct an original observation vector sequence. Calculate the time interval residual Construct a decay map to generate an interpolation confidence mask. And the interpolation confidence mask With the original observation vector sequence The concatenation mapping is used as the initial input tensor. S2, Regarding this initial input tensor and its latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. Combined with the interpolation confidence mask For the global covariance matrix Normalization is performed to calculate the attention weight matrix. And using this attention weight matrix For this initial input tensor Perform weighted recombination to generate normalized feature stream variables S3. Normalize the feature stream variable The input is fed into the temporal refinement module, which compresses the high-dimensional feature space into a funnel-shaped latent vector by progressively decreasing the neuron node parameters. Using the projection matrix to analyze the latent vector of the funnel Perform an up-dimensional mapping and connect it to a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Perform constraint truncation to generate deep prediction components. Inter-level residual tensor S4, Aggregate all deep prediction components Generate aggregated predicted trajectory tensors The aggregation prediction trajectory tensor is analyzed using higher-order dynamic numerical integration operators. Generate the physical trajectory tensor through time-domain integral derivation. A spatial topological difference measure mechanism is introduced to compare the aggregated predicted trajectory tensor. With this integral, the physical trajectory tensor is derived. Generate spatial topological difference penalty term And combined with the inter-level residual tensor Constructing a customized physical constraint loss function Based on this customized physical constraint loss function Perform high-dimensional parameter updates.

[0177] Furthermore, computer equipment also includes: The Communications Interface (CI) is used for communication between the memory and the processor.

[0178] The memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0179] If the memory, processor, and communication interface are implemented independently, they can be interconnected via a bus to communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0180] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0181] Display devices are used to display images, videos, etc. Display devices may include display panels, which may employ liquid crystal displays (LCDs), organic light-emitting diodes (OLEDs), active-matrix organic light-emitting diodes (AMOLEDs), flexible light-emitting diodes (FLEDs), MiniLEDs, MicroLEDs, Micro-OLEDs, quantum dot light-emitting diodes (QLEDs), etc.

[0182] Alternatively, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0183] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described Attention-N-beats-based ship trajectory prediction method. This method includes: S1, acquiring discrete multidimensional dynamic signals to construct an original observation vector sequence. Calculate the time interval residual Construct a decay map to generate an interpolation confidence mask. And the interpolation confidence mask With the original observation vector sequence The concatenation mapping is used as the initial input tensor. S2, Regarding this initial input tensor and its latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. Combined with the interpolation confidence mask For the global covariance matrix Normalization is performed to calculate the attention weight matrix. And using this attention weight matrix For this initial input tensor Perform weighted recombination to generate normalized feature stream variables S3. Normalize the feature stream variable The input is fed into the temporal refinement module, which compresses the high-dimensional feature space into a funnel-shaped latent vector by progressively decreasing the neuron node parameters. Using the projection matrix to analyze the latent vector of the funnel Perform an up-dimensional mapping and connect it to a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Perform constraint truncation to generate deep prediction components. Inter-level residual tensor S4, Aggregate all deep prediction components Generate aggregated predicted trajectory tensors The aggregation prediction trajectory tensor is analyzed using higher-order dynamic numerical integration operators. Generate the physical trajectory tensor through time-domain integral derivation. A spatial topological difference measure mechanism is introduced to compare the aggregated predicted trajectory tensor. With this integral, the physical trajectory tensor is derived. Generate spatial topological difference penalty term And combined with the inter-level residual tensor Constructing a customized physical constraint loss function Based on this customized physical constraint loss function Perform high-dimensional parameter updates.

[0184] Furthermore, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a computer-readable storage medium. The computer program can execute computer instructions, and when executed by a processor, the computer can perform the Attention-N-beats-based ship trajectory prediction method provided by the methods described above. This method includes: S1, acquiring discrete multidimensional dynamic signals to construct an original observation vector sequence. Calculate the time interval residual Construct a decay map to generate an interpolation confidence mask. And the interpolation confidence mask With the original observation vector sequence The concatenation mapping is used as the initial input tensor. S2, Regarding this initial input tensor and its latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. Combined with the interpolation confidence mask For the global covariance matrix Normalization is performed to calculate the attention weight matrix. And using this attention weight matrix For this initial input tensor Perform weighted recombination to generate normalized feature stream variables S3. Normalize the feature stream variable The input is fed into the temporal refinement module, which compresses the high-dimensional feature space into a funnel-shaped latent vector by progressively decreasing the neuron node parameters. Using the projection matrix to analyze the latent vector of the funnel Perform an up-dimensional mapping and connect it to a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Perform constraint truncation to generate deep prediction components. Inter-level residual tensor S4, Aggregate all deep prediction components Generate aggregated predicted trajectory tensors The aggregation prediction trajectory tensor is analyzed using higher-order dynamic numerical integration operators. Generate the physical trajectory tensor through time-domain integral derivation. A spatial topological difference measure mechanism is introduced to compare the aggregated predicted trajectory tensor. With this integral, the physical trajectory tensor is derived. Generate spatial topological difference penalty term And combined with the inter-level residual tensor Constructing a customized physical constraint loss function Based on this customized physical constraint loss function Perform high-dimensional parameter updates.

[0185] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0186] For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. Furthermore, a computer-readable medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0187] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0188] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0189] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0190] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0191] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0192] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A ship trajectory prediction method based on Attention-N-beats, characterized in that, Includes the following steps: S1. Obtain discrete multidimensional dynamic signals to construct the original observation vector sequence. Calculate the time interval residual Construct a decay map to generate an interpolation confidence mask. And the interpolation confidence mask With the original observation vector sequence The concatenation mapping is used as the initial input tensor. ; S2, for the initial input tensor and its latent space transpose matrix Generate the global covariance matrix by performing matrix inner product operations. Combined with interpolation confidence mask For the global covariance matrix Normalization is performed to calculate the attention weight matrix. And using the attention weight matrix For the initial input tensor Perform weighted recombination to generate normalized feature stream variables ; S3, Normalize the feature stream variables The input is fed into the temporal refinement module, which compresses the high-dimensional feature space into a funnel-shaped latent vector by progressively decreasing the neuron node parameters. Using the projection matrix to analyze the latent vector of the funnel Perform an up-dimensional mapping and connect it to a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Perform constraint truncation to generate deep prediction components. Inter-level residual tensor ; S4, Aggregate all deep prediction components Generate aggregated predicted trajectory tensors Using higher-order dynamic numerical integration operators to aggregate the predicted trajectory tensor Generate the physical trajectory tensor through time-domain integral derivation. A spatial topological difference measure mechanism is introduced to compare and aggregate predicted trajectory tensors. With integral derivation of physical trajectory tensor Generate spatial topological difference penalty term And combined with inter-level residual tensors Constructing a customized physical constraint loss function Based on customized physical constraint loss function Perform high-dimensional parameter updates.

2. The ship trajectory prediction method based on Attention-N-beats according to claim 1, characterized in that, Obtain discrete multidimensional dynamic signals to construct the original observation vector sequence ,include: Acquire historical navigation automatic identification system data from receiver terminals deployed on large lithium and hydrogen fuel cell ships; Based on historical Automatic Identification System (AIS) data, a six-dimensional set of physical quantities, including longitudinal position coordinates, lateral position coordinates, longitudinal velocity, lateral velocity, turning angular velocity, and navigation angle, is extracted. Under conditions of VHF channel congestion and base station packet loss, this six-dimensional set of physical quantities is used to construct an unaligned original observation vector sequence. ; Using a zero-order hold or a basic linear interpolation algorithm, the original observation vector sequence is interpolated according to a set reference time step. Mechanical fill-in is performed at missing moments to generate a time-equidistant basic interpolation observation vector sequence. .

3. The ship trajectory prediction method based on Attention-N-beats according to claim 1, characterized in that, In step S2, based on the global covariance matrix Normalization is performed to calculate the attention weight matrix. Generate normalized feature stream variables A multi-level time series refining module is constructed, from coarse to fine, and the normalized feature stream variables are... Input to the timing refinement module includes: Extracting latent space feature dimensions The square root of the value is used to construct a scaling penalty factor, which is then applied to the global covariance matrix. Perform element-wise numerical smoothing to generate a scaled covariance matrix. ; Scaling the covariance matrix using the exponentially normalized activation function Perform feature mapping processing and calculate the attention weight matrix. This is to cut off the computing memory path that transmits high-frequency noise to deep networks through a dynamic spatiotemporal low-pass filter. A layer normalization mechanism is introduced to perform feature alignment processing on the feature dimensions, generating normalized feature stream variables. ; The time series refining module is determined to be the basic module of N-Beats, which will normalize the feature stream variables. Input to In the N-Beats basic module composed of stack modules, the cascaded structure within each stack module is utilized. Each Block submodule, and the 4-layer fully connected network FC Stack contained within each Block submodule, performs feature parsing processing to achieve decoupling between macroscopic physical baseline extraction and microscopic cumulative drift cancellation in the network topology. By utilizing the progressively decreasing neuron node parameters in the temporal refinement module, computational dimensionality compression is performed on the high-dimensional feature space, reducing the latent space feature dimension. Forced compression into funnel-shaped hidden vectors .

4. The ship trajectory prediction method based on Attention-N-beats according to any one of claims 1 to 3, characterized in that, Based on customized physical constraint loss function Perform high-dimensional parameter updates, including: In the initial stage of model training, the basic loss function is constructed using the standard mean square error (MSE) and accelerated warm-up processing is performed on the network parameters. The Adaptive Moment Estimator (Adam) is introduced, utilizing a customized physical constraint loss function. The feedback gradient is used to perform the update calculation of model parameters, and the optimal combination of key hyperparameters is determined through cross-validation throughout the entire cycle.

5. The ship trajectory prediction method based on Attention-N-beats according to claim 1, characterized in that, Based on time interval residuals Process the attenuation mapping to generate an interpolation confidence mask. and the initial input tensor and initial input tensor The corresponding latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. ,include: Determine the physical attenuation penalty coefficient Based on physical attenuation penalty coefficient With time interval residual Perform exponential decay mapping and calculate the interpolation confidence mask. ; For the initial input tensor and the latent space transpose matrix Perform global matrix inner product operations to construct high-order cross-product terms involving heading angle, velocity, and displacement, generating a global covariance matrix. .

6. The ship trajectory prediction method based on Attention-N-beats according to claim 1 or 5, characterized in that, Based on the physical extremum mask tensor including multidimensional kinematic extremum thresholds Perform constraint truncation to generate inter-level residual tensors. ,include: Determine the physical extremum mask tensor This includes the absolute maximum acceleration threshold and the maximum steering angular velocity threshold for ships under extreme operating conditions; Introducing the Hadamard product operator, which multiplies corresponding elements of representation matrices. Based on the Hadamard product operator Hidden vectors of the funnel The output after dimensionality-upgrading mapping and the physical extremum mask tensor Perform feature truncation mapping processing; The inter-stage network input parameters of the temporal refinement module during the feature transition process are obtained. With backward reconstruction parameters Input parameters for inter-level networks With backward reconstruction parameters Perform residual stripping calculations to generate inter-level residual tensors. .

7. The ship trajectory prediction method based on Attention-N-beats according to claim 6, characterized in that, Using higher-order dynamical numerical integration operators to aggregate the predicted trajectory tensor Perform time-domain integral derivation to generate the physical trajectory tensor for integral derivation. A spatial topological difference measure mechanism is introduced to compare and aggregate predicted trajectory tensors. With integral derivation of physical trajectory tensor Generate spatial topological difference penalty term And constructing a customized physical constraint loss function ,include: The higher-order dynamics numerical integration operator was determined to be the fourth-order Runge-Kutta numerical integration operator RK4. The fourth-order Runge-Kutta numerical integration operator RK4 was used in conjunction with the rigid body kinematic differential equations to converge and predict the trajectory tensor. Perform a second consecutive derivation process in the physical domain to generate an integral derivation physical trajectory tensor. ; The spatial topological difference measure mechanism is determined to be the Fraser distance algorithm, which is used to aggregate the predicted trajectory tensor. With integral derivation of physical trajectory tensor Perform dynamic search and comparison processing to calculate the aggregated predicted trajectory tensor. With integral derivation of physical trajectory tensor The maximum physical offset in the worst case generates a spatial topology difference penalty term. ; For all inter-level residual tensors Perform the extraction operation of the Frobenius norm square, combined with the set progressive weights. The results of the extraction operation are subjected to a weighted summation to generate an inter-level residual dissipation penalty term. To conduct rigorous physical energy dissipation assessment and monitoring of the disordered drift characteristic energy that has failed to be absorbed by networks at all levels; Obtain the fundamental coordinate mean square error characterizing the error of the conventional model. Based on the mean square error of the basic coordinates Spatial topological difference penalty item The corresponding first constant weighting coefficient and inter-level residual dissipation penalty The corresponding second constant weighting coefficient Mean square error of basic coordinates Spatial topological difference penalty item Inter-level residual dissipation penalty term Perform full mathematical aggregation calculations to construct a customized physical constraint loss function. This generates a combined penalized gradient and locks the descent solution space of the inverse gradient.

8. A device for ship trajectory prediction based on Attention-N-beats, characterized in that, include: The initial input building block is used to acquire discrete multidimensional dynamic signals to construct the original observation vector sequence. Based on time interval residuals Process the attenuation mapping to generate an interpolation confidence mask. To characterize the original observation vector sequence Physical distortion characteristics; The initial input building block is also used to construct the interpolation confidence mask. With the original observation vector sequence Perform concatenation mapping to generate the initial input tensor. To provide a high-dimensional data base for pre-embedded penalty targets for downstream models; The normalized feature reconstruction module is used to normalize the initial input tensor. and initial input tensor The corresponding latent space transpose matrix Perform matrix inner product operations to generate the global covariance matrix. To construct higher-order cross-product terms across physical quantities; The normalized feature recombination module is also used to combine interpolation confidence masks. For the global covariance matrix Perform normalization and calculate the attention weight matrix. This is to achieve low-pass filtering of high-frequency disordered features; The normalized feature reorganization module is also used to utilize the attention weight matrix. For the initial input tensor Perform weighted reorganization to generate normalized feature stream variables. In order to complete the feature fusion of physical priors and spatiotemporal context; The deep feature mapping module is used to normalize feature stream variables. The input is fed into the temporal refinement module, which utilizes the neuron node parameters configured in a progressively decreasing manner to refine the normalized feature stream variables. Compressive mapping is performed on the high-dimensional feature space to generate funnel-shaped latent vectors. This is to eliminate high-frequency nonlinear mutation features by reducing the computational power required for feature representation; The deep feature mapping module is also used to apply the projection matrix to the funnel-shaped latent vectors. Perform up-dimensional mapping processing and base it on a physical extremum mask tensor containing multidimensional kinematic extremum thresholds. Constraint truncation is performed on the output of the dimensionality-upgrading mapping to generate deep prediction components. Inter-level residual tensor This is to forcibly remove abnormal data that violates the physical laws of ship rigidity; The parameter prediction update module is used to update all deep prediction components. Perform aggregation processing to generate an aggregated predicted trajectory tensor. ; The parameter prediction update module is also used to aggregate the predicted trajectory tensor using higher-order dynamic numerical integration operators. Perform time-domain integral derivation to generate the physical trajectory tensor for integral derivation. To generate a theoretical continuous trajectory reference that obeys the integral equations of rigid body dynamics; The parameter prediction update module is also used to perform aggregated prediction trajectory tensor analysis based on a spatial topological difference metric mechanism. With integral derivation of physical trajectory tensor Perform continuous parameterized alignment processing to generate a spatial topological difference penalty term. To quantify the geometric distortions during the global evolution process; The parameter prediction and update module is also used to extract the residual tensors of all levels. The physical energy dissipation characteristics are compared with the spatial topological difference penalty term. Perform aggregation processing to construct a customized physical constraint loss function. ; The parameter prediction and update module is also used for loss functions based on customized physical constraints. The high-dimensional parameter update of the execution time-series refining module is performed to aggregate the predicted trajectory tensor. It converges and is constrained within the defined physical safety envelope boundary.

9. A computer device comprising at least one processor coupled to at least one memory storing at least one computer program or instruction, characterized in that, The computer program or instructions are loaded and executed by the processor to implement the steps of the Attention-N-beats-based ship trajectory prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, implement the steps of the Attention-N-beats-based ship trajectory prediction method as described in any one of claims 1 to 7.

Images