A method and system for predicting transient dynamics of an ice class vessel

Abstract

The application discloses a kind of ice area ship transient dynamics prediction method and system, wherein method includes obtaining the automatic identification system AIS historical navigation data of ship and inertial measurement unit IMU data;The AIS historical navigation data and IMU data are time synchronization, abnormality is eliminated, and historical observation window is constructed to downsampling;At least one angular velocity, acceleration or attitude related signal in the IMU data is carried out discrete wavelet transform, and low-frequency approximation component and high-frequency detail component are obtained;Transient ice exciting energy feature is calculated based on the high-frequency detail component;According to the transient ice exciting energy feature, generate environment gate weight;Based on the environment gate weight, adaptive fusion is carried out between water area dynamics expert and ice area collision dynamics expert, and the hidden representation of fused IMU is input into Transform prediction model with AIS feature;The ship speed, heading or position trajectory prediction result in future one or more prediction time domains is output by using the Transform prediction model.

Description

Technical Field

[0001] This invention relates to the fields of ship navigation, intelligent navigation in polar regions, inertial measurement data processing, and deep learning time series prediction. Specifically, it relates to a method and system for predicting the transient dynamic state of ships in ice-covered areas by utilizing the frequency domain decoupling characteristics of inertial measurement units (IMUs), historical navigation data from AIS, an environment-adaptive hybrid expert Transformer network, and cross-domain knowledge transfer strategies. Background Technology

[0002] With the increasing demand for Arctic shipping routes, polar resource development, and ice-covered transportation, the safety requirements for ships navigating in areas of broken ice, ice ridges, and mixed ice and water are constantly rising. The ice-covered navigation environment is characterized by strong nonlinearity, high randomness, and suddenness. During icebreaking or tracking, ships are not only affected by the continuous resistance from smooth ice, but may also experience transient impacts from hidden ice ridges, giant floating ice, or asymmetric icebreaking loads. Such impacts can cause rapid speed loss, lateral yaw, and yaw. If the predictive and control systems cannot identify these transient responses in time, it may lead to increased yaw, the ship becoming stuck on ice ridges, deviation from the course, or even a collision.

[0003] Existing ship trajectory prediction methods largely rely on Automatic Identification System (AIS) data. AIS data describes a ship's macroscopic navigation status, such as position, speed, and heading, and is suitable for trajectory prediction in ports, open waters, or conventional waterways. However, AIS data has a low update frequency and often fails to capture the high-frequency mechanical oscillations and short-term sideslip responses generated during transient collisions between the ship and ice. Traditional AIS-based deep learning models tend to degenerate into inertial extrapolation models under sudden disturbances in ice-covered areas, meaning they only extrapolate future trajectories based on historical headings and speeds, making it difficult to promptly identify asymmetric yaw risks caused by ice resistance.

[0004] Inertial measurement units (IMUs) can provide high-frequency angular velocity, acceleration, and attitude information, naturally including the vibration response of the ship's structure when subjected to ice impact. Therefore, fusing IMU data with AIS data can compensate for the lack of low-frequency sensing capabilities in AIS. However, the original IMU signal also contains low-frequency swaying caused by active rudder control, periodic motion caused by waves, mechanical noise, and high-frequency oscillations generated by ice impact. If the original signals are directly spliced ​​and input into a deep network, it is easy to cause aliasing of physical features, causing key ice impact features to be submerged by background noise.

[0005] Furthermore, acquiring real ice-induced collisions and other extreme ice-induced collision samples is costly and risky, and real-world ship data in ice-covered areas often suffers from limited sample size, class imbalance, and labeling difficulties. Existing complex deep neural networks, when directly trained on limited ice-covered data, are prone to overfitting normal navigation segments, resulting in insufficient generalization ability for low-frequency, high-risk ice-induced collision samples. Therefore, there is an urgent need for a ship transient dynamics prediction technology capable of extracting ice-induced physical features from IMU high-frequency signals, adaptively adjusting the prediction model according to environmental conditions, and transferring knowledge from clean water area data to ice-covered scenarios. Summary of the Invention The purpose of this invention is to provide a method and system for predicting the transient dynamics of ships in ice-covered areas, which addresses the problems of insufficient perception, prediction lag, lack of physically interpretable features, and overfitting caused by the scarcity of extreme samples in ice-covered areas in existing AIS trajectory prediction models in high-frequency transient impact scenarios.

[0006] To achieve the above objectives, the technical solution provided by the present invention includes the following steps: Acquire historical navigation data from the vessel's Automatic Identification System (AIS) and Inertial Measurement Unit (IMU) data; The AIS historical navigation data and IMU data are time-synchronized, anomaly removal is performed, downsampling is performed, and historical observation windows are constructed. Perform discrete wavelet transform on at least one angular velocity, acceleration, or attitude-related signal in the IMU data to obtain low-frequency approximate components and high-frequency detail components; Calculate transient ice-induced energy characteristics based on the high-frequency detail components; Environmental gating weights are generated based on the transient ice-induced energy characteristics. Based on the environmental gating weights, adaptive fusion is performed between water dynamics experts and ice collision dynamics experts, and the fused IMU features and AIS features are input into the Transformer prediction model. The Transformer prediction model is used to output the predicted ship speed, heading, or position trajectory for one or more future prediction time domains.

[0007] Furthermore, the AIS historical navigation data includes one or more of the following: ship position, speed, heading, and timestamp; the IMU data includes one or more of the following: three-axis acceleration, three-axis angular velocity, roll angle, pitch angle, bow angle, or bow angular velocity.

[0008] Furthermore, the time synchronization, anomaly removal, downsampling, and historical observation window construction include: mapping AIS data and IMU data to a unified time axis; interpolating or removing missing values ​​and anomalous jump values; downsampling high-frequency IMU data to a preset frequency; and using data of a preset duration prior to the current time as the historical observation window.

[0009] Furthermore, the discrete wavelet transform employs Daubechies wavelet, Symlet wavelet, Coiflet wavelet, or other orthogonal / biorthogonal wavelet basis functions with compact support characteristics for multi-level decomposition, preferably using the db4 wavelet basis function for three-level decomposition; wherein, the low-frequency approximation component is used to characterize active ruddering, ocean current drift, or macroscopic heading maneuvering, and the high-frequency detail component is used to characterize ice impact, ice ridge disturbance, or hull mechanical oscillation.

[0010] Furthermore, the transient ice shock energy feature is obtained by calculating the local variance of the reconstructed high-frequency detail components within a sliding time window, wherein the sliding time window is 5 to 30 seconds, preferably 10 seconds; the transient ice shock energy feature is used to characterize the transient intensity of ice resistance or ice shock collision experienced by the ship.

[0011] Furthermore, the environmental gating weights are calculated by a nonlinear gating function based on the normalized transient ice-induced energy characteristics. The nonlinear gating function is preferably a Sigmoid function. When the transient ice-induced energy characteristics increase, the weight of the ice-zone collision dynamics expert is increased and the weight of the water-zone dynamics expert is decreased; when the transient ice-induced energy characteristics decrease, the weight of the water-zone dynamics expert is increased and the weight of the ice-zone collision dynamics expert is decreased.

[0012] Furthermore, the adaptive fusion includes: inputting the IMU input features into the water region dynamics expert and the ice region collision dynamics expert respectively to obtain the corresponding water region feature representation and ice region feature representation; and performing weighted fusion of the water region feature representation and ice region feature representation according to the environmental gating weight to obtain the adaptively fused IMU hidden representation.

[0013] Furthermore, the Transformer prediction model additively fuses or splices AIS features, adaptively fused IMU hidden representations, and transient ice shock energy features, and injects position encoding. Through a multi-head self-attention mechanism, it establishes a long-range temporal correlation between historical ice shock vibrations, macroscopic navigation status, and future yaw trajectory.

[0014] Furthermore, the prediction results include one or more of the following relative to the current time: change in speed, change in heading, change in latitude and longitude, and a two-dimensional trajectory point sequence; wherein, an angle modulus operation is performed on the change in heading to eliminate numerical abrupt changes caused by a 360-degree span.

[0015] Furthermore, the Transformer prediction model is trained using a cross-domain knowledge transfer approach, including: firstly, pre-training using navigation data from clean water areas or open water areas to enable the model to learn the basic hydrodynamics and inertial laws under ice-free resistance; then, transferring the pre-trained model to the ice-covered target domain, freezing at least a portion of the encoder parameters, and fine-tuning only one or more of the prediction mapping head, environmental gating network, or ice-covered collision dynamics expert.

[0016] Furthermore, the fine-tuning is performed using real measured data from ice-covered areas, and employs one or more of the following strategies: Smooth L1 loss function, extreme condition sample reweighting, or few-sample sampling, to suppress overfitting and abnormal gradient effects during training with small samples in ice-covered areas.

[0017] Furthermore, the method also includes defining or identifying extreme ice skid conditions based on the actual speed changes and actual heading changes during the training or validation phase, and / or the predicted speed changes, predicted heading changes, and environmental gating weights during the online output phase, and independently identifying or providing risk warnings for the extreme ice skid conditions during the model training, validation, or output phases; wherein, the speed change threshold is preferably 1.0 knots, and the heading change threshold is preferably 10 degrees.

[0018] A transient dynamics prediction system for ships in ice-covered areas includes: a data acquisition and preprocessing module for acquiring and synchronizing historical AIS navigation data and IMU data; a frequency domain decoupling module for performing discrete wavelet transform on the IMU data and outputting low-frequency approximate components and high-frequency detail components; an ice-induced energy extraction module for calculating transient ice-induced energy characteristics based on the high-frequency detail components; an environmental gating and hybrid expert module for generating environmental gating weights based on the transient ice-induced energy characteristics and adaptively fusing water dynamics experts and ice-induced collision dynamics experts; a Transformer prediction module for outputting the future ship motion state or trajectory based on AIS features, IMU features, and transient ice-induced energy characteristics; and a trajectory and risk output module for outputting prediction results and ice-induced sideslip risk warnings.

[0019] It also includes a transfer learning training module, which is configured to perform one or more of the following: pre-training with clean water area data, fine-tuning with measured data from ice areas, freezing model parameters, updating the prediction head, updating the expert layer, or robustly optimizing the loss function.

[0020] An electronic device includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the transient dynamics prediction method for ships in ice-covered areas as described in any one of claims 1 to 12.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting transient dynamics of ships in ice-covered areas.

[0022] Compared with existing technologies, this invention has at least the following beneficial effects: First, by using discrete wavelet transform to decouple the IMU signal in the frequency domain, it separates low-frequency macroscopic maneuvers such as active steering from high-frequency oscillations generated by ice ridge impacts, thereby improving the model's ability to perceive ice-induced danger signals. Second, through transient ice-induced energy characteristics and environmental gating mechanisms, the model can adaptively switch between water dynamics experts and ice-induced collision dynamics experts, reducing misjudgments by a single model in cross-environment scenarios. Third, through knowledge transfer methods such as pre-training in clean water areas and fine-tuning in ice areas, a stable prediction model can still be obtained under conditions of scarce real ice-induced data, suppressing overfitting of small samples. Fourth, through the Transformer long-time attention mechanism, it can simultaneously utilize historical macroscopic navigation states and transient high-frequency ice-induced responses, improving the robustness of ship sideslip, yaw, and avoidance trajectory prediction in ice areas. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a method for predicting transient dynamics of ships in ice-covered areas, provided by an embodiment of the present invention.

[0025] Figure 2 This is a functional module block diagram of a transient dynamics prediction system for ships in ice-covered areas provided by an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of IMU frequency domain decoupling and environmental gating feature generation provided by an embodiment of the present invention.

[0027] Figure 4 This is a structural diagram of an environment-adaptive MoE-Transformer prediction model provided in an embodiment of the present invention.

[0028] Figure 5 This is a flowchart of cross-domain transfer learning for pre-training in a clean water zone and fine-tuning in an ice zone, provided by an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] like Figure 1 The method for predicting transient dynamics of ships in ice-covered areas, as shown, includes the following steps: Data acquisition and preprocessing: Acquire historical navigation data from ship AIS and inertial measurement data from IMU, and perform time synchronization, outlier processing, interpolation, downsampling, normalization, and construction of historical observation windows for the two types of data.

[0032] IMU frequency domain decoupling: Perform discrete wavelet transform on the IMU's pitch angular velocity, lateral acceleration, or other attitude-related signals to separate low-frequency approximation components and high-frequency detail components. The low-frequency approximation components are used to describe macroscopic maneuvering intentions, while the high-frequency detail components are used to describe ice impacts and structural oscillations.

[0033] Transient ice-cold energy extraction: Calculate the local variance or energy index of high-frequency detail components within a preset sliding time window, and form continuous transient ice-cold energy characteristics.

[0034] Environmental gating calculation: The environmental gating weight is calculated based on the transient ice shock energy characteristics. The environmental gating weight represents the degree to which the current sample is dominated by the dynamic state of the water zone or the collision dynamic state of the ice zone.

[0035] Environmental Adaptive Hybrid Expert Prediction: Water dynamics experts and ice collision dynamics experts are set up. The outputs of the two types of experts are fused according to environmental gating weights. The fused IMU hidden representation, along with AIS features and ice-induced energy features, is input into the Transformer network to obtain future speed, heading, and trajectory prediction results.

[0036] Cross-domain knowledge transfer training: First, pre-train the model using data from clean water areas or open water areas, then transfer the model to the ice area dataset, freeze some encoders or basic feature extraction layers, and only fine-tune the prediction head, environmental gating network or ice area collision dynamics expert layer.

[0037] Track and Risk Output: Outputs speed, heading, two-dimensional position trajectory and ice slip risk warnings for one or more future time scales, which can be used by ship assisted driving, automatic heading or collision avoidance systems.

[0038] Example 1: Transient Ice Cracking Feature Extraction Based on Frequency Domain Decoupling Reference Figure 1 and Figure 3 This embodiment first acquires macroscopic navigation data such as ship latitude and longitude, speed (SOG), heading (COG), and timestamps through shipborne AIS receiving equipment or historical AIS database. Simultaneously, it collects high-frequency inertial data such as three-axis acceleration, three-axis angular velocity, roll angle, pitch angle, and bow angular velocity through shipborne IMU. Since the AIS and IMU sampling frequencies differ, the system unifies both types of data onto the same time axis. For high-frequency IMU data, it can be downsampled to 0.5Hz to 10Hz, preferably 1Hz, depending on computing resources and prediction requirements, to balance transient response preservation and model calculation efficiency.

[0039] The system constructs an observation window based on a preset historical length up to the current time t. For example, in a preferred embodiment, the historical observation window length is 60 seconds, and the prediction time domain is 10 seconds, 30 seconds, or 60 seconds. The invention is not limited to these values; other window lengths can be set according to ship type, speed, sensor frequency, and control system response time.

[0040] When performing discrete wavelet transform on the IMU signal, the system preferably selects the bow roll rate. As the primary decoupling target, lateral acceleration, longitudinal acceleration, or multiple IMU channels can also be selected for decoupling. For any discrete IMU sequence The system performs multi-level decomposition using low-pass and high-pass filters to obtain the first... Low-frequency approximate components of the layer and multiple high-frequency detail components Its decomposition relationship can be expressed as:

[0041] in, The original discrete signal of the IMU. For the first Low-frequency approximation components of the layer, For the first High-frequency detail components of the layer, This represents the wavelet decomposition level. Low-frequency approximation components. Primarily used to characterize ship active steering, ocean current drift, or macroscopic heading maneuvering trends, high-frequency detail components It is mainly used to characterize ice ridge impact, ice fragment impact, asymmetric ice load, and high-frequency oscillation of ship hull structure.

[0042] In a preferred embodiment, a three-level decomposition is performed using the db4 wavelet basis function. Alternatively, a multi-level decomposition can be performed using Daubechies wavelet, Symlet wavelet, Coiflet wavelet, or other orthogonal / biorthogonal wavelet basis functions with compact support properties. The system then calculates the decomposition based on a preset high-frequency detail layer set. Reconstructing high-frequency detail components:

[0043] in, For the reconstructed high-frequency detail components, This is the selected set of high-frequency detail layers. This reconstruction method allows the removal of low-frequency macroscopic maneuvering components from the raw IMU signal, while retaining high-frequency response components related to ice collisions, ice fragmentation, and mechanical oscillations.

[0044] Subsequently, the system has a length of The local mean and local variance of the high-frequency detail components are calculated within the sliding time window. The local mean of the sliding window is expressed as:

[0045] The transient ice-cold energy characteristics are represented as follows:

[0046] in, For a moment The corresponding transient ice-cold energy characteristics, The length of the sliding time window. This is the local mean of the high-frequency detail components within the sliding window. In a preferred embodiment, The duration can be set from 5 to 30 seconds, preferably 10 seconds. The transient ice-induced energy characteristic is used to convert discrete high-frequency mechanical oscillations into a continuous environmental drag intensity index. When a ship encounters ice ridge impact, ice fragmentation, or asymmetric ice load, the transient ice-induced energy characteristic increases and serves as the input to the subsequent environmental gating network.

[0047] Example 2: Environment-Adaptive Hybrid Expert Transformer Prediction Reference Figure 2 and Figure 4 This embodiment sets up an environmental gating and hybrid expert module, and constructs an environmentally adaptive MoE-Transformer prediction model. The system will incorporate transient icing energy characteristics. After normalization, the input is given to the gating network to obtain the environment gating weights. The gating weights can be calculated using the Sigmoid function: Ice cream energy normalization:

[0048]

[0049] in, The normalized characteristics of ice-induced energy. To prevent constants with a denominator of zero, For gating weights for ice zone collision dynamics experts, For water dynamics experts, gating weights This is the gating sensitivity coefficient. This is the ice-induced energy threshold or a learnable bias. The larger the value, the more likely the current sample is to be in a state dominated by ice zone collision dynamics; The smaller the value, the more likely the current sample is to be in a clean water area or a stable navigation state.

[0050] The hybrid expert module includes water zone dynamics experts. Collision dynamics experts in ice zones In this invention, the water zone includes clean water areas, open water areas, or navigation areas without significant ice resistance. Water zone dynamics experts are used to learn smooth hydrodynamic laws in the aforementioned water zones or during conventional navigation, while ice zone collision dynamics experts are used to learn yaw laws caused by ice ridge impacts, ice fragmentation, and asymmetric ice loads. The system performs weighted fusion of expert outputs based on gating weights.

[0051] in, For the adaptively fused IMU hidden representation, As a water dynamics expert, An expert in collision dynamics in ice zones, This is a feature for IMU decoupling.

[0052] The system will use AIS features Adaptive fusion Hidden representations of H_IMU and transient icing energy characteristics The fusion process can be performed. Fusion methods include additive fusion, linear mapping after splicing, or attention fusion.

[0053] in, The fused multimodal input sequence For AIS features, [;] indicates a concatenation operation. For linear or nonlinear mapping functions, For position encoding.

[0054] The fused sequence, after adding position encoding, is input into the Transformer encoder. The Transformer encoder utilizes a multi-head self-attention mechanism to globally model the macroscopic navigation state, low-frequency maneuvering trends, and high-frequency ice shock energy peaks within the historical observation window, thereby establishing a temporal correlation between historical impact vibrations and future yaw trajectories.

[0055] in, , , These are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector.

[0056] The prediction head can output the speed change over multiple future time steps. Change in heading , Latitude and longitude changes or two-dimensional trajectory point sequences:

[0057] in, To predict the mapping header, For the change in speed, For the change in heading, This refers to the change in position or the change in a two-dimensional trajectory point.

[0058] To avoid numerical discontinuities in the heading angle near 0 degrees and 360 degrees, the system performs modulo operations or angle normalization on the heading change.

[0059] in, Restricted to Within the interval, to eliminate heading angle crossings. Numerical mutations caused by time.

[0060] The final predicted trajectory can be reconstructed based on the ship's current position, speed, and predicted changes.

[0061] Example 3: Knowledge Transfer Training for Few-Sample Ice Zone Scenarios Reference Figure 5 Due to the high cost and danger of acquiring real-world extreme collision data in ice-covered areas, this invention employs a two-stage cross-domain transfer learning strategy combining pre-training in clean water areas and fine-tuning in ice-covered areas. The first stage is the pre-training stage in clean water areas: using a large amount of data from clean water areas, open waters, or conventional waterways, the Transformer encoder, water dynamics expert, and basic feature mapping layer are trained, enabling the model to learn the inertial laws, course-maintaining laws, and smooth hydrodynamic motion patterns of ships under ice-free resistance conditions.

[0062] in, The training dataset is for clean water areas or open water areas. For model parameters, These are the model parameters obtained after pre-training.

[0063] The second stage is the ice-region fine-tuning stage: the pre-trained model is transferred to the ice-region target domain, freezing some encoder layers, basic embedding layers, or water dynamics expert parameters, and only releasing some parameters in the prediction mapping head, environment gating network, and ice-region collision dynamics expert for fine-tuning. This allows the model to quickly adapt to the high drag, high yaw, and high oscillation characteristics of the ice-region target domain while maintaining basic hydrodynamic knowledge.

[0064] in, This is a dataset based on actual measurements in the ice-covered area. For predicting fine-tunable parameters in mapping heads, environmental gating networks, or ice collision dynamics experts, Some encoder parameters remain frozen.

[0065] During fine-tuning, the Smooth L1 loss function can be used to train the model. This loss function has convergence accuracy equal to the squared error when the error is small, and robustness to the absolute error when the error is large, making it suitable for handling abnormally large yaw errors caused by ice ridge impacts. Furthermore, extreme ice-induced sideslip samples can be resampled or loss-weighted to avoid diluting the training contribution of high-risk samples by a large number of stable samples.

[0066] in, The error between the predicted value and the actual value, This is the smoothing threshold.

[0067] Example 4: Identification and Risk Output of Extreme Ice Slippage Conditions In a preferred embodiment, during the training or validation phase, the system can label extreme ice-induced sideslip samples based on future real-time speed and heading changes. During the online prediction phase, risk warnings can be generated based on predicted speed changes, predicted heading changes, environmental gating weights, or ice-induced energy characteristics. For example, if the speed decreases by more than 1.0 knots or the heading changes by more than 10 degrees within the future prediction time domain, the sample is labeled as an extreme ice-induced sideslip condition. These thresholds can be adjusted based on ship size, speed, ice condition level, and control strategy.

[0068] in, Marked for extreme ice-induced sideslip conditions. For indicator functions, The speed change threshold, The heading change threshold, This is the environmental gating threshold.

[0069] The risk warnings output by the system may include one or more of the following: peak ice energy, environmental gating weight, predicted heading deviation, predicted lateral displacement, and future trajectory deviation.

[0070]

[0071] in, Assess the risk of ice cream skidding. To predict lateral displacement, , , , These are the weighting coefficients. This is a bias term. When... When the preset risk threshold is exceeded, the system outputs a warning about the risk of ice cream slipping.

[0072] When the environmental gating weight remains above the threshold and the predicted heading change increases rapidly, the system can send a risk warning to the driver assistance system or automatic heading control system, so that the downstream control system can execute deceleration, rudder angle correction or avoidance strategies in advance.

[0073] This invention is applicable not only to icebreakers but also to ice-class cargo ships, research vessels, polar supply ships, and other vessels navigating in ice break zones, ice ridge zones, or mixed ice-water zones. The method can be deployed on shipboard edge computing devices, shore-based monitoring platforms, or ship-shore collaborative intelligent navigation systems. The risk scoring described above is only one possible implementation; this invention can also employ threshold determination, rule fusion, or a learning classifier to output risk warnings.

[0074] Example 5: System Modules and Deployment Methods Reference Figure 2This invention also provides a transient dynamics prediction system for ships in ice-covered areas. This system may include a data acquisition and preprocessing module, a frequency domain decoupling module, an ice-induced energy extraction module, an environmental gating and hybrid expert module, a Transformer prediction module, a transfer learning training module, and a trajectory and risk output module. These modules can be implemented using software programs, hardware circuits, programmable logic devices, shipborne edge computing devices, shore-based servers, or combinations thereof.

[0075] The data acquisition and preprocessing module is used to acquire ship AIS and IMU data. AIS data may include one or more of the following: ship position, speed, heading, and timestamp. IMU data may include one or more of the following: three-axis acceleration, three-axis angular velocity, roll angle, pitch angle, bow angle, or bow angular velocity. This module also performs time synchronization, outlier removal, interpolation, downsampling, normalization, and historical observation window construction on the AIS and IMU data, thus forming a multi-source time-series input for subsequent model processing.

[0076] The frequency domain decoupling module performs discrete wavelet transform on at least one angular velocity, acceleration, or attitude-related signal in the IMU data to obtain low-frequency approximate components and high-frequency detail components. The low-frequency approximate components characterize active ship steering, ocean current drift, or macroscopic heading maneuvers, while the high-frequency detail components characterize ice impact, ice ridge disturbance, ice fragmentation, or ship mechanical oscillations. This module can perform three-level decomposition using the db4 wavelet basis function, or multi-level decomposition using Daubechies wavelet, Symlet wavelet, Coiflet wavelet, or other orthogonal / biorthogonal wavelet basis functions with compact support properties.

[0077] The ice-induced energy extraction module is used to calculate transient ice-induced energy characteristics based on the high-frequency detail components. In one embodiment, the ice-induced energy extraction module calculates the local variance or energy index of the high-frequency detail components within a preset sliding time window and uses it as a feature characterizing the transient intensity of ice drag, ice ridge impact, or asymmetric ice load. The transient ice-induced energy characteristics can also be used as input to an environmental gating and hybrid expert module to determine whether the current ship motion state is closer to a smooth navigation state in the water or a collision disturbance state in the ice zone.

[0078] The environmental gating and hybrid expert module generates environmental gating weights based on transient ice-induced energy characteristics and adaptively fuses the outputs of the water zone dynamics expert and the ice zone collision dynamics expert based on these environmental gating weights. In this invention, the water zone includes clear water areas, open water areas, or navigation areas without significant ice resistance. The water zone dynamics expert is used to learn the smooth hydrodynamics and inertial motion laws of ships under the aforementioned water zone conditions; the ice zone collision dynamics expert is used to learn the speed loss, yaw, and sideslip laws caused by ice ridge impact, ice fragmentation, and asymmetric ice loads.

[0079] The Transformer prediction module receives AIS features, adaptively fused IMU hidden representations, and transient ice-induced energy features. It then outputs changes in ship speed, heading, position, or a sequence of two-dimensional trajectory points for one or more future prediction time domains through position encoding, a multi-head self-attention mechanism, and a prediction mapping head. Based on the prediction results, this module can further reconstruct the future ship trajectory and provide input for assisted navigation, automatic heading control, or collision avoidance control.

[0080] The transfer learning training module is used to perform pre-training in clean water areas and fine-tuning in ice-covered areas. Specifically, in the first stage, this module uses data from clean water areas, open water areas, or regular waterways to train the basic feature mapping layer, the Transformer encoder, and the water dynamics expert, enabling the model to learn the basic hydrodynamics and inertial motion laws under ice-free resistance. In the second stage, this module uses a small amount of measured data from ice-covered areas to fine-tune the pre-trained model, freezing at least some encoder parameters and opening one or more parameters from the prediction mapping head, the environment gating network, or the ice-covered collision dynamics expert for updates, thereby improving the model's adaptability to ice-covered scenarios with few samples.

[0081] The trajectory and risk output module is used to output future speed, heading, two-dimensional position trajectory, and ice blast risk warnings. In one embodiment, when the predicted speed change, predicted heading change, environmental gating weight, or transient ice blast energy characteristics exceed a preset threshold, the trajectory and risk output module generates an ice blast risk warning. The risk warning can be sent to the ship's assisted driving system, automatic heading system, collision avoidance decision system, shipboard display terminal, shore-based monitoring platform, or ship-shore cooperative intelligent navigation system.

[0082] The aforementioned modules can be integrated into the same shipboard computing device or distributed across the ship, shore, or cloud. For example, the data acquisition and preprocessing module, frequency domain decoupling module, and ice shock energy extraction module can be deployed on a shipboard edge computing device to achieve real-time processing of AIS and IMU data; the transfer learning training module can be deployed on a shore-based server or cloud platform to train models using historical navigation data and measured data from ice zones; and the Transformer prediction module and trajectory and risk output module can be deployed on the ship or in a ship-shore collaborative system to achieve online prediction and risk warning. This invention does not limit the specific hardware deployment of each module; any module that can achieve the aforementioned data processing, frequency domain decoupling, environmental gating, prediction output, and risk warning functions should fall within the scope of protection of this invention.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the transient dynamics of ships in ice-covered areas, characterized in that, include: Acquire historical navigation data from the vessel's Automatic Identification System (AIS) and Inertial Measurement Unit (IMU) data; The AIS historical navigation data and IMU data are time-synchronized, anomaly removal is performed, downsampling is performed, and historical observation windows are constructed. Perform discrete wavelet transform on at least one angular velocity, acceleration, or attitude-related signal in the IMU data to obtain low-frequency approximate components and high-frequency detail components; Calculate transient ice-induced energy characteristics based on the high-frequency detail components; Environmental gating weights are generated based on the transient ice-induced energy characteristics. Based on the environmental gating weights, adaptive fusion is performed between water dynamics experts and ice collision dynamics experts, and the fused IMU features and AIS features are input into the Transformer prediction model. The Transformer prediction model is used to output the predicted ship speed, heading, or position trajectory for one or more future prediction time domains.

2. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The AIS historical navigation data includes one or more of the following: ship position, speed, heading, and timestamp; the IMU data includes one or more of the following: three-axis acceleration, three-axis angular velocity, roll angle, pitch angle, bow angle, or bow angular velocity.

3. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The time synchronization, anomaly removal, downsampling, and historical observation window construction include: mapping AIS data and IMU data to a unified time axis; interpolating or removing missing values ​​and anomalous jump values; downsampling high-frequency IMU data to a preset frequency; and using data of a preset duration prior to the current time as the historical observation window.

4. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The discrete wavelet transform employs Daubechies wavelet, Symlet wavelet, Coiflet wavelet, or other orthogonal / biorthogonal wavelet basis functions with compact support characteristics for multi-level decomposition, preferably using the db4 wavelet basis function for three-level decomposition; wherein, the low-frequency approximation component is used to characterize active ruddering, ocean current drift, or macroscopic heading maneuvering, and the high-frequency detail component is used to characterize ice impact, ice ridge disturbance, or hull mechanical oscillation.

5. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The transient ice shock energy feature is obtained by calculating the local variance of the reconstructed high-frequency detail components within a sliding time window, wherein the sliding time window is 5 to 30 seconds, preferably 10 seconds; the transient ice shock energy feature is used to characterize the transient intensity of ice resistance or ice shock impact experienced by the ship.

6. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The environmental gating weights are calculated by a nonlinear gating function based on the normalized transient ice-induced energy characteristics. The nonlinear gating function is preferably a Sigmoid function. When the transient ice-induced energy characteristics increase, the weight of the ice-zone collision dynamics expert is increased and the weight of the water-zone dynamics expert is decreased; when the transient ice-induced energy characteristics decrease, the weight of the water-zone dynamics expert is increased and the weight of the ice-zone collision dynamics expert is decreased.

7. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The adaptive fusion process is as follows: the IMU input features are respectively input into the water zone dynamics expert and the ice zone collision dynamics expert to obtain the corresponding water zone feature representation and ice zone feature representation; The water region feature representation and the ice region feature representation are weighted and fused according to the environmental gating weights to obtain the adaptively fused IMU hidden representation.

8. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The Transformer prediction model additively fuses or splices AIS features, adaptively fused IMU hidden representations, and transient ice shock energy features, and injects position encoding. It establishes a long-range temporal correlation between historical ice shock vibrations, macroscopic navigation status, and future yaw trajectory through a multi-head self-attention mechanism.

9. The method for predicting transient dynamics of ships in ice-covered areas according to claim 1, characterized in that: The Transformer prediction model is trained using a cross-domain knowledge transfer approach, which includes: firstly, pre-training using navigation data from clean water areas or open water areas to enable the model to learn the basic hydrodynamics and inertial laws under ice-free resistance; then, transferring the pre-trained model to the ice-covered target domain, freezing at least a portion of the encoder parameters, and fine-tuning only one or more of the prediction mapping head, environmental gating network, or ice-covered collision dynamics expert.

10. A transient dynamics prediction system for ships in ice-covered areas, characterized in that, include: The data acquisition and preprocessing module is used to acquire and synchronize historical AIS navigation data and IMU data; The frequency domain decoupling module is used to perform discrete wavelet transform on IMU data and output low-frequency approximate components and high-frequency detail components; The ice-cold energy extraction module is used to calculate transient ice-cold energy characteristics based on high-frequency detail components; The environmental gating and hybrid expert module is used to generate environmental gating weights based on transient ice-induced energy characteristics and adaptively fuse water zone dynamics experts and ice zone collision dynamics experts. The Transformer prediction module is used to output the future motion state or trajectory of a ship based on AIS features, IMU features, and transient ice shock energy features. The trajectory and risk output module is used to output prediction results and ice skating side-slip risk warnings.

Images