Deep learning-based wave force sensing modeling ship control motion forecasting method

Abstract

The invention discloses a deep learning-based wave force sensing modeling ship control motion forecasting method, which comprises the following steps of S1, pre-data processing: cleaning and standardizing original data; s2, secondary interpolation: estimating an intermediate value through the secondary interpolation step so as to better represent the continuous effect of the wave force and improve the time resolution of input data; and S3, LSTM network prediction: processing refined data to predict ship maneuvering motion, utilizing the advantage of the LSTM network in the aspect of capturing time dependence and having robustness for interpolation input, providing a robust mathematical model to predict ship maneuvering characteristics under wave interference under deep water and shallow water conditions, and improving the ship maneuvering performance. By adopting an improved LSTM model of three times of splicing interpolation, the method proves reliable prediction capability in the aspects of eliminating wave height, direction and current speed, and the method has relatively high accuracy and can improve navigation safety and reduce economic risks.

Description

Technical Field

[0001] The present invention relates to the technical field of ship maneuvering motion prediction, and in particular to a ship maneuvering motion prediction method based on deep learning and wave force perception modeling. Background Art

[0002] Prediction of ship maneuvering motions is a key aspect of marine navigation and involves the study of the motion of a ship under various forces and environmental conditions. Traditional approaches rely heavily on mathematical models and hydrodynamic theory to predict ship behavior, taking into account factors such as hull shape, rudder dynamics, and propulsion dynamics. These models often incorporate complex differential equations to simulate the motion of a ship in still water and riverbeds. However, these traditional approaches often fall short when dealing with the nonlinear and random elements inherent in the real marine environment. Accurate prediction of ship maneuvering motions is essential to enhance navigation safety, optimize route planning, and improve over-fidelity of marine operations.

[0003] With the development of big data and advanced computing power, deep learning techniques have become a powerful tool for predicting ship maneuvers. These methods can process large amounts of data and capture complex patterns that traditional models may miss. However, the implementation of deep learning in this field is not without challenges. One of the main issues is the lack of high-quality labeled data required to effectively train these models. In addition, the inherent complexity of ship dynamics, which is affected by different coasts, wind conditions, and wave forces, poses a significant obstacle. Current deep learning models often have difficulty generalizing to different environmental conditions, which leads to predictions that may be unreliable in practice.

[0004] Several deep learning approaches have been proposed to address the prediction of ship maneuvering motions. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs), including long short-term memory (LSTM) networks, are among the most commonly used architectures. CNNs have been applied to extract spatial features from ship motion data, while CNNs, especially LSTMs, are highly valued for their ability to model temporal dependencies. Despite these advances, significant challenges remain. These include the sensitivity of the models to noise in the data, the difficulty in capturing long-range dependencies, and the challenge of integrating different data sources (e.g., weather conditions, ocean state). In addition, most existing models do not fully consider the nonlinear effects of wave forces, which are critical for accurate maneuvering predictions.

[0005] To address these challenges, a new approach is proposed for ship maneuvering motion prediction using a quadratic interpolation-enhanced LSTM model that explicitly considers the effects of wave forces. This method improves existing methods by adopting a quadratic interpolation step, which improves the model's ability to capture the nonlinear and dynamic characteristics of wave-induced forces in ship motion. Summary of the invention

[0006] In view of the defects of the above-mentioned prior art, the object of the present invention is to provide a method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning.

[0007] To achieve the above object, the present invention provides a method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning, comprising the following steps:

[0008] S1, pre-data processing, cleaning and standardizing the raw data;

[0009] S2, quadratic interpolation, estimates intermediate values ​​through quadratic interpolation steps to better represent the continuous effects of wave forces and improve the temporal resolution of the input data;

[0010] S3, LSTM network prediction, processes the refined data to predict the ship maneuvering motion, taking advantage of its strengths in capturing temporal dependencies while being robust to interpolated inputs.

[0011] Furthermore, the LSTM network includes the following components:

[0012] The forget gate is used to decide to discard information from the cell state;

[0013] The input gate, which determines the new information that is added to the cell state;

[0014] Update the cell state to combine the previous cell state with the new information;

[0015] Output gate, used to determine the output of the unit state.

[0016] Furthermore, the prediction method further comprises:

[0017] Batch Normalization and Dropout layers to enhance stability and prevent overfitting;

[0018] Activation function layer, used for nonlinear processing;

[0019] Linear layer, used for regression analysis.

[0020] Furthermore, it includes computer program instructions and a storage medium, wherein the storage medium stores a computer program, and the instructions are steps for a computer to execute the above method.

[0021] Further, the pre-data processing step includes extracting key data, including latitude, longitude, ground speed, engine speed, engine angle, wind speed and wind direction.

[0022] Furthermore, the quadratic interpolation step includes cubic spline interpolation, which is used to process discontinuous flight path data and ensure the continuity and accuracy of the data.

[0023] Furthermore, accurate prediction and control of ship maneuvers is based on various maneuvering motion models, including the following models:

[0024] Modeling of ship maneuvers under wave forces: Sine wave motion model, mathematically expressed as: y = Asin (kx_ωt + φ);

[0025] Where y represents the amplitude, A represents the amplitude, k represents the wave number, represents the initial position of the wave, ω represents the frequency, t represents the time, and φ represents the phase;

[0026] The mathematical modeling of vascular motion under the action of wave-induced force is expressed mathematically as follows: ε(x, y, t) = acos[kxcosx_kysinx_(ω_Vkcosx)t];

[0027] Where x represents the wave direction angle, y represents the wave amplitude, k represents the wave number, ω represents the frequency, and t represents the time;

[0028] Mathematical modeling of wave-induced moments and moments: For deep-water regular waves, the dynamic pressure function is, Δp(ξ,ζ,t)=_ρgae -kξ cos(kζ_ωt);

[0029] Mathematical model of random waves: The variance and energy expressions of long-peak random waves are:

[0030] Furthermore, the pre-data processing step uses min-max normalization to scale the data to a fixed range, and the method also includes using cubic spline interpolation to process discontinuous flight path data, and applying a moving average with a window size of 9 to address data latency issues.

[0031] Furthermore, the LSTM network prediction step uses mean squared error (MSE) loss and adds a penalty for negative values ​​in the left and right engine speeds to ensure accurate predictions.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The prediction method of the present invention improves the accuracy of ship motion prediction under different wave conditions and enhances the model to better capture the nonlinear effects of wave forces, which are usually not adequately addressed by existing methods.

[0034] 2. The interpolation technique of the present invention improves the temporal resolution of the input data, enabling the LSTM network to learn more detailed and subtle patterns in ship motion. This improvement leads to more accurate predictions, especially in dynamic and rapidly changing marine environments.

[0035] 3. The prediction method of the present invention ensures robustness to the variability and noise present in real-world maritime data. By preprocessing and normalizing the data and using advanced interpolation techniques, the model maintains high performance under different environmental conditions and operating scenarios.

[0036] 4. The present invention proposes a robust mathematical model to predict the maneuverability of ships under wave interference in deep and shallow water conditions. By adopting an improved LSTM model with cubic splicing interpolation, this method has demonstrated reliable prediction capabilities in terms of excluding wave height, direction and current speed. This method has high accuracy, can improve navigation safety and reduce economic risks. The adaptability and accuracy of the model contribute to its potential application in real-world ocean scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the scheme in the present invention, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 is a schematic diagram of the regular wave motion provided by the present invention;

[0039] Figure 2 is a graphical representation of the wave direction angle provided by the present invention;

[0040] Figure 3 It is an overview diagram of the pipeline provided by the present invention;

[0041] Figure 4 It is a detailed diagram of the pipeline provided by the present invention;

[0042] Figure 5 is a result graph of cubic spline interpolation provided by the present invention;

[0043] Figure 6 is a visualization diagram of the data set provided by the present invention;

[0044] Figure 7 It is a schematic diagram of the LSTM-Δs result provided by the present invention;

[0045] Figure 8 It is a schematic diagram of the LSTM-Δsc result provided by the present invention;

[0046] Fig. 9 It is a schematic diagram of the LSTM-Δsdc results provided by the present invention;

[0047] Fig.10 is a first ablation experiment result diagram of the network configuration provided by the present invention;

[0048] Fig.11 is a second ablation experiment result diagram of the network configuration provided by the present invention;

[0049] Fig.12 This is a third ablation experiment result diagram of the network configuration provided by the present invention. DETAILED DESCRIPTION

[0050] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more explicit definition of the protection scope of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0051] The terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present invention or the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order.

[0052] See also Figure 1-Figure 12 , a method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning, comprising the following steps:

[0053] S1, pre-data processing, cleaning and standardizing the raw data;

[0054] S2, quadratic interpolation, estimates intermediate values ​​through quadratic interpolation steps to better represent the continuous effects of wave forces and improve the temporal resolution of the input data;

[0055] S3, LSTM network prediction, processes the refined data to predict the ship maneuvering motion, taking advantage of its strengths in capturing temporal dependencies while being robust to interpolated inputs.

[0056] As an improvement of the above technical solution, the LSTM network includes the following components:

[0057] The forget gate is used to decide to discard information from the cell state;

[0058] The input gate, which determines the new information that is added to the cell state;

[0059] Update the cell state to combine the previous cell state with the new information;

[0060] Output gate, used to determine the output of the unit state;

[0061] The forget gate determines which information from the cell state should be discarded as follows: ft = (wf.[ht-1,xt]+bf);

[0062] Where is the sigmoid activation function, Wf is the weight matrix of the forget gate, [ht-1, xt] represents the concatenation of the previous hidden state ht-1 and the current input xt. bf is the bias term of the forget gate;

[0063] Input gate (it). The input gate determines how much new information will be added to the cell state, and C represents the candidate value to be added to the cell state; Ct = tanh (WC. [ht_1, xt] + bC);

[0064] Where Wi and WC are the weight matrices of the input gate and candidate cell state, respectively. bi and dbC are the main ranges of the input gate and candidate cell state. tanh is the hyperbolic tangent activation function.

[0065] Update cell state (Ct). Update the cell state by combining the previous cell state and the candidate cell state, modulated by the forget gate and the input gate: Ct = ft.Ct_1 + it.~Ct;

[0066] Where Ct_1 is the cell state at the previous time step. This is the updated state.

[0067] Output gate (ot). The output gate determines which part of the cell state will be output as the hidden state. The hidden state is calculated as: ot = (Wo.[ht_1,xt] + b; o) ht = ot.tanh(Ct);

[0068] Where is the weight matrix of the output gate. b represents the bias term of the output gate.

[0069] The network structure consistency of batch normalization LSTM++ dropout (0.5) + ReLU + linear layer. This design balances the complexity and generalization ability of the model. The LSTM layer accepts an input dimension [16, 8, 12], where 16 represents the ebatch size, 8 is the time step, and 12 is the input feature size. The output from the LSTM layer has the dimension [16, 1, 8], indicating 8 output features after processing the time step.

[0070] To enhance stability and prevent overfitting, batch normalization and Dropout (0.5) are applied after the LSTM layer. Batch normalization speeds up training and alleviates the gradient vanishing problem, while Dropout improves generalization ability by randomly disabling some neurons. An activation function is then used for nonlinear processing, followed by a linear layer for regression. The hidden layer size is set to 512 to ensure sufficient learning capacity. The network structure is designed to handle the time series nature and complexity of ship maneuvering reliability prediction, providing more accurate predictions.

[0071] As an improvement of the above technical solution, the prediction method further includes:

[0072] Batch Normalization and Dropout layers to enhance stability and prevent overfitting;

[0073] Activation function layer, used for nonlinear processing;

[0074] Linear layer, used for regression analysis.

[0075] As an improvement of the above technical solution, it includes computer program instructions and a storage medium, wherein the storage medium stores a computer program, and the instructions are steps for a computer to execute the above method.

[0076] As an improvement to the above technical solution, the pre-data processing step includes extracting key data, including latitude, longitude, ground speed, engine speed, engine angle, wind speed and wind direction.

[0077] It is not the key to predict and analyze the ship's maneuverability, but to accurately acquire and process dynamic operating parameters. This process involves extracting and reading key data from raw data sets, which provides a comprehensive view of the ship's state and its contact with the external environment.

[0078] Specifically, the following ten columns of data were extracted: latitude (lat), longitude (lng), ground speed, engine speed (sr, r / min), engine right angle (r, -180 to 180.), wind speed (sw, m / s) and wind direction (w, 0.-360.). (sog, kn), these parameters are recorded at one-second intervals, thus ensuring the accuracy of real-time data and the continuity of sub-results t. Analyzing each parameter plays a vital role in predicting the maneuverability of the ship. Latitude and longitude provide the geographic location of the ship, which is essential for tracking its heading and position changes. Ground speed reflects the instantaneous speed of the ship, while ground heading records the current heading of the ship. These data are essential for evaluating the maneuverability of the eship and predicting its future movement. The speed and angle of the left and right engines directly affect the propulsion and steering capabilities of the ship. The engine speed determines the propulsion speed of the ship, and the angle affects its steering force and direction, so accurate monitoring of these data is essential for stabilizing ship control. Wind speed and wind direction describe external environmental conditions. Changes in these factors directly affect the navigation safety and maneuvering difficulty of the ship, especially in complex weather conditions. Therefore, real-time collection and analysis of these data are essential for optimizing the ship's route and maneuvering strategy. In the analysis of intelligent ship maneuverability, the accurate calculation of key parameters is the key to ensuring stability and safety. This section details the process of calculating ten basic parameters based on the original data: latitude difference (altitude), longitude difference (range), ground speed (range), arc (radius), ground speed, power (main 1 speed cmd), left engine angle in arc (main 1 angle), right engine power (main 2 speed cmd), right engine arc angle (main 2 angle), wind speed (wind speed), wind direction in radius (wind direction angle). These calculations provide a comprehensive understanding of its opera status and external environmental conditions, and the latitude and longitude differences (length) measure its displacement, which is essential for tracking its geographical location. The ground speed (sog) represents the average speed of the hip, which is a key indicator of propulsion efficiency. The radian course on the ground (gear-rad) is derived from the angle of the course and is used to evaluate the actual heading of the course. The left and right engine power (main 1 speed cmd, main 2 speed cmd) reflects the output power of each engine, which directly affects the propulsion and speed. The radian angles of the left and right engines (main 1 angle, main 2 angle) are converted according to their respective angle data to evaluate the rotational ability of the engines. Environmental factors such as wind speed (wind-speed) and wind direction (in radians) (wind direction angle-rad) quantify the impact of wind conditions on ship stability and response to external forces. The calculation and analysis of these parameters are essential for optimizing ship maneuverability and developing predictive models.

[0079] As an improvement of the above technical solution, the quadratic interpolation step includes cubic spline interpolation, which is used to process discontinuous flight path data and ensure the continuity and accuracy of the data.

[0080] As an improvement of the above technical solution, accurate prediction and control of ship maneuvers are basically based on various maneuvering motion models, including the following models:

[0081] Modeling of ship maneuvers under wave forces: Sine wave motion model, mathematically expressed as: y = Asin (kx_ωt + φ);

[0082] Where y represents the amplitude, A represents the amplitude, k represents the wave number, represents the initial position of the wave, ω represents the frequency, t represents the time, and φ represents the phase;

[0083] Small amplitude waves can be decomposed into simple harmonic motions in the horizontal and vertical directions, allowing a mathematical description of the wave propagation and evolution. The key assumptions of small amplitude wave theory are: (1) the fluid is inviscid and incompressible, with no internal friction or deformation; (2) the wave motion is described by a potential function, the flow field does not rotate, i.e. the angular velocity of the fluid particles is zero; (3) the wave height is small relative to the wavelength, defining a minimum wave amplitude, e.g. Figure 1 As shown, regular waves propagate in the ξ direction in an inertial reference frame Oξηζ fixed to the surface of the Earth, perpendicular to the still water surface. For an inviscid, incompressible fluid, the kinetic energy and pressure depend only on the velocity and the square of the fluid density, and have nothing to do with the state or shape of the fluid.

[0084] The mathematical modeling of vascular motion under the action of wave-induced force is expressed mathematically as follows: ε(x, y, t) = acos[kxcosx_kysinx_(ω_Vkcosx)t];

[0085] Where x is the wave direction angle, y is the wave amplitude, k is the wave number, ω is the frequency, and t is the time; it is defined as the angle between the wave propagation direction relative to the stern and the ship's forward direction. This angle is usually measured counterclockwise and the positive range is as follows: Figure 2 As shown in Figure 1, the interaction between a ship and water waves can be divided into five cases according to the different wave direction angles: follow-the-wave, the wave is consistent with the forward motion of the ship; stern oblique wave, the wave propagates obliquely from the stern to the bow; beam sea, the wave is perpendicular to the ship and produces a lateral impact; bow oblique wave, the wave propagates obliquely from the bow to the stern; and head sea, the wave is generally opposite to the forward direction of the ship.

[0086] Mathematical modeling of wave-induced moments and moments: For deep-water regular waves, the dynamic pressure function is, Δp(ξ,ζ,t)=_ρgae -kξ cos(kζ_ωt);

[0087] Mathematical model of random waves: The variance and energy expressions of long-peak random waves are:

[0088] As an improvement to the above technical solution, the pre-data processing step uses min-max normalization to scale the data to a fixed range, a process that ensures that features with different units or scales contribute equally to the analysis or machine learning model. Normalization helps improve the performance and convergence speed of the algorithm by avoiding bias caused by features of different sizes.

[0089] Two common data normalization methods include min-max normalization and Z-score normalization. Using min-max normalization, the data is scaled to a fixed range, usually [0, 1], using the following formula:

[0090] Where X is the original value, Xmin is the minimum value, and Xmax is the maximum value of the feature. To this end, the above-mentioned normalization method is applied to all data inputs in the network. This ensures that each feature is scaled to a consistent degree, which helps improve the performance and stability of the model. By normalizing the data, potential problems associated with different feature sizes and units are alleviated, resulting in more efficient training and better model convergence. The method also includes using cubic spline interpolation to handle discontinuous flight path data, and applying a moving average with a window size of 9 to address the delay problem of the data. Since the dataset includes multiple discontinuous flight paths, there may be cases where the latter part of one flight path is connected to the beginning of the next flight path. To address this issue, the boundary points between adjacent flight paths are identified by checking the values ​​of "lat-sub" and "lng-sub". Specifically, if the absolute value of "lat-sub" and "lng-sub" exceeds 1000 (normal values ​​are usually below 100), it is considered to be a boundary point. At these points, the "lat-sub" and "lng-sub" values ​​are replaced with the values ​​of subsequent time steps. The method ensures the continuity and accuracy of the data by effectively handling the transition between flight paths. In order to solve the delay problem of "lat-sub" and "lng-sub" data, a moving average with a window size of 9 is applied. The method involves calculating the average of the current value and the values ​​of the previous 8 time steps. This method improves the stability and accuracy of the data by smoothing fluctuations and reducing noise. The specific calculation process is as follows: For each data point, the average is calculated by adding the values ​​within the 9 time step window and dividing by 9, thereby mitigating the impact of transient anomalies.

[0091] To refine the "Ling Sub Smooth" and "Lng Sub Smooth" data, cubic spline interpolation is performed using known points spaced at 120 second intervals. This process generates the interpolated "Sub Interpolation" and "Liquify Interpolation". Cubic spline interpolation is used here because it is able to provide a smooth and continuous curve that accurately represents the data trend between the known points. By applying this technique, it is ensured that the interpolation maintains the consistency and consistency of the original data set while filling the interpolation between the known points.

[0092] As an improvement to the above technical solution, the LSTM network prediction step uses mean square error (MSE) loss and adds a penalty for negative values ​​in the left and right engine speeds to ensure accurate predictions. The optimizer is optimized using mean square error (MSE) loss and a penalty for negative values ​​in the left and right engine speeds to ensure accurate predictions.

[0093] experiment

[0094] Dataset

[0095] The ship's turning data (Rotation.xlsx) includes 1202 entries. Thhhiy's zigzag data (z.xlsx) also includes 1202 entries. The data for heavy wind navigation in port includes 7 routes, totaling 18,308 entries. The heavy wind navigation data outside the port includes one route, totaling 6,602 entries. The Train.xlsx file consists of heavy wind navigation data from inside and outside the port (567 routes), rotation data, and zigzag data, totaling 14,347 entries. The Val.xlsx file consists of 24 routes from heavy wind navigation in port, totaling 6,963 entries.

[0096] set up

[0097] For initial training, a dataset of ship turning maneuvers with high maneuvering volume and low wave vectors was used. The dataset spanned 1200 seconds with a learning rate (lr) of 0.001 and was trained for 50 epochs. This approach was chosen to capture effective dynamics under stable conditions. Subsequently, a dataset with low maneuvering volume and high wave vector characteristics was used for retraining, reflecting more realistic and challenging operational scenarios. The dataset was extended over 15,000 seconds with a learning rate of less than 0.00001 and a large number of learners, set to 300. This retraining phase was crucial to adapt the model to handle various and less predictable conditions. The training was performed using Python 3.8 and an RTX 3090 GPU, which allows for efficient computation and processing. The entire training process for both stages takes approximately 20 minutes, ensuring timely and effective model development. This comprehensive approach helps to improve the model to perform well in different operational scenarios,

[0098] Enhance its robustness and accuracy.

[0099] Evaluation Metrics

[0100] In trajectory prediction, the final displacement error (FDE) and the average displacement error (ADE) are usually used to evaluate the prediction quality. To evaluate early performance, the average displacement error of the first 10 moments (ADE-10) is also used. Given that most predictions cover 100 seconds, while our predictions extend longer, ADE-100 is also calculated for a more thorough evaluation.

[0101] Comparison with the State-of-the-Art

[0102] Two different network configurations are available for comparison:

[0103] LSTM-Δs: This configuration uses the LSTM model to train and predict the changes in longitude and latitude, and then calculates the latitude and longitude from a time point to this time. The validation set is a dataset of large wind power sailing routes in the port, with an average rate of return (lr) of 0.00001.

[0104] LSTM-Δsc: This configuration extends LSTM-Δs by incorporating cubic spline interpolation to predict both latitudinal and longitudinal ranges. The validation set is data from large wind farm routes within Port 1, and the learning rate (lr) is 0.00001.

[0105] LSTM-Δsdc (method): This method adopts the two-step training method used in LSTM-Δsd and adds cubic spline interpolation to predict changes in latitude and longitude. The initial training set is the turning maneuver dataset with a learning rate of 0.001. The second stage training phase uses the large wind direction route dataset from Port Test 2-Test 7 and is validated on the turning maneuver dataset with an LR of 0.00001.

[0106] Experimental Results

[0107] The experimental results show that although LSTM-Δs has a certain degree of inefficiency in predicting the actual ship's navigation trajectory, its performance shows significant variability in different situations. For example, it achieved satisfactory fitting accuracy in the test 1 and test 7 scenarios, but showed a large deviation in the test 4 scenario. Similarly, although LSTM-Δsc showed obvious improvements and LSTM-Δs, it still failed to achieve accurate prediction in the test 4 scenario. In contrast, the method of always achieves trajectory fitting for all scenarios, especially in test 4, thus verifying the robustness and advanced capabilities of the method. In order to provide a more detailed analysis, the final displacement error (FDE) and average displacement error (ADE) indicators are used to provide quantitative results, as shown in Table XX. The results show that the method of significantly outperforms the comparison methods on all indicators, further verifying the superiority and advanced performance of the method of.

[0108] Ablation studies

[0109] In order to verify the influence of wave force model and interpolation method on the experimental prediction of maneuverability, corresponding ablation experiments were carried out. The experimental objects are as follows:

[0110] Baseline: This configuration uses an LSTM model to train the latitude and length data in a single step and then makes multi-step predictions, eliminating the wave force model and interpolation methods.

[0111] LSTM-ΔSDC+w / o. Wave force model: This model adopts two-step training using LSTM, which just removes the wave force model from the full method. LSTM-ΔSDC (full model): This method adopts the two-step training method used in LSTM-ΔSD, adopts cubic spline interpolation method to predict longitude and latitude changes, and adopts a baseline-based wave force model.

[0112] Figure 10-12 : The results of ablation experiments for different network configurations are shown, and it is clear that removing the wave force model or the interpolation method leads to a significant drop in the AC curves of trajectory prediction. These experiments clearly verify the robustness and effectiveness of the proposed method.

[0113] in conclusion

[0114] This paper proposes a robust mathematical model to predict the maneuverability of ships under wave disturbance in deep and shallow water conditions. By adopting an improved LSTM model with cubic splicing interpolation, the method demonstrates reliable prediction capabilities in terms of excluding wave height, direction, and current speed. Experimental results show that the proposed method has high accuracy, can improve navigation safety, and reduce economic risks. The adaptability and accuracy of the model contribute to its potential application in real-world marine scenarios.

[0115] The above description is only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solution described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. Any equivalent structure or equivalent process transformation made using the contents of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning, characterized in that: The following steps are involved: S1, pre-data processing, cleaning and standardizing the raw data; S2, quadratic interpolation, estimates intermediate values ​​through quadratic interpolation steps to better represent the continuous effects of wave forces and improve the temporal resolution of the input data; S3, LSTM network prediction, processes the refined data to predict the ship maneuvering motion, taking advantage of its strengths in capturing temporal dependencies while being robust to interpolated inputs.

2. According to claim 1, a method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning is characterized in that: The LSTM network consists of the following components: The forget gate is used to decide to discard information from the cell state; The input gate, which determines the new information that is added to the cell state; Update the cell state to combine the previous cell state with the new information; Output gate, used to determine the output of the unit state.

3. According to a method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning in claim 1, it is characterized in that: The prediction method further includes: Batch Normalization and Dropout layers to enhance stability and prevent overfitting; Activation function layer, used for nonlinear processing; Linear layer, used for regression analysis.

4. A computer program product for executing the method according to any one of claims 1 to 3, characterized in that: Features: It includes computer program instructions and storage media, the storage medium stores a computer program, and the instructions are steps for a computer to execute the above method.

5. The method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning according to claim 4, characterized in that: The pre-data processing step includes extracting key data including latitude, longitude, ground speed, engine speed, engine angle, wind speed and wind direction.

6. A method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning according to claim 5, characterized in that: The quadratic interpolation step includes cubic spline interpolation, which is used to process discontinuous flight path data and ensure the continuity and accuracy of the data.

7. A method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning according to claim 6, characterized in that: Accurate prediction and control of ship maneuvers is based on various maneuvering motion models, including the following models: Modeling of ship maneuvers under wave forces: Sine wave motion model, mathematically expressed as: y = Asin (kx-ωt+φ); Where y represents the amplitude, A represents the amplitude, k represents the wave number, represents the initial position of the wave, ω represents the frequency, t represents the time, and φ represents the phase; Mathematical modeling of vascular motion under the action of wave-induced force is expressed mathematically as follows: ε(x, y, t) = acos[kxcosx-kysinx-(ω-Vkcosx)t]; Where x represents the wave direction angle, y represents the wave amplitude, k represents the wave number, ω represents the frequency, and t represents the time; Mathematical modeling of wave-induced moments and moments: For deep-water regular waves, the dynamic pressure function is, Δp(ξ,ζ,t)=-ρgae -kξ cos(kζ-ωt); Mathematical model of random waves: The variance and energy expressions of long-peak random waves are:

8. The method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning according to claim 7, characterized in that: The pre-data processing step uses min-max normalization to scale the data to a fixed range, uses cubic spline interpolation to handle discontinuous flight path data, and applies a moving average with a window size of 9 to account for data latency.

9. A method for predicting ship maneuvering motion based on wave force perception modeling based on deep learning according to claim 8, characterized in that: The LSTM network prediction step uses mean squared error (MSE) loss and adds a penalty for negative values ​​in the left and right engine speeds to ensure accurate predictions.

Images