Frequency domain analysis and deep learning-based ship navigation trajectory prediction method and system

Abstract

The invention discloses a ship navigation trajectory prediction method and system based on frequency domain analysis and deep learning, and relates to the technical field of ship navigation and maritime traffic management, and the method comprises the steps: collecting ship navigation AIS data in real time, and carrying out the preprocessing operation; segmenting the preprocessed time domain data by adopting a dynamic sliding window and converting the preprocessed time domain data into frequency domain representation, and updating the frequency component of the current window by adopting a sliding updating strategy so as to construct a frequency domain feature vector and realize real-time processing and noise suppression of high-frequency data; introducing a multi-modal gating attention fusion mechanism, fusing the frequency domain feature vector and the time domain transient feature vector, and dynamically balancing the weight of the frequency domain periodic feature and the weight of the time domain burst feature; and inputting the fusion feature vector into a deep learning prediction model, converting a frequency domain prediction result output by the model back to a time domain, and generating future position, navigational speed and course information of the ship. According to the method, the prediction precision and efficiency are greatly improved by combining the frequency domain analysis and the deep learning technology.

Description

Technical Field

[0001] The present invention relates to the technical field of ship navigation and maritime traffic management, and in particular to a method and system for predicting ship navigation trajectories based on frequency domain analysis and deep learning. Background Art

[0002] With the rapid development of global maritime trade, the volume of maritime traffic continues to increase, and the complexity of ship navigation and traffic management has also increased. Accurate track prediction is of great significance for avoiding collisions, optimizing route planning, and improving the efficiency of maritime traffic management. At present, ship trajectory prediction methods mainly rely on time series analysis technology based on the time domain. These methods usually use deep learning models to model time series features such as position, speed, and heading in the ship's automatic identification system (AIS) data. Through these models, it is possible to capture complex patterns in time series to a certain extent and predict the future trajectory of the ship. However, these methods have the following shortcomings when dealing with high-frequency noise and data non-stationarity:

[0003] Traditional modeling methods primarily focus on time-domain signal processing, failing to fully exploit the periodic characteristics of the data. Furthermore, the non-stationary nature of ship trajectory data makes it difficult for existing methods to effectively extract patterns, limiting their predictive capabilities. The presence of high-frequency noise in AIS data not only affects the model's ability to extract time series patterns but also increases the training complexity of deep learning models, leading to decreased prediction accuracy. Furthermore, as the complexity of AIS data increases, direct modeling of time-domain features can lead to time-consuming and unstable training processes, particularly under varying sea conditions and ship maneuvers, resulting in poor model robustness.

[0004] In recent years, although some studies have attempted to combine frequency domain analysis technology to improve the shortcomings of time domain modeling, most of these methods remain at the basic research level and have failed to be organically combined with deep learning technology, making it difficult to achieve high-precision prediction of complex ship trajectories. Summary of the Invention

[0005] This invention addresses the above-mentioned issues by combining frequency domain analysis with deep learning technology, proposing a method and system for ship trajectory prediction based on frequency domain analysis and deep learning, which significantly improves prediction accuracy, model efficiency, and adaptability. The technical solutions of this invention are as follows:

[0006] In a first aspect, the present application provides a method for predicting ship navigation trajectories based on frequency domain analysis and deep learning, comprising the following steps:

[0007] Real-time collection and pre-processing of ship navigation AIS data, including ship position, speed and heading time series data;

[0008] The pre-processed continuous time domain data is segmented using a dynamic sliding window and converted into a frequency domain representation, and the frequency component of the current window is updated using a sliding update strategy;

[0009] Select frequency components to construct frequency domain feature vectors;

[0010] A multimodal gated attention fusion mechanism is introduced to fuse the frequency domain feature vector and the time domain transient feature vector;

[0011] The fused feature vector is input into the deep learning prediction model, and the frequency domain prediction results output by the model are converted back to the time domain representation to generate the ship's future position, speed and heading information.

[0012] A further technical solution is to pre-process the ship navigation AIS data, including:

[0013] Convert the raw AIS data into a unified format, including timestamp, vessel unique identifier, position, speed, and heading fields;

[0014] Use the 3σ principle to identify and remove speeds and headings outside the specified range, and use a low-pass filter to filter out high-frequency noise;

[0015] For heading data with jumps after filtering, linear interpolation is used to smooth the transition.

[0016] A further technical solution is to segment the pre-processed continuous time domain data using a dynamic sliding window, including:

[0017] Based on the pre-processed speed and heading time domain data, the heading change rate and acceleration of each time stamp are obtained;

[0018] When the heading change rate is greater than the set heading threshold, or the acceleration is greater than the set acceleration threshold, the sliding window switches from the default window length to the minimum window length.

[0019] A further technical solution is to use a sliding update strategy to update the frequency component of the current window. The expression is:

[0020]

[0021] Among them, X old [k] is the frequency component corresponding to the initial window, X new [k] is the frequency component corresponding to the current window, x new 、x old are the newly added data points and the removed data points after the window slides, respectively. k represents the kth frequency component, and N represents the current window length, which represents the number of data points in this frequency domain transformation.

[0022] A further technical solution is to select frequency components to construct a frequency domain feature vector, including:

[0023] Calculate the amplitude spectrum of the frequency component corresponding to each window;

[0024] Calculate the cumulative energy based on the amplitude spectrum of each frequency component and select the frequency component whose cumulative energy exceeds the set percentage;

[0025] The amplitude of the selected frequency component is normalized to form a frequency domain feature vector.

[0026] Its further technical solution is to introduce a multimodal gated attention fusion mechanism to fuse the frequency domain feature vector and the time domain transient feature vector, including:

[0027] Based on the time domain data of speed and heading in the current window, a time domain transient feature vector is constructed;

[0028] The frequency domain feature vector and time domain transient feature vector are input into the fully connected layer for feature mapping;

[0029] The mapping features of the frequency domain and time domain are spliced together, and the dynamic weights are calculated through the gated attention mechanism;

[0030] The mapping features of frequency domain and time domain are fused based on dynamic weights.

[0031] A further technical solution is to construct a time domain transient feature vector based on the speed and heading time domain data of the current window, including:

[0032] Based on the time domain data of speed and heading in the current window, the heading change rate, acceleration, speed skewness and kurtosis of the current window are obtained to form the time domain transient feature vector.

[0033] Its further technical solution is that the deep learning prediction model is obtained based on model training of the long short-term memory network LSTM, where the LSTM is a single-layer structure or a double-layer structure.

[0034] Its further technical solution is that the method further comprises:

[0035] Load the physical parameters of interest according to the ship type and perform normalization processing;

[0036] The physical parameter eigenvector is concatenated with the frequency domain eigenvector to construct an enhanced frequency domain eigenvector;

[0037] A multimodal gated attention fusion mechanism is introduced to fuse the enhanced frequency domain feature vector and time domain transient feature vector.

[0038] Secondly, the present application also provides a ship navigation trajectory prediction system based on frequency domain analysis and deep learning, including:

[0039] The data processing module is used to collect and pre-process the ship's navigation AIS data in real time. The AIS data includes time series data of the ship's position, speed and heading;

[0040] The frequency domain conversion module is used to segment the pre-processed continuous time domain data using a dynamic sliding window and convert it into a frequency domain representation, and update the frequency component of the current window using a sliding update strategy;

[0041] Feature extraction module, used to select frequency components to construct frequency domain feature vectors;

[0042] Feature fusion module, which is used to introduce a multimodal gated attention fusion mechanism to fuse frequency domain feature vectors and time domain transient feature vectors;

[0043] The trajectory prediction module is used to input the fused feature vector into the deep learning prediction model, convert the frequency domain prediction results output by the model back to the time domain representation, and generate the future position, speed and heading information of the ship.

[0044] Compared with the existing technology, the present invention has the following significant beneficial effects by combining frequency domain analysis with deep learning models:

[0045] 1) Significantly Improved Prediction Accuracy: By online updating of frequency-domain features within a dynamic sliding window, and by effectively extracting periodicity and trend information from ship tracks through frequency-domain analysis, this approach enables real-time processing and noise suppression of high-frequency AIS data. A multimodal gated attention fusion mechanism (MTF-AGU) is introduced to dynamically balance the weights of frequency-domain periodicity and time-domain burst features. The resulting deep learning prediction model captures long- and short-term dependencies within time series, enhancing understanding and prediction capabilities for complex track patterns.

[0046] 2) Accelerate model convergence: Frequency domain feature enhancement and optimization processing simplifies the complexity of input data, making it easier for deep learning models to learn effective features during training, shortening model training time and optimizing prediction computational efficiency, making it suitable for real-time track prediction needs.

[0047] 3) Enhanced model robustness: Frequency domain processing effectively addresses data non-stationarity, ensuring the model maintains high predictive performance under varying sea conditions and vessel behavior. De-noising and feature extraction reduce random fluctuations in the data and improve the model's ability to withstand anomalous data.

[0048] 4) Improving model stability and reliability: Frequency domain analysis reduces the volatility of input data by extracting key features, making the deep learning prediction model more stable under varying data conditions. Combining frequency domain analysis with the multi-level feature processing mechanism of the deep learning model improves the reliability of the overall prediction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the ship navigation trajectory prediction method based on frequency domain analysis and deep learning provided by this application.

[0050] Figure 2 This is a schematic diagram of the composition of the ship navigation trajectory prediction system based on frequency domain analysis and deep learning provided by this application. DETAILED DESCRIPTION

[0051] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0052] In one embodiment of the present application, a method for predicting ship navigation trajectory based on frequency domain analysis and deep learning is provided. Figure 1 As shown, the method specifically includes the following steps:

[0053] Step 1: Collect ship navigation AIS data in real time and perform pre-processing operations.

[0054] In this embodiment, the AIS data includes time series data such as ship position (latitude and longitude), speed and heading, and the sampling frequency is set to 1Hz. The following preprocessing operations are performed on these time series data: the original AIS data is converted into a unified format (such as Comma-Separated Values, CSV), which contains timestamp, ship unique identifier MMSI, position, speed, and heading fields. Then the 3σ principle is used to identify and remove speeds (such as >20 knots) and headings (such as ±360° jumps) that exceed the specified range. A low-pass filter is used to filter out high-frequency noise, retain ship maneuvering related signals, and ensure the stability of the data. Finally, for the heading data that has jumps after filtering (such as 359°→0°), linear interpolation is used for smooth transition to reduce random fluctuations in the data.

[0055] Optionally, the low-pass filter is a Butterworth low-pass filter (cut-off frequency 0.5 Hz).

[0056] Step 2: Segment the preprocessed continuous time domain data using a dynamic sliding window.

[0057] In this embodiment, dynamic sliding window segmentation refers to the default window length N default and the minimum window length N minThe switching is performed according to the preset dynamic adjustment rules, and the sliding step is set to 1 / 4 of the window length (for example, if the window length is 64, the step is 16) to ensure data continuity. In specific applications, based on the pre-processed speed and heading time domain data, the heading change rate Δθ and acceleration |a of each timestamp are obtained. t |, and usually uses the default window length N default The pre-processed continuous time domain data is segmented, and the dynamic adjustment rule is: when Δθ is greater than the set heading threshold (for example, Δθ>15° / min), or |a t | is greater than the set acceleration threshold (for example |a t |>0.2m / s 2 ), the sliding window length changes from N default Switch to N min .

[0058] It should be noted that the default window length N default and the minimum window length N min (ie, the number of data points in the window) is usually a power of 2 (such as 256, 512, 1024, etc.) to improve the efficiency of subsequent frequency domain conversion calculations. default =256 (covering approximately 4.27 minutes of data), N min = 64 (covering approximately 1.07 minutes of data). The dynamic sliding strategy can switch to a small window length when the time domain data changes drastically or suddenly, thereby capturing the data details within the window.

[0059] Step 3: Perform frequency domain conversion and feature extraction on the segmented time domain data, which specifically includes the following steps:

[0060] Step 3.1: Convert the time domain data covered by the initial window into a frequency domain representation, and use a sliding update strategy to update the frequency components corresponding to the sliding window. In this embodiment, a frequency analysis algorithm such as Fast Fourier Transform (FFT) or wavelet analysis is used on the time domain data covered by the initial window to obtain the frequency domain representation X old [k] FFT can effectively separate the periodic components and random noise in the signal and extract the main frequency characteristics of the data. Using the sliding update strategy, when the window slides, the new data point x new and remove the data point x old When , the frequency component of the current window is updated as:

[0061]

[0062] Among them, k represents the kth frequency component, N represents the current window length, and represents the number of data points of this frequency domain transformation. In this example, N = N default Or N=N min .

[0063] Step 3.2: Select frequency components to construct frequency domain feature vectors, specifically including: first extract frequency component features, including main frequency components, amplitude spectrum, etc., as input for subsequent models. The amplitude spectrum of the frequency component corresponding to each window is calculated as:

[0064]

[0065] Where Re(X[k]) represents the real part of the kth frequency component, Im(X[k]) represents the imaginary part of the kth frequency component, and |X[k]| represents the amplitude of the kth frequency component.

[0066] Then the cumulative energy E_total = Σ|X[k]| is calculated based on the amplitude spectrum of each frequency component. 2 , and select the frequency components whose cumulative energy E_total exceeds the set proportion (for example, 90%). The amplitude of the selected frequency components is normalized and scaled to between 0 and 1 to ensure that the amplitudes of different frequency components are at the same level, thereby forming the frequency domain feature vector F freq =[|X(1)|,|X(2)|,…,|X(k)|].

[0067] In this embodiment, in order to construct a joint representation of ship dynamics enhancement, physical parameters such as ship inertia moment are embedded into the frequency domain feature vector F freq Then step 3.2 also includes: loading the physical parameters of interest (such as moment of inertia I) according to the ship type (MMSI associated database), and performing normalization processing to obtain the physical parameter feature vector:

[0068]

[0069] Then the physical parameter eigenvector F physics and the frequency domain eigenvector F freq Splicing is performed to construct the enhanced frequency domain feature vector: F enhanced =[F freq ; F physics ], enhancing the representation of frequency domain features can improve the predictive ability of the model.

[0070] Step 4: Introduce the multimodal gated attention fusion mechanism (MTF-AGU) to fuse the enhanced frequency domain feature vector and the time domain transient feature vector. The specific steps include the following:

[0071] Step 4.1: Based on the speed and heading time domain data of the current window, the following can be performed: (1) First-order derivative calculation of the heading change rate of the current window Δθ = θ t -θ t-1 and acceleration a t=v t -v t-1 (2) High-order statistics are used to calculate the skewness and kurtosis of the current window to capture the sudden behavior pattern. Based on the calculated parameters, the time domain transient feature vector is constructed as: F time =[Δθ,a t ,Skewness,Kurtosis], then F time and F enhanced This is then fed into the MTF-AGU pipeline, which includes the feature mapping layer, attention weight generation, and feature fusion. Note that the calculations for speed skewness and kurtosis can be referenced from existing formulas and will not be repeated here.

[0072] Step 4.2: The frequency domain feature vector F enhanced and the time domain transient eigenvector F time Input the fully connected layer for feature mapping. In this example, F enhanced For example, the mapping characteristics of frequency domain and time domain are expressed as:

[0073] H freq =RELU(W f ·F enhanced +b f )

[0074] H time =RELU(W t ·F time +b t )

[0075] Among them, RELU(·) represents the activation function, W f and b f represents the frequency domain learning weights and biases in the fully connected layer, W t and b t Represents the temporal learning weights and biases in the fully connected layer.

[0076] Step 4.3: Combine the mapping features of frequency domain and time domain [H freq ;H time ] and calculate the dynamic weights through the gated attention mechanism:

[0077] g=σ(W g ·[H freq ;H time ]+b g )

[0078] Where σ(·) represents the Sigmoid function, W g and b gRepresents the learned weights and biases of the gated attention module.

[0079] Step 4.4: Fusion of frequency domain and time domain mapping features based on dynamic weight g:

[0080] F fused =g·H freq +(1-g)·H time

[0081] Step 5: The fused feature vector F fused The frequency domain prediction results output by the model are converted back to the time domain representation to generate the ship's future position, speed and heading information.

[0082] In this embodiment, the deep learning prediction model is obtained by training the long short-term memory network LSTM. The LSTM model can be used to predict the fused feature vector F. fused Modeling is performed to predict future trajectory characteristics. The LSTM model can be a single-layer lightweight structure or a high-precision two-layer structure. Bayesian optimization and other methods can be used to optimize the structure and parameters of the deep learning model, including the number of network layers, number of hidden units, and regularization parameters, based on actual needs. The LSTM model's loss function is preferably the mean squared error (MSE), which calculates the difference between the predicted value and the true value. The model training strategy is designed as follows: the Adam optimizer is used, with a learning rate of 0.0001 and a batch size of 128. An early stopping mechanism is implemented: if the validation loss does not decrease (change < 1e-5) for 10 consecutive rounds, training is terminated early. Regularization: A dropout rate of 0.2 and an L2 regularization coefficient of 1e-4 are used.

[0083] The frequency domain prediction results output by the model can be converted back to the time domain through inverse fast Fourier transform (IFFT), and the time domain prediction results can be displayed in a graphical manner. For example, an interface can be built through PyQt to display the comparison between the predicted trajectory and the actual trajectory in real time, which can be used as a reference for navigation and maritime traffic management systems to assist in decision-making and planning.

[0084] The following is a supplementary explanation of the solution of the embodiment of the present application with reference to specific examples:

[0085] Example 1: Prediction of the speed of a 100,000-ton oil tanker entering and leaving a port

[0086] Step 1: Collect tanker navigation AIS data in real time and perform pre-processing operations.

[0087] Data source: AIS dataset of a certain port (10 hours of data, sampling frequency 1 Hz). Speed time series data is obtained from the ship automatic identification system (AIS). A total of 10 hours of data is used for modeling and testing.

[0088] Cleaning rules: Remove abnormal data points with speed greater than 25 knots (the typical speed range of oil tankers is 0-20 knots); heading jumps (such as 359°→0°) are corrected by linear interpolation.

[0089] Step 2: Segment the preprocessed continuous speed time domain data using a dynamic sliding window.

[0090] Set the default window length N default =256, if the heading change rate Δθ>15° / min or the acceleration|a t |>0.2m / s 2 , then switch to the minimum window length N min =64.

[0091] Step 3: Perform frequency domain conversion and feature extraction on the segmented speed time domain data.

[0092] After applying FFT transformation to the time domain data covered by the initial window, the corresponding formula of the sliding update strategy is used to update the frequency component X corresponding to the sliding window new [k] Calculate the amplitude spectrum of the frequency components corresponding to each window, and then obtain the cumulative energy E_total, and select the frequency components with E_total>90%. Normalize the amplitude of the selected frequency components to form the frequency domain feature vector F freq The moment of inertia of the loaded tanker I is normalized to F freq The enhanced frequency domain feature vector is obtained by concatenation:

[0093]

[0094] Step 4: Introduce the multimodal gated attention fusion mechanism (MTF-AGU) to enhance the frequency domain feature vector F enhanced and the time domain transient eigenvector F time to integrate.

[0095] Calculate the heading change rate Δθ and acceleration a t , the speed skewness (Skewness = 0.8) and kurtosis (Kurtosis = 3.2) in the current window are used to construct the time domain transient feature vector F time When the ship is not less than 5 nautical miles from the port, the dynamic weight g is calculated according to the relevant formulas in steps 4.2 and 4.3 above. Otherwise, g = 0.3 is set to enhance the time domain burst response. The fused feature vector F fused =0.3·H freq +0.7·H time .

[0096] Step 5: Training of deep learning prediction model and speed prediction.

[0097] The network architecture adopts a single-layer LSTM model, and the input layer receives the fusion feature F fused , with a sequence length of 20 and a feature dimension of 16. A single-layer LSTM contains 32 hidden units, uses the tanh activation function, and returns the output of the last time step. A 1D convolutional layer (kernel size 3, stride 1) is added between this output and the fully connected layer to extract local temporal features. The output layer uses a linear activation function to output the predicted speed, heading, and position coordinates. In this example, this is primarily used to predict the tanker's future speed.

[0098] The loss function of a single-layer LSTM model is:

[0099] L=MSE+0.1·|Δv pred -a·Δt| 2

[0100] Where Δv pred represents the speed change predicted by the model, a represents the real-time acceleration of the ship (calculated through AIS data), and Δt represents the time step (related to the AIS sampling frequency, Δt = 1s when 1Hz).

[0101] Training results: RMSE = 0.11 knots (traditional LSTM is 0.15 knots), indicating that this method has a higher accuracy in predicting the future speed of the ship.

[0102] Example 2: Ship track prediction for the next 10 minutes

[0103] For steps 1 to 4, please refer to the relevant contents in Example 1. The difference from Example 1 is that the network architecture of the deep learning prediction model in step 5 of this example adopts a two-layer LSTM model, in which the first layer LSTM contains 64 hidden units and the activation function is the tanh function. The Dropout rate of the Dropout (random inactivation regularization) layer is 0.2 to prevent overfitting. The second layer LSTM contains 32 hidden units and the activation function is the tanh function, which returns the output of the last time step. The output is connected to the fully connected layer, which contains 16 neural units and uses linear activation to predict the track (longitude and latitude) of the next time step.

[0104] The loss function of the two-layer LSTM model is:

[0105] L=(Δlat) 2 +(Δlon) 2

[0106] Among them, Δlat represents the longitude prediction error, and Δlon represents the latitude prediction error.

[0107] Training results: The RMSE of the comprehensive prediction of longitude and latitude is 0.012° (the traditional LSTM is 0.084°), indicating that this method has a higher accuracy in predicting the future trajectory of the ship.

[0108] Based on the above-mentioned disclosed method for ship navigation trajectory prediction based on frequency domain analysis and deep learning, it can be seen that the core invention points of this application include: proposing a dynamic sliding window frequency domain feature online update algorithm, and realizing real-time processing and noise suppression of high-frequency AIS data through exponentially weighted moving window and incremental FFT calculation; introducing a multimodal time-frequency gated attention fusion mechanism (MTF-AGU) to dynamically balance the weights of frequency domain periodic features and time domain burst features; constructing a joint representation of ship dynamics enhancement, that is, embedding physical parameters such as the ship's moment of inertia into the frequency domain feature vector, and then transmitting it to the prediction model, thereby improving the physical rationality of the ship navigation trajectory prediction.

[0109] Based on the same inventive concept, another embodiment of the present application provides a ship navigation trajectory prediction system based on frequency domain analysis and deep learning. The implementation solution for solving the problem provided by this system is similar to the implementation solution recorded in the above-mentioned ship navigation trajectory prediction method based on frequency domain analysis and deep learning. Therefore, the specific limitations in the embodiment of a ship navigation trajectory prediction system provided below can be referred to the limitations of steps 1 to 5 in the above-mentioned ship navigation trajectory prediction method, and will not be repeated here.

[0110] like Figure 2 As shown, the ship trajectory prediction system includes a data processing module, a frequency domain conversion module, a feature extraction module, a feature fusion module, and a trajectory prediction module, which are connected in sequence. The data processing module collects and preprocesses ship navigation AIS data in real time. AIS data includes time series data of the ship's position, speed, and heading. The frequency domain conversion module segments the preprocessed continuous time domain data using a dynamic sliding window and converts it into a frequency domain representation. It updates the frequency components of the current window using a sliding update strategy. The feature extraction module selects frequency components to construct frequency domain feature vectors and also constructs enhanced frequency domain feature vectors based on physical parameters. The feature fusion module introduces a multimodal gated attention fusion mechanism to fuse the frequency domain feature vector / enhanced frequency domain feature vector with the time domain transient feature vector. The trajectory prediction module inputs the fused feature vector into a deep learning prediction model and converts the frequency domain prediction output back into a time domain representation to generate the ship's future position, speed, and heading information.

[0111] The above description is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiment. It is understood that other improvements and variations directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the scope of protection of the present invention.

Claims

1. A ship navigation trajectory prediction method based on frequency domain analysis and deep learning, characterized in that: The method comprises: Real-time collection of ship navigation AIS data and pre-processing operations, the AIS data includes ship position, speed and heading time series data; The pre-processed continuous time domain data is segmented using a dynamic sliding window and converted into a frequency domain representation, and the frequency component of the current window is updated using a sliding update strategy; Select frequency components to construct frequency domain feature vectors; A multimodal gated attention fusion mechanism is introduced to fuse the frequency domain feature vector and the time domain transient feature vector; The fused feature vector is input into the deep learning prediction model, and the frequency domain prediction results output by the model are converted back to the time domain representation to generate the ship's future position, speed and heading information.

2. The ship navigation trajectory prediction method based on frequency domain analysis and deep learning according to claim 1 is characterized in that: The ship navigation AIS data is pre-processed, including: Convert the raw AIS data into a unified format, including timestamp, vessel unique identifier, position, speed, and heading fields; Use the 3σ principle to identify and remove speeds and headings outside the specified range, and use a low-pass filter to filter out high-frequency noise; For heading data with jumps after filtering, linear interpolation is used to smooth the transition.

3. The ship navigation trajectory prediction method based on frequency domain analysis and deep learning according to claim 1 is characterized in that: The method of segmenting the pre-processed continuous time domain data using a dynamic sliding window includes: Based on the pre-processed speed and heading time domain data, the heading change rate and acceleration of each time stamp are obtained; When the heading change rate is greater than a set heading threshold, or the acceleration is greater than a set acceleration threshold, the sliding window is switched from a default window length to a minimum window length.

4. The ship navigation trajectory prediction method based on frequency domain analysis and deep learning according to claim 1 is characterized in that: The expression for updating the frequency component of the current window using the sliding update strategy is: Among them, X old [k] is the frequency component corresponding to the initial window, X new [k] is the frequency component corresponding to the current window, x new 、x old are the newly added data points and the removed data points after the window slides, respectively. k represents the kth frequency component, and N represents the current window length, which represents the number of data points in this frequency domain transformation.

5. The method for predicting ship navigation trajectory based on frequency domain analysis and deep learning according to claim 1, characterized in that: The selecting frequency components to construct a frequency domain feature vector includes: Calculate the amplitude spectrum of the frequency component corresponding to each window; Calculate the cumulative energy based on the amplitude spectrum of each frequency component, and select the frequency component whose cumulative energy exceeds a set proportion; The amplitude of the selected frequency component is normalized to form a frequency domain feature vector.

6. The method for predicting ship navigation trajectory based on frequency domain analysis and deep learning according to claim 1, characterized in that: The multimodal gated attention fusion mechanism is introduced to fuse the frequency domain feature vector and the time domain transient feature vector, including: Based on the time domain data of speed and heading in the current window, a time domain transient feature vector is constructed; Inputting the frequency domain feature vector and the time domain transient feature vector into a fully connected layer for feature mapping respectively; The mapping features of the frequency domain and time domain are spliced together, and the dynamic weights are calculated through the gated attention mechanism; The mapping features of the frequency domain and the time domain are fused based on the dynamic weight.

7. The method for predicting ship navigation trajectory based on frequency domain analysis and deep learning according to claim 6, characterized in that: The process of constructing a time domain transient feature vector based on the speed and heading time domain data of the current window includes: Based on the time domain data of speed and heading in the current window, the heading change rate, acceleration, speed skewness and kurtosis of the current window are obtained to form the time domain transient feature vector.

8. The method for predicting ship navigation trajectory based on frequency domain analysis and deep learning according to claim 1, characterized in that: The deep learning prediction model is obtained based on model training of a long short-term memory network LSTM, wherein the LSTM is a single-layer structure or a double-layer structure.

9. The ship navigation trajectory prediction method based on frequency domain analysis and deep learning according to any one of claims 1 to 8, characterized in that: The method further comprises: Load the physical parameters of interest according to the ship type and perform normalization processing; Concatenating the physical parameter feature vector with the frequency domain feature vector to construct an enhanced frequency domain feature vector; A multimodal gated attention fusion mechanism is introduced to fuse the enhanced frequency domain feature vector and the time domain transient feature vector.

10. A ship navigation trajectory prediction system based on frequency domain analysis and deep learning, characterized in that: include: A data processing module is used to collect ship navigation AIS data in real time and perform pre-processing operations. The AIS data includes time series data of ship position, speed and heading; The frequency domain conversion module is used to segment the pre-processed continuous time domain data using a dynamic sliding window and convert it into a frequency domain representation, and update the frequency component of the current window using a sliding update strategy; Feature extraction module, used to select frequency components to construct frequency domain feature vectors; A feature fusion module, used to introduce a multimodal gated attention fusion mechanism to fuse the frequency domain feature vector and the time domain transient feature vector; The trajectory prediction module is used to input the fused feature vector into the deep learning prediction model, convert the frequency domain prediction results output by the model back to the time domain representation, and generate the future position, speed and heading information of the ship.

Images