A deep learning method for multi-step prediction of two-ship navigation states

Abstract

The application discloses a kind of deep learning methods for two ship navigation state multi-step prediction, comprising: according to AIS data features, from mass data, quickly extract, screen, interpolate effective ship data, according to ship potential risk formula, obtain the ship sequence data pair of maximum potential risk in the same space-time range, construct ship navigation state prediction model, effectively connect encoder and decoder through interactive module, and realize the intercommunication between ship historical data information, when model training, according to the loss value of calculation, carry out self-feedback regulation, help model quickly converge.The application realizes the multi-step prediction of the navigation state of two ships by combining the ship potential risk field and the long short-term memory network model, overcomes the limitation that the single ship model prediction ignores the influence of surrounding ships, and provides a theoretical basis for the selection of multi-ship model data.

Description

Technical Field

[0001] This invention relates to the field of ship state prediction technology, specifically to a deep learning method for multi-step prediction of the navigation state of two ships. Background Technology

[0002] In maritime safety research, collision avoidance has always been a challenging task that urgently needs to be addressed. With the increasing number of ships in the shipping industry and the rising demands of shipping operations, solving this problem has become increasingly important and urgent. In traditional shipboard environments, collision avoidance is accomplished by the ship's operators based on a comprehensive assessment of ship instrument information, the navigation environment, and their own operational experience.

[0003] Research on ship state prediction mainly falls into two categories: motion-based models and learning-based models. Due to the diverse properties of ships, existing motion-based models are not suitable for big data analysis. Among existing learning-based models, the Kalman filter method struggles to select a suitable kernel function and requires other methods to determine model parameters to aid convergence. Gaussian process regression, another learning-based method, estimates regression model parameters based on posterior distribution assumptions and sample data. Its distributional characteristics make this method widely applicable, but it suffers from limitations in scaling and computation when dealing with massive datasets, and it also has the limitation of neglecting the influence of surrounding ships in single-ship model predictions. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning method for multi-step prediction of the navigation states of two ships. It aims to address the limitations of existing ship state prediction methods, such as their inapplicability to big data analysis and the neglect of the influence of surrounding ships in single-ship model predictions.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] This invention provides a deep learning method for multi-step prediction of the navigation states of two ships, comprising:

[0007] Step S1: Obtain raw data from the Automatic Identification System (AIS) and preprocess the raw data to obtain preprocessed data;

[0008] Step S2: Based on the preprocessed data, match the Automatic Identification System (AIS) data sequence pairs according to time, calculate the sum of the weighted potential risk values ​​of the data sequence pairs, retain the ship sequence data pair with the highest current potential risk, and obtain a set of ship sequence data pairs;

[0009] Step S3: Based on the set of ship sequence data pairs, construct a ship navigation state prediction network model to predict the future state of the ship;

[0010] Step S4: Based on the ship navigation state prediction network model, calculate the error between the output value of the ship navigation state prediction network model and the true value, dynamically adjust the learning rate of the ship navigation state prediction model, and obtain the prediction result;

[0011] Step S5: Visualize the prediction results.

[0012] Preferably, in step S1, the preprocessing of the raw data of the Automatic Identification System (AIS) includes:

[0013] Step S1.1: For the acquired raw data from the Automatic Identification System (AIS), determine its data type and adjust its data format, and initially extract the data using a binary search algorithm;

[0014] Step S1.2: Based on the channel constraint boundary and the Laida criterion, further refine the data range of the initially extracted data and remove abnormal data to obtain the missing data;

[0015] Step S1.3: Complete the missing data by segmented cubic spline interpolation to obtain the preprocessed data.

[0016] Preferably, in step S1.1, the preliminary data extraction using the binary search algorithm specifically includes: extracting ship navigation status sequence data within the geographical range involved in the ship navigation status prediction from the original data of the Automatic Identification System (AIS).

[0017] Preferably, in step S2, the step of matching the Automatic Identification System (AIS) data sequence pairs according to time includes: extracting a set of ship data matrices that meet the time conditions based on the preprocessed data, and matching the corresponding data matrices according to the AIS data sequence pairs.

[0018] Preferably, the time t condition for extracting the ship data matrix set that satisfies the time condition is as follows:

[0019] t o -t m ≤t≤t o +t n -1

[0020] In the formula: t m To predict the required historical data period,

[0021] t o The predicted start time,

[0022] t n To predict the time frame for future data.

[0023] Preferably, in step S2, calculating the sum of weighted potential risk values ​​for the data sequence pairs includes: extracting latitude and longitude information, navigation information, and ship length information from the Automatic Identification System (AIS) data sequence pairs, and calculating the sum of the weighted potential risk values ​​for the two ships from t=t o -t m Time until t=t o The potential risk value at time -1.

[0024] Preferably, in step S2, retaining the ship sequence data pair with the highest potential risk specifically means:

[0025] If the two ships are from t=t o -t m Time until t=t o If the potential risk value at time -1 is not greater than the current preset maximum value, then the corresponding ship sequence data pair will be removed.

[0026] If the two ships are from t=t o -t m Time until t=t o If the potential risk value at time -1 is greater than the current preset maximum value, then the maximum value is updated, and the corresponding ship sequence data pair is retained.

[0027] Preferably, in step S3, the ship navigation state prediction network model consists of an encoder, an interaction module, and a decoder.

[0028] The encoder consists of two sub-encoders with the same LSTM unit, used to acquire historical ship data information.

[0029] The interaction module is used to integrate motion information between ships;

[0030] The decoder consists of two independent sub-decoders with the same LSTM unit, used to predict the future state of the ship.

[0031] Preferably, in step S4, calculating the error between the output value of the ship navigation state prediction network model and the true value includes:

[0032] Step S4.1: Map any ship sequence data in the set of ship sequence data pairs to the range [0, 1] by using maximum and minimum value normalization;

[0033] Step S4.2: Obtain training data from the ship sequence data set, input the training data into the ship navigation state prediction network model, calculate the output value and the true value of the ship navigation state prediction network model, and calculate the error between the output value and the true value.

[0034] Step S4.3: Determine the result of the error, and dynamically adjust the learning rate of the ship navigation state prediction network model according to the determination result to obtain the prediction result.

[0035] Preferably, the prediction results are visualized by displaying them using the matplotlib toolkit in Python.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] This invention utilizes AIS data characteristics to rapidly extract, filter, and interpolate effective ship data from massive datasets. Based on the ship potential risk formula, it acquires ship sequence data pairs with the highest potential risk within the same spatiotemporal range, constructing a ship navigation status prediction model. An interactive module effectively connects the encoder and decoder, enabling the exchange of historical ship data. During model training, self-feedback adjustments are made based on the calculated loss value to facilitate rapid model convergence. This invention combines the ship potential risk domain with a long short-term memory network model to achieve multi-step prediction of the navigation status of two ships, overcoming the limitation of single-ship model prediction ignoring the influence of surrounding ships and providing a theoretical basis for selecting data for multiple ship models. Attached Figure Description

[0038] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the drawings described below are one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort:

[0039] Figure 1 A flowchart of a deep learning method for multi-step prediction of navigation status of two ships provided in an embodiment of the present invention;

[0040] Figure 2 A flowchart of AIS (Automatic Identification System) data preprocessing provided in an embodiment of the present invention;

[0041] Figure 3 A flowchart of ship data matching provided in an embodiment of the present invention;

[0042] Figure 4 This is a flowchart of a self-feedback training process provided in an embodiment of the present invention. Detailed Implementation

[0043] The following is in conjunction with the appendix Figures 1 to 4The detailed embodiments of this invention provide a further detailed description of the deep learning method for multi-step prediction of the navigation states of two ships. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clarify the explanation of the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes and to enable those skilled in the art to understand and read them, and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0044] To address the limitations of existing ship condition prediction methods, such as their inapplicability to big data analysis and the neglect of the influence of surrounding ships in single-ship model predictions, Figure 1 As shown, this embodiment provides a deep learning method for multi-step prediction of the navigation states of two ships, including:

[0045] Step S1: AIS (Automatic Identification System) data preprocessing.

[0046] Step S1.1: Obtain raw data from the Automatic Identification System (AIS). For the obtained raw data from the AIS, determine its data type and adjust its data format, and initially extract the data using a binary search algorithm.

[0047] Step S1.2: Based on the channel constraint boundary and the Laida criterion, further refine the data range of the initially extracted data and remove abnormal data to obtain the missing data.

[0048] Step S1.3: Complete the missing data by segmented cubic spline interpolation to obtain the preprocessed data.

[0049] refer to Figure 2 As shown, steps S1.1 to S1.3 specifically include: considering the characteristics of ship AIS data such as measurement errors, inconsistent time intervals, rich ship navigation status information but lack of navigation environment information, determining the input data and prediction data of the ship navigation status prediction method.

[0050] Q t =[x t ,y t ,v t ,r t ,L i ] T t∈[0,1440] (1)

[0051] A i =[Q j …,Q k j,k∈[0,1440] (2)

[0052] In equations (1) and (2): x t ,y t ,v t ,r t ,L i Let Q represent the ship's longitude, latitude, speed, heading, and length at time t. t Let i be the data vector at time t; i is the ship's MMSI number, A i Let i be the data matrix for ship i.

[0053] Based on the geographical area involved in the prediction of ship navigation status, the ship navigation status sequence data within the geographical area involved is extracted from the massive AIS dataset. The processing procedure is as follows:

[0054] Sort the original data in the database in ascending order, and determine the maximum and minimum values ​​of ship latitude and longitude based on geographical range: x max x min y max and y min Based on the sorted data, a binary search method is used to find the data segment corresponding to the geographical range, avoiding the disadvantages of excessive memory consumption and low running efficiency caused by repeated database queries and data extraction.

[0055] Calculate the boundary function f1(x) on the channel based on the electronic nautical chart. t ,y t ) and the lower boundary function f2(x) t ,y t As shown in formula (3). When matrix A i Column vector Q t Does not satisfy f1(x) t ,y t f2(x) ≤ 0 or f2(x) t ,y t When )≥0, then the ship data A that deviates from the course will be... i Remove.

[0056] f n (x t ,y t ) = a n x t +d n y t +c n n∈{1,2} (3)

[0057] an ,d n ,c n Describe the boundary function f n coefficient, x t ,y t These are the horizontal and vertical coordinates, respectively.

[0058] According to the Raida criterion, the ship data matrix A is calculated using formulas (4) and (5). i The arithmetic mean of the data in row I With standard deviation δ I Matrix A i The residuals ε of each data point IJ It can be obtained according to formula (6). When the absolute value of the residual does not satisfy the range given in formula (7), matrix A i Then it does not belong to the reserved data matrix set M.

[0059]

[0060]

[0061]

[0062]

[0063] Based on the missing data obtained from the above calculations, cubic spline interpolation is performed to fill in A. i Data with missing time points was used to obtain preprocessed data.

[0064] Step S2: Ship data matching.

[0065] refer to Figure 3 As shown, based on the preprocessed data, extract data that satisfies t o -t m ≤t≤t o +t n The set of ship data matrices N, t, is -1. m To predict the required historical data duration, t o t is the predicted start time. n To predict the time horizon of future data, the initial data matrices for the two ships p and p′ are S, respectively. p ∈N and S p′ ∈N.

[0066] From S respectively p With S p′ Extract latitude and longitude information x pt ,y pt and x p′t ,y p′t Heading information ptCaptain Information L p And the calculation of the two ships from t=t using formulas (8)-(11) is performed. o -t m Time until t=t o The potential risk value at time -1; finally, retain the two ship data matrices that satisfy formula (12), and represent the retained ship data matrices as follows: Obtain the set φ of ship sequence data pairs.

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] In equations (8) to (12): D represents the azimuth angle of ship p′ relative to ship p at time t. p→p′,t L represents the semi-versus distance between ships p and p′ at time t. p Let y represent the length of ship p. pt With x pt Let y represent the longitude and latitude of ship p at time t, respectively. p′t With x p′t Let R represent the longitude and latitude of ship p′ at time t, respectively, and let Z represent the Earth's radius. p→p′,t ω represents the potential risk value posed by ship p′ to ship p at time t. t Let be the weighting factor at time t.

[0073] Continue to refer to Figure 3 As shown, ship sequence data pairs are obtained, and the potential risk value between ships is calculated. In this embodiment, Z is determined. p→p′,t Is it greater than the current preset maximum value? If Z p→p′,t If Z is not greater than the current preset maximum value, then discard the current data pair; if Z p→p′,t If the value is greater than the current preset maximum, update the maximum value and retain the current data pair. Then determine whether all data pairs have been fully traversed. If they have been fully traversed, the calculation ends; otherwise, return to recalculate the potential risk values ​​between ships until all ships have been traversed. Perform the above calculations sequentially for all ships in dataset N to obtain the data pair set φ.

[0074] Step S3: Construct a ship state prediction network model.

[0075] The ship state prediction network model consists of three main parts: an encoder, an interaction module, and a decoder.

[0076] The encoder consists of two sub-encoders with identical LSTM (Long Short-Term Memory) units to acquire historical ship data. The interaction module is used to fuse motion information between ships. The decoder consists of two independent sub-decoders with identical LSTM units to predict the future state of the ship.

[0077] Based on the assumption that ship motion states are similar when potential risks exist, the encoder section uses the same LSTM unit to learn the ship navigation characteristics from historical data. The LSTM unit calculation formula is as follows:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] h t =o t ⊙tanh(C t (18)

[0084] Based on the LSTM unit input at time t and the hidden state h of the LSTM cell at time t-1 t-1 Update input matrix to determine cell state Forgetting Gating Matrix f t Input gating matrix i t Output gate matrix o t Cell state C at time t t Subject to f t and i t Control and C t-1 and The impact; the hidden state h at time t t Received o t Control and C t The impact; W f W i W o and W c These are the weight matrices for the forget gate, input gate, output gate, and updating the input matrix, respectively. f b i b o and bc These are their biases; σ is the sigmoid function, and tanh is the hyperbolic tangent function. The outputs of the two sub-encoders are shown in equation (19):

[0085]

[0086]

[0087] In equation (19): LSTM represents the calculation of equations (13) to (18). and To predict the input data for the ship's states p and p′, h p with h p′ This refers to the hidden state information of the two ships.

[0088] The information fusion layer captures information from the sub-encoder u through four different connection layers, thereby extracting the motion features of different ships. and and Its weight matrix, and The bias is calculated using the following formula:

[0089]

[0090]

[0091] Using a batch normalization layer, learnable parameters γ and β are introduced to initialize the i-th dimension of the neural network unit with the initial output value k. (i) z is obtained after calculation using formulas (21)-(24). (i) This controls the output range of neural network units and helps with the backpropagation of weight gradients during network training.

[0092]

[0093]

[0094]

[0095]

[0096] In equations (21) to (24): μ B The initial output value k (i) The arithmetic mean of m and k (i) Dimension size, σ B For k (i) The variance is ε, which is a very small value to avoid the denominator being zero, and γ and β are the scaling factor and offset coefficient, respectively.

[0097] The motion characteristics of different ships after batch standardization and To merge, that is and The elements of each dimension are added together to obtain the relative motion characteristics D between the ships. u The calculation formulas are as follows:

[0098]

[0099]

[0100]

[0101] In equation (25): BN represents batch standardization operation.

[0102] The decoder uses the same LSTM module, independent of the encoder, to capture temporal information from historical data and output the predicted ship motion state O. u .

[0103] Step S4: Self-feedback training of the model.

[0104] refer to Figure 4 As shown, to avoid the influence of different data units on model prediction, the data with different ranges of variation are mapped to the range [0,1] by the maximum and minimum value normalization. The calculation formula is shown in (26):

[0105]

[0106] In equation (26): τ represents any ship sequence data in the data pair set φ. This is the normalized ship sequence data.

[0107] Based on the ship state prediction network model obtained in step S3, the data set φ obtained after processing the collected AIS system data in steps S1 and S2 is randomly divided into training datasets φ. train and validation dataset φ valid The percentages were 70% and 30%, respectively. For the training dataset φ train All data pairs (S) train_p ,S train_p′ During the self-feedback training process, t o Sequence data from -m to t0-1 Normalized calculations are performed and used as input values ​​to the ship state prediction network model; from t0 to t0+t n -1 sequence data and Normalized calculations are performed and used as the true values ​​of the ship state prediction network model output. The network forward propagation obtains the output values, and the output values ​​derived from the ship state prediction network model are calculated. Its true value Y i The error is shown in formula (27). When the prediction network model is on the validation dataset φ valid If the error on the training dataset does not decrease over consecutive training epochs (α), reduce the learning rate of the ship state prediction network model by half until it reaches the minimum learning rate ξ; when on the validation dataset φ valid If the error does not decrease within ω consecutive training cycles, then the network model training is terminated.

[0108]

[0109] Step S5: Visualize the prediction results.

[0110] The data to be predicted is input into the trained network model, and the output of the network model needs to be inversely normalized as shown in formula (28):

[0111]

[0112] The processed data can be visualized using Python's matplotlib toolkit.

[0113] In summary, this embodiment rapidly extracts, filters, and interpolates effective ship data from massive datasets based on AIS data characteristics. According to the ship potential risk formula, it obtains ship sequence data pairs with the highest potential risk within the same spatiotemporal range, constructs a ship navigation status prediction model, and effectively connects the encoder and decoder through an interactive module, enabling communication between historical ship data. During model training, it performs self-feedback adjustments based on the calculated loss value to help the model converge quickly. This embodiment combines the ship potential risk domain with a long short-term memory network model to achieve multi-step prediction of the navigation status of two ships, overcoming the limitation of single-ship model prediction ignoring the influence of surrounding ships and providing a theoretical basis for the selection of data for multiple ship models.

[0114] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0115] It should be noted that the apparatus and methods disclosed in the embodiments herein can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments herein. In this regard, each block in a flowchart or block diagram may represent a module, program, or part of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system to perform the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0116] In addition, the functional modules in the various embodiments of this article can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0117] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A deep learning method for multi-step prediction of the navigation states of two ships, characterized in that, include: Step S1: Obtain raw data from the Automatic Identification System (AIS) and preprocess the raw data to obtain preprocessed data; Step S2: Based on the preprocessed data, match the Automatic Identification System (AIS) data sequence pairs according to time, calculate the sum of the weighted potential risk values ​​of the data sequence pairs, retain the ship sequence data pair with the highest current potential risk, and obtain a set of ship sequence data pairs; The method of matching the automatic identification system data sequence pairs according to time includes: extracting a set of ship data matrices that meet the time conditions based on the preprocessed data, and matching the corresponding data matrices according to the automatic identification system data sequence pairs; The extraction of time from the ship data matrix set that meets the time condition. t The limiting conditions are: In the formula: To predict the required historical data period, The predicted start time, To predict the time frame of future data; Step S3: Based on the set of ship sequence data pairs, construct a ship navigation state prediction network model to predict the future state of the ship; The ship navigation status prediction network model consists of an encoder, an interaction module, and a decoder. The encoder consists of two sub-encoders with the same LSTM unit, used to acquire historical ship data information. The interaction module is used to integrate motion information between ships; The decoder consists of two independent sub-decoders with the same LSTM unit, used to predict the future state of the ship. Step S4: Based on the ship navigation state prediction network model, calculate the error between the output value of the ship navigation state prediction network model and the true value, dynamically adjust the learning rate of the ship navigation state prediction model, and obtain the prediction result; Step S5: Visualize the prediction results.

2. The deep learning method for multi-step prediction of navigation states of two ships as described in claim 1, characterized in that, In step S1, the preprocessing of the raw data from the Automatic Identification System (AIS) includes: Step S1.1: For the acquired raw data from the Automatic Identification System (AIS), determine its data type and adjust its data format, and initially extract the data using a binary search algorithm; Step S1.2: Based on the channel constraint boundary and the Laida criterion, further refine the data range of the initially extracted data and remove abnormal data to obtain the missing data; Step S1.3: Complete the missing data by segmented cubic spline interpolation to obtain the preprocessed data.

3. The deep learning method for multi-step prediction of navigation states of two ships as described in claim 2, characterized in that, In step S1.1, the preliminary data extraction using the binary search algorithm specifically includes: extracting the ship navigation status sequence data within the geographical range involved in the ship navigation status prediction from the original data of the Automatic Identification System (AIS).

4. The deep learning method for multi-step prediction of navigation states of two ships as described in claim 1, characterized in that, In step S2, calculating the sum of weighted potential risk values ​​for the data sequence pairs includes: extracting latitude and longitude information, navigation information, and ship length information from the Automatic Identification System (AIS) data sequence pairs, and calculating the weighted potential risk values ​​for the two ships from... Time to The potential risk value at any given moment.

5. The deep learning method for multi-step prediction of navigation states of two ships as described in claim 4, characterized in that, In step S2, retaining the ship sequence data pair with the highest potential risk specifically means: If the two ships mentioned above are from Time to If the potential risk value at any given time is not greater than the current preset maximum value, then the corresponding ship sequence data pair will be removed. If the two ships mentioned above are from Time to If the potential risk value at a given moment is greater than the current preset maximum value, then the maximum value is updated, and the corresponding ship sequence data pair is retained.

6. The deep learning method for multi-step prediction of navigation states of two ships as described in claim 1, characterized in that, In step S4, calculating the error between the output value of the ship navigation state prediction network model and the true value includes: Step S4.1: Map any ship sequence data in the ship sequence data pair set to a maximum-minimum value normalization method. Within the range; Step S4.2: Obtain training data from the ship sequence data set, input the training data into the ship navigation state prediction network model, calculate the output value and the true value of the ship navigation state prediction network model, and calculate the error between the output value and the true value. Step S4.3: Determine the result of the error, and dynamically adjust the learning rate of the ship navigation state prediction network model according to the determination result to obtain the prediction result.

7. The deep learning method for multi-step prediction of navigation states of two ships as described in claim 1, characterized in that, In step S5, the prediction results are visualized, specifically by displaying the prediction results using the matplotlib toolkit in Python.

Images