Ship navigation trajectory prediction method based on dynamic scoring mechanism time-varying weight

Abstract

The invention discloses a ship navigation trajectory prediction method based on a dynamic scoring mechanism time-varying weight. The ship navigation trajectory prediction method comprises the following steps: extracting a navigation trajectory and defining navigation features; establishing a ship navigation trajectory prediction model based on the dynamic scoring mechanism time-varying weight; and evaluating the precision of the ship trajectory prediction model. According to the method, the navigation characteristic variables of ship navigation are defined according to the position information, the navigation state and the environment information of the ship in the navigation process, so that the input information can sufficiently and comprehensively contain factors influencing ship navigation, and the ship navigation behavior mode can be conveniently learned and recognized. According to the invention, an information extraction mechanism capable of carrying out dynamic weight distribution from two dimensions of a coding end and a decoding end at the same time is designed, the weight of an input sequence can be dynamically adjusted to adapt to various prediction tasks, and the advantages of global connection and local connection of a dynamic scoring mechanism are combined; and the information mining capability of the ship trajectory prediction model in navigation characteristic variable learning is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of intelligent shipping and artificial intelligence, and particularly relates to a method for predicting ship navigation trajectories based on time-varying weights of a dynamic scoring mechanism. Background Art

[0002] Currently, ship trajectory prediction technologies mainly include kinematic models and machine learning methods. Kinematic models predict trajectories through equations, but complex environments and non-linear trajectories increase the difficulty. Machine learning, especially deep learning, has been introduced to solve this problem due to its powerful non-linear fitting ability, but it still faces performance challenges in complex environments, mainly due to the limited ability of the model to filter information. The innovative application of time-varying weights of the dynamic scoring mechanism provides an effective solution for this. Although it has been widely applied in the field of trajectory prediction, in-depth exploration of its stability and adaptability in various navigation environments remains the focus of researchers. In the future, the evolution direction of ship trajectory prediction technology focuses on improving the accuracy, efficiency of prediction, and the ability to handle complex uncertain scenarios, which is an inevitable trend in the development of maritime traffic management. Despite years of research, numerous mathematical models and algorithms, including cutting-edge machine learning and deep learning technologies, have been developed for trajectory prediction. However, by comprehensively reviewing the existing literature, it is not difficult to find that strategies for improving prediction accuracy and coping with complex maritime environments are still a scientific research topic that urgently needs to be further explored. Summary of the Invention

[0003] In order to solve the above problems existing in the prior art, the present invention proposes a method for predicting ship navigation trajectories based on time-varying weights of a dynamic scoring mechanism, which can effectively extract the behavioral characteristics of ship navigation to provide effective and accurate trajectory prediction services.

[0004] In order to achieve the above object, the technical means adopted by the present invention are as follows: A method for predicting ship navigation trajectories based on time-varying weights of a dynamic scoring mechanism, comprising the following steps:

[0005] S1. Extract the navigation trajectory and define the navigation features

[0006] S11. Preprocess the navigation trajectory

[0007] In order to solve the problems of the original AIS system data and on-board sensor data due to observation errors and time intervals, the original AIS system data and on-board sensor data are preprocessed according to the following steps:

[0008] S111. Segment and extract the navigation trajectory

[0009] In the AIS system data, different vessels are distinguished by their Maritime Mobile Service Identity (MMSI). For vessels with the same MMSI, if the time interval between two adjacent time points of the vessel in the AIS system exceeds the set threshold λ, this time interval point is recognized as a segmentation point of the vessel's navigation track. According to the above rules, a continuous segment of AIS data is divided into several independent and continuous navigation track segments executed by the vessel.

[0010] S112. Process outliers

[0011] In the extracted navigation track data, data points with repeated longitude and latitude and abnormal speed and course at the same moment among vessels with the same MMSI are deleted.

[0012] S113. Process missing data

[0013] For the situation where there is missing data in the navigation track data, the method of linear interpolation is used to fill in the missing values. The method is as follows:

[0014]

[0015] Where, represents the interpolation of variable a at moment, and a j respectively represent the data values of variable a at and j moments. The moment is between moment j and , and j is less than a represents any continuous variable in the vessel's navigation track data;

[0016] S114. Correlate navigation track data

[0017] According to different data sources, the data collected by the AIS system and on-board sensors are represented as follows:

[0018]

[0019] Where, and respectively represent the i-th information collected at the t-th moment in the data from the AIS system and on-board sensors. i represents the information collection sequence number, and n represents the total number of collected information. The information includes longitude Lng, latitude Lat, speed S, course C, wind speed WS, and wind direction WA; in data ais , it contains at least Lng and Lat information, that is, n≥2; in data sen n≥1; T ais and T senrespectively represent the sets of data acquisition times of the AIS system and the on-board sensors. The data sampling frequencies of the AIS system and the on-board sensors are c and b respectively, where c≠b, and both c and b are constants; among them, T sen ∈{1, 1 + c, …, t, t + c, …} and T ais ∈{1, 2 + b, …, t, t + b, …}; align the data sampling frequencies of the data with the larger sampling frequency through interpolation, then take the intersection of the sampling times of the two independent data sources, and correlate the two parts of data. The formula is as follows:

[0020]

[0021] After interpolation on the sampling frequency, T sen′ ∈{1, 2, …, t, t + 1, t + 2, …}, and the interpolated on-board sensor data is expressed as Let data ais and data sensor′ correlate the two parts of data according to the time data to obtain the correlated data data inter as follows:

[0022] T inter = T ais ∩T sen′

[0023]

[0024] S12. Define the navigation trajectory features

[0025] First, define the navigation position information, navigation state information, and navigation environment information during the ship's navigation through the correlated data, and then define the navigation trajectory features. The steps are as follows:

[0026] S121. Define the navigation position features

[0027] Obtain the geographical location information of the ship during navigation according to the data after correlating the AIS system and the on-board sensors, and represent the geographical location information as P = {p1, p2, …, p t , …, p N}; among them, p t represents the position feature of the ship at the t-th moment, p t = (Lng t , Lat t ), Lng t , Lat t respectively represent the longitude and latitude at the t-th moment, and N represents the total number of position points recorded in the entire navigation trajectory of the ship, that is, the value of the ship's navigation trajectory length;

[0028] S122. Define the navigation state features

[0029] According to the ship's speed and heading information during navigation, the ship's navigation state characteristics are defined as S = {s1, s2, …, s t , …, s N}; where s t represents the navigation state characteristics of the ship at the t-th moment, and s t = (C t , S t ), and C t , S t represent the heading and speed at the t-th moment respectively;

[0030] S123. Define the navigation environment characteristics

[0031] According to the environmental information around the ship during navigation, the ship's navigation environment characteristics are defined as E = {e1, e2, …, e t , …, e N}; where e t represents the navigation environment characteristics of the ship at the t-th moment, and e t = (WS t , WA t ), and WA t , WS t represent the wind direction and wind speed at the t-th moment respectively;

[0032] S124. Define the navigation trajectory characteristics

[0033] According to the definitions of the navigation trajectory-related characteristics in steps S121 to S123, the ship's navigation trajectory characteristics are defined as:

[0034]

[0035] In the formula, G represents the entire navigation characteristics of the ship during navigation, and g t represents the navigation trajectory characteristics at the t-th moment;

[0036] S2. Establish a ship navigation trajectory prediction model based on the time-varying weights of the dynamic scoring mechanism

[0037] S21. Construct the input matrix I and output matrix O of the ship navigation trajectory prediction model

[0038] According to the navigation characteristic trajectory G, use its first h moments in the time series to predict the ship navigation trajectory position information at the h + 1-th moment, which is expressed as follows:

[0039]

[0040] In the formula, I t represents the input matrix of the first h moments before the t-th moment, and the matrix size is h × 6; Ot+1 Indicates the use of I t The predicted longitude and latitude at the t+1th moment, the matrix size is 1×2; f(·) represents the spatiotemporal mapping function of the dual dynamic weight allocation ship trajectory prediction model based on the end-to-end network; the sliding window method is adopted on G to obtain the experimental sample set, which is expressed as follows:

[0041] I={I1,…,I N-h-1}

[0042] O={O1,…,O N-h-1}

[0043] Sample={I,O}={(I1,O1),…,(I N-h-1 ,O N-h-1 )}

[0044] In the formula, I, O and Sample represent the input data set, output data set and sample data set respectively;

[0045] S22, dynamic scoring mechanism with time-varying weights;

[0046] The dynamic scoring mechanism time-varying weight includes the dynamic scoring of time series information and the dynamic weight allocation of time series information; its function is to effectively extract time series information and dynamically score the sequence information over time and redistribute the weight. The steps are as follows:

[0047] S221, extract sequence information and encode

[0048] The time series information is extracted and encoded using a double-layer memory storage neuron; the memory storage neuron is used to extract the sequence history information c and the sequence current encoding information According to the input sequence I t , the input navigation characteristic variable g at time t t , historical information c before time t t-1 and the current information at time t-1 Update to get c at time t t and The update calculation of the sequence history information c at time t is as follows:

[0049]

[0050] In the formula, m t , k t Respectively represent the information that needs to be remembered and forgotten at time t; is the candidate state that needs to be added to the historical information at time t; B is the sample batch size during the training of the ship trajectory prediction model, d is the set output information dimension; f(·) is a non-linear mapping function, and its expression is:

[0051] f(x) = max{αx, e x}

[0052] where α is a constant, x represents the unknown input information, and e x represents the exponential mapping of the unknown input information, where e is an irrational number with a value of 2.71828;

[0053] The current encoded information is updated at time t as:

[0054]

[0055] where W *i , W *f , W *c and b * are all parameters learned through iteration. Among them, b * = {b i , b f , b c , b o}; tanh(·) is a mapping function. Using g t , c t-1 , The process of calculating and obtaining is the calculation process of a single-layer memory storage neuron, and the simplified expression is as follows:

[0056]

[0057] In the formula, MN(·) represents the memory storage neuron, and θ = {W *i , W *f , W *c , W *o , b *} represents the set of all learning parameters. output mn represents the encoded output of the time series information of a single-layer memory storage neuron; through the memory storage neuron to encode and extract the time series information, the encoded information of the second-layer memory storage neuron is expressed as:

[0058]

[0059] In the formula, output′ lstmRepresents the output of the second - layer memory - storing neurons encoding information. By extracting the feature information at each moment of the input sequence at time t, a real - time information set at each moment is obtained It is expressed as:

[0060]

[0061] S222. Assign time - series information weights

[0062] For the real - time information set at each moment in the time - series information Perform dynamic weight assignment to obtain the importance with respect to each moment of the output sequence and the input sequence, filter the sequence information to obtain the dynamic weight information vector of the input sequence To extract information to identify the association between the input sequence and the output sequence;

[0063] At the decoding end, use the history and real - time coding information blocks at time t in the encoder network block as the history and real - time information of the previous moment of the memory - storing neurons at the decoding end of the decoder, and expand the input sequence I t into a matrix of 1×(h×6) as the input of the decoding information of the decoder network block; At this time, the calculation expression of the double - layer memory - storing neurons at the decoding end is:

[0064] x dec = flatten(I t )

[0065]

[0066] In the formula, and are the outputs of the double - layer memory - storing neurons at the decoding end respectively. Among them, and For the convenience of subsequent calculations, convert into a three - dimensional vector; flatten(·) is the expansion function.

[0067] In the process of dynamic weight assignment, calculate the dynamic weight of the real - time information set at the encoding end with respect to the decoding end to express the association between the input - sequence feature information and the output sequence; The calculation formula is as follows:

[0068]

[0069] Among them, and are the learned weight values, and s μ is the correlation score between the task vector and the μ-th dynamic information vector, where s = {s1, s2, …, s μ}, μ ∈ {1, 2, …, 2h}; α μ is the normalization function value of the μ-th real-time information vector for the real-time decoding information at the decoding end, T is a constant function value greater than 0 and used to regulate the size of the weight mechanism, a is a constant with a value between [1, 10]. Finally, the dynamic weight of the obtained real-time information is multiplied by the new representation vector of the real-time information set at the encoding end to obtain the dynamic weight allocation value as follows:

[0070]

[0071] In the formula, is the output value of the time-varying weight of the dynamic scoring mechanism, For ease of calculation, is transformed into two-dimensional vector; the calculation process of the time-varying weight of the dynamic scoring mechanism is simplified and expressed as follows:

[0072]

[0073] In the formula, is denoted as the simplified expression function of the time-varying weight of the dynamic scoring mechanism.

[0074] S23. Implement the time-varying weight of the dynamic scoring mechanism at the decoding end

[0075] The time-varying weight of the dynamic scoring mechanism at the decoding end includes the dynamic weight allocation of a group dynamic scoring mechanism and an information splicing and fusion block from the encoding end to the decoding end; the time-varying weight of the group dynamic scoring mechanism is used to extract significant features between time series, and the time-varying weight of the group dynamic scoring mechanism at the decoding end and the dynamic weight information vector of the input sequence at the encoding end are spliced into a new matrix information through a decoding information splicing and fusion block for ship trajectory prediction; the steps are as follows:

[0076] S231. Allocate the time-varying weight of the group dynamic scoring mechanism

[0077] To obtain the significant features between sequence feature information, the time-varying weight of the group dynamic scoring mechanism is used to extract the significant features in the time series feature information; in calculating the time-varying weight of the group dynamic scoring mechanism, the input sequence feature information I t is respectively intercepted into 6 sub-vectors according to its dimensions; and each sub-vector is respectively multiplied by the learned weight vectors and to obtain and Among them, and The calculation process of dynamic weight allocation is consistent with the time-varying weight calculation process of the dynamic scoring mechanism at the encoding end. The time-varying weight of the dynamic scoring mechanism at the decoding end The calculation formula is as follows:

[0078]

[0079] In the formula, represents the τ-th dynamic weight allocation value, τ ∈ {1, …, n}, is the group dynamic weight allocation value, concat(·) represents the concatenation function, σ represents the sigmoid function, and is expressed as follows:

[0080]

[0081] To prevent the network gradient from vanishing, a regularization normalization intermediate layer is added. By summing the input and output of the group dynamic weight allocation and then performing layer normalization, the normalized vector is obtained Finally, is stretched into a matrix of B × (h × m) Then, significant feature information is obtained through a fully connected layer The formula is as follows:

[0082]

[0083]

[0084] In the formula, is the significant feature information between the time series output by the fully connected layer, W fn and b fn are the learned parameters;

[0085] S232. Concatenate and fuse the information from the encoding end to the decoding end

[0086] After extracting information through the dynamic scoring mechanism at the encoding-decoding end of the time series feature information, the encoding-decoding information is concatenated into a new matrix And the predicted value of the ship trajectory prediction model is output through a fully connected operation as follows:

[0087]

[0088] In the formula, is the predicted value at time t + 1,W out and b out are both learned parameters.

[0089] S3. Evaluate the accuracy of the ship trajectory prediction model

[0090] Match the predicted vector of the output of the last hidden layer with the target vector O t+1 Perform matching; use the mean square error MSE as the loss function to quantify the difference between the predicted value and the true value; after completing the construction of the ship trajectory prediction model, use the root mean square error RMSE to evaluate the ship trajectory prediction model, and continue to adjust the parameters, features or algorithms of the ship trajectory prediction model according to the evaluation results to achieve satisfactory results; the smaller the mean square error and the root mean square error, the better the effect of the ship trajectory prediction model. The calculation formulas are as follows:

[0091]

[0092] In the formula, P is the size of the sample batch during training; finally, compare the error between the predicted trajectory and the true trajectory according to the prediction results, and analyze the relevant reasons based on the structure of the ship trajectory prediction model.

[0093] Further, the set threshold λ in step S111 is 4 - 8 hours.

[0094] Further, the data sampling frequencies of the AIS system and the on - board sensors in step S114 are 1 - 2 s and 2 - 4 s respectively.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] First, the present invention takes the position information, navigation state and environmental information of the ship during navigation as the main factors affecting ship navigation, and designs a model with time - varying weights based on a dynamic scoring mechanism to predict the ship navigation trajectory. The present invention defines the navigation characteristic variables of the ship navigation from the position information, navigation state and environmental information of the ship during navigation, so that the input information can comprehensively contain the factors affecting ship navigation, which is convenient for learning and identifying the ship navigation behavior pattern.

[0097] Second, the present invention designs an information extraction mechanism that can perform dynamic weight allocation from two dimensions of the encoding end and the decoding end. The time - varying weights of the dynamic scoring mechanism at the encoding end help to identify the correlation between the input sequence and the output sequence, so as to dynamically adjust the input sequence weights to adapt to various prediction tasks. When generating the output, the input information is used through the time - varying weights of the group dynamic scoring mechanism at the decoding end to obtain the local significant features between time series. The present invention combines the advantages of the global connection and the local connection of the dynamic scoring mechanism, and improves the information mining ability of the ship trajectory prediction model in learning navigation characteristic variables. Description of the Drawings

[0098] Figure 1It is the overall flowchart of the present invention.

[0099] Figure 2 It is the diagram of the method for associating the navigation track data of the present invention.

[0100] Figure 3 It is the flowchart of the memory storage neuron calculation of the present invention.

[0101] Figure 4 It is the structural diagram of the time-varying weight calculation of the dynamic scoring mechanism of the present invention.

[0102] Figure 5 It is the structural diagram of the time-varying weight calculation of the group dynamic scoring mechanism of the present invention.

[0103] Figure 6 It is the structural diagram of the ship trajectory prediction model of the present invention. Detailed implementation manners

[0104] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. As Figure 1 shown, a ship navigation trajectory prediction method based on the time-varying weight of a dynamic scoring mechanism has the following steps:

[0105] S1. Extraction of the navigation trajectory and definition of navigation features based on AIS data and on-board sensor data

[0106] To solve the problems of the original AIS system data and on-board sensor data due to observation errors and time intervals, the navigation trajectory data of the ship is preprocessed, and the navigation feature variables are defined, which mainly includes two parts: preprocessing of the navigation trajectory based on AIS data and definition of navigation features. In the part of preprocessing of the navigation trajectory based on AIS data, data information from different sources is associated. As Figure 2 shown, the associated data is obtained, and the missing data is interpolated. In the part of definition of navigation features, the geographical location feature is represented as P = {p1, p2,..., p t ,..., p N}. The navigation state feature is represented as S = {s1, s2,..., s t ,..., s N}. The navigation environment feature is defined as E = {e1, e2,..., e t ,..., e N}. According to the definition of the navigation-related information of the associated data, combined with the data after cleaning the navigation trajectory data, the navigation trajectory characteristics of the ship's navigation are defined as: G = {P, S, E|T inter}.

[0107] S2. Establish a ship navigation trajectory prediction model based on the time-varying weight of the dynamic scoring mechanism

[0108] The ship navigation trajectory prediction model based on the time-varying weight of the dynamic scoring mechanism is as Figure 6 shown. First, according to the input trajectory feature information sample set Sample = {I, O}, at any time t, construct the input matrix I t = [g t-h+1 , g t-h+2 , …, g t T . Second, at the encoding end, through the memory storage neuron, as Figure 3 shown, encode the sequence feature information of the input sequence to generate the sequence encoding information h t , h′ t , and collect the real-time information encoding at each moment At the decoding end, decode the encoding information through the group dynamic scoring mechanism. The weight allocation block of the dynamic scoring mechanism at the encoding end can identify the association between the input sequence and the output sequence, as Figure 4 shown, to obtain the dynamic weight information vector which can dynamically adjust the weight of the input sequence feature information. When generating the output, the time-varying weight of the group dynamic scoring mechanism is used to obtain the local significant features between time series as Figure 5 shown. Finally, output the trajectory prediction value through the fully connected network

[0109] S3. Evaluate the performance of the ship trajectory prediction model

[0110] The mean square error is used as the measurement standard to measure the deviation between the prediction result and the actual situation. After the model is constructed, the root mean square error is used for performance evaluation. According to the evaluation feedback, continuously fine-tune the model parameters, feature selection or algorithm strategy until the expected optimization goal is achieved.

[0111] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention. ​

[0112] The present invention is not limited to this embodiment. Any equivalent conceptions or changes within the technical scope disclosed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A ship navigation trajectory prediction method based on time-varying weights of a dynamic scoring mechanism, characterized in that: It includes the following steps: S1. Extract the navigation trajectory and define the navigation features S11. Preprocess the navigation trajectory To solve the problems of the original AIS system data and on-board sensor data due to observation errors and time intervals, the original AIS system data and on-board sensor data are preprocessed according to the following steps: S111. Segment and extract the navigation trajectory In the AIS system data, different ships are distinguished by the Maritime Mobile Service Identity (MMSI) of the ship; for ships with the same MMSI identifier, if the interval between two adjacent time points of the ship in the AIS system exceeds the set threshold λ, then this time interval point is recognized as a segmentation point of the ship's navigation trajectory; according to the above rules, a continuous segment of AIS data is divided into several independent and continuous navigation trajectory segments executed by the ship; S112. Process outliers In the extracted navigation trajectory data, delete the data points with repeated longitude and latitude and abnormal speed and heading at the same moment among the ships with the same MMSI; S113. Process missing data For the situation where there is missing data in the navigation trajectory data, the method of linear interpolation is used to complete the filling of the missing values. The method is as follows: Among them, represents the interpolation of variable a at moment, and a j respectively represent the data values of variable a at and moment j. The moment is between moment j and , and j is less than a represents any continuous variable in the ship's navigation trajectory data; S114. Associate the navigation trajectory data According to different data sources, the data collected by the AIS system and on-board sensors are respectively expressed as follows: Among them, and respectively represent the i-th information collected at the t-th moment in the source data of the AIS system and the on-board sensor. i represents the serial number of the collected information, n represents the total number of collected information, and the information includes longitude Lng, latitude Lat, speed S, course C, wind speed WS, and wind direction WA; in data ais at least includes Lng and Lat information, that is, n≥2; in data sen n≥1; T ais and T sen respectively represent the sets of data acquisition times of the AIS system and the on-board sensor. The data sampling frequencies of the AIS system and the on-board sensor are c and b respectively, where c≠b, and both c and b are constants; among them, T sen ∈{1, 1 + c, …, t, t + c, …} and T ais ∈{1, 2 + b, …, t, t + b, …}; Align the data sampling frequencies of the data with the larger sampling frequency by interpolation, then take the intersection of the sampling times of the two independent data sources, and correlate the two parts of the data. The formula is as follows: After interpolation at the sampling frequency, T sen′ ∈ {1, 2, …, t, t + 1, t + 2, …}, and the interpolated shipborne sensor data is expressed as Let data ais and data sensor′ be associated according to the time data to obtain the associated data data inter as follows: T inter = T ais ∩ T sen′ S12. Define the navigation trajectory features First, define the navigation position information, navigation state information, and navigation environment information during the ship's navigation process through the associated data, and then define the navigation trajectory features. The steps are as follows: S121. Define the navigation position features Obtain the geographical location information of the ship during navigation according to the data associated with the AIS system and the on-board sensors, and represent the geographical location information as p = {p1, p2, …, p t , …, p N}; where p t represents the position feature of the ship at the t-th moment, p t = (lng t , lat t ), lng t and Lat t respectively represent the longitude and latitude at the t-th moment, and N represents the total number of position points recorded in the entire navigation trajectory of the ship, that is, the value of the length of the ship's navigation trajectory; S122. Define the navigation state features According to the speed and heading information of the ship during navigation, the ship navigation state characteristics are defined as S = {s1, s2, …, s t , …, s N}; where s t represents the navigation state characteristics of the ship at the t-th moment, and s t = (C t , S t ), and C t , S t represent the heading and speed at the t-th moment respectively; S123. Define the navigation environment features According to the environmental information around the ship during navigation, the ship navigation environment characteristics are defined as E = {e1, e2, …, e t , …, e N}; where, e t represents the navigation environment characteristics of the ship at the t-th moment, and e t = (WS t , WA t ), where WA t and WS t respectively represent the wind direction and wind speed at the t-th moment; S124. Define the navigation trajectory features According to the definitions of the navigation trajectory-related features in steps S121 to S123, the ship's navigation trajectory features are defined as: where G represents the entire navigation characteristics during ship navigation, and g t represents the navigation trajectory characteristics at time t; S2. Establish a ship navigation trajectory prediction model based on the time-varying weights of the dynamic scoring mechanism S21. Construct the input matrix I and output matrix O of the ship navigation trajectory prediction model According to the navigation feature trajectory feature G, use its first h moments in the time series to predict the ship's navigation trajectory position information at the h + 1 moment, which is expressed as follows: O t+1 = [Lng t+1 , Lat t+1 ​ O t+1 = f(I t ) where, I t represents the input matrix at h moments before the t-th moment, and the matrix size is h×6; O t+1 represents the longitude and latitude at the (t + 1)-th moment predicted by using I t and the matrix size is 1×2; f(·) represents the spatio-temporal mapping function of the ship trajectory prediction model with double dynamic weight allocation based on the end-to-end network; the experimental sample set is obtained by adopting the sliding window method on G, which is expressed as follows: I = {I1, …, I N-h-1} O = {O1, …, O N-h-1} Sample={I,O}={(I1,O1),…,(I N-h-1 ,O N-h-1 )} In the formula, I, O, and Sample respectively represent the input data set, output data set, and sample data set; S22. Time-varying weights of the dynamic scoring mechanism The time-varying weights of the dynamic scoring mechanism include the dynamic scoring of time series information and the dynamic weight allocation of time series information; its function is to effectively extract time series information, dynamically score the series information as it changes over time, and reallocate weights. The steps are as follows: S221. Extract sequence information and encode it Extract and encode time series information using double-layer memory storage neurons; the memory storage neurons are used to extract sequence historical information c and sequence current encoding information According to the input sequence I t and the input navigation feature variable g at time t t and the historical information c before time t t-1 and the current information at time t - 1 Update to obtain c at time t t and The update calculation of the sequence historical information c at time t is as follows: where m t , k t respectively represent the information to be memorized and forgotten at time t; is the candidate state to be added to the historical information at time t; where B is the sample batch size during the training of the ship trajectory prediction model, d is the set output information dimension; f(·) is a non-linear mapping function, and its expression is: f(x) = max{αx, e x} where α is a constant, x represents the unknown input information, and e x represents the exponential mapping of the unknown input information, where e is an irrational number with a value of 2.71828; Current coding information The update at time t is as follows: Among them, W *i 、W *f 、W *c and b * are all parameters obtained through iterative learning. Among them, b * ={b i , b f , b c , b o}; tanh(·) is a mapping function; The process of calculating t , c t-1 , to obtain is the calculation process of a single-layer memory storage neuron, and the simplified expression is as follows: Wherein, MN(·) represents a memory storage neuron, and θ = {W *i , W *f , W *c , W *o , b *} represents the set of all learning parameters; output mn represents the encoded output of the time series information of a single-layer memory storage neuron; through the encoding and extraction of the time series information by the memory storage neuron, the encoded information of the second-layer memory storage neuron is expressed as: where output′ lstm represents the output of the second-layer memory storage neuron encoding information; by extracting the feature information of each moment of the input sequence at time t, a real-time information set for each moment is obtained is expressed as: S222. Allocate weights for time series information The real-time information set at each moment in the time series information Perform dynamic weight allocation to obtain the importance of each moment relative to the output sequence and the input sequence, and filter the sequence information to obtain the dynamic weight information vector of the input sequence For extracting information to identify the association between the input sequence and the output sequence; At the decoding end, the historical and real-time coding information blocks at time t in the encoder network block are used as the historical and real-time information of the previous moment of the memory storage neurons at the decoding end of the decoder, and the input sequence I t is expanded into a matrix of 1×(h×6) as the input of the decoding information of the decoder network block; at this time, the computational expression of the double-layer memory storage neurons at the decoding end is: x dec = flatten(I t ) In the formula, and are the outputs of the double-layer memory storage neurons at the decoding end respectively; among them, and For the convenience of subsequent calculations, is transformed into a three-dimensional vector; flatten(·) is an unfolding function; In the process of dynamic weight allocation, calculate the real-time information set at the encoding end relative to the decoding end of the dynamic weight to express the correlation between the input sequence feature information and the output sequence; the calculation formula is as follows: Among them, and are the weight values for learning, and s μ is the correlation score between the task vector and the μ-th dynamic information vector, where s = {s1, s2, …, s μ}, μ ∈ {1, 2, …, 2h}; α μ is the normalized function value of the μ-th real-time information vector for the real-time decoded information at the decoding end, is a constant function value greater than 0 and is used to regulate the size of the weight mechanism. a is a constant with a value between [1, 10]. Finally, the dynamic weight of the obtained real-time information is multiplied by the new representation vector of the real-time information set at the encoding end to obtain the dynamic weight allocation value as follows: In the formula, is the output value of the time-varying weight of the dynamic scoring mechanism, For the convenience of calculation, is transformed into a two-dimensional vector; The calculation process of the time-varying weight of the dynamic scoring mechanism is simplified and expressed as follows: In the formula, is denoted as the simplified expression function of the time-varying weight of the dynamic scoring mechanism; S23. Implement the time-varying weights of the dynamic scoring mechanism at the decoding end The time-varying weight of the dynamic scoring mechanism at the decoding end includes the dynamic weight allocation of a group dynamic scoring mechanism and an information splicing and fusion block from the encoding end to the decoding end; the time-varying weight of the group dynamic scoring mechanism is used to extract significant features between time series, and the time-varying weight of the group dynamic scoring mechanism at the decoding end and the dynamic weight information vector of the input sequence at the encoding end are spliced into a new matrix information through a decoding information splicing and fusion block for ship trajectory prediction; the steps are as follows: S231. Allocate the time-varying weight of the group dynamic scoring mechanism To obtain the significant features among the sequence feature information, the time-varying weights of the group dynamic scoring mechanism are used to extract the significant features in the time series feature information; when calculating the time-varying weights of the group dynamic scoring mechanism, the input sequence feature information I t is intercepted into 6 sub-vectors respectively according to its dimensions; and each sub-vector is multiplied by the learned weight vector and to obtain and where and The calculation process of dynamic weight allocation is consistent with the calculation process of the time-varying weights of the dynamic scoring mechanism at the encoding end. The time-varying weights of the group dynamic scoring mechanism at the decoding end The calculation formula is as follows: In the formula, represents the τ-th dynamic weight allocation value, where τ ∈ {1, …, n}, is the group dynamic weight allocation value, concat(·) represents the concatenation function, and σ represents the sigmoid function, which are expressed as follows: To prevent the disappearance of network gradients, a regularization normalization intermediate layer is added. By summing the input and output of the group dynamic weight distribution and then performing layer normalization, a normalized vector is obtained. Finally, is stretched into a matrix of B×(h×m) and significant feature information is obtained through a fully connected layer. The formula is as follows: In the formula, is the significant feature information between the time series output by the fully connected layer, W fn and b fn are the learned parameters; S232. Splice and fuse the information from the encoding end to the decoding end After extracting information through the dynamic scoring mechanism at the encoding-decoding end of time series feature information, the encoding-decoding information is concatenated into a new matrix And the predicted values of the ship trajectory prediction model are output through a fully connected operation as follows: Wherein, is the predicted value at time t+1, and W out and b out are both parameters learned; S3. Evaluate the accuracy of the ship trajectory prediction model The predicted vector of the output of the last hidden layer is matched with the target vector O t+1 ; the mean square error MSE is used as the loss function to quantify the difference between the predicted value and the true value; after the construction of the ship trajectory prediction model is completed, the root mean square error RMSE is used to evaluate the ship trajectory prediction model, and the parameters, features or algorithms of the ship trajectory prediction model are continuously adjusted according to the evaluation results to achieve satisfactory results; the smaller the mean square error and the root mean square error, the better the effect of the ship trajectory prediction model, and the calculation formulas are as follows: In the formula, P is the size of the sample batch during training; finally, the error between the predicted trajectory and the true trajectory is compared according to the prediction result, and the relevant reasons are analyzed according to the ship trajectory prediction model structure.

2. The method for predicting the ship navigation trajectory based on the time-varying weight of the dynamic scoring mechanism according to claim 1, wherein: The set threshold λ in step S111 is 4 - 8 hours.

3. The ship navigation trajectory prediction method based on the time-varying weight of the dynamic scoring mechanism according to claim 1, wherein: The data sampling frequencies of the AIS system and the on-board sensors described in step S114 are 1 - 2 s and 2 - 4 s respectively.

Images