A ship navigation behavior prediction method and system based on explainable artificial intelligence

Abstract

The application relates to the technical field of ship navigation behavior prediction, and discloses a ship navigation behavior prediction method and system based on an interpretable artificial intelligence. The method comprises the following steps: acquiring ship AIS historical data and historical marine environment data in a target sea area; fusing the ship AIS historical data and the historical marine environment data based on a time scale and a space scale, and constructing a space-time data set; decoupling static variables and time-varying variables of the space-time data set, and determining historical input data, known future input data and static input data; constructing an interpretable space-time fusion neural network model based on a dynamic gate residual neural network, a space-time feature variable selection network, a Fourier analysis gate neural unit and an interpretable multi-head attention mechanism, so as to accurately predict ship navigation behaviors such as a ship trajectory, a ground speed and a ground heading, and output a navigation behavior influence factor weight visualization result, and the method has interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship navigation behavior prediction, and in particular to a ship navigation behavior prediction method and system based on explainable artificial intelligence. BACKGROUND

[0002] With the rapid development of the shipping industry, the sea traffic volume is showing explosive growth, which makes the sea navigation safety particularly important. Maritime traffic safety is not only directly related to the safety of life and property of the crew, but also has a profound impact on the stable operation of the global supply chain. Under the complex and changeable marine environmental conditions and the coupling challenges of various influencing factors, the safe and efficient navigation of ships in coastal port waters is facing severe tests.

[0003] In related technologies, the ship navigation behavior prediction method using deep learning does not comprehensively consider marine environment data, and static variables and time-varying variables are fused together and input into the model, which leads to that the prediction of ship navigation behavior cannot adapt to environmental interference and is difficult to capture complex spatio-temporal patterns. Moreover, the long-term dependence modeling capability is insufficient, thereby leading to relatively low accuracy of ship navigation behavior prediction under complex environment. In addition, the existing ship navigation behavior prediction model does not have the explainability of the prediction result, and cannot convert the prediction ability of the model into the ability to enhance the understanding and support management decision of safety guarantee.

[0004] Therefore, there is an urgent need for a ship navigation behavior prediction model with characteristics of explainability, high precision, high efficiency and high robustness, which can provide technical support for ship safety supervision and intelligent navigation of ships under complex environment. SUMMARY

[0005] In order to solve the above technical problems, the present application provides a ship navigation behavior prediction method based on explainable artificial intelligence, comprising the following steps:

[0006] Obtaining ship AIS historical data and historical marine environment data in a target sea area;

[0007] Based on time scale and space scale, the ship AIS historical data and the historical marine environment data are fused to construct a spatio-temporal data set;

[0008] The static variables and time-varying variables of the spatio-temporal data set are decoupled in feature to determine historical input data, known future input data and static input data;

[0009] An interpretable spatio-temporal fusion neural network model is constructed based on a dynamic gated residual neural network, a spatio-temporal feature variable selection network, a Fourier analysis gated neural unit and an interpretable multi-head attention mechanism; wherein the spatio-temporal feature variable selection network takes the dynamic gated residual neural network as a bottom layer architecture, the spatio-temporal feature variable selection network, the Fourier analysis gated neural unit and the interpretable multi-head attention mechanism are connected in sequence, and the interpretable spatio-temporal fusion neural network model is trained based on the historical input data, the known future input data and the static input data to obtain a ship navigation behavior prediction model;

[0010] The AIS measured data of ships in the target sea area and the measured marine environment data are input into the ship navigation behavior prediction model, and the navigation behavior influence factor weight visualization result, the ship trajectory, the prediction information of the ground speed and the ground course are output.

[0011] Further, the interpretable spatio-temporal fusion neural network model is constructed based on the dynamic gated residual neural network, the spatio-temporal feature variable selection network, the Fourier analysis gated neural unit and the interpretable multi-head attention mechanism, comprising:

[0012] The historical input data, the known future input data and the static input data are respectively constructed into corresponding dynamic gated residual neural networks;

[0013] The spatio-temporal feature variable selection network takes the dynamic gated residual neural network as a bottom layer architecture; the dynamic gated residual neural network performs noise reduction processing on the historical input data, the known future input data and the static input data, and the spatio-temporal feature variable selection network realizes weighted processing on the noise-reduced static variables and time-varying variables based on the dynamic gated residual neural network;

[0014] According to the spatio-temporal feature variable selection network, the noise-reduced historical input data, the known future input data and the static input data are weighted processed to determine the variable weights of the static variables and the time-varying variables; according to the variable weights, the target variables are determined in the static variables and the time-varying variables; the Fourier analysis gated neural unit is used to analyze the time-varying variables to determine the dependence relationship of the ship navigation behavior in the time series data;

[0015] The dynamic characteristics of the spatio-temporal data set are determined, the static variables and the dynamic characteristics are combined based on the interpretable multi-head attention mechanism to determine the multi-head attention weights of the global characteristics of the spatio-temporal data set; the target function for predicting the ship navigation behavior is determined by taking the target variables, the dependence relationship and the multi-head attention weights as calculation parameters, and the interpretable spatio-temporal fusion neural network model is constructed.

[0016] Further, the space-time feature variable selection network is used to perform weighted processing on the denoised historical input data, the known future input data and the static input data, to determine variable weights corresponding to the static variable and the time-varying variable, including:

[0017] Features of the historical input data, the known future input data and the static input data are obtained, importance scores of the features are calculated based on a space-time feature variable selection network, and the static variable and the time-varying variable are weighted processed according to the importance scores to determine variable weights corresponding to the static variable and the time-varying variable.

[0018] Further, the space-time feature variable selection network is used to perform weighted processing on the denoised historical input data, the known future input data and the static input data, to determine variable weights corresponding to the static variable and the time-varying variable, including:

[0019] The static variable and the time-varying variable are embedded and converted to generate static variable features and time-varying variable features, and the static variable features and the time-varying variable features are weighted according to importance scores corresponding to the static variable features and the time-varying variable features to determine variable weights corresponding to the static variable features and the time-varying variable features.

[0020] Further, the Fourier analysis gated neural unit is used to analyze the time-varying variable to determine a dependence relationship of the ship navigation behavior in time series data, including:

[0021] The Fourier analysis gated neural unit is used to analyze the time-varying variable to determine a time series, and a change of the time-varying variable on the time series is obtained; when the change is an abnormal change, a time point of the abnormal change is determined based on the time series, a reset gate of the Fourier analysis gated neural unit is closed, a ship navigation behavior state of the time-varying variable at a previous time point of the time point is hidden, a temporary state of the ship navigation behavior corresponding to the time point is generated, and a short-term dependence relationship of the ship navigation behavior in the time series is determined; when the change is a normal change, a ship navigation behavior state of the time-varying variable is retained based on an update gate of the Fourier analysis gated neural unit, and a long-term dependence relationship of the ship navigation behavior in the time series is determined.

[0022] Further, the space-time feature variable selection network is used to perform weighted processing on the denoised historical input data, the known future input data and the static input data, to determine variable weights corresponding to the static variable and the time-varying variable, including:

[0023] Linearly transform the static variable and the dynamic feature to generate a query matrix, a key value and a numerical matrix; perform linear transformation on the query matrix, the key value and the numerical matrix multiple times based on different parameters of an interpretable multi-head attention mechanism to determine multiple attention heads; connect each of the attention heads based on an output weight matrix to determine an attention weight of each of the attention heads; and perform weighted fusion on the attention weights to determine a multi-head attention weight of a global feature of the spatio-temporal data set.

[0024] Further, the output navigation behavior influence factor weight visualization result, the ship trajectory, the prediction information of the ground speed and the ground course include:

[0025] Output a single-step prediction result of the prediction information; and / or, output a multi-step prediction result of the prediction information; and / or output the prediction information based on quantile regression.

[0026] Another aspect of the present disclosure provides a ship navigation behavior prediction system based on an interpretable artificial intelligence, which comprises:

[0027] A data acquisition module is configured to acquire ship AIS historical data and historical marine environment data in a target sea area; a data construction module is configured to fuse the ship AIS historical data and the historical marine environment data based on a time scale and a space scale to construct a spatio-temporal data set; a data processing module is configured to decouple static variables and time-varying variables of the spatio-temporal data set to determine historical input data, known future input data and static input data; a model construction module is configured to construct an interpretable spatio-temporal fusion neural network model based on a dynamic gated residual neural network, a spatio-temporal feature variable selection network, a Fourier analysis gated neural unit and an interpretable multi-head attention mechanism; a model training module is configured to train the interpretable spatio-temporal fusion neural network model based on the historical input data, the known future input data and the static input data to obtain a ship navigation behavior prediction model; and a model prediction module is configured to input ship AIS measured data and measured marine environment data in the target sea area into the ship navigation behavior prediction model to output a navigation behavior influence factor weight visualization result, prediction information of a ship trajectory, a ground speed and a ground course.

[0028] Further, the model construction module is configured to:

[0029] The dynamic gated residual neural network constructs a spatio-temporal feature variable selection network through its built-in feature selection mechanism after noise reduction processing of the above data;

[0030] The dynamic gated residual neural network constructs a spatio-temporal feature variable selection network through its built-in feature selection mechanism after noise reduction processing of the above data;

[0031] According to the spatio-temporal feature variable selection network, the historical input data, the known future input data and the static input data after noise reduction are weighted processed to determine the variable weight corresponding to the static variable and the time-varying variable; according to the variable weight, a target variable is determined from the static variable and the time-varying variable; a Fourier analysis gated neural unit is used to analyze the time-varying variable to determine the dependency relationship of the ship navigation behavior in the time series data; the dynamic characteristics of the spatio-temporal data set are determined, the static variable and the dynamic characteristics are combined based on an interpretable multi-head attention mechanism to determine the multi-head attention weight of the global characteristics of the spatio-temporal data set; and the target variable, the dependency relationship and the multi-head attention weight are taken as calculation parameters to determine a target function for predicting the ship navigation behavior, and an interpretable spatio-temporal fusion neural network model is constructed.

[0032] The present disclosure has the following technical effects:

[0033] The ship navigation behavior prediction method based on the interpretable artificial intelligence provided by the present disclosure fuses the obtained ship AIS historical data and historical marine environment data based on the time scale and the space scale, constructs a spatio-temporal data set, comprehensively considers the influence of multiple factors such as marine environment, marine weather, and the interaction between ships on ship behavior, and provides rich data features. The static variable and the time-varying variable of the spatio-temporal data set are decoupled, and the historical input data, the known future input data and the static input data are determined, which can balance the effective information extraction and noise robustness in complex data, and support multi-dimensional interpretable analysis from feature importance to time pattern. Further, an interpretable spatio-temporal fusion neural network model is constructed based on a dynamic gated residual neural network, a spatio-temporal feature variable selection network, a Fourier analysis gated neural unit and an interpretable multi-head attention mechanism, and the interpretable spatio-temporal fusion neural network model is trained based on the historical input data, the known future input data and the static input data to obtain a ship navigation behavior prediction model, so that the ship navigation behavior prediction model can learn more extensive data features, thereby enhancing its generalization ability in different application scenarios. Further, the ship AIS measured data and the measured marine environment data in a target sea area are input into the ship navigation behavior prediction model, and the navigation behavior influence factor weight visualization result, the ship trajectory, the prediction information of the ground speed and the ground heading are output, which significantly improves the efficient and high-precision prediction of the ship navigation behavior in complex environment, and the prediction result has interpretability. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the specific embodiments or the prior art of the present application, the drawings required to be used in the specific embodiments or the prior art description will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0035] Figure 1 is the flow chart of the ship navigation behavior prediction method based on the interpretable artificial intelligence provided by the embodiment of the present application.

[0036] Figure 2 is the flow chart of the interpretable spatio-temporal fusion neural network model construction method provided by the embodiment of the present application.

[0037] Figure 3 is a schematic diagram of a ship navigation behavior prediction system based on the interpretable artificial intelligence provided by the embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present application more clear, the technical solutions of the present application will be clearly and completely described below. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the present application.

[0039] With the rapid development of the shipping industry, the sea traffic volume is growing rapidly, thereby accumulating rich ship navigation data. In order to improve the accuracy and robustness of ship navigation behavior prediction in complex environment, the present application proposes a ship navigation behavior prediction method and system based on interpretable artificial intelligence. By constructing a spatio-temporal data set of ship AIS historical data and historical marine environment data, and decoupling the static variables and time-varying variables of the spatio-temporal data set, the historical input data, known future input data and static input data are determined. According to the historical input data, known future input data and static input data, an interpretable spatio-temporal fusion neural network model is constructed based on a dynamic gated residual neural network, a spatio-temporal feature variable selection network, a Fourier analysis gated neural unit and an interpretable multi-head attention mechanism, and a ship navigation behavior prediction model is obtained, so as to realize the visualization result of the weight of the ship navigation behavior influencing factor, the ship trajectory, the accurate prediction information of the ground speed and the ground heading by using the ship navigation behavior prediction model.

[0040] Figure 1 is the flow chart of the ship navigation behavior prediction method based on the interpretable artificial intelligence provided by the embodiment of the present application. Referring to Figure 1 , specifically comprising the following steps:

[0041] In step S1, the ship AIS historical data and historical marine environment data in the target sea area are acquired.

[0042] In the embodiments of the present disclosure, the ship AIS historical data is real-time navigation information exchanged by ships and shore-based base stations through radio broadcasting, mainly used for maritime traffic management, collision avoidance and navigation safety, including static data: ship number (unique identification code), ship name, call sign, ship type (such as cargo ship, oil tanker), ship length / width, etc. Dynamic data: real-time position (latitude and longitude), speed, heading, turning rate, timestamp, etc. Voyage-related data: destination, estimated time of arrival (ETA), draft, etc. The historical marine environment data includes wind, wave, current and other data types related to the sea. For example, the wind type data includes the component of wind direction in the U direction, the component of wind direction in the V direction, the wind direction, the wind speed, etc. The wave type data includes the wavelength, the wave height, the wave period, etc. The current type data includes the component of flow direction in the U direction, the component of flow direction in the V direction, the flow direction, the flow speed, etc.

[0043] In step S2, the ship AIS historical data and the historical marine environment data are fused based on the time scale and the space scale to construct a spatio-temporal data set.

[0044] It should be noted that the ship AIS historical data in the present specification is the Automatic Identification System (AIS) historical data of the ship.

[0045] In step S3, the static variables and the time-varying variables of the spatio-temporal data set are decoupled to determine the historical input data, the known future input data and the static input data.

[0046] The historical input data contains the dynamic data in the ship AIS historical data and the historical marine environment data, which provides historical information about how the ship trajectory and the ship speed change with time and environment. These information, through dynamic selection, local modeling, long-term dependence capture and multi-source collaborative mechanism, form a key information source that can better enable the model to understand the complex patterns of spatio-temporal data and quantify uncertainty.

[0047] The known future input data refers to the time-varying variables whose future values are explicitly known at the time of prediction, including the destination, the estimated time of arrival (ETA), the draft, etc. Voyage-related data. By combining with the historical input data, the potential patterns of future ship behavior can be captured more accurately.

[0048] The static input data is a variable that does not change over time, such as the ship name, call sign, ship type, ship length / width, etc. This variable is not only a simple feature supplement, but also a core component of the model through the closed-loop mechanism of encoding-fusion-interpretation, which captures entity differences, optimizes dynamic processing, and improves interpretability.

[0049] In step S4, an interpretable spatio-temporal fusion neural network model is constructed based on a dynamic gated residual neural network, a spatio-temporal feature variable selection network, a Fourier analysis gated neural unit, and an interpretable multi-head attention mechanism. The spatio-temporal feature variable selection network uses the dynamic gated residual neural network as the underlying architecture. The spatio-temporal feature variable selection network, the Fourier analysis gated neural unit, and the interpretable multi-head attention mechanism are connected in sequence. The interpretable spatio-temporal fusion neural network model is trained based on historical input data, known future input data, and static input data to obtain a ship navigation behavior prediction model.

[0050] The output of the dynamic gated residual neural network is the input of the spatio-temporal feature variable selection network. The former focuses on noise suppression, and the latter focuses on feature importance quantification. They are different functional modules connected in series.

[0051] The spatio-temporal feature variable selection network is used to automatically identify and weight the importance of input features and select key variables. The dynamic gated residual neural network controls information flow through a dynamic gating mechanism (such as Tanh activation) to enhance non-linear feature extraction capabilities. The Fourier analysis gated neural unit is a lightweight recurrent neural network for processing sequence data, which controls information transmission through update and reset gates. The interpretable multi-head attention mechanism is essentially an information weight distribution model that filters irrelevant information while focusing on key information, allowing the model to adapt to large-scale sequence data learning.

[0052] Further, the interpretable spatio-temporal fusion neural network model constructed by the above network algorithm is trained based on historical input data, known future input data, and static input data to obtain a ship navigation behavior prediction model.

[0053] In step S5, the ship AIS measured data and measured ocean environment data in the target sea area are input into the ship navigation behavior prediction model, and the navigation behavior influence factor weight visualization result, ship trajectory, prediction information of ground speed and ground course are output.

[0054] In the embodiments of the present disclosure, a single-step prediction result of the prediction information can be output; and / or, a multi-step prediction result of the prediction information can be output; and / or, the prediction information can be output based on quantile regression.

[0055] where quantile prediction is a method to quantify the predicted future distribution by outputting multiple quantile values, which can not only provide single-point estimates such as the median, but also characterize the uncertainty range of the prediction, as shown in the following formula:

[0056]

[0057] wherein, denotes the predicted value at quantile q , time point t , and prediction step τ denotes the weight coefficient vector corresponding to quantile q , which maps high-dimensional features to a scalar,

[0058] denotes the bias term of quantile q , which adjusts the prediction reference value, denotes the feature representation corresponding to time t and prediction step τ , R represents the set of real numbers, and d represents the dimension of the feature vector.

[0059] At quantile q , the predicted value of the future t step at time point τ is obtained by linear transformation of the feature representation . Among them, the prediction only covers the future time range ∈{1,…, τ τ max , wherein denotes the maximum prediction step.

[0060] The ship navigation behavior prediction model can also predict key parameters such as future position, energy consumption, and arrival time of the ship.

[0061] ​​The ship navigation behavior prediction method based on the interpretable artificial intelligence provided by the present disclosure is based on time scale and space scale, fuses the obtained ship AIS historical data and historical marine environment data, constructs a space-time data set, comprehensively considers the influence of multiple factors such as marine environment, marine weather, and the interaction between ships on ship behavior, and provides rich data features. The static variables and time-varying variables of the space-time data set are decoupled in feature, the historical input data, known future input data and static input data are determined, the effective information extraction in complex data and the noise robustness can be balanced, and multi-dimensional interpretable analysis from feature importance to time pattern is supported. Further, based on the dynamic gated residual neural network, the space-time feature variable selection network, the Fourier analysis gated neural unit and the interpretable multi-head attention mechanism, an interpretable space-time fusion neural network model is constructed, and based on the historical input data, the known future input data and the static input data, the interpretable space-time fusion neural network model is trained to obtain a ship navigation behavior prediction model, so that the ship navigation behavior prediction model can learn more extensive data features, thereby enhancing its generalization ability in different application scenarios. Further, the ship AIS measured data and measured marine environment data in the target sea area are input into the ship navigation behavior prediction model, and the attention weight visualization result, the ship trajectory, the prediction information of the ground speed and the ground heading are output, thereby significantly improving the prediction accuracy of the model in complex environment.

[0062] The present disclosure constructs a space-time data set associated with ships and marine environment, which contains rich data features and lays a solid foundation for ship trajectory, ground speed and ground heading prediction in complex and variable marine environment. Based on this, the present disclosure further constructs an interpretable space-time fusion neural network model based on a dynamic gated residual neural network, a space-time feature variable selection network, a Fourier analysis gated neural unit and an interpretable multi-head attention mechanism, Figure 2 is a flowchart of the method for constructing the interpretable space-time fusion neural network model provided by the embodiments of the present disclosure. The following embodiments of the present disclosure will be combined with Figure 2 The construction of the interpretable space-time fusion neural network model will be described in detail. Specifically, it includes the following implementation:

[0063] The dynamic gated residual neural network is constructed for historical input data, known future input data and static input data respectively; the dynamic gated residual neural network is used for noise reduction processing of the historical input data, the known future input data and the static input data respectively, a spatio-temporal feature variable selection network is constructed; the historical input data, the known future input data and the static input data after noise reduction are weighted processed according to the spatio-temporal feature variable selection network, variable weights corresponding to static variables and time-varying variables are determined; the target variable is determined from the static variables and the time-varying variables according to the variable weights; the time-varying variable is analyzed by using the Fourier analysis gated neural unit to determine the dependence relationship of the ship navigation behavior in the time series data; the dynamic characteristics of the spatio-temporal data set are determined, the static variables and the dynamic characteristics are combined based on the interpretable multi-head attention mechanism, and the multi-head attention weight of the global feature of the spatio-temporal data set is determined; the target function used for predicting the ship navigation behavior is determined by taking the target variable, the dependence relationship and the multi-head attention weight as the calculation parameters, and the interpretable spatio-temporal fusion neural network model is constructed.

[0064] The present disclosure further describes an embodiment of the spatio-temporal feature variable selection network.

[0065] In the present disclosure, the historical input data, the known future input data and the static input data correspond to the construction of the spatio-temporal feature variable selection network with different weights, and each spatio-temporal feature variable selection network at the bottom layer uses the dynamic gated residual neural network to realize the filtering function. Thus, the features of the historical input data, the known future input data and the static input data are obtained, and the importance score of each feature is calculated based on the dynamic gated residual neural network at the bottom layer of each spatio-temporal feature variable selection network. It should be noted that the "feature" in the feature importance score specifically refers to various specific attribute information extracted from the historical input data, the known future input data and the static input data, which can be divided into static variable features and time-varying variable features. The static variable features are from the static input data, which refer to the inherent attribute features of the ship that do not change with time, such as ship number, ship name, ship type, cargo ship, oil tanker, ship length, ship width, etc. These features are inherent and stable attribute information of the ship. The time-varying variable features are from the historical input data and the known future input data, which refer to the attribute features that change dynamically with time.

[0066] Further need to explain is that the wind, wave, flow and other environmental factors interference received by the ship in the navigation process, and the dynamic changes of the loading quality and the ship speed will cause the uncertainty of the model parameters, and increase the nonlinear characteristics of the ship motion. Therefore, in order to accurately measure the correlation between the multi-dimensional feature input of the space-time data set and the target variable, a dynamic gated residual neural network is proposed; wherein the dynamic gated residual neural network includes exponential linear unit and gated linear unit. Before the exponential linear unit layer, we introduce the dynamic hyperbolic tangent function, realize the dynamic calibration of the characteristics by optimizing the parameter a end to end, and have the normalization effect and the calculation efficiency.

[0067] In the present disclosure, the implementation of the dynamic gated residual neural network of the network bottom layer based on each space-time feature variable selection is as follows:

[0068] According to the dynamic scaling feature range of the characteristics, the numerical range of the optimized characteristics is adaptively adjusted. When the linear transformation result of the characteristics in the numerical range is much larger than zero, the exponential linear unit of the positive activation dynamic gated residual neural network is determined. Based on the exponential linear unit, the negative value in the numerical range is suppressed, the dynamic characteristics are determined, and the nonlinear relationship between the dynamic characteristics and the marine environment is determined. Thus, the gated linear unit weight of the dynamic gated residual neural network is increased, and the contribution degree of the nonlinear relationship between the dynamic characteristics and the marine environment is determined based on the gated linear unit weight. Further, the importance score of each characteristic is calculated according to the contribution degree.

[0069] The static variable and the time-varying variable are embedded and converted to generate static variable characteristics and time-varying variable characteristics. According to the importance scores corresponding to the static variable characteristics and the time-varying variable characteristics, the static variable characteristics and the time-varying variable characteristics are weighted to determine the variable weights corresponding to the static variable characteristics and the time-varying variable characteristics. And, according to the variable weights, the target variable is determined in the static variable and the time-varying variable.

[0070] It should be noted that, as described above, the selection of the target variable is not to exclude unimportant variables, but to highlight important variables by weighting. The greater the weight, the higher the importance. The reference formula is as follows:

[0071]

[0072]

[0073]

[0074] In the formula, represents the variable weight; represents the feature of the current time step original input; represents the static input encoding; represents a dynamic gated residual neural network designed for variable selection tasks, used to calculate the original importance score score of each feature; represents normalizing the original score into a probability distribution; represents the first component of the features of the original input; j represents the first component of the features of the original input; represents an independent dynamic gated residual neural network, which performs nonlinear transformation and enhancement on each feature. represents the transformed feature. represents the importance weight of the feature to the current prediction task, represents the fused variable feature, represents the first component, represents the total number of feature space components, represents the state of the jth component at time t.

[0075] In the present disclosure, it is also necessary to analyze the time-varying variable by using the Fourier analysis gated neural unit to determine the dependence of the ship navigation behavior in the time series data. The Fourier analysis gated neural unit with simpler structure and fewer parameters can be used to analyze the time-varying variable. The Fourier analysis gated neural unit includes an update gate (Update Gate) and a reset gate (Reset Gate) and a Fourier analysis neural unit.

[0076] wherein the reset gate controls the information of the previous time with a certain probability, which is beneficial to obtain the short-term dependence in the time series data of the ship behavior, and the calculation formula is:

[0077]

[0078]

[0079] In the formula, represents the reset gate, represents the definition symbol; represents the hidden state at the previous time; represents the variable feature at the current time, represents the reset gate weight matrix, represents the Fourier reset gate analysis neural unit. DyT represents the dynamic layer, 、 represents the interpretable parameter, σ represents the activation function, cos represents the cosine function, and sin represents the sine function.

[0080] The update gate controls the state information of the previous time to be substituted into the current state, which is helpful to obtain the long-term dependence in the time series data of the ship behavior analysis, and the calculation formula is:

[0081]

[0082]

[0083] In the formula, represents an update gate, represents an update gate weight matrix, represents a Fourier update gate analysis neural unit, represents an interpretable parameter.

[0084] The candidate hidden state controls the hidden state by using a reset gate, and the reset gate determines whether to discard the past hidden state irrelevant to the future. The calculation formula is:

[0085]

[0086]

[0087] In the formula, represents a hidden state, represents a candidate hidden state weight matrix, represents a Fourier reset gate hidden state analysis neural unit, and W represents an input weight matrix, represents a Hadamard product, represents an interpretable parameter.

[0088] The hidden state of the jth layer at the time step t uses an update gate to update the previous jth layer hidden state and the hidden state . The update gate determines the importance of the past hidden state at the current time. It can solve the problem of gradient decay and better obtain the long-range dependence relationship in hydrological time series data.

[0089]

[0090] The specific embodiment is that the Fourier analysis gated neural unit is used to analyze the time characteristics in the time-varying variable characteristics, and the time sequence is determined. Thus, the change of the variable characteristics on the time sequence is obtained.

[0091] When the change is an abnormal change, the time point of the abnormal change is determined based on the time sequence, the reset gate of the Fourier analysis gated neural unit decoder is closed, the ship navigation behavior state of the variable characteristics at the previous time point of the hidden time point is generated, the temporary state of the ship navigation behavior corresponding to the time point is generated, and the short-term dependence relationship of the ship navigation behavior in the time sequence is determined.

[0092] ​​When the change is a normal change, the update gate of the Fourier analysis gated neural unit decoder is based on the ship navigation behavior state of the reserved variable feature, and the long-term dependence of the ship navigation behavior in the time sequence is determined.

[0093] The multi-head attention mechanism in the present disclosure is essentially an information weight distribution model, which can enable the model to filter irrelevant information while focusing on key information, thereby adapting to learning of large-scale sequence data and determining the multi-head attention weight of the global feature of the spatiotemporal data set.

[0094] Further, the static variable and the dynamic feature are linearly transformed to generate a query matrix, a key value and a numerical value. The query matrix, the key value and the numerical value are linearly transformed multiple times based on different parameters of the interpretable multi-head attention mechanism to determine multiple attention heads. Each attention head is connected based on a weight matrix to determine the attention weight of each attention head. The attention weights are weighted and fused to determine the multi-head attention weight of the global feature of the spatiotemporal data set.

[0095] Further, the attention head, i.e., the attention branch, can be processed by a scaled dot product attention (SDPA), and the processing formula is as follows:

[0096]

[0097] In the formula, represents the attention branch, Q represents the query matrix, K represents the key value, V represents the numerical value matrix, d is the number of hidden layer nodes, and is the dimension of the Q and K vectors, represents the dimension of the key value K.

[0098] In the multi-head attention mechanism, the query matrix Q, the key value K and the numerical value matrix V are linearly transformed multiple times by different parameters, and the relevant information of the water level feature of the channel can be learned from different dimensions and representation subspaces. The transformed results are input into the SDPA to obtain different attention outputs.

[0099] Therefore, the attention output of each attention head is shown in the formula.

[0100]

[0101] wherein, , and represent linear transformation parameters, and the different parameter linear transformations of Q, K and V represent the essence of "multi-head".

[0102] n is the input length, and after n times of scaled dot product attention calculation of different parameters, the output weight matrix is used to connect n parallel heads to The multi-head attention weights are obtained through linear transformation, as shown in the following formula.

[0103]

[0104] Indicates the multi-head attention weight. This represents the output weight matrix. This represents a vector concatenation operation.

[0105]

[0106] This indicates that the multi-head attention weights can be explained. This represents the weight values ​​for a uniform distribution. Used for the final linear mapping, Represents a set of matrices.

[0107] Based on and Figure 1 Using the same principle as the method shown, this disclosure also provides a ship navigation behavior prediction system based on interpretable artificial intelligence. Figure 3 This is a schematic diagram of a ship navigation behavior prediction system based on interpretable artificial intelligence, provided in an embodiment of the present invention. See also... Figure 3 The system 300 may include:

[0108] Data acquisition module 301 is used to acquire historical AIS data and historical marine environment data of ships within the target sea area; data construction module 302 is used to fuse the historical AIS data and historical marine environment data based on time and spatial scales to construct a spatiotemporal dataset; data processing module 303 is used to decouple the static variables and time-varying variables of the spatiotemporal dataset to determine historical input data, known future input data, and static input data; model construction module 304 is used to construct a model based on a dynamic gated residual neural network, a spatiotemporal feature variable selection network, and Fourier transform... An interpretable spatiotemporal fusion neural network model is constructed using gating neural units and an interpretable multi-head attention mechanism. A model training module 305 is used to train the interpretable spatiotemporal fusion neural network model based on the historical input data, the known future input data, and the static input data to obtain a ship navigation behavior prediction model. A model prediction module 306 is used to input the ship's AIS measured data and measured marine environment data within the target sea area into the ship navigation behavior prediction model, and output the visualization results of the weights of navigation behavior influencing factors, ship trajectory, and predicted information on ground speed and ground heading.

[0109] In the present disclosure, the model construction module 304 is configured to acquire features of the historical input data, the known future input data and the static input data; calculate importance scores of the features based on a spatio-temporal feature variable selection network; and perform weighting processing on the static variable and the time-varying variable according to the importance scores to determine variable weights corresponding to the static variable and the time-varying variable.

[0110] In the present disclosure, the model construction module 304 is configured to perform embedding conversion on the static variable and the time-varying variable to generate static variable features and time-varying variable features; and perform weighting on the static variable features and the time-varying variable features according to importance scores corresponding to the static variable features and the time-varying variable features to determine variable weights corresponding to the static variable features and the time-varying variable features.

[0111] In the present disclosure, the model construction module 304 is configured to:

[0112] construct corresponding dynamic gated residual neural networks for the historical input data, the known future input data and the static input data respectively; the spatio-temporal feature variable selection network has a dynamic gated residual neural network as a bottom layer architecture; the dynamic gated residual neural network performs noise reduction processing on the historical input data, the known future input data and the static input data, and the spatio-temporal feature variable selection network performs weighting processing on the noise-reduced static variable and time-varying variable based on the dynamic gated residual neural network; the dynamic gated residual neural network performs noise reduction processing on the historical input data, the known future input data and the static input data respectively to construct the spatio-temporal feature variable selection network; the spatio-temporal feature variable selection network performs weighting processing on the noise-reduced historical input data, known future input data and static input data to determine variable weights corresponding to the static variable and the time-varying variable; according to the variable weights, a target variable is determined from the static variable and the time-varying variable; a Fourier analysis gated neural unit is used to analyze the time-varying variable to determine a dependency relationship of the ship navigation behavior in the time series data; a dynamic feature of the spatio-temporal data set is determined, the static variable and the dynamic feature are combined based on an interpretable multi-head attention mechanism to determine a multi-head attention weight of the global feature of the spatio-temporal data set; the target variable, the dependency relationship and the multi-head attention weight are used as calculation parameters to determine a target function for predicting the ship navigation behavior, and an interpretable spatio-temporal fusion neural network model is constructed.

[0113] In the present disclosure, the model construction module 304 is configured to analyze the time-varying variable by using a Fourier analysis gated neural unit to determine a time sequence, acquire a change of the time-varying variable on the time sequence, when the change is an abnormal change, determine a time point of the abnormal change based on the time sequence, close a reset gate of the Fourier analysis gated neural unit, implicitly a ship navigation behavior state of the time-varying variable at a previous time point of the time point, generate a temporary state of ship navigation behavior corresponding to the time point, and determine a short-term dependence relationship of ship navigation behavior in the time sequence; when the change is a normal change, retain a ship navigation behavior state of the time-varying variable based on an update gate of the Fourier analysis gated neural unit, and determine a long-term dependence relationship of ship navigation behavior in the time sequence.

[0114] In the present disclosure, the model construction module 304 is configured to perform linear transformation on the static variable and the dynamic feature to generate a query matrix, a key value and a numerical matrix, perform linear transformation on the query matrix, the key value and the numerical matrix multiple times based on different parameters of an interpretable multi-head attention mechanism to determine multiple attention heads, connect each of the attention heads based on an output weight matrix to determine an attention weight of each of the attention heads, and perform weighted fusion on the attention weights to determine a multi-head attention weight of a global feature of a spatio-temporal data set.

[0115] In the present disclosure, the model prediction module 306 is configured to output a single-step prediction result of the prediction information, and / or output a multi-step prediction result of the prediction information, and / or output the prediction information based on quantile regression.

[0116] Through simulation verification, the top 6 features in weight ranking of the ship in normal navigation are longitude, latitude, ground speed, wind direction, flow direction and flow rate. When the ship has abnormal behavior, the top 6 features in weight ranking are longitude, latitude, ground heading, flow direction, flow rate and wind direction.

[0117] The target variable is the trajectory, ground speed and ground heading of the ship in the next 1-6 hours, and the target function adopts a weighted loss function:

[0118] , wherein is determined through cross-validation, wherein MSE represents mean square error, a represents the weight of trajectory prediction, β represents the weight of speed prediction, γ represents the weight of heading prediction, and λ represents the weight of a regularization term. 轨迹 MSE 航速 represents the mean square error of the speed, and MSE 航向 represents the mean square error of the heading.

[0119] The results after simulation verification include: superimposed comparison of the predicted trajectory of the ship and the actual trajectory; risk area labeling and early warning information, which directly shows the influence weight of each factor on the navigation behavior.

[0120] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting ship navigation behavior based on interpretable artificial intelligence, characterized in that, Includes the following steps: Acquire historical AIS data and historical marine environmental data of vessels within the target sea area; Based on time and space scales, the historical AIS data of the ships and the historical marine environment data are fused to construct a spatiotemporal dataset; The static and time-varying variables of the spatiotemporal dataset are decoupled by feature analysis to determine historical input data, known future input data, and static input data. An interpretable spatiotemporal fusion neural network model is constructed based on a dynamic gated residual neural network, a spatiotemporal feature variable selection network, a Fourier analysis gated neural unit, and an interpretable multi-head attention mechanism. The spatiotemporal feature variable selection network uses a dynamic gated residual neural network as its underlying architecture. The spatiotemporal feature variable selection network, the Fourier analysis gated neural unit, and the interpretable multi-head attention mechanism are sequentially connected. The interpretable spatiotemporal fusion neural network model is trained based on historical input data, known future input data, and static input data to obtain a ship navigation behavior prediction model. The ship navigation behavior prediction model is input into the ship AIS measured data and measured marine environment data in the target sea area, and outputs the visualization results of the weight of navigation behavior influencing factors, ship trajectory, and predicted information of ground speed and ground heading. The interpretable spatiotemporal fusion neural network model constructed based on a dynamic gated residual neural network, a spatiotemporal feature variable selection network, a Fourier analysis gated neural unit, and an interpretable multi-head attention mechanism includes: The historical input data, the known future input data, and the static input data are used to construct corresponding dynamic gated residual neural networks. The spatiotemporal feature variable selection network uses a dynamically gated residual neural network as its underlying architecture. The dynamically gated residual neural network performs noise reduction on historical input data, known future input data, and static input data. The spatiotemporal feature variable selection network implements weighted processing of the denoised static variables and time-varying variables based on the dynamically gated residual neural network. Based on the spatiotemporal feature variables, the network selects to perform weighted processing on the denoised historical input data, the known future input data, and the static input data to determine the variable weights corresponding to the static variables and the time-varying variables; Based on the variable weights, the target variable is determined from the static variables and the time-varying variables; Fourier analysis-gated neural units are used to analyze the time-varying variables and determine the dependencies of ship navigation behavior in time-series data; The dynamic features of the spatiotemporal dataset are determined, and the static variables and the dynamic features are combined based on an interpretable multi-head attention mechanism to determine the multi-head attention weights of the global features of the spatiotemporal dataset. Using the target variable, the dependency relationship, and the multi-head attention weight as calculation parameters, an objective function for predicting ship navigation behavior is determined, and an interpretable spatiotemporal fusion neural network model is constructed.

2. The ship navigation behavior prediction method based on interpretable artificial intelligence according to claim 1, characterized in that, The step of selecting the network based on the spatiotemporal feature variables to perform weighted processing on the denoised historical input data, the known future input data, and the static input data, and determining the variable weights corresponding to the static variables and the time-varying variables, includes: Acquire the features of the historical input data, the known future input data, and the static input data; The importance score of the feature is calculated based on the spatiotemporal feature variable selection network. Based on the importance scores, the static variables and the time-varying variables are weighted to determine the variable weights corresponding to the static variables and the time-varying variables.

3. The ship navigation behavior prediction method based on interpretable artificial intelligence according to claim 2, characterized in that, The step of weighting the static variable and the time-varying variable according to the importance score to determine the variable weights corresponding to the static variable and the time-varying variable includes: The static variables and the time-varying variables are embedded and transformed to generate static variable features and time-varying variable features; Based on the importance scores corresponding to the static variable features and the time-varying variable features, the static variable features and the time-varying variable features are weighted to determine the variable weights corresponding to the static variable features and the time-varying variable features.

4. The ship navigation behavior prediction method based on interpretable artificial intelligence according to claim 1, characterized in that, The step of using a Fourier analysis-gated neural unit to analyze the time-varying variables and determine the dependencies of ship navigation behavior in the time-series data includes: Fourier analysis-gated neural units are used to analyze the time-varying variables and determine the time series; Obtain the changes of the time-varying variable over the time series; When the change is an abnormal change, the time of the abnormal change is determined based on the time series, the reset gate of the Fourier analysis gated neural unit is closed, the ship navigation behavior state of the time-varying variable is implied to be the time before the time of the abnormal change, a temporary state of ship navigation behavior corresponding to the time is generated, and the short-term dependency of ship navigation behavior in the time series is determined. When the change is a normal change, the ship navigation behavior state of the time-varying variable is retained based on the update gate of the Fourier analysis gated neural unit to determine the long-term dependency of ship navigation behavior in the time series.

5. The ship navigation behavior prediction method based on interpretable artificial intelligence according to claim 1, characterized in that, The method of combining the static variables and dynamic features based on an interpretable multi-head attention mechanism to determine the multi-head attention weights of the global features of the spatiotemporal dataset includes: A linear transformation is performed on the static variables and the dynamic features to generate a query matrix, a key-value matrix, and a numerical matrix. Multiple attention heads are determined by performing multiple linear transformations on the query matrix, the key value, and the numerical matrix based on different parameters of an interpretable multi-head attention mechanism. Each attention head is connected based on the output weight matrix to determine the attention weight of each attention head; The attention weights are weighted and fused to determine the multi-head attention weights of the global features of the spatiotemporal dataset.

6. The ship navigation behavior prediction method based on interpretable artificial intelligence according to claim 1, characterized in that, The output includes the visualization results of the weights of the factors influencing navigation behavior, the ship's trajectory, and the predicted information of its speed and heading to the ground. Output the single-step prediction result of the prediction information; and / or Output the multi-step prediction results of the prediction information; and / or The predicted information is output based on quantile regression.

7. A ship navigation behavior prediction system based on interpretable artificial intelligence, characterized in that, The system includes: The data acquisition module is used to acquire historical AIS data and historical marine environmental data of ships within the target sea area; The data construction module is used to fuse the ship's historical AIS data and the historical marine environment data based on time and space scales to construct a spatiotemporal dataset; The data processing module is used to decouple the static variables and time-varying variables of the spatiotemporal dataset by feature decoupling, and to determine historical input data, known future input data and static input data; The model building module is used to construct interpretable spatiotemporal fusion neural network models based on dynamic gated residual neural networks, spatiotemporal feature variable selection networks, Fourier analysis gated neural units, and interpretable multi-head attention mechanisms. The model training module is used to train the interpretable spatiotemporal fusion neural network model based on the historical input data, the known future input data, and the static input data to obtain a ship navigation behavior prediction model. The model prediction module is used to input the ship's AIS measured data and measured marine environment data in the target sea area into the ship navigation behavior prediction model, and output the visualization results of the weights of navigation behavior influencing factors and the prediction information of ship trajectory, speed to land and heading to land. The model building module is used for: Construct corresponding dynamic gated residual neural networks for the historical input data, the known future input data, and the static input data, respectively; After the dynamic gated residual neural network performs noise reduction on the above data, it constructs a spatiotemporal feature variable selection network through its built-in feature selection mechanism. Based on the spatiotemporal feature variables, the network selects to perform weighted processing on the denoised historical input data, the known future input data, and the static input data to determine the variable weights corresponding to the static variables and the time-varying variables; Based on the variable weights, the target variable is determined from the static variables and the time-varying variables; Fourier analysis-gated neural units are used to analyze the time-varying variables to determine the dependencies of ship navigation behavior in time-series data; the dynamic features of the spatiotemporal dataset are determined, and the static variables and dynamic features are combined based on an interpretable multi-head attention mechanism to determine the multi-head attention weights of the global features of the spatiotemporal dataset; Using the target variable, the dependency relationship, and the multi-head attention weight as calculation parameters, an objective function for predicting ship navigation behavior is determined, and an interpretable spatiotemporal fusion neural network model is constructed.

Images