A ship wake prediction method with self-adaptive time interval based on mirror gate network

Abstract

A self-adaptive time interval ship wake prediction method based on mirror gate network is provided. [Solution] The method includes the steps of building a mirror gate network, extracting trajectory feature points at equal time intervals, constructing input trajectory data, and building a self-adaptive trajectory prediction model with variable long-term intervals. A new type of mirror gate network is built, and the mirror gates are trained from both ends of the data respectively, and the time distance is adjusted, so that when the error obtained by the model meets the convergence condition, the training of the trajectory prediction model is achieved. [Effects] The mirror network constructed by the present invention achieves higher trajectory prediction accuracy because the learning capability from both ends allows the model to spontaneously discover and utilize hidden modes in the data to fully utilize its own autonomous learning and adaptation capabilities.

Description

[Technical field]

[0001] The present invention relates to the technical field of ship wake prediction, and in particular to a ship wake prediction method with self-adaptive time interval based on mirror gate network. [Background technology]

[0002] In recent years, with the continuous development of artificial intelligence technology, deep learning has made great strides in the field of ship trajectory prediction. Representative trajectory prediction methods include Artificial Neural Network (ANN), Recurrent Neural Network (RNN), Long Short Term Memory (LSTM), Gate Recurrent Unit (GRU), etc. Besides, Temporal Convolutional Network (TCN) has attracted more and more attention due to its advantages in time series modeling and feature capture. Then, the proposed attention mechanism became a new trend, and this state-of-the-art method has gradually been applied to the field of trajectory prediction and has led to many variants. The introduction of attention mechanism makes the model more useful for extracting important information in the data.

[0003] Patent Documents 1 and 2 disclose a predictive model that uses a loss function for model learning in order to minimize errors in trajectory prediction. However, it is difficult to capture dependencies and dynamic patterns before and after the trajectory, a loss function is required to evaluate the superiority or inferiority of the predictive model, and there are problems such as generalized learning and lack of adaptability. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] China Patent Application Publication No. 117093889 [Patent Document 2] China Patent Application Publication No. 116629116 Summary of the Invention [Problem to be solved by the invention]

[0005] The present invention has been made in consideration of the above problems existing in the prior art, and aims to propose a ship wake prediction method with self-adaptive time intervals based on a mirror gate network, which can understand and capture dependencies and dynamic patterns in time series data. The present invention is the first to propose a neural network technology based on mirror learning, adopting a method of self-learning of the model and self-matching of the results, moving in both directions until the mirrors overlap from both ends, thereby completing the training of the prediction model without evaluating the loss function. [Means for solving the problem]

[0006] In order to achieve the above object, the solution of the present invention is as follows.

[0007] A method for predicting ship trajectories with self-adaptive time intervals based on a mirror gate network, which is one embodiment of the present invention, includes the following steps S1 to S4: S1, the process of constructing a mirror gate network; S2, a step of extracting trajectory characteristic points at equal time intervals; S3, constructing input track data; and S4, constructing a self-adaptive trajectory prediction model with variable long time interval; In S1, the mirror gate network comprises a forward gate unit and a reverse gate unit; S1 includes the following steps S11 to S14, S11, setting the forward gate unit of the mirror gate network; In S11, the forward gate unit includes a forward merging gate, a forward splitting gate and a forward regenerating gate, and the forward gate unit is used to control the input and output of track information; S11 includes the following steps S111 to S113, S111, the forward merging gate is in the current track information input state x t and the unit status of the previous gate unit

number

number

number

number

number

[0008] S112, the forward branch gate converts the information passing through the forward merge gate into the unit state of the previous gate unit.

number

number

number

number

number

number

number

[0009] S113, the forward regeneration gate converts the information passed through the forward merging gate into the current track information input state x t and the unit state of the current unit

number

number

number

number

number

number

number

[0010] S12, setting the reverse gate unit of the mirror gate network; In S12, the reverse gate unit includes a reverse merging gate, a reverse splitting gate and a reverse regenerating gate, and the reverse gate unit performs a reverse operation on the input information; S12 includes the following steps S121 to S123, S121, the reverse merging gate is the current track information input state x t and the unit state of the next gate unit

number

number

number

number

number

[0011] S122, the reverse branch gate converts the information passing through the reverse merge gate into the unit state of the next gate unit according to the following formula:

number

number

number

number

number

number

number

[0012] S123, the reverse regeneration gate converts the information passed through the reverse merging gate into the current track information input state x t and the unit state of the current unit

number

number

number

number

number

number

number

[0013] S13, the forward gate units are sequentially connected in the forward direction, and the backward gate units are sequentially connected in the backward direction, and the states of the forward gate units and the backward gate units at the same time distance are merged to construct the mirror gate network, and the calculation formula is as follows:

number

number

[0014] S14, perform calculations in the mirror gate network to obtain the following predicted output information:

number

number

[0015] S2 includes the following steps S21 to S24: S21, by extracting the feature points of the self-adaptive time interval, the wake P in the AIS data collected in the same sea area is k is defined and expressed as a track feature vector, and the calculation formula is as follows: P k ={p1 k ,p2 k ,...,p i k ,...,p l k} Here, p i k is the trajectory P of the kth ship in the sea area at timestamp i. k represents the track point of p i k The characteristics of are expressed by the following formula: p i k =(lon ik ,lat i k ,cog i k ,sog i k ) Here, lon i k represents the ship's longitude value, and lat i k represents the ship's latitude value, and cog i k represents the ship's course over the ground, and sog i k represents the ship's ground speed, The self-adaptation time interval is defined as t and expressed as follows: t=(t1,t2,...,t i ,...,t n ) Here, t i , P k represents the time distance of the i-th track point in the

[0016] S22, interval [t i ,t i+1 ], we define an interpolation function expressed by the following formula, X i (t i )=a i t i 3 +b i t i 2 +c i t i +d i X i+1 (t i+1 )=a i t i+1 3 +b i t i+1 2 +c i t i+1 +d i Taking the first derivative of the above function gives the following equation: X i (t i )′=3a i t i 2+2b i t i +c i X i+1 (t i+1 )′=3a i t i+1 2 +2b i t i+1 +c i Combining the interpolation function and its derivative and solving the equation gives a i , b i , c i and d i Find the value of .

[0017] S23, X i (t i )' and X i+1 (t i+1 )′, where all intermediate points of the interpolation function are calculated using M j (t j ,x j ) and M according to the following formula: j The first-order difference quotient of the left and right coordinate points of j and k j+1 It is expressed as

number

number

number

[0018] S24: Substituting the time series obtained by equally dividing the track recording period into an interpolation function of the longitude and latitude in each period, feature points extracted at equal time intervals of the track are obtained.

[0019] S3 includes the following steps S31 to S32: S31. In order to retain the time series characteristics, dynamic patterns and correlation information of the track data and allow the model to better learn the change / evolution process of the track data, a moving time window process is performed to establish a moving time window, where the length of the moving time window is set as L and the moving distance is set as L / q, where q is an integer that can be flexibly selected, and the smaller q is, the stronger the correlation between the data before and after the data is.

[0020] S32, input the AIS data that has undergone feature extraction and moving time window processing at self-adaptive time intervals into the mirror gate network, respectively train the forward gate unit and the backward gate unit, and train and predict one or more trajectories by the mirror gate network, and input data I t and output data O t is expressed by the following formula:

number

[0021] In S4, each input data is trained through N forward gate units and N backward gate units, and a self-adaptive variable length N is determined according to the final result of the mirror gate network, and the optimal trajectory prediction result is obtained. Effect of the Invention

[0022] The present invention has the following beneficial effects compared to the prior art:

[0023] 1. The present invention establishes a new type of mirror gate network, which allows the mirror gate to learn from both ends of the data, and adjusts the time distance, so that when the error obtained by the model meets the convergence condition, the trajectory prediction model can be trained. This new gate structure can deeply capture the dependency and dynamic pattern before and after the wake.

[0024] 2. The mirror network constructed by the present invention has the ability to learn from both ends, allowing the model to spontaneously discover and utilize hidden modes in the data to fully utilize its autonomous learning and adaptation capabilities, thereby achieving higher trajectory prediction accuracy.

[0025] 3. The present invention uses the method of self-adaptive time interval feature extraction to achieve self-adaptive alignment of different track data, and can also construct input data for multiple track prediction. [Brief description of the drawings]

[0026] [Figure 1] FIG. 1 is a schematic flow chart of the present invention. [Diagram 2] FIG. 1 is a schematic diagram showing a forward gate and a reverse gate; [Diagram 3] FIG. 2 is a schematic diagram of a mirror gate network. [Figure 4] FIG. 1 illustrates a schematic diagram of the extraction of equally spaced AIS points. [Diagram 5] FIG. 2 shows a schematic diagram of the construction of an input wake; [Figure 6] FIG. 2 is a diagram illustrating a trajectory prediction model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] We will further describe the present invention in combination with the following drawings: As shown in Figure 1, the self-adaptive time interval ship wake prediction method based on Miller Gate network includes the following steps:

[0028] S1, construct a mirror gate network. Figure 2 shows the specific structure of the forward gate and the reverse gate. Figure 3 shows the details of the mirror gate network constructed based on the forward gate and the reverse gate. This structure is used as the basic model.

[0029] S2. Aligning the equal time interval data. As shown in Figure 4, we achieve data alignment by constructing interpolation functions in four different dimensions: lon, lat, cog, and sog.

[0030] S3, construct input data. As shown in Fig. 5, a moving time window process with length L and stride length L / q is performed on the aligned track data, and various track data are merged to construct the input of the multiple track prediction model.

[0031] S4: Construct a self-adaptive trajectory prediction model with variable long-term interval. As shown in Fig. 6, train N forward and backward gate networks, stop training when the loss of the forward and backward gates is minimized, and determine the self-adaptive variable length N according to the final result.

[0032] The present invention is not limited to the present embodiment, and any equivalent ideas or modifications within the technical scope disclosed in the present invention are included in the scope of protection of the present invention.

[0033] (Additional Note) (Appendix 1) A step S1 of constructing a mirror gate network; A step S2 of extracting trajectory characteristic points at equal time intervals; A step S3 of constructing input track data; and and (S4) constructing a self-adaptive trajectory prediction model with a variable long time interval; In S1, the mirror gate network comprises a forward gate unit and a reverse gate unit; S1 includes the following steps S11 to S14, S11, setting the forward gate unit of the mirror gate network; In S11, the forward gate unit includes a forward merging gate, a forward splitting gate and a forward regenerating gate, and the forward gate unit is used to control the input and output of track information; S11 includes the following steps S111 to S113, S111, the forward merging gate is in the current track information input state x t and the unit status of the previous gate unit

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

Claims

[Claim 1] A step S1 of constructing a mirror gate network; A step S2 of extracting trajectory characteristic points at equal time intervals; A step S3 of constructing input track data; and and (S4) constructing a self-adaptive track prediction model with a variable long time interval; In S1, the mirror gate network comprises a forward gate unit and a reverse gate unit; S1 includes the following steps S11 to S14: S11, setting the forward gate unit of the mirror gate network; In S11, the forward gate unit includes a forward merging gate, a forward splitting gate, and a forward regenerating gate, and the forward gate unit is used to control the input and output of track information; S11 includes the following steps S111 to S113: S111, the forward merging gate is in the current track information input state x ｔ and the unit status of the previous gate unit [0010] and the hidden layer state of the previous gate unit [0025] The forward merging gate is responsible for aggregating all input information, and the calculation formula is as follows: [0030] Where: [0045] are respectively [0050] represents the weight of , sigmoid represents the activation function, S112, the forward branch gate converts the information passed through the forward merge gate into the unit state of the previous gate unit. [006] and the hidden layer state of the previous gate unit [0070] Together, they form a branch and create a new unit state. [0080] and the hidden layer state [0090] The calculation formula is as follows: [0010] Where: ##EQU00011## are respectively ##EQU00012## represents the weights corresponding to the process of calculating S113, the forward regeneration gate converts the information passed through the forward merging gate into the current track information input state x ｔ and the unit state of the current unit ##EQU00013## and the hidden layer state ##EQU14## With new condition ##EQU00015## The formula is as follows: ##EQU00016## Where: ##EQU00017## are respectively [0018] In the process of generating [0019] represents the weight corresponding to S12, setting the reverse gate unit of the mirror gate network; In S12, the reverse gate unit includes a reverse merging gate, a reverse splitting gate and a reverse regenerating gate, and the reverse gate unit performs a reverse operation on the input information; S12 includes the following steps S121 to S123: S121, the reverse merging gate is in the current track information input state x ｔ and the unit state of the next gate unit [0020] and the hidden layer state of the next gate unit ##EQU00021## The reverse merging gate is responsible for aggregating all input information, and the calculation formula is as follows: [0022] Where: [0023] are respectively [0024] represents the weight of S122, the reverse branch gate converts the information passed through the reverse merge gate into the unit state of the next gate unit according to the following formula: [0025] and the hidden layer state of the next gate unit [0026] Together, they form a branch and create a new unit state. [0027] and the hidden layer state [0028] The calculation formula is as follows: [0029] Where: [0030] are calculated in the process of calculating the reverse shunt gate. [0031] represents the weight corresponding to S123, the reverse regeneration gate converts the information passed through the reverse merging gate into the current track information input state x ｔ and the unit state of the current unit [0032] and the hidden layer state [Equation 33] With new condition [0034] The formula is as follows: [Equation 35] Where: [0036] are respectively [Equation 37] In the process of generating [Equation 38] represents the weight corresponding to S13, the forward gate units are sequentially connected in the forward direction, and the backward gate units are sequentially connected in the backward direction, and the states of the forward gate units and the backward gate units at the same time distance are merged to construct the mirror gate network, and the calculation formula is as follows: [0039] Where: [0040] represents the merge operation, S14, perform calculations in the mirror gate network to obtain the following predicted output information: [0041] Here, w ｇｙ teeth, [0042] represents the weight of S2 includes the following steps S21 to S24: S21, by extracting the feature points of the self-adaptation time interval, the wake P in the AIS data collected in the same sea area is ｋ is defined and expressed as a track feature vector, and the calculation formula is as follows: P ｋ ={p １ ｋ ,p ２ ｋ ,...,p ｉ ｋ ,...,p ｌ ｋ } Here, p ｉ ｋ is the track P of the kth ship in the sea area at time stamp i. ｋ represents the track point of p ｉ ｋ The characteristics of are expressed by the following formula: p ｉ ｋ =(lon ｉ ｋ , lat ｉ ｋ , cog ｉ ｋ ,sog ｉ ｋ ) Here, lon ｉ ｋ represents the ship's longitude value, and lat ｉ ｋ represents the ship's latitude value, and cog ｉ ｋ represents the ship's course over the ground, and sog ｉ ｋ represents the ship's ground speed, The self-adaptation time interval is defined as t and expressed as follows: t=(t １ ,t ２ ,...,t ｉ ,...,t ｎ ) Here, t ｉ , P ｋ represents the time distance of the i-th track point in S22, section [t ｉ , t ｉ＋１ ], an interpolation function represented by the following formula is defined: X ｉ (t ｉ )=a ｉ t ｉ ３ +b ｉ t ｉ ２ +c ｉ t ｉ +d ｉ X ｉ＋１ (t ｉ＋１ )=a ｉ t ｉ＋１ ３ +b ｉ t ｉ＋１ ２ +c ｉ t ｉ＋１ +d ｉ Taking the first derivative of the above function gives the following equation: X ｉ (t ｉ )′=3a ｉ t ｉ ２ +2b ｉ t ｉ +c ｉ X ｉ＋１ (t ｉ＋１ )′=3a ｉ t ｉ＋１ ２ +2b ｉ t ｉ＋１ +c ｉ Combining the interpolation function and the derivative of the interpolation function and solving the equation, a ｉ , b ｉ , c ｉ and d ｉ Find the value of S23, X ｉ (t ｉ ) ' and X ｉ＋１ (t ｉ＋１ )′, where all intermediate points of the interpolation function are calculated using M ｊ (t ｊ , x ｊ ) and M according to the following formula: ｊ The first-order difference quotient of the left and right coordinate points of k ｊ Tok ｊ＋１ It is expressed as [0043] According to the following formula, the midpoint M ｊ The weights of the two adjacent periods to the left and right of are ω ｊ and ω ｊ＋１ Define it as: [0044] According to the following formula, the midpoint M ｊ The first derivative in X ｊ (t)' is approximated, [0045] S24, substituting the time series obtained by equally dividing the track recording period into an interpolation function of the longitude and latitude in each period to obtain feature points extracted at equal time intervals of the track; S3 includes the following steps S31 to S32: S31. Perform a moving time window process to establish a moving time window to retain the time series characteristics, dynamic patterns and correlation information of the track data and allow the model to better learn the change / evolution process of the track data, and set the length of the moving time window as L and the moving distance as L / q, where q is a flexibly selected integer, and the smaller q is, the stronger the correlation between the data before and after; S32, the AIS data that has undergone feature extraction and moving time window processing for self-adaptive time intervals is input to the mirror gate network to train the forward gate unit and the backward gate unit respectively, and the mirror gate network is used to train and predict one or more trajectories; and the input data I ｔ and output data O ｔ is expressed by the following formula: [0046] where r represents the number of predicted tracks and S ｔ ｋ represents the track data of the kth ship processed at time t, and lon ｔ ｋ represents the interpolated longitude at time t, and lat ｔ ｋ represents the interpolated latitude at time t, and cog ｔ ｋ represents the course corresponding to the interpolated latitude and longitude at time t, and sog ｔ ｋ represents the navigation speed corresponding to the interpolated latitude and longitude at time t, L denotes the length of the moving time window at time t, and u represents the length of the predicted trajectory of the model. In S4, each input data is trained through N forward gate units and N backward gate units, and a self-adaptive variable length N is determined according to the final result of the mirror gate network, and an optimal trajectory prediction result is obtained; A self-adaptive time interval ship wake prediction method based on mirror gate network, characterized by:

Images