CN117933492A - Ship track long-term prediction method based on space-time feature fusion - Google Patents
Ship track long-term prediction method based on space-time feature fusion Download PDFInfo
- Publication number
- CN117933492A CN117933492A CN202410323216.5A CN202410323216A CN117933492A CN 117933492 A CN117933492 A CN 117933492A CN 202410323216 A CN202410323216 A CN 202410323216A CN 117933492 A CN117933492 A CN 117933492A
- Authority
- CN
- China
- Prior art keywords
- time
- ship
- ship track
- space
- prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 74
- 230000004927 fusion Effects 0.000 title claims abstract description 46
- 230000007774 longterm Effects 0.000 title claims abstract description 37
- 239000013598 vector Substances 0.000 claims abstract description 81
- 230000008569 process Effects 0.000 claims abstract description 13
- 238000012549 training Methods 0.000 claims description 18
- 230000001364 causal effect Effects 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 13
- 238000009826 distribution Methods 0.000 claims description 9
- 238000011176 pooling Methods 0.000 claims description 9
- 238000004140 cleaning Methods 0.000 claims description 8
- 230000000694 effects Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 5
- 230000010339 dilation Effects 0.000 claims description 5
- 230000003044 adaptive effect Effects 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000007499 fusion processing Methods 0.000 claims description 2
- 230000006872 improvement Effects 0.000 description 9
- 230000007246 mechanism Effects 0.000 description 7
- 230000008901 benefit Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000006399 behavior Effects 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 108010076504 Protein Sorting Signals Proteins 0.000 description 1
- 238000004873 anchoring Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a ship track long-term prediction method based on space-time feature fusion, and belongs to the field of data prediction. The method predicts based on a ship track prediction model TCNformer, and the prediction process is as follows: the method comprises the steps of obtaining a known ship track, converting AIS information observed values of all times in the known ship track into corresponding four-independent heat vectors, converting the four-independent heat vectors of all times in the known ship track into corresponding high-dimensional embedded vectors, extracting time dimension features and space dimension features from a high-dimensional feature vector sequence, fusing the time dimension features and the space dimension features to obtain space-time fusion features, and predicting track information of the next time point according to the space-time fusion features until a prediction task is completed. The method can effectively solve the problems of complexity of the ship motion mode and heterogeneity of AIS data, and realize long-term prediction of ship tracks.
Description
Technical Field
The invention belongs to the field of data prediction, and particularly relates to a track prediction method.
Background
In recent years, along with the continuous increase of global trade and the increasing frequency of maritime transportation activities, the maritime traffic density is increased, and the problems of route congestion, high collision risk, difficult path planning and the like of the traditional ship navigation management face the traditional ship navigation management, so that the importance of ship track prediction technology is increasingly highlighted. The ship track prediction can be performed by using the ship automatic identification system (Automatic Identification System, AIS) to monitor the ship position in real time, predict potential collision risk, optimize ship route planning and regulate and control marine traffic flow, and improve the overall shipping efficiency.
The ship dynamic information, static information and voyage related information included in the ship AIS data contain space-time distribution of ship behavior modes, ship operating behavior characteristics, ship traffic flow characteristics and ship habit voyage distribution characteristics. How to use the methods of geographic data mining and machine learning to mine and analyze the characteristics, the potential ship behavior knowledge and law hidden behind the ship AIS data are found, and further the future track prediction of the ship is of great significance.
In the prior art, the ship track prediction method based on the traditional physical model and the machine learning model has high calculation cost, is greatly influenced by data quality, and has low ship track prediction precision; the ship track prediction method based on deep learning can obtain a high-precision result when complex and dynamic AIS track data are processed by virtue of robustness and generalization.
However, the existing prediction method still cannot fully mine the time sequence characteristics among the motion tracks of the ship, and can only reach the relevant prediction performance in a short time range, namely a few minutes to half an hour or a predefined offshore route. Meanwhile, due to the complexity of the ship motion mode and the heterogeneity of AIS data, namely the AIS data has various sources and complex motion modes, the ship navigation state can change along with the time environmental condition, and in the ship track prediction, the heterogeneity leads to two ship paths with very similar motion modes, and different tracks can be displayed due to factors such as different destinations and different navigation strategies. Predicting the position of a ship for a long period of time in the future remains a great challenge.
Disclosure of Invention
The invention provides a ship track long-term prediction method based on space-time feature fusion, which aims at: the problem that the existing prediction method cannot accurately predict the future long-period track is solved.
The technical scheme of the invention is as follows:
a ship track long-term prediction method based on space-time feature fusion predicts based on a ship track prediction model TCNformer, and the prediction process of the ship track prediction model TCNformer is as follows:
step 1, acquiring and cleaning AIS data, and extracting known AIS data of voyages to be predicted from the AIS data after cleaning treatment to obtain a known ship track;
Step 2, converting AIS information observed values of all times in the known ship track into corresponding four independent heat vectors;
step 3, converting four independent heat vectors at each time in the known ship track into corresponding high-dimensional embedded vectors;
step 4, extracting time dimension features from the high-dimensional feature vector sequence by using a TCN network;
Step 5, extracting space dimension features from the high-dimensional feature vector sequence by using a transducer network;
step 6, fusing the time dimension features and the space dimension features to obtain space-time fusion features;
Step 7, predicting track information of the next time point according to the space-time fusion characteristics;
Step 8, if the predicted time length does not reach the preset length, adding the track information which is just predicted into the current known ship track, and then returning to the step 2 to predict the next time; and if the predicted time length reaches the preset length, the task is predicted to be completed.
As a further improvement of the ship track long-term prediction method based on space-time feature fusion: in step 2, it is assumed that the known ship trajectory is: Wherein/> Representing the number of currently known AIS data,/>Expressed in time/>Is expressed as: /(I); Wherein/>Is the latitude of the ship,/>Is the longitude of ship,/>For the speed of the vessel,/>Is the course of the ship; at this time/>In (a)The four attribute values are continuous values;
Then for all : Converting the attribute values into discrete values according to preset interval values through a segmentation method for each attribute; for each attribute/>Respectively obtain the whole/>The minimum value and the maximum value of the attribute are divided by the interval value corresponding to the attribute by the difference between the maximum value and the minimum value to obtain the attribute block value corresponding to the attribute; for each/>Each attribute/>The corresponding discrete value, minimum value, maximum value and attribute block value are taken as the/>Independent heat vector/>, corresponding to the attributeThen construct/>, using the unithermal vectors of the four attributesCorresponding four independent heat vectors/>：
。
As a further improvement of the ship track long-term prediction method based on space-time feature fusion: in step 3, for timeFour independent heat vectors/>The unique heat vector/>, of four attributesRespectively input to the embedded network/>Obtaining the corresponding attribute embedded vector/>Then concatenating the embedded vectors of each attribute as time/>Is embedded in high-dimensional feature vectorsFinally, the known ship track/>, is obtainedCorresponding high-dimensional feature vector sequence。
As a further improvement of the ship track long-term prediction method based on space-time feature fusion: in step 4, for an input high-dimensional feature vector sequenceDefine the filter as/>，/>Is the filter size, and thus the input high-dimensional feature vector sequence/>In/>The causal convolution operation of the dilation at the time point is defined as:
,/>；
wherein, Is the expansion factor,/>The past direction is described, in order to avoid the time index being negative, pair/>Filling with 0, filling length is/>；
Time point 0 to time pointAll/>Combining to obtain a high-dimensional feature vector sequence/>Expansion causal convolution results/>And the expansion cause and effect convolution operation is described as/>；
And then extracting time dimension characteristics:
；
；
；
wherein, Representation/>Convolution,/>Representing a non-linear activation function,Representing a weight normalization layer,/>Representation/>Layer/>Representing the time dimension characteristics of the TCN network output.
As a further improvement of the ship track long-term prediction method based on space-time feature fusion, the calculation process of step5 is as follows: will beHigh-dimensional feature vector/>Through a transducer network, a vector/>, the dimension of which is the same as that of input, can be obtainedThen each/>Corresponding/>Constitute spatial dimension feature/>The process is as follows:
。
As a further improvement of the ship track long-term prediction method based on space-time feature fusion, the fusion process of the step 6 is as follows:
Firstly, introducing a self-adaptive average pooling layer to obtain the time dimension characteristics after the change:
；
wherein, Is a feature of TCN network output,/>Representing an adaptive average pooling layer,Is the output characteristic after the self-adaptive average pooling layer;
Then the features are And the spatial dimension characteristics obtained in the step 5/>And (3) performing splicing fusion to obtain a high-dimensional feature vector sequence/>Corresponding spatiotemporal fusion features/>：
；
Spatiotemporal fusion featuresFor prediction of subsequent trails.
As a further improvement of the ship track long-term prediction method based on space-time feature fusion, the specific prediction mode of the step 7 is as follows:
Fusing spatio-temporal features The method is divided into four parts according to attributes:
；
wherein, Representation pair/>According to the number of divided blocks/>The division is carried out in such a way that,，/>Dividing the block number into four attributes of latitude, longitude, speed and heading respectively;
then, the classification distribution is obtained through logarithmic calculation to obtain the next time The predicted independent heat vector for each of the corresponding four attributes:
；
wherein, Representing a discrete probability distribution,/>Representing the original unnormalized fraction of the network output for prediction;
And then splicing the four predicted independent heat vectors to obtain predicted four independent heat vectors:
；
Then four independent heat vectors Pseudo-inversion to obtain time/>Corresponding AIS information predicted value:，/> Representing a pseudo-inverse calculation.
As a further improvement of the ship track long-term prediction method based on space-time feature fusion, the training mode of the ship track prediction model TCNformer is as follows:
step T1, collecting and cleaning AIS data, obtaining AIS data of a plurality of voyages based on the processed AIS data, and sorting the AIS data of each voyage into the AIS data of each voyage Format, input and output sample pairs are obtained from the obtained voyage AIS data: firstly randomly cutting/>Observed value/>And/>Are all preset integers, and then cut out 0 th to/>Personal/>Individual observations/>As input, the/>Personal arrival/>/>The individual observations are noted/>As an output, the true value of the track to be predicted is obtained; AIS data of each voyage obtain a group of corresponding input/output sample pairs,/>>；
Selecting a training set from the input and output sample sets;
step T2, centering the first group of input/output samples in the training set Inputting the model into a ship track prediction model;
step T3, calculating a loss value when the ship track prediction model predicts;
Step T4, per complete set Corresponding/>After the prediction of the AIS information, the final loss value/>, corresponding to the group of inputs, is obtainedThen, new network parameters are obtained through AdamW functions:
；
wherein, For the current network parameter set,/>To obtain a new set of network parameters;
Will be Assigning values to a ship track prediction model;
Step T5, if there are still input/output samples in the training set that are not input, the next set of input/output samples is processed Inputting the model into a ship track prediction model, and returning to the step T3; otherwise, the training is ended.
As a further improvement of the ship track long-term prediction method based on space-time feature fusion, the specific processing mode of the step T3 is as follows:
Each input set of When the ship track prediction model is reached, the loss value/>The initial value is set to 0; then, the time/>, is predicted in turnTo/>AIS information of (2); each time an AIS information is predicted, the loss value/>, is updated once。
As a further improvement of the ship track long-term prediction method based on space-time feature fusion, the loss value is updatedThe method comprises the following steps:
step T3-1, setting the currently predicted time The corresponding AIS information predictors are noted/>，/>Representing the number of currently known AIS data, prediction/>The spatiotemporal fusion feature used is/>Dividing it by the number of divided blocksThe method comprises the following steps:
；
step T3-2, setting the current input/output sample centering time The corresponding real AIS information observation value isIts corresponding four independent heat vectors are/>(Refer to the conversion approach of step 2), thenAccording to the number of divided blocks/>The method comprises the following steps:
；
Step T3-3, calculating the cross entropy corresponding to each of the four attributes:
；
；
；
；
wherein, Is a cross entropy function;
step T3-4, calculating the total cross entropy ：
；
The current total cross entropyAccumulated to loss value/>The method comprises the following steps: /(I)。
Compared with the prior art, the invention has the following beneficial effects:
1. The method can effectively solve the problems of complexity of the ship motion mode and heterogeneity of AIS data, and realize long-term prediction of ship tracks:
1. According to the invention, the time dimension characteristics and the space dimension characteristics are fused, the representation of AIS data can be fully extracted, and the representation which is more suitable for future track prediction tasks is formed on input data through parameter optimization.
2. The prediction problem is effectively modeled as a classification problem when the loss is calculated, so that the problem that the similar motion mode targets are difficult to distinguish due to common mean square error loss is avoided, and the model is better trained, so that the prediction precision is improved, and long-term track prediction is realized.
3. Through iterative optimization loss, construction of four-attribute classification distribution and repeated sampling, different possible prediction paths can be fully considered, and the problem of multiple states of ship motion is effectively solved.
2. According to the invention, AIS data are cleaned, and extra noise and uncertainty caused by abnormal values and missing data are eliminated, so that the convergence of the model training process is improved; on the other hand, eliminating the data loss value helps to calculate the prediction error during the model evaluation phase.
3. According to the invention, the cleaned AIS data are mapped to the high-dimensional feature space so as to increase the distinction of the motion modes of two ships with similar historical tracks, thereby avoiding the situation that the track prediction results are similar due to the fact that the two ship paths with similar motion modes are mapped to similar tracks, and further better processing the ship targets with complex motion models.
4. The invention constructs a time dimension feature extraction module based on a TCN network, and captures the local dependency relationship in the sequence data more effectively by introducing causal convolution; by introducing expansion convolution to skip part of input, the filter can apply an area larger than the length of the filter, so that the problem of information leakage caused by the receptive field size of the CNN in the calculation process is solved; by introducing residual connection, the network can transmit information in a cross-layer mode, and the possible overfitting problem of the network is effectively solved.
5. According to the invention, the space characteristics are extracted through the transducer module, so that the possible long-term correlation in the historical AIS observation data can be effectively captured, and the capability of the algorithm for processing a large amount of input data is improved through a multi-head attention mechanism.
Drawings
FIG. 1 is a schematic diagram of TCNformer prediction model according to the present invention;
FIG. 2 is a schematic diagram of AIS data sparse high-dimensional mapping;
FIG. 3 is a schematic diagram of an expansion convolution structure;
FIG. 4 is a schematic diagram of a multi-headed attention structure;
fig. 5 is a graph comparing prediction errors of different models.
Detailed Description
The technical scheme of the invention is described in detail below with reference to the accompanying drawings. It will be apparent that the described embodiments are only some, but not all, embodiments of the invention.
A ship track long-term prediction method based on space-time feature fusion defines a prediction problem as a classification problem through an AIS data sparse high-dimensional representation method, and then captures long-term dependency features in a ship track through a ship track prediction model TCNformer (shown in figure 1) combining time dimension features and space dimension features. The result of testing TCNformer predictive models on a real AIS data set shows that the prediction error of the method on long-term prediction is significantly lower than that of other models.
The long-term prediction method for the ship track comprises the following steps:
and step1, collecting and cleaning AIS data.
For the collected original AIS data, abnormal value and missing value processing is carried out according to the following steps:
(1) AIS information with unrealistic speed value (SOG more than or equal to 30 sections) is deleted;
(2) Removing mooring or anchoring vessel information;
(3) AIS observation information within a coastline 1 is deleted;
(4) Dividing discontinuous voyages into continuous voyages, continuous voyages referring to voyages in which the maximum interval between two consecutive AIS messages is less than a predefined value (here, 2 hours);
(5) Deleting AIS voyages with a length less than 20 or a duration less than 4 hours;
(6) Deleting the anomaly message, if the speed (calculated by dividing the distance travelled by the corresponding interval between two consecutive messages) is too great, as anomaly, where the speed threshold locates 40 knots;
(7) The long range is divided into short ranges, and the maximum range length is 20 hours.
And then extracting known AIS data of the voyage number to be predicted from the processed AIS data: Wherein/> Representing the number of currently known AIS data,/>Expressed in time/>Is expressed as:
；
wherein, Is the latitude of the ship,/>Is the longitude of ship,/>For the speed of the vessel,/>Is the heading of the ship. At this time/>Middle/>The four attribute values are consecutive values.
And 2, converting AIS information observed values of all times in the known ship track into corresponding four independent heat vectors.
For all of: Converting the attribute values into discrete values according to preset interval values through a segmentation method for each attribute; for each attribute/>Respectively obtain the whole/>The minimum value and the maximum value of the attribute are divided by the interval value corresponding to the attribute by the difference between the maximum value and the minimum value to obtain the attribute block value corresponding to the attribute; for each/>Each attribute/>The corresponding discrete value, minimum value, maximum value and attribute block value are taken as the/>Independent heat vector/>, corresponding to the attributeThen construct/>, using the unithermal vectors of the four attributesCorresponding four independent heat vectors/>：
。
And step 3, converting four independent heat vectors at each time in the known ship track into corresponding high-dimensional embedded vectors.
As in fig. 2, for timeFour independent heat vectors/>The single heat vector of four attributes is calculatedRespectively input to the embedded network/>Obtaining the corresponding attribute embedded vector/>Then concatenating the embedded vectors of each attribute as time/>Is embedded in high-dimensional feature vector/>. Embedded network/>The method comprises a self-attention layer and a linear layer, wherein the self-attention layer weights information in the input independent heat vector, so that richer characteristic representations are extracted, and the information is input into the linear layer to realize high-dimensional embedding.
From this, a known ship track is obtainedCorresponding high-dimensional feature vector sequence/>。
And 4, extracting time dimension features from the high-dimensional feature vector sequence by using a TCN network (Temporal Convolutional Networks).
The high-dimensional feature vector sequence is input into causal convolution, which is a unidirectional structure, the output is only determined by current and past inputs, and the causal convolution is learnedFuture data cannot be seen when historical data in the causal convolution is obtained, and only the previous data has the effect of the latter data, so that the data points learned by the causal convolution have strict constraint in time, but the problem of calculation amount surge and overfitting caused by the increase of the layer number is caused. For this purpose, an expansion convolution and residual connection structure is introduced.
The expansion convolution structure enables the filter to apply an area larger than the length of the filter by skipping part of input, so that the problem of information leakage caused by the fact that the CNN is subjected to receptive field size in the calculation process is solved, meanwhile, information of some neurons is ignored by means of crossing connection, and therefore the calculation complexity of the model is effectively reduced. The causal convolution structure of the dilation is shown in fig. 3 for a sequence of high-dimensional eigenvectors of the inputDefine the filter as/>，/>Is the filter size, and thus the input high-dimensional feature vector sequence/>In/>The causal convolution operation of the dilation at the time point is:
,/>；
wherein, Is the expansion factor,/>The past direction is described, in order to avoid the time index being negative, pair/>Filling with 0 with a proper amount, wherein the filling length is/>。
Time point 0 to time pointAll/>Combining to obtain a high-dimensional feature vector sequence/>Expansion causal convolution results/>And the expansion cause and effect convolution operation is described as/>。
In order to solve the over-fitting problem that may result as the number of layers increases, a residual module is introduced so that the network can pass information in a cross-layer fashion. The residual module constructed by the method comprises two layers of expansion causal convolution and nonlinear activation functions, weightNorm and Dropout are added into each layer to regularize a network, and one is adopted to solve the problem that input and output have different characteristic dimensionsConvolution reduces the dimension of the input to realize the addition of the feature graphs of the input and the output, and the process of extracting the time dimension features is expressed as follows:
；
；
；
wherein, Representation/>Convolution,/>For the output eigenvectors through the first dilation causal convolution, weightNorm and Dropout layers,/>Representing a nonlinear activation function,/>Representing a weight normalization layer,/>Representation/>Layer/>Representing the time dimension characteristics of the TCN network output.
And 5, extracting the space dimension characteristics by using a transducer network.
In order to be able to correctly predict the trajectory of the ship, the prediction model constructed needs to capture long-term correlations that may exist in the historical AIS observations. Therefore, the application selects a transducer module to extract the spatial characteristics, and a transducer network model adopts a framework similar to a GPT sub-model, and the paper 'Improving Language Understanding by GENERATIVE PRE-Training' can be referred to specifically. The model consists of 8 block structures, each containing 2 linear layers, one multi-headed attention layer and one MLP layer. The input of the transducer network is the high-dimensional feature vector sequence constructed in the step 2The output is the spatial dimension feature/>. The transducer network is composed of a series of attention layers stacked together, each layer can function as an autoregressive model, and a multi-head attention mechanism is mainly adopted in the attention layers, the multi-head attention mechanism is shown in fig. 4, the multi-head attention mechanism can effectively extract long-distance characteristics of a signal sequence and has the capability of processing a large amount of data, and the multi-head attention mechanism can be represented by the following formula:
；
Wherein the method comprises the steps of 、/>、/>Is a linear projection of the input sequence, in the formula/>For/>Or/>Is a dimension of (c). At each layer, the input sequence is projected to a new space/>And the output of the attention block is/>Wherein weights represent the relative contribution of each time step, these weights being calculated as/>And/>Dot product of (1)/>, useNormalized Softmax values. During the training phase, the network will learn/>、/>、/>And performs parallel computation. Parallel processing capability is a key advantage of the Transformer model over recursive networks, which allows the model to retrieve information directly from multiple past time steps simultaneously, where the model must process the data in order and may not be able to retrieve long-term information.
High-dimensional feature vector/>Through a transducer network, a vector/>, the dimension of which is the same as that of input, can be obtainedThen each/>Corresponding/>Constitute spatial dimension feature/>The process is as follows:
。
and step 6, fusing the time dimension features and the space dimension features to obtain space-time fusion features.
In order to fully extract characteristics of input AIS track information data, TCNformer provided by the patent application comprises two networks of TCN and transducer, which are respectively used for extracting time dimension characteristics and space dimension characteristics of the AIS track information data. The two features are further processed and fused respectively, and full characterization of AIS data is completed.
The input of the module is the characteristics obtained by learning two models of TCN and transducer, and the output is the space-time fusion characteristics. Specifically, since the sequence length of the spatial dimension feature extracted in the step4 changes along with the parameters of the embedding function, after the time dimension feature is extracted by the TCN network, an adaptive average pooling layer needs to be introduced to adaptively adjust the sequence length of the time dimension feature, so as to obtain the changed time dimension feature:
；
wherein, Is a feature of TCN network output,/>Representing an adaptive average pooling layer,Is the characteristic output after the self-adaptive average pooling layer, and the characteristic is compared with the space dimension characteristic/>, which is output after the conversion networkAnd (3) performing splicing fusion to obtain a high-dimensional feature vector sequence/>Corresponding spatiotemporal fusion features/>：
；
Spatiotemporal fusion featuresFor prediction of subsequent trails.
And 7, predicting the track information of the next time point according to the space-time fusion characteristics.
Fusing spatio-temporal featuresThe method is divided into four parts according to attributes:
；
wherein, Representation pair/>According to the number of divided blocks/>The division is carried out in such a way that,，/>The number of divided blocks is respectively the latitude, longitude, speed and heading.
Then, the classification distribution is obtained through logarithmic calculation to obtain the next timeThe predicted independent heat vector for each of the corresponding four attributes:
；
wherein, Representing a discrete probability distribution,/>Representing the original unnormalized fraction of the network output that was predicted, i.e., the feature.
And then splicing the four predicted independent heat vectors to obtain predicted four independent heat vectors:
；
Then four independent heat vectors Pseudo-inversion to obtain time/>Corresponding AIS information predicted value: . Wherein/> Representing a pseudo-inverse calculation.
Step 8, if the predicted time length does not reach the preset lengthWill/>As/>Added to the currently known ship track as/>In order/>Obtaining a new known ship track as/>Then returning to the step 2 to predict the next time; if the predicted length of time reaches the preset length/>The task is predicted to be completed.
The training mode of the ship track prediction model TCNformer is as follows:
And step T1, acquiring and cleaning AIS data, and obtaining AIS data of a plurality of voyages based on the processed AIS data. AIS data of each voyage is organized into Format. Obtaining input and output sample pairs from the obtained voyage AIS data: firstly randomly cutting/>Observed value/>And/>Are all preset integers, and then cut out 0 th to/>Personal/>Individual observations/>As input, the/>Personal arrival/>/>The individual observations are noted/>As an output, the true value of the trajectory to be predicted is obtained. AIS data of each voyage obtain a group of corresponding input/output sample pairs,/>>。
And dividing the input and output sample set according to the proportion of 8:2, and respectively serving as a training set and a testing set for obtaining AIS data.
Step T2, centering the first group of input/output samples in the training setAnd inputting the model into a ship track prediction model.
And T3, calculating a loss value when the ship track prediction model predicts.
Specifically, each input is a group ofWhen the ship track prediction model is reached, the loss value/>The initial value is set to 0. Then, the time/>, is predicted in turnTo/>Is set in the database. Each time an AIS information is predicted, the loss value/>, is updated once：
Step T3-1, setting the currently predicted timeThe corresponding AIS information predictors are noted/>，/>Representing the number of currently known AIS data, prediction/>The spatiotemporal fusion feature used is/>Dividing it by the number of divided blocksThe method comprises the following steps:
。
step T3-2, setting the current input/output sample centering time The corresponding real AIS information observation value isIts corresponding four independent heat vectors are/>(Refer to the conversion approach of step 2), thenAccording to the number of divided blocks/>The method comprises the following steps:
。
Step T3-3, calculating the cross entropy corresponding to each of the four attributes:
；
；
；
；
wherein, Is a cross entropy function.
Step T3-4, calculating the total cross entropy：
。
The current total cross entropyAccumulated to loss value/>The method comprises the following steps: /(I)。
Step T4, per complete setCorresponding/>After the prediction of the AIS information, the final loss value/>, corresponding to the group of inputs, is obtainedThen, new network parameters are obtained through AdamW functions:
；/>
wherein, For the current network parameter set,/>To obtain a new set of network parameters.
Will beAnd assigning values to the ship track prediction model.
Step T5, if there are still input/output samples in the training set that are not input, the next set of input/output samples is processedInputting the model into a ship track prediction model, and returning to the step T3; otherwise, the training is ended.
To further verify the effect of the proposed algorithm, the present embodiment tested the model TCNformer provided by the present application on a common AIS dataset provided by the danish maritime office (DMA). The dataset comprises AIS information data of cargo ships and wheels ranging from (55.5 DEG, 10.3 DEG) to (58.0 DEG, 13.0 DEG) during the period from 1 st 1 nd 2019 to 31 rd 3 nd 2019, the dataset comprises about 7.12 hundred million AIS messages prior to preprocessing, model and trim super parameters are respectively trained using data from 1 st 1 nd 2019 to 10 th3 nd 2019 and from 11 nd 3 nd 2019 to 20 th3 nd 2019, and the test dataset comprises data from 21 nd 3 nd 2019 to 31 nd 2019.
The TCN network in this embodiment uses 2 residual blocks, each consisting of 2 layers of causal convolutional layers of expansion and activation functions, the transducer network is an 8-layer structure, each layer contains 8 attention headers, and the design isCorresponding sizes of 256, 128 and 128, respectively, resulting in an embedding vector/> of 768 dimensionsHistory sequence length/>The model is trained by using a AdamW optimizer and a cyclic cosine decay learning rate scheduler after being set for 3 hours, the initial learning rate is set to be 5e-4, the performance of the proposed model and the comparison model is tested on an RTX 3080GPU, the epoch is set to be 100, and an early-stop mechanism is introduced.
Comparative model this patent application uses lstm_seq2seq, conv_seq2seq, lstm_seq2seq with attention mechanism, and transfomer models.
The evaluation criterion adopted in this embodiment is a prediction error for each predicted point, specifically, a baseline distance between a real position and a predicted position is calculated, and a calculation formula is as follows:
；
Wherein the method comprises the steps of Is the radius of the earth,/>，/>，/>And/>Respectively represent latitude,/>And/>Longitude representing the predicted position and the true position, respectively.
Table 1 shows the average prediction errors (sea) of the different models over 1, 2 and 3 hours on the extracted AIS dataset with only the wheel track, the prediction errors of the different models are gradually improved over time, the model designed in this example achieves the lowest prediction error, and the performance is improved by a factor of 2 compared to the other optimal models.
Table 1 average predicted performance of different models over a subset
。
The performance advantages of the proposed algorithm are mainly due to two aspects: 1) The AIS data similarity sparse high-dimensional representation is adopted for input, and the prediction problem is converted into the classification problem; 2) The model combines the time dimension feature and the space dimension feature, and can well extract the long-term dependency feature of AIS data. Compared with a transducer model, the method has the advantages that the performance is improved, the addition of the time dimension characteristics is proved to be capable of better representing AIS historical data, and the track position characteristics of AIS information data are learned.
In addition, one important factor to consider in using a predictive model is the prediction horizon, which refers to how long the predicted value is valid. To demonstrate the long-term predictive advantage of the present model, long-term predictive experiments were performed on the entire dataset, with the experimental results shown in fig. 5.
As can be seen from fig. 5, overall, the prediction performance gradually decreases as the prediction time becomes longer, indicating the diversity of the ship motion patterns. Compared with other models, the method has lower prediction error in the same time, and proves the effectiveness of the algorithm.
In search and rescue activities, prediction is considered helpful if the prediction error is less than visibility. The visibility is generally assumed to be 10 knots in sunny weather conditions, and the longest prediction horizon of the model can reach 7.21 hours under the visibility, which is prolonged by about 3.8 times compared with a time sequence model and about 1.4 times compared with a transducer model.
It should be noted that it will be apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, but may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the invention is indicated by the appended claims rather than by the foregoing description.
Claims (10)
1. A ship track long-term prediction method based on space-time feature fusion is characterized in that: based on the ship track prediction model TCNformer, the prediction process of the ship track prediction model TCNformer is as follows:
step 1, acquiring and cleaning AIS data, and extracting known AIS data of voyages to be predicted from the AIS data after cleaning treatment to obtain a known ship track;
Step 2, converting AIS information observed values of all times in the known ship track into corresponding four independent heat vectors;
step 3, converting four independent heat vectors at each time in the known ship track into corresponding high-dimensional embedded vectors;
step 4, extracting time dimension features from the high-dimensional feature vector sequence by using a TCN network;
Step 5, extracting space dimension features from the high-dimensional feature vector sequence by using a transducer network;
step 6, fusing the time dimension features and the space dimension features to obtain space-time fusion features;
Step 7, predicting track information of the next time point according to the space-time fusion characteristics;
Step 8, if the predicted time length does not reach the preset length, adding the track information which is just predicted into the current known ship track, and then returning to the step 2 to predict the next time; and if the predicted time length reaches the preset length, the task is predicted to be completed.
2. The ship track long-term prediction method based on space-time feature fusion as claimed in claim 1, wherein: in step 2, it is assumed that the known ship trajectory is:
Wherein/> Representing the number of currently known AIS data,/>Expressed in time/>Is expressed as: /(I); Wherein/>Is the latitude of the ship,/>Is the longitude of ship,/>For the speed of the vessel,/>Is the course of the ship; at this time/>Middle/>The four attribute values are continuous values;
Then for all : Converting the attribute values into discrete values according to preset interval values through a segmentation method for each attribute; for each attribute/>Respectively obtain the whole/>The minimum value and the maximum value of the attribute are divided by the interval value corresponding to the attribute by the difference between the maximum value and the minimum value to obtain the attribute block value corresponding to the attribute; for each/>Each attribute/>The corresponding discrete value, minimum value, maximum value and attribute block value are taken as the/>Independent heat vector/>, corresponding to the attributeThen construct/>, using the unithermal vectors of the four attributesCorresponding four independent heat vectors/>：
。
3. The ship track long-term prediction method based on space-time feature fusion as claimed in claim 2, wherein: in step 3, for timeFour independent heat vectors/>The unique heat vector/>, of four attributes Respectively input to the embedded network/>Obtaining the corresponding attribute embedded vector/>Then concatenating the embedded vectors of each attribute as time/>Is embedded in high-dimensional feature vector/>Finally, the known ship track/>, is obtainedCorresponding high-dimensional feature vector sequence/>。
4. A ship track long-term prediction method based on space-time feature fusion as claimed in claim 3, wherein: in step 4, for an input high-dimensional feature vector sequenceDefine the filter as/>，/>Is the filter size, and thus the input high-dimensional feature vector sequence/>In/>The causal convolution operation of the dilation at the time point is defined as:
,/>；
wherein, Is the expansion factor,/>The past direction is described, in order to avoid the time index being negative, pair/>Filling with 0, filling length is/>；
Time point 0 to time pointAll/>Combining to obtain a high-dimensional feature vector sequence/>Expansion causal convolution results/>And the expansion cause and effect convolution operation is described as/>；
And then extracting time dimension characteristics:
；
；
；
wherein, Representation/>Convolution,/>Representing a non-linear activation function,Representing a weight normalization layer,/>Representation/>Layer/>Representing the time dimension characteristics of the TCN network output.
5. The ship track long-term prediction method based on space-time feature fusion as claimed in claim 3, wherein the calculation process of the step 5 is as follows: will beHigh-dimensional feature vector/>Through a transducer network, a vector/>, the dimension of which is the same as that of input, can be obtainedThen each/>Corresponding/>Constitute spatial dimension feature/>The process is as follows:
。
6. the ship track long-term prediction method based on space-time feature fusion as claimed in claim 3, wherein the fusion process of step 6 is as follows:
Firstly, introducing a self-adaptive average pooling layer to obtain the time dimension characteristics after the change:
；
wherein, Is a feature of TCN network output,/>Representing an adaptive average pooling layer,/>Is the output characteristic after the self-adaptive average pooling layer;
Then the features are And the spatial dimension characteristics obtained in the step 5/>And (3) performing splicing fusion to obtain a high-dimensional feature vector sequence/>Corresponding spatiotemporal fusion features/>：
；
Spatiotemporal fusion featuresFor prediction of subsequent trails.
7. The method for long-term prediction of ship track based on space-time feature fusion as claimed in claim 6, wherein the specific prediction mode in step 7 is as follows:
Fusing spatio-temporal features The method is divided into four parts according to attributes:
；
wherein, Representation pair/>According to the number of divided blocks/>The division is carried out in such a way that,，/>Dividing the block number into four attributes of latitude, longitude, speed and heading respectively;
then, the classification distribution is obtained through logarithmic calculation to obtain the next time The predicted independent heat vector for each of the corresponding four attributes:
；
wherein, Representing a discrete probability distribution,/>Representing the original unnormalized fraction of the network output for prediction;
And then splicing the four predicted independent heat vectors to obtain predicted four independent heat vectors:
；
Then four independent heat vectors Pseudo-inversion to obtain time/>Corresponding AIS information predicted value:，/> Representing a pseudo-inverse calculation.
8. The long-term prediction method of ship track based on space-time feature fusion as claimed in claim 7, wherein the training mode of the ship track prediction model TCNformer is as follows:
step T1, collecting and cleaning AIS data, obtaining AIS data of a plurality of voyages based on the processed AIS data, and sorting the AIS data of each voyage into the AIS data of each voyage Format, input and output sample pairs are obtained from the obtained voyage AIS data: firstly randomly cutting/>Observed value/>And/>Are all preset integers, and then cut out 0 th to/>Personal/>Individual observations/>As input, the/>Personal arrival/>/>The individual observations are noted/>As an output, the true value of the track to be predicted is obtained; AIS data of each voyage obtain a group of corresponding input/output sample pairs,/>>；
Selecting a training set from the input and output sample sets;
step T2, centering the first group of input/output samples in the training set Inputting the model into a ship track prediction model;
step T3, calculating a loss value when the ship track prediction model predicts;
Step T4, per complete set Corresponding/>After the prediction of the AIS information, the final loss value/>, corresponding to the group of inputs, is obtainedThen, new network parameters are obtained through AdamW functions:
；
wherein, For the current network parameter set,/>To obtain a new set of network parameters;
Will be Assigning values to a ship track prediction model;
Step T5, if there are still input/output samples in the training set that are not input, the next set of input/output samples is processed Inputting the model into a ship track prediction model, and returning to the step T3; otherwise, the training is ended.
9. The ship track long-term prediction method based on space-time feature fusion as claimed in claim 8, wherein the specific processing manner of step T3 is as follows:
Each input set of When the ship track prediction model is reached, the loss value/>The initial value is set to 0; then, the time/>, is predicted in turnTo/>AIS information of (2); each time an AIS information is predicted, the loss value/>, is updated once。
10. The method for long-term prediction of ship track based on space-time feature fusion according to claim 9, wherein the loss value is updatedThe method comprises the following steps:
step T3-1, setting the currently predicted time The corresponding AIS information predictors are noted/>，/>Representing the number of currently known AIS data, prediction/>The spatiotemporal fusion feature used is/>Dividing the block into blocks/>The method comprises the following steps:
；
step T3-2, setting the current input/output sample centering time The corresponding real AIS information observation value isIts corresponding four independent heat vectors are/>Will/>According to the number of divided blocks/>The method comprises the following steps:
；
Step T3-3, calculating the cross entropy corresponding to each of the four attributes:
；
；
；
；
wherein, Is a cross entropy function;
step T3-4, calculating the total cross entropy ：
；
The current total cross entropyAccumulated to loss value/>The method comprises the following steps: /(I)。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410323216.5A CN117933492B (en) | 2024-03-21 | 2024-03-21 | Ship track long-term prediction method based on space-time feature fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410323216.5A CN117933492B (en) | 2024-03-21 | 2024-03-21 | Ship track long-term prediction method based on space-time feature fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117933492A true CN117933492A (en) | 2024-04-26 |
| CN117933492B CN117933492B (en) | 2024-06-11 |
Family
ID=90752246
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410323216.5A Active CN117933492B (en) | 2024-03-21 | 2024-03-21 | Ship track long-term prediction method based on space-time feature fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117933492B (en) |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837148A (en) * | 2021-11-04 | 2021-12-24 | 昆明理工大学 | Pedestrian trajectory prediction method based on self-adjusting sparse graph transform |
| GB202208732D0 (en) * | 2022-06-14 | 2022-07-27 | Five Ai Ltd | Motion prediction for mobile agents |
| CN114997067A (en) * | 2022-06-30 | 2022-09-02 | 南京航空航天大学 | Trajectory prediction method based on space-time diagram and space-domain aggregation Transformer network |
| CN115079116A (en) * | 2022-04-14 | 2022-09-20 | 杭州电子科技大学 | Radar target identification method based on Transformer and time convolution network |
| WO2022252398A1 (en) * | 2021-05-31 | 2022-12-08 | 武汉理工大学 | Ship trajectory feature point extraction-based spatio-temporal dp method |
| CN116342657A (en) * | 2023-03-29 | 2023-06-27 | 西安电子科技大学 | TCN-GRU ship track prediction method, system, equipment and medium based on coding-decoding structure |
| CN116541708A (en) * | 2023-05-16 | 2023-08-04 | 上海船舶运输科学研究所有限公司 | Ship track prediction method and system integrating data quality control and converter network |
| CN116613740A (en) * | 2023-05-23 | 2023-08-18 | 浪潮云信息技术股份公司 | Intelligent load prediction method based on transform and TCN combined model |
| CN116721538A (en) * | 2023-05-18 | 2023-09-08 | 汕头大学 | Method for adaptively learning traffic flow prediction under dynamic traffic condition |
| CN117390506A (en) * | 2023-09-24 | 2024-01-12 | 北京工业大学 | Ship path classification method based on grid coding and textRCNN |
| CN117392830A (en) * | 2023-08-28 | 2024-01-12 | 湖北工业大学 | Space-time normalization graph convolution neural network traffic flow prediction method, system and medium |
| CN117474184A (en) * | 2023-11-08 | 2024-01-30 | 中国科学院国家空间科学中心 | Ship track prediction method and system driven by dynamics knowledge |
| CN117522920A (en) * | 2023-11-10 | 2024-02-06 | 南通大学 | Pedestrian track prediction method based on improved space-time diagram attention network |
-
2024
- 2024-03-21 CN CN202410323216.5A patent/CN117933492B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022252398A1 (en) * | 2021-05-31 | 2022-12-08 | 武汉理工大学 | Ship trajectory feature point extraction-based spatio-temporal dp method |
| CN113837148A (en) * | 2021-11-04 | 2021-12-24 | 昆明理工大学 | Pedestrian trajectory prediction method based on self-adjusting sparse graph transform |
| CN115079116A (en) * | 2022-04-14 | 2022-09-20 | 杭州电子科技大学 | Radar target identification method based on Transformer and time convolution network |
| GB202208732D0 (en) * | 2022-06-14 | 2022-07-27 | Five Ai Ltd | Motion prediction for mobile agents |
| CN114997067A (en) * | 2022-06-30 | 2022-09-02 | 南京航空航天大学 | Trajectory prediction method based on space-time diagram and space-domain aggregation Transformer network |
| CN116342657A (en) * | 2023-03-29 | 2023-06-27 | 西安电子科技大学 | TCN-GRU ship track prediction method, system, equipment and medium based on coding-decoding structure |
| CN116541708A (en) * | 2023-05-16 | 2023-08-04 | 上海船舶运输科学研究所有限公司 | Ship track prediction method and system integrating data quality control and converter network |
| CN116721538A (en) * | 2023-05-18 | 2023-09-08 | 汕头大学 | Method for adaptively learning traffic flow prediction under dynamic traffic condition |
| CN116613740A (en) * | 2023-05-23 | 2023-08-18 | 浪潮云信息技术股份公司 | Intelligent load prediction method based on transform and TCN combined model |
| CN117392830A (en) * | 2023-08-28 | 2024-01-12 | 湖北工业大学 | Space-time normalization graph convolution neural network traffic flow prediction method, system and medium |
| CN117390506A (en) * | 2023-09-24 | 2024-01-12 | 北京工业大学 | Ship path classification method based on grid coding and textRCNN |
| CN117474184A (en) * | 2023-11-08 | 2024-01-30 | 中国科学院国家空间科学中心 | Ship track prediction method and system driven by dynamics knowledge |
| CN117522920A (en) * | 2023-11-10 | 2024-02-06 | 南通大学 | Pedestrian track prediction method based on improved space-time diagram attention network |
Non-Patent Citations (2)
| Title |
|---|
| ZIJING DONG等: "TCN-Informer-Based Flight Trajectory Prediction for Aircraft in the Approach Phase", SUSTAINABILITY, vol. 15, 27 November 2023 (2023-11-27), pages 3 * |
| 耿磊 等: "基于TCN和Transformer的鸡胚心跳混淆信号分类方法", 《农业机械学报》, vol. 54, no. 8, 31 August 2023 (2023-08-31), pages 296 - 308 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117933492B (en) | 2024-06-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dinkel et al. | Towards duration robust weakly supervised sound event detection | |
| CN111832814A (en) | Air pollutant concentration prediction method based on graph attention machine mechanism | |
| CN116342657B (en) | TCN-GRU ship track prediction method, system, equipment and medium based on coding-decoding structure | |
| CN114202120A (en) | Urban traffic travel time prediction method aiming at multi-source heterogeneous data | |
| CN114565124A (en) | Ship traffic flow prediction method based on improved graph convolution neural network | |
| CN115599779B (en) | Urban road traffic missing data interpolation method and related equipment | |
| CN115510174A (en) | Road network pixelation-based Wasserstein generation countermeasure flow data interpolation method | |
| CN110737267A (en) | Multi-objective optimization method for unmanned ships and intelligent comprehensive management and control system for unmanned ships | |
| CN116307152A (en) | Traffic prediction method for space-time interactive dynamic graph attention network | |
| CN115862319A (en) | Traffic flow prediction method for space-time diagram self-encoder | |
| CN114090718A (en) | Bi-LSTM prediction and fuzzy analysis based interrupted track correlation method | |
| CN115099328A (en) | Traffic flow prediction method, system, device and storage medium based on countermeasure network | |
| Xu et al. | Improved Vessel Trajectory Prediction Model Based on Stacked‐BiGRUs | |
| CN117933492B (en) | Ship track long-term prediction method based on space-time feature fusion | |
| CN113033410A (en) | Domain generalization pedestrian re-identification method, system and medium based on automatic data enhancement | |
| CN117474184A (en) | Ship track prediction method and system driven by dynamics knowledge | |
| Wang et al. | Filling gaps in significant wave height time series records using bidirectional gated recurrent unit and cressman analysis | |
| CN116578661A (en) | Vehicle track time-space reconstruction method and system based on attention mechanism | |
| CN116541708A (en) | Ship track prediction method and system integrating data quality control and converter network | |
| CN114595770B (en) | Long time sequence prediction method for ship track | |
| Zhou et al. | Sa-sgan: A vehicle trajectory prediction model based on generative adversarial networks | |
| Zhang et al. | A novel ship trajectory clustering analysis and anomaly detection method based on AIS data | |
| CN113792919B (en) | Wind power prediction method based on combination of transfer learning and deep learning | |
| Zhang et al. | Prediction of Vessel Arrival Time to Pilotage Area Using Multi-Data Fusion and Deep Learning | |
| CN111967593A (en) | Method and system for processing abnormal data based on modeling |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |