Spatial-temporal trajectory-oriented multi-dimensional feature fusion coding method and device

Abstract

The invention discloses a spatial-temporal trajectory-oriented multi-dimensional feature fusion coding method and device, and belongs to the technical field of target information fusion, and the method comprises the steps: firstly, carrying out the tag coding of the numeric features and category features of a target spatial-temporal trajectory, and obtaining the integer coding value of each feature; then, vector embedding is carried out on the coded value of each feature, and an embedded vector of each feature is obtained; and finally, splicing the embedded vectors of all the features, and converting the spliced embedded vectors into a multi-dimensional feature fusion vector through a nonlinear transformation network. The method is high in mobility and wider in application range.

Description

Technical Field

[0001] The present invention relates to the technical field of target information fusion, and more specifically, to a multi-dimensional feature fusion coding method and device for spatio-temporal trajectories. Background Art

[0002] With the increasing development and maturity of sensor technology, wireless communication technology, and global positioning technology, a large amount of spatio-temporal trajectory data of various types of targets such as vehicles, ships, and airplanes has been generated. A spatio-temporal trajectory is a curve in a multi-dimensional space formed by adding a time axis to the geographical space, which can represent the position change of a moving object over a long period of time and consists of a series of trajectory points. Currently, the academic community has carried out a large amount of research work on spatio-temporal trajectories, including motion pattern mining, clustering analysis, trajectory classification, behavior recognition, anomaly detection, etc.

[0003] Currently, spatio-temporal trajectory data encoding, intelligent cognitive model construction, output pattern space expression, and intelligent model knowledge acquisition are the key technologies for realizing the intelligent cognition of spatio-temporal trajectories. The construction of an intelligent cognitive model is mainly based on machine learning models such as Bayesian networks, decision trees, and artificial neural networks; the output pattern space description is mainly used to define the output space of the intelligent model, that is, what the meaning of the output result of the intelligent model is, and to achieve an organic combination of the model output result and human understanding and cognition; intelligent model knowledge acquisition mainly includes issues such as the construction of a training sample set, data annotation, and model training optimization. Spatio-temporal trajectory data encoding is used to convert complex and heterogeneous spatio-temporal trajectory data into input information that an intelligent cognitive model can accept, and is the basis for the intelligent cognition of spatio-temporal trajectories.

[0004] Traditional machine learning models mainly rely on manual feature engineering to achieve the encoding of spatio-temporal trajectory data. In recent years, with the rapid development of artificial intelligence technology represented by deep learning, the reliability and scientificity of intelligent cognitive models have been greatly improved, and a variety of deep neural networks such as convolutional neural networks, recurrent neural networks, and Transformers have emerged. Due to the natural temporality of spatio-temporal trajectories and the need for no complex feature engineering, recurrent neural networks represented by LSTM and GRU have been widely used in spatio-temporal trajectory analysis fields such as trajectory prediction and trajectory classification, and these models have strong feature expression, extraction, and abstraction capabilities.

[0005] However, the spatio-temporal trajectory analysis method represented by the recurrent neural network still uses the traditional normalization method to process spatio-temporal trajectory features. It directly normalizes numerical features such as longitude, latitude, speed, and heading and uses them as the input of the intelligent model, which has the following problems: First, since the target trajectory is globally distributed, when the global distribution of the target trajectory is uneven, directly normalizing the longitude and latitude will result in relatively large changes in the normalized features, which will affect the analysis performance of the intelligent model; Second, the meanings corresponding to the feature values of each dimension obtained by the existing normalization encoding method are related to the distribution of the training data set used and are not fixed, which causes great problems in the generality and transferability of the intelligent model; Third, the existing normalization encoding method lacks a fusion encoding mechanism for numerical features and categorical features of spatio-temporal trajectories, is not well applicable to intelligent models represented by deep learning, and is also difficult to introduce a self-attention mechanism to establish long-distance dependencies. It can only establish short-distance dependencies, affecting the model performance. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a multi-dimensional feature fusion encoding method and device for spatio-temporal trajectories, which has strong transferability and a wider scope of application.

[0007] The purpose of the present invention is achieved through the following solutions:

[0008] A multi-dimensional feature fusion encoding method for spatio-temporal trajectories includes the following steps:

[0009] First, perform label encoding on the numerical features and categorical features of the target spatio-temporal trajectory to obtain the integer encoding value of each feature;

[0010] Then, perform vector embedding on the encoding value of each feature to obtain the embedding vector of each feature;

[0011] Finally, after splicing the embedding vectors of all features, convert them into a multi-dimensional feature fusion vector through a non-linear transformation network.

[0012] Further, the label encoding of the numerical features and categorical features of the target spatio-temporal trajectory specifically includes the time feature label encoding process:

[0013] Extract the hour, minute, and second information in the time information respectively, and then convert them into the corresponding label encoding. The calculation method is as follows:

[0014] h f = h + 1

[0015] m f = m + 1;

[0016] s f = s + 1

[0017] Among them, h, m, and s are the specific values of hours, minutes, and seconds respectively, and their value ranges are 0 to 23, 0 to 59, and 0 to 59 respectively. h f , m f , s f represent the label encodings of hours, minutes, and seconds respectively, and their value ranges are 0 to 24, 0 to 60, and 0 to 60 respectively. When the label encoding value is 0, it means the encoding is a padding value.

[0018] Furthermore, the label encoding of the numerical features and categorical features of the target spatio-temporal trajectory specifically includes the longitude and latitude feature label encoding process:

[0019] Using the equal-width discretization method, divide the value ranges of the longitude and latitude continuous variables into equal-width intervals respectively, and then map the variable values to the corresponding intervals; when performing equal-width discretization, determine the number of intervals. Among them, if the number of intervals for the target longitude division is set to N lon , and the number of intervals for the latitude division is N lat , then the calculation methods for the label encodings of longitude and latitude are as follows:

[0020]

[0021] Among them, lon and lat are the specific values of the longitude and latitude features of the spatio-temporal trajectory respectively, lon f and lat f are the label encodings of the longitude and latitude features respectively, and floor(·) represents rounding down; when the label encoding value is 0, it means the encoding is a padding value.

[0022] Furthermore, the label encoding of the numerical features and categorical features of the target spatio-temporal trajectory specifically includes the speed feature label encoding process:

[0023] Using the equal-width discretization method to discretize the continuous speed values into label encodings. When performing specific discretization, set the speed division interval to replace the setting of the number of speed intervals; according to the requirements of the actual application scenario, set the upper limit of the speed value. Among them, if the upper limit of the speed value is set to V max , and the speed division interval is V p , then the calculation method for the target speed feature label encoding is as follows:

[0024]

[0025] Among them, v is the specific value of the target speed feature, and floor(·) represents rounding down; when the label encoding value is 0, it means the encoding is a padding value.

[0026] Further, the label encoding of the numerical features and categorical features of the target spatio-temporal trajectory specifically includes the heading feature label encoding process:

[0027] The continuous heading values are discretized into label encodings using the equal-width discretization method. The target heading feature values are between 0 and 360°, and the heading division interval is set to C p , then the encoding calculation method of the target speed feature encoding is as follows:

[0028] c f = floor(c / C p ) + 1;

[0029] where c is the specific value of the target heading feature, and floor(·) represents rounding down; when the label encoding value is 0, it means it is encoded as a padding value.

[0030] Further, the label encoding of the numerical features and categorical features of the target spatio-temporal trajectory specifically includes the type feature label encoding process:

[0031] Map each target type feature value to an integer starting from 0; among them, if the target type feature value set is expressed as:

[0032] A = {a1, a2,... a i ,...., a n};

[0033] Then the label encoding corresponding to the target type feature value a i is i, denoted as t f ; when the target type feature value is not in the set A, the label encoding is directly set to 0.

[0034] Further, the vector embedding of the encoding values of each feature specifically includes the following sub-steps:

[0035] Assume that after the label encoding step of the numerical features and categorical features of the target spatio-temporal trajectory, a label encoding set {h f , m f , s f , lon f , lat f , v f , c f , t f} is obtained; perform a vector embedding operation on the label encodings obtained after the label encoding step of the numerical features and categorical features of the target spatio-temporal trajectory, and convert them into a dense vector form. For each label encoding, generate its embedding vector using the following method:

[0036] em = M m δ m ;

[0037] Among them, e m is the embedding vector corresponding to the label encoding, m ∈ {h f , m f , s f , lon f , lat f , v f , c f , t f}, is the embedding matrix, D m represents the dimension size of the embedding vector corresponding to the label encoding m, S m represents the size of the value set corresponding to the label encoding m, represents the one-hot vector corresponding to the label encoding m;

[0038] Through the above method, the embedding vectors corresponding to each label encoding are obtained, which are respectively

[0039] Furthermore, after concatenating the embedding vectors of all features, they are converted into a multi-dimensional feature fusion vector through a non-linear transformation network, which specifically includes the following sub-steps:

[0040] First, concatenate the embedding vectors corresponding to multiple label encodings to obtain a concatenated vector e:

[0041] e = ||e m , m ∈ {h f , m f , s f , lon f , lat f , v f , c f , t f};

[0042] Among them, || represents the vector concatenation operation;

[0043] Then, through the non-linear transformation network, convert the concatenated vector e into a multi-dimensional feature fusion vector e F .

[0044] Furthermore, the non-linear transformation network includes 1 layer of network, and the specific implementation method is as follows:

[0045] e F = σ(We + b);

[0046] Among them, is a network parameter, and σ(·) is an activation function.

[0047] A multi-dimensional feature fusion encoding device for spatio-temporal trajectories includes a processor and a memory. A computer program is stored in the memory, and when the computer program is loaded by the processor, the method described in any one of the above is executed.

[0048] The beneficial effects of the present invention include:

[0049] The multi-dimensional feature fusion encoding method for spatio-temporal trajectories proposed by the present invention can achieve deep fusion encoding of numerical features and categorical features, can provide a more unified and suitable input form for spatio-temporal trajectory applications based on deep learning models, is convenient for introducing the popular self-attention mechanism, and has a wider scope of application. In addition, compared with the normalization encoding method, the method proposed by the present invention ensures the stability of the corresponding meanings of feature values through label encoding, does not change with the change of the distribution of the training data set, and can improve the transferability of intelligent models. Description of the Drawings

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0051] Figure 1 is a flowchart of the steps of a multi-dimensional feature fusion encoding method for spatio-temporal trajectories in an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of the generation of a multi-dimensional feature fusion vector for spatio-temporal trajectories in an embodiment of the present invention. Detailed Embodiments

[0053] All features disclosed in all embodiments in this specification, or all steps in the methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or extended and replaced in any way.

[0054] In the inventive concept, aiming at the problem of multi-dimensional feature encoding of spatio-temporal trajectories, an intelligent model construction mainly based on the Transformer architecture is considered, and a spatio-temporal trajectory feature fusion encoding method is proposed to describe the multi-dimensional features of the target spatio-temporal trajectory in the form of vectors, which has a wider applicability. More specifically, in order to better solve the problem of multi-dimensional feature encoding of the target spatio-temporal trajectory, the method of the present invention fuses multi-dimensional features such as time, longitude, latitude, speed, heading, type, and country by using label encoding and vector embedding, and obtains a fused vector representation of the multi-dimensional features. This fused vector can be used as data for the intelligent model and can support various applications based on spatio-temporal trajectories.

[0055] In a more specific embodiment, as Figure 1 shown, a multi-dimensional feature fusion encoding method for spatio-temporal trajectories is specifically provided, including the following steps:

[0056] Step 1, label encoding of multi-dimensional features of spatio-temporal trajectories. Further, the features of the target spatio-temporal trajectory include but are not limited to numerical features such as time, longitude, latitude, speed, altitude, and heading, as well as categorical features such as country and type. Hereinafter, this embodiment takes 6 features of time, longitude, latitude, speed, heading, and type as examples for implementation description. In this embodiment, different features are respectively label-encoded to obtain the label-encoded values of each feature, where label encoding refers to mapping the feature values to a continuous integer starting from 0.

[0057] Step 1.1, label encoding of time features. Further, the time feature of the target spatio-temporal trajectory refers to the acquisition time of each trajectory point, which is generally composed of year, month, day, hour, minute, and second, such as 14:44:32 on November 15, 2024. The label encoding of the time feature can be based on the original year, month, day, hour, minute, and second information, or on information such as the number of weeks, day of the week, and holidays obtained from the original time data. The selection of time information is related to the actual application. This embodiment takes the encoding of hour, minute, and second information as an example for description.

[0058] First, the hour, minute, and second information in the time information is respectively extracted, and then it is converted into the corresponding label encoding. The calculation method is as follows:

[0059] h f =h + 1

[0060] m f =m + 1;

[0061] s f =s + 1

[0062] where h, m, and s are the specific values of hour, minute, and second respectively, and the value ranges are 0 to 23, 0 to 59, and 0 to 59 respectively, h f ,mf , s f are the tag encodings representing hours, minutes, and seconds respectively, and their value ranges are 0 to 24, 0 to 60, and 0 to 60 respectively. When the tag encoding value is 0, it means the encoding is a fill value.

[0063] Step 1.2, longitude and latitude feature tag encoding. Further, the target longitude feature value is between -180° and 180°, and the latitude feature value is between -90° and 90°. In order to perform tag encoding on the longitude and latitude values, in this embodiment, an equal-width discretization method is adopted. The value ranges of the continuous longitude and latitude variables are respectively divided into equal-width intervals, and then the variable values are mapped into the corresponding intervals. When performing equal-width discretization, the number of intervals needs to be determined, which is related to the specific application scenario. However, generally, the following rules need to be followed: after discretization, the adjacent longitude and latitude tag encodings should be as different as possible to avoid continuous repeating patterns.

[0064] If the number of intervals for dividing the target longitude is set to N lon , and the number of intervals for dividing the latitude is set to N lat , then the calculation methods for the longitude and latitude tag encodings are as follows:

[0065]

[0066] where lon and lat are the specific values of the longitude and latitude features of the spatio-temporal trajectory, lon f and lat f are the tag encodings of the longitude and latitude features respectively, and floor(·) represents rounding down. When the tag encoding value is 0, it means the encoding is a fill value.

[0067] Step 1.3, speed feature tag encoding. Speed feature tag encoding refers to discretizing continuous speed values into tag encodings, using the equal-width discretization method. When performing specific discretization, the setting of the speed interval number can be replaced by setting the speed division interval. Since the value range of the target speed is not fixed and can only be determined to be greater than or equal to 0, in actual processing, it is also necessary to set the upper limit of the speed value according to the requirements of the actual application scenario, that is, the maximum allowable speed value. If the upper limit of the speed value is set to V max , and the speed division interval is V p , then the calculation method for the target speed feature tag encoding is as follows:

[0068]

[0069] where v is the specific value of the target speed feature, and floor(·) represents rounding down. When the tag encoding value is 0, it means the encoding is a fill value.

[0070] Step 1.4, Heading Feature Label Encoding. Further, the heading feature label encoding refers to discretizing continuous heading values into label encodings using the equal-width discretization method. The target heading feature value ranges from 0 to 360°, and the heading division interval is set to C p , then the calculation method of the target speed feature encoding is as follows:

[0071] c f = floor(c / C p );

[0072] where c is the specific value of the target heading feature, and floor(·) represents rounding down. When the label encoding value is 0, it means the encoding is a padding value.

[0073] Step 1.5, Type Feature Label Encoding. Further, the target type feature has a limited number of values and is usually a non-numeric set. For example, the set of aircraft target types can be {passenger aircraft, cargo aircraft}. The target type label encoding refers to mapping each target type feature value to an integer starting from 0. If the set of target type feature values is represented as:

[0074] A = {a1, a2,... a i ,...., a n}

[0075] then the label encoding corresponding to the target type feature value a i is i, denoted as t f . It should be noted that when the target type feature value is not in the set A, the label encoding is directly set to 0.

[0076] Step 2, Generation of Embedding Vectors for Label Encodings. Further, through Step 1, this embodiment obtains a set of label encodings {h f , m f , s f , lon f , lat f , v f , c f , t f} for 6 features including time, longitude, latitude, speed, heading, and type, totaling 8 label encodings. Since neural network models generally require the input form to be a real-valued vector, in order to enable the neural network model to process label encodings, it is necessary to perform a vector embedding operation on the label encodings to convert them into a dense vector form, also known as an embedding vector. For each label encoding, its embedding vector is generated using the following method:

[0077] e m = M m δ m ;

[0078] Among them, e m is the embedding vector corresponding to the label encoding, m ∈ {h f , m f , s f , lon f , lat f , v f , c f , t f}, is the embedding matrix, D m represents the dimension size of the embedding vector corresponding to the label encoding m, S m represents the size of the value set corresponding to the label encoding m, represents the one-hot vector corresponding to the label encoding m.

[0079] Through the above method, the embedding vector corresponding to each label encoding can be obtained, which are respectively

[0080] Step 3, generation of the multi-dimensional feature fusion vector. Further, the generation of the multi-dimensional feature fusion vector means fusing the embedding vectors corresponding to multiple label encodings into one embedding vector for use as the input of the neural network model. Refer to Figure 2 , and the specific generation method is as follows:

[0081] First, splice the embedding vectors corresponding to multiple label encodings to obtain the spliced vector e:

[0082] e = ||e m , m ∈ {h f , m f , s f , lon f , lat f , v f , c f , t f};

[0083] Among them, || represents the vector splicing operation.

[0084] Then, through a non-linear transformation network, convert the spliced vector e into the multi-dimensional feature fusion vector e F . It should be noted that the non-linear transformation network can be a multi-layer network. Taking a 1-layer network as an example, the implementation method is as follows:

[0085] e F = σ(We + b)

[0086] Among them, are the network parameters, and σ(·) is the activation function.

[0087] The units involved in the embodiments of the present invention can be implemented in software or in hardware, and the described units can also be provided in a processor. Among them, the names of these units do not constitute a limitation to the units themselves in some cases.

[0088] According to one aspect of the embodiments of the present invention, there is provided a computer program product or a computer program, the computer program product or the computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the above various alternative implementation manners.

[0089] As another aspect, the embodiments of the present invention further provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or may exist separately without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the one or more programs are executed by an electronic device, the electronic device implements the methods described in the above embodiments.

Claims

1. A multi-dimensional feature fusion coding method for spatiotemporal trajectories, characterized in that: The following steps are involved: First, the numerical features and categorical features of the target spatiotemporal trajectory are label-encoded to obtain the integer encoding value of each feature; Then, the encoding value of each feature is vector-embedded to obtain the embedding vector of each feature; Finally, the embedding vectors of all features are concatenated and converted into multi-dimensional feature fusion vectors through a nonlinear transformation network.

2. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 1 is characterized in that: The label encoding of the numerical features and categorical features of the target spatiotemporal trajectory specifically includes the time feature label encoding process: Extract the hour, minute, and second information from the time information respectively, and then convert them into the corresponding label encoding. The calculation method is as follows: h f =h+1 m f =m+1; s f =s+1 Among them, h, m, and s are the specific values ​​of hours, minutes, and seconds, and the value ranges are 0 to 23, 0 to 59, and 0 to 59 respectively. f ,m f ,s f The label codes for hours, minutes, and seconds are in the range of 0 to 24, 0 to 60, and 0 to 60, respectively. When the label code value is 0, it indicates that the code is a fill value.

3. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 1 is characterized in that: The label encoding of the numerical features and categorical features of the target spatiotemporal trajectory specifically includes the latitude and longitude feature label encoding process: The equal-width discretization method is used to divide the value range of the longitude and latitude continuous variables into intervals of equal width, and then map the variable values ​​to the corresponding intervals. When the equal-width discretization is performed, the number of intervals is determined. If the number of target longitude intervals is set to N, lon , the number of latitude division intervals is N lat , then the label encoding of longitude and latitude is calculated as follows: Among them, lon and lat are the specific values ​​of the longitude and latitude characteristics of the space-time trajectory, lon f and lat f are the label codes of longitude and latitude features respectively. floor(·) means rounding down. When the label code value is 0, it means the code is a fill value.

4. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 1 is characterized in that: The label encoding of the numerical features and categorical features of the target spatiotemporal trajectory specifically includes the speed feature label encoding process: The continuous speed value is discretized into label code by equal width discretization method. When performing specific discretization, the speed division interval is set instead of the number of speed intervals. According to the actual application scenario requirements, the upper limit of the speed value is set. If the upper limit of the speed value is set to V max , the speed division interval is V p , then the target speed feature label encoding calculation method is as follows: Where v is the specific value of the target speed feature, floor(·) means rounding down; when the label code value is 0, it means that the code is a filler value.

5. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 1 is characterized in that: The label encoding of the numerical features and categorical features of the target spatiotemporal trajectory specifically includes the heading feature label encoding process: The equal-width discretization method is used to discretize the continuous heading value into label code. The target heading feature value is between 0 and 360°, and the heading division interval is set to C. p , then the target speed characteristic encoding is calculated as follows: c f =floor(c / C p )+1; Where c is the specific value of the target heading feature, floor(·) means rounding down; when the label code value is 0, it means that the code is a fill value.

6. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to any one of claims 1 to 5, characterized in that: The label encoding of the numerical features and categorical features of the target spatiotemporal trajectory specifically includes the following process: Map each target type feature value to an integer starting from 0; if the target type feature value set is represented as: A={a1,a2,...a i ,....,a n }; Then the target type feature takes the value a i The corresponding label is encoded as i and recorded as t f ; When the target type feature value is not in set A, the label code is directly set to 0.

7. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 6 is characterized in that: The step of embedding the encoding value of each feature into a vector comprises the following sub-steps: Suppose that after the label encoding step of the numerical features and categorical features of the target spatiotemporal trajectory, the label encoding set {h f ,m f ,s f ,lon f ,lat f ,v f ,c f ,t f }; Perform a vector embedding operation on the label code obtained by the label encoding step of the numerical features and categorical features of the target spatiotemporal trajectory, convert it into a dense vector form, and for each label code, generate its embedding vector using the following method: e m =M m d m ; Among them, e m is the embedding vector corresponding to the label encoding, m∈{h f ,m f ,s f ,lon f ,lat f ,v f ,c f ,t f }, is the embedding matrix, D m Indicates the dimension size of the embedding vector corresponding to the label encoding m, S m Indicates the size of the value set corresponding to the label code m, Represents the one-hot vector corresponding to the label encoding m; Through the above method, the embedding vectors corresponding to each label encoding are obtained, which are 8. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 7 is characterized in that: After concatenating all feature embedding vectors, the vectors are converted into multi-dimensional feature fusion vectors through a nonlinear transformation network, which specifically includes the following sub-steps: First, concatenate the embedding vectors corresponding to multiple label encodings to obtain the concatenated vector e: e=||e m ,m∈{h f ,m f ,s f ,lon f ,lat f ,v f ,c f ,t f }; in, || represents vector concatenation operation; Then, through the nonlinear transformation network, the concatenated vector e is converted into a multi-dimensional feature fusion vector e F .

9. The multi-dimensional feature fusion coding method for spatiotemporal trajectories according to claim 8 is characterized in that: The nonlinear transformation network includes a 1-layer network, and the specific implementation method is as follows: e F =σ(We+b); in, is the network parameter, and σ(·) is the activation function.

10. A multi-dimensional feature fusion encoding device for spatiotemporal trajectories, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, the method according to any one of claims 1 to 9 is executed.

Images