Co-traveling vehicle analysis method based on shared prototype space

Abstract

The invention belongs to the technical field of data analysis, and relates to a shared prototype space-based accompanying vehicle analysis method, which comprises the following steps of: collecting and preprocessing data, extracting vehicle information by using a feature extractor, aligning vehicles and spatio-temporal information by using a shared prototype space on the basis of extracted features, and taking the vehicle and spatio-temporal information as modeling for habits of vehicle owners. And according to the learned prototype space, habit characteristics of the vehicle owner are further analyzed, so that more sufficient characteristic modeling and fusion are realized, potential behavior information can be more sufficiently mined, and services such as recommendation for vehicle owners with similar habits can be performed.

Description

Technical Field

[0001] The invention belongs to the technical field of data analysis and relates to a method for analyzing traveling vehicles based on a shared prototype space. Background Art

[0002] In recent years, with the development of intelligent transportation systems, the collection and analysis of urban vehicle driving data has become increasingly important. By analyzing vehicle driving data, we can explore the hidden habits of car owners behind the data and find car owners with similar living and working habits, providing strong support for automotive aftermarket recommendations and other businesses.

[0003] The vehicle peer analysis algorithm usually analyzes the owner's life and work habits based on the vehicle's structured spatiotemporal data (vehicle information, time information, driving area information), and then clusters the owners through these habit information to make recommendations for owners with similar habits. Existing peer vehicle analysis methods usually use methods based on path overlap and methods based on deep learning. Among them, the method based on path overlap uses the vehicle's driving trajectory as a feature to calculate similarity and cluster, which fails to fully utilize spatiotemporal data and lacks interaction between information; the method based on deep learning uses vehicle information and spatiotemporal data as input, uses neural networks to learn corresponding features, and only splices the features together for learning. The interaction between features is insufficient, and the potential behavioral habit information cannot be fully mined. Therefore, it is urgent to design a new peer vehicle analysis method based on shared prototype space. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a peer vehicle analysis method based on a shared prototype space. The method uses a multi-source fusion Transformer to extract vehicle information and model and fuse time information. Based on the learned features, the shared prototype space is used to align vehicle and spatiotemporal information as a model for the owner's habits. The owner's habit characteristics are further analyzed based on the learned prototype space.

[0005] To achieve the above object, the present invention adopts the following technical solutions: A peer vehicle analysis method based on a shared prototype space specifically includes the following steps: S1. Data collection: Collect the structured spatiotemporal data of vehicles, which includes vehicle information, time information and space information; S2, data preprocessing: preprocess the structured spatiotemporal data to obtain vehicle features, time features and spatial features; S3, vehicle feature extraction: The vehicle features and time features are concatenated and input into the feature extractor to extract the vehicle and time features again, and the time features are fused into the vehicle features using cross attention. The feature extractor uses a multi-source feature fusion Transformer backbone network; S4, shared prototype space learning and training: embed the vehicle features obtained in step S3 into the shared prototype space, capture the habit features contained in the vehicle features and time feature information, and learn and train by aligning the vehicle and space information with the prototype features; S5. Analysis of peer vehicles: Different clustering is performed based on date types, and automotive aftermarket business recommendations are made for car owners with similar habits.

[0006] As a further technical solution of the present invention, the vehicle information described in step S1 is the information of the vehicle itself, including the license plate number, vehicle model and type; the time information is the specific time, and also includes holidays and morning and evening rush hour information; the spatial information is the road section and area where the vehicle is located and the basic information of the area itself.

[0007] As a further technical solution of the present invention, the vehicle characteristics in step S2 are recorded as , including vehicle price, vehicle type and license plate number features, where vehicle price and vehicle type are input as discrete features, and the license plate number is obtained through the embedding layer; the time feature is recorded as , taking the special marks of the date as discrete features, the special marks include whether it is a working day or a holiday. Each date in a year is embedded to obtain the date feature, and the time is divided into 1 interval every hour, and the time feature is obtained by embedding; the spatial feature is recorded as , the city is pre-divided into different areas, and the type, size, and traffic volume of each area are used as discrete features.

[0008] As a further technical solution of the present invention, the multi-source feature fusion Transformer backbone network in step S3 includes Layer, The layer contains two self-attention modules, a cross-attention module and an FFN layer; the working process of each layer is as follows: The two self-attention modules are used to extract the interaction between sub-features within the same feature, which can be expressed as follows: , ,in , Respectively Vehicle characteristics and time characteristics of the layer; The cross attention module is used to fuse the temporal features into the vehicle features. As a key to the attention mechanism Sum , vehicle characteristics As a query , extracting from time features to key features for fusion, expressed as: , , ,in represents a linear transformation, represents the activation function; The FFN layer further extracts features through multi-layer MLP to obtain the output of this layer ; go through After layer calculation, the final vehicle characteristics and time characteristics are obtained .

[0009] As a further technical solution of the present invention, the specific process of step S4 is: before the learning and training begins, first initialize Prototype Features , where the dimension of each prototype feature is and same, Represents a certain habit, combines time and area as an index, and randomly selects The prototype features are used as the expression of the index. Specifically, assuming that areas, with an interval of 1 hour, there are a total of indexes, each index Representative Region No. The index of the prototype feature corresponding to the time region. During learning and training, the feature extractor and the prototype feature are optimized respectively; the process of optimizing the feature extractor is: First select one of the data. The index of the prototype feature corresponding to the time zone in the data is , after step B S3, the feature is obtained ,Will Concatenate them and send them into two linear layers to learn a two-dimensional weight vector ,use Get the final fusion features : ,in They are The corresponding two-dimensional weight vector; Recalculate Similarity scores to each prototype: ,maximize The sum of the similarities of the corresponding positions in , the loss function is: At the same time, according to Update each prototype feature ,in is a hyperparameter; The multi-task learning method is used to predict the corresponding spatial features based on the prototype features and time features of the vehicle, and the type of the area where the vehicle is currently located is predicted based on the learned features: , ,in For splicing operation, is the owner's habit feature; and uses cross entropy loss to guide learning: ,in To predict the results, is the predicted probability value; The specific process of optimizing prototype features is as follows: after updating the parameters of the feature extractor, each index is updated according to the owner information, that is, a new prototype is reassigned to each index. , calculate its difference with each prototype Score : ,in For car owners With prototype The similarity score of =1 means the owner In time Passing area , otherwise 0, select the one with the highest score The prototype is used as the new prototype feature corresponding to the index.

[0010] As a further technical solution of the present invention, the specific process of step S5 is: for For a certain period of time, the vehicle information is divided into The time information is stitched together and input into the prototype shared space to get this Score of time period , based on the obtained scores and prototype features, the owner's habitual feature expression is obtained : ,in , and then The data are stored in the database as the characteristics of the car owners, and the car owners are clustered through the K-means clustering algorithm; or for a certain car owner, the nearest neighbor search algorithm is used to find car owners with similar habits for further analysis of the car owner.

[0011] Compared with the prior art, the present invention has the following advantages: (1) A multi-source fusion Transformer is proposed to extract vehicle and time information separately, and use cross-attention to fuse the time information into the vehicle information to achieve more comprehensive feature modeling and fusion.

[0012] (2) Use the shared prototype space to represent different behavioral habit information, align the vehicle and space information with the prototype, and more fully explore the potential behavioral information. From the extracted features, we can mine the habit characteristics of the car owner and make recommendations for car owners with similar habits. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a workflow diagram of the peer vehicle analysis method based on shared prototype space described in the present invention. DETAILED DESCRIPTION

[0014] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings.

[0015] Example: like Figure 1 As shown, this embodiment provides a method for analyzing traveling vehicles based on a shared prototype space, which specifically includes the following steps: S1. Data collection: Collect structured spatiotemporal data of vehicles. Structured spatiotemporal data includes vehicle information, time information and spatial information. Vehicle information refers to the information of the vehicle itself, including license plate number, vehicle model and type. Time information refers to the specific time, including holidays and morning and evening rush hour information. Spatial information refers to the road section and area where the vehicle is located and the basic information of the area itself.

[0016] S2. Data preprocessing: Preprocess structured spatiotemporal data to obtain vehicle features , Time characteristics and spatial characteristics ; Vehicle features Including vehicle price, vehicle type and license plate number features, where vehicle price and vehicle type are input as discrete features, and the license plate number is obtained through the embedding layer; time features The special marks of the date are used as discrete features. The special marks include whether it is a working day or a holiday. Each date in a year is embedded to obtain the date feature. The time is divided into an interval every hour, and the time feature is obtained by embedding. Spatial features The city is divided into different areas in advance, and the type (such as commercial area, etc.), size (small, medium and large), and traffic volume of each area are used as discrete features. These discrete features are used as labels to guide model learning in subsequent training.

[0017] S3. Vehicle feature extraction: extract vehicle features and time characteristics After splicing, the input feature extractor extracts vehicle and time features again, and uses cross attention to fuse the time features into vehicle features. The feature extractor uses a multi-source feature fusion Transformer backbone network; the multi-source feature fusion Transformer backbone network includes Layer, The layer contains two self-attention modules, a cross-attention module and an FFN layer; the working process of each layer is as follows: The two self-attention modules are used to extract the interaction between sub-features within the same feature, which can be expressed as follows: , ,in , Respectively Vehicle characteristics and time characteristics of the layer; The cross attention module is used to fuse the temporal features into the vehicle features. As a key to the attention mechanism Sum , vehicle characteristics As a query , extracting from time features to key features for fusion, expressed as: , , ,in represents a linear transformation, represents the activation function; The FFN layer further extracts features through multi-layer MLP to obtain the output of this layer ; go through After layer calculation, the final vehicle characteristics and time characteristics are obtained .

[0018] S4, shared prototype space learning and training: embed the vehicle features obtained in step S3 into the shared prototype space, capture the habit features contained in the vehicle features and time feature information, and learn and train by aligning the vehicle and space information with the prototype features; the specific process is: before the learning and training begins, initialize Prototype Features , where the dimension of each prototype feature is and same, Represents a certain habit, such as going to work, entertainment, etc. If a car appears at a specific place and time, it contains the possible work and life habits. The time and area are combined as an index, and a random selection is made under each index. The prototype features are used as the expression of the index. Specifically, assuming that areas, with an interval of 1 hour, there are a total of indexes, each index Representative Region No. The index of the prototype feature corresponding to the time region. During learning and training, the feature extractor and the prototype feature are optimized respectively; the process of optimizing the feature extractor is: First select one of the data. The index of the prototype feature corresponding to the time zone in the data is , after step B S3, the feature is obtained ,Will Concatenate them and send them into two linear layers to learn a two-dimensional weight vector ,use Get the final fusion features : ,in They are The corresponding two-dimensional weight vector; Recalculate Similarity scores to each prototype: , since the goal is to index Will Aligned with the prototype space, passing through the same area on weekends and weekdays may correspond to different habits, so only maximizing The sum of the similarities of the corresponding positions in the , rather than maximizing the score of each prototype, the loss function is: At the same time, according to Update each prototype feature ,in is a hyperparameter; To ensure that the prototype features include habit features, multi-task learning is used to predict the corresponding spatial features based on the prototype features and time features corresponding to the vehicle. For example, if the vehicle owner a and the time 9 a.m. on a weekday are input, the type of area where the vehicle is currently located (such as a commercial area) is predicted based on the learned features: , ,in For splicing operation, is the owner's habit feature; and uses cross entropy loss to guide learning: ,in To predict the results, is the predicted probability value; The specific process of optimizing prototype features is as follows: after updating the parameters of the feature extractor, each index is updated according to the owner information, that is, a new prototype is reassigned to each index. , calculate its difference with each prototype Score : ,in For car owners With prototype The similarity score of =1 means the owner In time Passing area , otherwise 0, select the one with the highest score The prototype is used as the new prototype feature corresponding to the index.

[0019] S5. Analysis of peer vehicles: Different clusters are performed based on date types to recommend aftermarket services to car owners with similar habits. For the working days in January, the vehicle information is combined with 24 time information (month is January, working day, time is ), the vehicle information is spliced ​​with the 24 time information and input into the prototype shared space to obtain the scores of these 24 time periods , based on the obtained scores and prototype features, the owner's habitual feature expression is obtained : ,in , and then The data are stored in the database as the characteristics of the car owners, and the car owners are clustered through the K-means clustering algorithm; or for a certain car owner, the nearest neighbor search algorithm is used to find car owners with similar habits for further analysis of the car owner.

[0020] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. The systems and algorithms not described in detail in the present invention are all general techniques. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0021] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A peer vehicle analysis method based on a shared prototype space, characterized in that: The specific steps include: S1. Data collection: Collect the structured spatiotemporal data of vehicles, which includes vehicle information, time information and space information; S2, data preprocessing: preprocess the structured spatiotemporal data to obtain vehicle features, time features and spatial features; S3, vehicle feature extraction: The vehicle features and time features are concatenated and input into the feature extractor to extract the vehicle and time features again, and the time features are fused into the vehicle features using cross attention. The feature extractor uses a multi-source feature fusion Transformer backbone network; S4, shared prototype space learning and training: embed the vehicle features obtained in step S3 into the shared prototype space, capture the habit features contained in the vehicle features and time feature information, and learn and train by aligning the vehicle and space information with the prototype features; S5. Analysis of peer vehicles: Different clustering is performed based on date types, and automotive aftermarket business recommendations are made for car owners with similar habits.

2. The method for analyzing traveling vehicles based on a shared prototype space according to claim 1 is characterized in that: The vehicle information in step S1 is the information of the vehicle itself, including the license plate number, vehicle model and type; the time information is the specific time, including holidays and morning and evening rush hour information; the spatial information is the road section and area where the vehicle is located and the basic information of the area itself.

3. The method for analyzing traveling vehicles based on a shared prototype space according to claim 2 is characterized in that: The vehicle characteristics in step S2 are recorded as , including vehicle price, vehicle type and license plate number features, where vehicle price and vehicle type are input as discrete features, and the license plate number is obtained through the embedding layer; the time feature is recorded as , taking the special marks of the date as discrete features, the special marks include whether it is a working day or a holiday. Each date in a year is embedded to obtain the date feature, and the time is divided into 1 interval every hour, and the time feature is obtained by embedding; the spatial feature is recorded as , the city is pre-divided into different areas, and the type, size, and traffic volume of each area are used as discrete features.

4. The method for analyzing traveling vehicles based on a shared prototype space according to claim 3 is characterized in that: The multi-source feature fusion Transformer backbone network in step S3 includes Layer, The layer contains two self-attention modules, a cross-attention module and an FFN layer; the working process of each layer is as follows: The two self-attention modules are used to extract the interaction between sub-features within the same feature, which can be expressed as follows: , ,in , Respectively Vehicle characteristics and time characteristics of the layer; The cross attention module is used to fuse the temporal features into the vehicle features. As a key to the attention mechanism Sum , vehicle characteristics As a query , extracting from time features to key features for fusion, expressed as: , , ,in represents a linear transformation, represents the activation function; The FFN layer further extracts features through multi-layer MLP to obtain the output of this layer ; go through After layer calculation, the final vehicle characteristics and time characteristics are obtained .

5. The method for analyzing traveling vehicles based on a shared prototype space according to claim 4 is characterized in that: The specific process of step S4 is as follows: before the learning and training begins, initialize Prototype Features , where the dimension of each prototype feature is and same, Represents a certain habit, combines time and area as an index, and randomly selects The prototype features are used as the expression of the index. Specifically, assuming that areas, with an interval of 1 hour, there are a total of indexes, each index Representative Region No. The index of the prototype feature corresponding to the time region. During learning and training, the feature extractor and the prototype feature are optimized respectively; the process of optimizing the feature extractor is: First select one of the data. The index of the prototype feature corresponding to the time zone in the data is , after step B S3, the feature is obtained ,Will Concatenate them and send them into two linear layers to learn a two-dimensional weight vector ,use Get the final fusion features : ,in They are The corresponding two-dimensional weight vector; Recalculate Similarity scores to each prototype: ,maximize The sum of the similarities of the corresponding positions in , the loss function is: At the same time, according to Update each prototype feature ,in is a hyperparameter; The multi-task learning method is used to predict the corresponding spatial features based on the prototype features and time features of the vehicle, and the type of the area where the vehicle is currently located is predicted based on the learned features: , ,in For splicing operation, is the owner's habit feature; and uses cross entropy loss to guide learning: ,in To predict the results, is the predicted probability value; The specific process of optimizing the prototype features is as follows: after updating the parameters of the feature extractor, each index is updated according to the owner information, that is, a new prototype is reassigned to each index. , calculate its difference with each prototype Score : ,in For car owners With prototype The similarity score of =1 means the owner In time Passing area , otherwise 0, select the one with the highest score The prototype is used as the new prototype feature corresponding to the index.

6. The method for analyzing traveling vehicles based on a shared prototype space according to claim 5 is characterized in that: The specific process of step S5 is: For a certain period of time, the vehicle information is divided into The time information is stitched together and input into the prototype shared space to get this Score of time period , based on the obtained scores and prototype features, the owner's habitual feature expression is obtained : ,in , and then The data are stored in the database as the characteristics of the car owners, and the car owners are clustered through the K-means clustering algorithm; or for a certain car owner, the nearest neighbor search algorithm is used to find car owners with similar habits for further analysis of the car owner.

Images