Motorcycle user riding portrait construction method and system based on time sequence characteristics
Through the motorcycle user cycling portrait construction method based on timing characteristics, the motorcycle user cycling data is analyzed using convolutional neural network and clustering model, the problem of timing feature processing of motorcycle riding data in the existing technology is solved, and more accurate user cycling portrait construction is achieved, and the competitiveness of motorcycle companies is enhanced.
Patent Information
- Application Number
- CN202411790257.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-06
- Publication Date
- 2025-03-28
AI Technical Summary
The prior art is difficult to effectively process the timing characteristics of motorcycle riding data, resulting in poor user cycling portrait construction results, and large data volume and complex parameters, making it difficult to perform effective timing processing and dimensionality reduction.
A method for constructing motorcycle user cycling portraits based on timing features is proposed. By obtaining user historical cycling data from the database, cleaning and timing processing is performed, the cycling behavior model constructed by convolutional neural network is used to model and analyze the cycling feature sequence, generate user cycling representation vectors, and classify users through clustering models to define cycling labels.
The timing characteristics of motorcycle user riding data have been fully explored, and a user riding portrait is built that is more in line with user riding preferences and styles, so that motorcycle companies can better understand user needs, optimize products and services, and enhance corporate competitiveness.
Smart Images
Figure CN119719830A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of motorcycle user riding, and in particular to a method and system for constructing a motorcycle user riding portrait based on time series features. Background Art
[0002] With the rapid development of my country's automobile industry and the transformation of transportation structure, the motorcycle industry has also ushered in a new development cycle. In addition to serving as a means of transportation for urban travel, it has also become a new facility for urban leisure and entertainment. In addition to paying attention to vehicle performance and quality, motorcycle consumers have increasingly diversified demands, which has promoted the personalized and intelligent development of the motorcycle market and the technological innovation of automobile companies.
[0003] Similar to the automotive industry, the motorcycle industry has also begun to use big data technology to analyze user riding data, explore user riding preferences, build user riding portraits, and provide users with personalized services that are more in line with their riding habits. At this stage, the construction of motorcycle user riding portraits is mostly achieved by compressing user riding information into independent statistical feature values and then clustering them. Although statistical feature values can reflect the user's riding style to a certain extent, they cannot reflect the essence of the time series of riding data, and the modeling effect is poor when input into the model. In addition, the data collected during motorcycle driving is huge in volume and complex in parameters. How to process it in time series and reduce its dimension is also a difficult problem for motorcycle user riding modeling. Summary of the invention
[0004] In view of the deficiencies in the prior art, the present invention proposes a method and system for constructing a motorcycle user riding portrait based on time series features.
[0005] The specific technical solutions are as follows:
[0006] A method for constructing a motorcycle user riding portrait based on time series features comprises the following steps:
[0007] S1: Obtain the key information data of the user's historical riding from the database, and clean the obtained riding data to obtain the cleaned key information data of the user's riding;
[0008] S2: Perform time series processing and preliminary feature extraction on the cleaned user riding key information data to obtain a riding feature sequence;
[0009] S3: Modeling and analyzing the cycling feature sequence using the cycling behavior model to obtain a user cycling representation vector; the cycling behavior model is constructed based on a convolutional neural network, and each convolution layer in the network extracts and analyzes local information of the sequence by sliding the convolution kernel on the sequence. Through the stacking of convolution layers, the receptive field of the convolution kernel is gradually expanded, thereby obtaining more comprehensive and deeper sequence features, and finally outputting a representation vector representing the user's cycling behavior;
[0010] S4: Use the clustering model to classify users according to the similarity between their riding representation vectors, and define user riding labels based on the classification results.
[0011] Furthermore, in S1, the key riding information data includes: frame number, reporting time, current riding time, speed, acceleration, throttle acceleration opening and torque.
[0012] Furthermore, in S1, data cleaning includes removing riding data containing null values, outliers and riding data in a parking state.
[0013] Furthermore, S2 is implemented by the following sub-steps:
[0014] S2.1: Aggregate the key cycling information data belonging to the same user into a group, and then sort the key cycling information data of each user according to time;
[0015] S2.2. Perform preliminary feature extraction on the grouped and sorted key riding information data: construct a one-dimensional feature extraction window, and set the window length and step size; the feature extraction window slides on each key riding information data according to the step size, and slices it according to the set window length, and calculates the maximum value, variance and average value of the data in each slice, and uses these statistical features to represent the riding information in the slice to achieve data dimensionality reduction; the statistical features of each slice are combined to obtain the riding feature sequence of each key riding information data of the user, and the riding feature sequence includes three columns of data: maximum value, variance and average value.
[0016] Furthermore, in S3, the riding feature sequence of each riding key information data of the user obtained in S2 is firstly input into the riding behavior model to obtain the representation vector of each riding key information data; then the representation vectors of all the riding key information of the user are mean-pooled and concatenated to obtain the representation vector representing the riding behavior of the user.
[0017] Furthermore, in S3, the convolutional neural network constituting the riding behavior model is optimized as follows:
[0018] (1) Setting a receptive field expansion mechanism: When training the neural network, a receptive field diffusion parameter d is set, d ≥ 1, and a zero matrix of size w × (d-1) is embedded between each two adjacent sets of weight parameters of the convolution kernel, so that the convolutional neural network can fully extract sequence information while saving the model's computational complexity. w is the width of the convolution kernel.
[0019] (2) Neuron pruning: Remove redundant neurons in each convolutional layer without affecting the accuracy of the cycling behavior model.
[0020] Furthermore, S4 is implemented by the following sub-steps:
[0021] S4.1: Input the user riding representation vector obtained in S3 into a clustering model, and the clustering model classifies the users based on a similarity measure between the riding representation vectors;
[0022] The clustering model is implemented based on the K-means algorithm. The cluster center is iteratively optimized with the goal of minimizing the sum of the distances between the riding representation vectors of all users in each cluster and the cluster center vector. Finally, users with high similarity in riding representation vectors are clustered in the same cluster.
[0023] S4.2: The user in each category that is closest to the cluster center vector is taken as the representative of the cycling users in that category. The cycling feature sequence of the representative user is analyzed and plotted. The user cycling label is defined according to the cycling style and characteristics of the representative user. The cycling label of the representative user is used as the label of the cycling behavior category to construct a cycling portrait of the cycling users in that category.
[0024] A motorcycle user riding portrait construction system based on time series features, used to implement the motorcycle user riding portrait construction method based on time series features, including: a data acquisition module, a time series feature extraction module, a time series feature modeling module, a user portrait construction module, and a user portrait storage module;
[0025] The data acquisition module is used to obtain key information data of user riding from the database and clean the data;
[0026] The time series feature extraction module is used to group the cleaned user riding key information data and sort them in time series, as well as to extract preliminary features to obtain a user riding feature sequence;
[0027] The temporal feature modeling module is used to model and analyze the user's riding feature sequence using the riding behavior model, and output a riding representation vector containing the user's riding habits and style;
[0028] The user portrait construction module is used to classify users and construct user portraits according to the similarity of riding representation vectors; the module first uses the built-in clustering model to cluster the user riding representation vectors, and clusters users with high riding similarity into one category; based on the clustering results, the user closest to the cluster center vector in each category is used as the representative user of the category, and a riding characteristic curve is generated according to the riding characteristic sequence of the user, and the user riding label is customized according to the riding characteristics of the representative user; the riding label of the representative user will also be used as the label of the category;
[0029] The user portrait storage module is used to store the riding categories output by the user portrait construction module, where the categories represent users, all users in the categories, and category labels.
[0030] The beneficial effects of the present invention are:
[0031] The present invention proposes a method and system for constructing a motorcycle user riding portrait based on time series characteristics, which solves the problem of huge amount of motorcycle riding data and complex parameters, fully mines the time series characteristics of motorcycle user riding data, and constructs a user riding portrait that is more in line with the user's riding preferences and style, so that motorcycle companies can better understand user needs, optimize products and services, and enhance corporate competitiveness, while promoting the intelligent development and technological innovation of the motorcycle industry. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a flowchart of a method for constructing a motorcycle user riding portrait based on time series features proposed in an embodiment of the present invention.
[0033] Figure 2 It is a schematic diagram of a feature extraction window for performing preliminary feature extraction on grouped and sorted riding key information data in an embodiment of the present invention.
[0034] Figure 3 Schematic diagram of the receptive field diffusion mechanism for optimizing the convolutional neural network in an embodiment of the present invention.
[0035] Figure 4 It is a structural diagram of a system for constructing a motorcycle user riding portrait based on time series features proposed in an embodiment of the present invention. DETAILED DESCRIPTION
[0036] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effect of the present invention will become more clear. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0037] like Figure 1 As shown, a method for constructing a motorcycle user riding portrait based on time series features includes the following steps:
[0038] S1: Obtain the key information data of the user's historical riding from the database, and clean the obtained riding data to obtain the cleaned key information data of the user's riding.
[0039] Among them, the key riding information data includes: frame number, reporting time, current riding time, speed, acceleration, throttle acceleration opening and torque.
[0040] Data cleaning mainly refers to removing riding data containing null values, outliers, and parking status.
[0041] S2: Perform time series processing and preliminary feature extraction on the cleaned user riding key information data to obtain a riding feature sequence. This is achieved through the following sub-steps:
[0042] S2.1: Group and sort the cleaned user riding key information data based on user and time dimensions.
[0043] Specifically, the key cycling information data of the same "frame number" is regarded as the cycling data of the same user. The key cycling information data belonging to the same user are first aggregated into a group using the "frame number" information, and then the key cycling information data of each user are arranged in descending order according to the "reporting time" to ensure that the key cycling information data of each user are arranged in chronological order.
[0044] S2.2: Perform preliminary feature extraction on the grouped and sorted riding key information data to obtain a riding feature sequence for each user.
[0045] The acquired user riding key information data is collected by the vehicle-side sensor and uploaded to the database. However, due to the high frequency of sensor collection, the amount of riding key information data is still relatively large and the distribution is too dense after cleaning. If all of these riding key information data are input into the subsequent riding behavior model, it will affect the model's operating performance and efficiency. Therefore, this embodiment constructs a feature extraction window to perform preliminary feature extraction on the grouped and sorted riding key information data to achieve data dimensionality reduction. The specific operations are as follows:
[0046] (2.2.1) Construct a one-dimensional feature extraction window and determine the length and moving step of the feature extraction window. The window length here refers to the time range covered by the window. In order to fully realize the data dimensionality reduction effect and reduce repeated calculations and redundant information, the step length is preferably the window length.
[0047] (2.2.2) Let the feature extraction window slide on a certain user's riding key information data arranged in time sequence with a set step length, and slice it according to the set window length.
[0048] (2.2.3) Calculate the maximum value, variance and mean value of the data in each slice, and use these statistical features to represent the riding information in the slice to achieve data dimensionality reduction.
[0049] (2.2.4) Collect the statistical features of each slice of the key information data of the user's ride to obtain a riding feature sequence of the key information data of the user's ride that includes three columns of data: maximum value, variance, and average value.
[0050] (2.2.5) Determine whether the data of each key riding information of the user has completed data dimensionality reduction. If so, execute the next step; if not, return to step (2.2.2) to perform the next data dimensionality reduction.
[0051] (2.2.6) Determine whether to traverse the key riding information data of all users. If so, end the preliminary feature extraction and output the riding feature sequence of each user; if not, return to step (2.2.2).
[0052] like Figure 2 As shown, in this embodiment, the data acquisition frequency of the sensor is 100ms / time, the time range covered by the feature extraction window (i.e., the window length) is 5s, and the step length is 5s. Taking the cycling speed in the key cycling information data as an example, the feature extraction window starts to slide on the cycling speed data of user A with the "reporting time" 8:00:00.000 as the initial position. At the initial position, the window covers the cycling speed data from 8:00:00.000 to 8:00:05.000. The maximum value, average value, and variance of the cycling speed data within 5s are calculated to obtain the first set of eigenvalues {71, 69, 0.12}, and the first set of eigenvalues is added to the first row position of the cycling speed feature sequence of user A. After the first set of eigenvalues is added to the feature sequence, the feature extraction window slides to the position from 8:00:05.100 to 8:00:10.100, and the maximum value, average value, and variance of the riding speed data in the window are calculated to obtain the second set of eigenvalues {80, 73, 0.2}, which are added to the second row of the riding speed feature sequence of user A. The subsequent feature extraction and feature sequence construction are similar.
[0053] S3: Use the riding behavior model to model and analyze the riding feature sequence to obtain the user's riding representation vector. Specifically, the riding feature sequence obtained in S2 is input into the riding behavior model, and each riding feature sequence of the user is modeled and analyzed to obtain the representation vector of the riding key information. Then, the representation vectors of all the riding key information are mean-pooled and concatenated to finally obtain the representation vector representing the user's riding behavior.
[0054] The cycling behavior model is built based on a convolutional neural network. Each convolution layer in the network extracts and analyzes the local information of the sequence by sliding the convolution kernel on the sequence. Through the stacking of convolution layers, the receptive field of the convolution kernel is gradually expanded, thereby obtaining more comprehensive and deeper sequence features, and finally outputting a representation vector representing the user's cycling behavior.
[0055] Furthermore, the present invention optimizes the convolutional neural network as follows:
[0056] (1) Setting the receptive field expansion mechanism: In order to enable the convolutional neural network to fully extract sequence information while saving the model's computational complexity, such as Figure 3 As shown in the figure, a receptive field diffusion parameter d (d ≥ 1) is set during neural network training. This parameter embeds a zero matrix of size w×(d-1) between each two adjacent weight parameters of the convolution kernel (w is the width of the convolution kernel). After the matrix is embedded, the receptive field in the convolution kernel is expanded by about (d-1) times without increasing the training parameters. At the same time, the number of stacked convolution layers of the model is controlled, which can effectively save computing resources when modeling long sequence data.
[0057] (2) Neuron pruning: Since convolutional neural networks are prone to overfitting, redundant neurons in each convolutional layer are removed without affecting the accuracy of the model to improve the efficiency of model operation and optimization.
[0058] S4: Use the clustering model to classify users according to the similarity between their riding representation vectors, and define user riding labels based on the classification results to complete the construction of user riding portraits. This is achieved through the following sub-steps:
[0059] S4.1: The user riding representation vector obtained in S3 is input into the clustering model, and the clustering model classifies the users based on the similarity measurement between the riding representation vectors.
[0060] The clustering model is implemented based on the K-means algorithm, which iteratively optimizes the cluster center vector with the goal of minimizing the sum of the distances from the riding representation vectors of all users in each class after clustering to the cluster center vector, and ultimately clusters users with high similarity in riding representation vectors into the same class.
[0061] The calculation formula of the distance from the user's riding characterization vector to its cluster center vector is:
[0062]
[0063] Among them, x is the user riding representation vector, C is the cluster center vector of the class to which x belongs, m is the vector dimension, and x j is the j-th attribute value of the user's riding characterization vector, C j is the attribute value of the j-th dimension of the cluster center vector.
[0064] S4.2: After the clustering model classifies users based on their riding characterization vectors, in order to more intuitively analyze the riding characteristics of each category, the user in each category that is closest to the cluster center vector is taken as the representative of the riding users in that category. The riding feature sequence of the representative user is analyzed and plotted, and the user riding label is defined based on the riding style and characteristics of the representative user. The riding label of the representative user is used as the label of the riding behavior category to construct a riding portrait of the riding users in that category.
[0065] In order to implement the above-mentioned method for constructing a motorcycle user portrait based on riding time sequence characteristics, this embodiment also proposes a motorcycle user portrait construction system based on riding time sequence characteristics, such as Figure 4 As shown, the system includes: a data acquisition module, a time series feature extraction module, a time series feature modeling module, a user portrait construction module, and a user portrait storage module.
[0066] The data acquisition module is used to obtain key information data of user riding from the database and clean the data.
[0067] The time series feature extraction module is used to group the cleaned user riding key information data and sort them in time series, as well as perform preliminary feature extraction to obtain the user riding feature sequence.
[0068] The time series feature modeling module is used to model and analyze the user's riding feature sequence using the riding behavior model, and output a riding representation vector that contains the user's riding habits and style.
[0069] The user portrait construction module is used to classify users and construct user portraits based on the similarity of riding representation vectors. The module first uses the built-in clustering model to cluster the user riding representation vectors and cluster users with high riding similarity into one category. Based on the clustering results, the user closest to the cluster center vector in each category is used as the representative user of the category, and a riding characteristic curve is generated based on its riding characteristic sequence. The user riding label is customized based on the riding characteristics of the representative user, and the riding label of the representative user will also be used as the label of the category.
[0070] The user portrait storage module is used to store the riding categories, category representative users, all users in the category, and category labels output by the user portrait construction module.
[0071] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.
Claims
1. A method for constructing a motorcycle user riding portrait based on time series features, characterized in that: The following steps are involved: S1: Obtain the key information data of the user's historical riding from the database, and clean the obtained riding data to obtain the cleaned key information data of the user's riding; S2: Perform time series processing and preliminary feature extraction on the cleaned user riding key information data to obtain a riding feature sequence; S3: Modeling and analyzing the cycling feature sequence using the cycling behavior model to obtain a user cycling representation vector; the cycling behavior model is constructed based on a convolutional neural network, and each convolution layer in the network extracts and analyzes local information of the sequence by sliding the convolution kernel on the sequence. Through the stacking of convolution layers, the receptive field of the convolution kernel is gradually expanded, thereby obtaining more comprehensive and deeper sequence features, and finally outputting a representation vector representing the user's cycling behavior; S4: Use the clustering model to classify users according to the similarity between their riding representation vectors, and define user riding labels based on the classification results.
2. The method for constructing a motorcycle user riding portrait based on time series features according to claim 1 is characterized in that: In S1, the key riding information data includes: frame number, reporting time, current riding time, speed, acceleration, throttle acceleration opening and torque.
3. The method for constructing a motorcycle user riding portrait based on time series features according to claim 1 is characterized in that: In S1, data cleaning includes removing riding data containing null values, outliers, and riding data in a parking state.
4. The method for constructing a motorcycle user riding portrait based on time series features according to claim 1, characterized in that: The S2 is implemented by the following sub-steps: S2.1: Aggregate the key cycling information data belonging to the same user into a group, and then sort the key cycling information data of each user according to time; S2.
2. Perform preliminary feature extraction on the grouped and sorted key riding information data: construct a one-dimensional feature extraction window, and set the window length and step size; the feature extraction window slides on each key riding information data according to the step size, and slices it according to the set window length, and calculates the maximum value, variance and average value of the data in each slice, and uses these statistical features to represent the riding information in the slice to achieve data dimensionality reduction; the statistical features of each slice are combined to obtain the riding feature sequence of each key riding information data of the user, and the riding feature sequence includes three columns of data: maximum value, variance and average value.
5. The method for constructing a motorcycle user riding portrait based on time series features according to claim 1, characterized in that: In S3, firstly, the riding feature sequence of each riding key information data of the user obtained in S2 is input into the riding behavior model to obtain the representation vector of each riding key information data; then, the representation vectors of all the riding key information of the user are mean-pooled and concatenated to obtain the representation vector representing the riding behavior of the user.
6. The method for constructing a motorcycle user riding portrait based on time series features according to claim 1, characterized in that: In S3, the convolutional neural network constituting the cycling behavior model is optimized as follows: (1) Setting a receptive field expansion mechanism: When training the neural network, a receptive field diffusion parameter d is set, d ≥ 1, and a zero matrix of size w × (d-1) is embedded between each two adjacent sets of weight parameters of the convolution kernel, so that the convolutional neural network can fully extract sequence information while saving the model's computational complexity. w is the width of the convolution kernel. (2) Neuron pruning: Remove redundant neurons in each convolutional layer without affecting the accuracy of the cycling behavior model.
7. The method for constructing a motorcycle user riding portrait based on time series features according to claim 1, characterized in that: The S4 is implemented by the following sub-steps: S4.1: Input the user riding representation vector obtained in S3 into a clustering model, and the clustering model classifies the users based on a similarity measure between the riding representation vectors; The clustering model is implemented based on the K-means algorithm. The cluster center is iteratively optimized with the goal of minimizing the sum of the distances between the riding representation vectors of all users in each cluster and the cluster center vector. Finally, users with high similarity in riding representation vectors are clustered in the same cluster. S4.2: The user in each category that is closest to the cluster center vector is taken as the representative of the cycling users in that category. The cycling feature sequence of the representative user is analyzed and plotted. The user cycling label is defined according to the cycling style and characteristics of the representative user. The cycling label of the representative user is used as the label of the cycling behavior category to construct a cycling portrait of the cycling users in that category.
8. A motorcycle user riding portrait construction system based on time series features, used to implement the motorcycle user riding portrait construction method based on time series features according to any one of claims 1 to 7, characterized in that: include: Data acquisition module, time series feature extraction module, time series feature modeling module, user portrait construction module, user portrait storage module; The data acquisition module is used to obtain key information data of user riding from the database and clean the data; The time series feature extraction module is used to group the cleaned user riding key information data and sort them in time series, as well as to extract preliminary features to obtain a user riding feature sequence; The temporal feature modeling module is used to model and analyze the user's riding feature sequence using the riding behavior model, and output a riding representation vector containing the user's riding habits and style; The user portrait building module is used to classify users and build user portraits according to the similarity of riding representation vectors; The module first uses the built-in clustering model to cluster the user's riding representation vectors, and groups users with high riding similarity into one category; Based on the clustering results, the user closest to the cluster center vector in each category is taken as the representative user of the category, and a riding characteristic curve is generated based on its riding characteristic sequence. The user riding label is customized based on the riding characteristics of the representative user; the riding label of the representative user will also be used as the label of the category; The user portrait storage module is used to store the riding categories output by the user portrait construction module, where the categories represent users, all users in the categories, and category labels.