Multi-modal data fusion crowd travel intelligent analysis method based on large model
By fusing multimodal data based on embedding technology and LSTM algorithm, the accuracy and efficiency issues of user travel mode prediction in existing technologies are solved, achieving efficient and secure user travel prediction, which is applicable to data analysis in fields such as transportation, culture and tourism, retail and finance.
Patent Information
- Application Number
- CN202511118785.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-11
- Publication Date
- 2025-11-14
AI Technical Summary
Existing methods for predicting user travel patterns suffer from low data accuracy, high data latency, high computational resource consumption, and privacy issues, making it difficult to achieve efficient and accurate predictions of user travel patterns.
By employing embedding technology combined with operator location trajectory data, the raw data is transformed into meaningful low-dimensional features. The model is trained through classification algorithms, combined with temporal and spatial representation learning, and regularization techniques are used to prevent overfitting. Finally, the LSTM algorithm is used to predict user travel.
It enables efficient and accurate prediction of user travel patterns, reduces computing resource requirements, improves data timeliness, and protects user privacy. It is suitable for data analysis in industries such as transportation, culture and tourism, retail, and finance.
Smart Images
Figure CN120952251A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of traffic group travel prediction methods, specifically involving an intelligent analysis method for population travel based on multimodal data fusion using a large model. Background Technology
[0002] Currently, with urban development, transportation has become more convenient and faster, leading to a surge in travelers. This places immense pressure on passenger flow management and safety assurance at major transportation hubs, especially during holidays. To address this potential hazard, it is necessary to anticipate user travel patterns in advance, plan the scheduling of various modes of transportation, and analyze and process population travel trajectories using big data technology. Combining a large volume of user travel trajectories with machine learning can predict user travel patterns. Commonly used travel pattern prediction methods include the following: 1. Government Transportation Survey Questionnaire: By distributing questionnaires to a large number of people, we collect information on the modes of transportation they choose to leave after arriving at major transportation hubs, as well as the modes of transportation they choose to travel to major transportation hubs. Then, by organizing and storing the data from a large number of questionnaires, we perform statistical analysis after data expansion, weighting, and verification to establish a predictive model for user travel modes.
[0003] 2. Data Analysis: Using traditional data analysis techniques and combining various transportation travel rules, statistical analysis is conducted on the travel population based on indicators such as OD (origin and destination) + speed + distance + regional stations to establish a predictive model for user travel modes.
[0004] Existing methods for predicting user travel patterns have shown some effectiveness in scenarios such as transportation capacity resource scheduling, route advancement, and route planning, but they also have drawbacks and limitations: 1. Low data accuracy: Data generated through a large number of questionnaires is one-sided, only applicable to travel modes at specific transportation hubs in a particular region or area, and cannot achieve comprehensiveness, resulting in low model accuracy.
[0005] 2. High data latency: Traditional data analysis methods require the collection of a large amount of user travel trajectory data. Generally, it is based on the user switching base stations to form base station zip data, and then the user's dwell status, travel distance and speed are calculated to further determine the user's travel mode. Finally, the data is summarized and statistically analyzed. The whole process involves massive data calculation and is done in a serial manner, which results in high latency of model output results, poor data timeliness and mediocre application effect.
[0006] 3. Privacy issues: Traditional data analysis techniques directly use all of a user's travel location data, which involves personal location privacy. Improper handling may involve the leakage of a large amount of user travel location information.
[0007] 4. High computational resource consumption: Predicting user travel patterns often involves a large amount of data, requiring a large amount of computational resources, which is very costly. Summary of the Invention
[0008] To address the technical problem of "low data accuracy in existing user travel mode prediction methods," and considering the shortcomings of existing technologies, this invention proposes to use embedding technology, combined with operator location trajectory data, to transform the raw data into meaningful low-dimensional features, construct multi-dimensional vector representation data, and train the model based on these representations using a classification algorithm. This reduces the demand for data and computing power, and through temporal and spatial representation learning, it can calculate, identify, and predict user travel modes.
[0009] This invention provides the following technical solution: A method for intelligent analysis of population travel based on multimodal data fusion using a large model includes the following steps: Step S1: Collect and label the true value data of users' historical travel trajectories. The collection methods include identification based on subway checkpoints, bus card readers, user APP content data combined with traditional methods, and volunteer data collection. Step S2: Based on the collected true data of the user's historical travel trajectory, perform time-series processing on all the base stations switched by the user from the start to the end of the trip to obtain the user's base station zipper data; combine the user's travel trajectory path reconstruction technology to fill in the base station zipper data and construct the full-chain base station data; Step S3: Based on the Chuanliu large model, the embedding technology is used to vectorize the discrete features in the full-chain base station data and normalize the continuous features to obtain multi-dimensional vector representation data of user travel. Step S4: Use regularization techniques to process the multidimensional vector representation data, and use the regularized multidimensional vector representation data as input to the time series algorithm to train and obtain the user travel prediction model. Step S5: Predict user travel patterns based on the user travel prediction model.
[0010] Further, in step S2, obtaining the user's base station zipper data includes: collecting base station data; obtaining the association data between the user and the base station based on the user's historical travel trajectory data and the base station data to form the user's base station zipper table data; and merging the user's historical travel trajectory data and the base station zipper table data to construct the user's travel time series base station zipper data.
[0011] Furthermore, in step S2, the user travel trajectory path reconstruction technology includes: based on the user's travel base station data, combined with neural network learning, using a time series model to learn the trajectory temporal features, predict missing segments, generate adversarial networks, and generate reasonable paths to fill data gaps.
[0012] Furthermore, in step S2, the user travel trajectory path reconstruction technology also includes: comparing the performance of different time series models through cross-validation, and using the time series model with the smallest error to train on complete base station data.
[0013] Furthermore, in step S2, the user travel trajectory path reconstruction technology also includes: finally, applying the trained time series model to predict missing values, and filling in the missing parts of the trajectory through an incremental sliding window method, thereby reconstructing the discontinuous trajectory into a continuous trajectory path.
[0014] Further, in step S2, multi-dimensional vector representation data of user travel is obtained, including: based on the Chuanliu travel big model, the embedding layer is defined as an Nx768 matrix, where N is the batch size and 768 is the dimension of the embedding vector. The embedding layer weights are initialized using random values. During the forward propagation of the Chuanliu travel big model, the corresponding 768-dimensional vector is extracted from the embedding layer according to the input category index. The weights of the embedding layer are updated through backpropagation so that similar categories are closer in the embedding space.
[0015] Furthermore, in step S4, regularization techniques are used to process the multidimensional vector representation data: L2 regularization is used to add an extra term to the loss function to limit the size of the user travel prediction model parameters in order to prevent overfitting.
[0016] Furthermore, L2 regularization adds the sum of squares of the user travel prediction model parameters to the loss function, as shown in the following formula:
[0017] J(θ): Regularized loss function; θ: Parameter; This is the regularization term, also known as the L2 regularization term; α: regularization coefficient; n is the number of parameters θ. i This represents the i-th parameter.
[0018] Furthermore, when applying regularization, the following steps are performed: Choose the regularization type: Select L1 or L2 regularization based on the nature of the problem and the characteristics of the data; Set the regularization coefficient: Select an appropriate regularization coefficient α through methods such as cross-validation; Modify the loss function: Add a regularization term to the original loss function; Optimize the modified loss function: Use gradient descent or other optimization algorithms to minimize the modified loss function.
[0019] Furthermore, step S4 also includes training and inferring the user travel prediction model using the LSTM algorithm based on multidimensional vector representation data, training the user travel prediction model using the training set, and evaluating it using the test set.
[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention introduces embedding technology, which is more efficient and accurate than traditional techniques. Combined with multi-dimensional vector representation data, it is safer and more efficient with shorter training computation time. Combined with trajectory path reconstruction technology, it fills in the spatiotemporal gaps of user travel trajectories and constructs the user's full-chain travel. Regularization technology effectively prevents model overfitting during training.
[0021] Therefore, the user travel pattern prediction method based on operator data fusion embedding technology has technical advantages such as high efficiency, high accuracy, low data latency, and strong performance. These advantages make this method highly adaptable and practical in various application scenarios.
[0022] (2) Based on embedding technology, massive spatiotemporal data of operators are transformed into a basic data base with general dynamic representation. Combined with trajectory path reconstruction technology and regularization technology, the basic representation is trained and predicted by LSTM algorithm, which empowers data analysis and mining tasks in key industries such as transportation, culture and tourism, retail and finance.
[0023] (3) Traffic management: It plays an important guiding role in real-time allocation of public transportation resources and the formulation of traffic control measures according to local conditions, and can effectively alleviate traffic congestion around key locations.
[0024] (4) Improve the connectivity of transportation hubs: Through comprehensive perception, prediction, deduction and evaluation of passenger flow at hubs, realize the scientific allocation of transportation capacity at transportation hubs and the timely release of travel information, quickly allocate transportation resources and accelerate the operation and turnover of hubs.
[0025] (5) The operation status of large passenger transport hubs, such as regional agglomeration passenger flow, inflow and outflow passenger flow, available transport capacity, inflow and outflow transport capacity, passenger flow direction, and dwell time, can realize multi-dimensional dynamic monitoring of passenger flow and transport capacity in the region. Attached Figure Description
[0026] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0027] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.
[0028] like Figure 1 As shown, this invention provides a method for intelligent analysis of population travel based on multimodal data fusion using a large model, comprising the following steps: Step S1: Collect and label the true value data of users' historical travel trajectories. The collection methods include identification based on subway checkpoints, bus card readers, user APP content data combined with traditional methods, and volunteer data collection. Step S2: Based on the collected true data of the user's historical travel trajectory, perform time-series processing on all the base stations switched by the user from the start to the end of the trip to obtain the user's base station zipper data; combine the user's travel trajectory path reconstruction technology to fill in the base station zipper data and construct the full-chain base station data; Step S3: Based on the Chuanliu large model, the embedding technology is used to vectorize the discrete features in the full-chain base station data and normalize the continuous features to obtain multi-dimensional vector representation data of user travel. Step S4: Use regularization techniques to process the multidimensional vector representation data, and use the regularized multidimensional vector representation data as input to the time series algorithm to train and obtain the user travel prediction model. Step S5: Predict user travel patterns based on the user travel prediction model.
[0029] Next, each step will be explained in detail: Step S1: Collect and label the true data of users' historical travel trajectories; 1. Truth Database Construction: By collecting user historical travel trajectory truth data from subway checkpoints, bus card readers, user APP content data, and traditional identification methods, as well as volunteer data collection, a truth database is constructed. The format is detailed in Table 1.
[0030] Table 1
[0031] The true value data of users' historical travel trajectory mainly consists of labeled user trajectory data. This user trajectory data may come from various channels, such as travel route information actively provided by users, trajectory data obtained through high-precision positioning devices, etc.
[0032] 2. Truth Data Labeling: The collected truth data of users' historical travel trajectories are labeled to mark the specific mode of each trip. Before labeling, the data is cleaned to remove or correct outliers and erroneous data.
[0033] The annotations include the user's origin and destination, key locations along the route, and mode of travel (walking, driving, public transportation, etc.), which are used for supervised learning during the subsequent training of the user travel prediction model to improve the accuracy of the user travel prediction model in judging the mode of travel.
[0034] Step S2: Based on the collected historical travel trajectory data of users, perform time-series processing on all base stations switched by the user from the start to the end of the trip to obtain the user's base station chain data; combine user travel trajectory path reconstruction technology to fill in the base station chain data and construct full-chain base station data; specifically, 1. Obtain user's base station zipper data: Collect base station data, and based on the user's historical travel trajectory data and base station data, obtain the association data between the user and the base station, that is, the sequence of base stations the user connected to at different times, forming the user's base station zipper table data. The base station zipper table data includes information such as user number, base station identifier, connection time, etc., which is used to record the user's switching between different base stations and reflect the user's movement trajectory.
[0035] The collected historical user travel trajectory data is fused with base station connection records: User travel trajectories are matched with corresponding base station connection records according to time series. The base stations switched during the user's trip from start to finish are time-series processed. Noise reduction and interpolation are performed on the base station data, and filtering algorithms (Kalman filtering, particle filtering) are used to eliminate base station drift, forming base station connection records for the user's travel time series. The format is detailed in Table 2. For example, based on the location information recorded in the user trajectory data at a certain time point, the base stations connected to the user at that time point are found, forming a comprehensive dataset (base station connection records) containing user trajectory and base station information.
[0036] Table 2
[0037] Collecting base station data involves gathering signal data generated by vehicles as they pass through mobile communication base stations deployed on roads. This data includes key information such as the user's phone number, the cell they are in, their CGI (Cell Global Identifier), and timestamps. This information is used to determine the location of the vehicle at different base stations at different times, thereby providing a preliminary estimate of the vehicle's location.
[0038] Base station chain data represents the association data between a user and a base station, that is, the sequence of base stations a user connects to at different times. Base station chain data includes information such as user number, base station identifier, and connection time, used to record the user's switching between different base stations and reflect the user's movement trajectory.
[0039] 2. Combining user travel trajectory reconstruction technology, spatiotemporal filling is performed on base station zipper data to construct full-chain base station data: When there are missing or incomplete true data of user historical travel trajectory, user travel trajectory reconstruction technology is used to restore the user's movement path.
[0040] The user travel trajectory path reconstruction technology specifically involves: based on base station data of user travel, combining neural network learning, using time series models such as LSTM and Transformer to learn trajectory temporal features, predict missing segments, generating a generative adversarial network (GAN) to generate reasonable paths to fill data gaps, comparing the performance of different time series models through cross-validation, and using the time series model with the smallest error to train on complete base station data; finally, applying the trained time series model to predict missing values, and filling in the missing parts of the trajectory through an incremental sliding window method, thereby reconstructing discontinuous trajectories into continuous trajectory paths.
[0041] Spatiotemporal completion constructs full-chain base station data to address potential spatiotemporal discontinuities in user trajectories. By analyzing user movement patterns and time intervals between adjacent base stations, gaps in user trajectories are filled in, making user trajectories more complete and continuous.
[0042] At the same time, a full-chain base station data is constructed, forming a continuous chain structure of the base stations connected to users in chronological order, so as to more accurately track users' travel paths and provide a higher quality data foundation for subsequent feature extraction and user travel prediction model training.
[0043] Step S3: Based on the large flow model, embedding technology is used to vectorize the discrete features in the full-chain base station data and normalize the continuous features to obtain multi-dimensional vector representation data of user travel; among which, Based on the Chuanliu large model, the embedding technology is used to vectorize the discrete features in the full-chain base station data and normalize the continuous features: (1) Vectorization of discrete features: The embedding technology is used to convert the original data into meaningful low-dimensional features such as base station ID and base station type in the full-chain base station data; (2) Normalization of continuous features: The continuous features such as time interval, distance between base stations, and speed in the full-chain base station data are normalized. The process of transforming the numerical range of continuous features to a uniform interval (usually [0, 1] or [-1, 1]). Here are two common normalization methods: 1) min-max normalization, which scales the feature values to the [0, 1] interval; and 2) standardization, which transforms the feature values into data with zero mean and unit variance.
[0044] Discrete feature embedding and continuous feature normalization involve embedding discrete features (such as base station identifiers and time period categories corresponding to timestamps) in user travel data, mapping them to a continuous vector space. This allows machine learning models to better process and understand these features while preserving the similarity and correlation between features. Normalization, on the other hand, maps continuous feature values from different ranges to the same scale, making different features comparable and consistent, thus improving the efficiency and accuracy of machine learning models in processing features.
[0045] Based on the large-scale model of Chuanliu Travel, its embedding layer is defined as an Nx768 matrix, where N is the batch size and 768 is the dimension of the embedding vector. The embedding layer weights are initialized with random values. During the forward propagation of the model, the corresponding 768-dimensional vector is extracted from the embedding layer according to the input category index. The weights of the embedding layer are updated through backpropagation, so that similar categories are closer in the embedding space.
[0046] The Chuanliu Travel Big Data Model adopts the Jiutian Chuanliu Travel Big Data Model, which is an artificial intelligence pre-trained model launched by China Mobile based on ultra-large-scale dynamic graph neural network technology.
[0047] This multidimensional vector representation data comprehensively and accurately describes users' travel patterns and characteristics, including information on users' movement patterns, preferred routes, and travel time distribution, providing rich feature inputs for subsequent user travel prediction model training.
[0048] Step S4: Use regularization techniques to process the multidimensional vector representation data, and use the regularized multidimensional vector representation data as input to the time series algorithm to train and obtain the user travel prediction model. 1. Using regularization techniques to process multi-dimensional vector representation data: Regularization techniques are used to treat overfitting during the training of the user travel prediction model. L2 regularization (Ridge) adds an extra term to the loss function to limit the magnitude of subsequent user travel prediction model parameters, thus preventing overfitting. L2 regularization adds the sum of squares of the user travel prediction model parameters to the loss function, as shown in the following formula:
[0049] J(θ): Regularized loss function; θ: Parameter; This is the regularization term, also known as the L2 regularization term; α: regularization coefficient; n is the number of parameters θ. i This represents the i-th parameter.
[0050] When applying regularization, perform the following steps: Set the regularization coefficient: Select an appropriate regularization coefficient α through methods such as cross-validation.
[0051] Modify the loss function: Add a regularization term to the original loss function.
[0052] Optimize the modified loss function: Use gradient descent or other optimization algorithms to minimize the modified loss function.
[0053] Regularization techniques can prevent user travel prediction models from overfitting to the multidimensional vector representation data used for training, thereby improving the generalization ability of user travel prediction models and enabling them to maintain good predictive performance when faced with new and unseen travel data.
[0054] 2. Using the regularized multidimensional vector representation data as input to a time series algorithm to train and obtain a user travel prediction model: The constructed multidimensional vector representation data of user travel is divided into training, validation, and test sets. A time series algorithm (such as a recurrent neural network, RNN) is used to model the multidimensional vector representation data of user travel to obtain a user travel prediction model. The user travel prediction model is trained, adjusted, and evaluated by constructing a recurrent neural network model. The time series algorithm can capture the time dependencies and sequence patterns in the multidimensional vector representation data.
[0055] Step S4 also includes training and inferring the user travel prediction model based on multidimensional vector representation data using the LSTM algorithm, training the user travel prediction model using the training set, evaluating it using the test set, and outputting evaluation metrics such as accuracy, recall, and F1 score.
[0056] During the training process of the user travel prediction model, the training set is used to optimize and learn the model's parameters. The model's fit on the training set is evaluated using metrics such as the loss function. The validation set is then used to validate and adjust the model, determining its hyperparameters and structure to prevent overfitting. Finally, the test set is used to comprehensively evaluate the trained model, calculating metrics such as accuracy, recall, and F1 score to assess the accuracy and reliability of the model's predictions of user travel modes. Based on the evaluation results, further optimization and adjustments are made to improve the model's performance.
[0057] Step S5: Predict user travel modes based on the user travel prediction model. Specifically, for new user base station handover records, construct multi-dimensional vector representation data, input it into the trained user travel prediction model, and predict the user's travel mode. The output is a multi-class probability distribution of the user's current travel mode, and the category with the highest probability is selected as the prediction result.
[0058] The trained user travel prediction model is deployed in real-world applications, inputting real-time base station snapshot data (including information about the user's current base station connection, timestamps, etc.). The model then infers and predicts the user's travel mode. Based on the user's current base station connection status and historical travel characteristics, the user travel prediction model can quickly determine whether the user is walking, driving, or using public transportation, providing decision support for subsequent traffic management and travel service applications.
[0059] The above technical features constitute the preferred embodiment of the present invention, which has strong adaptability and optimal implementation effect. Non-essential technical features can be added or removed according to actual needs to meet the needs of different situations.
[0060] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.
Claims
1. A method for intelligent analysis of population travel based on multimodal data fusion using a large model, characterized in that, Including the following steps: Step S1: Collect and label the true value data of users' historical travel trajectories. The collection methods include identification based on subway checkpoints, bus card readers, user APP content data combined with traditional methods, and volunteer data collection. Step S2: Based on the collected true data of the user's historical travel trajectory, perform time-series processing on all the base stations switched by the user from the start to the end of the trip to obtain the user's base station zipper data. By combining user travel trajectory reconstruction technology, base station zipper data is filled in to construct full-chain base station data; Step S3: Based on the Chuanliu large model, the discrete features in the full-chain base station data are vectorized and the continuous features are normalized using the embedding technology to obtain multi-dimensional vector representation data of user travel. Step S4: Use regularization techniques to process the multidimensional vector representation data, and use the regularized multidimensional vector representation data as input to the time series algorithm to train and obtain the user travel prediction model. Step S5: Predict user travel patterns based on the user travel prediction model.
2. The intelligent analysis method for population travel based on multimodal data fusion using a large model, as described in claim 1, is characterized in that... In step S2, obtaining the user's base station zipper data includes: collecting base station data; obtaining the association data between the user and the base station based on the user's historical travel trajectory data and the base station data to form the user's base station zipper table data; and merging the user's historical travel trajectory data and the base station zipper table data to construct the user's travel time series base station zipper data.
3. The intelligent analysis method for population travel based on multimodal data fusion using a large model, as described in claim 2, is characterized in that... In step S2, the user travel trajectory path reconstruction technology includes: based on the user's travel base station data, combined with neural network learning, using a time series model to learn the trajectory temporal features, predict missing segments, generate adversarial networks, and generate reasonable paths to fill data gaps.
4. The intelligent analysis method for population travel based on multimodal data fusion using a large model, as described in claim 3, is characterized in that... In step S2, the user travel trajectory path reconstruction technology also includes: comparing the performance of different time series models through cross-validation, and training the time series model with the smallest error on complete base station data.
5. The intelligent analysis method for population travel based on multimodal data fusion using a large model, as described in claim 4, is characterized in that... In step S2, the user travel trajectory path reconstruction technology also includes: finally, applying the trained time series model to predict missing values, and filling in the missing parts of the trajectory through an incremental sliding window method, thereby reconstructing the discontinuous trajectory into a continuous trajectory path.
6. The intelligent analysis method for population travel based on multimodal data fusion using a large model, as described in claim 1, is characterized in that... In step S2, multidimensional vector representation data of user travel is obtained, including: based on the Chuanliu travel big model, the embedding layer is defined as an Nx768 matrix, where N is the batch size and 768 is the dimension of the embedding vector. The embedding layer weights are initialized using random values. During the forward propagation of the Chuanliu travel big model, the corresponding 768-dimensional vector is extracted from the embedding layer according to the input category index. The weights of the embedding layer are updated through backpropagation so that similar categories are closer in the embedding space.
7. A method for intelligent analysis of population travel based on multimodal data fusion using a large model, as described in any one of claims 1-6, characterized in that... In step S4, regularization techniques are used to process the multidimensional vector representation data: L2 regularization is used to add an extra term to the loss function to limit the size of the parameters of the user travel prediction model in order to prevent overfitting.
8. The intelligent analysis method for crowd travel based on multimodal data fusion using a large model, as described in claim 7, is characterized in that... L2 regularization adds the sum of squares of the user travel prediction model parameters to the loss function, as shown in the following formula: J(θ): Regularized loss function; θ: Parameter; This is the regularization term, also known as the L2 regularization term; α: regularization coefficient; n is the number of parameters θ. i This represents the i-th parameter.
9. The intelligent analysis method for crowd travel based on multimodal data fusion using a large model, as described in claim 8, is characterized in that... When applying regularization, perform the following steps: Set the regularization coefficient: Choose an appropriate regularization coefficient α; Modify the loss function: Add a regularization term to the original loss function; Optimize the modified loss function: Use an optimization algorithm to minimize the modified loss function.
10. The intelligent analysis method for crowd travel based on multimodal data fusion using a large model, as described in claim 8, is characterized in that... Step S4 also includes training and inferring the user travel prediction model using the LSTM algorithm based on multidimensional vector representation data, training the user travel prediction model using the training set, and evaluating it using the test set.