A self-driven identification method for passenger travel intention based on dynamic state augmentation network

Abstract

The present invention relates to the fields of intelligent rail transit technology and spatiotemporal data mining, and more specifically to a self-driven identification method for passenger travel intentions based on a dynamic state augmentation network. This method is used to deeply identify passengers' potential travel intentions and behavioral patterns from massive, anonymous rail transit travel trajectories. The core of this method is to construct a dynamic state augmented rail topology network. Based on this, multi-dimensional and interpretable behavioral indicators are extracted in an unsupervised manner. Micro-decision preference indicators in tensor flow are innovatively introduced, including a composite transfer structure indicator, destination and route selection concentration, temporal rhythmicity indicators, and a transfer route preference index, to accurately characterize passenger intentions.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent rail transit technology and spatiotemporal data mining, and more specifically, to a self-driven identification method for passenger travel intention based on a dynamic state augmentation network. Background Art

[0002] Urban rail transit systems are the arteries of modern cities. Refined mining of passenger travel data is key to achieving intelligent operations, optimizing resource allocation, and improving service quality. Existing passenger intent recognition methods mostly rely on statistics and traditional machine learning clustering or classification models, but they generally suffer from the following insurmountable flaws:

[0003] 1. Dependence on external labels and privacy risks: This approach relies heavily on personal attributes such as the user's occupation and age, and may require the integration of auxiliary data such as mobile phone signaling. This limits its application in the context of increasingly stringent data desensitization and privacy protection.

[0004] 2. Distortion of topological structure and micro-behavior: Complex rail networks are generally simplified into homogeneous "station-section" diagrams, or directly degenerated into origin-destination (OD) pairs, completely ignoring key micro-behaviors such as passengers' transfer decisions and route selection preferences within the station, resulting in serious distortion of behavioral information.

[0005] 3. Static modeling ignores environmental dynamics: Using static network parameters (such as fixed operating hours and transfer times) cannot reflect the time-varying characteristics of the rail system during peak and off-peak periods, and under normal and abnormal conditions, and is seriously out of touch with the actual traffic flow conversion characteristics.

[0006] 4. Poor model interpretability: Although methods such as deep learning have impressive performance, their "black box" nature makes their decision-making process opaque and cannot provide operations managers with intuitive and reliable decision-making basis.

[0007] Based on the above analysis, there is an urgent need for an unsupervised passenger intention recognition method that does not rely on external labels, is interpretable, and can integrately model the rail transit topology, passenger micro-decision-making behavior and system dynamic characteristics, so as to break through the technical bottlenecks of traditional methods in behavior restoration and generalized deployment. Summary of the Invention

[0008] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a self-driving identification method for passenger travel intentions based on the Dynamic State-Augmented Topology Network (DSATN). Through trajectory mapping modeling, self-driving behavior indicator extraction and comparative learning embedding, it realizes label-free and interpretable passenger behavior profiling and intention recognition.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A self-driven identification method for passenger travel intention based on a dynamic state augmentation network is described. The process of the method is as follows:

[0011] Step 1: Construct a dynamic state augmented track topology network.

[0012] Abstract the rail system as a state network that changes over time ,in, is the state node set; It is an edge set used to connect any two reachable state nodes, representing the possibility and cost of riding or transferring behavior, with time slicing Related attribute changes;

[0013] According to the rail transit line structure, first generate the state node set , each node is defined as a two-tuple , which represents the status of a physical station on a specific line. For all physical stations, lines, and all transfer stations, corresponding status nodes are constructed accordingly.

[0014] Furthermore, two new types of network connections (edge ​​sets) are added according to the structure table: one is the ride connection, which is the connection between related status nodes of physically adjacent stations on the same line; the other is the transfer connection, which is the connection between different line status nodes at the same physical station.

[0015] All connections are assigned a time attribute: based on historical operational data and business division, time slices are created. The travel time (travel_time(t)) (corresponding to boarding connections) and transfer time (transfer_time(t)) of each connection in different time periods are counted and assigned values, thus achieving the dynamic and spatiotemporal adaptability of DSATN.

[0016] Step 2: Passenger travel tensor flow mapping and construction.

[0017] For each passenger, their original travel records within a specific period are mapped into the above-mentioned DSATN to form a complete path sequence consisting of state nodes. The passenger travel tensor flow is constructed according to the three dimensions of [time slice × state path sequence × multi-dimensional behavior vector].

[0018] Step 3: Extraction of self-driven multi-dimensional behavioral indicator system.

[0019] This paper designs and calculates a set of behavioral indicators that can be directly obtained from the original trajectory data. This system not only includes traditional macro indicators, but also innovatively introduces micro decision preference indicators from tensor flow: including transfer structural load index (TSLI), destination and choice local concentration (DCLC), regularity score (RS), and transfer path preference index (TPI).

[0020] Step 4: Intent recognition and label generation based on state sequence contrast learning.

[0021] S41, State Sequence Encoding: Map the passenger's travel trajectory into a state node sequence and generate a trajectory representation vector through the embedding layer and sequence modeling network ;

[0022] S42, positive and negative sample construction: Positive samples are similar trips of the same passenger or trajectories of the same OD on different days, while negative samples are trajectories of different passengers, different ODs, and with significantly different paths;

[0023] S43, contrast learning objective function: The present invention adopts NT-Xent (Normalized Temperature-scaled Cross Entropy) loss function as the optimization objective; for an anchor point sample , the positive sample is , and within a batch negative samples , the loss function is defined as:

[0024] ;

[0025] in, Represents the cosine similarity function, which calculates the similarity of two vectors in direction, with a range of ; represents the temperature coefficient, which is a positive scalar hyperparameter;

[0026] S44, Behavior Pattern Clustering and Label Generation: Perform unsupervised clustering on all passenger trajectory embedding vectors to automatically discover behavior pattern clusters and automatically generate descriptive intent labels based on metrics such as TSLI, DCLC, RS, and TPI.

[0027] S45, Confidence and Explanation Output: Output clustering probability, typical trajectory of cluster center, mean value of main behavioral indicators and their explanation;

[0028] Furthermore, the calculation method of the transfer structure composite index is as follows:

[0029] Set passenger The state sequence of the next trip is , define the transfer point set as , then The TSLI for this trip is:

[0030] ;

[0031] in, is the conversion coefficient, is the total number of state nodes.

[0032] Furthermore, the transfer coefficient The setting is used to perform graded adjustment and index smoothing of the structural load of different types of transfers. It is divided into three levels, and the specific values ​​are as follows:

[0033] First-level transfer (similar to platform transfer, but takes less time): ;

[0034] Second-level transfer (if you need to walk a long distance and it takes a medium time): ;

[0035] Three-level transfer (if you need to exit and then re-enter the station, it will take a long time): .

[0036] Furthermore, the destination and route selection concentration is based on entropy to quantify the stability of passengers in their destination and route selection. It is specifically divided into destination concentration and route selection concentration:

[0037] If the set of all the terminal stations of passengers in the cycle is , is the number of terminal stations, and the probability of each terminal station appearing is , the destination concentration calculation formula is:

[0038] ;

[0039] If the set of all complete paths is , is the number of complete paths, and the probability of each path appearing is , the calculation formula of path selection centrality is:

[0040] .

[0041] Furthermore, the temporal rhythmicity index is used to measure the regularity of passengers’ departure times on high-frequency routes:

[0042] If the departure time series of passengers on the high-frequency route is , the data type of the departure time series is timestamp, and the calculation formula of RS is:

[0043] ;

[0044] Among them, RSstd is the standard deviation of the departure time series; RSacf is the periodic autocorrelation coefficient; is the weight, and its value range is [0,1].

[0045] Furthermore, the transfer route preference index is calculated by the proportion of the most commonly used transfer mode:

[0046] ;

[0047] Among them, c is a specific transfer type; is the number of its occurrences, is the total number of transfers; the TPI range is .

[0048] Furthermore, the specific method of constructing positive and negative samples is as follows:

[0049] Positive samples :

[0050] Method 1: For two trips of the same passenger at different times, if the normalized edit distance of their state sequences is less than a preset threshold , they are considered to be similar trips and are positive samples of each other;

[0051] Method 2: Randomly perform small perturbations on the anchor point trajectory sequence to generate positive samples;

[0052] Negative samples :

[0053] Method 1: Randomly sample other passengers’ trajectories from the training batch as negative samples;

[0054] Method 2: For the same passenger, if the normalized edit distance between the state sequences of two trips is greater than the preset threshold , or their OD pairs are completely different, they are considered as trips with large differences and can be used as negative samples.

[0055] Furthermore, the confidence and interpretability outputs include:

[0056] Cluster belonging probability: For any new travel trajectory, calculate the distance between its representation vector and each cluster center, and convert the distance into the probability of belonging to each cluster through the Softmax function;

[0057] Typical trajectories of cluster centers: For each cluster, find one or more true trajectories within it that are closest to the cluster centroid as the "prototype" or "example" of the pattern;

[0058] Profiling and interpretation of main behavioral indicators: The present invention automatically generates labels by establishing a "indicator-pattern" mapping rule library.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. Accurate characterization of micro-behavior: Through the innovative design of "state augmentation nodes (station-route pairs)" and "multimodal edges (ride / transfer connections)", passenger transfer decisions are transformed from implicit behaviors to explicit path characteristics. Leveraging micro-indicators such as the transfer structure composite index and the transfer path preference index, key behaviors such as transfer costs and path selection preferences are accurately quantified. This overcomes the behavioral distortion caused by the simplified "station-interval" modeling of traditional models, making the passenger micro-decision-making process perceptible and quantifiable, significantly improving the granularity and accuracy of the behavioral model.

[0061] 2. High fidelity of spatiotemporal dynamics: The rail network is innovatively constructed as a stateful network that changes dynamically over time slices. By discretely partitioning the time dimension, network edge attributes (running time, transfer time) can reflect dynamic characteristics such as system congestion and time period differences in real time. Compared with static network models, this method is more closely aligned with the actual operating patterns of rail transit systems, ensuring that intent recognition results are highly consistent with actual passenger behavior, and improving adaptability in dynamic scenarios.

[0062] 3. Excellent explainability and insight: An innovative self-driven multi-dimensional behavioral indicator system, combined with unsupervised clustering and automatic label generation mechanisms, can directly output concrete intent labels. At the same time, through clustering probability, typical trajectory examples, and indicator mean interpretation, abstract model outputs are transformed into understandable decision-making basis for operators, solving the "black box" problem of traditional deep learning.

[0063] 4. Unsupervised and highly universal: The system relies solely on the structure of anonymous trajectory data, eliminating the need for privacy labels such as user occupation and age, or manual annotation of information. This mitigates data privacy risks while reducing reliance on external data. This feature enables direct adaptation to rail transit networks of varying sizes in different cities, eliminating the need for secondary development for specific scenarios. Its strong universality across systems, time, and space significantly enhances the feasibility and scalability of technology implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flowchart of a self-driven identification method for passenger travel intention based on dynamic state augmentation network;

[0065] Figure 2 To calculate the destination and path selection concentration, confirm Experimental diagram of the values. DETAILED DESCRIPTION

[0066] Example, see Figure 1 The present embodiment provides a self-driven identification method for passenger travel intention based on a dynamic state augmentation network, specifically comprising the following steps:

[0067] Step 1: Construct a dynamic state augmented track topology network (DSATN).

[0068] In order to support the structured modeling and path semantic expression of rail transit travel behavior, this paper abstracts the rail system into a state network that changes over time. ,in, is a set of state nodes, each node represents the state of a passenger at a certain station and the line he is on; It is an edge set used to connect any two reachable state nodes, representing the possibility and cost of riding or transferring behavior, with time slicing Related attribute changes;

[0069] S11. Status node construction method:

[0070] Each state node is defined as a two-tuple ,in, Indicates the physical station number in the rail transit system (such as "People's Square"); Indicates the rail line number where the station is located (such as "Line 2"); if a station supports multiple line transfers, such as , then multiple state nodes are generated accordingly: (People's Square, Line 1), (People's Square, Line 2), (People's Square, Line 8); state node set The structured relationship table from the station-line ensures its physical and semantic interpretation in the graph;

[0071] S12, edge construction method:

[0072] Edges represent the connection relationship in the network, representing the legal behavior path of passengers migrating from one state to another. Each connection has a time-varying attribute, which is used to characterize its operation or transfer status in different time slices, forming a time-aware state network sequence. To adapt to the differences in peak and flat operating conditions, edges are divided into two categories:

[0073] S121, Bus Connection:

[0074] Connects the status nodes corresponding to adjacent stations on the same line; the data source is the line structure sequence table (such as GTFS data); the edge attribute is travel_time(t): which represents the travel time of the section in different time periods; the format is: (Nanjing West Road, Line 2) (People's Square, Line 2);

[0075] S122, transfer connection:

[0076] Connects status nodes belonging to different lines at the same physical station; the data source is the transfer station structure in the data table; the edge attribute is transfer_time(t)=base_walk_time+wait_time(t); the format is: (People's Square, Line 1) (People's Square, Line 2);

[0077] S13. Network construction logic:

[0078] State nodes and edges are defined independently in data and structurally combined to form a state network, which is endowed with path semantics and behavioral attributes. This structure provides the following advantages for rail network modeling:

[0079] a. Clearly distinguish between riding and transfer behaviors to improve route interpretability;

[0080] b. The graph structure is operational and can be used for tasks such as path planning, behavior analysis, and graph embedding;

[0081] c. Edge attributes can be used for indicator modeling (such as transfer load) to facilitate the identification of high-frequency patterns and intention types.

[0082] Step 2: Passenger travel tensor flow mapping and construction.

[0083] For each passenger, their original travel records within a specific period are mapped into the above-mentioned DSATN to form a complete path sequence consisting of "state nodes". According to the three dimensions of [time slice × state path sequence × multi-dimensional behavior vector], a passenger travel tensor flow is constructed to fully characterize their spatiotemporal behavior.

[0084] In order to reflect the dynamic differences of passenger travel behavior in different time periods, the present invention discretely divides the time dimension during the tensor construction process to generate a unified set of time slices. ,in, Indicates the number of slices. The present invention adopts the following fixed partitioning strategy:

[0085] The time slice granularity is set based on the operating day. As shown in Table 1, the full-day operating period is divided into four fixed time periods according to the business period standard:

[0086] Table 1

[0087]

[0088] State network time update mechanism:

[0089] State network constructed by the present invention In the network structure, the network structure has dynamic characteristics at different time slices: the node set Represents all valid state nodes (station-line pairs); edge set Indicates time slice The effective boarding connections and transfer connections are obtained; the attribute values ​​of all edges are counted and assigned according to historical data within the corresponding time period; each time slice corresponds to a set of independent network connection states, ensuring that the state network can accurately reflect the system's operating rules and passenger behavior characteristics within the time period.

[0090] Trajectory mapping and time discretization:

[0091] The multiple trajectories of each passenger in a specific statistical period are sequentially mapped to the DSATN under the corresponding time slice to form a state path sequence:

[0092] ;

[0093] in, Indicates the number of paths;

[0094] The time dimension is divided according to the above time periods, and the final travel behavior tensor structure is:

[0095] ;

[0096] in, Indicates the number of time slices; represents the number of passenger travel paths; Indicates the behavioral feature dimension. In this embodiment ; Each element of the tensor records the passenger in the time slice The travel paths under the above conditions and their corresponding multidimensional behavior indicators.

[0097] Step 3: Extraction of self-driven multi-dimensional behavioral indicator system.

[0098] This paper designs and calculates a set of behavioral indicators that can be directly obtained from the original trajectory data. Based on the tensor flow, a new self-driven indicator system that can deeply characterize the inherent behavior patterns of passengers is designed and calculated. This system not only includes traditional macro indicators, but also innovatively introduces micro decision preference indicators from tensor flow:

[0099] S31. Transfer structure composite index: measures the structural complexity density caused by frequent transfers in rail transit travel. This index can effectively reflect the density of transfer behavior in the passenger path and is an important parameter for describing the micro-decision-making characteristics of passengers. The state sequence of the next trip is , define the transfer point set as , then The TSLI for this trip is:

[0100] ;

[0101] in, is the conversion coefficient, is the total number of state nodes; The setting is used to perform graded adjustment and index smoothing of the structural loads of different types of transfers. In order to ensure the significant distinction of the structural loads under different transfer scenarios, the present invention will It is divided into three levels, and the specific values ​​are as follows:

[0102] First-level transfer (similar to platform transfer, but takes less time): ;

[0103] Second-level transfer (if you need to walk a long distance and it takes a medium time): ;

[0104] Three-level transfer (if you need to exit and then re-enter the station, it will take a long time): ;

[0105] The introduction of TSLI ensures the comparability and rationality of the indicators in different transfer scenarios, can reflect the differences in transfer difficulty in the actual rail transit network, and smoothes the indicators to improve their applicability in cross-network and cross-temporal comparisons.

[0106] Boundary condition processing:

[0107] If there is no transfer behavior in the passenger's travel trajectory, the transfer point set , then define:

[0108] ;

[0109] This definition means that the path does not require transfers and the structural load is minimal;

[0110] More specifically, suppose a passenger The state sequence of secondary rail transit trips is: In this state sequence, the passenger departs from Station A on Line 1, passes Station B on Line 1, arrives at Station B on Line 2 (transfer occurs at Station B), then arrives at Station C on Line 2, and finally arrives at Station D on Line 10. According to the transfer point determination rule of the present invention, if two consecutive state nodes in the state sequence have the same station but different lines, it is considered a transfer. Therefore, the transfer point set of this trip is for , number of transfers 1; total number of status nodes Defined as the number of state nodes for this trip, in this example n3=5;

[0111] Assuming that Station B is a secondary transfer station, the TSLI for this trip is calculated as follows:

[0112] ;

[0113] The results show that the transfer structure load of passengers in this trip is low and the path structure is relatively simple. All parameters of this indicator are directly obtained from the original trajectory data, which is highly objective and operational.

[0114] If there is no transfer behavior in the passenger's travel trajectory, the transfer point set Is an empty set, then the present invention defines , indicating that the route does not require transfer and the structural load is minimal.

[0115] S32, Destination and route selection concentration: Quantify the stability of passengers in destination and route selection based on entropy value;

[0116] Destination Convergence:

[0117] The set of all the terminal stations of passengers in the cycle is , is the number of terminal stations, and the probability of each terminal station appearing is , the destination concentration calculation formula is:

[0118] ;

[0119] Boundary condition description: If there is only one terminal station in a cycle, then , indicating that destination choice is highly concentrated;

[0120] Path Choice Convergence:

[0121] The set of all complete paths is , is the number of complete paths, and the probability of each path appearing is , the calculation formula of path selection centrality is:

[0122] ;

[0123] Boundary condition description: If there is only one path in a cycle, then , indicating that path selection is highly concentrated;

[0124] Furthermore, the DCLC total index calculation formula is as follows:

[0125] ;

[0126] In the present invention The value of is 0.7, which is the optimal result obtained through experiments. Specifically, the discrimination score is used to measure the DCLC indicator's ability to distinguish between different passenger behavior patterns. This indicator can effectively reflect the distribution differences of DCLC among different passenger groups. The larger the value, the stronger the DCLC indicator's ability to distinguish different behavior patterns. Figure 2 As shown, when When it is 0.7, the discrimination ability of DCLC reaches the highest.

[0127] S33, Regularity Score (RS): This is used to measure the regularity of passengers’ departure times on high-frequency routes. Specifically, all the travel routes of passengers during the statistical period are counted, and the routes with a frequency of occurrence exceeding 80% of the total number of trips are defined as high-frequency routes. The departure time series of passengers on high-frequency routes is: The data type of the departure time series is a timestamp, which can be in the "hour:minute" format (such as 07:10) or converted to the number of minutes since midnight (such as 430 minutes) to facilitate subsequent statistical analysis and calculations;

[0128] The calculation formula for RS is:

[0129] ;

[0130] Among them, RSstd is the standard deviation of the departure time series, which reflects the concentration of the departure time distribution; RSacf is the periodic autocorrelation coefficient, which reflects the periodic regularity of the departure time; is a weight used to adjust the relative importance of the standard deviation and the autocorrelation coefficient in the RS index, and its value range is [0,1]. In this embodiment, the value is 0.5. The value is based on the following:

[0131] The regularity of departure time is reflected in both the concentration of time distribution (standard deviation) and the intensity of periodic repetition (autocorrelation coefficient). To ensure the balance of influence of the two in the indicators, It is set to 0.5, that is, the standard deviation and autocorrelation coefficient each have a weight of 50%; this setting can ensure that the RS indicator has good discrimination and explanatory power under different passenger behavior patterns, and is suitable for most behavioral analysis scenarios of urban rail transit networks.

[0132] More specifically, suppose a passenger's departure time series on a high-frequency route is {07:10, 07:12, 07:09, 07:11, 07:13}, which is expressed in minutes as {430, 432, 429, 431, 433}; RSstd = 1.58 minutes, RSacf = 0.92 (calculated based on a one-week cycle), , substituting into the formula we get:

[0133] ;

[0134] Judgment basis: the smaller the standard deviation, the more concentrated the departure time, and the stronger the regularity; the larger the periodic autocorrelation coefficient, the closer the departure time is to the periodicity, and the stronger the regularity; the lower the RS value, the higher the regularity.

[0135] S34. Transfer Preference Index (TPI): The percentage of the most commonly used transfer method. The higher the TPI, the stronger the preference. The calculation formula is as follows:

[0136] ;

[0137] Where c is a specific transfer type (e.g., "change from Line 1 to Line 2 at People's Square Station"), is the number of its occurrences, is the total number of transfers; the TPI range is ; It means that all the passengers' transfers are of the same type, and their preference is very strong; the closer the TPI value is, the higher the preference is. ( is the total number of transfer types), indicating a weaker preference;

[0138] Step 4: Intent recognition and label generation based on state sequence contrast learning.

[0139] S41, State Sequence Encoding: Map the passenger's travel trajectory into a state node sequence and generate a trajectory representation vector through the embedding layer and sequence modeling network (Transformer Encoder) The encoder consists of two parts:

[0140] Embedding layer: the trained state node Mapped into a continuous vector;

[0141] Sequence Modeling Network: This model can capture the long-range dependencies between nodes in the sequence, such as the global structure of the path, and generate the final trajectory representation vector ;

[0142] S42, positive and negative sample construction: Positive samples are similar trips of the same passenger or trajectories of the same OD on different days, while negative samples are trajectories of different passengers, different ODs, and with significantly different paths;

[0143] Specifically, the positive sample (Similar trips):

[0144] Method 1 (same passenger, similar route): For two trips of the same passenger occurring at different times, if the normalized edit distance of their state sequences is less than a preset threshold (0.1), they are considered similar and are positive samples of each other. The edit distance calculates the minimum number of insertions, deletions, and substitutions required to transform one sequence into another;

[0145] Method 2 (data augmentation): Randomly perform small perturbations on the anchor point trajectory sequence (such as randomly masking a non-critical intermediate station) to generate positive samples;

[0146] Negative samples (Travel with big differences):

[0147] Method 1 (different passengers): Randomly sample the trajectories of other passengers from the training batch as negative samples;

[0148] Method 2 (same passenger, different routes): For the same passenger, if the normalized edit distance between the state sequences of the two trips is greater than the preset threshold , or their OD pairs are completely different, they are considered as trips with large differences and can be used as negative samples;

[0149] S43, contrast learning objective function: This invention adopts NT-Xent (Normalized Temperature-scaled Cross Entropy) loss function as the optimization objective. , the positive sample is , and within a batch negative samples , the loss function is defined as:

[0150] ;

[0151] in, Represents the cosine similarity function, which calculates the similarity of two vectors in direction, with a range of ; Represents the temperature coefficient, which is a positive scalar hyperparameter. Its function is to adjust the model's discrimination of negative samples. A smaller It will widen the distance between samples of different categories, making the model learn more discriminative features; in this embodiment, through cross-validation, The empirical value range of , the typical value is 0.1;

[0152] S44, Behavior Pattern Clustering and Label Generation: Perform unsupervised clustering on all passenger trajectory embedding vectors to automatically discover behavior pattern clusters and automatically generate descriptive intent labels based on metrics such as TSLI, DCLC, RS, and TPI.

[0153] Specifically, after the model training converges, all the travel trajectories of all passengers are input into the encoder to obtain their respective representation vectors ;

[0154] Clustering algorithm: Characterization vector of all trajectories The K-means clustering algorithm is used for unsupervised clustering of the dataset. The number of clusters, K, is automatically determined using the silhouette coefficient analysis method. The average silhouette coefficient is calculated for different K values ​​(e.g., 2 to 20), and the K that maximizes this coefficient is selected as the final number of clusters. The silhouette coefficient measures both the compactness within a cluster and the separation between clusters, making it an effective method for determining the value of K.

[0155] Automatic label generation: After clustering is complete, each cluster represents a unique behavioral pattern. This method calculates the mean and variance of the behavioral indicators (TSLI, DCLC, RS, TPI) of all trajectories within each cluster and automatically generates an interpretable intent label for each cluster based on pre-set rules.

[0156] S45, Confidence and Explanability Output: Output clustering probability, typical trajectory of cluster centers, mean values ​​of main behavioral indicators and their explanations. Providing transparent and reliable decision support for operators, the output of this invention includes:

[0157] Cluster belonging probability: For any new travel trajectory, the distance between its representation vector and each cluster center (centroid) is calculated, and the distance is converted into the probability of belonging to each cluster through the Softmax function; this provides the confidence level of the model decision;

[0158] Typical trajectories of cluster centers: For each cluster, find one or more real trajectories within it that are closest to the cluster centroid as the "prototype" or "example" of the pattern for intuitive understanding;

[0159] Key behavioral indicator profiling and interpretation: This is the key to achieving interpretability. This invention automatically generates labels by establishing a "indicator-pattern" mapping rule base. The details are shown in Table 2:

[0160] Table 2

[0161]

[0162] Through the detailed introduction of the above embodiments, the present invention provides a self-driven identification method for passenger travel intentions based on a dynamic state augmentation network, which is used to identify passengers' potential travel intentions and behavior patterns from massive anonymous rail transit travel trajectories. Through the innovative design of "state augmentation nodes" and "multimodal edges", with the help of micro-indicators such as transfer structure composite indicators and transfer path preference index, key behaviors such as transfer costs and path selection preferences are accurately quantified, breaking through the behavioral distortion problem caused by the simplified modeling of "station-interval" by traditional models, making the passenger's micro-decision-making process perceptible and quantifiable. At the same time, an innovative self-driven multi-dimensional behavior indicator system is constructed, combined with unsupervised clustering and automatic label generation mechanism, which can directly output concrete intention labels.

[0163] The above formulas are all dimensionless and numerically calculated, and the preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0164] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0165] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0166] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0167] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0168] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0169] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0170] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A self-driven identification method for passenger travel intention based on dynamic state augmentation network, characterized by: The method flow is as follows: Step 1: Construct a dynamic state augmented track topology network: abstract the track system into a state network that changes over time; Step 2: Passenger travel tensor flow mapping and construction: For each passenger, their original travel records within a specific period are mapped to the dynamic state augmented track topology network to form a complete path sequence consisting of state nodes; the passenger travel tensor flow is constructed according to the three dimensions of [time slice × state path sequence × multi-dimensional behavior vector]; Step 3: Extraction of a self-driven multi-dimensional behavior indicator system: Directly obtain behavioral indicators from the original trajectory data. These indicators include not only traditional macro indicators but also micro-decision preference indicators from TensorFlow. These micro-decision preference indicators include a transfer structure composite indicator, destination and route selection concentration, time rhythmicity indicators, and a transfer route preference index. The calculation method of the transfer structure composite index is as follows: Set passenger The state sequence of the next trip is , define the transfer point set as , Indicates the physical station number in the rail transit system. Indicates the track line number where the station is located. The TSLI for this trip is: ； in, is the conversion coefficient, is the total number of state nodes; Destination and route selection concentration is specifically divided into destination concentration and route selection concentration: If the set of all the terminal stations of passengers in the cycle is , is the number of terminal stations, and the probability of each terminal station appearing is , the destination concentration calculation formula is: ； If the set of all complete paths is , is the number of complete paths, and the probability of each path appearing is , the calculation formula of path selection centrality is: ； Time rhythmicity index, RS is calculated based on the departure time series of passengers on high-frequency routes: ； Among them, RSstd is the standard deviation of the departure time series; RSacf is the periodic autocorrelation coefficient; is the weight, the value range is [0,1]; Step 4: Intent recognition and label generation based on state sequence contrastive learning: Map the passenger's travel trajectory into a state node sequence and generate a trajectory representation vector; use the NT-Xent loss function as the optimization objective; define the loss function based on the anchor point samples and constructed positive and negative samples; perform unsupervised clustering on the trajectory embedding vectors of all passengers, and automatically generate descriptive intent labels based on micro-decision preference indicators; finally, output confidence and interpretability.

2. The self-driven identification method of passenger travel intention based on dynamic state augmentation network according to claim 1 is characterized in that: Transfer coefficient The specific values ​​are as follows: First-level transfer, ; Secondary transfer, ; Level 3 transfer, .

3. The self-driven identification method of passenger travel intention based on dynamic state augmentation network according to claim 1 is characterized in that: The transfer route preference index is calculated by the proportion of the most commonly used transfer mode: ； Among them, c is a specific transfer type; is the number of its occurrences, is the total number of transfers; the TPI range is .

4. The self-driven identification method of passenger travel intention based on dynamic state augmentation network according to claim 1 is characterized in that: The specific method of constructing positive and negative samples is as follows: For positive samples, Method 1: Two trips of the same passenger occurring at different times are considered similar if the normalized edit distance of their state sequences is less than a preset threshold, and are mutually positive samples. Method 2: Randomly perturb the anchor point trajectory sequence to generate positive samples. For negative samples, Method 1: randomly select other passengers' trajectories from the training batch as negative samples; Method 2: for the same passenger, if the normalized edit distance of the state sequences of two trips is greater than the preset threshold, or their OD pairs are completely different, they are regarded as trips with large differences and used as negative samples.

5. The self-driven identification method of passenger travel intention based on dynamic state augmentation network according to claim 1 is characterized in that: The output confidence and interpretability are as follows: Cluster belonging probability: For any new travel trajectory, calculate the distance between its representation vector and each cluster center, and convert the distance into the probability of belonging to each cluster through the Softmax function; Typical trajectories of cluster centers: For each cluster, find one or more real trajectories within it that are closest to the cluster centroid as the "prototype" or "example" of the trajectory; Profiling and interpretation of key behavioral indicators: Automatically generate labels by establishing a "indicator-pattern" mapping rule library.

Images